source: mainline/kernel/arch/sparc64/src/smp/sun4v/smp.c@ 95c4776

/*
 * Copyright (c) 2006 Jakub Jermar
 * Copyright (c) 2009 Pavel Rimsky 
 * 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
 */

#include <smp/smp.h>
#include <smp/ipi.h>
#include <genarch/ofw/ofw_tree.h>
#include <cpu.h>
#include <arch/cpu.h>
#include <arch/boot/boot.h>
#include <arch.h>
#include <config.h>
#include <macros.h>
#include <func.h>
#include <arch/types.h>
#include <synch/synch.h>
#include <synch/waitq.h>
#include <print.h>
#include <arch/sun4v/hypercall.h>
#include <arch/sun4v/md.h>
#include <arch/sun4v/ipi.h>
#include <time/delay.h>
#include <arch/smp/sun4v/smp.h>
#include <string.h>

/** hypervisor code of the "running" state of the CPU */
#define CPU_STATE_RUNNING       2

/** maximum possible number of processor cores */
#define MAX_NUM_CORES           8

/** needed in the CPU_START hypercall */
extern void kernel_image_start(void);

/** needed in the CPU_START hypercall */
extern void *trap_table;

/** number of execution units detected */
uint8_t exec_unit_count = 0;

/** execution units (processor cores) */
exec_unit_t exec_units[MAX_NUM_CORES];

/** CPU structures */
extern cpu_t *cpus;

/** maximum number of strands per a physical core detected */
unsigned int max_core_strands = 0;

#ifdef CONFIG_SIMICS_SMP_HACK
/**
 * Copies a piece of HelenOS code to the place where OBP had its IPI handler.
 * By sending an IPI by the BSP to the AP the code will be executed.
 * The code will jump to the first instruction of the kernel. This is
 * a workaround how to make APs execute HelenOS code on Simics.
 */
static void simics_smp_hack_init(void) {
        asm volatile (
                "setx temp_cpu_mondo_handler, %g4, %g6 \n"
                "setx 0x80200f80, %g4, %g7 \n"

                "ldx [%g6], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x8], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x10], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x18], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x20], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x28], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x30], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x38], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "add %g7, 0x8, %g7 \n"
                "ldx [%g6 + 0x40], %g4 \n"
                "stxa %g4, [%g7] 0x14 \n"
                "membar #Sync \n"

                "flush %i7"

                );
}
#endif

#if 0
/**
 * Proposes the optimal number of ready threads for each virtual processor
 * in the given processor core so that the processor core is as busy as the
 * average processor core. The proposed number of ready threads will be
 * stored to the proposed_nrdy variable of the cpu_arch_t struture.
 */
bool calculate_optimal_nrdy(exec_unit_t *exec_unit) {

        /* calculate the number of threads the core will steal */
        int avg = atomic_get(&nrdy) / exec_unit_count;
        int to_steal = avg - atomic_get(&(exec_units->nrdy));
        if (to_steal < 0) {
                return true;
        } else if (to_steal == 0) {
                return false;
        }

        /* initialize the proposals with the real numbers of ready threads */
        unsigned int k;
        for (k = 0; k < exec_unit->strand_count; k++) {
                exec_units->cpus[k]->arch.proposed_nrdy =
                        atomic_get(&(exec_unit->cpus[k]->nrdy));
        }

        /* distribute the threads to be stolen to the core's CPUs */
        int j;
        for (j = to_steal; j > 0; j--) {
                unsigned int k;
                unsigned int least_busy = 0;
                unsigned int least_busy_nrdy =
                        exec_unit->cpus[0]->arch.proposed_nrdy;

                /* for each stolen thread, give it to the least busy CPU */
                for (k = 0; k < exec_unit->strand_count; k++) {
                        if (exec_unit->cpus[k]->arch.proposed_nrdy
                                        < least_busy_nrdy) {
                                least_busy = k;
                                least_busy_nrdy =
                                        exec_unit->cpus[k]->arch.proposed_nrdy;
                        }
                }
                exec_unit->cpus[least_busy]->arch.proposed_nrdy++;
        }

        return false;
}
#endif

/**
 * Finds out which execution units belong to particular CPUs. By execution unit
 * we mean the physical core the logical processor is backed by. Since each
 * Niagara physical core has just one integer execution unit and we will
 * ignore other execution units than the integer ones, we will use the terms
 * "integer execution unit", "execution unit" and "physical core"
 * interchangeably.
 *
 * The physical cores are detected by browsing the children of the CPU node
 * in the machine description and looking for a node representing an integer
 * execution unit. Once the integer execution unit of a particular CPU is
 * known, the ID of the CPU is added to the list of cpuids of the corresponding
 * execution unit structure (exec_unit_t). If an execution unit is encountered
 * for the first time, a new execution unit structure (exec_unit_t) must be
 * created first and added to the execution units array (exec_units).
 *
 * If the function fails to find an execution unit for a CPU (this may happen
 * on machines with older firmware or on Simics), it performs a fallback code
 * which pretends there exists just one execution unit and all CPUs belong to
 * it.
 *
 * Finally, the array of all execution units is reordered such that its element
 * which represents the physical core of the the bootstrap CPU is at index 0.
 * Moreover, the array of CPU IDs within the BSP's physical core structure is
 * reordered such that the element which represents the ID of the BSP is at
 * index 0. This is done because we would like the CPUs to be woken up
 * such that the 0-index CPU of the 0-index execution unit is
 * woken up first. And since the BSP is already woken up, we would like it to be
 * at 0-th position of the 0-th execution unit structure.
 *
 * Apart from that, the code also counts the total number of CPUs and stores
 * it to the global config.cpu_count variable.
 */
static void detect_execution_units(void)
{
        /* ID of the bootstrap processor */
        uint64_t myid;

        /* total number of CPUs detected */
        size_t cpu_count = 0;

        /* will be set to 1 if detecting the physical cores fails */
        bool exec_unit_assign_error = 0;

        /* index of the bootstrap physical core in the array of cores */
        unsigned int bsp_exec_unit_index = 0;

        /* index of the BSP ID inside the array of bootstrap core's cpuids */
        unsigned int bsp_core_strand_index = 0;

        __hypercall_fast_ret1(0, 0, 0, 0, 0, CPU_MYID, &myid);
        md_node_t node = md_get_root();

        /* walk through all the CPU nodes in the MD*/
        while (md_next_node(&node, "cpu")) {

                uint64_t cpuid;
                md_get_integer_property(node, "id", &cpuid);
                cpu_count++;

                /*
                 * if failed in previous CPUs, don't try
                 * to detect physical cores any more
                 */
                if (exec_unit_assign_error)
                        continue;

                /* detect exec. unit for the CPU represented by current node */
                uint64_t exec_unit_id = 0;
                md_child_iter_t it = md_get_child_iterator(node);

                while (md_next_child(&it)) {
                        md_node_t child = md_get_child_node(it);
                        const char *exec_unit_type;
                        md_get_string_property(child, "type", &exec_unit_type);

                        /* each physical core has just 1 integer exec. unit */
                        if (str_cmp(exec_unit_type, "integer") == 0) {
                                exec_unit_id = child;
                                break;
                        }
                }

                /* execution unit detected successfully */
                if (exec_unit_id != 0) {

                        /* find the exec. unit in array of existing units */
                        unsigned int i = 0;
                        for (i = 0; i < exec_unit_count; i++) {
                                if (exec_units[i].exec_unit_id == exec_unit_id)
                                        break;
                        }

                        /*
                         * execution unit just met has not been met before, so
                         * create a new entry in array of all execution units
                         */
                        if (i == exec_unit_count) {
                                exec_units[i].exec_unit_id = exec_unit_id;
                                exec_units[i].strand_count = 0;
                                atomic_set(&(exec_units[i].nrdy), 0);
                                spinlock_initialize(&(exec_units[i].proposed_nrdy_lock), "proposed nrdy lock");
                                exec_unit_count++;
                        }

                        /*
                         * remember the exec. unit and strand of the BSP
                         */
                        if (cpuid == myid) {
                                bsp_exec_unit_index = i;
                                bsp_core_strand_index = exec_units[i].strand_count;
                        }

                        /* add the CPU just met to the exec. unit's list */
                        exec_units[i].cpuids[exec_units[i].strand_count] = cpuid;
                        exec_units[i].strand_count++;
                        max_core_strands =
                                exec_units[i].strand_count > max_core_strands ?
                                exec_units[i].strand_count : max_core_strands;

                /* detecting execution unit failed */
                } else {
                        exec_unit_assign_error = 1;
                }
        }               

        /* save the number of CPUs to a globally accessible variable */
        config.cpu_count = cpu_count;

        /*
         * A fallback code which will be executed if finding out which
         * execution units belong to particular CPUs fails. Pretend there
         * exists just one execution unit and all CPUs belong to it.
         */
        if (exec_unit_assign_error) {
                bsp_exec_unit_index = 0;
                exec_unit_count = 1;
                exec_units[0].strand_count = cpu_count;
                exec_units[0].exec_unit_id = 1;
                spinlock_initialize(&(exec_units[0].proposed_nrdy_lock), "proposed nrdy lock");
                atomic_set(&(exec_units[0].nrdy), 0);
                max_core_strands = cpu_count;

                /* browse CPUs again, assign them the fictional exec. unit */
                node = md_get_root();
                unsigned int i = 0;

                while (md_next_node(&node, "cpu")) {
                        uint64_t cpuid;
                        md_get_integer_property(node, "id", &cpuid);
                        if (cpuid == myid) {
                                bsp_core_strand_index = i;
                        }
                        exec_units[0].cpuids[i++] = cpuid;
                }
        }

        /*
         * Reorder the execution units array elements and the cpuid array
         * elements so that the BSP will always be the very first CPU of
         * the very first execution unit.
         */
        exec_unit_t temp_exec_unit = exec_units[0];
        exec_units[0] = exec_units[bsp_exec_unit_index];
        exec_units[bsp_exec_unit_index] = temp_exec_unit;

        uint64_t temp_cpuid = exec_units[0].cpuids[0];
        exec_units[0].cpuids[0] = exec_units[0].cpuids[bsp_exec_unit_index];
        exec_units[0].cpuids[bsp_core_strand_index] = temp_cpuid;

}

/**
 * Determine number of processors and detect physical cores. On Simics
 * copy the code which will be executed by the AP when the BSP sends an
 * IPI to it in order to make it execute HelenOS code.
 */
void smp_init(void)
{
        detect_execution_units();
#ifdef CONFIG_SIMICS_SMP_HACK
        simics_smp_hack_init();
#endif
}

/**
 * For each CPU sets the value of cpus[i].arch.id, where i is the
 * index of the CPU in the cpus variable, to the cpuid of the i-th processor
 * to be run. The CPUs are run such that the CPU represented by cpus[0]
 * is run first, cpus[1] is run after it, and cpus[cpu_count - 1] is run as the
 * last one.
 *
 * The CPU IDs are set such that during waking the CPUs up the
 * processor cores will be alternated, i.e. first one CPU from the first core
 * will be run, after that one CPU from the second CPU core will be run,...
 * then one CPU from the last core will be run, after that another CPU
 * from the first core will be run, then another CPU from the second core
 * will be run,... then another CPU from the last core will be run, and so on.
 */
static void init_cpuids(void)
{
        unsigned int cur_core_strand;
        unsigned int cur_core;
        unsigned int cur_cpu = 0;

        for (cur_core_strand = 0; cur_core_strand < max_core_strands; cur_core_strand++) {
                for (cur_core = 0; cur_core < exec_unit_count; cur_core++) {
                        if (cur_core_strand > exec_units[cur_core].strand_count)
                                continue;

                        cpus[cur_cpu].arch.exec_unit = &(exec_units[cur_core]);
                        atomic_add(&(exec_units[cur_core].nrdy), atomic_get(&(cpus[cur_cpu].nrdy)));
                        cpus[cur_cpu].arch.id = exec_units[cur_core].cpuids[cur_core_strand];
                        exec_units[cur_core].cpus[cur_core_strand] = &(cpus[cur_cpu]);
                        cur_cpu++;
                }
        }
}

/**
 * Wakes up a single CPU.
 *
 * @param cpuid ID of the CPU to be woken up
 */
static bool wake_cpu(uint64_t cpuid)
{

#ifdef CONFIG_SIMICS_SMP_HACK
        ipi_unicast_to((void (*)(void)) 1234, cpuid);
#else
        /* stop the CPU before making it execute our code */
        if (__hypercall_fast1(CPU_STOP, cpuid) != EOK)
                return false;

        /* wait for the CPU to stop */
        uint64_t state;
        __hypercall_fast_ret1(cpuid, 0, 0, 0, 0,
                CPU_STATE, &state);
        while (state == CPU_STATE_RUNNING) {
                __hypercall_fast_ret1(cpuid, 0, 0, 0, 0,
                        CPU_STATE, &state);
        }

        /* make the CPU run again and execute HelenOS code */
        if (__hypercall_fast4(
                CPU_START, cpuid,
                (uint64_t) KA2PA(kernel_image_start),
                KA2PA(trap_table), bootinfo.physmem_start                       
                ) != EOK)
                        return false;
#endif

        if (waitq_sleep_timeout(&ap_completion_wq, 10000000, SYNCH_FLAGS_NONE) ==
                        ESYNCH_TIMEOUT)
                printf("%s: waiting for processor (cpuid = %" PRIu32
                ") timed out\n", __func__, cpuid);

        return true;
}

/** Wake application processors up. */
void kmp(void *arg)
{
        init_cpuids();

        unsigned int i;

        for (i = 1; i < config.cpu_count; i++) {
                wake_cpu(cpus[i].arch.id);
        }
}

/** @}
 */