source: mainline/kernel/generic/src/synch/rcu.c@ 9306cd7

/*
 * 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 <assert.h>
#include <synch/rcu.h>
#include <synch/condvar.h>
#include <synch/semaphore.h>
#include <synch/spinlock.h>
#include <synch/mutex.h>
#include <proc/thread.h>
#include <cpu/cpu_mask.h>
#include <cpu.h>
#include <smp/smp_call.h>
#include <compiler/barrier.h>
#include <atomic.h>
#include <arch.h>
#include <macros.h>

/* 
 * 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");
}

/** @}
 */