Index: HelenOS.config =================================================================== --- HelenOS.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ HelenOS.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -399,4 +399,9 @@ ! [COMPILER=gcc_cross|COMPILER=gcc_native] CONFIG_LTO (n/y) +% Kernel RCU algorithm +@ "PREEMPT_PODZIMEK" Preemptible Podzimek-RCU +@ "PREEMPT_A" Preemptible A-RCU +! RCU (choice) + ## Hardware support Index: abi/include/syscall.h =================================================================== --- abi/include/syscall.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ abi/include/syscall.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -55,4 +55,5 @@ SYS_FUTEX_WAKEUP, SYS_SMC_COHERENCE, + SYS_SMP_MEMORY_BARRIER, SYS_AS_AREA_CREATE, Index: boot/Makefile.common =================================================================== --- boot/Makefile.common (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ boot/Makefile.common (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -174,4 +174,5 @@ $(USPACE_PATH)/app/mkexfat/mkexfat \ $(USPACE_PATH)/app/mkmfs/mkmfs \ + $(USPACE_PATH)/app/rcutest/rcutest \ $(USPACE_PATH)/app/sbi/sbi \ $(USPACE_PATH)/app/sportdmp/sportdmp \ Index: defaults/amd64/Makefile.config =================================================================== --- defaults/amd64/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/amd64/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -38,4 +38,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/arm32/Makefile.config =================================================================== --- defaults/arm32/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/arm32/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -29,4 +29,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # What is your input device? CONFIG_HID_IN = generic Index: defaults/ia32/Makefile.config =================================================================== --- defaults/ia32/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/ia32/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -44,4 +44,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/ia64/Makefile.config =================================================================== --- defaults/ia64/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/ia64/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/mips32/Makefile.config =================================================================== --- defaults/mips32/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/mips32/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -35,4 +35,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/mips64/Makefile.config =================================================================== --- defaults/mips64/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/mips64/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -35,4 +35,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/ppc32/Makefile.config =================================================================== --- defaults/ppc32/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/ppc32/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -29,4 +29,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/sparc64/Makefile.config =================================================================== --- defaults/sparc64/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/sparc64/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -44,4 +44,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Input device class CONFIG_HID_IN = generic Index: defaults/special/Makefile.config =================================================================== --- defaults/special/Makefile.config (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ defaults/special/Makefile.config (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -29,4 +29,7 @@ CONFIG_TEST = y +# Kernel RCU implementation +RCU = PREEMPT_A + # Load disk drivers on startup CONFIG_START_BD = n Index: kernel/Makefile =================================================================== --- kernel/Makefile (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/Makefile (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -192,4 +192,5 @@ generic/src/adt/bitmap.c \ generic/src/adt/btree.c \ + generic/src/adt/cht.c \ generic/src/adt/hash_table.c \ generic/src/adt/list.c \ @@ -198,4 +199,5 @@ generic/src/console/prompt.c \ generic/src/cpu/cpu.c \ + generic/src/cpu/cpu_mask.c \ generic/src/ddi/ddi.c \ generic/src/ddi/irq.c \ @@ -251,8 +253,12 @@ generic/src/synch/semaphore.c \ generic/src/synch/smc.c \ + generic/src/synch/smp_memory_barrier.c \ generic/src/synch/waitq.c \ generic/src/synch/futex.c \ + generic/src/synch/workqueue.c \ + generic/src/synch/rcu.c \ generic/src/smp/ipi.c \ generic/src/smp/smp.c \ + generic/src/smp/smp_call.c \ generic/src/ipc/ipc.c \ generic/src/ipc/sysipc.c \ @@ -304,4 +310,5 @@ test/atomic/atomic1.c \ test/btree/btree1.c \ + test/cht/cht1.c \ test/avltree/avltree1.c \ test/fault/fault1.c \ @@ -313,4 +320,7 @@ test/synch/semaphore1.c \ test/synch/semaphore2.c \ + test/synch/workqueue2.c \ + test/synch/workqueue3.c \ + test/synch/rcu1.c \ test/print/print1.c \ test/print/print2.c \ @@ -318,5 +328,6 @@ test/print/print4.c \ test/print/print5.c \ - test/thread/thread1.c + test/thread/thread1.c \ + test/smpcall/smpcall1.c ifeq ($(KARCH),mips32) Index: kernel/arch/amd64/Makefile.inc =================================================================== --- kernel/arch/amd64/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -85,4 +85,5 @@ arch/$(KARCH)/src/smp/ipi.c \ arch/$(KARCH)/src/smp/mps.c \ + arch/$(KARCH)/src/smp/smp_call.c \ arch/$(KARCH)/src/smp/smp.c endif Index: kernel/arch/amd64/include/atomic.h =================================================================== --- kernel/arch/amd64/include/atomic.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/include/atomic.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -1,4 +1,5 @@ /* * Copyright (c) 2001-2004 Jakub Jermar + * Copyright (c) 2012 Adam Hraska * All rights reserved. * @@ -140,6 +141,124 @@ } + +#define _atomic_cas_impl(pptr, exp_val, new_val, old_val, prefix) \ +({ \ + switch (sizeof(typeof(*(pptr)))) { \ + case 1: \ + asm volatile ( \ + prefix " cmpxchgb %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + case 2: \ + asm volatile ( \ + prefix " cmpxchgw %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + case 4: \ + asm volatile ( \ + prefix " cmpxchgl %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + case 8: \ + asm volatile ( \ + prefix " cmpxchgq %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + } \ +}) + + +#ifndef local_atomic_cas + +#define local_atomic_cas(pptr, exp_val, new_val) \ +({ \ + /* Use proper types and avoid name clashes */ \ + typeof(*(pptr)) _old_val_cas; \ + typeof(*(pptr)) _exp_val_cas = exp_val; \ + typeof(*(pptr)) _new_val_cas = new_val; \ + _atomic_cas_impl(pptr, _exp_val_cas, _new_val_cas, _old_val_cas, ""); \ + \ + _old_val_cas; \ +}) + +#else +/* Check if arch/atomic.h does not accidentally include /atomic.h .*/ +#error Architecture specific cpu local atomics already defined! Check your includes. #endif + +#ifndef local_atomic_exchange +/* + * Issuing a xchg instruction always implies lock prefix semantics. + * Therefore, it is cheaper to use a cmpxchg without a lock prefix + * in a loop. + */ +#define local_atomic_exchange(pptr, new_val) \ +({ \ + /* Use proper types and avoid name clashes */ \ + typeof(*(pptr)) _exp_val_x; \ + typeof(*(pptr)) _old_val_x; \ + typeof(*(pptr)) _new_val_x = new_val; \ + \ + do { \ + _exp_val_x = *pptr; \ + _old_val_x = local_atomic_cas(pptr, _exp_val_x, _new_val_x); \ + } while (_old_val_x != _exp_val_x); \ + \ + _old_val_x; \ +}) + +#else +/* Check if arch/atomic.h does not accidentally include /atomic.h .*/ +#error Architecture specific cpu local atomics already defined! Check your includes. +#endif + + +#endif + /** @} */ Index: kernel/arch/amd64/include/cpu.h =================================================================== --- kernel/arch/amd64/include/cpu.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/include/cpu.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -73,4 +73,6 @@ tss_t *tss; + unsigned int id; /** CPU's local, ie physical, APIC ID. */ + size_t iomapver_copy; /** Copy of TASK's I/O Permission bitmap generation count. */ } cpu_arch_t; Index: kernel/arch/amd64/include/interrupt.h =================================================================== --- kernel/arch/amd64/include/interrupt.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/include/interrupt.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -69,4 +69,5 @@ #define VECTOR_TLB_SHOOTDOWN_IPI (IVT_FREEBASE + 1) #define VECTOR_DEBUG_IPI (IVT_FREEBASE + 2) +#define VECTOR_SMP_CALL_IPI (IVT_FREEBASE + 3) extern void (* disable_irqs_function)(uint16_t); Index: kernel/arch/amd64/src/amd64.c =================================================================== --- kernel/arch/amd64/src/amd64.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/src/amd64.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -171,5 +171,5 @@ } -void arch_post_cpu_init() +void arch_post_cpu_init(void) { #ifdef CONFIG_SMP Index: kernel/arch/amd64/src/cpu/cpu.c =================================================================== --- kernel/arch/amd64/src/cpu/cpu.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/src/cpu/cpu.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -158,7 +158,7 @@ void cpu_print_report(cpu_t* m) { - printf("cpu%d: (%s family=%d model=%d stepping=%d) %dMHz\n", + printf("cpu%d: (%s family=%d model=%d stepping=%d apicid=%u) %dMHz\n", m->id, vendor_str[m->arch.vendor], m->arch.family, m->arch.model, - m->arch.stepping, m->frequency_mhz); + m->arch.stepping, m->arch.id, m->frequency_mhz); } Index: kernel/arch/amd64/src/interrupt.c =================================================================== --- kernel/arch/amd64/src/interrupt.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/amd64/src/interrupt.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -54,4 +54,5 @@ #include #include +#include /* @@ -161,4 +162,10 @@ tlb_shootdown_ipi_recv(); } + +static void arch_smp_call_ipi_recv(unsigned int n, istate_t *istate) +{ + trap_virtual_eoi(); + smp_call_ipi_recv(); +} #endif @@ -222,4 +229,6 @@ exc_register(VECTOR_TLB_SHOOTDOWN_IPI, "tlb_shootdown", true, (iroutine_t) tlb_shootdown_ipi); + exc_register(VECTOR_SMP_CALL_IPI, "smp_call", true, + (iroutine_t) arch_smp_call_ipi_recv); #endif } Index: kernel/arch/amd64/src/smp/smp_call.c =================================================================== --- kernel/arch/amd64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/amd64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,1 @@ +../../../ia32/src/smp/smp_call.c Index: kernel/arch/arm32/Makefile.inc =================================================================== --- kernel/arch/arm32/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/arm32/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -60,4 +60,5 @@ arch/$(KARCH)/src/mm/tlb.c \ arch/$(KARCH)/src/mm/page_fault.c \ + arch/$(KARCH)/src/atomic.c \ arch/$(KARCH)/src/ras.c Index: kernel/arch/arm32/src/atomic.c =================================================================== --- kernel/arch/arm32/src/atomic.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/arm32/src/atomic.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup arm32 + * @{ + */ +/** @file + * @brief Atomic operations emulation. + */ + +#include + + +IRQ_SPINLOCK_STATIC_INITIALIZE_NAME(cas_lock, "arm-cas-lock"); + +/** Implements GCC's missing compare-and-swap intrinsic for ARM. + * + * Sets \a *ptr to \a new_val if it is equal to \a expected. In any case, + * returns the previous value of \a *ptr. + */ +void * __sync_val_compare_and_swap_4(void **ptr, void *expected, void *new_val) +{ + /* + * Using an interrupt disabling spinlock might still lead to deadlock + * if CAS() is used in an exception handler. Eg. if a CAS() results + * in a page fault exception and the exception handler again tries + * to invoke CAS() (even for a different memory location), the spinlock + * would deadlock. + */ + irq_spinlock_lock(&cas_lock, true); + + void * cur_val = *ptr; + + if (cur_val == expected) { + *ptr = new_val; + } + + irq_spinlock_unlock(&cas_lock, true); + + return cur_val; +} + + +/** @} + */ Index: kernel/arch/ia32/Makefile.inc =================================================================== --- kernel/arch/ia32/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -86,4 +86,5 @@ arch/$(KARCH)/src/smp/mps.c \ arch/$(KARCH)/src/smp/smp.c \ + arch/$(KARCH)/src/smp/smp_call.c \ arch/$(KARCH)/src/atomic.S \ arch/$(KARCH)/src/smp/ipi.c \ Index: kernel/arch/ia32/include/atomic.h =================================================================== --- kernel/arch/ia32/include/atomic.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/include/atomic.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -1,4 +1,5 @@ /* * Copyright (c) 2001-2004 Jakub Jermar + * Copyright (c) 2012 Adam Hraska * All rights reserved. * @@ -113,4 +114,5 @@ } + /** ia32 specific fast spinlock */ NO_TRACE static inline void atomic_lock_arch(atomic_t *val) @@ -142,4 +144,106 @@ } + +#define _atomic_cas_impl(pptr, exp_val, new_val, old_val, prefix) \ +({ \ + switch (sizeof(typeof(*(pptr)))) { \ + case 1: \ + asm volatile ( \ + prefix " cmpxchgb %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + case 2: \ + asm volatile ( \ + prefix " cmpxchgw %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + case 4: \ + asm volatile ( \ + prefix " cmpxchgl %[newval], %[ptr]\n" \ + : /* Output operands. */ \ + /* Old/current value is returned in eax. */ \ + [oldval] "=a" (old_val), \ + /* (*ptr) will be read and written to, hence "+" */ \ + [ptr] "+m" (*pptr) \ + : /* Input operands. */ \ + /* Expected value must be in eax. */ \ + [expval] "a" (exp_val), \ + /* The new value may be in any register. */ \ + [newval] "r" (new_val) \ + : "memory" \ + ); \ + break; \ + } \ +}) + + +#ifndef local_atomic_cas + +#define local_atomic_cas(pptr, exp_val, new_val) \ +({ \ + /* Use proper types and avoid name clashes */ \ + typeof(*(pptr)) _old_val_cas; \ + typeof(*(pptr)) _exp_val_cas = exp_val; \ + typeof(*(pptr)) _new_val_cas = new_val; \ + _atomic_cas_impl(pptr, _exp_val_cas, _new_val_cas, _old_val_cas, ""); \ + \ + _old_val_cas; \ +}) + +#else +/* Check if arch/atomic.h does not accidentally include /atomic.h .*/ +#error Architecture specific cpu local atomics already defined! Check your includes. +#endif + + +#ifndef local_atomic_exchange +/* + * Issuing a xchg instruction always implies lock prefix semantics. + * Therefore, it is cheaper to use a cmpxchg without a lock prefix + * in a loop. + */ +#define local_atomic_exchange(pptr, new_val) \ +({ \ + /* Use proper types and avoid name clashes */ \ + typeof(*(pptr)) _exp_val_x; \ + typeof(*(pptr)) _old_val_x; \ + typeof(*(pptr)) _new_val_x = new_val; \ + \ + do { \ + _exp_val_x = *pptr; \ + _old_val_x = local_atomic_cas(pptr, _exp_val_x, _new_val_x); \ + } while (_old_val_x != _exp_val_x); \ + \ + _old_val_x; \ +}) + +#else +/* Check if arch/atomic.h does not accidentally include /atomic.h .*/ +#error Architecture specific cpu local atomics already defined! Check your includes. +#endif + + #endif Index: kernel/arch/ia32/include/cpu.h =================================================================== --- kernel/arch/ia32/include/cpu.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/include/cpu.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -60,4 +60,6 @@ unsigned int stepping; cpuid_feature_info fi; + + unsigned int id; /** CPU's local, ie physical, APIC ID. */ tss_t *tss; Index: kernel/arch/ia32/include/interrupt.h =================================================================== --- kernel/arch/ia32/include/interrupt.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/include/interrupt.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -69,4 +69,5 @@ #define VECTOR_TLB_SHOOTDOWN_IPI (IVT_FREEBASE + 1) #define VECTOR_DEBUG_IPI (IVT_FREEBASE + 2) +#define VECTOR_SMP_CALL_IPI (IVT_FREEBASE + 3) extern void (* disable_irqs_function)(uint16_t); Index: kernel/arch/ia32/include/smp/apic.h =================================================================== --- kernel/arch/ia32/include/smp/apic.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/include/smp/apic.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -353,4 +353,5 @@ extern void l_apic_init(void); extern void l_apic_eoi(void); +extern int l_apic_send_custom_ipi(uint8_t, uint8_t); extern int l_apic_broadcast_custom_ipi(uint8_t); extern int l_apic_send_init_ipi(uint8_t); Index: kernel/arch/ia32/src/cpu/cpu.c =================================================================== --- kernel/arch/ia32/src/cpu/cpu.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/src/cpu/cpu.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -160,7 +160,7 @@ void cpu_print_report(cpu_t* cpu) { - printf("cpu%u: (%s family=%u model=%u stepping=%u) %" PRIu16 " MHz\n", - cpu->id, vendor_str[cpu->arch.vendor], cpu->arch.family, - cpu->arch.model, cpu->arch.stepping, cpu->frequency_mhz); + printf("cpu%u: (%s family=%u model=%u stepping=%u apicid=%u) %" PRIu16 + " MHz\n", cpu->id, vendor_str[cpu->arch.vendor], cpu->arch.family, + cpu->arch.model, cpu->arch.stepping, cpu->arch.id, cpu->frequency_mhz); } Index: kernel/arch/ia32/src/ia32.c =================================================================== --- kernel/arch/ia32/src/ia32.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/src/ia32.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -125,5 +125,5 @@ } -void arch_post_cpu_init() +void arch_post_cpu_init(void) { #ifdef CONFIG_SMP Index: kernel/arch/ia32/src/interrupt.c =================================================================== --- kernel/arch/ia32/src/interrupt.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/src/interrupt.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -54,4 +54,6 @@ #include #include +#include +#include /* @@ -170,4 +172,10 @@ tlb_shootdown_ipi_recv(); } + +static void arch_smp_call_ipi_recv(unsigned int n, istate_t *istate) +{ + trap_virtual_eoi(); + smp_call_ipi_recv(); +} #endif @@ -230,4 +238,6 @@ exc_register(VECTOR_TLB_SHOOTDOWN_IPI, "tlb_shootdown", true, (iroutine_t) tlb_shootdown_ipi); + exc_register(VECTOR_SMP_CALL_IPI, "smp_call", true, + (iroutine_t) arch_smp_call_ipi_recv); #endif } Index: kernel/arch/ia32/src/smp/apic.c =================================================================== --- kernel/arch/ia32/src/smp/apic.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/src/smp/apic.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -259,22 +259,44 @@ } -#define DELIVS_PENDING_SILENT_RETRIES 4 - +/* Waits for the destination cpu to accept the previous ipi. */ static void l_apic_wait_for_delivery(void) { icr_t icr; - unsigned retries = 0; - + do { - if (retries++ > DELIVS_PENDING_SILENT_RETRIES) { - retries = 0; -#ifdef CONFIG_DEBUG - printf("IPI is pending.\n"); -#endif - delay(20); - } icr.lo = l_apic[ICRlo]; - } while (icr.delivs == DELIVS_PENDING); - + } while (icr.delivs != DELIVS_IDLE); +} + +/** Send one CPU an IPI vector. + * + * @param apicid Physical APIC ID of the destination CPU. + * @param vector Interrupt vector to be sent. + * + * @return 0 on failure, 1 on success. + */ +int l_apic_send_custom_ipi(uint8_t apicid, uint8_t vector) +{ + icr_t icr; + + /* Wait for a destination cpu to accept our previous ipi. */ + l_apic_wait_for_delivery(); + + icr.lo = l_apic[ICRlo]; + icr.hi = l_apic[ICRhi]; + + icr.delmod = DELMOD_FIXED; + icr.destmod = DESTMOD_PHYS; + icr.level = LEVEL_ASSERT; + icr.shorthand = SHORTHAND_NONE; + icr.trigger_mode = TRIGMOD_LEVEL; + icr.vector = vector; + icr.dest = apicid; + + /* Send the IPI by writing to l_apic[ICRlo]. */ + l_apic[ICRhi] = icr.hi; + l_apic[ICRlo] = icr.lo; + + return apic_poll_errors(); } @@ -289,4 +311,7 @@ { icr_t icr; + + /* Wait for a destination cpu to accept our previous ipi. */ + l_apic_wait_for_delivery(); icr.lo = l_apic[ICRlo]; @@ -299,6 +324,4 @@ l_apic[ICRlo] = icr.lo; - - l_apic_wait_for_delivery(); return apic_poll_errors(); Index: kernel/arch/ia32/src/smp/smp.c =================================================================== --- kernel/arch/ia32/src/smp/smp.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia32/src/smp/smp.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -55,4 +55,5 @@ #include #include +#include #ifdef CONFIG_SMP @@ -77,4 +78,14 @@ io_apic = (uint32_t *) km_map((uintptr_t) io_apic, PAGE_SIZE, PAGE_WRITE | PAGE_NOT_CACHEABLE); + } +} + +static void cpu_arch_id_init(void) +{ + ASSERT(ops != NULL); + ASSERT(cpus != NULL); + + for (unsigned int i = 0; i < config.cpu_count; ++i) { + cpus[i].arch.id = ops->cpu_apic_id(i); } } @@ -92,4 +103,10 @@ ASSERT(ops != NULL); + + /* + * SMP initialized, cpus array allocated. Assign each CPU its + * physical APIC ID. + */ + cpu_arch_id_init(); /* Index: kernel/arch/ia32/src/smp/smp_call.c =================================================================== --- kernel/arch/ia32/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/ia32/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,14 @@ +#include +#include +#include +#include + +#ifdef CONFIG_SMP + +void arch_smp_call_ipi(unsigned int cpu_id) +{ + (void) l_apic_send_custom_ipi(cpus[cpu_id].arch.id, VECTOR_SMP_CALL_IPI); +} + +#endif /* CONFIG_SMP */ + Index: kernel/arch/ia64/Makefile.inc =================================================================== --- kernel/arch/ia64/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/ia64/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -71,4 +71,5 @@ arch/$(KARCH)/src/ddi/ddi.c \ arch/$(KARCH)/src/smp/smp.c \ + arch/$(KARCH)/src/smp/smp_call.c \ arch/$(KARCH)/src/drivers/it.c Index: kernel/arch/ia64/src/smp/smp_call.c =================================================================== --- kernel/arch/ia64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/ia64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup ia64 + * @{ + */ +/** @file + */ + +#include +#include + +#ifdef CONFIG_SMP + +void arch_smp_call_ipi(unsigned int cpu_id) +{ + panic("smp_call IPI not implemented."); +} + +#endif /* CONFIG_SMP */ + +/** @} + */ Index: kernel/arch/mips32/Makefile.inc =================================================================== --- kernel/arch/mips32/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/mips32/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -70,3 +70,4 @@ arch/$(KARCH)/src/ddi/ddi.c \ arch/$(KARCH)/src/smp/dorder.c \ + arch/$(KARCH)/src/smp/smp_call.c \ arch/$(KARCH)/src/smp/smp.c Index: kernel/arch/mips32/src/smp/smp_call.c =================================================================== --- kernel/arch/mips32/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/mips32/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup mips32 + * @{ + */ +/** @file + */ + +#include +#include + +#ifdef CONFIG_SMP + +void arch_smp_call_ipi(unsigned int cpu_id) +{ + panic("smp_call IPI not implemented."); +} + +#endif /* CONFIG_SMP */ + +/** @} + */ Index: kernel/arch/mips64/Makefile.inc =================================================================== --- kernel/arch/mips64/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/mips64/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -62,3 +62,4 @@ arch/$(KARCH)/src/ddi/ddi.c \ arch/$(KARCH)/src/smp/dorder.c \ + arch/$(KARCH)/src/smp/smp_call.c \ arch/$(KARCH)/src/smp/smp.c Index: kernel/arch/mips64/src/smp/smp_call.c =================================================================== --- kernel/arch/mips64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/mips64/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup mips64 + * @{ + */ +/** @file + */ + +#include +#include + +#ifdef CONFIG_SMP + +void arch_smp_call_ipi(unsigned int cpu_id) +{ + panic("smp_call IPI not implemented."); +} + +#endif /* CONFIG_SMP */ + +/** @} + */ Index: kernel/arch/sparc64/Makefile.inc =================================================================== --- kernel/arch/sparc64/Makefile.inc (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/sparc64/Makefile.inc (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -101,4 +101,5 @@ ARCH_SOURCES += \ arch/$(KARCH)/src/smp/$(USARCH)/smp.c \ + arch/$(KARCH)/src/smp/$(USARCH)/smp_call.c \ arch/$(KARCH)/src/smp/$(USARCH)/ipi.c endif Index: kernel/arch/sparc64/include/barrier.h =================================================================== --- kernel/arch/sparc64/include/barrier.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/sparc64/include/barrier.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -37,4 +37,10 @@ #include + +#ifdef KERNEL +#include +#else +#include +#endif /* Index: kernel/arch/sparc64/include/interrupt.h =================================================================== --- kernel/arch/sparc64/include/interrupt.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/sparc64/include/interrupt.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -47,5 +47,6 @@ enum { - IPI_TLB_SHOOTDOWN = VECTOR_TLB_SHOOTDOWN_IPI + IPI_TLB_SHOOTDOWN = VECTOR_TLB_SHOOTDOWN_IPI, + IPI_SMP_CALL }; Index: kernel/arch/sparc64/include/smp/sun4u/ipi.h =================================================================== --- kernel/arch/sparc64/include/smp/sun4u/ipi.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/sparc64/include/smp/sun4u/ipi.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,45 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup sparc64 + * @{ + */ +/** + * @file + * @brief IPI functions specific to Sun4U. + */ + +#ifndef KERN_sparc64_sun4u_IPI_H_ +#define KERN_sparc64_sun4u_IPI_H_ + +extern void ipi_unicast_arch(unsigned int, int); + +#endif + +/** @} + */ Index: kernel/arch/sparc64/src/debug/stacktrace.c =================================================================== --- kernel/arch/sparc64/src/debug/stacktrace.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/sparc64/src/debug/stacktrace.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -36,4 +36,5 @@ #include #include +#include #include Index: kernel/arch/sparc64/src/smp/sun4u/ipi.c =================================================================== --- kernel/arch/sparc64/src/smp/sun4u/ipi.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/arch/sparc64/src/smp/sun4u/ipi.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -34,4 +34,5 @@ #include +#include #include #include @@ -40,4 +41,5 @@ #include #include +#include #include #include @@ -171,4 +173,26 @@ } + +/* + * Deliver an IPI to the specified processors (except the current one). + * + * Interrupts must be disabled. + * + * @param cpu_id Destination cpu id (index into cpus array). Must not + * be the current cpu. + * @param ipi IPI number. + */ +void ipi_unicast_arch(unsigned int cpu_id, int ipi) +{ + ASSERT(&cpus[cpu_id] != CPU); + + if (ipi == IPI_SMP_CALL) { + cross_call(cpus[cpu_id].arch.mid, smp_call_ipi_recv); + } else { + panic("Unknown IPI (%d).\n", ipi); + return; + } +} + /** @} */ Index: kernel/arch/sparc64/src/smp/sun4u/smp_call.c =================================================================== --- kernel/arch/sparc64/src/smp/sun4u/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/arch/sparc64/src/smp/sun4u/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,55 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup sparc64 + * @{ + */ + +/** + * @file + * @brief Sun4u specific smp call support. + */ + +#include +#include +#include + +void arch_smp_call_ipi(unsigned int cpu_id) +{ + /* + * Required by ipi_unicast_arch(). That functions resolves a potential + * deadlock should both the destination and source cpus be sending + * unicast ipis to each other with interrupts disabled. + */ + ipl_t ipl = interrupts_disable(); + ipi_unicast_arch(cpu_id, IPI_SMP_CALL); + interrupts_restore(ipl); +} + +/** @} + */ Index: kernel/generic/include/adt/cht.h =================================================================== --- kernel/generic/include/adt/cht.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/adt/cht.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,153 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup genericadt + * @{ + */ +/** @file + */ + +#ifndef KERN_CONC_HASH_TABLE_H_ +#define KERN_CONC_HASH_TABLE_H_ + +#include +#include +#include +#include +#include + +typedef uintptr_t cht_ptr_t; + +/** Concurrent hash table node link. */ +typedef struct cht_link { + /* Must be placed first. + * + * The function pointer (rcu_link.func) is used to store the item's + * mixed memoized hash. If in use by RCU (ie waiting for deferred + * destruction) the hash will contain the value of + * cht_t.op->remove_callback. + */ + union { + rcu_item_t rcu_link; + size_t hash; + }; + /** Link to the next item in the bucket including any marks. */ + cht_ptr_t link; +} cht_link_t; + +/** Set of operations for a concurrent hash table. */ +typedef struct cht_ops { + /** Returns the hash of the item. + * + * Applicable also to items that were logically deleted from the table + * but have yet to be physically removed by means of remove_callback(). + */ + size_t (*hash)(const cht_link_t *item); + /** Returns the hash value of the key used to search for entries. */ + size_t (*key_hash)(void *key); + /** Returns true if the two items store equal search keys. */ + bool (*equal)(const cht_link_t *item1, const cht_link_t *item2); + /** Returns true if the item contains an equal search key. */ + bool (*key_equal)(void *key, const cht_link_t *item); + /** Invoked to free a removed item once all references to it are dropped. */ + void (*remove_callback)(cht_link_t *item); +} cht_ops_t; + +/** Groups hash table buckets with their count. + * + * It allows both the number of buckets as well as the bucket array + * to be swapped atomically when resing the table. + */ +typedef struct cht_buckets { + /** The number of buckets is 2^order. */ + size_t order; + /** Array of single linked list bucket heads along with any marks. */ + cht_ptr_t head[1]; +} cht_buckets_t; + +/** Concurrent hash table structure. */ +typedef struct { + /** Item specific operations. */ + cht_ops_t *op; + + /** Buckets currently in use. */ + cht_buckets_t *b; + /** Resized table buckets that will replace b once resize is complete. */ + cht_buckets_t *new_b; + /** Invalid memoized hash value. + * + * If cht_link.hash contains this value the item had been logically + * removed and is waiting to be freed. Such hashes (and the associated + * items) are disregarded and skipped or the actual hash must be + * determined via op->hash(). + */ + size_t invalid_hash; + + /** Minimum number of buckets is 2^min_order. */ + size_t min_order; + /** Maximum number of items per bucket before the table grows. */ + size_t max_load; + /** Table is resized in the background in a work queue. */ + work_t resize_work; + /** If positive the table should grow or shrink. + * + * If not 0 resize work had already been posted to the system work queue. + */ + atomic_t resize_reqs; + + /** Number of items in the table that have not been logically deleted. */ + atomic_t item_cnt; +} cht_t; + +#define cht_get_inst(item, type, member) \ + member_to_inst((item), type, member) + + +#define cht_read_lock() rcu_read_lock() +#define cht_read_unlock() rcu_read_unlock() + +extern bool cht_create_simple(cht_t *h, cht_ops_t *op); +extern bool cht_create(cht_t *h, size_t init_size, size_t min_size, + size_t max_load, bool can_block, cht_ops_t *op); +extern void cht_destroy(cht_t *h); +extern void cht_destroy_unsafe(cht_t *h); + +extern cht_link_t *cht_find(cht_t *h, void *key); +extern cht_link_t *cht_find_lazy(cht_t *h, void *key); +extern cht_link_t *cht_find_next(cht_t *h, const cht_link_t *item); +extern cht_link_t *cht_find_next_lazy(cht_t *h, const cht_link_t *item); + +extern void cht_insert(cht_t *h, cht_link_t *item); +extern bool cht_insert_unique(cht_t *h, cht_link_t *item, cht_link_t **dup_item); +extern size_t cht_remove_key(cht_t *h, void *key); +extern bool cht_remove_item(cht_t *h, cht_link_t *item); + +#endif + +/** @} + */ Index: kernel/generic/include/adt/hash.h =================================================================== --- kernel/generic/include/adt/hash.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/adt/hash.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,112 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup genericadt + * @{ + */ +/** @file + */ +#ifndef KERN_HASH_H_ +#define KERN_HASH_H_ + +#include + +/** Produces a uniform hash affecting all output bits from the skewed input. */ +static inline uint32_t hash_mix32(uint32_t hash) +{ + /* + * Thomas Wang's modification of Bob Jenkin's hash mixing function: + * http://www.concentric.net/~Ttwang/tech/inthash.htm + * Public domain. + */ + hash = ~hash + (hash << 15); + hash = hash ^ (hash >> 12); + hash = hash + (hash << 2); + hash = hash ^ (hash >> 4); + hash = hash * 2057; + hash = hash ^ (hash >> 16); + return hash; +} + +/** Produces a uniform hash affecting all output bits from the skewed input. */ +static inline uint64_t hash_mix64(uint64_t hash) +{ + /* + * Thomas Wang's public domain 64-bit hash mixing function: + * http://www.concentric.net/~Ttwang/tech/inthash.htm + */ + hash = (hash ^ 61) ^ (hash >> 16); + hash = hash + (hash << 3); + hash = hash ^ (hash >> 4); + hash = hash * 0x27d4eb2d; + hash = hash ^ (hash >> 15); + /* + * Lower order bits are mixed more thoroughly. Swap them with + * the higher order bits and make the resulting higher order bits + * more usable. + */ + return (hash << 32) | (hash >> 32); +} + +/** Produces a uniform hash affecting all output bits from the skewed input. */ +static inline size_t hash_mix(size_t hash) +{ +#ifdef __32_BITS__ + return hash_mix32(hash); +#elif defined(__64_BITS__) + return hash_mix64(hash); +#else +#error Unknown size_t size - cannot select proper hash mix function. +#endif +} + +/** Use to create a hash from multiple values. + * + * Typical usage: + * @code + * int car_id; + * bool car_convertible; + * // .. + * size_t hash = 0; + * hash = hash_combine(hash, car_id); + * hash = hash_combine(hash, car_convertible); + * // Now use hash as a hash of both car_id and car_convertible. + * @endcode + */ +static inline size_t hash_combine(size_t seed, size_t hash) +{ + /* + * todo: use Bob Jenkin's proper mixing hash pass: + * http://burtleburtle.net/bob/c/lookup3.c + */ + seed ^= hash + 0x9e3779b9 + + ((seed << 5) | (seed >> (sizeof(size_t) * 8 - 5))); + return seed; +} + +#endif Index: kernel/generic/include/adt/list.h =================================================================== --- kernel/generic/include/adt/list.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/adt/list.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -51,4 +51,10 @@ } list_t; + +extern int list_member(const link_t *, const list_t *); +extern void list_splice(list_t *, link_t *); +extern unsigned int list_count(const list_t *); + + /** Declare and initialize statically allocated list. * @@ -71,4 +77,36 @@ iterator != &(list).head; iterator = iterator->next) +/** Unlike list_foreach(), allows removing items while traversing a list. + * + * @code + * list_t mylist; + * typedef struct item { + * int value; + * link_t item_link; + * } item_t; + * + * //.. + * + * // Print each list element's value and remove the element from the list. + * list_foreach_safe(mylist, cur_link, next_link) { + * item_t *cur_item = list_get_instance(cur_link, item_t, item_link); + * printf("%d\n", cur_item->value); + * list_remove(cur_link); + * } + * @endcode + * + * @param list List to traverse. + * @param iterator Iterator to the current element of the list. + * The item this iterator points may be safely removed + * from the list. + * @param next_iter Iterator to the next element of the list. + */ +#define list_foreach_safe(list, iterator, next_iter) \ + for (link_t *iterator = (list).head.next, \ + *next_iter = iterator->next; \ + iterator != &(list).head; \ + iterator = next_iter, next_iter = iterator->next) + + #define assert_link_not_used(link) \ ASSERT(((link)->prev == NULL) && ((link)->next == NULL)) @@ -85,4 +123,15 @@ link->prev = NULL; link->next = NULL; +} + +/** Returns true if the initialized link is already in use by any list. + * + * @param link Link to examine whether if belongs to a list or not. + * @return 1 if the link is part of a list. + * @return 0 otherwise. + */ +NO_TRACE static inline int link_used(const link_t *link) +{ + return link->prev != NULL || link->next != NULL; } @@ -256,4 +305,19 @@ { headless_list_split_or_concat(part1, part2); +} + +/** Concatenate two lists + * + * Concatenate lists @a list1 and @a list2, producing a single + * list @a list1 containing items from both (in @a list1, @a list2 + * order) and empty list @a list2. + * + * @param list1 First list and concatenated output + * @param list2 Second list and empty output. + * + */ +NO_TRACE static inline void list_concat(list_t *list1, list_t *list2) +{ + list_splice(list2, list1->head.prev); } @@ -281,8 +345,4 @@ } -extern int list_member(const link_t *, const list_t *); -extern void list_concat(list_t *, list_t *); -extern unsigned int list_count(const list_t *); - #endif Index: kernel/generic/include/arch.h =================================================================== --- kernel/generic/include/arch.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/arch.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -36,8 +36,8 @@ #define KERN_ARCH_H_ -#include -#include -#include -#include +#include /* arch_pre_main() */ +#include /* get_stack_base() */ +#include + /* @@ -49,9 +49,4 @@ #define THE ((the_t * )(get_stack_base())) -#define CPU THE->cpu -#define THREAD THE->thread -#define TASK THE->task -#define AS THE->as -#define PREEMPTION_DISABLED THE->preemption_disabled #define MAGIC UINT32_C(0xfacefeed) @@ -62,4 +57,10 @@ ((THE->task) ? (THE->task->container) : (DEFAULT_CONTAINER)) +/* Fwd decl. to avoid include hell. */ +struct thread; +struct task; +struct cpu; +struct as; + /** * For each possible kernel stack, structure @@ -68,10 +69,13 @@ */ typedef struct { - size_t preemption_disabled; /**< Preemption disabled counter. */ - thread_t *thread; /**< Current thread. */ - task_t *task; /**< Current task. */ - cpu_t *cpu; /**< Executing cpu. */ - as_t *as; /**< Current address space. */ - uint32_t magic; /**< Magic value */ + size_t preemption; /**< Preemption disabled counter and flag. */ +#ifdef RCU_PREEMPT_A + size_t rcu_nesting; /**< RCU nesting count and flag. */ +#endif + struct thread *thread; /**< Current thread. */ + struct task *task; /**< Current task. */ + struct cpu *cpu; /**< Executing cpu. */ + struct as *as; /**< Current address space. */ + uint32_t magic; /**< Magic value */ } the_t; @@ -91,4 +95,5 @@ extern void *arch_construct_function(fncptr_t *, void *, void *); + #endif Index: kernel/generic/include/atomic.h =================================================================== --- kernel/generic/include/atomic.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/atomic.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -53,4 +53,18 @@ } + +/* + * If the architecture does not provide operations that are atomic + * only with respect to the local cpu (eg exception handlers) and + * not other cpus, implement these cpu local atomic operations with + * full blown smp-safe atomics. + */ +#ifndef local_atomic_exchange +#define local_atomic_exchange(var_addr, new_val) \ + __atomic_exchange_n((var_addr), (new_val), __ATOMIC_RELAXED) +#endif + + + #endif Index: kernel/generic/include/compiler/barrier.h =================================================================== --- kernel/generic/include/compiler/barrier.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/compiler/barrier.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,37 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#ifndef KERN_COMPILER_BARRIER_H_ +#define KERN_COMPILER_BARRIER_H_ + +#define compiler_barrier() asm volatile ("" ::: "memory") + +/** Forces the compiler to access (ie load/store) the variable only once. */ +#define ACCESS_ONCE(var) (*((volatile typeof(var)*)&(var))) + +#endif /* KERN_COMPILER_BARRIER_H_ */ Index: kernel/generic/include/cpu.h =================================================================== --- kernel/generic/include/cpu.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/cpu.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -38,7 +38,13 @@ #include #include +#include #include #include #include +#include +#include + +#define CPU THE->cpu + /** CPU structure. @@ -94,4 +100,13 @@ /** + * SMP calls to invoke on this CPU. + */ + SPINLOCK_DECLARE(smp_calls_lock); + list_t smp_pending_calls; + + /** RCU per-cpu data. Uses own locking. */ + rcu_cpu_data_t rcu; + + /** * Stack used by scheduler when there is no running thread. */ Index: kernel/generic/include/cpu/cpu_mask.h =================================================================== --- kernel/generic/include/cpu/cpu_mask.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/cpu/cpu_mask.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,79 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ +/** @file + */ +#ifndef KERN_CPU_CPU_MASK_H_ +#define KERN_CPU_CPU_MASK_H_ + +#include +#include +#include + +/** Iterates over all cpu id's whose bit is included in the cpu mask. + * + * Example usage: + * @code + * DEFINE_CPU_MASK(cpu_mask); + * cpu_mask_active(&cpu_mask); + * + * cpu_mask_for_each(cpu_mask, cpu_id) { + * printf("Cpu with logical id %u is active.\n", cpu_id); + * } + * @endcode + */ +#define cpu_mask_for_each(mask, cpu_id) \ + for (unsigned int (cpu_id) = 0; (cpu_id) < config.cpu_count; ++(cpu_id)) \ + if (cpu_mask_is_set(&(mask), (cpu_id))) + +/** Allocates a cpu_mask_t on stack. */ +#define DEFINE_CPU_MASK(cpu_mask) \ + cpu_mask_t *(cpu_mask) = (cpu_mask_t*) alloca(cpu_mask_size()) + +/** If used with DEFINE_CPU_MASK, the mask is large enough for all detected cpus.*/ +typedef struct cpu_mask { + unsigned int mask[1]; +} cpu_mask_t; + + +extern size_t cpu_mask_size(void); +extern void cpu_mask_active(cpu_mask_t *); +extern void cpu_mask_all(cpu_mask_t *); +extern void cpu_mask_none(cpu_mask_t *); +extern void cpu_mask_set(cpu_mask_t *, unsigned int); +extern void cpu_mask_reset(cpu_mask_t *, unsigned int); +extern bool cpu_mask_is_set(cpu_mask_t *, unsigned int); +extern bool cpu_mask_is_none(cpu_mask_t *); + +#endif /* KERN_CPU_CPU_MASK_H_ */ + +/** @} + */ Index: kernel/generic/include/lib/memfnc.h =================================================================== --- kernel/generic/include/lib/memfnc.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/lib/memfnc.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,6 @@ extern void *memcpy(void *, const void *, size_t); +#define alloca(size) __builtin_alloca((size)) + #endif Index: kernel/generic/include/macros.h =================================================================== --- kernel/generic/include/macros.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/macros.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -150,4 +150,11 @@ }) + +#ifndef member_to_inst +#define member_to_inst(ptr_member, type, member_identif) \ + ((type*) (((void*)(ptr_member)) - ((void*)&(((type*)0)->member_identif)))) +#endif + + #endif Index: kernel/generic/include/memstr.h =================================================================== --- kernel/generic/include/memstr.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/memstr.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -40,4 +40,5 @@ #define memset(dst, val, cnt) __builtin_memset((dst), (val), (cnt)) #define memcpy(dst, src, cnt) __builtin_memcpy((dst), (src), (cnt)) +#define bzero(dst, cnt) memset((dst), 0, (cnt)) extern void memsetb(void *, size_t, uint8_t); Index: kernel/generic/include/mm/as.h =================================================================== --- kernel/generic/include/mm/as.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/mm/as.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -48,4 +48,8 @@ #include #include +#include + +#define AS THE->as + /** Index: kernel/generic/include/preemption.h =================================================================== --- kernel/generic/include/preemption.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/preemption.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -36,6 +36,27 @@ #define KERN_PREEMPTION_H_ -extern void preemption_disable(void); -extern void preemption_enable(void); +#include +#include +#include + +#define PREEMPTION_INC (1 << 0) +#define PREEMPTION_DISABLED (PREEMPTION_INC <= THE->preemption) +#define PREEMPTION_ENABLED (!PREEMPTION_DISABLED) + +/** Increment preemption disabled counter. */ +#define preemption_disable() \ + do { \ + THE->preemption += PREEMPTION_INC; \ + compiler_barrier(); \ + } while (0) + +/** Restores preemption but never reschedules. */ +#define preemption_enable() \ + do { \ + ASSERT(PREEMPTION_DISABLED); \ + compiler_barrier(); \ + THE->preemption -= PREEMPTION_INC; \ + } while (0) + #endif Index: kernel/generic/include/proc/task.h =================================================================== --- kernel/generic/include/proc/task.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/proc/task.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -43,6 +43,8 @@ #include #include +#include #include #include +#include #include #include @@ -57,4 +59,8 @@ #include #include +#include + +#define TASK THE->task + struct thread; @@ -123,11 +129,13 @@ task_arch_t arch; - /** - * Serializes access to the B+tree of task's futexes. This mutex is - * independent on the task spinlock. - */ - mutex_t futexes_lock; - /** B+tree of futexes referenced by this task. */ - btree_t futexes; + struct futex_cache { + /** CHT mapping virtual addresses of futex variables to futex objects.*/ + cht_t ht; + /** Serializes access to futex_list.*/ + mutex_t list_lock; + /** List of all futexes accesses by this task. */ + list_t list; + work_t destroy_work; + } *futexes; /** Accumulated accounting. */ Index: kernel/generic/include/proc/thread.h =================================================================== --- kernel/generic/include/proc/thread.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/proc/thread.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,5 @@ #include #include +#include #include #include @@ -48,4 +49,8 @@ #include #include +#include + + +#define THREAD THE->thread #define THREAD_NAME_BUFLEN 20 @@ -180,4 +185,16 @@ /** Thread ID. */ thread_id_t tid; + + /** Work queue this thread belongs to or NULL. Immutable. */ + struct work_queue *workq; + /** Links work queue threads. Protected by workq->lock. */ + link_t workq_link; + /** True if the worker was blocked and is not running. Use thread->lock. */ + bool workq_blocked; + /** True if the worker will block in order to become idle. Use workq->lock. */ + bool workq_idling; + + /** RCU thread related data. Protected by its own locks. */ + rcu_thread_data_t rcu; /** Architecture-specific data. */ @@ -217,4 +234,6 @@ extern void thread_ready(thread_t *); extern void thread_exit(void) __attribute__((noreturn)); +extern void thread_interrupt(thread_t *); +extern bool thread_interrupted(thread_t *); #ifndef thread_create_arch Index: kernel/generic/include/smp/smp_call.h =================================================================== --- kernel/generic/include/smp/smp_call.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/smp/smp_call.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,71 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ +/** @file + */ + +#ifndef KERN_SMP_CALL_H_ +#define KERN_SMP_CALL_H_ + +#include +#include +#include + +typedef void (*smp_call_func_t)(void *); + +typedef struct smp_call { + smp_call_func_t func; + void *arg; + link_t calls_link; + atomic_t pending; +} smp_call_t; + + + +extern void smp_call(unsigned int, smp_call_func_t, void *); +extern void smp_call_async(unsigned int, smp_call_func_t, void *, smp_call_t *); +extern void smp_call_wait(smp_call_t *); + +extern void smp_call_init(void); + +#ifdef CONFIG_SMP +extern void smp_call_ipi_recv(void); +extern void arch_smp_call_ipi(unsigned int); +#endif + + + + +#endif /* KERN_SMP_CALL_H_ */ + +/** @} + */ + Index: kernel/generic/include/synch/condvar.h =================================================================== --- kernel/generic/include/synch/condvar.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/synch/condvar.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -39,4 +39,5 @@ #include #include +#include #include @@ -50,4 +51,12 @@ _condvar_wait_timeout((cv), (mtx), (usec), SYNCH_FLAGS_NONE) +#ifdef CONFIG_SMP +#define _condvar_wait_timeout_spinlock(cv, lock, usec, flags) \ + _condvar_wait_timeout_spinlock_impl((cv), (lock), (usec), (flags)) +#else +#define _condvar_wait_timeout_spinlock(cv, lock, usec, flags) \ + _condvar_wait_timeout_spinlock_impl((cv), NULL, (usec), (flags)) +#endif + extern void condvar_initialize(condvar_t *cv); extern void condvar_signal(condvar_t *cv); @@ -55,4 +64,9 @@ extern int _condvar_wait_timeout(condvar_t *cv, mutex_t *mtx, uint32_t usec, int flags); +extern int _condvar_wait_timeout_spinlock_impl(condvar_t *cv, spinlock_t *lock, + uint32_t usec, int flags); +extern int _condvar_wait_timeout_irq_spinlock(condvar_t *cv, + irq_spinlock_t *irq_lock, uint32_t usec, int flags); + #endif Index: kernel/generic/include/synch/futex.h =================================================================== --- kernel/generic/include/synch/futex.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/synch/futex.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -55,5 +55,7 @@ extern sysarg_t sys_futex_wakeup(uintptr_t); -extern void futex_cleanup(void); +extern void futex_task_cleanup(void); +extern void futex_task_init(struct task *); +extern void futex_task_deinit(struct task *); #endif Index: kernel/generic/include/synch/rcu.h =================================================================== --- kernel/generic/include/synch/rcu.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/synch/rcu.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,247 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup sync + * @{ + */ +/** @file + */ + +#ifndef KERN_RCU_H_ +#define KERN_RCU_H_ + +#include +#include + + +/** Use to assign a pointer to newly initialized data to a rcu reader + * accessible pointer. + * + * Example: + * @code + * typedef struct exam { + * struct exam *next; + * int grade; + * } exam_t; + * + * exam_t *exam_list; + * // .. + * + * // Insert at the beginning of the list. + * exam_t *my_exam = malloc(sizeof(exam_t), 0); + * my_exam->grade = 5; + * my_exam->next = exam_list; + * rcu_assign(exam_list, my_exam); + * + * // Changes properly propagate. Every reader either sees + * // the old version of exam_list or the new version with + * // the fully initialized my_exam. + * rcu_synchronize(); + * // Now we can be sure every reader sees my_exam. + * + * @endcode + */ +#define rcu_assign(ptr, value) \ + do { \ + memory_barrier(); \ + (ptr) = (value); \ + } while (0) + +/** Use to access RCU protected data in a reader section. + * + * Example: + * @code + * exam_t *exam_list; + * // ... + * + * rcu_read_lock(); + * exam_t *first_exam = rcu_access(exam_list); + * // We can now safely use first_exam, it won't change + * // under us while we're using it. + * + * // .. + * rcu_read_unlock(); + * @endcode + */ +#define rcu_access(ptr) ACCESS_ONCE(ptr) + + + + +#include +#include +#include +#include + + +extern bool rcu_read_locked(void); +extern void rcu_synchronize(void); +extern void rcu_synchronize_expedite(void); +extern void rcu_call(rcu_item_t *rcu_item, rcu_func_t func); +extern void rcu_barrier(void); + +extern void rcu_print_stat(void); + +extern void rcu_init(void); +extern void rcu_stop(void); +extern void rcu_cpu_init(void); +extern void rcu_kinit_init(void); +extern void rcu_thread_init(struct thread*); +extern void rcu_thread_exiting(void); +extern void rcu_after_thread_ran(void); +extern void rcu_before_thread_runs(void); + +extern uint64_t rcu_completed_gps(void); +extern void _rcu_call(bool expedite, rcu_item_t *rcu_item, rcu_func_t func); +extern void _rcu_synchronize(bool expedite); + + +#ifdef RCU_PREEMPT_A + +#define RCU_CNT_INC (1 << 1) +#define RCU_WAS_PREEMPTED (1 << 0) + +/* Fwd. decl. because of inlining. */ +void _rcu_preempted_unlock(void); + +/** Delimits the start of an RCU reader critical section. + * + * Reader sections may be nested and are preemptible. You must not + * however block/sleep within reader sections. + */ +static inline void rcu_read_lock(void) +{ + THE->rcu_nesting += RCU_CNT_INC; + compiler_barrier(); +} + +/** Delimits the end of an RCU reader critical section. */ +static inline void rcu_read_unlock(void) +{ + compiler_barrier(); + THE->rcu_nesting -= RCU_CNT_INC; + + if (RCU_WAS_PREEMPTED == THE->rcu_nesting) { + _rcu_preempted_unlock(); + } +} + +#elif defined(RCU_PREEMPT_PODZIMEK) + +/* Fwd decl. required by the inlined implementation. Not part of public API. */ +extern rcu_gp_t _rcu_cur_gp; +extern void _rcu_signal_read_unlock(void); + + +/** Unconditionally records a quiescent state for the local cpu. */ +static inline void _rcu_record_qs(void) +{ + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + + /* + * A new GP was started since the last time we passed a QS. + * Notify the detector we have reached a new QS. + */ + if (CPU->rcu.last_seen_gp != _rcu_cur_gp) { + rcu_gp_t cur_gp = ACCESS_ONCE(_rcu_cur_gp); + /* + * Contain memory accesses within a reader critical section. + * If we are in rcu_lock() it also makes changes prior to the + * start of the GP visible in the reader section. + */ + memory_barrier(); + /* + * Acknowledge we passed a QS since the beginning of rcu.cur_gp. + * Cache coherency will lazily transport the value to the + * detector while it sleeps in gp_sleep(). + * + * Note that there is a theoretical possibility that we + * overwrite a more recent/greater last_seen_gp here with + * an older/smaller value. If this cpu is interrupted here + * while in rcu_lock() reader sections in the interrupt handler + * will update last_seen_gp to the same value as is currently + * in local cur_gp. However, if the cpu continues processing + * interrupts and the detector starts a new GP immediately, + * local interrupt handlers may update last_seen_gp again (ie + * properly ack the new GP) with a value greater than local cur_gp. + * Resetting last_seen_gp to a previous value here is however + * benign and we only have to remember that this reader may end up + * in cur_preempted even after the GP ends. That is why we + * append next_preempted to cur_preempted rather than overwriting + * it as if cur_preempted were empty. + */ + CPU->rcu.last_seen_gp = cur_gp; + } +} + +/** Delimits the start of an RCU reader critical section. + * + * Reader sections may be nested and are preemptable. You must not + * however block/sleep within reader sections. + */ +static inline void rcu_read_lock(void) +{ + ASSERT(CPU); + preemption_disable(); + + /* Record a QS if not in a reader critical section. */ + if (0 == CPU->rcu.nesting_cnt) + _rcu_record_qs(); + + ++CPU->rcu.nesting_cnt; + + preemption_enable(); +} + +/** Delimits the end of an RCU reader critical section. */ +static inline void rcu_read_unlock(void) +{ + ASSERT(CPU); + preemption_disable(); + + if (0 == --CPU->rcu.nesting_cnt) { + _rcu_record_qs(); + + /* + * The thread was preempted while in a critical section or + * the detector is eagerly waiting for this cpu's reader to finish. + */ + if (CPU->rcu.signal_unlock) { + /* Rechecks with disabled interrupts. */ + _rcu_signal_read_unlock(); + } + } + + preemption_enable(); +} +#endif + +#endif + +/** @} + */ Index: kernel/generic/include/synch/rcu_types.h =================================================================== --- kernel/generic/include/synch/rcu_types.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/synch/rcu_types.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,171 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup sync + * @{ + */ +/** @file + */ + +#ifndef KERN_RCU_TYPES_H_ +#define KERN_RCU_TYPES_H_ + +#include +#include + +#if !defined(RCU_PREEMPT_PODZIMEK) && !defined(RCU_PREEMPT_A) +#error You must select an RCU algorithm. +#endif + + +/* Fwd decl. */ +struct thread; +struct rcu_item; + +/** Grace period number typedef. */ +typedef uint64_t rcu_gp_t; + +/** RCU callback type. The passed rcu_item_t maybe freed. */ +typedef void (*rcu_func_t)(struct rcu_item *rcu_item); + +typedef struct rcu_item { + rcu_func_t func; + struct rcu_item *next; +} rcu_item_t; + + +/** RCU related per-cpu data. */ +typedef struct rcu_cpu_data { + /** The cpu recorded a quiescent state last time during this grace period.*/ + rcu_gp_t last_seen_gp; + +#ifdef RCU_PREEMPT_PODZIMEK + /** This cpu has not yet passed a quiescent state and it is delaying the + * detector. Once it reaches a QS it must sema_up(rcu.remaining_readers). + */ + bool is_delaying_gp; + + /** True if we should signal the detector that we exited a reader section. + * + * Equal to (THREAD->rcu.was_preempted || CPU->rcu.is_delaying_gp). + */ + bool signal_unlock; + + /** The number of times an RCU reader section is nested on this cpu. + * + * If positive, it is definitely executing reader code. If zero, + * the thread might already be executing reader code thanks to + * cpu instruction reordering. + */ + size_t nesting_cnt; +#endif + + /** Callbacks to invoke once the current grace period ends, ie cur_cbs_gp. + * Accessed by the local reclaimer only. + */ + rcu_item_t *cur_cbs; + /** Number of callbacks in cur_cbs. */ + size_t cur_cbs_cnt; + /** Callbacks to invoke once the next grace period ends, ie next_cbs_gp. + * Accessed by the local reclaimer only. + */ + rcu_item_t *next_cbs; + /** Number of callbacks in next_cbs. */ + size_t next_cbs_cnt; + /** New callbacks are place at the end of this list. */ + rcu_item_t *arriving_cbs; + /** Tail of arriving_cbs list. Disable interrupts to access. */ + rcu_item_t **parriving_cbs_tail; + /** Number of callbacks currently in arriving_cbs. + * Disable interrupts to access. + */ + size_t arriving_cbs_cnt; + + /** At the end of this grace period callbacks in cur_cbs will be invoked.*/ + rcu_gp_t cur_cbs_gp; + /** At the end of this grace period callbacks in next_cbs will be invoked. + * + * Should be the next grace period but it allows the reclaimer to + * notice if it missed a grace period end announcement. In that + * case it can execute next_cbs without waiting for another GP. + * + * Invariant: next_cbs_gp >= cur_cbs_gp + */ + rcu_gp_t next_cbs_gp; + + /** Positive if there are callbacks pending in arriving_cbs. */ + semaphore_t arrived_flag; + + /** The reclaimer should expedite GPs for cbs in arriving_cbs. */ + bool expedite_arriving; + + /** Protected by global rcu.barrier_mtx. */ + rcu_item_t barrier_item; + + /** Interruptable attached reclaimer thread. */ + struct thread *reclaimer_thr; + + /* Some statistics. */ + size_t stat_max_cbs; + size_t stat_avg_cbs; + size_t stat_missed_gps; + size_t stat_missed_gp_in_wait; + size_t stat_max_slice_cbs; + size_t last_arriving_cnt; +} rcu_cpu_data_t; + + +/** RCU related per-thread data. */ +typedef struct rcu_thread_data { + /** + * Nesting count of the thread's RCU read sections when the thread + * is not running. + */ + size_t nesting_cnt; + +#ifdef RCU_PREEMPT_PODZIMEK + + /** True if the thread was preempted in a reader section. + * + * The thread is placed into rcu.cur_preempted or rcu.next_preempted + * and must remove itself in rcu_read_unlock(). + * + * Access with interrupts disabled. + */ + bool was_preempted; +#endif + + /** Preempted threads link. Access with rcu.prempt_lock.*/ + link_t preempt_link; +} rcu_thread_data_t; + + +#endif + +/** @} + */ Index: kernel/generic/include/synch/semaphore.h =================================================================== --- kernel/generic/include/synch/semaphore.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/synch/semaphore.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -53,4 +53,8 @@ _semaphore_down_timeout((s), (usec), SYNCH_FLAGS_NONE) +#define semaphore_down_interruptable(s) \ + (ESYNCH_INTERRUPTED != _semaphore_down_timeout((s), SYNCH_NO_TIMEOUT, \ + SYNCH_FLAGS_INTERRUPTIBLE)) + extern void semaphore_initialize(semaphore_t *, int); extern int _semaphore_down_timeout(semaphore_t *, uint32_t, unsigned int); Index: kernel/generic/include/synch/smp_memory_barrier.h =================================================================== --- kernel/generic/include/synch/smp_memory_barrier.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/synch/smp_memory_barrier.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,45 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup sync + * @{ + */ +/** @file + */ + +#ifndef KERN_SMP_MEM_BAR_H_ +#define KERN_SMP_MEM_BAR_H_ + +#include + +extern sysarg_t sys_smp_memory_barrier(void); + +#endif + +/** @} + */ Index: kernel/generic/include/synch/spinlock.h =================================================================== --- kernel/generic/include/synch/spinlock.h (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/include/synch/spinlock.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -45,5 +45,5 @@ #ifdef CONFIG_SMP -typedef struct { +typedef struct spinlock { atomic_t val; @@ -162,4 +162,7 @@ /* On UP systems, spinlocks are effectively left out. */ +/* Allow the use of spinlock_t as an incomplete type. */ +typedef struct spinlock spinlock_t; + #define SPINLOCK_DECLARE(name) #define SPINLOCK_EXTERN(name) @@ -176,5 +179,5 @@ #define spinlock_lock(lock) preemption_disable() -#define spinlock_trylock(lock) (preemption_disable(), 1) +#define spinlock_trylock(lock) ({ preemption_disable(); 1; }) #define spinlock_unlock(lock) preemption_enable() #define spinlock_locked(lock) 1 Index: kernel/generic/include/synch/workqueue.h =================================================================== --- kernel/generic/include/synch/workqueue.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/include/synch/workqueue.h (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,83 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ +/** @file + */ + +#ifndef KERN_WORKQUEUE_H_ +#define KERN_WORKQUEUE_H_ + +#include + +/* Fwd decl. */ +struct thread; +struct work_item; +struct work_queue; +typedef struct work_queue work_queue_t; + +typedef void (*work_func_t)(struct work_item *); + +typedef struct work_item { + link_t queue_link; + work_func_t func; + +#ifdef CONFIG_DEBUG + /* Magic number for integrity checks. */ + uint32_t cookie; +#endif +} work_t; + + + +extern void workq_global_init(void); +extern void workq_global_worker_init(void); +extern void workq_global_stop(void); +extern int workq_global_enqueue_noblock(work_t *, work_func_t); +extern int workq_global_enqueue(work_t *, work_func_t); + +extern struct work_queue * workq_create(const char *); +extern void workq_destroy(struct work_queue *); +extern int workq_init(struct work_queue *, const char *); +extern void workq_stop(struct work_queue *); +extern int workq_enqueue_noblock(struct work_queue *, work_t *, work_func_t); +extern int workq_enqueue(struct work_queue *, work_t *, work_func_t); + +extern void workq_print_info(struct work_queue *); +extern void workq_global_print_info(void); + + +extern void workq_after_thread_ran(void); +extern void workq_before_thread_is_ready(struct thread *); + +#endif /* KERN_WORKQUEUE_H_ */ + +/** @} + */ Index: kernel/generic/src/adt/cht.c =================================================================== --- kernel/generic/src/adt/cht.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/adt/cht.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,2704 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + + +/** @addtogroup genericadt + * @{ + */ + +/** + * @file + * @brief Scalable resizable concurrent lock-free hash table. + * + * CHT is a concurrent hash table that is scalable resizable and lock-free. + * resizable = the number of buckets of the table increases or decreases + * depending on the average number of elements per bucket (ie load) + * scalable = accessing the table from more cpus increases performance + * almost linearly + * lock-free = common operations never block; even if any of the operations + * is preempted or interrupted at any time, other operations will still + * make forward progress + * + * CHT is designed for read mostly scenarios. Performance degrades as the + * fraction of updates (insert/remove) increases. Other data structures + * significantly outperform CHT if the fraction of updates exceeds ~40%. + * + * CHT tolerates hardware exceptions and may be accessed from exception + * handlers as long as the underlying RCU implementation is exception safe. + * + * @par Caveats + * + * 0) Never assume an item is still in the table. + * The table may be accessed concurrently; therefore, other threads may + * insert or remove an item at any time. Do not assume an item is still + * in the table if cht_find() just returned it to you. Similarly, an + * item may have already been inserted by the time cht_find() returns NULL. + * + * 1) Always use RCU read locks when searching the table. + * Holding an RCU lock guarantees that an item found in the table remains + * valid (eg is not freed) even if the item was removed from the table + * in the meantime by another thread. + * + * 2) Never update values in place. + * Do not update items in the table in place, ie directly. The changes + * will not propagate to other readers (on other cpus) immediately or even + * correctly. Some readers may then encounter items that have only some + * of their fields changed or are completely inconsistent. + * + * Instead consider inserting an updated/changed copy of the item and + * removing the original item. Or contact the maintainer to provide + * you with a function that atomically replaces an item with a copy. + * + * 3) Use cht_insert_unique() instead of checking for duplicates with cht_find() + * The following code is prone to race conditions: + * @code + * if (NULL == cht_find(&h, key)) { + * // If another thread inserts and item here, we'll insert a duplicate. + * cht_insert(&h, item); + * } + * @endcode + * See cht_insert_unique() on how to correctly fix this. + * + * + * @par Semantics + * + * Lazy readers = cht_find_lazy(), cht_find_next_lazy() + * Readers = lazy readers, cht_find(), cht_find_next() + * Updates = cht_insert(), cht_insert_unique(), cht_remove_key(), + * cht_remove_item() + * + * Readers (but not lazy readers) are guaranteed to see the effects + * of @e completed updates. In other words, if cht_find() is invoked + * after a cht_insert() @e returned eg on another cpu, cht_find() is + * guaranteed to see the inserted item. + * + * Similarly, updates see the effects of @e completed updates. For example, + * issuing cht_remove() after a cht_insert() for that key returned (even + * on another cpu) is guaranteed to remove the inserted item. + * + * Reading or updating the table concurrently with other updates + * always returns consistent data and never corrupts the table. + * However the effects of concurrent updates may or may not be + * visible to all other concurrent readers or updaters. Eg, not + * all readers may see that an item has already been inserted + * if cht_insert() has not yet returned. + * + * Lazy readers are guaranteed to eventually see updates but it + * may take some time (possibly milliseconds) after the update + * completes for the change to propagate to lazy readers on all + * cpus. + * + * @par Implementation + * + * Collisions in CHT are resolved with chaining. The number of buckets + * is always a power of 2. Each bucket is represented with a single linked + * lock-free list [1]. Items in buckets are sorted by their mixed hashes + * in ascending order. All buckets are terminated with a single global + * sentinel node whose mixed hash value is the greatest possible. + * + * CHT with 2^k buckets uses the k most significant bits of a hash value + * to determine the bucket number where an item is to be stored. To + * avoid storing all items in a single bucket if the user supplied + * hash function does not produce uniform hashes, hash values are + * mixed first so that the top bits of a mixed hash change even if hash + * values differ only in the least significant bits. The mixed hash + * values are cached in cht_link.hash (which is overwritten once the + * item is scheduled for removal via rcu_call). + * + * A new item is inserted before all other existing items in the bucket + * with the same hash value as the newly inserted item (a la the original + * lock-free list [2]). Placing new items at the start of a same-hash + * sequence of items (eg duplicates) allows us to easily check for duplicates + * in cht_insert_unique(). The function can first check that there are + * no duplicates of the newly inserted item amongst the items with the + * same hash as the new item. If there were no duplicates the new item + * is linked before the same-hash items. Inserting a duplicate while + * the function is checking for duplicates is detected as a change of + * the link to the first checked same-hash item (and the search for + * duplicates can be restarted). + * + * @par Table resize algorithm + * + * Table resize is based on [3] and [5]. First, a new bucket head array + * is allocated and initialized. Second, old bucket heads are moved + * to the new bucket head array with the protocol mentioned in [5]. + * At this point updaters start using the new bucket heads. Third, + * buckets are split (or joined) so that the table can make use of + * the extra bucket head slots in the new array (or stop wasting space + * with the unnecessary extra slots in the old array). Splitting + * or joining buckets employs a custom protocol. Last, the new array + * replaces the original bucket array. + * + * A single background work item (of the system work queue) guides + * resizing of the table. If an updater detects that the bucket it + * is about to access is undergoing a resize (ie its head is moving + * or it needs to be split/joined), it helps out and completes the + * head move or the bucket split/join. + * + * The table always grows or shrinks by a factor of 2. Because items + * are assigned a bucket based on the top k bits of their mixed hash + * values, when growing the table each bucket is split into two buckets + * and all items of the two new buckets come from the single bucket in the + * original table. Ie items from separate buckets in the original table + * never intermix in the new buckets. Moreover + * since the buckets are sorted by their mixed hash values the items + * at the beginning of the old bucket will end up in the first new + * bucket while all the remaining items of the old bucket will end up + * in the second new bucket. Therefore, there is a single point where + * to split the linked list of the old bucket into two correctly sorted + * linked lists of the new buckets: + * .- bucket split + * | + * <-- first --> v <-- second --> + * [old] --> [00b] -> [01b] -> [10b] -> [11b] -> sentinel + * ^ ^ + * [new0] -- -+ | + * [new1] -- -- -- -- -- -- -- -+ + * + * Resize in greater detail: + * + * a) First, a resizer (a single background system work queue item + * in charge of resizing the table) allocates and initializes a new + * bucket head array. New bucket heads are pointed to the sentinel + * and marked Invalid (in the lower order bits of the pointer to the + * next item, ie the sentinel in this case): + * + * [old, N] --> [00b] -> [01b] -> [10b] -> [11b] -> sentinel + * ^ ^ + * [new0, Inv] -------------------------------------+ | + * [new1, Inv] ---------------------------------------+ + * + * + * b) Second, the resizer starts moving old bucket heads with the following + * lock-free protocol (from [5]) where cas(variable, expected_val, new_val) + * is short for compare-and-swap: + * + * old head new0 head transition to next state + * -------- --------- ------------------------ + * addr, N sentinel, Inv cas(old, (addr, N), (addr, Const)) + * .. mark the old head as immutable, so that + * updaters do not relink it to other nodes + * until the head move is done. + * addr, Const sentinel, Inv cas(new0, (sentinel, Inv), (addr, N)) + * .. move the address to the new head and mark + * the new head normal so updaters can start + * using it. + * addr, Const addr, N cas(old, (addr, Const), (addr, Inv)) + * .. mark the old head Invalid to signify + * the head move is done. + * addr, Inv addr, N + * + * Notice that concurrent updaters may step in at any point and correctly + * complete the head move without disrupting the resizer. At worst, the + * resizer or other concurrent updaters will attempt a number of CAS() that + * will correctly fail. + * + * [old, Inv] -> [00b] -> [01b] -> [10b] -> [11b] -> sentinel + * ^ ^ + * [new0, N] ----+ | + * [new1, Inv] --------------------------------------+ + * + * + * c) Third, buckets are split if the table is growing; or joined if + * shrinking (by the resizer or updaters depending on whoever accesses + * the bucket first). See split_bucket() and join_buckets() for details. + * + * 1) Mark the last item of new0 with JOIN_FOLLOWS: + * [old, Inv] -> [00b] -> [01b, JF] -> [10b] -> [11b] -> sentinel + * ^ ^ + * [new0, N] ----+ | + * [new1, Inv] ------------------------------------------+ + * + * 2) Mark the first item of new1 with JOIN_NODE: + * [old, Inv] -> [00b] -> [01b, JF] -> [10b, JN] -> [11b] -> sentinel + * ^ ^ + * [new0, N] ----+ | + * [new1, Inv] ----------------------------------------------+ + * + * 3) Point new1 to the join-node and mark new1 NORMAL. + * [old, Inv] -> [00b] -> [01b, JF] -> [10b, JN] -> [11b] -> sentinel + * ^ ^ + * [new0, N] ----+ | + * [new1, N] --------------------------+ + * + * + * d) Fourth, the resizer cleans up extra marks added during bucket + * splits/joins but only when it is sure all updaters are accessing + * the table via the new bucket heads only (ie it is certain there + * are no delayed updaters unaware of the resize and accessing the + * table via the old bucket head). + * + * [old, Inv] ---+ + * v + * [new0, N] --> [00b] -> [01b, N] ---+ + * v + * [new1, N] --> [10b, N] -> [11b] -> sentinel + * + * + * e) Last, the resizer publishes the new bucket head array for everyone + * to see and use. This signals the end of the resize and the old bucket + * array is freed. + * + * + * To understand details of how the table is resized, read [1, 3, 5] + * and comments in join_buckets(), split_bucket(). + * + * + * [1] High performance dynamic lock-free hash tables and list-based sets, + * Michael, 2002 + * http://www.research.ibm.com/people/m/michael/spaa-2002.pdf + * [2] Lock-free linked lists using compare-and-swap, + * Valois, 1995 + * http://people.csail.mit.edu/bushl2/rpi/portfolio/lockfree-grape/documents/lock-free-linked-lists.pdf + * [3] Resizable, scalable, concurrent hash tables via relativistic programming, + * Triplett, 2011 + * http://www.usenix.org/event/atc11/tech/final_files/Triplett.pdf + * [4] Split-ordered Lists: Lock-free Extensible Hash Tables, + * Shavit, 2006 + * http://www.cs.ucf.edu/~dcm/Teaching/COT4810-Spring2011/Literature/SplitOrderedLists.pdf + * [5] Towards a Scalable Non-blocking Coding Style, + * Click, 2008 + * http://www.azulsystems.com/events/javaone_2008/2008_CodingNonBlock.pdf + */ + + +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +/* Logarithm of the min bucket count. Must be at least 3. 2^6 == 64 buckets. */ +#define CHT_MIN_ORDER 6 +/* Logarithm of the max bucket count. */ +#define CHT_MAX_ORDER (8 * sizeof(size_t)) +/* Minimum number of hash table buckets. */ +#define CHT_MIN_BUCKET_CNT (1 << CHT_MIN_ORDER) +/* Does not have to be a power of 2. */ +#define CHT_MAX_LOAD 2 + +typedef cht_ptr_t marked_ptr_t; +typedef bool (*equal_pred_t)(void *arg, const cht_link_t *item); + +/** The following mark items and bucket heads. + * + * They are stored in the two low order bits of the next item pointers. + * Some marks may be combined. Some marks share the same binary value and + * are distinguished only by context (eg bucket head vs an ordinary item), + * in particular by walk_mode_t. + */ +typedef enum mark { + /** Normal non-deleted item or a valid bucket head. */ + N_NORMAL = 0, + /** Logically deleted item that might have already been unlinked. + * + * May be combined with N_JOIN and N_JOIN_FOLLOWS. Applicable only + * to items; never to bucket heads. + * + * Once marked deleted an item remains marked deleted. + */ + N_DELETED = 1, + /** Immutable bucket head. + * + * The bucket is being moved or joined with another and its (old) head + * must not be modified. + * + * May be combined with N_INVALID. Applicable only to old bucket heads, + * ie cht_t.b and not cht_t.new_b. + */ + N_CONST = 1, + /** Invalid bucket head. The bucket head must not be modified. + * + * Old bucket heads (ie cht_t.b) are marked invalid if they have + * already been moved to cht_t.new_b or if the bucket had already + * been merged with another when shrinking the table. New bucket + * heads (ie cht_t.new_b) are marked invalid if the old bucket had + * not yet been moved or if an old bucket had not yet been split + * when growing the table. + */ + N_INVALID = 3, + /** The item is a join node, ie joining two buckets + * + * A join node is either the first node of the second part of + * a bucket to be split; or it is the first node of the bucket + * to be merged into/appended to/joined with another bucket. + * + * May be combined with N_DELETED. Applicable only to items, never + * to bucket heads. + * + * Join nodes are referred to from two different buckets and may, + * therefore, not be safely/atomically unlinked from both buckets. + * As a result join nodes are not unlinked but rather just marked + * deleted. Once resize completes join nodes marked deleted are + * garbage collected. + */ + N_JOIN = 2, + /** The next node is a join node and will soon be marked so. + * + * A join-follows node is the last node of the first part of bucket + * that is to be split, ie it is the last node that will remain + * in the same bucket after splitting it. + * + * May be combined with N_DELETED. Applicable to items as well + * as to bucket heads of the bucket to be split (but only in cht_t.new_b). + */ + N_JOIN_FOLLOWS = 2, + /** Bit mask to filter out the address to the next item from the next ptr. */ + N_MARK_MASK = 3 +} mark_t; + +/** Determines */ +typedef enum walk_mode { + /** The table is not resizing. */ + WM_NORMAL = 4, + /** The table is undergoing a resize. Join nodes may be encountered. */ + WM_LEAVE_JOIN, + /** The table is growing. A join-follows node may be encountered. */ + WM_MOVE_JOIN_FOLLOWS +} walk_mode_t; + +/** Bucket position window. */ +typedef struct wnd { + /** Pointer to cur's predecessor. */ + marked_ptr_t *ppred; + /** Current item. */ + cht_link_t *cur; + /** Last encountered item. Deleted or not. */ + cht_link_t *last; +} wnd_t; + + +/* Sentinel node used by all buckets. Stores the greatest possible hash value.*/ +static const cht_link_t sentinel = { + /* NULL and N_NORMAL */ + .link = 0 | N_NORMAL, + .hash = -1 +}; + + +static size_t size_to_order(size_t bucket_cnt, size_t min_order); +static cht_buckets_t *alloc_buckets(size_t order, bool set_invalid, + bool can_block); +static inline cht_link_t *find_lazy(cht_t *h, void *key); +static cht_link_t *search_bucket(cht_t *h, marked_ptr_t head, void *key, + size_t search_hash); +static cht_link_t *find_resizing(cht_t *h, void *key, size_t hash, + marked_ptr_t old_head, size_t old_idx); +static bool insert_impl(cht_t *h, cht_link_t *item, cht_link_t **dup_item); +static bool insert_at(cht_link_t *item, const wnd_t *wnd, walk_mode_t walk_mode, + bool *resizing); +static bool has_duplicate(cht_t *h, const cht_link_t *item, size_t hash, + cht_link_t *cur, cht_link_t **dup_item); +static cht_link_t *find_duplicate(cht_t *h, const cht_link_t *item, size_t hash, + cht_link_t *start); +static bool remove_pred(cht_t *h, size_t hash, equal_pred_t pred, void *pred_arg); +static bool delete_at(cht_t *h, wnd_t *wnd, walk_mode_t walk_mode, + bool *deleted_but_gc, bool *resizing); +static bool mark_deleted(cht_link_t *cur, walk_mode_t walk_mode, bool *resizing); +static bool unlink_from_pred(wnd_t *wnd, walk_mode_t walk_mode, bool *resizing); +static bool find_wnd_and_gc_pred(cht_t *h, size_t hash, walk_mode_t walk_mode, + equal_pred_t pred, void *pred_arg, wnd_t *wnd, bool *resizing); +static bool find_wnd_and_gc(cht_t *h, size_t hash, walk_mode_t walk_mode, + wnd_t *wnd, bool *resizing); +static bool gc_deleted_node(cht_t *h, walk_mode_t walk_mode, wnd_t *wnd, + bool *resizing); +static bool join_completed(cht_t *h, const wnd_t *wnd); +static void upd_resizing_head(cht_t *h, size_t hash, marked_ptr_t **phead, + bool *join_finishing, walk_mode_t *walk_mode); +static void item_removed(cht_t *h); +static void item_inserted(cht_t *h); +static void free_later(cht_t *h, cht_link_t *item); +static void help_head_move(marked_ptr_t *psrc_head, marked_ptr_t *pdest_head); +static void start_head_move(marked_ptr_t *psrc_head); +static void mark_const(marked_ptr_t *psrc_head); +static void complete_head_move(marked_ptr_t *psrc_head, marked_ptr_t *pdest_head); +static void split_bucket(cht_t *h, marked_ptr_t *psrc_head, + marked_ptr_t *pdest_head, size_t split_hash); +static void mark_join_follows(cht_t *h, marked_ptr_t *psrc_head, + size_t split_hash, wnd_t *wnd); +static void mark_join_node(cht_link_t *join_node); +static void join_buckets(cht_t *h, marked_ptr_t *psrc_head, + marked_ptr_t *pdest_head, size_t split_hash); +static void link_to_join_node(cht_t *h, marked_ptr_t *pdest_head, + cht_link_t *join_node, size_t split_hash); +static void resize_table(work_t *arg); +static void grow_table(cht_t *h); +static void shrink_table(cht_t *h); +static void cleanup_join_node(cht_t *h, marked_ptr_t *new_head); +static void clear_join_and_gc(cht_t *h, cht_link_t *join_node, + marked_ptr_t *new_head); +static void cleanup_join_follows(cht_t *h, marked_ptr_t *new_head); +static marked_ptr_t make_link(const cht_link_t *next, mark_t mark); +static cht_link_t * get_next(marked_ptr_t link); +static mark_t get_mark(marked_ptr_t link); +static void next_wnd(wnd_t *wnd); +static bool same_node_pred(void *node, const cht_link_t *item2); +static size_t calc_key_hash(cht_t *h, void *key); +static size_t node_hash(cht_t *h, const cht_link_t *item); +static size_t calc_node_hash(cht_t *h, const cht_link_t *item); +static void memoize_node_hash(cht_t *h, cht_link_t *item); +static size_t calc_split_hash(size_t split_idx, size_t order); +static size_t calc_bucket_idx(size_t hash, size_t order); +static size_t grow_to_split_idx(size_t old_idx); +static size_t grow_idx(size_t idx); +static size_t shrink_idx(size_t idx); +static marked_ptr_t cas_link(marked_ptr_t *link, const cht_link_t *cur_next, + mark_t cur_mark, const cht_link_t *new_next, mark_t new_mark); +static marked_ptr_t _cas_link(marked_ptr_t *link, marked_ptr_t cur, + marked_ptr_t new); +static void cas_order_barrier(void); + +static void dummy_remove_callback(cht_link_t *item) +{ + /* empty */ +} + +/** Creates a concurrent hash table. + * + * @param h Valid pointer to a cht_t instance. + * @param op Item specific operations. All operations are compulsory. + * @return True if successfully created the table. False otherwise. + */ +bool cht_create_simple(cht_t *h, cht_ops_t *op) +{ + return cht_create(h, 0, 0, 0, false, op); +} + +/** Creates a concurrent hash table. + * + * @param h Valid pointer to a cht_t instance. + * @param init_size The initial number of buckets the table should contain. + * The table may be shrunk below this value if deemed necessary. + * Uses the default value if 0. + * @param min_size Minimum number of buckets that the table should contain. + * The number of buckets never drops below this value, + * although it may be rounded up internally as appropriate. + * Uses the default value if 0. + * @param max_load Maximum average number of items per bucket that allowed + * before the table grows. + * @param can_block If true creating the table blocks until enough memory + * is available (possibly indefinitely). Otherwise, + * table creation does not block and returns immediately + * even if not enough memory is available. + * @param op Item specific operations. All operations are compulsory. + * @return True if successfully created the table. False otherwise. + */ +bool cht_create(cht_t *h, size_t init_size, size_t min_size, size_t max_load, + bool can_block, cht_ops_t *op) +{ + ASSERT(h); + ASSERT(op && op->hash && op->key_hash && op->equal && op->key_equal); + /* Memoized hashes are stored in the rcu_link.func function pointer. */ + ASSERT(sizeof(size_t) == sizeof(rcu_func_t)); + ASSERT(sentinel.hash == (uintptr_t)sentinel.rcu_link.func); + + /* All operations are compulsory. */ + if (!op || !op->hash || !op->key_hash || !op->equal || !op->key_equal) + return false; + + size_t min_order = size_to_order(min_size, CHT_MIN_ORDER); + size_t order = size_to_order(init_size, min_order); + + h->b = alloc_buckets(order, false, can_block); + + if (!h->b) + return false; + + h->max_load = (max_load == 0) ? CHT_MAX_LOAD : max_load; + h->min_order = min_order; + h->new_b = NULL; + h->op = op; + atomic_set(&h->item_cnt, 0); + atomic_set(&h->resize_reqs, 0); + + if (NULL == op->remove_callback) { + h->op->remove_callback = dummy_remove_callback; + } + + /* + * Cached item hashes are stored in item->rcu_link.func. Once the item + * is deleted rcu_link.func will contain the value of invalid_hash. + */ + h->invalid_hash = (uintptr_t)h->op->remove_callback; + + /* Ensure the initialization takes place before we start using the table. */ + write_barrier(); + + return true; +} + +/** Allocates and initializes 2^order buckets. + * + * All bucket heads are initialized to point to the sentinel node. + * + * @param order The number of buckets to allocate is 2^order. + * @param set_invalid Bucket heads are marked invalid if true; otherwise + * they are marked N_NORMAL. + * @param can_block If true memory allocation blocks until enough memory + * is available (possibly indefinitely). Otherwise, + * memory allocation does not block. + * @return Newly allocated and initialized buckets or NULL if not enough memory. + */ +static cht_buckets_t *alloc_buckets(size_t order, bool set_invalid, bool can_block) +{ + size_t bucket_cnt = (1 << order); + size_t bytes = + sizeof(cht_buckets_t) + (bucket_cnt - 1) * sizeof(marked_ptr_t); + cht_buckets_t *b = malloc(bytes, can_block ? 0 : FRAME_ATOMIC); + + if (!b) + return NULL; + + b->order = order; + + marked_ptr_t head_link = set_invalid + ? make_link(&sentinel, N_INVALID) + : make_link(&sentinel, N_NORMAL); + + for (size_t i = 0; i < bucket_cnt; ++i) { + b->head[i] = head_link; + } + + return b; +} + +/** Returns the smallest k such that bucket_cnt <= 2^k and min_order <= k.*/ +static size_t size_to_order(size_t bucket_cnt, size_t min_order) +{ + size_t order = min_order; + + /* Find a power of two such that bucket_cnt <= 2^order */ + do { + if (bucket_cnt <= ((size_t)1 << order)) + return order; + + ++order; + } while (order < CHT_MAX_ORDER); + + return order; +} + +/** Destroys a CHT successfully created via cht_create(). + * + * Waits for all outstanding concurrent operations to complete and + * frees internal allocated resources. The table is however not cleared + * and items already present in the table (if any) are leaked. + */ +void cht_destroy(cht_t *h) +{ + cht_destroy_unsafe(h); + + /* You must clear the table of items. Otherwise cht_destroy will leak. */ + ASSERT(atomic_get(&h->item_cnt) == 0); +} + +/** Destroys a successfully created CHT but does no error checking. */ +void cht_destroy_unsafe(cht_t *h) +{ + /* Wait for resize to complete. */ + while (0 < atomic_get(&h->resize_reqs)) { + rcu_barrier(); + } + + /* Wait for all remove_callback()s to complete. */ + rcu_barrier(); + + free(h->b); + h->b = NULL; +} + +/** Returns the first item equal to the search key or NULL if not found. + * + * The call must be enclosed in a rcu_read_lock() unlock() pair. The + * returned item is guaranteed to be allocated until rcu_read_unlock() + * although the item may be concurrently removed from the table by another + * cpu. + * + * Further items matching the key may be retrieved via cht_find_next(). + * + * cht_find() sees the effects of any completed cht_remove(), cht_insert(). + * If a concurrent remove or insert had not yet completed cht_find() may + * or may not see the effects of it (eg it may find an item being removed). + * + * @param h CHT to operate on. + * @param key Search key as defined by cht_ops_t.key_equal() and .key_hash(). + * @return First item equal to the key or NULL if such an item does not exist. + */ +cht_link_t *cht_find(cht_t *h, void *key) +{ + /* Make the most recent changes to the table visible. */ + read_barrier(); + return cht_find_lazy(h, key); +} + +/** Returns the first item equal to the search key or NULL if not found. + * + * Unlike cht_find(), cht_find_lazy() may not see the effects of + * cht_remove() or cht_insert() even though they have already completed. + * It may take a couple of milliseconds for those changes to propagate + * and become visible to cht_find_lazy(). On the other hand, cht_find_lazy() + * operates a bit faster than cht_find(). + * + * See cht_find() for more details. + */ +cht_link_t *cht_find_lazy(cht_t *h, void *key) +{ + return find_lazy(h, key); +} + +/** Finds the first item equal to the search key. */ +static inline cht_link_t *find_lazy(cht_t *h, void *key) +{ + ASSERT(h); + /* See docs to cht_find() and cht_find_lazy(). */ + ASSERT(rcu_read_locked()); + + size_t hash = calc_key_hash(h, key); + + cht_buckets_t *b = rcu_access(h->b); + size_t idx = calc_bucket_idx(hash, b->order); + /* + * No need for access_once. b->head[idx] will point to an allocated node + * even if marked invalid until we exit rcu read section. + */ + marked_ptr_t head = b->head[idx]; + + /* Undergoing a resize - take the slow path. */ + if (N_INVALID == get_mark(head)) + return find_resizing(h, key, hash, head, idx); + + return search_bucket(h, head, key, hash); +} + +/** Returns the next item matching \a item. + * + * Must be enclosed in a rcu_read_lock()/unlock() pair. Effects of + * completed cht_remove(), cht_insert() are guaranteed to be visible + * to cht_find_next(). + * + * See cht_find() for more details. + */ +cht_link_t *cht_find_next(cht_t *h, const cht_link_t *item) +{ + /* Make the most recent changes to the table visible. */ + read_barrier(); + return cht_find_next_lazy(h, item); +} + +/** Returns the next item matching \a item. + * + * Must be enclosed in a rcu_read_lock()/unlock() pair. Effects of + * completed cht_remove(), cht_insert() may or may not be visible + * to cht_find_next_lazy(). + * + * See cht_find_lazy() for more details. + */ +cht_link_t *cht_find_next_lazy(cht_t *h, const cht_link_t *item) +{ + ASSERT(h); + ASSERT(rcu_read_locked()); + ASSERT(item); + + return find_duplicate(h, item, calc_node_hash(h, item), get_next(item->link)); +} + +/** Searches the bucket at head for key using search_hash. */ +static inline cht_link_t *search_bucket(cht_t *h, marked_ptr_t head, void *key, + size_t search_hash) +{ + /* + * It is safe to access nodes even outside of this bucket (eg when + * splitting the bucket). The resizer makes sure that any node we + * may find by following the next pointers is allocated. + */ + + cht_link_t *cur = NULL; + marked_ptr_t prev = head; + +try_again: + /* Filter out items with different hashes. */ + do { + cur = get_next(prev); + ASSERT(cur); + prev = cur->link; + } while (node_hash(h, cur) < search_hash); + + /* + * Only search for an item with an equal key if cur is not the sentinel + * node or a node with a different hash. + */ + while (node_hash(h, cur) == search_hash) { + if (h->op->key_equal(key, cur)) { + if (!(N_DELETED & get_mark(cur->link))) + return cur; + } + + cur = get_next(cur->link); + ASSERT(cur); + } + + /* + * In the unlikely case that we have encountered a node whose cached + * hash has been overwritten due to a pending rcu_call for it, skip + * the node and try again. + */ + if (node_hash(h, cur) == h->invalid_hash) { + prev = cur->link; + goto try_again; + } + + return NULL; +} + +/** Searches for the key while the table is undergoing a resize. */ +static cht_link_t *find_resizing(cht_t *h, void *key, size_t hash, + marked_ptr_t old_head, size_t old_idx) +{ + ASSERT(N_INVALID == get_mark(old_head)); + ASSERT(h->new_b); + + size_t new_idx = calc_bucket_idx(hash, h->new_b->order); + marked_ptr_t new_head = h->new_b->head[new_idx]; + marked_ptr_t search_head = new_head; + + /* Growing. */ + if (h->b->order < h->new_b->order) { + /* + * Old bucket head is invalid, so it must have been already + * moved. Make the new head visible if still not visible, ie + * invalid. + */ + if (N_INVALID == get_mark(new_head)) { + /* + * We should be searching a newly added bucket but the old + * moved bucket has not yet been split (its marked invalid) + * or we have not yet seen the split. + */ + if (grow_idx(old_idx) != new_idx) { + /* + * Search the moved bucket. It is guaranteed to contain + * items of the newly added bucket that were present + * before the moved bucket was split. + */ + new_head = h->new_b->head[grow_idx(old_idx)]; + } + + /* new_head is now the moved bucket, either valid or invalid. */ + + /* + * The old bucket was definitely moved to new_head but the + * change of new_head had not yet propagated to this cpu. + */ + if (N_INVALID == get_mark(new_head)) { + /* + * We could issue a read_barrier() and make the now valid + * moved bucket head new_head visible, but instead fall back + * on using the old bucket. Although the old bucket head is + * invalid, it points to a node that is allocated and in the + * right bucket. Before the node can be freed, it must be + * unlinked from the head (or another item after that item + * modified the new_head) and a grace period must elapse. + * As a result had the node been already freed the grace + * period preceeding the free() would make the unlink and + * any changes to new_head visible. Therefore, it is safe + * to use the node pointed to from the old bucket head. + */ + + search_head = old_head; + } else { + search_head = new_head; + } + } + + return search_bucket(h, search_head, key, hash); + } else if (h->b->order > h->new_b->order) { + /* Shrinking. */ + + /* Index of the bucket in the old table that was moved. */ + size_t move_src_idx = grow_idx(new_idx); + marked_ptr_t moved_old_head = h->b->head[move_src_idx]; + + /* + * h->b->head[move_src_idx] had already been moved to new_head + * but the change to new_head had not yet propagated to us. + */ + if (N_INVALID == get_mark(new_head)) { + /* + * new_head is definitely valid and we could make it visible + * to this cpu with a read_barrier(). Instead, use the bucket + * in the old table that was moved even though it is now marked + * as invalid. The node it points to must be allocated because + * a grace period would have to elapse before it could be freed; + * and the grace period would make the now valid new_head + * visible to all cpus. + * + * Note that move_src_idx may not be the same as old_idx. + * If move_src_idx != old_idx then old_idx is the bucket + * in the old table that is not moved but instead it is + * appended to the moved bucket, ie it is added at the tail + * of new_head. In that case an invalid old_head notes that + * it had already been merged into (the moved) new_head. + * We will try to search that bucket first because it + * may contain some newly added nodes after the bucket + * join. Moreover, the bucket joining link may already be + * visible even if new_head is not. Therefore, if we're + * lucky we'll find the item via moved_old_head. In any + * case, we'll retry in proper old_head if not found. + */ + search_head = moved_old_head; + } + + cht_link_t *ret = search_bucket(h, search_head, key, hash); + + if (ret) + return ret; + /* + * Bucket old_head was already joined with moved_old_head + * in the new table but we have not yet seen change of the + * joining link (or the item is not in the table). + */ + if (move_src_idx != old_idx && get_next(old_head) != &sentinel) { + /* + * Note that old_head (the bucket to be merged into new_head) + * points to an allocated join node (if non-null) even if marked + * invalid. Before the resizer lets join nodes to be unlinked + * (and freed) it sets old_head to NULL and waits for a grace period. + * So either the invalid old_head points to join node; or old_head + * is null and we would have seen a completed bucket join while + * traversing search_head. + */ + ASSERT(N_JOIN & get_mark(get_next(old_head)->link)); + return search_bucket(h, old_head, key, hash); + } + + return NULL; + } else { + /* + * Resize is almost done. The resizer is waiting to make + * sure all cpus see that the new table replaced the old one. + */ + ASSERT(h->b->order == h->new_b->order); + /* + * The resizer must ensure all new bucket heads are visible before + * replacing the old table. + */ + ASSERT(N_NORMAL == get_mark(new_head)); + return search_bucket(h, new_head, key, hash); + } +} + +/** Inserts an item. Succeeds even if an equal item is already present. */ +void cht_insert(cht_t *h, cht_link_t *item) +{ + insert_impl(h, item, NULL); +} + +/** Inserts a unique item. Returns false if an equal item was already present. + * + * Use this function to atomically check if an equal/duplicate item had + * not yet been inserted into the table and to insert this item into the + * table. + * + * The following is @e NOT thread-safe, so do not use: + * @code + * if (!cht_find(h, key)) { + * // A concurrent insert here may go unnoticed by cht_find() above. + * item = malloc(..); + * cht_insert(h, item); + * // Now we may have two items with equal search keys. + * } + * @endcode + * + * Replace such code with: + * @code + * item = malloc(..); + * if (!cht_insert_unique(h, item, &dup_item)) { + * // Whoops, someone beat us to it - an equal item 'dup_item' + * // had already been inserted. + * free(item); + * } else { + * // Successfully inserted the item and we are guaranteed that + * // there are no other equal items. + * } + * @endcode + * + */ +bool cht_insert_unique(cht_t *h, cht_link_t *item, cht_link_t **dup_item) +{ + ASSERT(rcu_read_locked()); + ASSERT(dup_item); + return insert_impl(h, item, dup_item); +} + +/** Inserts the item into the table and checks for duplicates if dup_item. */ +static bool insert_impl(cht_t *h, cht_link_t *item, cht_link_t **dup_item) +{ + rcu_read_lock(); + + cht_buckets_t *b = rcu_access(h->b); + memoize_node_hash(h, item); + size_t hash = node_hash(h, item); + size_t idx = calc_bucket_idx(hash, b->order); + marked_ptr_t *phead = &b->head[idx]; + + bool resizing = false; + bool inserted = false; + + do { + walk_mode_t walk_mode = WM_NORMAL; + bool join_finishing; + + resizing = resizing || (N_NORMAL != get_mark(*phead)); + + /* The table is resizing. Get the correct bucket head. */ + if (resizing) { + upd_resizing_head(h, hash, &phead, &join_finishing, &walk_mode); + } + + wnd_t wnd = { + .ppred = phead, + .cur = get_next(*phead), + .last = NULL + }; + + if (!find_wnd_and_gc(h, hash, walk_mode, &wnd, &resizing)) { + /* Could not GC a node; or detected an unexpected resize. */ + continue; + } + + if (dup_item && has_duplicate(h, item, hash, wnd.cur, dup_item)) { + rcu_read_unlock(); + return false; + } + + inserted = insert_at(item, &wnd, walk_mode, &resizing); + } while (!inserted); + + rcu_read_unlock(); + + item_inserted(h); + return true; +} + +/** Inserts item between wnd.ppred and wnd.cur. + * + * @param item Item to link to wnd.ppred and wnd.cur. + * @param wnd The item will be inserted before wnd.cur. Wnd.ppred + * must be N_NORMAL. + * @param walk_mode + * @param resizing Set to true only if the table is undergoing resize + * and it was not expected (ie walk_mode == WM_NORMAL). + * @return True if the item was successfully linked to wnd.ppred. False + * if whole insert operation must be retried because the predecessor + * of wnd.cur has changed. + */ +inline static bool insert_at(cht_link_t *item, const wnd_t *wnd, + walk_mode_t walk_mode, bool *resizing) +{ + marked_ptr_t ret; + + if (walk_mode == WM_NORMAL) { + item->link = make_link(wnd->cur, N_NORMAL); + /* Initialize the item before adding it to a bucket. */ + memory_barrier(); + + /* Link a clean/normal predecessor to the item. */ + ret = cas_link(wnd->ppred, wnd->cur, N_NORMAL, item, N_NORMAL); + + if (ret == make_link(wnd->cur, N_NORMAL)) { + return true; + } else { + /* This includes an invalid head but not a const head. */ + *resizing = ((N_JOIN_FOLLOWS | N_JOIN) & get_mark(ret)); + return false; + } + } else if (walk_mode == WM_MOVE_JOIN_FOLLOWS) { + /* Move JOIN_FOLLOWS mark but filter out the DELETED mark. */ + mark_t jf_mark = get_mark(*wnd->ppred) & N_JOIN_FOLLOWS; + item->link = make_link(wnd->cur, jf_mark); + /* Initialize the item before adding it to a bucket. */ + memory_barrier(); + + /* Link the not-deleted predecessor to the item. Move its JF mark. */ + ret = cas_link(wnd->ppred, wnd->cur, jf_mark, item, N_NORMAL); + + return ret == make_link(wnd->cur, jf_mark); + } else { + ASSERT(walk_mode == WM_LEAVE_JOIN); + + item->link = make_link(wnd->cur, N_NORMAL); + /* Initialize the item before adding it to a bucket. */ + memory_barrier(); + + mark_t pred_mark = get_mark(*wnd->ppred); + /* If the predecessor is a join node it may be marked deleted.*/ + mark_t exp_pred_mark = (N_JOIN & pred_mark) ? pred_mark : N_NORMAL; + + ret = cas_link(wnd->ppred, wnd->cur, exp_pred_mark, item, exp_pred_mark); + return ret == make_link(wnd->cur, exp_pred_mark); + } +} + +/** Returns true if the chain starting at cur has an item equal to \a item. + * + * @param h CHT to operate on. + * @param item Item whose duplicates the function looks for. + * @param hash Hash of \a item. + * @param[in] cur The first node with a hash greater to or equal to item's hash. + * @param[out] dup_item The first duplicate item encountered. + * @return True if a non-deleted item equal to \a item exists in the table. + */ +static inline bool has_duplicate(cht_t *h, const cht_link_t *item, size_t hash, + cht_link_t *cur, cht_link_t **dup_item) +{ + ASSERT(cur); + ASSERT(cur == &sentinel || hash <= node_hash(h, cur) + || node_hash(h, cur) == h->invalid_hash); + + /* hash < node_hash(h, cur) */ + if (hash != node_hash(h, cur) && h->invalid_hash != node_hash(h, cur)) + return false; + + /* + * Load the most recent node marks. Otherwise we might pronounce a + * logically deleted node for a duplicate of the item just because + * the deleted node's DEL mark had not yet propagated to this cpu. + */ + read_barrier(); + + *dup_item = find_duplicate(h, item, hash, cur); + return NULL != *dup_item; +} + +/** Returns an item that is equal to \a item starting in a chain at \a start. */ +static cht_link_t *find_duplicate(cht_t *h, const cht_link_t *item, size_t hash, + cht_link_t *start) +{ + ASSERT(hash <= node_hash(h, start) || h->invalid_hash == node_hash(h, start)); + + cht_link_t *cur = start; + +try_again: + ASSERT(cur); + + while (node_hash(h, cur) == hash) { + ASSERT(cur != &sentinel); + + bool deleted = (N_DELETED & get_mark(cur->link)); + + /* Skip logically deleted nodes. */ + if (!deleted && h->op->equal(item, cur)) + return cur; + + cur = get_next(cur->link); + ASSERT(cur); + } + + /* Skip logically deleted nodes with rcu_call() in progress. */ + if (h->invalid_hash == node_hash(h, cur)) { + cur = get_next(cur->link); + goto try_again; + } + + return NULL; +} + +/** Removes all items matching the search key. Returns the number of items removed.*/ +size_t cht_remove_key(cht_t *h, void *key) +{ + ASSERT(h); + + size_t hash = calc_key_hash(h, key); + size_t removed = 0; + + while (remove_pred(h, hash, h->op->key_equal, key)) + ++removed; + + return removed; +} + +/** Removes a specific item from the table. + * + * The called must hold rcu read lock. + * + * @param item Item presumably present in the table and to be removed. + * @return True if the item was removed successfully; or false if it had + * already been deleted. + */ +bool cht_remove_item(cht_t *h, cht_link_t *item) +{ + ASSERT(h); + ASSERT(item); + /* Otherwise a concurrent cht_remove_key might free the item. */ + ASSERT(rcu_read_locked()); + + /* + * Even though we know the node we want to delete we must unlink it + * from the correct bucket and from a clean/normal predecessor. Therefore, + * we search for it again from the beginning of the correct bucket. + */ + size_t hash = calc_node_hash(h, item); + return remove_pred(h, hash, same_node_pred, item); +} + +/** Removes an item equal to pred_arg according to the predicate pred. */ +static bool remove_pred(cht_t *h, size_t hash, equal_pred_t pred, void *pred_arg) +{ + rcu_read_lock(); + + bool resizing = false; + bool deleted = false; + bool deleted_but_gc = false; + + cht_buckets_t *b = rcu_access(h->b); + size_t idx = calc_bucket_idx(hash, b->order); + marked_ptr_t *phead = &b->head[idx]; + + do { + walk_mode_t walk_mode = WM_NORMAL; + bool join_finishing = false; + + resizing = resizing || (N_NORMAL != get_mark(*phead)); + + /* The table is resizing. Get the correct bucket head. */ + if (resizing) { + upd_resizing_head(h, hash, &phead, &join_finishing, &walk_mode); + } + + wnd_t wnd = { + .ppred = phead, + .cur = get_next(*phead), + .last = NULL + }; + + if (!find_wnd_and_gc_pred( + h, hash, walk_mode, pred, pred_arg, &wnd, &resizing)) { + /* Could not GC a node; or detected an unexpected resize. */ + continue; + } + + /* + * The item lookup is affected by a bucket join but effects of + * the bucket join have not been seen while searching for the item. + */ + if (join_finishing && !join_completed(h, &wnd)) { + /* + * Bucket was appended at the end of another but the next + * ptr linking them together was not visible on this cpu. + * join_completed() makes this appended bucket visible. + */ + continue; + } + + /* Already deleted, but delete_at() requested one GC pass. */ + if (deleted_but_gc) + break; + + bool found = (wnd.cur != &sentinel && pred(pred_arg, wnd.cur)); + + if (!found) { + rcu_read_unlock(); + return false; + } + + deleted = delete_at(h, &wnd, walk_mode, &deleted_but_gc, &resizing); + } while (!deleted || deleted_but_gc); + + rcu_read_unlock(); + return true; +} + +/** Unlinks wnd.cur from wnd.ppred and schedules a deferred free for the item. + * + * Ignores nodes marked N_JOIN if walk mode is WM_LEAVE_JOIN. + * + * @param h CHT to operate on. + * @param wnd Points to the item to delete and its N_NORMAL predecessor. + * @param walk_mode Bucket chaing walk mode. + * @param deleted_but_gc Set to true if the item had been logically deleted, + * but a garbage collecting walk of the bucket is in order for + * it to be fully unlinked. + * @param resizing Set to true if the table is undergoing an unexpected + * resize (ie walk_mode == WM_NORMAL). + * @return False if the wnd.ppred changed in the meantime and the whole + * delete operation must be retried. + */ +static inline bool delete_at(cht_t *h, wnd_t *wnd, walk_mode_t walk_mode, + bool *deleted_but_gc, bool *resizing) +{ + ASSERT(wnd->cur && wnd->cur != &sentinel); + + *deleted_but_gc = false; + + if (!mark_deleted(wnd->cur, walk_mode, resizing)) { + /* Already deleted, or unexpectedly marked as JOIN/JOIN_FOLLOWS. */ + return false; + } + + /* Marked deleted. Unlink from the bucket. */ + + /* Never unlink join nodes. */ + if (walk_mode == WM_LEAVE_JOIN && (N_JOIN & get_mark(wnd->cur->link))) + return true; + + cas_order_barrier(); + + if (unlink_from_pred(wnd, walk_mode, resizing)) { + free_later(h, wnd->cur); + } else { + *deleted_but_gc = true; + } + + return true; +} + +/** Marks cur logically deleted. Returns false to request a retry. */ +static inline bool mark_deleted(cht_link_t *cur, walk_mode_t walk_mode, + bool *resizing) +{ + ASSERT(cur && cur != &sentinel); + + /* + * Btw, we could loop here if the cas fails but let's not complicate + * things and let's retry from the head of the bucket. + */ + + cht_link_t *next = get_next(cur->link); + + if (walk_mode == WM_NORMAL) { + /* Only mark clean/normal nodes - JF/JN is used only during resize. */ + marked_ptr_t ret = cas_link(&cur->link, next, N_NORMAL, next, N_DELETED); + + if (ret != make_link(next, N_NORMAL)) { + *resizing = (N_JOIN | N_JOIN_FOLLOWS) & get_mark(ret); + return false; + } + } else { + ASSERT(N_JOIN == N_JOIN_FOLLOWS); + + /* Keep the N_JOIN/N_JOIN_FOLLOWS mark but strip N_DELETED. */ + mark_t cur_mark = get_mark(cur->link) & N_JOIN_FOLLOWS; + + marked_ptr_t ret = + cas_link(&cur->link, next, cur_mark, next, cur_mark | N_DELETED); + + if (ret != make_link(next, cur_mark)) + return false; + } + + return true; +} + +/** Unlinks wnd.cur from wnd.ppred. Returns false if it should be retried. */ +static inline bool unlink_from_pred(wnd_t *wnd, walk_mode_t walk_mode, + bool *resizing) +{ + ASSERT(wnd->cur != &sentinel); + ASSERT(wnd->cur && (N_DELETED & get_mark(wnd->cur->link))); + + cht_link_t *next = get_next(wnd->cur->link); + + if (walk_mode == WM_LEAVE_JOIN) { + /* Never try to unlink join nodes. */ + ASSERT(!(N_JOIN & get_mark(wnd->cur->link))); + + mark_t pred_mark = get_mark(*wnd->ppred); + /* Succeed only if the predecessor is clean/normal or a join node. */ + mark_t exp_pred_mark = (N_JOIN & pred_mark) ? pred_mark : N_NORMAL; + + marked_ptr_t pred_link = make_link(wnd->cur, exp_pred_mark); + marked_ptr_t next_link = make_link(next, exp_pred_mark); + + if (pred_link != _cas_link(wnd->ppred, pred_link, next_link)) + return false; + } else { + ASSERT(walk_mode == WM_MOVE_JOIN_FOLLOWS || walk_mode == WM_NORMAL); + /* Move the JF mark if set. Clear DEL mark. */ + mark_t cur_mark = N_JOIN_FOLLOWS & get_mark(wnd->cur->link); + + /* The predecessor must be clean/normal. */ + marked_ptr_t pred_link = make_link(wnd->cur, N_NORMAL); + /* Link to cur's successor keeping/copying cur's JF mark. */ + marked_ptr_t next_link = make_link(next, cur_mark); + + marked_ptr_t ret = _cas_link(wnd->ppred, pred_link, next_link); + + if (pred_link != ret) { + /* If we're not resizing the table there are no JF/JN nodes. */ + *resizing = (walk_mode == WM_NORMAL) + && (N_JOIN_FOLLOWS & get_mark(ret)); + return false; + } + } + + return true; +} + +/** Finds the first non-deleted item equal to \a pred_arg according to \a pred. + * + * The function returns the candidate item in \a wnd. Logically deleted + * nodes are garbage collected so the predecessor will most likely not + * be marked as deleted. + * + * Unlike find_wnd_and_gc(), this function never returns a node that + * is known to have already been marked N_DELETED. + * + * Any logically deleted nodes (ie those marked N_DELETED) are garbage + * collected, ie free in the background via rcu_call (except for join-nodes + * if walk_mode == WM_LEAVE_JOIN). + * + * @param h CHT to operate on. + * @param hash Hash the search for. + * @param walk_mode Bucket chain walk mode. + * @param pred Predicate used to find an item equal to pred_arg. + * @param pred_arg Argument to pass to the equality predicate \a pred. + * @param[in,out] wnd The search starts with wnd.cur. If the desired + * item is found wnd.cur will point to it. + * @param resizing Set to true if the table is resizing but it was not + * expected (ie walk_mode == WM_NORMAL). + * @return False if the operation has to be retried. True otherwise + * (even if an equal item had not been found). + */ +static bool find_wnd_and_gc_pred(cht_t *h, size_t hash, walk_mode_t walk_mode, + equal_pred_t pred, void *pred_arg, wnd_t *wnd, bool *resizing) +{ + ASSERT(wnd->cur); + + if (wnd->cur == &sentinel) + return true; + + /* + * A read barrier is not needed here to bring up the most recent + * node marks (esp the N_DELETED). At worst we'll try to delete + * an already deleted node; fail in delete_at(); and retry. + */ + + size_t cur_hash; + +try_again: + cur_hash = node_hash(h, wnd->cur); + + while (cur_hash <= hash) { + ASSERT(wnd->cur && wnd->cur != &sentinel); + + /* GC any deleted nodes on the way. */ + if (N_DELETED & get_mark(wnd->cur->link)) { + if (!gc_deleted_node(h, walk_mode, wnd, resizing)) { + /* Retry from the head of a bucket. */ + return false; + } + } else { + /* Is this the node we were looking for? */ + if (cur_hash == hash && pred(pred_arg, wnd->cur)) + return true; + + next_wnd(wnd); + } + + cur_hash = node_hash(h, wnd->cur); + } + + if (cur_hash == h->invalid_hash) { + next_wnd(wnd); + ASSERT(wnd->cur); + goto try_again; + } + + /* The searched for node is not in the current bucket. */ + return true; +} + +/** Find the first item (deleted or not) with a hash greater or equal to \a hash. + * + * The function returns the first item with a hash that is greater or + * equal to \a hash in \a wnd. Moreover it garbage collects logically + * deleted node that have not yet been unlinked and freed. Therefore, + * the returned node's predecessor will most likely be N_NORMAL. + * + * Unlike find_wnd_and_gc_pred(), this function may return a node + * that is known to had been marked N_DELETED. + * + * @param h CHT to operate on. + * @param hash Hash of the item to find. + * @param walk_mode Bucket chain walk mode. + * @param[in,out] wnd wnd.cur denotes the first node of the chain. If the + * the operation is successful, \a wnd points to the desired + * item. + * @param resizing Set to true if a table resize was detected but walk_mode + * suggested the table was not undergoing a resize. + * @return False indicates the operation must be retried. True otherwise + * (even if an item with exactly the same has was not found). + */ +static bool find_wnd_and_gc(cht_t *h, size_t hash, walk_mode_t walk_mode, + wnd_t *wnd, bool *resizing) +{ +try_again: + ASSERT(wnd->cur); + + while (node_hash(h, wnd->cur) < hash) { + /* GC any deleted nodes along the way to our desired node. */ + if (N_DELETED & get_mark(wnd->cur->link)) { + if (!gc_deleted_node(h, walk_mode, wnd, resizing)) { + /* Failed to remove the garbage node. Retry. */ + return false; + } + } else { + next_wnd(wnd); + } + + ASSERT(wnd->cur); + } + + if (node_hash(h, wnd->cur) == h->invalid_hash) { + next_wnd(wnd); + goto try_again; + } + + /* wnd->cur may be NULL or even marked N_DELETED. */ + return true; +} + +/** Garbage collects the N_DELETED node at \a wnd skipping join nodes. */ +static bool gc_deleted_node(cht_t *h, walk_mode_t walk_mode, wnd_t *wnd, + bool *resizing) +{ + ASSERT(N_DELETED & get_mark(wnd->cur->link)); + + /* Skip deleted JOIN nodes. */ + if (walk_mode == WM_LEAVE_JOIN && (N_JOIN & get_mark(wnd->cur->link))) { + next_wnd(wnd); + } else { + /* Ordinary deleted node or a deleted JOIN_FOLLOWS. */ + ASSERT(walk_mode != WM_LEAVE_JOIN + || !((N_JOIN | N_JOIN_FOLLOWS) & get_mark(wnd->cur->link))); + + /* Unlink an ordinary deleted node, move JOIN_FOLLOWS mark. */ + if (!unlink_from_pred(wnd, walk_mode, resizing)) { + /* Retry. The predecessor was deleted, invalid, const, join_follows. */ + return false; + } + + free_later(h, wnd->cur); + + /* Leave ppred as is. */ + wnd->last = wnd->cur; + wnd->cur = get_next(wnd->cur->link); + } + + return true; +} + +/** Returns true if a bucket join had already completed. + * + * May only be called if upd_resizing_head() indicates a bucket join + * may be in progress. + * + * If it returns false, the search must be retried in order to guarantee + * all item that should have been encountered have been seen. + */ +static bool join_completed(cht_t *h, const wnd_t *wnd) +{ + /* + * The table is shrinking and the searched for item is in a bucket + * appended to another. Check that the link joining these two buckets + * is visible and if not, make it visible to this cpu. + */ + + /* + * Resizer ensures h->b->order stays the same for the duration of this + * func. We got here because there was an alternative head to search. + * The resizer waits for all preexisting readers to finish after + * it + */ + ASSERT(h->b->order > h->new_b->order); + ASSERT(wnd->cur); + + /* Either we did not need the joining link or we have already followed it.*/ + if (wnd->cur != &sentinel) + return true; + + /* We have reached the end of a bucket. */ + + if (wnd->last != &sentinel) { + size_t last_seen_hash = node_hash(h, wnd->last); + + if (last_seen_hash == h->invalid_hash) { + last_seen_hash = calc_node_hash(h, wnd->last); + } + + size_t last_old_idx = calc_bucket_idx(last_seen_hash, h->b->order); + size_t move_src_idx = grow_idx(shrink_idx(last_old_idx)); + + /* + * Last node seen was in the joining bucket - if the searched + * for node is there we will find it. + */ + if (move_src_idx != last_old_idx) + return true; + } + + /* + * Reached the end of the bucket but no nodes from the joining bucket + * were seen. There should have at least been a JOIN node so we have + * definitely not seen (and followed) the joining link. Make the link + * visible and retry. + */ + read_barrier(); + return false; +} + +/** When resizing returns the bucket head to start the search with in \a phead. + * + * If a resize had been detected (eg cht_t.b.head[idx] is marked immutable). + * upd_resizing_head() moves the bucket for \a hash from the old head + * to the new head. Moreover, it splits or joins buckets as necessary. + * + * @param h CHT to operate on. + * @param hash Hash of an item whose chain we would like to traverse. + * @param[out] phead Head of the bucket to search for \a hash. + * @param[out] join_finishing Set to true if a bucket join might be + * in progress and the bucket may have to traversed again + * as indicated by join_completed(). + * @param[out] walk_mode Specifies how to interpret node marks. + */ +static void upd_resizing_head(cht_t *h, size_t hash, marked_ptr_t **phead, + bool *join_finishing, walk_mode_t *walk_mode) +{ + cht_buckets_t *b = rcu_access(h->b); + size_t old_idx = calc_bucket_idx(hash, b->order); + size_t new_idx = calc_bucket_idx(hash, h->new_b->order); + + marked_ptr_t *pold_head = &b->head[old_idx]; + marked_ptr_t *pnew_head = &h->new_b->head[new_idx]; + + /* In any case, use the bucket in the new table. */ + *phead = pnew_head; + + /* Growing the table. */ + if (b->order < h->new_b->order) { + size_t move_dest_idx = grow_idx(old_idx); + marked_ptr_t *pmoved_head = &h->new_b->head[move_dest_idx]; + + /* Complete moving the bucket from the old to the new table. */ + help_head_move(pold_head, pmoved_head); + + /* The hash belongs to the moved bucket. */ + if (move_dest_idx == new_idx) { + ASSERT(pmoved_head == pnew_head); + /* + * move_head() makes the new head of the moved bucket visible. + * The new head may be marked with a JOIN_FOLLOWS + */ + ASSERT(!(N_CONST & get_mark(*pmoved_head))); + *walk_mode = WM_MOVE_JOIN_FOLLOWS; + } else { + ASSERT(pmoved_head != pnew_head); + /* + * The hash belongs to the bucket that is the result of splitting + * the old/moved bucket, ie the bucket that contains the second + * half of the split/old/moved bucket. + */ + + /* The moved bucket has not yet been split. */ + if (N_NORMAL != get_mark(*pnew_head)) { + size_t split_hash = calc_split_hash(new_idx, h->new_b->order); + split_bucket(h, pmoved_head, pnew_head, split_hash); + /* + * split_bucket() makes the new head visible. No + * JOIN_FOLLOWS in this part of split bucket. + */ + ASSERT(N_NORMAL == get_mark(*pnew_head)); + } + + *walk_mode = WM_LEAVE_JOIN; + } + } else if (h->new_b->order < b->order ) { + /* Shrinking the table. */ + + size_t move_src_idx = grow_idx(new_idx); + + /* + * Complete moving the bucket from the old to the new table. + * Makes a valid pnew_head visible if already moved. + */ + help_head_move(&b->head[move_src_idx], pnew_head); + + /* Hash belongs to the bucket to be joined with the moved bucket. */ + if (move_src_idx != old_idx) { + /* Bucket join not yet completed. */ + if (N_INVALID != get_mark(*pold_head)) { + size_t split_hash = calc_split_hash(old_idx, b->order); + join_buckets(h, pold_head, pnew_head, split_hash); + } + + /* + * The resizer sets pold_head to &sentinel when all cpus are + * guaranteed to see the bucket join. + */ + *join_finishing = (&sentinel != get_next(*pold_head)); + } + + /* move_head() or join_buckets() makes it so or makes the mark visible.*/ + ASSERT(N_INVALID == get_mark(*pold_head)); + /* move_head() makes it visible. No JOIN_FOLLOWS used when shrinking. */ + ASSERT(N_NORMAL == get_mark(*pnew_head)); + + *walk_mode = WM_LEAVE_JOIN; + } else { + /* + * Final stage of resize. The resizer is waiting for all + * readers to notice that the old table had been replaced. + */ + ASSERT(b == h->new_b); + *walk_mode = WM_NORMAL; + } +} + + +#if 0 +static void move_head(marked_ptr_t *psrc_head, marked_ptr_t *pdest_head) +{ + start_head_move(psrc_head); + cas_order_barrier(); + complete_head_move(psrc_head, pdest_head); +} +#endif + +/** Moves an immutable head \a psrc_head of cht_t.b to \a pdest_head of cht_t.new_b. + * + * The function guarantees the move will be visible on this cpu once + * it completes. In particular, *pdest_head will not be N_INVALID. + * + * Unlike complete_head_move(), help_head_move() checks if the head had already + * been moved and tries to avoid moving the bucket heads if possible. + */ +static inline void help_head_move(marked_ptr_t *psrc_head, + marked_ptr_t *pdest_head) +{ + /* Head move has to in progress already when calling this func. */ + ASSERT(N_CONST & get_mark(*psrc_head)); + + /* Head already moved. */ + if (N_INVALID == get_mark(*psrc_head)) { + /* Effects of the head move have not yet propagated to this cpu. */ + if (N_INVALID == get_mark(*pdest_head)) { + /* Make the move visible on this cpu. */ + read_barrier(); + } + } else { + complete_head_move(psrc_head, pdest_head); + } + + ASSERT(!(N_CONST & get_mark(*pdest_head))); +} + +/** Initiates the move of the old head \a psrc_head. + * + * The move may be completed with help_head_move(). + */ +static void start_head_move(marked_ptr_t *psrc_head) +{ + /* Mark src head immutable. */ + mark_const(psrc_head); +} + +/** Marks the head immutable. */ +static void mark_const(marked_ptr_t *psrc_head) +{ + marked_ptr_t ret, src_link; + + /* Mark src head immutable. */ + do { + cht_link_t *next = get_next(*psrc_head); + src_link = make_link(next, N_NORMAL); + + /* Mark the normal/clean src link immutable/const. */ + ret = cas_link(psrc_head, next, N_NORMAL, next, N_CONST); + } while(ret != src_link && !(N_CONST & get_mark(ret))); +} + +/** Completes moving head psrc_head to pdest_head (started by start_head_move()).*/ +static void complete_head_move(marked_ptr_t *psrc_head, marked_ptr_t *pdest_head) +{ + ASSERT(N_JOIN_FOLLOWS != get_mark(*psrc_head)); + ASSERT(N_CONST & get_mark(*psrc_head)); + + cht_link_t *next = get_next(*psrc_head); + marked_ptr_t ret; + + ret = cas_link(pdest_head, &sentinel, N_INVALID, next, N_NORMAL); + ASSERT(ret == make_link(&sentinel, N_INVALID) || (N_NORMAL == get_mark(ret))); + cas_order_barrier(); + + ret = cas_link(psrc_head, next, N_CONST, next, N_INVALID); + ASSERT(ret == make_link(next, N_CONST) || (N_INVALID == get_mark(ret))); + cas_order_barrier(); +} + +/** Splits the bucket at psrc_head and links to the remainder from pdest_head. + * + * Items with hashes greater or equal to \a split_hash are moved to bucket + * with head at \a pdest_head. + * + * @param h CHT to operate on. + * @param psrc_head Head of the bucket to split (in cht_t.new_b). + * @param pdest_head Head of the bucket that points to the second part + * of the split bucket in psrc_head. (in cht_t.new_b) + * @param split_hash Hash of the first possible item in the remainder of + * psrc_head, ie the smallest hash pdest_head is allowed + * to point to.. + */ +static void split_bucket(cht_t *h, marked_ptr_t *psrc_head, + marked_ptr_t *pdest_head, size_t split_hash) +{ + /* Already split. */ + if (N_NORMAL == get_mark(*pdest_head)) + return; + + /* + * L == Last node of the first part of the split bucket. That part + * remains in the original/src bucket. + * F == First node of the second part of the split bucket. That part + * will be referenced from the dest bucket head. + * + * We want to first mark a clean L as JF so that updaters unaware of + * the split (or table resize): + * - do not insert a new node between L and F + * - do not unlink L (that is why it has to be clean/normal) + * - do not unlink F + * + * Then we can safely mark F as JN even if it has been marked deleted. + * Once F is marked as JN updaters aware of table resize will not + * attempt to unlink it (JN will have two predecessors - we cannot + * safely unlink from both at the same time). Updaters unaware of + * ongoing resize can reach F only via L and that node is already + * marked JF, so they won't unlink F. + * + * Last, link the new/dest head to F. + * + * + * 0) ,-- split_hash, first hash of the dest bucket + * v + * [src_head | N] -> .. -> [L] -> [F] + * [dest_head | Inv] + * + * 1) ,-- split_hash + * v + * [src_head | N] -> .. -> [JF] -> [F] + * [dest_head | Inv] + * + * 2) ,-- split_hash + * v + * [src_head | N] -> .. -> [JF] -> [JN] + * [dest_head | Inv] + * + * 3) ,-- split_hash + * v + * [src_head | N] -> .. -> [JF] -> [JN] + * ^ + * [dest_head | N] -----------------' + */ + wnd_t wnd; + + rcu_read_lock(); + + /* Mark the last node of the first part of the split bucket as JF. */ + mark_join_follows(h, psrc_head, split_hash, &wnd); + cas_order_barrier(); + + /* There are nodes in the dest bucket, ie the second part of the split. */ + if (wnd.cur != &sentinel) { + /* + * Mark the first node of the dest bucket as a join node so + * updaters do not attempt to unlink it if it is deleted. + */ + mark_join_node(wnd.cur); + cas_order_barrier(); + } else { + /* + * Second part of the split bucket is empty. There are no nodes + * to mark as JOIN nodes and there never will be. + */ + } + + /* Link the dest head to the second part of the split. */ + marked_ptr_t ret = + cas_link(pdest_head, &sentinel, N_INVALID, wnd.cur, N_NORMAL); + ASSERT(ret == make_link(&sentinel, N_INVALID) || (N_NORMAL == get_mark(ret))); + cas_order_barrier(); + + rcu_read_unlock(); +} + +/** Finds and marks the last node of psrc_head w/ hash less than split_hash. + * + * Finds a node in psrc_head with the greatest hash that is strictly less + * than split_hash and marks it with N_JOIN_FOLLOWS. + * + * Returns a window pointing to that node. + * + * Any logically deleted nodes along the way are + * garbage collected; therefore, the predecessor node (if any) will most + * likely not be marked N_DELETED. + * + * @param h CHT to operate on. + * @param psrc_head Bucket head. + * @param split_hash The smallest hash a join node (ie the node following + * the desired join-follows node) may have. + * @param[out] wnd Points to the node marked with N_JOIN_FOLLOWS. + */ +static void mark_join_follows(cht_t *h, marked_ptr_t *psrc_head, + size_t split_hash, wnd_t *wnd) +{ + /* See comment in split_bucket(). */ + + bool done; + do { + bool resizing = false; + wnd->ppred = psrc_head; + wnd->cur = get_next(*psrc_head); + + /* + * Find the split window, ie the last node of the first part of + * the split bucket and the its successor - the first node of + * the second part of the split bucket. Retry if GC failed. + */ + if (!find_wnd_and_gc(h, split_hash, WM_MOVE_JOIN_FOLLOWS, wnd, &resizing)) + continue; + + /* Must not report that the table is resizing if WM_MOVE_JOIN_FOLLOWS.*/ + ASSERT(!resizing); + /* + * Mark the last node of the first half of the split bucket + * that a join node follows. It must be clean/normal. + */ + marked_ptr_t ret + = cas_link(wnd->ppred, wnd->cur, N_NORMAL, wnd->cur, N_JOIN_FOLLOWS); + + /* + * Successfully marked as a JF node or already marked that way (even + * if also marked deleted - unlinking the node will move the JF mark). + */ + done = (ret == make_link(wnd->cur, N_NORMAL)) + || (N_JOIN_FOLLOWS & get_mark(ret)); + } while (!done); +} + +/** Marks join_node with N_JOIN. */ +static void mark_join_node(cht_link_t *join_node) +{ + /* See comment in split_bucket(). */ + + bool done; + do { + cht_link_t *next = get_next(join_node->link); + mark_t mark = get_mark(join_node->link); + + /* + * May already be marked as deleted, but it won't be unlinked + * because its predecessor is marked with JOIN_FOLLOWS or CONST. + */ + marked_ptr_t ret + = cas_link(&join_node->link, next, mark, next, mark | N_JOIN); + + /* Successfully marked or already marked as a join node. */ + done = (ret == make_link(next, mark)) + || (N_JOIN & get_mark(ret)); + } while(!done); +} + +/** Appends the bucket at psrc_head to the bucket at pdest_head. + * + * @param h CHT to operate on. + * @param psrc_head Bucket to merge with pdest_head. + * @param pdest_head Bucket to be joined by psrc_head. + * @param split_hash The smallest hash psrc_head may contain. + */ +static void join_buckets(cht_t *h, marked_ptr_t *psrc_head, + marked_ptr_t *pdest_head, size_t split_hash) +{ + /* Buckets already joined. */ + if (N_INVALID == get_mark(*psrc_head)) + return; + /* + * F == First node of psrc_head, ie the bucket we want to append + * to (ie join with) the bucket starting at pdest_head. + * L == Last node of pdest_head, ie the bucket that psrc_head will + * be appended to. + * + * (1) We first mark psrc_head immutable to signal that a join is + * in progress and so that updaters unaware of the join (or table + * resize): + * - do not insert new nodes between the head psrc_head and F + * - do not unlink F (it may already be marked deleted) + * + * (2) Next, F is marked as a join node. Updaters aware of table resize + * will not attempt to unlink it. We cannot safely/atomically unlink + * the join node because it will be pointed to from two different + * buckets. Updaters unaware of resize will fail to unlink the join + * node due to the head being marked immutable. + * + * (3) Then the tail of the bucket at pdest_head is linked to the join + * node. From now on, nodes in both buckets can be found via pdest_head. + * + * (4) Last, mark immutable psrc_head as invalid. It signals updaters + * that the join is complete and they can insert new nodes (originally + * destined for psrc_head) into pdest_head. + * + * Note that pdest_head keeps pointing at the join node. This allows + * lookups and updaters to determine if they should see a link between + * the tail L and F when searching for nodes originally in psrc_head + * via pdest_head. If they reach the tail of pdest_head without + * encountering any nodes of psrc_head, either there were no nodes + * in psrc_head to begin with or the link between L and F did not + * yet propagate to their cpus. If psrc_head was empty, it remains + * NULL. Otherwise psrc_head points to a join node (it will not be + * unlinked until table resize completes) and updaters/lookups + * should issue a read_barrier() to make the link [L]->[JN] visible. + * + * 0) ,-- split_hash, first hash of the src bucket + * v + * [dest_head | N]-> .. -> [L] + * [src_head | N]--> [F] -> .. + * ^ + * ` split_hash, first hash of the src bucket + * + * 1) ,-- split_hash + * v + * [dest_head | N]-> .. -> [L] + * [src_head | C]--> [F] -> .. + * + * 2) ,-- split_hash + * v + * [dest_head | N]-> .. -> [L] + * [src_head | C]--> [JN] -> .. + * + * 3) ,-- split_hash + * v + * [dest_head | N]-> .. -> [L] --+ + * v + * [src_head | C]-------------> [JN] -> .. + * + * 4) ,-- split_hash + * v + * [dest_head | N]-> .. -> [L] --+ + * v + * [src_head | Inv]-----------> [JN] -> .. + */ + + rcu_read_lock(); + + /* Mark src_head immutable - signals updaters that bucket join started. */ + mark_const(psrc_head); + cas_order_barrier(); + + cht_link_t *join_node = get_next(*psrc_head); + + if (join_node != &sentinel) { + mark_join_node(join_node); + cas_order_barrier(); + + link_to_join_node(h, pdest_head, join_node, split_hash); + cas_order_barrier(); + } + + marked_ptr_t ret = + cas_link(psrc_head, join_node, N_CONST, join_node, N_INVALID); + ASSERT(ret == make_link(join_node, N_CONST) || (N_INVALID == get_mark(ret))); + cas_order_barrier(); + + rcu_read_unlock(); +} + +/** Links the tail of pdest_head to join_node. + * + * @param h CHT to operate on. + * @param pdest_head Head of the bucket whose tail is to be linked to join_node. + * @param join_node A node marked N_JOIN with a hash greater or equal to + * split_hash. + * @param split_hash The least hash that is greater than the hash of any items + * (originally) in pdest_head. + */ +static void link_to_join_node(cht_t *h, marked_ptr_t *pdest_head, + cht_link_t *join_node, size_t split_hash) +{ + bool done; + do { + wnd_t wnd = { + .ppred = pdest_head, + .cur = get_next(*pdest_head) + }; + + bool resizing = false; + + if (!find_wnd_and_gc(h, split_hash, WM_LEAVE_JOIN, &wnd, &resizing)) + continue; + + ASSERT(!resizing); + + if (wnd.cur != &sentinel) { + /* Must be from the new appended bucket. */ + ASSERT(split_hash <= node_hash(h, wnd.cur) + || h->invalid_hash == node_hash(h, wnd.cur)); + return; + } + + /* Reached the tail of pdest_head - link it to the join node. */ + marked_ptr_t ret = + cas_link(wnd.ppred, &sentinel, N_NORMAL, join_node, N_NORMAL); + + done = (ret == make_link(&sentinel, N_NORMAL)); + } while (!done); +} + +/** Instructs RCU to free the item once all preexisting references are dropped. + * + * The item is freed via op->remove_callback(). + */ +static void free_later(cht_t *h, cht_link_t *item) +{ + ASSERT(item != &sentinel); + + /* + * remove_callback only works as rcu_func_t because rcu_link is the first + * field in cht_link_t. + */ + rcu_call(&item->rcu_link, (rcu_func_t)h->op->remove_callback); + + item_removed(h); +} + +/** Notes that an item had been unlinked from the table and shrinks it if needed. + * + * If the number of items in the table drops below 1/4 of the maximum + * allowed load the table is shrunk in the background. + */ +static inline void item_removed(cht_t *h) +{ + size_t items = (size_t) atomic_predec(&h->item_cnt); + size_t bucket_cnt = (1 << h->b->order); + + bool need_shrink = (items == h->max_load * bucket_cnt / 4); + bool missed_shrink = (items == h->max_load * bucket_cnt / 8); + + if ((need_shrink || missed_shrink) && h->b->order > h->min_order) { + atomic_count_t resize_reqs = atomic_preinc(&h->resize_reqs); + /* The first resize request. Start the resizer. */ + if (1 == resize_reqs) { + workq_global_enqueue_noblock(&h->resize_work, resize_table); + } + } +} + +/** Notes an item had been inserted and grows the table if needed. + * + * The table is resized in the background. + */ +static inline void item_inserted(cht_t *h) +{ + size_t items = (size_t) atomic_preinc(&h->item_cnt); + size_t bucket_cnt = (1 << h->b->order); + + bool need_grow = (items == h->max_load * bucket_cnt); + bool missed_grow = (items == 2 * h->max_load * bucket_cnt); + + if ((need_grow || missed_grow) && h->b->order < CHT_MAX_ORDER) { + atomic_count_t resize_reqs = atomic_preinc(&h->resize_reqs); + /* The first resize request. Start the resizer. */ + if (1 == resize_reqs) { + workq_global_enqueue_noblock(&h->resize_work, resize_table); + } + } +} + +/** Resize request handler. Invoked on the system work queue. */ +static void resize_table(work_t *arg) +{ + cht_t *h = member_to_inst(arg, cht_t, resize_work); + +#ifdef CONFIG_DEBUG + ASSERT(h->b); + /* Make resize_reqs visible. */ + read_barrier(); + ASSERT(0 < atomic_get(&h->resize_reqs)); +#endif + + bool done; + do { + /* Load the most recent h->item_cnt. */ + read_barrier(); + size_t cur_items = (size_t) atomic_get(&h->item_cnt); + size_t bucket_cnt = (1 << h->b->order); + size_t max_items = h->max_load * bucket_cnt; + + if (cur_items >= max_items && h->b->order < CHT_MAX_ORDER) { + grow_table(h); + } else if (cur_items <= max_items / 4 && h->b->order > h->min_order) { + shrink_table(h); + } else { + /* Table is just the right size. */ + atomic_count_t reqs = atomic_predec(&h->resize_reqs); + done = (reqs == 0); + } + } while (!done); +} + +/** Increases the number of buckets two-fold. Blocks until done. */ +static void grow_table(cht_t *h) +{ + if (h->b->order >= CHT_MAX_ORDER) + return; + + h->new_b = alloc_buckets(h->b->order + 1, true, false); + + /* Failed to alloc a new table - try next time the resizer is run. */ + if (!h->new_b) + return; + + /* Wait for all readers and updaters to see the initialized new table. */ + rcu_synchronize(); + size_t old_bucket_cnt = (1 << h->b->order); + + /* + * Give updaters a chance to help out with the resize. Do the minimum + * work needed to announce a resize is in progress, ie start moving heads. + */ + for (size_t idx = 0; idx < old_bucket_cnt; ++idx) { + start_head_move(&h->b->head[idx]); + } + + /* Order start_head_move() wrt complete_head_move(). */ + cas_order_barrier(); + + /* Complete moving heads and split any buckets not yet split by updaters. */ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + marked_ptr_t *move_dest_head = &h->new_b->head[grow_idx(old_idx)]; + marked_ptr_t *move_src_head = &h->b->head[old_idx]; + + /* Head move not yet completed. */ + if (N_INVALID != get_mark(*move_src_head)) { + complete_head_move(move_src_head, move_dest_head); + } + + size_t split_idx = grow_to_split_idx(old_idx); + size_t split_hash = calc_split_hash(split_idx, h->new_b->order); + marked_ptr_t *split_dest_head = &h->new_b->head[split_idx]; + + split_bucket(h, move_dest_head, split_dest_head, split_hash); + } + + /* + * Wait for all updaters to notice the new heads. Once everyone sees + * the invalid old bucket heads they will know a resize is in progress + * and updaters will modify the correct new buckets. + */ + rcu_synchronize(); + + /* Clear the JOIN_FOLLOWS mark and remove the link between the split buckets.*/ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + size_t new_idx = grow_idx(old_idx); + + cleanup_join_follows(h, &h->new_b->head[new_idx]); + } + + /* + * Wait for everyone to notice that buckets were split, ie link connecting + * the join follows and join node has been cut. + */ + rcu_synchronize(); + + /* Clear the JOIN mark and GC any deleted join nodes. */ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + size_t new_idx = grow_to_split_idx(old_idx); + + cleanup_join_node(h, &h->new_b->head[new_idx]); + } + + /* Wait for everyone to see that the table is clear of any resize marks. */ + rcu_synchronize(); + + cht_buckets_t *old_b = h->b; + rcu_assign(h->b, h->new_b); + + /* Wait for everyone to start using the new table. */ + rcu_synchronize(); + + free(old_b); + + /* Not needed; just for increased readability. */ + h->new_b = NULL; +} + +/** Halfs the number of buckets. Blocks until done. */ +static void shrink_table(cht_t *h) +{ + if (h->b->order <= h->min_order) + return; + + h->new_b = alloc_buckets(h->b->order - 1, true, false); + + /* Failed to alloc a new table - try next time the resizer is run. */ + if (!h->new_b) + return; + + /* Wait for all readers and updaters to see the initialized new table. */ + rcu_synchronize(); + + size_t old_bucket_cnt = (1 << h->b->order); + + /* + * Give updaters a chance to help out with the resize. Do the minimum + * work needed to announce a resize is in progress, ie start moving heads. + */ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + size_t new_idx = shrink_idx(old_idx); + + /* This bucket should be moved. */ + if (grow_idx(new_idx) == old_idx) { + start_head_move(&h->b->head[old_idx]); + } else { + /* This bucket should join the moved bucket once the move is done.*/ + } + } + + /* Order start_head_move() wrt to complete_head_move(). */ + cas_order_barrier(); + + /* Complete moving heads and join buckets with the moved buckets. */ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + size_t new_idx = shrink_idx(old_idx); + size_t move_src_idx = grow_idx(new_idx); + + /* This bucket should be moved. */ + if (move_src_idx == old_idx) { + /* Head move not yet completed. */ + if (N_INVALID != get_mark(h->b->head[old_idx])) { + complete_head_move(&h->b->head[old_idx], &h->new_b->head[new_idx]); + } + } else { + /* This bucket should join the moved bucket. */ + size_t split_hash = calc_split_hash(old_idx, h->b->order); + join_buckets(h, &h->b->head[old_idx], &h->new_b->head[new_idx], + split_hash); + } + } + + /* + * Wait for all updaters to notice the new heads. Once everyone sees + * the invalid old bucket heads they will know a resize is in progress + * and updaters will modify the correct new buckets. + */ + rcu_synchronize(); + + /* Let everyone know joins are complete and fully visible. */ + for (size_t old_idx = 0; old_idx < old_bucket_cnt; ++old_idx) { + size_t move_src_idx = grow_idx(shrink_idx(old_idx)); + + /* Set the invalid joinee head to NULL. */ + if (old_idx != move_src_idx) { + ASSERT(N_INVALID == get_mark(h->b->head[old_idx])); + + if (&sentinel != get_next(h->b->head[old_idx])) + h->b->head[old_idx] = make_link(&sentinel, N_INVALID); + } + } + + /* todo comment join node vs reset joinee head*/ + rcu_synchronize(); + + size_t new_bucket_cnt = (1 << h->new_b->order); + + /* Clear the JOIN mark and GC any deleted join nodes. */ + for (size_t new_idx = 0; new_idx < new_bucket_cnt; ++new_idx) { + cleanup_join_node(h, &h->new_b->head[new_idx]); + } + + /* Wait for everyone to see that the table is clear of any resize marks. */ + rcu_synchronize(); + + cht_buckets_t *old_b = h->b; + rcu_assign(h->b, h->new_b); + + /* Wait for everyone to start using the new table. */ + rcu_synchronize(); + + free(old_b); + + /* Not needed; just for increased readability. */ + h->new_b = NULL; +} + +/** Finds and clears the N_JOIN mark from a node in new_head (if present). */ +static void cleanup_join_node(cht_t *h, marked_ptr_t *new_head) +{ + rcu_read_lock(); + + cht_link_t *cur = get_next(*new_head); + + while (cur != &sentinel) { + /* Clear the join node's JN mark - even if it is marked as deleted. */ + if (N_JOIN & get_mark(cur->link)) { + clear_join_and_gc(h, cur, new_head); + break; + } + + cur = get_next(cur->link); + } + + rcu_read_unlock(); +} + +/** Clears the join_node's N_JOIN mark frees it if marked N_DELETED as well. */ +static void clear_join_and_gc(cht_t *h, cht_link_t *join_node, + marked_ptr_t *new_head) +{ + ASSERT(join_node != &sentinel); + ASSERT(join_node && (N_JOIN & get_mark(join_node->link))); + + bool done; + + /* Clear the JN mark. */ + do { + marked_ptr_t jn_link = join_node->link; + cht_link_t *next = get_next(jn_link); + /* Clear the JOIN mark but keep the DEL mark if present. */ + mark_t cleared_mark = get_mark(jn_link) & N_DELETED; + + marked_ptr_t ret = + _cas_link(&join_node->link, jn_link, make_link(next, cleared_mark)); + + /* Done if the mark was cleared. Retry if a new node was inserted. */ + done = (ret == jn_link); + ASSERT(ret == jn_link || (get_mark(ret) & N_JOIN)); + } while (!done); + + if (!(N_DELETED & get_mark(join_node->link))) + return; + + /* The join node had been marked as deleted - GC it. */ + + /* Clear the JOIN mark before trying to unlink the deleted join node.*/ + cas_order_barrier(); + + size_t jn_hash = node_hash(h, join_node); + do { + bool resizing = false; + + wnd_t wnd = { + .ppred = new_head, + .cur = get_next(*new_head) + }; + + done = find_wnd_and_gc_pred(h, jn_hash, WM_NORMAL, same_node_pred, + join_node, &wnd, &resizing); + + ASSERT(!resizing); + } while (!done); +} + +/** Finds a non-deleted node with N_JOIN_FOLLOWS and clears the mark. */ +static void cleanup_join_follows(cht_t *h, marked_ptr_t *new_head) +{ + ASSERT(new_head); + + rcu_read_lock(); + + wnd_t wnd = { + .ppred = NULL, + .cur = NULL + }; + marked_ptr_t *cur_link = new_head; + + /* + * Find the non-deleted node with a JF mark and clear the JF mark. + * The JF node may be deleted and/or the mark moved to its neighbors + * at any time. Therefore, we GC deleted nodes until we find the JF + * node in order to remove stale/deleted JF nodes left behind eg by + * delayed threads that did not yet get a chance to unlink the deleted + * JF node and move its mark. + * + * Note that the head may be marked JF (but never DELETED). + */ + while (true) { + bool is_jf_node = N_JOIN_FOLLOWS & get_mark(*cur_link); + + /* GC any deleted nodes on the way - even deleted JOIN_FOLLOWS. */ + if (N_DELETED & get_mark(*cur_link)) { + ASSERT(cur_link != new_head); + ASSERT(wnd.ppred && wnd.cur && wnd.cur != &sentinel); + ASSERT(cur_link == &wnd.cur->link); + + bool dummy; + bool deleted = gc_deleted_node(h, WM_MOVE_JOIN_FOLLOWS, &wnd, &dummy); + + /* Failed to GC or collected a deleted JOIN_FOLLOWS. */ + if (!deleted || is_jf_node) { + /* Retry from the head of the bucket. */ + cur_link = new_head; + continue; + } + } else { + /* Found a non-deleted JF. Clear its JF mark. */ + if (is_jf_node) { + cht_link_t *next = get_next(*cur_link); + marked_ptr_t ret = + cas_link(cur_link, next, N_JOIN_FOLLOWS, &sentinel, N_NORMAL); + + ASSERT(next == &sentinel + || ((N_JOIN | N_JOIN_FOLLOWS) & get_mark(ret))); + + /* Successfully cleared the JF mark of a non-deleted node. */ + if (ret == make_link(next, N_JOIN_FOLLOWS)) { + break; + } else { + /* + * The JF node had been deleted or a new node inserted + * right after it. Retry from the head. + */ + cur_link = new_head; + continue; + } + } else { + wnd.ppred = cur_link; + wnd.cur = get_next(*cur_link); + } + } + + /* We must encounter a JF node before we reach the end of the bucket. */ + ASSERT(wnd.cur && wnd.cur != &sentinel); + cur_link = &wnd.cur->link; + } + + rcu_read_unlock(); +} + +/** Returns the first possible hash following a bucket split point. + * + * In other words the returned hash is the smallest possible hash + * the remainder of the split bucket may contain. + */ +static inline size_t calc_split_hash(size_t split_idx, size_t order) +{ + ASSERT(1 <= order && order <= 8 * sizeof(size_t)); + return split_idx << (8 * sizeof(size_t) - order); +} + +/** Returns the bucket head index given the table size order and item hash. */ +static inline size_t calc_bucket_idx(size_t hash, size_t order) +{ + ASSERT(1 <= order && order <= 8 * sizeof(size_t)); + return hash >> (8 * sizeof(size_t) - order); +} + +/** Returns the bucket index of destination*/ +static inline size_t grow_to_split_idx(size_t old_idx) +{ + return grow_idx(old_idx) | 1; +} + +/** Returns the destination index of a bucket head when the table is growing. */ +static inline size_t grow_idx(size_t idx) +{ + return idx << 1; +} + +/** Returns the destination index of a bucket head when the table is shrinking.*/ +static inline size_t shrink_idx(size_t idx) +{ + return idx >> 1; +} + +/** Returns a mixed hash of the search key.*/ +static inline size_t calc_key_hash(cht_t *h, void *key) +{ + /* Mimic calc_node_hash. */ + return hash_mix(h->op->key_hash(key)) & ~(size_t)1; +} + +/** Returns a memoized mixed hash of the item. */ +static inline size_t node_hash(cht_t *h, const cht_link_t *item) +{ + ASSERT(item->hash == h->invalid_hash + || item->hash == sentinel.hash + || item->hash == calc_node_hash(h, item)); + + return item->hash; +} + +/** Calculates and mixed the hash of the item. */ +static inline size_t calc_node_hash(cht_t *h, const cht_link_t *item) +{ + ASSERT(item != &sentinel); + /* + * Clear the lowest order bit in order for sentinel's node hash + * to be the greatest possible. + */ + return hash_mix(h->op->hash(item)) & ~(size_t)1; +} + +/** Computes and memoizes the hash of the item. */ +static inline void memoize_node_hash(cht_t *h, cht_link_t *item) +{ + item->hash = calc_node_hash(h, item); +} + +/** Packs the next pointer address and the mark into a single pointer. */ +static inline marked_ptr_t make_link(const cht_link_t *next, mark_t mark) +{ + marked_ptr_t ptr = (marked_ptr_t) next; + + ASSERT(!(ptr & N_MARK_MASK)); + ASSERT(!((unsigned)mark & ~N_MARK_MASK)); + + return ptr | mark; +} + +/** Strips any marks from the next item link and returns the next item's address.*/ +static inline cht_link_t * get_next(marked_ptr_t link) +{ + return (cht_link_t*)(link & ~N_MARK_MASK); +} + +/** Returns the current node's mark stored in the next item link. */ +static inline mark_t get_mark(marked_ptr_t link) +{ + return (mark_t)(link & N_MARK_MASK); +} + +/** Moves the window by one item so that is points to the next item. */ +static inline void next_wnd(wnd_t *wnd) +{ + ASSERT(wnd); + ASSERT(wnd->cur); + + wnd->last = wnd->cur; + wnd->ppred = &wnd->cur->link; + wnd->cur = get_next(wnd->cur->link); +} + +/** Predicate that matches only exactly the same node. */ +static bool same_node_pred(void *node, const cht_link_t *item2) +{ + const cht_link_t *item1 = (const cht_link_t*) node; + return item1 == item2; +} + +/** Compare-and-swaps a next item link. */ +static inline marked_ptr_t cas_link(marked_ptr_t *link, const cht_link_t *cur_next, + mark_t cur_mark, const cht_link_t *new_next, mark_t new_mark) +{ + return _cas_link(link, make_link(cur_next, cur_mark), + make_link(new_next, new_mark)); +} + +/** Compare-and-swaps a next item link. */ +static inline marked_ptr_t _cas_link(marked_ptr_t *link, marked_ptr_t cur, + marked_ptr_t new) +{ + ASSERT(link != &sentinel.link); + /* + * cas(x) on the same location x on one cpu must be ordered, but do not + * have to be ordered wrt to other cas(y) to a different location y + * on the same cpu. + * + * cas(x) must act as a write barrier on x, ie if cas(x) succeeds + * and is observed by another cpu, then all cpus must be able to + * make the effects of cas(x) visible just by issuing a load barrier. + * For example: + * cpu1 cpu2 cpu3 + * cas(x, 0 -> 1), succeeds + * cas(x, 0 -> 1), fails + * MB, to order load of x in cas and store to y + * y = 7 + * sees y == 7 + * loadMB must be enough to make cas(x) on cpu3 visible to cpu1, ie x == 1. + * + * If cas() did not work this way: + * a) our head move protocol would not be correct. + * b) freeing an item linked to a moved head after another item was + * inserted in front of it, would require more than one grace period. + * + * Ad (a): In the following example, cpu1 starts moving old_head + * to new_head, cpu2 completes the move and cpu3 notices cpu2 + * completed the move before cpu1 gets a chance to notice cpu2 + * had already completed the move. Our requirements for cas() + * assume cpu3 will see a valid and mutable value in new_head + * after issuing a load memory barrier once it has determined + * the old_head's value had been successfully moved to new_head + * (because it sees old_head marked invalid). + * + * cpu1 cpu2 cpu3 + * cas(old_head, , ), succeeds + * cas-order-barrier + * // Move from old_head to new_head started, now the interesting stuff: + * cas(new_head, <0, Inv>, ), succeeds + * + * cas(new_head, <0, Inv>, ), but fails + * cas-order-barrier + * cas(old_head, , ), succeeds + * + * Sees old_head marked Inv (by cpu2) + * load-MB + * assert(new_head == ) + * + * cas-order-barrier + * + * Even though cpu1 did not yet issue a cas-order-barrier, cpu1's store + * to new_head (successful cas()) must be made visible to cpu3 with + * a load memory barrier if cpu1's store to new_head is visible + * on another cpu (cpu2) and that cpu's (cpu2's) store to old_head + * is already visible to cpu3. * + */ + void *expected = (void*)cur; + + /* + * Use the acquire-release model, although we could probably + * get away even with the relaxed memory model due to our use + * of explicit memory barriers. + */ + __atomic_compare_exchange_n((void**)link, &expected, (void *)new, false, + __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE); + + return (marked_ptr_t) expected; +} + +/** Orders compare-and-swaps to different memory locations. */ +static inline void cas_order_barrier(void) +{ + /* Make sure CAS to different memory locations are ordered. */ + write_barrier(); +} + + +/** @} + */ Index: kernel/generic/src/adt/list.c =================================================================== --- kernel/generic/src/adt/list.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/adt/list.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -68,24 +68,26 @@ } -/** Concatenate two lists - * - * Concatenate lists @a list1 and @a list2, producing a single - * list @a list1 containing items from both (in @a list1, @a list2 - * order) and empty list @a list2. - * - * @param list1 First list and concatenated output - * @param list2 Second list and empty output. - * +/** Moves items of one list into another after the specified item. + * + * Inserts all items of @a list after item at @a pos in another list. + * Both lists may be empty. + * + * @param list Source list to move after pos. Empty afterwards. + * @param pos Source items will be placed after this item. */ -void list_concat(list_t *list1, list_t *list2) +void list_splice(list_t *list, link_t *pos) { - if (list_empty(list2)) + if (list_empty(list)) return; - - list2->head.next->prev = list1->head.prev; - list2->head.prev->next = &list1->head; - list1->head.prev->next = list2->head.next; - list1->head.prev = list2->head.prev; - list_initialize(list2); + + /* Attach list to destination. */ + list->head.next->prev = pos; + list->head.prev->next = pos->next; + + /* Link destination list to the added list. */ + pos->next->prev = list->head.prev; + pos->next = list->head.next; + + list_initialize(list); } Index: kernel/generic/src/console/chardev.c =================================================================== --- kernel/generic/src/console/chardev.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/console/chardev.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -39,5 +39,5 @@ #include #include -#include +#include /** Initialize input character device. Index: kernel/generic/src/console/cmd.c =================================================================== --- kernel/generic/src/console/cmd.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/console/cmd.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -69,4 +69,6 @@ #include #include +#include +#include #include @@ -525,4 +527,22 @@ }; +/* Data and methods for the 'workq' command */ +static int cmd_workq(cmd_arg_t *argv); +static cmd_info_t workq_info = { + .name = "workq", + .description = "Show global workq information.", + .func = cmd_workq, + .argc = 0 +}; + +/* Data and methods for the 'workq' command */ +static int cmd_rcu(cmd_arg_t *argv); +static cmd_info_t rcu_info = { + .name = "rcu", + .description = "Show RCU run-time statistics.", + .func = cmd_rcu, + .argc = 0 +}; + /* Data and methods for 'ipc' command */ static int cmd_ipc(cmd_arg_t *argv); @@ -588,4 +608,5 @@ &physmem_info, &reboot_info, + &rcu_info, &sched_info, &set4_info, @@ -598,4 +619,5 @@ &uptime_info, &version_info, + &workq_info, &zones_info, &zone_info, @@ -1280,4 +1302,28 @@ } +/** Prints information about the global work queue. + * + * @param argv Ignores + * + * @return Always 1 + */ +int cmd_workq(cmd_arg_t *argv) +{ + workq_global_print_info(); + return 1; +} + +/** Prints RCU statistics. + * + * @param argv Ignores + * + * @return Always 1 + */ +int cmd_rcu(cmd_arg_t *argv) +{ + rcu_print_stat(); + return 1; +} + /** Command for listing memory zones * Index: kernel/generic/src/console/console.c =================================================================== --- kernel/generic/src/console/console.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/console/console.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -52,4 +52,6 @@ #include #include +#include /* SIZE2FRAMES */ +#include /* malloc */ #define KLOG_PAGES 8 Index: kernel/generic/src/console/kconsole.c =================================================================== --- kernel/generic/src/console/kconsole.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/console/kconsole.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -60,4 +60,5 @@ #include #include +#include /** Simple kernel console. Index: kernel/generic/src/cpu/cpu.c =================================================================== --- kernel/generic/src/cpu/cpu.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/cpu/cpu.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -50,4 +50,5 @@ #include #include +#include cpu_t *cpus; @@ -102,4 +103,5 @@ cpu_identify(); cpu_arch_init(); + rcu_cpu_init(); } Index: kernel/generic/src/cpu/cpu_mask.c =================================================================== --- kernel/generic/src/cpu/cpu_mask.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/cpu/cpu_mask.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,139 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ + +/** + * @file + * @brief CPU mask manipulation functions. + */ +#include +#include +#include + +static const size_t word_size = sizeof(unsigned int); +static const size_t word_bit_cnt = 8 * sizeof(unsigned int); + +/** Returns the size of cpu_mask_t for the detected number of cpus in bytes. */ +size_t cpu_mask_size(void) +{ + size_t word_cnt = (config.cpu_count + word_bit_cnt - 1) / word_bit_cnt; + return word_cnt * word_size; +} + +/** Add first cpu_cnt cpus to the mask, ie sets the first cpu_cnt bits. */ +static void cpu_mask_count(cpu_mask_t *cpus, size_t cpu_cnt) +{ + ASSERT(NULL != cpus); + ASSERT(cpu_cnt <= config.cpu_count); + + for (size_t active_word = 0; + (active_word + 1) * word_bit_cnt <= cpu_cnt; + ++active_word) { + /* Set all bits in the cell/word. */ + cpus->mask[active_word] = -1; + } + + size_t remaining_bits = (cpu_cnt % word_bit_cnt); + if (0 < remaining_bits) { + /* Set lower remaining_bits of the last word. */ + cpus->mask[cpu_cnt / word_bit_cnt] = (1 << remaining_bits) - 1; + } +} + +/** Sets bits corresponding to the active cpus, ie the first + * config.cpu_active cpus. + */ +void cpu_mask_active(cpu_mask_t *cpus) +{ + cpu_mask_none(cpus); + cpu_mask_count(cpus, config.cpu_active); +} + +/** Sets bits for all cpus of the mask. */ +void cpu_mask_all(cpu_mask_t *cpus) +{ + cpu_mask_count(cpus, config.cpu_count); +} + +/** Resets/removes all bits. */ +void cpu_mask_none(cpu_mask_t *cpus) +{ + ASSERT(cpus); + + size_t word_cnt = cpu_mask_size() / word_size; + + for (size_t word = 0; word < word_cnt; ++word) { + cpus->mask[word] = 0; + } +} + +/** Sets the bit corresponding to cpu_id to true. */ +void cpu_mask_set(cpu_mask_t *cpus, unsigned int cpu_id) +{ + size_t word = cpu_id / word_bit_cnt; + size_t word_pos = cpu_id % word_bit_cnt; + + cpus->mask[word] |= (1U << word_pos); +} + +/** Resets the bit corresponding to cpu_id to false. */ +void cpu_mask_reset(cpu_mask_t *cpus, unsigned int cpu_id) +{ + size_t word = cpu_id / word_bit_cnt; + size_t word_pos = cpu_id % word_bit_cnt; + + cpus->mask[word] &= ~(1U << word_pos); +} + +/** Returns true if the bit corresponding to cpu_id is set. */ +bool cpu_mask_is_set(cpu_mask_t *cpus, unsigned int cpu_id) +{ + size_t word = cpu_id / word_bit_cnt; + size_t word_pos = cpu_id % word_bit_cnt; + + return 0 != (cpus->mask[word] & (1U << word_pos)); +} + +/** Returns true if no bits are set. */ +bool cpu_mask_is_none(cpu_mask_t *cpus) +{ + size_t word_cnt = cpu_mask_size() / word_size; + + for (size_t word = 0; word < word_cnt; ++word) { + if (cpus->mask[word]) + return false; + } + + return true; +} + +/** @} + */ Index: kernel/generic/src/debug/panic.c =================================================================== --- kernel/generic/src/debug/panic.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/debug/panic.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -97,5 +97,5 @@ if (THE != NULL) { printf("pe=%" PRIun " thr=%p task=%p cpu=%p as=%p" - " magic=%#" PRIx32 "\n", THE->preemption_disabled, + " magic=%#" PRIx32 "\n", THE->preemption, THE->thread, THE->task, THE->cpu, THE->as, THE->magic); } else Index: kernel/generic/src/interrupt/interrupt.c =================================================================== --- kernel/generic/src/interrupt/interrupt.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/interrupt/interrupt.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -111,13 +111,11 @@ } - /* Account CPU usage if it has waked up from sleep */ - if (CPU) { + /* Account CPU usage if it woke up from sleep */ + if (CPU && CPU->idle) { irq_spinlock_lock(&CPU->lock, false); - if (CPU->idle) { - uint64_t now = get_cycle(); - CPU->idle_cycles += now - CPU->last_cycle; - CPU->last_cycle = now; - CPU->idle = false; - } + uint64_t now = get_cycle(); + CPU->idle_cycles += now - CPU->last_cycle; + CPU->last_cycle = now; + CPU->idle = false; irq_spinlock_unlock(&CPU->lock, false); } Index: kernel/generic/src/ipc/kbox.c =================================================================== --- kernel/generic/src/ipc/kbox.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/ipc/kbox.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -44,4 +44,5 @@ #include #include +#include void ipc_kbox_cleanup(void) Index: kernel/generic/src/lib/str.c =================================================================== --- kernel/generic/src/lib/str.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/lib/str.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -111,4 +111,5 @@ #include #include +#include /** Check the condition if wchar_t is signed */ @@ -567,4 +568,5 @@ /* There must be space for a null terminator in the buffer. */ ASSERT(size > 0); + ASSERT(src != NULL); size_t src_off = 0; Index: kernel/generic/src/main/kinit.c =================================================================== --- kernel/generic/src/main/kinit.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/main/kinit.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -78,4 +78,6 @@ #include #include +#include +#include #define ALIVE_CHARS 4 @@ -104,6 +106,14 @@ */ thread_detach(THREAD); - + interrupts_disable(); + + /* Start processing RCU callbacks. RCU is fully functional afterwards. */ + rcu_kinit_init(); + + /* + * Start processing work queue items. Some may have been queued during boot. + */ + workq_global_worker_init(); #ifdef CONFIG_SMP Index: kernel/generic/src/main/main.c =================================================================== --- kernel/generic/src/main/main.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/main/main.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -75,4 +75,6 @@ #include #include +#include +#include #include #include @@ -244,6 +246,9 @@ cpu_init(); - calibrate_delay_loop(); + arch_post_cpu_init(); + + smp_call_init(); + workq_global_init(); clock_counter_init(); timeout_init(); @@ -347,4 +352,6 @@ void main_ap_separated_stack(void) { + smp_call_init(); + /* * Configure timeouts for this cpu. Index: kernel/generic/src/main/shutdown.c =================================================================== --- kernel/generic/src/main/shutdown.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/main/shutdown.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -37,4 +37,5 @@ #include +#include #include #include Index: kernel/generic/src/mm/frame.c =================================================================== --- kernel/generic/src/mm/frame.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/mm/frame.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -61,4 +61,5 @@ #include #include +#include /* THREAD */ zones_t zones; Index: kernel/generic/src/mm/km.c =================================================================== --- kernel/generic/src/mm/km.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/mm/km.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -49,4 +49,5 @@ #include #include +#include static ra_arena_t *km_ni_arena; Index: kernel/generic/src/mm/slab.c =================================================================== --- kernel/generic/src/mm/slab.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/mm/slab.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -114,4 +114,5 @@ #include #include +#include IRQ_SPINLOCK_STATIC_INITIALIZE(slab_cache_lock); Index: kernel/generic/src/preempt/preemption.c =================================================================== --- kernel/generic/src/preempt/preemption.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/preempt/preemption.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -37,23 +37,5 @@ #include -#include -#include -#include -#include -/** Increment preemption disabled counter. */ -void preemption_disable(void) -{ - THE->preemption_disabled++; - memory_barrier(); -} - -/** Decrement preemption disabled counter. */ -void preemption_enable(void) -{ - ASSERT(PREEMPTION_DISABLED); - memory_barrier(); - THE->preemption_disabled--; -} /** @} Index: kernel/generic/src/proc/scheduler.c =================================================================== --- kernel/generic/src/proc/scheduler.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/proc/scheduler.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -52,4 +52,6 @@ #include #include +#include +#include #include #include @@ -63,4 +65,5 @@ #include #include +#include static void scheduler_separated_stack(void); @@ -86,4 +89,5 @@ { before_thread_runs_arch(); + rcu_before_thread_runs(); #ifdef CONFIG_FPU_LAZY @@ -126,4 +130,6 @@ static void after_thread_ran(void) { + workq_after_thread_ran(); + rcu_after_thread_ran(); after_thread_ran_arch(); } @@ -218,4 +224,6 @@ goto loop; } + + ASSERT(!CPU->idle); unsigned int i; @@ -397,4 +405,5 @@ ASSERT((!THREAD) || (irq_spinlock_locked(&THREAD->lock))); ASSERT(CPU != NULL); + ASSERT(interrupts_disabled()); /* @@ -420,4 +429,5 @@ case Exiting: + rcu_thread_exiting(); repeat: if (THREAD->detached) { Index: kernel/generic/src/proc/task.c =================================================================== --- kernel/generic/src/proc/task.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/proc/task.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,5 @@ #include #include +#include #include #include @@ -153,5 +154,4 @@ irq_spinlock_initialize(&task->lock, "task_t_lock"); - mutex_initialize(&task->futexes_lock, MUTEX_PASSIVE); list_initialize(&task->threads); @@ -165,5 +165,5 @@ spinlock_initialize(&task->active_calls_lock, "active_calls_lock"); list_initialize(&task->active_calls); - + #ifdef CONFIG_UDEBUG /* Init kbox stuff */ @@ -221,5 +221,5 @@ (void) ipc_phone_connect(&task->phones[0], ipc_phone_0); - btree_create(&task->futexes); + futex_task_init(task); /* @@ -262,5 +262,5 @@ * Free up dynamically allocated state. */ - btree_destroy(&task->futexes); + futex_task_deinit(task); /* Index: kernel/generic/src/proc/the.c =================================================================== --- kernel/generic/src/proc/the.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/proc/the.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -43,4 +43,5 @@ #include +#include /** Initialize THE structure @@ -53,5 +54,5 @@ void the_initialize(the_t *the) { - the->preemption_disabled = 0; + the->preemption = 0; the->cpu = NULL; the->thread = NULL; @@ -59,4 +60,7 @@ the->as = NULL; the->magic = MAGIC; +#ifdef RCU_PREEMPT_A + the->rcu_nesting = 0; +#endif } Index: kernel/generic/src/proc/thread.c =================================================================== --- kernel/generic/src/proc/thread.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/proc/thread.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -46,4 +46,6 @@ #include #include +#include +#include #include #include @@ -260,4 +262,11 @@ } +/** Invoked right before thread_ready() readies the thread. thread is locked. */ +static void before_thread_is_ready(thread_t *thread) +{ + ASSERT(irq_spinlock_locked(&thread->lock)); + workq_before_thread_is_ready(thread); +} + /** Make thread ready * @@ -272,13 +281,20 @@ ASSERT(thread->state != Ready); + + before_thread_is_ready(thread); int i = (thread->priority < RQ_COUNT - 1) ? ++thread->priority : thread->priority; - - cpu_t *cpu; - if (thread->wired || thread->nomigrate || thread->fpu_context_engaged) { - ASSERT(thread->cpu != NULL); - cpu = thread->cpu; - } else + + /* Check that thread->cpu is set whenever it needs to be. */ + ASSERT(thread->cpu != NULL || + (!thread->wired && !thread->nomigrate && !thread->fpu_context_engaged)); + + /* + * Prefer to run on the same cpu as the last time. Used by wired + * threads as well as threads with disabled migration. + */ + cpu_t *cpu = thread->cpu; + if (cpu == NULL) cpu = CPU; @@ -374,4 +390,6 @@ thread->task = task; + thread->workq = NULL; + thread->fpu_context_exists = false; thread->fpu_context_engaged = false; @@ -388,4 +406,6 @@ /* Might depend on previous initialization */ thread_create_arch(thread); + + rcu_thread_init(thread); if ((flags & THREAD_FLAG_NOATTACH) != THREAD_FLAG_NOATTACH) @@ -498,5 +518,5 @@ */ ipc_cleanup(); - futex_cleanup(); + futex_task_cleanup(); LOG("Cleanup of task %" PRIu64" completed.", TASK->taskid); } @@ -518,4 +538,52 @@ /* Not reached */ while (true); +} + +/** Interrupts an existing thread so that it may exit as soon as possible. + * + * Threads that are blocked waiting for a synchronization primitive + * are woken up with a return code of ESYNCH_INTERRUPTED if the + * blocking call was interruptable. See waitq_sleep_timeout(). + * + * The caller must guarantee the thread object is valid during the entire + * function, eg by holding the threads_lock lock. + * + * Interrupted threads automatically exit when returning back to user space. + * + * @param thread A valid thread object. The caller must guarantee it + * will remain valid until thread_interrupt() exits. + */ +void thread_interrupt(thread_t *thread) +{ + ASSERT(thread != NULL); + + irq_spinlock_lock(&thread->lock, true); + + thread->interrupted = true; + bool sleeping = (thread->state == Sleeping); + + irq_spinlock_unlock(&thread->lock, true); + + if (sleeping) + waitq_interrupt_sleep(thread); +} + +/** Returns true if the thread was interrupted. + * + * @param thread A valid thread object. User must guarantee it will + * be alive during the entire call. + * @return true if the thread was already interrupted via thread_interrupt(). + */ +bool thread_interrupted(thread_t *thread) +{ + ASSERT(thread != NULL); + + bool interrupted; + + irq_spinlock_lock(&thread->lock, true); + interrupted = thread->interrupted; + irq_spinlock_unlock(&thread->lock, true); + + return interrupted; } Index: kernel/generic/src/smp/smp_call.c =================================================================== --- kernel/generic/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/smp/smp_call.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,278 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ + +/** + * @file + * @brief Facility to invoke functions on other cpus via IPIs. + */ + +#include +#include +#include /* interrupt_disable */ +#include +#include +#include +#include +#include + +static void call_start(smp_call_t *call_info, smp_call_func_t func, void *arg); +static void call_done(smp_call_t *call_info); +static void call_wait(smp_call_t *call_info); + + +/** Init smp_call() on the local cpu. */ +void smp_call_init(void) +{ + ASSERT(CPU); + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + + spinlock_initialize(&CPU->smp_calls_lock, "cpu[].smp_calls_lock"); + list_initialize(&CPU->smp_pending_calls); +} + +/** Invokes a function on a specific cpu and waits for it to complete. + * + * Calls @a func on the CPU denoted by its logical id @cpu_id . + * The function will execute with interrupts disabled. It should + * be a quick and simple function and must never block. + * + * If @a cpu_id is the local CPU, the function will be invoked + * directly. + * + * All memory accesses of prior to smp_call() will be visible + * to @a func on cpu @a cpu_id. Similarly, any changes @a func + * makes on cpu @a cpu_id will be visible on this cpu once + * smp_call() returns. + * + * Invoking @a func on the destination cpu acts as a memory barrier + * on that cpu. + * + * @param cpu_id Destination CPU's logical id (eg CPU->id) + * @param func Function to call. + * @param arg Argument to pass to the user supplied function @a func. + */ +void smp_call(unsigned int cpu_id, smp_call_func_t func, void *arg) +{ + smp_call_t call_info; + smp_call_async(cpu_id, func, arg, &call_info); + smp_call_wait(&call_info); +} + +/** Invokes a function on a specific cpu asynchronously. + * + * Calls @a func on the CPU denoted by its logical id @cpu_id . + * The function will execute with interrupts disabled. It should + * be a quick and simple function and must never block. + * + * Pass @a call_info to smp_call_wait() in order to wait for + * @a func to complete. + * + * @a call_info must be valid until/after @a func returns. Use + * smp_call_wait() to wait until it is safe to free @a call_info. + * + * If @a cpu_id is the local CPU, the function will be invoked + * directly. If the destination cpu id @a cpu_id is invalid + * or denotes an inactive cpu, the call is discarded immediately. + * + * All memory accesses of the caller prior to smp_call_async() + * will be made visible to @a func on the other cpu. Similarly, + * any changes @a func makes on cpu @a cpu_id will be visible + * to this cpu when smp_call_wait() returns. + * + * Invoking @a func on the destination cpu acts as a memory barrier + * on that cpu. + * + * Interrupts must be enabled. Otherwise you run the risk + * of a deadlock. + * + * @param cpu_id Destination CPU's logical id (eg CPU->id). + * @param func Function to call. + * @param arg Argument to pass to the user supplied function @a func. + * @param call_info Use it to wait for the function to complete. Must + * be valid until the function completes. + */ +void smp_call_async(unsigned int cpu_id, smp_call_func_t func, void *arg, + smp_call_t *call_info) +{ + /* + * Interrupts must not be disabled or you run the risk of a deadlock + * if both the destination and source cpus try to send an IPI to each + * other with interrupts disabled. Because the interrupts are disabled + * the IPIs cannot be delivered and both cpus will forever busy wait + * for an acknowledgment of the IPI from the other cpu. + */ + ASSERT(!interrupts_disabled()); + ASSERT(call_info != NULL); + + /* Discard invalid calls. */ + if (config.cpu_count <= cpu_id || !cpus[cpu_id].active) { + call_start(call_info, func, arg); + call_done(call_info); + return; + } + + /* Protect cpu->id against migration. */ + preemption_disable(); + + call_start(call_info, func, arg); + + if (cpu_id != CPU->id) { +#ifdef CONFIG_SMP + spinlock_lock(&cpus[cpu_id].smp_calls_lock); + list_append(&call_info->calls_link, &cpus[cpu_id].smp_pending_calls); + spinlock_unlock(&cpus[cpu_id].smp_calls_lock); + + /* + * If a platform supports SMP it must implement arch_smp_call_ipi(). + * It should issue an IPI on cpu_id and invoke smp_call_ipi_recv() + * on cpu_id in turn. + * + * Do not implement as just an empty dummy function. Instead + * consider providing a full implementation or at least a version + * that panics if invoked. Note that smp_call_async() never + * calls arch_smp_call_ipi() on uniprocessors even if CONFIG_SMP. + */ + arch_smp_call_ipi(cpu_id); +#endif + } else { + /* Invoke local smp calls in place. */ + ipl_t ipl = interrupts_disable(); + func(arg); + interrupts_restore(ipl); + + call_done(call_info); + } + + preemption_enable(); +} + +/** Waits for a function invoked on another CPU asynchronously to complete. + * + * Does not sleep but rather spins. + * + * Example usage: + * @code + * void hello(void *p) { + * puts((char*)p); + * } + * + * smp_call_t call_info; + * smp_call_async(cpus[2].id, hello, "hi!\n", &call_info); + * // Do some work. In the meantime, hello() is executed on cpu2. + * smp_call_wait(&call_info); + * @endcode + * + * @param call_info Initialized by smp_call_async(). + */ +void smp_call_wait(smp_call_t *call_info) +{ + call_wait(call_info); +} + +#ifdef CONFIG_SMP + +/** Architecture independent smp call IPI handler. + * + * Interrupts must be disabled. Tolerates spurious calls. + */ +void smp_call_ipi_recv(void) +{ + ASSERT(interrupts_disabled()); + ASSERT(CPU); + + list_t calls_list; + list_initialize(&calls_list); + + /* + * Acts as a load memory barrier. Any changes made by the cpu that + * added the smp_call to calls_list will be made visible to this cpu. + */ + spinlock_lock(&CPU->smp_calls_lock); + list_concat(&calls_list, &CPU->smp_pending_calls); + spinlock_unlock(&CPU->smp_calls_lock); + + /* Walk the list manually, so that we can safely remove list items. */ + for (link_t *cur = calls_list.head.next, *next = cur->next; + !list_empty(&calls_list); cur = next, next = cur->next) { + + smp_call_t *call_info = list_get_instance(cur, smp_call_t, calls_link); + list_remove(cur); + + call_info->func(call_info->arg); + call_done(call_info); + } +} + +#endif /* CONFIG_SMP */ + +static void call_start(smp_call_t *call_info, smp_call_func_t func, void *arg) +{ + link_initialize(&call_info->calls_link); + call_info->func = func; + call_info->arg = arg; + + /* + * We can't use standard spinlocks here because we want to lock + * the structure on one cpu and unlock it on another (without + * messing up the preemption count). + */ + atomic_set(&call_info->pending, 1); + + /* Let initialization complete before continuing. */ + memory_barrier(); +} + +static void call_done(smp_call_t *call_info) +{ + /* + * Separate memory accesses of the called function from the + * announcement of its completion. + */ + memory_barrier(); + atomic_set(&call_info->pending, 0); +} + +static void call_wait(smp_call_t *call_info) +{ + do { + /* + * Ensure memory accesses following call_wait() are ordered + * after completion of the called function on another cpu. + * Also, speed up loading of call_info->pending. + */ + memory_barrier(); + } while (atomic_get(&call_info->pending)); +} + + +/** @} + */ Index: kernel/generic/src/synch/condvar.c =================================================================== --- kernel/generic/src/synch/condvar.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/condvar.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -38,4 +38,5 @@ #include #include +#include #include #include @@ -90,4 +91,5 @@ ipl = waitq_sleep_prepare(&cv->wq); + /* Unlock only after the waitq is locked so we don't miss a wakeup. */ mutex_unlock(mtx); @@ -95,10 +97,93 @@ rc = waitq_sleep_timeout_unsafe(&cv->wq, usec, flags); + waitq_sleep_finish(&cv->wq, rc, ipl); + /* Lock only after releasing the waitq to avoid a possible deadlock. */ mutex_lock(mtx); - waitq_sleep_finish(&cv->wq, rc, ipl); return rc; } +/** Wait for the condition to become true with a locked spinlock. + * + * The function is not aware of irq_spinlock. Therefore do not even + * try passing irq_spinlock_t to it. Use _condvar_wait_timeout_irq_spinlock() + * instead. + * + * @param cv Condition variable. + * @param lock Locked spinlock. + * @param usec Timeout value in microseconds. + * @param flags Select mode of operation. + * + * For exact description of meaning of possible combinations of usec and flags, + * see comment for waitq_sleep_timeout(). Note that when + * SYNCH_FLAGS_NON_BLOCKING is specified here, ESYNCH_WOULD_BLOCK is always + * returned. + * + * @return See comment for waitq_sleep_timeout(). + */ +int _condvar_wait_timeout_spinlock_impl(condvar_t *cv, spinlock_t *lock, + uint32_t usec, int flags) +{ + int rc; + ipl_t ipl; + + ipl = waitq_sleep_prepare(&cv->wq); + + /* Unlock only after the waitq is locked so we don't miss a wakeup. */ + spinlock_unlock(lock); + + cv->wq.missed_wakeups = 0; /* Enforce blocking. */ + rc = waitq_sleep_timeout_unsafe(&cv->wq, usec, flags); + + waitq_sleep_finish(&cv->wq, rc, ipl); + /* Lock only after releasing the waitq to avoid a possible deadlock. */ + spinlock_lock(lock); + + return rc; +} + +/** Wait for the condition to become true with a locked irq spinlock. + * + * @param cv Condition variable. + * @param lock Locked irq spinlock. + * @param usec Timeout value in microseconds. + * @param flags Select mode of operation. + * + * For exact description of meaning of possible combinations of usec and flags, + * see comment for waitq_sleep_timeout(). Note that when + * SYNCH_FLAGS_NON_BLOCKING is specified here, ESYNCH_WOULD_BLOCK is always + * returned. + * + * @return See comment for waitq_sleep_timeout(). + */ +int _condvar_wait_timeout_irq_spinlock(condvar_t *cv, irq_spinlock_t *irq_lock, + uint32_t usec, int flags) +{ + int rc; + /* Save spinlock's state so we can restore it correctly later on. */ + ipl_t ipl = irq_lock->ipl; + bool guard = irq_lock->guard; + + irq_lock->guard = false; + + /* + * waitq_prepare() restores interrupts to the current state, + * ie disabled. Therefore, interrupts will remain disabled while + * it spins waiting for a pending timeout handler to complete. + * Although it spins with interrupts disabled there can only + * be a pending timeout if we failed to cancel an imminent + * timeout (on another cpu) during a wakeup. As a result the + * timeout handler is guaranteed to run (it is most likely already + * running) and there is no danger of a deadlock. + */ + rc = _condvar_wait_timeout_spinlock(cv, &irq_lock->lock, usec, flags); + + irq_lock->guard = guard; + irq_lock->ipl = ipl; + + return rc; +} + + /** @} */ Index: kernel/generic/src/synch/futex.c =================================================================== --- kernel/generic/src/synch/futex.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/futex.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -1,4 +1,5 @@ /* * Copyright (c) 2006 Jakub Jermar + * Copyright (c) 2012 Adam Hraska * All rights reserved. * @@ -34,4 +35,28 @@ * @file * @brief Kernel backend for futexes. + * + * Kernel futex objects are stored in a global hash table futex_ht + * where the physical address of the futex variable (futex_t.paddr) + * is used as the lookup key. As a result multiple address spaces + * may share the same futex variable. + * + * A kernel futex object is created the first time a task accesses + * the futex (having a futex variable at a physical address not + * encountered before). Futex object's lifetime is governed by + * a reference count that represents the number of all the different + * user space virtual addresses from all tasks that map to the + * physical address of the futex variable. A futex object is freed + * when the last task having accessed the futex exits. + * + * Each task keeps track of the futex objects it accessed in a list + * of pointers (futex_ptr_t, task->futex_list) to the different futex + * objects. + * + * To speed up translation of futex variables' virtual addresses + * to their physical addresses, futex pointers accessed by the + * task are furthermore stored in a concurrent hash table (CHT, + * task->futexes->ht). A single lookup without locks or accesses + * to the page table translates a futex variable's virtual address + * into its futex kernel object. */ @@ -39,4 +64,5 @@ #include #include +#include #include #include @@ -46,4 +72,5 @@ #include #include +#include #include #include @@ -52,26 +79,55 @@ #include #include -#include #define FUTEX_HT_SIZE 1024 /* keep it a power of 2 */ -static void futex_initialize(futex_t *futex); - -static futex_t *futex_find(uintptr_t paddr); +/** Task specific pointer to a global kernel futex object. */ +typedef struct futex_ptr { + /** CHT link. */ + cht_link_t cht_link; + /** List of all futex pointers used by the task. */ + link_t all_link; + /** Kernel futex object. */ + futex_t *futex; + /** User space virtual address of the futex variable in the task. */ + uintptr_t uaddr; +} futex_ptr_t; + + +static void destroy_task_cache(work_t *work); + +static void futex_initialize(futex_t *futex, uintptr_t paddr); +static void futex_add_ref(futex_t *futex); +static void futex_release_ref(futex_t *futex); +static void futex_release_ref_locked(futex_t *futex); + +static futex_t *get_futex(uintptr_t uaddr); +static futex_t *find_cached_futex(uintptr_t uaddr); +static futex_t *get_and_cache_futex(uintptr_t phys_addr, uintptr_t uaddr); +static bool find_futex_paddr(uintptr_t uaddr, uintptr_t *phys_addr); + static size_t futex_ht_hash(sysarg_t *key); static bool futex_ht_compare(sysarg_t *key, size_t keys, link_t *item); static void futex_ht_remove_callback(link_t *item); -/** - * Mutex protecting global futex hash table. - * It is also used to serialize access to all futex_t structures. - * Must be acquired before the task futex B+tree lock. +static size_t task_fut_ht_hash(const cht_link_t *link); +static size_t task_fut_ht_key_hash(void *key); +static bool task_fut_ht_equal(const cht_link_t *item1, const cht_link_t *item2); +static bool task_fut_ht_key_equal(void *key, const cht_link_t *item); + + +/** Mutex protecting the global futex hash table. + * + * Acquire task specific TASK->futex_list_lock before this mutex. */ static mutex_t futex_ht_lock; -/** Futex hash table. */ +/** Global kernel futex hash table. Lock futex_ht_lock before accessing. + * + * Physical address of the futex variable is the lookup key. + */ static hash_table_t futex_ht; -/** Futex hash table operations. */ +/** Global kernel futex hash table operations. */ static hash_table_operations_t futex_ht_ops = { .hash = futex_ht_hash, @@ -80,4 +136,13 @@ }; +/** Task futex cache CHT operations. */ +static cht_ops_t task_futex_ht_ops = { + .hash = task_fut_ht_hash, + .key_hash = task_fut_ht_key_hash, + .equal = task_fut_ht_equal, + .key_equal = task_fut_ht_key_equal, + .remove_callback = NULL +}; + /** Initialize futex subsystem. */ void futex_init(void) @@ -87,14 +152,233 @@ } -/** Initialize kernel futex structure. - * - * @param futex Kernel futex structure. - */ -void futex_initialize(futex_t *futex) +/** Initializes the futex structures for the new task. */ +void futex_task_init(struct task *task) +{ + task->futexes = malloc(sizeof(struct futex_cache), 0); + + cht_create(&task->futexes->ht, 0, 0, 0, true, &task_futex_ht_ops); + + list_initialize(&task->futexes->list); + mutex_initialize(&task->futexes->list_lock, MUTEX_PASSIVE); +} + +/** Destroys the futex structures for the dying task. */ +void futex_task_deinit(task_t *task) +{ + /* Interrupts are disabled so we must not block (cannot run cht_destroy). */ + if (interrupts_disabled()) { + /* Invoke the blocking cht_destroy in the background. */ + workq_global_enqueue_noblock(&task->futexes->destroy_work, + destroy_task_cache); + } else { + /* We can block. Invoke cht_destroy in this thread. */ + destroy_task_cache(&task->futexes->destroy_work); + } +} + +/** Deallocates a task's CHT futex cache (must already be empty). */ +static void destroy_task_cache(work_t *work) +{ + struct futex_cache *cache = + member_to_inst(work, struct futex_cache, destroy_work); + + /* + * Destroy the cache before manually freeing items of the cache in case + * table resize is in progress. + */ + cht_destroy_unsafe(&cache->ht); + + /* Manually free futex_ptr cache items. */ + list_foreach_safe(cache->list, cur_link, next_link) { + futex_ptr_t *fut_ptr = member_to_inst(cur_link, futex_ptr_t, all_link); + + list_remove(cur_link); + free(fut_ptr); + } + + free(cache); +} + +/** Remove references from futexes known to the current task. */ +void futex_task_cleanup(void) +{ + struct futex_cache *futexes = TASK->futexes; + + /* All threads of this task have terminated. This is the last thread. */ + mutex_lock(&futexes->list_lock); + + list_foreach_safe(futexes->list, cur_link, next_link) { + futex_ptr_t *fut_ptr = member_to_inst(cur_link, futex_ptr_t, all_link); + + /* + * The function is free to free the futex. All other threads of this + * task have already terminated, so they have also definitely + * exited their CHT futex cache protecting rcu reader sections. + * Moreover release_ref() only frees the futex if this is the + * last task referencing the futex. Therefore, only threads + * of this task may have referenced the futex if it is to be freed. + */ + futex_release_ref_locked(fut_ptr->futex); + } + + mutex_unlock(&futexes->list_lock); +} + + +/** Initialize the kernel futex structure. + * + * @param futex Kernel futex structure. + * @param paddr Physical address of the futex variable. + */ +static void futex_initialize(futex_t *futex, uintptr_t paddr) { waitq_initialize(&futex->wq); link_initialize(&futex->ht_link); - futex->paddr = 0; + futex->paddr = paddr; futex->refcount = 1; +} + +/** Increments the counter of tasks referencing the futex. */ +static void futex_add_ref(futex_t *futex) +{ + ASSERT(mutex_locked(&futex_ht_lock)); + ASSERT(0 < futex->refcount); + ++futex->refcount; +} + +/** Decrements the counter of tasks referencing the futex. May free the futex.*/ +static void futex_release_ref(futex_t *futex) +{ + ASSERT(mutex_locked(&futex_ht_lock)); + ASSERT(0 < futex->refcount); + + --futex->refcount; + + if (0 == futex->refcount) { + hash_table_remove(&futex_ht, &futex->paddr, 1); + } +} + +/** Decrements the counter of tasks referencing the futex. May free the futex.*/ +static void futex_release_ref_locked(futex_t *futex) +{ + mutex_lock(&futex_ht_lock); + futex_release_ref(futex); + mutex_unlock(&futex_ht_lock); +} + +/** Returns a futex for the virtual address @a uaddr (or creates one). */ +static futex_t *get_futex(uintptr_t uaddr) +{ + futex_t *futex = find_cached_futex(uaddr); + + if (futex) + return futex; + + uintptr_t paddr; + + if (!find_futex_paddr(uaddr, &paddr)) + return 0; + + return get_and_cache_futex(paddr, uaddr); +} + + +/** Finds the physical address of the futex variable. */ +static bool find_futex_paddr(uintptr_t uaddr, uintptr_t *paddr) +{ + page_table_lock(AS, true); + + bool found = false; + pte_t *t = page_mapping_find(AS, ALIGN_DOWN(uaddr, PAGE_SIZE), false); + + if (t && PTE_VALID(t) && PTE_PRESENT(t)) { + found = true; + *paddr = PTE_GET_FRAME(t) + (uaddr - ALIGN_DOWN(uaddr, PAGE_SIZE)); + } + + page_table_unlock(AS, true); + + return found; +} + +/** Returns the futex cached in this task with the virtual address uaddr. */ +static futex_t *find_cached_futex(uintptr_t uaddr) +{ + cht_read_lock(); + + futex_t *futex; + cht_link_t *futex_ptr_link = cht_find_lazy(&TASK->futexes->ht, &uaddr); + + if (futex_ptr_link) { + futex_ptr_t *futex_ptr + = member_to_inst(futex_ptr_link, futex_ptr_t, cht_link); + + futex = futex_ptr->futex; + } else { + futex = NULL; + } + + cht_read_unlock(); + + return futex; +} + + +/** + * Returns a kernel futex for the physical address @a phys_addr and caches + * it in this task under the virtual address @a uaddr (if not already cached). + */ +static futex_t *get_and_cache_futex(uintptr_t phys_addr, uintptr_t uaddr) +{ + futex_t *futex = malloc(sizeof(futex_t), 0); + + /* + * Find the futex object in the global futex table (or insert it + * if it is not present). + */ + mutex_lock(&futex_ht_lock); + + link_t *fut_link = hash_table_find(&futex_ht, &phys_addr); + + if (fut_link) { + free(futex); + futex = member_to_inst(fut_link, futex_t, ht_link); + futex_add_ref(futex); + } else { + futex_initialize(futex, phys_addr); + hash_table_insert(&futex_ht, &phys_addr, &futex->ht_link); + } + + mutex_unlock(&futex_ht_lock); + + /* + * Cache the link to the futex object for this task. + */ + futex_ptr_t *fut_ptr = malloc(sizeof(futex_ptr_t), 0); + cht_link_t *dup_link; + + fut_ptr->futex = futex; + fut_ptr->uaddr = uaddr; + + cht_read_lock(); + + /* Cache the mapping from the virtual address to the futex for this task. */ + if (cht_insert_unique(&TASK->futexes->ht, &fut_ptr->cht_link, &dup_link)) { + mutex_lock(&TASK->futexes->list_lock); + list_append(&fut_ptr->all_link, &TASK->futexes->list); + mutex_unlock(&TASK->futexes->list_lock); + } else { + /* Another thread of this task beat us to it. Use that mapping instead.*/ + free(fut_ptr); + futex_release_ref_locked(futex); + + futex_ptr_t *dup = member_to_inst(dup_link, futex_ptr_t, cht_link); + futex = dup->futex; + } + + cht_read_unlock(); + + return futex; } @@ -109,27 +393,13 @@ sysarg_t sys_futex_sleep(uintptr_t uaddr) { - futex_t *futex; - uintptr_t paddr; - pte_t *t; - int rc; - - /* - * Find physical address of futex counter. - */ - page_table_lock(AS, true); - t = page_mapping_find(AS, ALIGN_DOWN(uaddr, PAGE_SIZE), false); - if (!t || !PTE_VALID(t) || !PTE_PRESENT(t)) { - page_table_unlock(AS, true); + futex_t *futex = get_futex(uaddr); + + if (!futex) return (sysarg_t) ENOENT; - } - paddr = PTE_GET_FRAME(t) + (uaddr - ALIGN_DOWN(uaddr, PAGE_SIZE)); - page_table_unlock(AS, true); - - futex = futex_find(paddr); #ifdef CONFIG_UDEBUG udebug_stoppable_begin(); #endif - rc = waitq_sleep_timeout(&futex->wq, 0, SYNCH_FLAGS_INTERRUPTIBLE); + int rc = waitq_sleep_timeout(&futex->wq, 0, SYNCH_FLAGS_INTERRUPTIBLE); #ifdef CONFIG_UDEBUG udebug_stoppable_end(); @@ -146,84 +416,14 @@ sysarg_t sys_futex_wakeup(uintptr_t uaddr) { - futex_t *futex; - uintptr_t paddr; - pte_t *t; - - /* - * Find physical address of futex counter. - */ - page_table_lock(AS, true); - t = page_mapping_find(AS, ALIGN_DOWN(uaddr, PAGE_SIZE), false); - if (!t || !PTE_VALID(t) || !PTE_PRESENT(t)) { - page_table_unlock(AS, true); + futex_t *futex = get_futex(uaddr); + + if (futex) { + waitq_wakeup(&futex->wq, WAKEUP_FIRST); + return 0; + } else { return (sysarg_t) ENOENT; } - paddr = PTE_GET_FRAME(t) + (uaddr - ALIGN_DOWN(uaddr, PAGE_SIZE)); - page_table_unlock(AS, true); - - futex = futex_find(paddr); - - waitq_wakeup(&futex->wq, WAKEUP_FIRST); - - return 0; -} - -/** Find kernel address of the futex structure corresponding to paddr. - * - * If the structure does not exist already, a new one is created. - * - * @param paddr Physical address of the userspace futex counter. - * - * @return Address of the kernel futex structure. - */ -futex_t *futex_find(uintptr_t paddr) -{ - link_t *item; - futex_t *futex; - btree_node_t *leaf; - - /* - * Find the respective futex structure - * or allocate new one if it does not exist already. - */ - mutex_lock(&futex_ht_lock); - item = hash_table_find(&futex_ht, &paddr); - if (item) { - futex = hash_table_get_instance(item, futex_t, ht_link); - - /* - * See if the current task knows this futex. - */ - mutex_lock(&TASK->futexes_lock); - if (!btree_search(&TASK->futexes, paddr, &leaf)) { - /* - * The futex is new to the current task. - * Upgrade its reference count and put it to the - * current task's B+tree of known futexes. - */ - futex->refcount++; - btree_insert(&TASK->futexes, paddr, futex, leaf); - } - mutex_unlock(&TASK->futexes_lock); - } else { - futex = (futex_t *) malloc(sizeof(futex_t), 0); - futex_initialize(futex); - futex->paddr = paddr; - hash_table_insert(&futex_ht, &paddr, &futex->ht_link); - - /* - * This is the first task referencing the futex. - * It can be directly inserted into its - * B+tree of known futexes. - */ - mutex_lock(&TASK->futexes_lock); - btree_insert(&TASK->futexes, paddr, futex, NULL); - mutex_unlock(&TASK->futexes_lock); - - } - mutex_unlock(&futex_ht_lock); - - return futex; -} +} + /** Compute hash index into futex hash table. @@ -268,28 +468,36 @@ } -/** Remove references from futexes known to the current task. */ -void futex_cleanup(void) -{ - mutex_lock(&futex_ht_lock); - mutex_lock(&TASK->futexes_lock); - - list_foreach(TASK->futexes.leaf_list, cur) { - btree_node_t *node; - unsigned int i; - - node = list_get_instance(cur, btree_node_t, leaf_link); - for (i = 0; i < node->keys; i++) { - futex_t *ftx; - uintptr_t paddr = node->key[i]; - - ftx = (futex_t *) node->value[i]; - if (--ftx->refcount == 0) - hash_table_remove(&futex_ht, &paddr, 1); - } - } - - mutex_unlock(&TASK->futexes_lock); - mutex_unlock(&futex_ht_lock); -} +/* + * Operations of a task's CHT that caches mappings of futex user space + * virtual addresses to kernel futex objects. + */ + +static size_t task_fut_ht_hash(const cht_link_t *link) +{ + const futex_ptr_t *fut_ptr = member_to_inst(link, futex_ptr_t, cht_link); + return fut_ptr->uaddr; +} + +static size_t task_fut_ht_key_hash(void *key) +{ + return *(uintptr_t*)key; +} + +static bool task_fut_ht_equal(const cht_link_t *item1, const cht_link_t *item2) +{ + const futex_ptr_t *fut_ptr1 = member_to_inst(item1, futex_ptr_t, cht_link); + const futex_ptr_t *fut_ptr2 = member_to_inst(item2, futex_ptr_t, cht_link); + + return fut_ptr1->uaddr == fut_ptr2->uaddr; +} + +static bool task_fut_ht_key_equal(void *key, const cht_link_t *item) +{ + const futex_ptr_t *fut_ptr = member_to_inst(item, futex_ptr_t, cht_link); + uintptr_t uaddr = *(uintptr_t*)key; + + return fut_ptr->uaddr == uaddr; +} + /** @} Index: kernel/generic/src/synch/mutex.c =================================================================== --- kernel/generic/src/synch/mutex.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/mutex.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,6 @@ #include #include +#include +#include /** Initialize mutex. Index: kernel/generic/src/synch/rcu.c =================================================================== --- kernel/generic/src/synch/rcu.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/synch/rcu.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,1873 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + + +/** @addtogroup sync + * @{ + */ + +/** + * @file + * @brief Preemptible read-copy update. Usable from interrupt handlers. + * + * @par Podzimek-preempt-RCU (RCU_PREEMPT_PODZIMEK) + * + * Podzimek-preempt-RCU is a preemptible variant of Podzimek's non-preemptible + * RCU algorithm [1, 2]. Grace period (GP) detection is centralized into a + * single detector thread. The detector requests that each cpu announces + * that it passed a quiescent state (QS), ie a state when the cpu is + * outside of an rcu reader section (CS). Cpus check for QSs during context + * switches and when entering and exiting rcu reader sections. Once all + * cpus announce a QS and if there were no threads preempted in a CS, the + * GP ends. + * + * The detector increments the global GP counter, _rcu_cur_gp, in order + * to start a new GP. Readers notice the new GP by comparing the changed + * _rcu_cur_gp to a locally stored value last_seen_gp which denotes the + * the last GP number for which the cpu noted an explicit QS (and issued + * a memory barrier). Readers check for the change in the outer-most + * (ie not nested) rcu_read_lock()/unlock() as these functions represent + * a QS. The reader first executes a memory barrier (MB) in order to contain + * memory references within a CS (and to make changes made by writers + * visible in the CS following rcu_read_lock()). Next, the reader notes + * that it reached a QS by updating the cpu local last_seen_gp to the + * global GP counter, _rcu_cur_gp. Cache coherency eventually makes + * the updated last_seen_gp visible to the detector cpu, much like it + * delivered the changed _rcu_cur_gp to all cpus. + * + * The detector waits a while after starting a GP and then reads each + * cpu's last_seen_gp to see if it reached a QS. If a cpu did not record + * a QS (might be a long running thread without an RCU reader CS; or cache + * coherency has yet to make the most current last_seen_gp visible to + * the detector; or the cpu is still in a CS) the cpu is interrupted + * via an IPI. If the IPI handler finds the cpu still in a CS, it instructs + * the cpu to notify the detector that it had exited the CS via a semaphore + * (CPU->rcu.is_delaying_gp). + * The detector then waits on the semaphore for any cpus to exit their + * CSs. Lastly, it waits for the last reader preempted in a CS to + * exit its CS if there were any and signals the end of the GP to + * separate reclaimer threads wired to each cpu. Reclaimers then + * execute the callbacks queued on each of the cpus. + * + * + * @par A-RCU algorithm (RCU_PREEMPT_A) + * + * A-RCU is based on the user space rcu algorithm in [3] utilizing signals + * (urcu) and Podzimek's rcu [1]. Like in Podzimek's rcu, callbacks are + * executed by cpu-bound reclaimer threads. There is however no dedicated + * detector thread and the reclaimers take on the responsibilities of the + * detector when they need to start a new GP. A new GP is again announced + * and acknowledged with _rcu_cur_gp and the cpu local last_seen_gp. Unlike + * Podzimek's rcu, cpus check explicitly for QS only during context switches. + * Like in urcu, rcu_read_lock()/unlock() only maintain the nesting count + * and never issue any memory barriers. This makes rcu_read_lock()/unlock() + * simple and fast. + * + * If a new callback is queued for a reclaimer and no GP is in progress, + * the reclaimer takes on the role of a detector. The detector increments + * _rcu_cur_gp in order to start a new GP. It waits a while to give cpus + * a chance to switch a context (a natural QS). Then, it examines each + * non-idle cpu that has yet to pass a QS via an IPI. The IPI handler + * sees the most current _rcu_cur_gp and last_seen_gp and notes a QS + * with a memory barrier and an update to last_seen_gp. If the handler + * finds the cpu in a CS it does nothing and let the detector poll/interrupt + * the cpu again after a short sleep. + * + * @par Caveats + * + * last_seen_gp and _rcu_cur_gp are always 64bit variables and they + * are read non-atomically on 32bit machines. Reading a clobbered + * value of last_seen_gp or _rcu_cur_gp or writing a clobbered value + * of _rcu_cur_gp to last_seen_gp will at worst force the detector + * to unnecessarily interrupt a cpu. Interrupting a cpu makes the + * correct value of _rcu_cur_gp visible to the cpu and correctly + * resets last_seen_gp in both algorithms. + * + * + * + * [1] Read-copy-update for opensolaris, + * 2010, Podzimek + * https://andrej.podzimek.org/thesis.pdf + * + * [2] (podzimek-rcu) implementation file "rcu.patch" + * http://d3s.mff.cuni.cz/projects/operating_systems/rcu/rcu.patch + * + * [3] User-level implementations of read-copy update, + * 2012, appendix + * http://www.rdrop.com/users/paulmck/RCU/urcu-supp-accepted.2011.08.30a.pdf + * + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* + * Number of milliseconds to give to preexisting readers to finish + * when non-expedited grace period detection is in progress. + */ +#define DETECT_SLEEP_MS 10 +/* + * Max number of pending callbacks in the local cpu's queue before + * aggressively expediting the current grace period + */ +#define EXPEDITE_THRESHOLD 2000 +/* + * Max number of callbacks to execute in one go with preemption + * enabled. If there are more callbacks to be executed they will + * be run with preemption disabled in order to prolong reclaimer's + * time slice and give it a chance to catch up with callback producers. + */ +#define CRITICAL_THRESHOLD 30000 +/* Half the number of values a uint32 can hold. */ +#define UINT32_MAX_HALF 2147483648U + +/** + * The current grace period number. Increases monotonically. + * Lock rcu.gp_lock or rcu.preempt_lock to get a current value. + */ +rcu_gp_t _rcu_cur_gp; + +/** Global RCU data. */ +typedef struct rcu_data { + /** Detector uses so signal reclaimers that a grace period ended. */ + condvar_t gp_ended; + /** Reclaimers use to notify the detector to accelerate GP detection. */ + condvar_t expedite_now; + /** + * Protects: req_gp_end_cnt, req_expedited_cnt, completed_gp, _rcu_cur_gp; + * or: completed_gp, _rcu_cur_gp + */ + SPINLOCK_DECLARE(gp_lock); + /** + * The number of the most recently completed grace period. At most + * one behind _rcu_cur_gp. If equal to _rcu_cur_gp, a grace period + * detection is not in progress and the detector is idle. + */ + rcu_gp_t completed_gp; + + /** Protects the following 3 fields. */ + IRQ_SPINLOCK_DECLARE(preempt_lock); + /** Preexisting readers that have been preempted. */ + list_t cur_preempted; + /** Reader that have been preempted and might delay the next grace period.*/ + list_t next_preempted; + /** + * The detector is waiting for the last preempted reader + * in cur_preempted to announce that it exited its reader + * section by up()ing remaining_readers. + */ + bool preempt_blocking_det; + +#ifdef RCU_PREEMPT_A + + /** + * The detector waits on this semaphore for any preempted readers + * delaying the grace period once all cpus pass a quiescent state. + */ + semaphore_t remaining_readers; + +#elif defined(RCU_PREEMPT_PODZIMEK) + + /** Reclaimers notify the detector when they request more grace periods.*/ + condvar_t req_gp_changed; + /** Number of grace period ends the detector was requested to announce. */ + size_t req_gp_end_cnt; + /** Number of consecutive grace periods to detect quickly and aggressively.*/ + size_t req_expedited_cnt; + /** + * Number of cpus with readers that are delaying the current GP. + * They will up() remaining_readers. + */ + atomic_t delaying_cpu_cnt; + /** + * The detector waits on this semaphore for any readers delaying the GP. + * + * Each of the cpus with readers that are delaying the current GP + * must up() this sema once they reach a quiescent state. If there + * are any readers in cur_preempted (ie preempted preexisting) and + * they are already delaying GP detection, the last to unlock its + * reader section must up() this sema once. + */ + semaphore_t remaining_readers; +#endif + + /** Excludes simultaneous rcu_barrier() calls. */ + mutex_t barrier_mtx; + /** Number of cpus that we are waiting for to complete rcu_barrier(). */ + atomic_t barrier_wait_cnt; + /** rcu_barrier() waits for the completion of barrier callbacks on this wq.*/ + waitq_t barrier_wq; + + /** Interruptible attached detector thread pointer. */ + thread_t *detector_thr; + + /* Some statistics. */ + size_t stat_expedited_cnt; + size_t stat_delayed_cnt; + size_t stat_preempt_blocking_cnt; + /* Does not contain self/local calls. */ + size_t stat_smp_call_cnt; +} rcu_data_t; + + +static rcu_data_t rcu; + +static void start_reclaimers(void); +static void synch_complete(rcu_item_t *rcu_item); +static inline void rcu_call_impl(bool expedite, rcu_item_t *rcu_item, + rcu_func_t func); +static void add_barrier_cb(void *arg); +static void barrier_complete(rcu_item_t *barrier_item); +static bool arriving_cbs_empty(void); +static bool next_cbs_empty(void); +static bool cur_cbs_empty(void); +static bool all_cbs_empty(void); +static void reclaimer(void *arg); +static bool wait_for_pending_cbs(void); +static bool advance_cbs(void); +static void exec_completed_cbs(rcu_gp_t last_completed_gp); +static void exec_cbs(rcu_item_t **phead); +static bool wait_for_cur_cbs_gp_end(bool expedite, rcu_gp_t *last_completed_gp); +static void upd_missed_gp_in_wait(rcu_gp_t completed_gp); + +#ifdef RCU_PREEMPT_PODZIMEK +static void start_detector(void); +static void read_unlock_impl(size_t *pnesting_cnt); +static void req_detection(size_t req_cnt); +static bool cv_wait_for_gp(rcu_gp_t wait_on_gp); +static void detector(void *); +static bool wait_for_detect_req(void); +static void end_cur_gp(void); +static bool wait_for_readers(void); +static bool gp_sleep(void); +static void interrupt_delaying_cpus(cpu_mask_t *cpu_mask); +static bool wait_for_delaying_cpus(void); +#elif defined(RCU_PREEMPT_A) +static bool wait_for_readers(bool expedite); +static bool gp_sleep(bool *expedite); +#endif + +static void start_new_gp(void); +static void rm_quiescent_cpus(cpu_mask_t *cpu_mask); +static void sample_cpus(cpu_mask_t *reader_cpus, void *arg); +static void sample_local_cpu(void *); +static bool wait_for_preempt_reader(void); +static void note_preempted_reader(void); +static void rm_preempted_reader(void); +static void upd_max_cbs_in_slice(size_t arriving_cbs_cnt); + + + +/** Initializes global RCU structures. */ +void rcu_init(void) +{ + condvar_initialize(&rcu.gp_ended); + condvar_initialize(&rcu.expedite_now); + + spinlock_initialize(&rcu.gp_lock, "rcu.gp_lock"); + _rcu_cur_gp = 0; + rcu.completed_gp = 0; + + irq_spinlock_initialize(&rcu.preempt_lock, "rcu.preempt_lock"); + list_initialize(&rcu.cur_preempted); + list_initialize(&rcu.next_preempted); + rcu.preempt_blocking_det = false; + + mutex_initialize(&rcu.barrier_mtx, MUTEX_PASSIVE); + atomic_set(&rcu.barrier_wait_cnt, 0); + waitq_initialize(&rcu.barrier_wq); + + semaphore_initialize(&rcu.remaining_readers, 0); + +#ifdef RCU_PREEMPT_PODZIMEK + condvar_initialize(&rcu.req_gp_changed); + + rcu.req_gp_end_cnt = 0; + rcu.req_expedited_cnt = 0; + atomic_set(&rcu.delaying_cpu_cnt, 0); +#endif + + rcu.detector_thr = NULL; + + rcu.stat_expedited_cnt = 0; + rcu.stat_delayed_cnt = 0; + rcu.stat_preempt_blocking_cnt = 0; + rcu.stat_smp_call_cnt = 0; +} + +/** Initializes per-CPU RCU data. If on the boot cpu inits global data too.*/ +void rcu_cpu_init(void) +{ + if (config.cpu_active == 1) { + rcu_init(); + } + + CPU->rcu.last_seen_gp = 0; + +#ifdef RCU_PREEMPT_PODZIMEK + CPU->rcu.nesting_cnt = 0; + CPU->rcu.is_delaying_gp = false; + CPU->rcu.signal_unlock = false; +#endif + + CPU->rcu.cur_cbs = NULL; + CPU->rcu.cur_cbs_cnt = 0; + CPU->rcu.next_cbs = NULL; + CPU->rcu.next_cbs_cnt = 0; + CPU->rcu.arriving_cbs = NULL; + CPU->rcu.parriving_cbs_tail = &CPU->rcu.arriving_cbs; + CPU->rcu.arriving_cbs_cnt = 0; + + CPU->rcu.cur_cbs_gp = 0; + CPU->rcu.next_cbs_gp = 0; + + semaphore_initialize(&CPU->rcu.arrived_flag, 0); + + /* BSP creates reclaimer threads before AP's rcu_cpu_init() runs. */ + if (config.cpu_active == 1) + CPU->rcu.reclaimer_thr = NULL; + + CPU->rcu.stat_max_cbs = 0; + CPU->rcu.stat_avg_cbs = 0; + CPU->rcu.stat_missed_gps = 0; + CPU->rcu.stat_missed_gp_in_wait = 0; + CPU->rcu.stat_max_slice_cbs = 0; + CPU->rcu.last_arriving_cnt = 0; +} + +/** Completes RCU init. Creates and runs the detector and reclaimer threads.*/ +void rcu_kinit_init(void) +{ +#ifdef RCU_PREEMPT_PODZIMEK + start_detector(); +#endif + + start_reclaimers(); +} + +/** Initializes any per-thread RCU structures. */ +void rcu_thread_init(thread_t *thread) +{ + thread->rcu.nesting_cnt = 0; + +#ifdef RCU_PREEMPT_PODZIMEK + thread->rcu.was_preempted = false; +#endif + + link_initialize(&thread->rcu.preempt_link); +} + + +/** Cleans up global RCU resources and stops dispatching callbacks. + * + * Call when shutting down the kernel. Outstanding callbacks will + * not be processed. Instead they will linger forever. + */ +void rcu_stop(void) +{ + /* Stop and wait for reclaimers. */ + for (unsigned int cpu_id = 0; cpu_id < config.cpu_active; ++cpu_id) { + ASSERT(cpus[cpu_id].rcu.reclaimer_thr != NULL); + + if (cpus[cpu_id].rcu.reclaimer_thr) { + thread_interrupt(cpus[cpu_id].rcu.reclaimer_thr); + thread_join(cpus[cpu_id].rcu.reclaimer_thr); + thread_detach(cpus[cpu_id].rcu.reclaimer_thr); + cpus[cpu_id].rcu.reclaimer_thr = NULL; + } + } + +#ifdef RCU_PREEMPT_PODZIMEK + /* Stop the detector and wait. */ + if (rcu.detector_thr) { + thread_interrupt(rcu.detector_thr); + thread_join(rcu.detector_thr); + thread_detach(rcu.detector_thr); + rcu.detector_thr = NULL; + } +#endif +} + +/** Returns the number of elapsed grace periods since boot. */ +uint64_t rcu_completed_gps(void) +{ + spinlock_lock(&rcu.gp_lock); + uint64_t completed = rcu.completed_gp; + spinlock_unlock(&rcu.gp_lock); + + return completed; +} + +/** Creates and runs cpu-bound reclaimer threads. */ +static void start_reclaimers(void) +{ + for (unsigned int cpu_id = 0; cpu_id < config.cpu_count; ++cpu_id) { + char name[THREAD_NAME_BUFLEN] = {0}; + + snprintf(name, THREAD_NAME_BUFLEN - 1, "rcu-rec/%u", cpu_id); + + cpus[cpu_id].rcu.reclaimer_thr = + thread_create(reclaimer, NULL, TASK, THREAD_FLAG_NONE, name); + + if (!cpus[cpu_id].rcu.reclaimer_thr) + panic("Failed to create RCU reclaimer thread on cpu%u.", cpu_id); + + thread_wire(cpus[cpu_id].rcu.reclaimer_thr, &cpus[cpu_id]); + thread_ready(cpus[cpu_id].rcu.reclaimer_thr); + } +} + +#ifdef RCU_PREEMPT_PODZIMEK + +/** Starts the detector thread. */ +static void start_detector(void) +{ + rcu.detector_thr = + thread_create(detector, NULL, TASK, THREAD_FLAG_NONE, "rcu-det"); + + if (!rcu.detector_thr) + panic("Failed to create RCU detector thread."); + + thread_ready(rcu.detector_thr); +} + +/** Returns true if in an rcu reader section. */ +bool rcu_read_locked(void) +{ + preemption_disable(); + bool locked = 0 < CPU->rcu.nesting_cnt; + preemption_enable(); + + return locked; +} + +/** Unlocks the local reader section using the given nesting count. + * + * Preemption or interrupts must be disabled. + * + * @param pnesting_cnt Either &CPU->rcu.tmp_nesting_cnt or + * THREAD->rcu.nesting_cnt. + */ +static void read_unlock_impl(size_t *pnesting_cnt) +{ + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + + if (0 == --(*pnesting_cnt)) { + _rcu_record_qs(); + + /* + * The thread was preempted while in a critical section or + * the detector is eagerly waiting for this cpu's reader + * to finish. + * + * Note that THREAD may be NULL in scheduler() and not just during boot. + */ + if ((THREAD && THREAD->rcu.was_preempted) || CPU->rcu.is_delaying_gp) { + /* Rechecks with disabled interrupts. */ + _rcu_signal_read_unlock(); + } + } +} + +/** If necessary, signals the detector that we exited a reader section. */ +void _rcu_signal_read_unlock(void) +{ + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + + /* + * If an interrupt occurs here (even a NMI) it may beat us to + * resetting .is_delaying_gp or .was_preempted and up the semaphore + * for us. + */ + + /* + * If the detector is eagerly waiting for this cpu's reader to unlock, + * notify it that the reader did so. + */ + if (local_atomic_exchange(&CPU->rcu.is_delaying_gp, false)) { + semaphore_up(&rcu.remaining_readers); + } + + /* + * This reader was preempted while in a reader section. + * We might be holding up the current GP. Notify the + * detector if so. + */ + if (THREAD && local_atomic_exchange(&THREAD->rcu.was_preempted, false)) { + ASSERT(link_used(&THREAD->rcu.preempt_link)); + + rm_preempted_reader(); + } + + /* If there was something to signal to the detector we have done so. */ + CPU->rcu.signal_unlock = false; +} + +#endif /* RCU_PREEMPT_PODZIMEK */ + +typedef struct synch_item { + waitq_t wq; + rcu_item_t rcu_item; +} synch_item_t; + +/** Blocks until all preexisting readers exit their critical sections. */ +void rcu_synchronize(void) +{ + _rcu_synchronize(false); +} + +/** Blocks until all preexisting readers exit their critical sections. */ +void rcu_synchronize_expedite(void) +{ + _rcu_synchronize(true); +} + +/** Blocks until all preexisting readers exit their critical sections. */ +void _rcu_synchronize(bool expedite) +{ + /* Calling from a reader section will deadlock. */ + ASSERT(!rcu_read_locked()); + + synch_item_t completion; + + waitq_initialize(&completion.wq); + _rcu_call(expedite, &completion.rcu_item, synch_complete); + waitq_sleep(&completion.wq); +} + +/** rcu_synchronize's callback. */ +static void synch_complete(rcu_item_t *rcu_item) +{ + synch_item_t *completion = member_to_inst(rcu_item, synch_item_t, rcu_item); + ASSERT(completion); + waitq_wakeup(&completion->wq, WAKEUP_FIRST); +} + +/** Waits for all outstanding rcu calls to complete. */ +void rcu_barrier(void) +{ + /* + * Serialize rcu_barrier() calls so we don't overwrite cpu.barrier_item + * currently in use by rcu_barrier(). + */ + mutex_lock(&rcu.barrier_mtx); + + /* + * Ensure we queue a barrier callback on all cpus before the already + * enqueued barrier callbacks start signaling completion. + */ + atomic_set(&rcu.barrier_wait_cnt, 1); + + DEFINE_CPU_MASK(cpu_mask); + cpu_mask_active(cpu_mask); + + cpu_mask_for_each(*cpu_mask, cpu_id) { + smp_call(cpu_id, add_barrier_cb, NULL); + } + + if (0 < atomic_predec(&rcu.barrier_wait_cnt)) { + waitq_sleep(&rcu.barrier_wq); + } + + mutex_unlock(&rcu.barrier_mtx); +} + +/** Issues a rcu_barrier() callback on the local cpu. + * + * Executed with interrupts disabled. + */ +static void add_barrier_cb(void *arg) +{ + ASSERT(interrupts_disabled() || PREEMPTION_DISABLED); + atomic_inc(&rcu.barrier_wait_cnt); + rcu_call(&CPU->rcu.barrier_item, barrier_complete); +} + +/** Local cpu's rcu_barrier() completion callback. */ +static void barrier_complete(rcu_item_t *barrier_item) +{ + /* Is this the last barrier callback completed? */ + if (0 == atomic_predec(&rcu.barrier_wait_cnt)) { + /* Notify rcu_barrier() that we're done. */ + waitq_wakeup(&rcu.barrier_wq, WAKEUP_FIRST); + } +} + +/** Adds a callback to invoke after all preexisting readers finish. + * + * May be called from within interrupt handlers or RCU reader sections. + * + * @param rcu_item Used by RCU to track the call. Must remain + * until the user callback function is entered. + * @param func User callback function that will be invoked once a full + * grace period elapsed, ie at a time when all preexisting + * readers have finished. The callback should be short and must + * not block. If you must sleep, enqueue your work in the system + * work queue from the callback (ie workq_global_enqueue()). + */ +void rcu_call(rcu_item_t *rcu_item, rcu_func_t func) +{ + rcu_call_impl(false, rcu_item, func); +} + +/** rcu_call() implementation. See rcu_call() for comments. */ +void _rcu_call(bool expedite, rcu_item_t *rcu_item, rcu_func_t func) +{ + rcu_call_impl(expedite, rcu_item, func); +} + +/** rcu_call() inline-able implementation. See rcu_call() for comments. */ +static inline void rcu_call_impl(bool expedite, rcu_item_t *rcu_item, + rcu_func_t func) +{ + ASSERT(rcu_item); + + rcu_item->func = func; + rcu_item->next = NULL; + + preemption_disable(); + + rcu_cpu_data_t *r = &CPU->rcu; + + rcu_item_t **prev_tail + = local_atomic_exchange(&r->parriving_cbs_tail, &rcu_item->next); + *prev_tail = rcu_item; + + /* Approximate the number of callbacks present. */ + ++r->arriving_cbs_cnt; + + if (expedite) { + r->expedite_arriving = true; + } + + bool first_cb = (prev_tail == &CPU->rcu.arriving_cbs); + + /* Added first callback - notify the reclaimer. */ + if (first_cb && !semaphore_count_get(&r->arrived_flag)) { + semaphore_up(&r->arrived_flag); + } + + preemption_enable(); +} + +static bool cur_cbs_empty(void) +{ + ASSERT(THREAD && THREAD->wired); + return NULL == CPU->rcu.cur_cbs; +} + +static bool next_cbs_empty(void) +{ + ASSERT(THREAD && THREAD->wired); + return NULL == CPU->rcu.next_cbs; +} + +/** Disable interrupts to get an up-to-date result. */ +static bool arriving_cbs_empty(void) +{ + ASSERT(THREAD && THREAD->wired); + /* + * Accessing with interrupts enabled may at worst lead to + * a false negative if we race with a local interrupt handler. + */ + return NULL == CPU->rcu.arriving_cbs; +} + +static bool all_cbs_empty(void) +{ + return cur_cbs_empty() && next_cbs_empty() && arriving_cbs_empty(); +} + + +/** Reclaimer thread dispatches locally queued callbacks once a GP ends. */ +static void reclaimer(void *arg) +{ + ASSERT(THREAD && THREAD->wired); + ASSERT(THREAD == CPU->rcu.reclaimer_thr); + + rcu_gp_t last_compl_gp = 0; + bool ok = true; + + while (ok && wait_for_pending_cbs()) { + ASSERT(CPU->rcu.reclaimer_thr == THREAD); + + exec_completed_cbs(last_compl_gp); + + bool expedite = advance_cbs(); + + ok = wait_for_cur_cbs_gp_end(expedite, &last_compl_gp); + } +} + +/** Waits until there are callbacks waiting to be dispatched. */ +static bool wait_for_pending_cbs(void) +{ + if (!all_cbs_empty()) + return true; + + bool ok = true; + + while (arriving_cbs_empty() && ok) { + ok = semaphore_down_interruptable(&CPU->rcu.arrived_flag); + } + + return ok; +} + +static void upd_stat_missed_gp(rcu_gp_t compl) +{ + if (CPU->rcu.cur_cbs_gp < compl) { + CPU->rcu.stat_missed_gps += (size_t)(compl - CPU->rcu.cur_cbs_gp); + } +} + +/** Executes all callbacks for the given completed grace period. */ +static void exec_completed_cbs(rcu_gp_t last_completed_gp) +{ + upd_stat_missed_gp(last_completed_gp); + + /* Both next_cbs and cur_cbs GP elapsed. */ + if (CPU->rcu.next_cbs_gp <= last_completed_gp) { + ASSERT(CPU->rcu.cur_cbs_gp <= CPU->rcu.next_cbs_gp); + + size_t exec_cnt = CPU->rcu.cur_cbs_cnt + CPU->rcu.next_cbs_cnt; + + if (exec_cnt < CRITICAL_THRESHOLD) { + exec_cbs(&CPU->rcu.cur_cbs); + exec_cbs(&CPU->rcu.next_cbs); + } else { + /* + * Getting overwhelmed with too many callbacks to run. + * Disable preemption in order to prolong our time slice + * and catch up with updaters posting new callbacks. + */ + preemption_disable(); + exec_cbs(&CPU->rcu.cur_cbs); + exec_cbs(&CPU->rcu.next_cbs); + preemption_enable(); + } + + CPU->rcu.cur_cbs_cnt = 0; + CPU->rcu.next_cbs_cnt = 0; + } else if (CPU->rcu.cur_cbs_gp <= last_completed_gp) { + + if (CPU->rcu.cur_cbs_cnt < CRITICAL_THRESHOLD) { + exec_cbs(&CPU->rcu.cur_cbs); + } else { + /* + * Getting overwhelmed with too many callbacks to run. + * Disable preemption in order to prolong our time slice + * and catch up with updaters posting new callbacks. + */ + preemption_disable(); + exec_cbs(&CPU->rcu.cur_cbs); + preemption_enable(); + } + + CPU->rcu.cur_cbs_cnt = 0; + } +} + +/** Executes callbacks in the single-linked list. The list is left empty. */ +static void exec_cbs(rcu_item_t **phead) +{ + rcu_item_t *rcu_item = *phead; + + while (rcu_item) { + /* func() may free rcu_item. Get a local copy. */ + rcu_item_t *next = rcu_item->next; + rcu_func_t func = rcu_item->func; + + func(rcu_item); + + rcu_item = next; + } + + *phead = NULL; +} + +static void upd_stat_cb_cnts(size_t arriving_cnt) +{ + CPU->rcu.stat_max_cbs = max(arriving_cnt, CPU->rcu.stat_max_cbs); + if (0 < arriving_cnt) { + CPU->rcu.stat_avg_cbs = + (99 * CPU->rcu.stat_avg_cbs + 1 * arriving_cnt) / 100; + } +} + +/** Prepares another batch of callbacks to dispatch at the nest grace period. + * + * @return True if the next batch of callbacks must be expedited quickly. + */ +static bool advance_cbs(void) +{ + /* Move next_cbs to cur_cbs. */ + CPU->rcu.cur_cbs = CPU->rcu.next_cbs; + CPU->rcu.cur_cbs_cnt = CPU->rcu.next_cbs_cnt; + CPU->rcu.cur_cbs_gp = CPU->rcu.next_cbs_gp; + + /* Move arriving_cbs to next_cbs. */ + + CPU->rcu.next_cbs_cnt = CPU->rcu.arriving_cbs_cnt; + CPU->rcu.arriving_cbs_cnt = 0; + + /* + * Too many callbacks queued. Better speed up the detection + * or risk exhausting all system memory. + */ + bool expedite = (EXPEDITE_THRESHOLD < CPU->rcu.next_cbs_cnt) + || CPU->rcu.expedite_arriving; + CPU->rcu.expedite_arriving = false; + + /* Start moving the arriving_cbs list to next_cbs. */ + CPU->rcu.next_cbs = CPU->rcu.arriving_cbs; + + /* + * At least one callback arrived. The tail therefore does not point + * to the head of arriving_cbs and we can safely reset it to NULL. + */ + if (CPU->rcu.next_cbs) { + ASSERT(CPU->rcu.parriving_cbs_tail != &CPU->rcu.arriving_cbs); + + CPU->rcu.arriving_cbs = NULL; + /* Reset arriving_cbs before updating the tail pointer. */ + compiler_barrier(); + /* Updating the tail pointer completes the move of arriving_cbs. */ + ACCESS_ONCE(CPU->rcu.parriving_cbs_tail) = &CPU->rcu.arriving_cbs; + } else { + /* + * arriving_cbs was null and parriving_cbs_tail pointed to it + * so leave it that way. Note that interrupt handlers may have + * added a callback in the meantime so it is not safe to reset + * arriving_cbs or parriving_cbs. + */ + } + + /* Update statistics of arrived callbacks. */ + upd_stat_cb_cnts(CPU->rcu.next_cbs_cnt); + + /* + * Make changes prior to queuing next_cbs visible to readers. + * See comment in wait_for_readers(). + */ + memory_barrier(); /* MB A, B */ + + /* At the end of next_cbs_gp, exec next_cbs. Determine what GP that is. */ + + if (!next_cbs_empty()) { + spinlock_lock(&rcu.gp_lock); + + /* Exec next_cbs at the end of the next GP. */ + CPU->rcu.next_cbs_gp = _rcu_cur_gp + 1; + + /* + * There are no callbacks to invoke before next_cbs. Instruct + * wait_for_cur_cbs_gp() to notify us of the nearest GP end. + * That could be sooner than next_cbs_gp (if the current GP + * had not yet completed), so we'll create a shorter batch + * of callbacks next time around. + */ + if (cur_cbs_empty()) { + CPU->rcu.cur_cbs_gp = rcu.completed_gp + 1; + } + + spinlock_unlock(&rcu.gp_lock); + } else { + CPU->rcu.next_cbs_gp = CPU->rcu.cur_cbs_gp; + } + + ASSERT(CPU->rcu.cur_cbs_gp <= CPU->rcu.next_cbs_gp); + + return expedite; +} + + +#ifdef RCU_PREEMPT_A + +/** Waits for the grace period associated with callbacks cub_cbs to elapse. + * + * @param expedite Instructs the detector to aggressively speed up grace + * period detection without any delay. + * @param completed_gp Returns the most recent completed grace period + * number. + * @return false if the thread was interrupted and should stop. + */ +static bool wait_for_cur_cbs_gp_end(bool expedite, rcu_gp_t *completed_gp) +{ + spinlock_lock(&rcu.gp_lock); + + ASSERT(CPU->rcu.cur_cbs_gp <= CPU->rcu.next_cbs_gp); + ASSERT(CPU->rcu.cur_cbs_gp <= _rcu_cur_gp + 1); + + while (rcu.completed_gp < CPU->rcu.cur_cbs_gp) { + /* GP has not yet started - start a new one. */ + if (rcu.completed_gp == _rcu_cur_gp) { + start_new_gp(); + spinlock_unlock(&rcu.gp_lock); + + if (!wait_for_readers(expedite)) + return false; + + spinlock_lock(&rcu.gp_lock); + /* Notify any reclaimers this GP had ended. */ + rcu.completed_gp = _rcu_cur_gp; + condvar_broadcast(&rcu.gp_ended); + } else { + /* GP detection is in progress.*/ + + if (expedite) + condvar_signal(&rcu.expedite_now); + + /* Wait for the GP to complete. */ + int ret = _condvar_wait_timeout_spinlock(&rcu.gp_ended, &rcu.gp_lock, + SYNCH_NO_TIMEOUT, SYNCH_FLAGS_INTERRUPTIBLE); + + if (ret == ESYNCH_INTERRUPTED) { + spinlock_unlock(&rcu.gp_lock); + return false; + } + } + } + + upd_missed_gp_in_wait(rcu.completed_gp); + + *completed_gp = rcu.completed_gp; + spinlock_unlock(&rcu.gp_lock); + + return true; +} + +static bool wait_for_readers(bool expedite) +{ + DEFINE_CPU_MASK(reader_cpus); + + cpu_mask_active(reader_cpus); + rm_quiescent_cpus(reader_cpus); + + while (!cpu_mask_is_none(reader_cpus)) { + /* Give cpus a chance to context switch (a QS) and batch callbacks. */ + if(!gp_sleep(&expedite)) + return false; + + rm_quiescent_cpus(reader_cpus); + sample_cpus(reader_cpus, reader_cpus); + } + + /* Update statistic. */ + if (expedite) { + ++rcu.stat_expedited_cnt; + } + + /* + * All cpus have passed through a QS and see the most recent _rcu_cur_gp. + * As a result newly preempted readers will associate with next_preempted + * and the number of old readers in cur_preempted will monotonically + * decrease. Wait for those old/preexisting readers. + */ + return wait_for_preempt_reader(); +} + +static bool gp_sleep(bool *expedite) +{ + if (*expedite) { + scheduler(); + return true; + } else { + spinlock_lock(&rcu.gp_lock); + + int ret = 0; + ret = _condvar_wait_timeout_spinlock(&rcu.expedite_now, &rcu.gp_lock, + DETECT_SLEEP_MS * 1000, SYNCH_FLAGS_INTERRUPTIBLE); + + /* rcu.expedite_now was signaled. */ + if (ret == ESYNCH_OK_BLOCKED) { + *expedite = true; + } + + spinlock_unlock(&rcu.gp_lock); + + return (ret != ESYNCH_INTERRUPTED); + } +} + +static void sample_local_cpu(void *arg) +{ + ASSERT(interrupts_disabled()); + cpu_mask_t *reader_cpus = (cpu_mask_t *)arg; + + bool locked = RCU_CNT_INC <= THE->rcu_nesting; + /* smp_call machinery makes the most current _rcu_cur_gp visible. */ + bool passed_qs = (CPU->rcu.last_seen_gp == _rcu_cur_gp); + + if (locked && !passed_qs) { + /* + * This cpu has not yet passed a quiescent state during this grace + * period and it is currently in a reader section. We'll have to + * try to sample this cpu again later. + */ + } else { + /* Either not in a reader section or already passed a QS. */ + cpu_mask_reset(reader_cpus, CPU->id); + /* Contain new reader sections and make prior changes visible to them.*/ + memory_barrier(); + CPU->rcu.last_seen_gp = _rcu_cur_gp; + } +} + +/** Called by the scheduler() when switching away from the current thread. */ +void rcu_after_thread_ran(void) +{ + ASSERT(interrupts_disabled()); + + /* + * In order not to worry about NMI seeing rcu_nesting change work + * with a local copy. + */ + size_t nesting_cnt = local_atomic_exchange(&THE->rcu_nesting, 0); + + /* + * Ensures NMIs see .rcu_nesting without the WAS_PREEMPTED mark and + * do not accidentally call rm_preempted_reader() from unlock(). + */ + compiler_barrier(); + + /* Preempted a reader critical section for the first time. */ + if (RCU_CNT_INC <= nesting_cnt && !(nesting_cnt & RCU_WAS_PREEMPTED)) { + nesting_cnt |= RCU_WAS_PREEMPTED; + note_preempted_reader(); + } + + /* Save the thread's nesting count when it is not running. */ + THREAD->rcu.nesting_cnt = nesting_cnt; + + if (CPU->rcu.last_seen_gp != _rcu_cur_gp) { + /* + * Contain any memory accesses of old readers before announcing a QS. + * Also make changes from the previous GP visible to this cpu. + * Moreover it separates writing to last_seen_gp from + * note_preempted_reader(). + */ + memory_barrier(); + /* + * The preempted reader has been noted globally. There are therefore + * no readers running on this cpu so this is a quiescent state. + * + * Reading the multiword _rcu_cur_gp non-atomically is benign. + * At worst, the read value will be different from the actual value. + * As a result, both the detector and this cpu will believe + * this cpu has not yet passed a QS although it really did. + * + * Reloading _rcu_cur_gp is benign, because it cannot change + * until this cpu acknowledges it passed a QS by writing to + * last_seen_gp. Since interrupts are disabled, only this + * code may to so (IPIs won't get through). + */ + CPU->rcu.last_seen_gp = _rcu_cur_gp; + } + + /* + * Forcefully associate the reclaimer with the highest priority + * even if preempted due to its time slice running out. + */ + if (THREAD == CPU->rcu.reclaimer_thr) { + THREAD->priority = -1; + } + + upd_max_cbs_in_slice(CPU->rcu.arriving_cbs_cnt); +} + +/** Called by the scheduler() when switching to a newly scheduled thread. */ +void rcu_before_thread_runs(void) +{ + ASSERT(!rcu_read_locked()); + + /* Load the thread's saved nesting count from before it was preempted. */ + THE->rcu_nesting = THREAD->rcu.nesting_cnt; +} + +/** Called from scheduler() when exiting the current thread. + * + * Preemption or interrupts are disabled and the scheduler() already + * switched away from the current thread, calling rcu_after_thread_ran(). + */ +void rcu_thread_exiting(void) +{ + ASSERT(THE->rcu_nesting == 0); + + /* + * The thread forgot to exit its reader critical section. + * It is a bug, but rather than letting the entire system lock up + * forcefully leave the reader section. The thread is not holding + * any references anyway since it is exiting so it is safe. + */ + if (RCU_CNT_INC <= THREAD->rcu.nesting_cnt) { + /* Emulate _rcu_preempted_unlock() with the proper nesting count. */ + if (THREAD->rcu.nesting_cnt & RCU_WAS_PREEMPTED) { + rm_preempted_reader(); + } + + printf("Bug: thread (id %" PRIu64 " \"%s\") exited while in RCU read" + " section.\n", THREAD->tid, THREAD->name); + } +} + +/** Returns true if in an rcu reader section. */ +bool rcu_read_locked(void) +{ + return RCU_CNT_INC <= THE->rcu_nesting; +} + +/** Invoked when a preempted reader finally exits its reader section. */ +void _rcu_preempted_unlock(void) +{ + ASSERT(0 == THE->rcu_nesting || RCU_WAS_PREEMPTED == THE->rcu_nesting); + + size_t prev = local_atomic_exchange(&THE->rcu_nesting, 0); + if (prev == RCU_WAS_PREEMPTED) { + /* + * NMI handlers are never preempted but may call rm_preempted_reader() + * if a NMI occurred in _rcu_preempted_unlock() of a preempted thread. + * The only other rcu code that may have been interrupted by the NMI + * in _rcu_preempted_unlock() is: an IPI/sample_local_cpu() and + * the initial part of rcu_after_thread_ran(). + * + * rm_preempted_reader() will not deadlock because none of the locks + * it uses are locked in this case. Neither _rcu_preempted_unlock() + * nor sample_local_cpu() nor the initial part of rcu_after_thread_ran() + * acquire any locks. + */ + rm_preempted_reader(); + } +} + +#elif defined(RCU_PREEMPT_PODZIMEK) + +/** Waits for the grace period associated with callbacks cub_cbs to elapse. + * + * @param expedite Instructs the detector to aggressively speed up grace + * period detection without any delay. + * @param completed_gp Returns the most recent completed grace period + * number. + * @return false if the thread was interrupted and should stop. + */ +static bool wait_for_cur_cbs_gp_end(bool expedite, rcu_gp_t *completed_gp) +{ + /* + * Use a possibly outdated version of completed_gp to bypass checking + * with the lock. + * + * Note that loading and storing rcu.completed_gp is not atomic + * (it is 64bit wide). Reading a clobbered value that is less than + * rcu.completed_gp is harmless - we'll recheck with a lock. The + * only way to read a clobbered value that is greater than the actual + * value is if the detector increases the higher-order word first and + * then decreases the lower-order word (or we see stores in that order), + * eg when incrementing from 2^32 - 1 to 2^32. The loaded value + * suddenly jumps by 2^32. It would take hours for such an increase + * to occur so it is safe to discard the value. We allow increases + * of up to half the maximum to generously accommodate for loading an + * outdated lower word. + */ + rcu_gp_t compl_gp = ACCESS_ONCE(rcu.completed_gp); + if (CPU->rcu.cur_cbs_gp <= compl_gp + && compl_gp <= CPU->rcu.cur_cbs_gp + UINT32_MAX_HALF) { + *completed_gp = compl_gp; + return true; + } + + spinlock_lock(&rcu.gp_lock); + + if (CPU->rcu.cur_cbs_gp <= rcu.completed_gp) { + *completed_gp = rcu.completed_gp; + spinlock_unlock(&rcu.gp_lock); + return true; + } + + ASSERT(CPU->rcu.cur_cbs_gp <= CPU->rcu.next_cbs_gp); + ASSERT(_rcu_cur_gp <= CPU->rcu.cur_cbs_gp); + + /* + * Notify the detector of how many GP ends we intend to wait for, so + * it can avoid going to sleep unnecessarily. Optimistically assume + * new callbacks will arrive while we're waiting; hence +1. + */ + size_t remaining_gp_ends = (size_t) (CPU->rcu.next_cbs_gp - _rcu_cur_gp); + req_detection(remaining_gp_ends + (arriving_cbs_empty() ? 0 : 1)); + + /* + * Ask the detector to speed up GP detection if there are too many + * pending callbacks and other reclaimers have not already done so. + */ + if (expedite) { + if(0 == rcu.req_expedited_cnt) + condvar_signal(&rcu.expedite_now); + + /* + * Expedite only cub_cbs. If there really is a surge of callbacks + * the arriving batch will expedite the GP for the huge number + * of callbacks currently in next_cbs + */ + rcu.req_expedited_cnt = 1; + } + + /* Wait for cur_cbs_gp to end. */ + bool interrupted = cv_wait_for_gp(CPU->rcu.cur_cbs_gp); + + *completed_gp = rcu.completed_gp; + spinlock_unlock(&rcu.gp_lock); + + if (!interrupted) + upd_missed_gp_in_wait(*completed_gp); + + return !interrupted; +} + +/** Waits for an announcement of the end of the grace period wait_on_gp. */ +static bool cv_wait_for_gp(rcu_gp_t wait_on_gp) +{ + ASSERT(spinlock_locked(&rcu.gp_lock)); + + bool interrupted = false; + + /* Wait until wait_on_gp ends. */ + while (rcu.completed_gp < wait_on_gp && !interrupted) { + int ret = _condvar_wait_timeout_spinlock(&rcu.gp_ended, &rcu.gp_lock, + SYNCH_NO_TIMEOUT, SYNCH_FLAGS_INTERRUPTIBLE); + interrupted = (ret == ESYNCH_INTERRUPTED); + } + + return interrupted; +} + +/** Requests the detector to detect at least req_cnt consecutive grace periods.*/ +static void req_detection(size_t req_cnt) +{ + if (rcu.req_gp_end_cnt < req_cnt) { + bool detector_idle = (0 == rcu.req_gp_end_cnt); + rcu.req_gp_end_cnt = req_cnt; + + if (detector_idle) { + ASSERT(_rcu_cur_gp == rcu.completed_gp); + condvar_signal(&rcu.req_gp_changed); + } + } +} + + +/** The detector thread detects and notifies reclaimers of grace period ends. */ +static void detector(void *arg) +{ + spinlock_lock(&rcu.gp_lock); + + while (wait_for_detect_req()) { + /* + * Announce new GP started. Readers start lazily acknowledging that + * they passed a QS. + */ + start_new_gp(); + + spinlock_unlock(&rcu.gp_lock); + + if (!wait_for_readers()) + goto unlocked_out; + + spinlock_lock(&rcu.gp_lock); + + /* Notify reclaimers that they may now invoke queued callbacks. */ + end_cur_gp(); + } + + spinlock_unlock(&rcu.gp_lock); + +unlocked_out: + return; +} + +/** Waits for a request from a reclaimer thread to detect a grace period. */ +static bool wait_for_detect_req(void) +{ + ASSERT(spinlock_locked(&rcu.gp_lock)); + + bool interrupted = false; + + while (0 == rcu.req_gp_end_cnt && !interrupted) { + int ret = _condvar_wait_timeout_spinlock(&rcu.req_gp_changed, + &rcu.gp_lock, SYNCH_NO_TIMEOUT, SYNCH_FLAGS_INTERRUPTIBLE); + + interrupted = (ret == ESYNCH_INTERRUPTED); + } + + return !interrupted; +} + + +static void end_cur_gp(void) +{ + ASSERT(spinlock_locked(&rcu.gp_lock)); + + rcu.completed_gp = _rcu_cur_gp; + --rcu.req_gp_end_cnt; + + condvar_broadcast(&rcu.gp_ended); +} + +/** Waits for readers that started before the current GP started to finish. */ +static bool wait_for_readers(void) +{ + DEFINE_CPU_MASK(reading_cpus); + + /* All running cpus have potential readers. */ + cpu_mask_active(reading_cpus); + + /* + * Give readers time to pass through a QS. Also, batch arriving + * callbacks in order to amortize detection overhead. + */ + if (!gp_sleep()) + return false; + + /* Non-intrusively determine which cpus have yet to pass a QS. */ + rm_quiescent_cpus(reading_cpus); + + /* Actively interrupt cpus delaying the current GP and demand a QS. */ + interrupt_delaying_cpus(reading_cpus); + + /* Wait for the interrupted cpus to notify us that they reached a QS. */ + if (!wait_for_delaying_cpus()) + return false; + /* + * All cpus recorded a QS or are still idle. Any new readers will be added + * to next_preempt if preempted, ie the number of readers in cur_preempted + * monotonically descreases. + */ + + /* Wait for the last reader in cur_preempted to notify us it is done. */ + if (!wait_for_preempt_reader()) + return false; + + return true; +} + +/** Sleeps a while if the current grace period is not to be expedited. */ +static bool gp_sleep(void) +{ + spinlock_lock(&rcu.gp_lock); + + int ret = 0; + while (0 == rcu.req_expedited_cnt && 0 == ret) { + /* minor bug: sleeps for the same duration if woken up spuriously. */ + ret = _condvar_wait_timeout_spinlock(&rcu.expedite_now, &rcu.gp_lock, + DETECT_SLEEP_MS * 1000, SYNCH_FLAGS_INTERRUPTIBLE); + } + + if (0 < rcu.req_expedited_cnt) { + --rcu.req_expedited_cnt; + /* Update statistic. */ + ++rcu.stat_expedited_cnt; + } + + spinlock_unlock(&rcu.gp_lock); + + return (ret != ESYNCH_INTERRUPTED); +} + +/** Actively interrupts and checks the offending cpus for quiescent states. */ +static void interrupt_delaying_cpus(cpu_mask_t *cpu_mask) +{ + atomic_set(&rcu.delaying_cpu_cnt, 0); + + sample_cpus(cpu_mask, NULL); +} + +/** Invoked on a cpu delaying grace period detection. + * + * Induces a quiescent state for the cpu or it instructs remaining + * readers to notify the detector once they finish. + */ +static void sample_local_cpu(void *arg) +{ + ASSERT(interrupts_disabled()); + ASSERT(!CPU->rcu.is_delaying_gp); + + /* Cpu did not pass a quiescent state yet. */ + if (CPU->rcu.last_seen_gp != _rcu_cur_gp) { + /* Interrupted a reader in a reader critical section. */ + if (0 < CPU->rcu.nesting_cnt) { + ASSERT(!CPU->idle); + /* + * Note to notify the detector from rcu_read_unlock(). + * + * ACCESS_ONCE ensures the compiler writes to is_delaying_gp + * only after it determines that we are in a reader CS. + */ + ACCESS_ONCE(CPU->rcu.is_delaying_gp) = true; + CPU->rcu.signal_unlock = true; + + atomic_inc(&rcu.delaying_cpu_cnt); + } else { + /* + * The cpu did not enter any rcu reader sections since + * the start of the current GP. Record a quiescent state. + * + * Or, we interrupted rcu_read_unlock_impl() right before + * it recorded a QS. Record a QS for it. The memory barrier + * contains the reader section's mem accesses before + * updating last_seen_gp. + * + * Or, we interrupted rcu_read_lock() right after it recorded + * a QS for the previous GP but before it got a chance to + * increment its nesting count. The memory barrier again + * stops the CS code from spilling out of the CS. + */ + memory_barrier(); + CPU->rcu.last_seen_gp = _rcu_cur_gp; + } + } else { + /* + * This cpu already acknowledged that it had passed through + * a quiescent state since the start of cur_gp. + */ + } + + /* + * smp_call() makes sure any changes propagate back to the caller. + * In particular, it makes the most current last_seen_gp visible + * to the detector. + */ +} + +/** Waits for cpus delaying the current grace period if there are any. */ +static bool wait_for_delaying_cpus(void) +{ + int delaying_cpu_cnt = atomic_get(&rcu.delaying_cpu_cnt); + + for (int i = 0; i < delaying_cpu_cnt; ++i){ + if (!semaphore_down_interruptable(&rcu.remaining_readers)) + return false; + } + + /* Update statistic. */ + rcu.stat_delayed_cnt += delaying_cpu_cnt; + + return true; +} + +/** Called by the scheduler() when switching away from the current thread. */ +void rcu_after_thread_ran(void) +{ + ASSERT(interrupts_disabled()); + + /* + * Prevent NMI handlers from interfering. The detector will be notified + * in this function if CPU->rcu.is_delaying_gp. The current thread is + * no longer running so there is nothing else to signal to the detector. + */ + CPU->rcu.signal_unlock = false; + /* + * Separates clearing of .signal_unlock from accesses to + * THREAD->rcu.was_preempted and CPU->rcu.nesting_cnt. + */ + compiler_barrier(); + + /* Save the thread's nesting count when it is not running. */ + THREAD->rcu.nesting_cnt = CPU->rcu.nesting_cnt; + + /* Preempted a reader critical section for the first time. */ + if (0 < THREAD->rcu.nesting_cnt && !THREAD->rcu.was_preempted) { + THREAD->rcu.was_preempted = true; + note_preempted_reader(); + } + + /* + * The preempted reader has been noted globally. There are therefore + * no readers running on this cpu so this is a quiescent state. + */ + _rcu_record_qs(); + + /* + * Interrupt handlers might use RCU while idle in scheduler(). + * The preempted reader has been noted globally, so the handlers + * may now start announcing quiescent states. + */ + CPU->rcu.nesting_cnt = 0; + + /* + * This cpu is holding up the current GP. Let the detector know + * it has just passed a quiescent state. + * + * The detector waits separately for preempted readers, so we have + * to notify the detector even if we have just preempted a reader. + */ + if (CPU->rcu.is_delaying_gp) { + CPU->rcu.is_delaying_gp = false; + semaphore_up(&rcu.remaining_readers); + } + + /* + * Forcefully associate the detector with the highest priority + * even if preempted due to its time slice running out. + * + * todo: Replace with strict scheduler priority classes. + */ + if (THREAD == rcu.detector_thr) { + THREAD->priority = -1; + } + else if (THREAD == CPU->rcu.reclaimer_thr) { + THREAD->priority = -1; + } + + upd_max_cbs_in_slice(CPU->rcu.arriving_cbs_cnt); +} + +/** Called by the scheduler() when switching to a newly scheduled thread. */ +void rcu_before_thread_runs(void) +{ + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + ASSERT(0 == CPU->rcu.nesting_cnt); + + /* Load the thread's saved nesting count from before it was preempted. */ + CPU->rcu.nesting_cnt = THREAD->rcu.nesting_cnt; + + /* + * Ensures NMI see the proper nesting count before .signal_unlock. + * Otherwise the NMI may incorrectly signal that a preempted reader + * exited its reader section. + */ + compiler_barrier(); + + /* + * In the unlikely event that a NMI occurs between the loading of the + * variables and setting signal_unlock, the NMI handler may invoke + * rcu_read_unlock() and clear signal_unlock. In that case we will + * incorrectly overwrite signal_unlock from false to true. This event + * is benign and the next rcu_read_unlock() will at worst + * needlessly invoke _rcu_signal_unlock(). + */ + CPU->rcu.signal_unlock = THREAD->rcu.was_preempted || CPU->rcu.is_delaying_gp; +} + +/** Called from scheduler() when exiting the current thread. + * + * Preemption or interrupts are disabled and the scheduler() already + * switched away from the current thread, calling rcu_after_thread_ran(). + */ +void rcu_thread_exiting(void) +{ + ASSERT(THREAD != NULL); + ASSERT(THREAD->state == Exiting); + ASSERT(PREEMPTION_DISABLED || interrupts_disabled()); + + /* + * The thread forgot to exit its reader critical section. + * It is a bug, but rather than letting the entire system lock up + * forcefully leave the reader section. The thread is not holding + * any references anyway since it is exiting so it is safe. + */ + if (0 < THREAD->rcu.nesting_cnt) { + THREAD->rcu.nesting_cnt = 1; + read_unlock_impl(&THREAD->rcu.nesting_cnt); + + printf("Bug: thread (id %" PRIu64 " \"%s\") exited while in RCU read" + " section.\n", THREAD->tid, THREAD->name); + } +} + + +#endif /* RCU_PREEMPT_PODZIMEK */ + +/** Announces the start of a new grace period for preexisting readers to ack. */ +static void start_new_gp(void) +{ + ASSERT(spinlock_locked(&rcu.gp_lock)); + + irq_spinlock_lock(&rcu.preempt_lock, true); + + /* Start a new GP. Announce to readers that a quiescent state is needed. */ + ++_rcu_cur_gp; + + /* + * Readers preempted before the start of this GP (next_preempted) + * are preexisting readers now that a GP started and will hold up + * the current GP until they exit their reader sections. + * + * Preempted readers from the previous GP have finished so + * cur_preempted is empty, but see comment in _rcu_record_qs(). + */ + list_concat(&rcu.cur_preempted, &rcu.next_preempted); + + irq_spinlock_unlock(&rcu.preempt_lock, true); +} + +/** Remove those cpus from the mask that have already passed a quiescent + * state since the start of the current grace period. + */ +static void rm_quiescent_cpus(cpu_mask_t *cpu_mask) +{ + /* + * Ensure the announcement of the start of a new GP (ie up-to-date + * cur_gp) propagates to cpus that are just coming out of idle + * mode before we sample their idle state flag. + * + * Cpus guarantee that after they set CPU->idle = true they will not + * execute any RCU reader sections without first setting idle to + * false and issuing a memory barrier. Therefore, if rm_quiescent_cpus() + * later on sees an idle cpu, but the cpu is just exiting its idle mode, + * the cpu must not have yet executed its memory barrier (otherwise + * it would pair up with this mem barrier and we would see idle == false). + * That memory barrier will pair up with the one below and ensure + * that a reader on the now-non-idle cpu will see the most current + * cur_gp. As a result, such a reader will never attempt to semaphore_up( + * pending_readers) during this GP, which allows the detector to + * ignore that cpu (the detector thinks it is idle). Moreover, any + * changes made by RCU updaters will have propagated to readers + * on the previously idle cpu -- again thanks to issuing a memory + * barrier after returning from idle mode. + * + * idle -> non-idle cpu | detector | reclaimer + * ------------------------------------------------------ + * rcu reader 1 | | rcu_call() + * MB X | | + * idle = true | | rcu_call() + * (no rcu readers allowed ) | | MB A in advance_cbs() + * MB Y | (...) | (...) + * (no rcu readers allowed) | | MB B in advance_cbs() + * idle = false | ++cur_gp | + * (no rcu readers allowed) | MB C | + * MB Z | signal gp_end | + * rcu reader 2 | | exec_cur_cbs() + * + * + * MB Y orders visibility of changes to idle for detector's sake. + * + * MB Z pairs up with MB C. The cpu making a transition from idle + * will see the most current value of cur_gp and will not attempt + * to notify the detector even if preempted during this GP. + * + * MB Z pairs up with MB A from the previous batch. Updaters' changes + * are visible to reader 2 even when the detector thinks the cpu is idle + * but it is not anymore. + * + * MB X pairs up with MB B. Late mem accesses of reader 1 are contained + * and visible before idling and before any callbacks are executed + * by reclaimers. + * + * In summary, the detector does not know of or wait for reader 2, but + * it does not have to since it is a new reader that will not access + * data from previous GPs and will see any changes. + */ + memory_barrier(); /* MB C */ + + cpu_mask_for_each(*cpu_mask, cpu_id) { + /* + * The cpu already checked for and passed through a quiescent + * state since the beginning of this GP. + * + * _rcu_cur_gp is modified by local detector thread only. + * Therefore, it is up-to-date even without a lock. + * + * cpu.last_seen_gp may not be up-to-date. At worst, we will + * unnecessarily sample its last_seen_gp with a smp_call. + */ + bool cpu_acked_gp = (cpus[cpu_id].rcu.last_seen_gp == _rcu_cur_gp); + + /* + * Either the cpu is idle or it is exiting away from idle mode + * and already sees the most current _rcu_cur_gp. See comment + * in wait_for_readers(). + */ + bool cpu_idle = cpus[cpu_id].idle; + + if (cpu_acked_gp || cpu_idle) { + cpu_mask_reset(cpu_mask, cpu_id); + } + } +} + +/** Serially invokes sample_local_cpu(arg) on each cpu of reader_cpus. */ +static void sample_cpus(cpu_mask_t *reader_cpus, void *arg) +{ + cpu_mask_for_each(*reader_cpus, cpu_id) { + smp_call(cpu_id, sample_local_cpu, arg); + + /* Update statistic. */ + if (CPU->id != cpu_id) + ++rcu.stat_smp_call_cnt; + } +} + +static void upd_missed_gp_in_wait(rcu_gp_t completed_gp) +{ + ASSERT(CPU->rcu.cur_cbs_gp <= completed_gp); + + size_t delta = (size_t)(completed_gp - CPU->rcu.cur_cbs_gp); + CPU->rcu.stat_missed_gp_in_wait += delta; +} + +/** Globally note that the current thread was preempted in a reader section. */ +static void note_preempted_reader(void) +{ + irq_spinlock_lock(&rcu.preempt_lock, false); + + if (CPU->rcu.last_seen_gp != _rcu_cur_gp) { + /* The reader started before the GP started - we must wait for it.*/ + list_append(&THREAD->rcu.preempt_link, &rcu.cur_preempted); + } else { + /* + * The reader started after the GP started and this cpu + * already noted a quiescent state. We might block the next GP. + */ + list_append(&THREAD->rcu.preempt_link, &rcu.next_preempted); + } + + irq_spinlock_unlock(&rcu.preempt_lock, false); +} + +/** Remove the current thread from the global list of preempted readers. */ +static void rm_preempted_reader(void) +{ + irq_spinlock_lock(&rcu.preempt_lock, true); + + ASSERT(link_used(&THREAD->rcu.preempt_link)); + + bool prev_empty = list_empty(&rcu.cur_preempted); + list_remove(&THREAD->rcu.preempt_link); + bool now_empty = list_empty(&rcu.cur_preempted); + + /* This was the last reader in cur_preempted. */ + bool last_removed = now_empty && !prev_empty; + + /* + * Preempted readers are blocking the detector and + * this was the last reader blocking the current GP. + */ + if (last_removed && rcu.preempt_blocking_det) { + rcu.preempt_blocking_det = false; + semaphore_up(&rcu.remaining_readers); + } + + irq_spinlock_unlock(&rcu.preempt_lock, true); +} + +/** Waits for any preempted readers blocking this grace period to finish.*/ +static bool wait_for_preempt_reader(void) +{ + irq_spinlock_lock(&rcu.preempt_lock, true); + + bool reader_exists = !list_empty(&rcu.cur_preempted); + rcu.preempt_blocking_det = reader_exists; + + irq_spinlock_unlock(&rcu.preempt_lock, true); + + if (reader_exists) { + /* Update statistic. */ + ++rcu.stat_preempt_blocking_cnt; + + return semaphore_down_interruptable(&rcu.remaining_readers); + } + + return true; +} + +static void upd_max_cbs_in_slice(size_t arriving_cbs_cnt) +{ + rcu_cpu_data_t *cr = &CPU->rcu; + + if (arriving_cbs_cnt > cr->last_arriving_cnt) { + size_t arrived_cnt = arriving_cbs_cnt - cr->last_arriving_cnt; + cr->stat_max_slice_cbs = max(arrived_cnt, cr->stat_max_slice_cbs); + } + + cr->last_arriving_cnt = arriving_cbs_cnt; +} + +/** Prints RCU run-time statistics. */ +void rcu_print_stat(void) +{ + /* + * Don't take locks. Worst case is we get out-dated values. + * CPU local values are updated without any locks, so there + * are no locks to lock in order to get up-to-date values. + */ + +#ifdef RCU_PREEMPT_PODZIMEK + const char *algo = "podzimek-preempt-rcu"; +#elif defined(RCU_PREEMPT_A) + const char *algo = "a-preempt-rcu"; +#endif + + printf("Config: expedite_threshold=%d, critical_threshold=%d," + " detect_sleep=%dms, %s\n", + EXPEDITE_THRESHOLD, CRITICAL_THRESHOLD, DETECT_SLEEP_MS, algo); + printf("Completed GPs: %" PRIu64 "\n", rcu.completed_gp); + printf("Expedited GPs: %zu\n", rcu.stat_expedited_cnt); + printf("Delayed GPs: %zu (cpus w/ still running readers after gp sleep)\n", + rcu.stat_delayed_cnt); + printf("Preempt blocked GPs: %zu (waited for preempted readers; " + "running or not)\n", rcu.stat_preempt_blocking_cnt); + printf("Smp calls: %zu\n", rcu.stat_smp_call_cnt); + + printf("Max arrived callbacks per GP and CPU:\n"); + for (unsigned int i = 0; i < config.cpu_count; ++i) { + printf(" %zu", cpus[i].rcu.stat_max_cbs); + } + + printf("\nAvg arrived callbacks per GP and CPU (nonempty batches only):\n"); + for (unsigned int i = 0; i < config.cpu_count; ++i) { + printf(" %zu", cpus[i].rcu.stat_avg_cbs); + } + + printf("\nMax arrived callbacks per time slice and CPU:\n"); + for (unsigned int i = 0; i < config.cpu_count; ++i) { + printf(" %zu", cpus[i].rcu.stat_max_slice_cbs); + } + + printf("\nMissed GP notifications per CPU:\n"); + for (unsigned int i = 0; i < config.cpu_count; ++i) { + printf(" %zu", cpus[i].rcu.stat_missed_gps); + } + + printf("\nMissed GP notifications per CPU while waking up:\n"); + for (unsigned int i = 0; i < config.cpu_count; ++i) { + printf(" %zu", cpus[i].rcu.stat_missed_gp_in_wait); + } + printf("\n"); +} + +/** @} + */ Index: kernel/generic/src/synch/smc.c =================================================================== --- kernel/generic/src/synch/smc.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/smc.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -41,4 +41,5 @@ #include #include +#include sysarg_t sys_smc_coherence(uintptr_t va, size_t size) Index: kernel/generic/src/synch/smp_memory_barrier.c =================================================================== --- kernel/generic/src/synch/smp_memory_barrier.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/synch/smp_memory_barrier.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,63 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ + +/** + * @file + * @brief Syscall implementation that issues a memory barrier on all cpus. + */ + +#include +#include +#include + + +static void issue_mem_bar(void *arg) +{ + /* smp_call already issues memory barriers on return from this function */ +} + +/** Issues a memory barrier on each cpu that is running a thread of the current + * task. + * + * @return Irrelevant. + */ +sysarg_t sys_smp_memory_barrier(void) +{ + for (unsigned int cpu_id = 0; cpu_id < config.cpu_active; ++cpu_id) { + smp_call(cpu_id, issue_mem_bar, NULL); + } + + return 0; +} + +/** @} + */ Index: kernel/generic/src/synch/spinlock.c =================================================================== --- kernel/generic/src/synch/spinlock.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/spinlock.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -45,4 +45,5 @@ #include #include +#include #ifdef CONFIG_SMP @@ -198,7 +199,6 @@ * * @param lock IRQ spinlock to be locked. - * @param irq_dis If true, interrupts are actually disabled - * prior locking the spinlock. If false, interrupts - * are expected to be already disabled. + * @param irq_dis If true, disables interrupts before locking the spinlock. + * If false, interrupts are expected to be already disabled. * */ Index: kernel/generic/src/synch/waitq.c =================================================================== --- kernel/generic/src/synch/waitq.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/synch/waitq.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -57,4 +57,6 @@ static void waitq_sleep_timed_out(void *); +static void waitq_complete_wakeup(waitq_t *); + /** Initialize wait queue @@ -330,4 +332,18 @@ break; default: + /* + * Wait for a waitq_wakeup() or waitq_unsleep() to complete + * before returning from waitq_sleep() to the caller. Otherwise + * the caller might expect that the wait queue is no longer used + * and deallocate it (although the wakeup on a another cpu has + * not yet completed and is using the wait queue). + * + * Note that we have to do this for ESYNCH_OK_BLOCKED and + * ESYNCH_INTERRUPTED, but not necessarily for ESYNCH_TIMEOUT + * where the timeout handler stops using the waitq before waking + * us up. To be on the safe side, ensure the waitq is not in use + * anymore in this case as well. + */ + waitq_complete_wakeup(wq); break; } @@ -357,5 +373,5 @@ } else { if (PARAM_NON_BLOCKING(flags, usec)) { - /* Return immediatelly instead of going to sleep */ + /* Return immediately instead of going to sleep */ return ESYNCH_WOULD_BLOCK; } @@ -442,4 +458,48 @@ irq_spinlock_unlock(&wq->lock, true); } + +/** If there is a wakeup in progress actively waits for it to complete. + * + * The function returns once the concurrently running waitq_wakeup() + * exits. It returns immediately if there are no concurrent wakeups + * at the time. + * + * Interrupts must be disabled. + * + * Example usage: + * @code + * void callback(waitq *wq) + * { + * // Do something and notify wait_for_completion() that we're done. + * waitq_wakeup(wq); + * } + * void wait_for_completion(void) + * { + * waitq wg; + * waitq_initialize(&wq); + * // Run callback() in the background, pass it wq. + * do_asynchronously(callback, &wq); + * // Wait for callback() to complete its work. + * waitq_sleep(&wq); + * // callback() completed its work, but it may still be accessing + * // wq in waitq_wakeup(). Therefore it is not yet safe to return + * // from waitq_sleep() or it would clobber up our stack (where wq + * // is stored). waitq_sleep() ensures the wait queue is no longer + * // in use by invoking waitq_complete_wakeup() internally. + * + * // waitq_sleep() returned, it is safe to free wq. + * } + * @endcode + * + * @param wq Pointer to a wait queue. + */ +static void waitq_complete_wakeup(waitq_t *wq) +{ + ASSERT(interrupts_disabled()); + + irq_spinlock_lock(&wq->lock, false); + irq_spinlock_unlock(&wq->lock, false); +} + /** Internal SMP- and IRQ-unsafe version of waitq_wakeup() Index: kernel/generic/src/synch/workqueue.c =================================================================== --- kernel/generic/src/synch/workqueue.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/generic/src/synch/workqueue.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,976 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +/** @addtogroup generic + * @{ + */ + +/** + * @file + * @brief Work queue/thread pool that automatically adjusts its size + * depending on the current load. Queued work functions may sleep.. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define WORKQ_MAGIC 0xf00c1333U +#define WORK_ITEM_MAGIC 0xfeec1777U + + +struct work_queue { + /* + * Protects everything except activate_worker. + * Must be acquired after any thread->locks. + */ + IRQ_SPINLOCK_DECLARE(lock); + + /* Activates a worker if new work arrives or if shutting down the queue. */ + condvar_t activate_worker; + + /* Queue of work_items ready to be dispatched. */ + list_t queue; + + /* List of worker threads. */ + list_t workers; + + /* Number of work items queued. */ + size_t item_cnt; + + /* Indicates the work queue is shutting down. */ + bool stopping; + const char *name; + + /* Total number of created worker threads. */ + size_t cur_worker_cnt; + /* Number of workers waiting for work to arrive. */ + size_t idle_worker_cnt; + /* Number of idle workers signaled that have not yet been woken up. */ + size_t activate_pending; + /* Number of blocked workers sleeping in work func() (ie not idle). */ + size_t blocked_worker_cnt; + + /* Number of pending signal_worker_op() operations. */ + size_t pending_op_cnt; + + link_t nb_link; + +#ifdef CONFIG_DEBUG + /* Magic cookie for integrity checks. Immutable. Accessed without lock. */ + uint32_t cookie; +#endif +}; + + +/** Min number of idle workers to keep. */ +static size_t min_worker_cnt; +/** Max total number of workers - be it blocked, idle, or active. */ +static size_t max_worker_cnt; +/** Max number of concurrently running active workers, ie not blocked nor idle. */ +static size_t max_concurrent_workers; +/** Max number of work items per active worker before a new worker is activated.*/ +static const size_t max_items_per_worker = 8; + +/** System wide work queue. */ +static struct work_queue g_work_queue; + +static int booting = true; + + +typedef struct { + IRQ_SPINLOCK_DECLARE(lock); + condvar_t req_cv; + thread_t *thread; + list_t work_queues; +} nonblock_adder_t; + +static nonblock_adder_t nonblock_adder; + + + +/** Typedef a worker thread signaling operation prototype. */ +typedef void (*signal_op_t)(struct work_queue *workq); + + +/* Fwd decl. */ +static void workq_preinit(struct work_queue *workq, const char *name); +static bool add_worker(struct work_queue *workq); +static void interrupt_workers(struct work_queue *workq); +static void wait_for_workers(struct work_queue *workq); +static int _workq_enqueue(struct work_queue *workq, work_t *work_item, + work_func_t func, bool can_block); +static void init_work_item(work_t *work_item, work_func_t func); +static signal_op_t signal_worker_logic(struct work_queue *workq, bool can_block); +static void worker_thread(void *arg); +static bool dequeue_work(struct work_queue *workq, work_t **pwork_item); +static bool worker_unnecessary(struct work_queue *workq); +static void cv_wait(struct work_queue *workq); +static void nonblock_init(void); +static bool workq_corrupted(struct work_queue *workq); +static bool work_item_corrupted(work_t *work_item); + + +/** Creates worker thread for the system-wide worker queue. */ +void workq_global_worker_init(void) +{ + /* + * No need for additional synchronization. Stores to word-sized + * variables are atomic and the change will eventually propagate. + * Moreover add_worker() includes the necessary memory barriers + * in spinlock lock/unlock(). + */ + booting = false; + + nonblock_init(); + + if (!add_worker(&g_work_queue)) + panic("Could not create a single global work queue worker!\n"); + +} + +/** Initializes the system wide work queue and support for other work queues. */ +void workq_global_init(void) +{ + /* Keep idle workers on 1/4-th of cpus, but at least 2 threads. */ + min_worker_cnt = max(2, config.cpu_count / 4); + /* Allow max 8 sleeping work items per cpu. */ + max_worker_cnt = max(32, 8 * config.cpu_count); + /* Maximum concurrency without slowing down the system. */ + max_concurrent_workers = max(2, config.cpu_count); + + workq_preinit(&g_work_queue, "kworkq"); +} + +/** Stops the system global work queue and waits for all work items to complete.*/ +void workq_global_stop(void) +{ + workq_stop(&g_work_queue); +} + +/** Creates and initializes a work queue. Returns NULL upon failure. */ +struct work_queue * workq_create(const char *name) +{ + struct work_queue *workq = malloc(sizeof(struct work_queue), 0); + + if (workq) { + if (workq_init(workq, name)) { + ASSERT(!workq_corrupted(workq)); + return workq; + } + + free(workq); + } + + return NULL; +} + +/** Frees work queue resources and stops it if it had not been done so already.*/ +void workq_destroy(struct work_queue *workq) +{ + ASSERT(!workq_corrupted(workq)); + + irq_spinlock_lock(&workq->lock, true); + bool stopped = workq->stopping; + size_t running_workers = workq->cur_worker_cnt; + irq_spinlock_unlock(&workq->lock, true); + + if (!stopped) { + workq_stop(workq); + } else { + ASSERT(0 == running_workers); + } + +#ifdef CONFIG_DEBUG + workq->cookie = 0; +#endif + + free(workq); +} + +/** Initializes workq structure without creating any workers. */ +static void workq_preinit(struct work_queue *workq, const char *name) +{ +#ifdef CONFIG_DEBUG + workq->cookie = WORKQ_MAGIC; +#endif + + irq_spinlock_initialize(&workq->lock, name); + condvar_initialize(&workq->activate_worker); + + list_initialize(&workq->queue); + list_initialize(&workq->workers); + + workq->item_cnt = 0; + workq->stopping = false; + workq->name = name; + + workq->cur_worker_cnt = 1; + workq->idle_worker_cnt = 0; + workq->activate_pending = 0; + workq->blocked_worker_cnt = 0; + + workq->pending_op_cnt = 0; + link_initialize(&workq->nb_link); +} + +/** Initializes a work queue. Returns true if successful. + * + * Before destroying a work queue it must be stopped via + * workq_stop(). + */ +int workq_init(struct work_queue *workq, const char *name) +{ + workq_preinit(workq, name); + return add_worker(workq); +} + +/** Add a new worker thread. Returns false if the thread could not be created. */ +static bool add_worker(struct work_queue *workq) +{ + ASSERT(!workq_corrupted(workq)); + + thread_t *thread = thread_create(worker_thread, workq, TASK, + THREAD_FLAG_NONE, workq->name); + + if (!thread) { + irq_spinlock_lock(&workq->lock, true); + + /* cur_worker_cnt proactively increased in signal_worker_logic() .*/ + ASSERT(0 < workq->cur_worker_cnt); + --workq->cur_worker_cnt; + + irq_spinlock_unlock(&workq->lock, true); + return false; + } + + /* Respect lock ordering. */ + irq_spinlock_lock(&thread->lock, true); + irq_spinlock_lock(&workq->lock, false); + + bool success; + + if (!workq->stopping) { + success = true; + + /* Try to distribute workers among cpus right away. */ + unsigned int cpu_id = (workq->cur_worker_cnt) % config.cpu_active; + + if (!cpus[cpu_id].active) + cpu_id = CPU->id; + + thread->workq = workq; + thread->cpu = &cpus[cpu_id]; + thread->workq_blocked = false; + thread->workq_idling = false; + link_initialize(&thread->workq_link); + + list_append(&thread->workq_link, &workq->workers); + } else { + /* + * Work queue is shutting down - we must not add the worker + * and we cannot destroy it without ready-ing it. Mark it + * interrupted so the worker exits right away without even + * touching workq. + */ + success = false; + + /* cur_worker_cnt proactively increased in signal_worker() .*/ + ASSERT(0 < workq->cur_worker_cnt); + --workq->cur_worker_cnt; + } + + irq_spinlock_unlock(&workq->lock, false); + irq_spinlock_unlock(&thread->lock, true); + + if (!success) { + thread_interrupt(thread); + } + + thread_ready(thread); + + return success; +} + +/** Shuts down the work queue. Waits for all pending work items to complete. + * + * workq_stop() may only be run once. + */ +void workq_stop(struct work_queue *workq) +{ + ASSERT(!workq_corrupted(workq)); + + interrupt_workers(workq); + wait_for_workers(workq); +} + +/** Notifies worker threads the work queue is shutting down. */ +static void interrupt_workers(struct work_queue *workq) +{ + irq_spinlock_lock(&workq->lock, true); + + /* workq_stop() may only be called once. */ + ASSERT(!workq->stopping); + workq->stopping = true; + + /* Respect lock ordering - do not hold workq->lock during broadcast. */ + irq_spinlock_unlock(&workq->lock, true); + + condvar_broadcast(&workq->activate_worker); +} + +/** Waits for all worker threads to exit. */ +static void wait_for_workers(struct work_queue *workq) +{ + ASSERT(!PREEMPTION_DISABLED); + + irq_spinlock_lock(&workq->lock, true); + + list_foreach_safe(workq->workers, cur_worker, next_worker) { + thread_t *worker = list_get_instance(cur_worker, thread_t, workq_link); + list_remove(cur_worker); + + /* Wait without the lock. */ + irq_spinlock_unlock(&workq->lock, true); + + thread_join(worker); + thread_detach(worker); + + irq_spinlock_lock(&workq->lock, true); + } + + ASSERT(list_empty(&workq->workers)); + + /* Wait for deferred add_worker_op(), signal_worker_op() to finish. */ + while (0 < workq->cur_worker_cnt || 0 < workq->pending_op_cnt) { + irq_spinlock_unlock(&workq->lock, true); + + scheduler(); + + irq_spinlock_lock(&workq->lock, true); + } + + irq_spinlock_unlock(&workq->lock, true); +} + +/** Queues a function into the global wait queue without blocking. + * + * See workq_enqueue_noblock() for more details. + */ +int workq_global_enqueue_noblock(work_t *work_item, work_func_t func) +{ + return workq_enqueue_noblock(&g_work_queue, work_item, func); +} + +/** Queues a function into the global wait queue; may block. + * + * See workq_enqueue() for more details. + */ +int workq_global_enqueue(work_t *work_item, work_func_t func) +{ + return workq_enqueue(&g_work_queue, work_item, func); +} + +/** Adds a function to be invoked in a separate thread without blocking. + * + * workq_enqueue_noblock() is guaranteed not to block. It is safe + * to invoke from interrupt handlers. + * + * Consider using workq_enqueue() instead if at all possible. Otherwise, + * your work item may have to wait for previously enqueued sleeping + * work items to complete if you are unlucky. + * + * @param workq Work queue where to queue the work item. + * @param work_item Work item bookkeeping structure. Must be valid + * until func() is entered. + * @param func User supplied function to invoke in a worker thread. + + * @return false if work queue is shutting down; function is not + * queued for further processing. + * @return true Otherwise. func() will be invoked in a separate thread. + */ +int workq_enqueue_noblock(struct work_queue *workq, work_t *work_item, + work_func_t func) +{ + return _workq_enqueue(workq, work_item, func, false); +} + +/** Adds a function to be invoked in a separate thread; may block. + * + * While the workq_enqueue() is unlikely to block, it may do so if too + * many previous work items blocked sleeping. + * + * @param workq Work queue where to queue the work item. + * @param work_item Work item bookkeeping structure. Must be valid + * until func() is entered. + * @param func User supplied function to invoke in a worker thread. + + * @return false if work queue is shutting down; function is not + * queued for further processing. + * @return true Otherwise. func() will be invoked in a separate thread. + */ +int workq_enqueue(struct work_queue *workq, work_t *work_item, work_func_t func) +{ + return _workq_enqueue(workq, work_item, func, true); +} + +/** Adds a work item that will be processed by a separate worker thread. + * + * func() will be invoked in another kernel thread and may block. + * + * Prefer to call _workq_enqueue() with can_block set. Otherwise + * your work item may have to wait for sleeping work items to complete. + * If all worker threads are blocked/sleeping a new worker thread cannot + * be create without can_block set because creating a thread might + * block due to low memory conditions. + * + * @param workq Work queue where to queue the work item. + * @param work_item Work item bookkeeping structure. Must be valid + * until func() is entered. + * @param func User supplied function to invoke in a worker thread. + * @param can_block May adding this work item block? + + * @return false if work queue is shutting down; function is not + * queued for further processing. + * @return true Otherwise. + */ +static int _workq_enqueue(struct work_queue *workq, work_t *work_item, + work_func_t func, bool can_block) +{ + ASSERT(!workq_corrupted(workq)); + + bool success = true; + signal_op_t signal_op = NULL; + + irq_spinlock_lock(&workq->lock, true); + + if (workq->stopping) { + success = false; + } else { + init_work_item(work_item, func); + list_append(&work_item->queue_link, &workq->queue); + ++workq->item_cnt; + success = true; + + if (!booting) { + signal_op = signal_worker_logic(workq, can_block); + } else { + /* + * During boot there are no workers to signal. Just queue + * the work and let future workers take care of it. + */ + } + } + + irq_spinlock_unlock(&workq->lock, true); + + if (signal_op) { + signal_op(workq); + } + + return success; +} + +/** Prepare an item to be added to the work item queue. */ +static void init_work_item(work_t *work_item, work_func_t func) +{ +#ifdef CONFIG_DEBUG + work_item->cookie = WORK_ITEM_MAGIC; +#endif + + link_initialize(&work_item->queue_link); + work_item->func = func; +} + +/** Returns the number of workers running work func() that are not blocked. */ +static size_t active_workers_now(struct work_queue *workq) +{ + ASSERT(irq_spinlock_locked(&workq->lock)); + + /* Workers blocked are sleeping in the work function (ie not idle). */ + ASSERT(workq->blocked_worker_cnt <= workq->cur_worker_cnt); + /* Idle workers are waiting for more work to arrive in condvar_wait. */ + ASSERT(workq->idle_worker_cnt <= workq->cur_worker_cnt); + + /* Idle + blocked workers == sleeping worker threads. */ + size_t sleeping_workers = workq->blocked_worker_cnt + workq->idle_worker_cnt; + + ASSERT(sleeping_workers <= workq->cur_worker_cnt); + /* Workers pending activation are idle workers not yet given a time slice. */ + ASSERT(workq->activate_pending <= workq->idle_worker_cnt); + + /* + * Workers actively running the work func() this very moment and + * are neither blocked nor idle. Exclude ->activate_pending workers + * since they will run their work func() once they get a time slice + * and are not running it right now. + */ + return workq->cur_worker_cnt - sleeping_workers; +} + +/** + * Returns the number of workers that are running or are about to run work + * func() and that are not blocked. + */ +static size_t active_workers(struct work_queue *workq) +{ + ASSERT(irq_spinlock_locked(&workq->lock)); + + /* + * Workers actively running the work func() and are neither blocked nor + * idle. ->activate_pending workers will run their work func() once they + * get a time slice after waking from a condvar wait, so count them + * as well. + */ + return active_workers_now(workq) + workq->activate_pending; +} + +static void add_worker_noblock_op(struct work_queue *workq) +{ + condvar_signal(&nonblock_adder.req_cv); +} + +static void add_worker_op(struct work_queue *workq) +{ + add_worker(workq); +} + +static void signal_worker_op(struct work_queue *workq) +{ + ASSERT(!workq_corrupted(workq)); + + condvar_signal(&workq->activate_worker); + + irq_spinlock_lock(&workq->lock, true); + ASSERT(0 < workq->pending_op_cnt); + --workq->pending_op_cnt; + irq_spinlock_unlock(&workq->lock, true); +} + +/** Determines how to signal workers if at all. + * + * @param workq Work queue where a new work item was queued. + * @param can_block True if we may block while signaling a worker or creating + * a new worker. + * + * @return Function that will notify workers or NULL if no action is needed. + */ +static signal_op_t signal_worker_logic(struct work_queue *workq, bool can_block) +{ + ASSERT(!workq_corrupted(workq)); + ASSERT(irq_spinlock_locked(&workq->lock)); + + /* Only signal workers if really necessary. */ + signal_op_t signal_op = NULL; + + /* + * Workers actively running the work func() and neither blocked nor idle. + * Including ->activate_pending workers that will run their work func() + * once they get a time slice. + */ + size_t active = active_workers(workq); + /* Max total allowed number of work items queued for active workers. */ + size_t max_load = active * max_items_per_worker; + + /* Active workers are getting overwhelmed - activate another. */ + if (max_load < workq->item_cnt) { + + size_t remaining_idle = + workq->idle_worker_cnt - workq->activate_pending; + + /* Idle workers still exist - activate one. */ + if (remaining_idle > 0) { + /* + * Directly changing idle_worker_cnt here would not allow + * workers to recognize spurious wake-ups. Change + * activate_pending instead. + */ + ++workq->activate_pending; + ++workq->pending_op_cnt; + signal_op = signal_worker_op; + } else { + /* No idle workers remain. Request that a new one be created. */ + bool need_worker = (active < max_concurrent_workers) + && (workq->cur_worker_cnt < max_worker_cnt); + + if (need_worker && can_block) { + signal_op = add_worker_op; + /* + * It may take some time to actually create the worker. + * We don't want to swamp the thread pool with superfluous + * worker creation requests so pretend it was already + * created and proactively increase the worker count. + */ + ++workq->cur_worker_cnt; + } + + /* + * We cannot create a new worker but we need one desperately + * because all workers are blocked in their work functions. + */ + if (need_worker && !can_block && 0 == active) { + ASSERT(0 == workq->idle_worker_cnt); + + irq_spinlock_lock(&nonblock_adder.lock, true); + + if (nonblock_adder.thread && !link_used(&workq->nb_link)) { + signal_op = add_worker_noblock_op; + ++workq->cur_worker_cnt; + list_append(&workq->nb_link, &nonblock_adder.work_queues); + } + + irq_spinlock_unlock(&nonblock_adder.lock, true); + } + } + } else { + /* + * There are enough active/running workers to process the queue. + * No need to signal/activate any new workers. + */ + signal_op = NULL; + } + + return signal_op; +} + +/** Executes queued work items. */ +static void worker_thread(void *arg) +{ + /* + * The thread has been created after the work queue was ordered to stop. + * Do not access the work queue and return immediately. + */ + if (thread_interrupted(THREAD)) { + thread_detach(THREAD); + return; + } + + ASSERT(arg != NULL); + + struct work_queue *workq = arg; + work_t *work_item; + + while (dequeue_work(workq, &work_item)) { + /* Copy the func field so func() can safely free work_item. */ + work_func_t func = work_item->func; + + func(work_item); + } +} + +/** Waits and retrieves a work item. Returns false if the worker should exit. */ +static bool dequeue_work(struct work_queue *workq, work_t **pwork_item) +{ + ASSERT(!workq_corrupted(workq)); + + irq_spinlock_lock(&workq->lock, true); + + /* Check if we should exit if load is low. */ + if (!workq->stopping && worker_unnecessary(workq)) { + /* There are too many workers for this load. Exit. */ + ASSERT(0 < workq->cur_worker_cnt); + --workq->cur_worker_cnt; + list_remove(&THREAD->workq_link); + irq_spinlock_unlock(&workq->lock, true); + + thread_detach(THREAD); + return false; + } + + bool stop = false; + + /* Wait for work to arrive. */ + while (list_empty(&workq->queue) && !workq->stopping) { + cv_wait(workq); + + if (0 < workq->activate_pending) + --workq->activate_pending; + } + + /* Process remaining work even if requested to stop. */ + if (!list_empty(&workq->queue)) { + link_t *work_link = list_first(&workq->queue); + *pwork_item = list_get_instance(work_link, work_t, queue_link); + +#ifdef CONFIG_DEBUG + ASSERT(!work_item_corrupted(*pwork_item)); + (*pwork_item)->cookie = 0; +#endif + list_remove(work_link); + --workq->item_cnt; + + stop = false; + } else { + /* Requested to stop and no more work queued. */ + ASSERT(workq->stopping); + --workq->cur_worker_cnt; + stop = true; + } + + irq_spinlock_unlock(&workq->lock, true); + + return !stop; +} + +/** Returns true if for the given load there are too many workers. */ +static bool worker_unnecessary(struct work_queue *workq) +{ + ASSERT(irq_spinlock_locked(&workq->lock)); + + /* No work is pending. We don't need too many idle threads. */ + if (list_empty(&workq->queue)) { + /* There are too many idle workers. Exit. */ + return (min_worker_cnt <= workq->idle_worker_cnt); + } else { + /* + * There is work but we are swamped with too many active workers + * that were woken up from sleep at around the same time. We + * don't need another worker fighting for cpu time. + */ + size_t active = active_workers_now(workq); + return (max_concurrent_workers < active); + } +} + +/** Waits for a signal to activate_worker. Thread marked idle while waiting. */ +static void cv_wait(struct work_queue *workq) +{ + ++workq->idle_worker_cnt; + THREAD->workq_idling = true; + + /* Ignore lock ordering just here. */ + ASSERT(irq_spinlock_locked(&workq->lock)); + + _condvar_wait_timeout_irq_spinlock(&workq->activate_worker, + &workq->lock, SYNCH_NO_TIMEOUT, SYNCH_FLAGS_NONE); + + ASSERT(!workq_corrupted(workq)); + ASSERT(irq_spinlock_locked(&workq->lock)); + + THREAD->workq_idling = false; + --workq->idle_worker_cnt; +} + + +/** Invoked from thread_ready() right before the thread is woken up. */ +void workq_before_thread_is_ready(thread_t *thread) +{ + ASSERT(thread); + ASSERT(irq_spinlock_locked(&thread->lock)); + + /* Worker's work func() is about to wake up from sleeping. */ + if (thread->workq && thread->workq_blocked) { + /* Must be blocked in user work func() and not be waiting for work. */ + ASSERT(!thread->workq_idling); + ASSERT(thread->state == Sleeping); + ASSERT(THREAD != thread); + ASSERT(!workq_corrupted(thread->workq)); + + /* Protected by thread->lock */ + thread->workq_blocked = false; + + irq_spinlock_lock(&thread->workq->lock, true); + --thread->workq->blocked_worker_cnt; + irq_spinlock_unlock(&thread->workq->lock, true); + } +} + +/** Invoked from scheduler() before switching away from a thread. */ +void workq_after_thread_ran(void) +{ + ASSERT(THREAD); + ASSERT(irq_spinlock_locked(&THREAD->lock)); + + /* Worker's work func() is about to sleep/block. */ + if (THREAD->workq && THREAD->state == Sleeping && !THREAD->workq_idling) { + ASSERT(!THREAD->workq_blocked); + ASSERT(!workq_corrupted(THREAD->workq)); + + THREAD->workq_blocked = true; + + irq_spinlock_lock(&THREAD->workq->lock, false); + + ++THREAD->workq->blocked_worker_cnt; + + bool can_block = false; + signal_op_t op = signal_worker_logic(THREAD->workq, can_block); + + irq_spinlock_unlock(&THREAD->workq->lock, false); + + if (op) { + ASSERT(add_worker_noblock_op == op || signal_worker_op == op); + op(THREAD->workq); + } + } +} + +/** Prints stats of the work queue to the kernel console. */ +void workq_print_info(struct work_queue *workq) +{ + irq_spinlock_lock(&workq->lock, true); + + size_t total = workq->cur_worker_cnt; + size_t blocked = workq->blocked_worker_cnt; + size_t idle = workq->idle_worker_cnt; + size_t active = active_workers(workq); + size_t items = workq->item_cnt; + bool stopping = workq->stopping; + bool worker_surplus = worker_unnecessary(workq); + const char *load_str = worker_surplus ? "decreasing" : + (0 < workq->activate_pending) ? "increasing" : "stable"; + + irq_spinlock_unlock(&workq->lock, true); + + printf( + "Configuration: max_worker_cnt=%zu, min_worker_cnt=%zu,\n" + " max_concurrent_workers=%zu, max_items_per_worker=%zu\n" + "Workers: %zu\n" + "Active: %zu (workers currently processing work)\n" + "Blocked: %zu (work functions sleeping/blocked)\n" + "Idle: %zu (idle workers waiting for more work)\n" + "Items: %zu (queued not yet dispatched work)\n" + "Stopping: %d\n" + "Load: %s\n", + max_worker_cnt, min_worker_cnt, + max_concurrent_workers, max_items_per_worker, + total, + active, + blocked, + idle, + items, + stopping, + load_str + ); +} + +/** Prints stats of the global work queue. */ +void workq_global_print_info(void) +{ + workq_print_info(&g_work_queue); +} + + +static bool dequeue_add_req(nonblock_adder_t *info, struct work_queue **pworkq) +{ + bool stop = false; + + irq_spinlock_lock(&info->lock, true); + + while (list_empty(&info->work_queues) && !stop) { + int ret = _condvar_wait_timeout_irq_spinlock(&info->req_cv, + &info->lock, SYNCH_NO_TIMEOUT, SYNCH_FLAGS_INTERRUPTIBLE); + + stop = (ret == ESYNCH_INTERRUPTED); + } + + if (!stop) { + *pworkq = list_get_instance(list_first(&info->work_queues), + struct work_queue, nb_link); + + ASSERT(!workq_corrupted(*pworkq)); + + list_remove(&(*pworkq)->nb_link); + } + + irq_spinlock_unlock(&info->lock, true); + + return !stop; +} + +static void thr_nonblock_add_worker(void *arg) +{ + nonblock_adder_t *info = arg; + struct work_queue *workq; + + while (dequeue_add_req(info, &workq)) { + add_worker(workq); + } +} + + +static void nonblock_init(void) +{ + irq_spinlock_initialize(&nonblock_adder.lock, "kworkq-nb.lock"); + condvar_initialize(&nonblock_adder.req_cv); + list_initialize(&nonblock_adder.work_queues); + + nonblock_adder.thread = thread_create(thr_nonblock_add_worker, + &nonblock_adder, TASK, THREAD_FLAG_NONE, "kworkq-nb"); + + if (nonblock_adder.thread) { + thread_ready(nonblock_adder.thread); + } else { + /* + * We won't be able to add workers without blocking if all workers + * sleep, but at least boot the system. + */ + printf("Failed to create kworkq-nb. Sleeping work may stall the workq.\n"); + } +} + +/** Returns true if the workq is definitely corrupted; false if not sure. + * + * Can be used outside of any locks. + */ +static bool workq_corrupted(struct work_queue *workq) +{ +#ifdef CONFIG_DEBUG + /* + * Needed to make the most current cookie value set by workq_preinit() + * visible even if we access the workq right after it is created but + * on a different cpu. Otherwise, workq_corrupted() would not work + * outside a lock. + */ + memory_barrier(); + return NULL == workq || workq->cookie != WORKQ_MAGIC; +#else + return false; +#endif +} + +/** Returns true if the work_item is definitely corrupted; false if not sure. + * + * Must be used with the work queue protecting spinlock locked. + */ +static bool work_item_corrupted(work_t *work_item) +{ +#ifdef CONFIG_DEBUG + return NULL == work_item || work_item->cookie != WORK_ITEM_MAGIC; +#else + return false; +#endif +} + +/** @} + */ Index: kernel/generic/src/syscall/syscall.c =================================================================== --- kernel/generic/src/syscall/syscall.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/syscall/syscall.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -50,4 +50,5 @@ #include #include +#include #include #include @@ -140,4 +141,6 @@ (syshandler_t) sys_futex_wakeup, (syshandler_t) sys_smc_coherence, + (syshandler_t) sys_smp_memory_barrier, + /* Address space related syscalls. */ Index: kernel/generic/src/time/clock.c =================================================================== --- kernel/generic/src/time/clock.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/time/clock.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -212,5 +212,5 @@ irq_spinlock_unlock(&THREAD->lock, false); - if ((!ticks) && (!PREEMPTION_DISABLED)) { + if (ticks == 0 && PREEMPTION_ENABLED) { scheduler(); #ifdef CONFIG_UDEBUG Index: kernel/generic/src/udebug/udebug.c =================================================================== --- kernel/generic/src/udebug/udebug.c (revision 7462674581270cbc4c5e0b2d1075ebff0d1aec55) +++ kernel/generic/src/udebug/udebug.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -44,4 +44,6 @@ #include #include +#include +#include /** Initialize udebug part of task structure. Index: kernel/test/cht/cht1.c =================================================================== --- kernel/test/cht/cht1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/test/cht/cht1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,573 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include +#include +#include +#include +#include + +typedef struct val { + /* Place at the top to simplify re-casting. */ + cht_link_t link; + size_t hash; + size_t unique_id; + bool deleted; + bool mark; +} val_t; + +static size_t val_hash(const cht_link_t *item) +{ + val_t *v = member_to_inst(item, val_t, link); + ASSERT(v->hash == (v->unique_id % 10)); + return v->hash; +} + +static size_t val_key_hash(void *key) +{ + return (uintptr_t)key % 10; +} + +static bool val_equal(const cht_link_t *item1, const cht_link_t *item2) +{ + val_t *v1 = member_to_inst(item1, val_t, link); + val_t *v2 = member_to_inst(item2, val_t, link); + return v1->unique_id == v2->unique_id; +} + +static bool val_key_equal(void *key, const cht_link_t *item2) +{ + val_t *v2 = member_to_inst(item2, val_t, link); + return (uintptr_t)key == v2->unique_id; +} + +static void val_rm_callback(cht_link_t *item) +{ + val_t *v = member_to_inst(item, val_t, link); + ASSERT(!v->deleted); + v->deleted = true; + free(v); +} + + +static cht_ops_t val_ops = { + .hash = val_hash, + .key_hash = val_key_hash, + .equal = val_equal, + .key_equal = val_key_equal, + .remove_callback = val_rm_callback, +}; + +static void set_val(val_t *v, size_t h, size_t uid) +{ + v->hash = h; + v->unique_id = uid; + v->deleted = false; + v->mark = false; +} + +/*-------------------------------------------------------------------*/ + + +static const char * do_sanity_test(cht_t *h) +{ + if (cht_find_lazy(h, (void*)0)) + return "Found lazy in empty table."; + + if (cht_find(h, (void*)0)) + return "Found in empty table."; + + if (cht_remove_key(h, (void*)0)) + return "Removed from empty table."; + + const int val_cnt = 6; + val_t *v[6] = { NULL }; + + for (int i = 0; i < val_cnt; ++i) + v[i] = malloc(sizeof(val_t), 0); + + size_t key[] = { 1, 1, 1, 11, 12, 13 }; + + /* First three are identical */ + for (int i = 0; i < 3; ++i) + set_val(v[i], 1, key[i]); + + /* Same hash, different key.*/ + set_val(v[3], 1, key[3]); + + /* Different hashes and keys. */ + set_val(v[4], 2, key[4]); + set_val(v[5], 3, key[5]); + + cht_link_t *dup; + + if (!cht_insert_unique(h, &v[0]->link, &dup)) + return "Duplicates in empty"; + + if (cht_insert_unique(h, &v[1]->link, &dup)) + return "Inserted a duplicate"; + + if (dup != &v[0]->link) + return "Returned wrong duplicate"; + + if (!cht_insert_unique(h, &v[3]->link, &dup)) + return "Refused non-equal item but with a hash in table."; + + cht_insert(h, &v[1]->link); + cht_insert(h, &v[2]->link); + + bool ok = true; + ok = ok && cht_insert_unique(h, &v[4]->link, &dup); + ok = ok && cht_insert_unique(h, &v[5]->link, &dup); + + if (!ok) + return "Refused unique ins 4, 5."; + + if (cht_find(h, (void*)0)) + return "Phantom find."; + + cht_link_t *item = cht_find(h, (void*)v[5]->unique_id); + if (!item || item != &v[5]->link) + return "Missing 5."; + + item = cht_find_next(h, &v[5]->link); + if (item) + return "Found nonexisting duplicate 5"; + + item = cht_find(h, (void*)v[3]->unique_id); + if (!item || item != &v[3]->link) + return "Missing 3."; + + item = cht_find_next(h, &v[3]->link); + if (item) + return "Found nonexisting duplicate 3, same hash as others."; + + item = cht_find(h, (void*)v[0]->unique_id); + ((val_t*)item)->mark = true; + + for (int k = 1; k < 3; ++k) { + item = cht_find_next(h, item); + if (!item) + return "Did not find an inserted duplicate"; + + val_t *val = ((val_t*)item); + + if (val->unique_id != v[0]->unique_id) + return "Found item with a different key."; + if (val->mark) + return "Found twice the same node."; + val->mark = true; + } + + for (int i = 0; i < 3; ++i) { + if (!v[i]->mark) + return "Did not find all duplicates"; + + v[i]->mark = false; + } + + if (cht_find_next(h, item)) + return "Found non-existing duplicate."; + + item = cht_find_next(h, cht_find(h, (void*)key[0])); + + ((val_t*)item)->mark = true; + if (!cht_remove_item(h, item)) + return "Failed to remove inserted item"; + + item = cht_find(h, (void*)key[0]); + if (!item || ((val_t*)item)->mark) + return "Did not find proper item."; + + item = cht_find_next(h, item); + if (!item || ((val_t*)item)->mark) + return "Did not find proper duplicate."; + + item = cht_find_next(h, item); + if (item) + return "Found removed duplicate"; + + if (2 != cht_remove_key(h, (void*)key[0])) + return "Failed to remove all duplicates"; + + if (cht_find(h, (void*)key[0])) + return "Found removed key"; + + if (!cht_find(h, (void*)key[3])) + return "Removed incorrect key"; + + for (size_t k = 0; k < sizeof(v) / sizeof(v[0]); ++k) { + cht_remove_key(h, (void*)key[k]); + } + + for (size_t k = 0; k < sizeof(v) / sizeof(v[0]); ++k) { + if (cht_find(h, (void*)key[k])) + return "Found a key in a cleared table"; + } + + return NULL; +} + +static const char * sanity_test(void) +{ + cht_t h; + if (!cht_create_simple(&h, &val_ops)) + return "Could not create the table."; + + rcu_read_lock(); + const char *err = do_sanity_test(&h); + rcu_read_unlock(); + + cht_destroy(&h); + + return err; +} + +/*-------------------------------------------------------------------*/ + +static size_t next_rand(size_t seed) +{ + return (seed * 1103515245 + 12345) & ((1U << 31) - 1); +} + +/*-------------------------------------------------------------------*/ +typedef struct { + cht_link_t link; + size_t key; + bool free; + bool inserted; + bool deleted; +} stress_t; + +typedef struct { + cht_t *h; + int *stop; + stress_t *elem; + size_t elem_cnt; + size_t upd_prob; + size_t wave_cnt; + size_t wave_elems; + size_t id; + bool failed; +} stress_work_t; + +static size_t stress_hash(const cht_link_t *item) +{ + return ((stress_t*)item)->key >> 8; +} +static size_t stress_key_hash(void *key) +{ + return ((size_t)key) >> 8; +} +static bool stress_equal(const cht_link_t *item1, const cht_link_t *item2) +{ + return ((stress_t*)item1)->key == ((stress_t*)item2)->key; +} +static bool stress_key_equal(void *key, const cht_link_t *item) +{ + return ((size_t)key) == ((stress_t*)item)->key; +} +static void stress_rm_callback(cht_link_t *item) +{ + if (((stress_t*)item)->free) + free(item); + else + ((stress_t*)item)->deleted = true; +} + +cht_ops_t stress_ops = { + .hash = stress_hash, + .key_hash = stress_key_hash, + .equal = stress_equal, + .key_equal = stress_key_equal, + .remove_callback = stress_rm_callback +}; + +static void resize_stresser(void *arg) +{ + stress_work_t *work = (stress_work_t *)arg; + + for (size_t k = 0; k < work->wave_cnt; ++k) { + TPRINTF("I{"); + for (size_t i = 0; i < work->wave_elems; ++i) { + stress_t *s = malloc(sizeof(stress_t), FRAME_ATOMIC); + if (!s) { + TPRINTF("[out-of-mem]\n"); + goto out_of_mem; + } + + s->free = true; + s->key = (i << 8) + work->id; + + cht_insert(work->h, &s->link); + } + TPRINTF("}"); + + thread_sleep(2); + + TPRINTF("R<"); + for (size_t i = 0; i < work->wave_elems; ++i) { + size_t key = (i << 8) + work->id; + + if (1 != cht_remove_key(work->h, (void*)key)) { + TPRINTF("Err: Failed to remove inserted item\n"); + goto failed; + } + } + TPRINTF(">"); + } + + /* Request that others stop. */ + *work->stop = 1; + return; + +failed: + work->failed = true; + +out_of_mem: + /* Request that others stop. */ + *work->stop = 1; + + /* Remove anything we may have inserted. */ + for (size_t i = 0; i < work->wave_elems; ++i) { + size_t key = (i << 8) + work->id; + cht_remove_key(work->h, (void*)key); + } +} + +static void op_stresser(void *arg) +{ + stress_work_t *work = (stress_work_t *)arg; + ASSERT(0 == *work->stop); + + size_t loops = 0; + size_t seed = work->id; + + while (0 == *work->stop && !work->failed) { + seed = next_rand(seed); + bool upd = ((seed % 100) <= work->upd_prob); + seed = next_rand(seed); + size_t elem_idx = seed % work->elem_cnt; + + ++loops; + if (0 == loops % (1024 * 1024)) { + /* Make the most current work->stop visible. */ + read_barrier(); + TPRINTF("*"); + } + + if (upd) { + seed = next_rand(seed); + bool item_op = seed & 1; + + if (work->elem[elem_idx].inserted) { + if (item_op) { + rcu_read_lock(); + cht_remove_item(work->h, &work->elem[elem_idx].link); + rcu_read_unlock(); + } else { + void *key = (void*)work->elem[elem_idx].key; + if (1 != cht_remove_key(work->h, key)) { + TPRINTF("Err: did not rm the key\n"); + work->failed = true; + } + } + work->elem[elem_idx].inserted = false; + } else if (work->elem[elem_idx].deleted) { + work->elem[elem_idx].deleted = false; + + if (item_op) { + rcu_read_lock(); + cht_link_t *dup; + if (!cht_insert_unique(work->h, &work->elem[elem_idx].link, + &dup)) { + TPRINTF("Err: already inserted\n"); + work->failed = true; + } + rcu_read_unlock(); + } else { + cht_insert(work->h, &work->elem[elem_idx].link); + } + + work->elem[elem_idx].inserted = true; + } + } else { + rcu_read_lock(); + cht_link_t *item = + cht_find(work->h, (void*)work->elem[elem_idx].key); + rcu_read_unlock(); + + if (item) { + if (!work->elem[elem_idx].inserted) { + TPRINTF("Err: found but not inserted!"); + work->failed = true; + } + if (item != &work->elem[elem_idx].link) { + TPRINTF("Err: found but incorrect item\n"); + work->failed = true; + } + } else { + if (work->elem[elem_idx].inserted) { + TPRINTF("Err: inserted but not found!"); + work->failed = true; + } + } + } + } + + + /* Remove anything we may have inserted. */ + for (size_t i = 0; i < work->elem_cnt; ++i) { + void *key = (void*) work->elem[i].key; + cht_remove_key(work->h, key); + } +} + +static bool do_stress(void) +{ + cht_t h; + + if (!cht_create_simple(&h, &stress_ops)) { + TPRINTF("Failed to create the table\n"); + return false; + } + + const size_t wave_cnt = 10; + const size_t max_thread_cnt = 8; + const size_t resize_thread_cnt = 2; + size_t op_thread_cnt = min(max_thread_cnt, 2 * config.cpu_active); + size_t total_thr_cnt = op_thread_cnt + resize_thread_cnt; + size_t items_per_thread = 1024; + + size_t work_cnt = op_thread_cnt + resize_thread_cnt; + size_t item_cnt = op_thread_cnt * items_per_thread; + + /* Alloc hash table items. */ + size_t size = item_cnt * sizeof(stress_t) + work_cnt * sizeof(stress_work_t) + + sizeof(int); + + TPRINTF("Alloc and init table items. \n"); + void *p = malloc(size, FRAME_ATOMIC); + if (!p) { + TPRINTF("Failed to alloc items\n"); + cht_destroy(&h); + return false; + } + + stress_t *pitem = p + work_cnt * sizeof(stress_work_t); + stress_work_t *pwork = p; + int *pstop = (int*)(pitem + item_cnt); + + *pstop = 0; + + /* Init work items. */ + for (size_t i = 0; i < op_thread_cnt; ++i) { + pwork[i].h = &h; + pwork[i].stop = pstop; + pwork[i].elem = &pitem[i * items_per_thread]; + pwork[i].upd_prob = (i + 1) * 100 / op_thread_cnt; + pwork[i].id = i; + pwork[i].elem_cnt = items_per_thread; + pwork[i].failed = false; + } + + for (size_t i = op_thread_cnt; i < op_thread_cnt + resize_thread_cnt; ++i) { + pwork[i].h = &h; + pwork[i].stop = pstop; + pwork[i].wave_cnt = wave_cnt; + pwork[i].wave_elems = item_cnt * 4; + pwork[i].id = i; + pwork[i].failed = false; + } + + /* Init table elements. */ + for (size_t k = 0; k < op_thread_cnt; ++k) { + for (size_t i = 0; i < items_per_thread; ++i) { + pwork[k].elem[i].key = (i << 8) + k; + pwork[k].elem[i].free = false; + pwork[k].elem[i].inserted = false; + pwork[k].elem[i].deleted = true; + } + } + + TPRINTF("Running %zu ins/del/find stress threads + %zu resizers.\n", + op_thread_cnt, resize_thread_cnt); + + /* Create and run threads. */ + thread_t *thr[max_thread_cnt + resize_thread_cnt]; + + for (size_t i = 0; i < total_thr_cnt; ++i) { + if (i < op_thread_cnt) + thr[i] = thread_create(op_stresser, &pwork[i], TASK, 0, "cht-op-stress"); + else + thr[i] = thread_create(resize_stresser, &pwork[i], TASK, 0, "cht-resize"); + + ASSERT(thr[i]); + thread_wire(thr[i], &cpus[i % config.cpu_active]); + thread_ready(thr[i]); + } + + bool failed = false; + + /* Wait for all threads to return. */ + TPRINTF("Joining resize stressers.\n"); + for (size_t i = op_thread_cnt; i < total_thr_cnt; ++i) { + thread_join(thr[i]); + thread_detach(thr[i]); + failed = pwork[i].failed || failed; + } + + TPRINTF("Joining op stressers.\n"); + for (int i = (int)op_thread_cnt - 1; i >= 0; --i) { + TPRINTF("%d threads remain\n", i); + thread_join(thr[i]); + thread_detach(thr[i]); + failed = pwork[i].failed || failed; + } + + cht_destroy(&h); + free(p); + + return !failed; +} + +/*-------------------------------------------------------------------*/ + + +const char *test_cht1(void) +{ + const char *err = sanity_test(); + if (err) + return err; + printf("Basic sanity test: ok.\n"); + + if (!do_stress()) + return "CHT stress test failed."; + else + return NULL; +} Index: kernel/test/cht/cht1.def =================================================================== --- kernel/test/cht/cht1.def (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/test/cht/cht1.def (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,6 @@ +{ + "cht", + "Concurrent hash table test", + &test_cht1, + true +}, Index: kernel/test/smpcall/smpcall1.c =================================================================== --- kernel/test/smpcall/smpcall1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/test/smpcall/smpcall1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,158 @@ +/* + */ + +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +/* + * Maximum total number of smp_calls in the system is: + * 162000 == 9^2 * 1000 * 2 + * == MAX_CPUS^2 * ITERATIONS * EACH_CPU_INC_PER_ITER + */ +#define MAX_CPUS 9 +#define ITERATIONS 1000 +#define EACH_CPU_INC_PER_ITER 2 + + +static void inc(void *p) +{ + ASSERT(interrupts_disabled()); + + size_t *pcall_cnt = (size_t*)p; + /* + * No synchronization. Tests if smp_calls makes changes + * visible to the caller. + */ + ++*pcall_cnt; +} + + +static void test_thread(void *p) +{ + size_t *pcall_cnt = (size_t*)p; + smp_call_t call_info[MAX_CPUS]; + + unsigned int cpu_count = min(config.cpu_active, MAX_CPUS); + + for (int iter = 0; iter < ITERATIONS; ++iter) { + /* Synchronous version. */ + for (unsigned cpu_id = 0; cpu_id < cpu_count; ++cpu_id) { + /* + * smp_call should make changes by inc() visible on this cpu. + * As a result we can pass it our pcall_cnt and not worry + * about other synchronization. + */ + smp_call(cpu_id, inc, pcall_cnt); + } + + /* + * Async calls run in parallel on different cpus, so passing the + * same counter would clobber it without additional synchronization. + */ + size_t local_cnt[MAX_CPUS] = {0}; + + /* Now start asynchronous calls. */ + for (unsigned cpu_id = 0; cpu_id < cpu_count; ++cpu_id) { + smp_call_async(cpu_id, inc, &local_cnt[cpu_id], &call_info[cpu_id]); + } + + /* And wait for all async calls to complete. */ + for (unsigned cpu_id = 0; cpu_id < cpu_count; ++cpu_id) { + smp_call_wait(&call_info[cpu_id]); + *pcall_cnt += local_cnt[cpu_id]; + } + + /* Give other threads a chance to run. */ + thread_usleep(10000); + } +} + +static size_t calc_exp_calls(size_t thread_cnt) +{ + return thread_cnt * ITERATIONS * EACH_CPU_INC_PER_ITER; +} + +const char *test_smpcall1(void) +{ + /* Number of received calls that were sent by cpu[i]. */ + size_t call_cnt[MAX_CPUS] = {0}; + thread_t *thread[MAX_CPUS] = { NULL }; + + unsigned int cpu_count = min(config.cpu_active, MAX_CPUS); + size_t running_thread_cnt = 0; + + TPRINTF("Spawning threads on %u cpus.\n", cpu_count); + + /* Create a wired thread on each cpu. */ + for (unsigned int id = 0; id < cpu_count; ++id) { + thread[id] = thread_create(test_thread, &call_cnt[id], TASK, + THREAD_FLAG_NONE, "smp-call-test"); + + if (thread[id]) { + thread_wire(thread[id], &cpus[id]); + ++running_thread_cnt; + } else { + TPRINTF("Failed to create thread on cpu%u.\n", id); + } + } + + size_t exp_calls = calc_exp_calls(running_thread_cnt); + size_t exp_calls_sum = exp_calls * cpu_count; + + TPRINTF("Running %zu wired threads. Expecting %zu calls. Be patient.\n", + running_thread_cnt, exp_calls_sum); + + for (unsigned int i = 0; i < cpu_count; ++i) { + if (thread[i] != NULL) { + thread_ready(thread[i]); + } + } + + /* Wait for threads to complete. */ + for (unsigned int i = 0; i < cpu_count; ++i) { + if (thread[i] != NULL) { + thread_join(thread[i]); + thread_detach(thread[i]); + } + } + + TPRINTF("Threads finished. Checking number of smp_call()s.\n"); + + bool ok = true; + size_t calls_sum = 0; + + for (size_t i = 0; i < cpu_count; ++i) { + if (thread[i] != NULL) { + if (call_cnt[i] != exp_calls) { + ok = false; + TPRINTF("Error: %zu instead of %zu cpu%zu's calls were" + " acknowledged.\n", call_cnt[i], exp_calls, i); + } + } + + calls_sum += call_cnt[i]; + } + + if (calls_sum != exp_calls_sum) { + TPRINTF("Error: total acknowledged sum: %zu instead of %zu.\n", + calls_sum, exp_calls_sum); + + ok = false; + } + + if (ok) { + TPRINTF("Success: number of received smp_calls is as expected (%zu).\n", + exp_calls_sum); + return NULL; + } else + return "Failed: incorrect acknowledged smp_calls.\n"; + +} Index: kernel/test/smpcall/smpcall1.def =================================================================== --- kernel/test/smpcall/smpcall1.def (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/test/smpcall/smpcall1.def (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,6 @@ +{ + "smpcall1", + "smp_call() test", + &test_smpcall1, + true +}, Index: kernel/test/synch/rcu1.c =================================================================== --- kernel/test/synch/rcu1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) +++ kernel/test/synch/rcu1.c (revision 69146b93ef5ca7beba675e46730ac8d6d5a24316) @@ -0,0 +1,1052 @@ +/* + * Copyright (c) 2012 Adam Hraska + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * - The name of the author may not be used to endorse or promote products + * derived from this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR + * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES + * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. + * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, + * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT + * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include