source: mainline/kernel/test/synch/rcu1.c@ 15f8079

/*
 * 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 <assert.h>
#include <test.h>
#include <arch.h>
#include <atomic.h>
#include <print.h>
#include <proc/thread.h>
#include <macros.h>
#include <str.h>
#include <errno.h>
#include <time/delay.h>

#include <synch/rcu.h>


#define MAX_THREADS 32

static int one_idx = 0;
static thread_t *thread[MAX_THREADS] = { NULL };

typedef struct {
        rcu_item_t rcu;
        bool exited;
} exited_t;

/* Callback raced with preexisting readers. */
#define ERACE   123
/* Waited for too long for the callback to exit; consider it lost. */
#define ECBLOST 432

/*-------------------------------------------------------------------*/
static void wait_for_cb_exit(size_t secs, exited_t *p, int *presult)
{
        size_t loops = 0;
        /* 4 secs max */
        size_t loop_ms_sec = 500;
        size_t max_loops = ((secs * 1000 + loop_ms_sec - 1) / loop_ms_sec);

        while (loops < max_loops && !p->exited) {
                ++loops;
                thread_usleep(loop_ms_sec * 1000);
                TPRINTF(".");
        }
        
        if (!p->exited) {
                *presult = ECBLOST;
        }
}

static size_t get_thread_cnt(void)
{
        return min(MAX_THREADS, config.cpu_active * 4);
}

static void run_thread(size_t k, void (*func)(void*), void *arg)
{
        assert(thread[k] == NULL);
        
        thread[k] = thread_create(func, arg, TASK, THREAD_FLAG_NONE, 
                "test-rcu-thread");
                
        if(thread[k]) {
                /* Distribute evenly. */
                thread_wire(thread[k], &cpus[k % config.cpu_active]);
                thread_ready(thread[k]);
        }
}

static void run_all(void (*func)(void*))
{
        size_t thread_cnt = get_thread_cnt();
        
        one_idx = 0;
        
        for (size_t i = 0; i < thread_cnt; ++i) {
                run_thread(i, func, NULL);
        }
}

static void join_all(void)
{
        size_t thread_cnt = get_thread_cnt();
        
        one_idx = 0;
        
        for (size_t i = 0; i < thread_cnt; ++i) {
                if (thread[i]) {
                        bool joined = false;
                        do {
                                int ret = thread_join_timeout(thread[i], 5 * 1000 * 1000, 0);
                                joined = (ret != ESYNCH_TIMEOUT);
                                
                                if (ret == ESYNCH_OK_BLOCKED) {
                                        TPRINTF("%zu threads remain\n", thread_cnt - i - 1);
                                }
                        } while (!joined);
                        
                        thread_detach(thread[i]);
                        thread[i] = NULL;
                }
        }
}

static void run_one(void (*func)(void*), void *arg)
{
        assert(one_idx < MAX_THREADS);
        run_thread(one_idx, func, arg);
        ++one_idx;
}


static void join_one(void)
{
        assert(0 < one_idx && one_idx <= MAX_THREADS);

        --one_idx;
        
        if (thread[one_idx]) {
                thread_join(thread[one_idx]);
                thread_detach(thread[one_idx]);
                thread[one_idx] = NULL;
        }
}

/*-------------------------------------------------------------------*/


static void nop_reader(void *arg)
{
        size_t nop_iters = (size_t)arg;
        
        TPRINTF("Enter nop-reader\n");
        
        for (size_t i = 0; i < nop_iters; ++i) {
                rcu_read_lock();
                rcu_read_unlock();
        }
        
        TPRINTF("Exit nop-reader\n");
}

static void get_seq(size_t from, size_t to, size_t steps, size_t *seq)
{
        assert(0 < steps && from <= to && 0 < to);
        size_t inc = (to - from) / (steps - 1);
        
        for (size_t i = 0; i < steps - 1; ++i) {
                seq[i] = i * inc + from;
        }
        
        seq[steps - 1] = to;
}

static bool do_nop_readers(void)
{
        size_t seq[MAX_THREADS] = {0};
        get_seq(100, 100000, get_thread_cnt(), seq);
        
        TPRINTF("\nRun %zu thr: repeat empty no-op reader sections\n", get_thread_cnt());
        
        for (size_t k = 0; k < get_thread_cnt(); ++k)
                run_one(nop_reader, (void*)seq[k]);
        
        TPRINTF("\nJoining %zu no-op readers\n", get_thread_cnt());
        join_all();
        
        return true;
}

/*-------------------------------------------------------------------*/



static void long_reader(void *arg)
{
        const size_t iter_cnt = 100 * 1000 * 1000;
        size_t nop_iters = (size_t)arg;
        size_t outer_iters = iter_cnt / nop_iters;
        
        TPRINTF("Enter long-reader\n");
        
        for (size_t i = 0; i < outer_iters; ++i) {
                rcu_read_lock();
                
                for (volatile size_t k = 0; k < nop_iters; ++k) {
                        /* nop, but increment volatile k */
                }
                
                rcu_read_unlock();
        }
        
        TPRINTF("Exit long-reader\n");
}

static bool do_long_readers(void)
{
        size_t seq[MAX_THREADS] = {0};
        get_seq(10, 1000 * 1000, get_thread_cnt(), seq);
        
        TPRINTF("\nRun %zu thr: repeat long reader sections, will preempt, no cbs.\n", 
                get_thread_cnt());
        
        for (size_t k = 0; k < get_thread_cnt(); ++k)
                run_one(long_reader, (void*)seq[k]);
        
        TPRINTF("\nJoining %zu readers with long reader sections.\n", get_thread_cnt());
        join_all();
        
        return true;
}

/*-------------------------------------------------------------------*/


static atomic_t nop_callbacks_cnt = {0};
/* Must be even. */
static const int nop_updater_iters = 10000;

static void count_cb(rcu_item_t *item)
{
        atomic_inc(&nop_callbacks_cnt);
        free(item);
}

static void nop_updater(void *arg)
{
        for (int i = 0; i < nop_updater_iters; i += 2){
                rcu_item_t *a = malloc(sizeof(rcu_item_t), FRAME_ATOMIC);
                rcu_item_t *b = malloc(sizeof(rcu_item_t), FRAME_ATOMIC);
                
                if (a && b) {
                        rcu_call(a, count_cb);
                        rcu_call(b, count_cb);
                } else {
                        TPRINTF("[out-of-mem]\n");
                        free(a);
                        free(b);
                        return;
                }
        }
}

static bool do_nop_callbacks(void)
{
        atomic_set(&nop_callbacks_cnt, 0);

        size_t exp_cnt = nop_updater_iters * get_thread_cnt();
        size_t max_used_mem = sizeof(rcu_item_t) * exp_cnt;
        
        TPRINTF("\nRun %zu thr: post %zu no-op callbacks (%zu B used), no readers.\n", 
                get_thread_cnt(), exp_cnt, max_used_mem);
        
        run_all(nop_updater);
        TPRINTF("\nJoining %zu no-op callback threads\n", get_thread_cnt());
        join_all();
        
        size_t loop_cnt = 0, max_loops = 15;

        while (exp_cnt != atomic_get(&nop_callbacks_cnt) && loop_cnt < max_loops) {
                ++loop_cnt;
                TPRINTF(".");
                thread_sleep(1);
        }
        
        return loop_cnt < max_loops;
}

/*-------------------------------------------------------------------*/

typedef struct {
        rcu_item_t rcu_item;
        int cookie;
} item_w_cookie_t;

const int magic_cookie = 0x01234567;
static int one_cb_is_done = 0;

static void one_cb_done(rcu_item_t *item)
{
        assert( ((item_w_cookie_t *)item)->cookie == magic_cookie);
        one_cb_is_done = 1;
        TPRINTF("Callback()\n");
        free(item);
}

static void one_cb_reader(void *arg)
{
        TPRINTF("Enter one-cb-reader\n");
        
        rcu_read_lock();
        
        item_w_cookie_t *item = malloc(sizeof(item_w_cookie_t), FRAME_ATOMIC);
        
        if (item) {
                item->cookie = magic_cookie;
                rcu_call(&item->rcu_item, one_cb_done);
        } else {
                TPRINTF("\n[out-of-mem]\n");
        }
        
        thread_sleep(1);
        
        rcu_read_unlock();
        
        TPRINTF("Exit one-cb-reader\n");
}

static bool do_one_cb(void)
{
        one_cb_is_done = 0;
        
        TPRINTF("\nRun a single reader that posts one callback.\n");
        run_one(one_cb_reader, NULL);
        join_one();
        
        TPRINTF("\nJoined one-cb reader, wait for callback.\n");
        size_t loop_cnt = 0;
        size_t max_loops = 4; /* 200 ms total */
        
        while (!one_cb_is_done && loop_cnt < max_loops) {
                thread_usleep(50 * 1000);
                ++loop_cnt;
        }
        
        return one_cb_is_done;
}

/*-------------------------------------------------------------------*/

typedef struct {
        size_t update_cnt;
        size_t read_cnt;
        size_t iters;
} seq_work_t;

typedef struct {
        rcu_item_t rcu;
        atomic_count_t start_time;
} seq_item_t;


static int seq_test_result = EOK;

static atomic_t cur_time = {1};
static atomic_count_t max_upd_done_time = {0};

static void seq_cb(rcu_item_t *rcu_item)
{
        seq_item_t *item = member_to_inst(rcu_item, seq_item_t, rcu);
        
        /* Racy but errs to the conservative side, so it is ok. */
        if (max_upd_done_time < item->start_time) {
                max_upd_done_time = item->start_time;
                
                /* Make updated time visible */
                memory_barrier();
        }

        free(item);
}

static void seq_func(void *arg)
{
        /*
         * Temporarily workaround GCC 7.1.0 internal
         * compiler error when compiling for riscv64.
         */
#ifndef KARCH_riscv64
        seq_work_t *work = (seq_work_t*)arg;
        
        /* Alternate between reader and updater roles. */
        for (size_t k = 0; k < work->iters; ++k) {
                /* Reader */
                for (size_t i = 0; i < work->read_cnt; ++i) {
                        rcu_read_lock();
                        atomic_count_t start_time = atomic_postinc(&cur_time);
                        
                        for (volatile size_t d = 0; d < 10 * i; ++d ){
                                /* no-op */
                        }
                        
                        /* Get most recent max_upd_done_time. */
                        memory_barrier();
                        
                        if (start_time < max_upd_done_time) {
                                seq_test_result = ERACE;
                        }
                        
                        rcu_read_unlock();
                        
                        if (seq_test_result != EOK) 
                                return;
                }
                
                /* Updater */
                for (size_t i = 0; i < work->update_cnt; ++i) {
                        seq_item_t *a = malloc(sizeof(seq_item_t), FRAME_ATOMIC);
                        seq_item_t *b = malloc(sizeof(seq_item_t), FRAME_ATOMIC);
                        
                        if (a && b) {
                                a->start_time = atomic_postinc(&cur_time);
                                rcu_call(&a->rcu, seq_cb);
                                
                                b->start_time = atomic_postinc(&cur_time);
                                rcu_call(&b->rcu, seq_cb);
                        } else {
                                TPRINTF("\n[out-of-mem]\n");
                                seq_test_result = ENOMEM;
                                free(a);
                                free(b);
                                return;
                        }
                }
                
        }
#else
        (void) seq_cb;
#endif
}

static bool do_seq_check(void)
{
        seq_test_result = EOK;
        max_upd_done_time = 0;
        atomic_set(&cur_time, 1);

        const size_t iters = 100;
        const size_t total_cnt = 1000;
        size_t read_cnt[MAX_THREADS] = {0};
        seq_work_t item[MAX_THREADS];
        
        size_t total_cbs = 0;
        size_t max_used_mem = 0;
        
        get_seq(0, total_cnt, get_thread_cnt(), read_cnt);
        

        for (size_t i = 0; i < get_thread_cnt(); ++i) {
                item[i].update_cnt = total_cnt - read_cnt[i];
                item[i].read_cnt = read_cnt[i];
                item[i].iters = iters;
                
                total_cbs += 2 * iters * item[i].update_cnt;
        }
        
        max_used_mem = total_cbs * sizeof(seq_item_t);

        const char *mem_suffix;
        uint64_t mem_units;
        bin_order_suffix(max_used_mem, &mem_units, &mem_suffix, false);
        
        TPRINTF("\nRun %zu th: check callback completion time in readers. "
                "%zu callbacks total (max %" PRIu64 " %s used). Be patient.\n", 
                get_thread_cnt(), total_cbs, mem_units, mem_suffix);
        
        for (size_t i = 0; i < get_thread_cnt(); ++i) {
                run_one(seq_func, &item[i]);
        }
        
        TPRINTF("\nJoining %zu seq-threads\n", get_thread_cnt());
        join_all();
        
        if (seq_test_result == ENOMEM) {
                TPRINTF("\nErr: out-of mem\n");
        } else if (seq_test_result == ERACE) {
                TPRINTF("\nERROR: race detected!!\n");
        } 
        
        return seq_test_result == EOK;
}

/*-------------------------------------------------------------------*/


static void reader_unlocked(rcu_item_t *item)
{
        exited_t *p = (exited_t*)item;
        p->exited = true;
}

static void reader_exit(void *arg)
{
        rcu_read_lock();
        rcu_read_lock();
        rcu_read_lock();
        rcu_read_unlock();
        
        rcu_call((rcu_item_t*)arg, reader_unlocked);
        
        rcu_read_lock();
        rcu_read_lock();
        
        /* Exit without unlocking the rcu reader section. */
}

static bool do_reader_exit(void)
{
        TPRINTF("\nReader exits thread with rcu_lock\n");
        
        exited_t *p = malloc(sizeof(exited_t), FRAME_ATOMIC);
        if (!p) {
                TPRINTF("[out-of-mem]\n");
                return false;
        }
                
        p->exited = false;
        
        run_one(reader_exit, p);        
        join_one();
        
        int result = EOK;
        wait_for_cb_exit(2 /* secs */, p, &result);
        
        if (result != EOK) {
                TPRINTF("Err: RCU locked up after exiting from within a reader\n");
                /* Leak the mem. */
        } else {
                free(p);
        }
        
        return result == EOK;
}

/*-------------------------------------------------------------------*/

/*-------------------------------------------------------------------*/

typedef struct preempt_struct {
        exited_t e;
        int result;
} preempt_t;


static void preempted_unlocked(rcu_item_t *item)
{
        preempt_t *p = member_to_inst(item, preempt_t, e.rcu);
        p->e.exited = true;
        TPRINTF("Callback().\n");
}

static void preempted_reader_prev(void *arg)
{
        preempt_t *p = (preempt_t*)arg;
        assert(!p->e.exited);

        TPRINTF("reader_prev{ ");
        
        rcu_read_lock();
        scheduler();
        rcu_read_unlock();

        /* 
         * Start GP after exiting reader section w/ preemption. 
         * Just check that the callback does not lock up and is not lost.
         */
        rcu_call(&p->e.rcu, preempted_unlocked);

        TPRINTF("}reader_prev\n");
}

static void preempted_reader_inside_cur(void *arg)
{
        preempt_t *p = (preempt_t*)arg;
        assert(!p->e.exited);
        
        TPRINTF("reader_inside_cur{ ");
        /* 
         * Start a GP and try to finish the reader before 
         * the GP ends (including preemption). 
         */
        rcu_call(&p->e.rcu, preempted_unlocked);

        /* Give RCU threads a chance to start up. */
        scheduler();
        scheduler();

        rcu_read_lock();
        /* Come back as soon as possible to complete before GP ends. */
        thread_usleep(2);
        rcu_read_unlock();

        TPRINTF("}reader_inside_cur\n");
}


static void preempted_reader_cur(void *arg)
{
        preempt_t *p = (preempt_t*)arg;
        assert(!p->e.exited);
        
        TPRINTF("reader_cur{ ");
        rcu_read_lock();

        /* Start GP. */
        rcu_call(&p->e.rcu, preempted_unlocked);

        /* Preempt while cur GP detection is running */
        thread_sleep(1);
        
        /* Err: exited before this reader completed. */
        if (p->e.exited)
                p->result = ERACE;

        rcu_read_unlock();
        TPRINTF("}reader_cur\n");
}

static void preempted_reader_next1(void *arg)
{
        preempt_t *p = (preempt_t*)arg;
        assert(!p->e.exited);
        
        TPRINTF("reader_next1{ ");
        rcu_read_lock();

        /* Preempt before cur GP detection starts. */
        scheduler();
        
        /* Start GP. */
        rcu_call(&p->e.rcu, preempted_unlocked);

        /* Err: exited before this reader completed. */
        if (p->e.exited)
                p->result = ERACE;

        rcu_read_unlock();
        TPRINTF("}reader_next1\n");
}

static void preempted_reader_next2(void *arg)
{
        preempt_t *p = (preempt_t*)arg;
        assert(!p->e.exited);
        
        TPRINTF("reader_next2{ ");
        rcu_read_lock();

        /* Preempt before cur GP detection starts. */
        scheduler();
        
        /* Start GP. */
        rcu_call(&p->e.rcu, preempted_unlocked);

        /* 
         * Preempt twice while GP is running after we've been known 
         * to hold up the GP just to make sure multiple preemptions
         * are properly tracked if a reader is delaying the cur GP.
         */
        thread_sleep(1);
        thread_sleep(1);

        /* Err: exited before this reader completed. */
        if (p->e.exited)
                p->result = ERACE;

        rcu_read_unlock();
        TPRINTF("}reader_next2\n");
}


static bool do_one_reader_preempt(void (*f)(void*), const char *err)
{
        preempt_t *p = malloc(sizeof(preempt_t), FRAME_ATOMIC);
        if (!p) {
                TPRINTF("[out-of-mem]\n");
                return false;
        }
        
        p->e.exited = false;
        p->result = EOK;
        
        run_one(f, p);  
        join_one();
        
        /* Wait at most 4 secs. */
        wait_for_cb_exit(4, &p->e, &p->result);
        
        if (p->result == EOK) {
                free(p);
                return true;
        } else {
                TPRINTF("%s", err);
                /* Leak a bit of mem. */
                return false;
        }
}

static bool do_reader_preempt(void)
{
        TPRINTF("\nReaders will be preempted.\n");
        
        bool success = true;
        bool ok = true;
        
        ok = do_one_reader_preempt(preempted_reader_prev, 
                "Err: preempted_reader_prev()\n");
        success = success && ok;
        
        ok = do_one_reader_preempt(preempted_reader_inside_cur, 
                "Err: preempted_reader_inside_cur()\n");
        success = success && ok;
        
        ok = do_one_reader_preempt(preempted_reader_cur, 
                "Err: preempted_reader_cur()\n");
        success = success && ok;
        
        ok = do_one_reader_preempt(preempted_reader_next1, 
                "Err: preempted_reader_next1()\n");
        success = success && ok;

        ok = do_one_reader_preempt(preempted_reader_next2, 
                "Err: preempted_reader_next2()\n");
        success = success && ok;
        
        return success;
}

/*-------------------------------------------------------------------*/
typedef struct {
        bool reader_done;
        bool reader_running;
        bool synch_running;
} synch_t;

static void synch_reader(void *arg)
{
        synch_t *synch = (synch_t *) arg;
        
        rcu_read_lock();

        /* Order accesses of synch after the reader section begins. */
        memory_barrier();
        
        synch->reader_running = true;
        
        while (!synch->synch_running) {
                /* 0.5 sec */
                delay(500 * 1000);
        }
        
        /* Run for 1 sec */
        delay(1000 * 1000);
        /* thread_join() propagates done to do_synch() */
        synch->reader_done = true;
        
        rcu_read_unlock();
}


static bool do_synch(void)
{
        TPRINTF("\nSynchronize with long reader\n");
        
        synch_t *synch = malloc(sizeof(synch_t), FRAME_ATOMIC);
        
        if (!synch) {
                TPRINTF("[out-of-mem]\n");
                return false;
        }
        
        synch->reader_done = false;
        synch->reader_running = false;
        synch->synch_running = false;
        
        run_one(synch_reader, synch);   
        
        /* Wait for the reader to enter its critical section. */
        scheduler();
        while (!synch->reader_running) {
                thread_usleep(500 * 1000);
        }
        
        synch->synch_running = true;
        
        rcu_synchronize();
        join_one();
        
        
        if (synch->reader_done) {
                free(synch);
                return true;
        } else {
                TPRINTF("Err: synchronize() exited prematurely \n");
                /* Leak some mem. */
                return false;
        }
}

/*-------------------------------------------------------------------*/
typedef struct {
        rcu_item_t rcu_item;
        atomic_t done;
} barrier_t;

static void barrier_callback(rcu_item_t *item)
{
        barrier_t *b = member_to_inst(item, barrier_t, rcu_item);
        atomic_set(&b->done, 1);
}

static bool do_barrier(void)
{
        TPRINTF("\nrcu_barrier: Wait for outstanding rcu callbacks to complete\n");
        
        barrier_t *barrier = malloc(sizeof(barrier_t), FRAME_ATOMIC);
        
        if (!barrier) {
                TPRINTF("[out-of-mem]\n");
                return false;
        }
        
        atomic_set(&barrier->done, 0);
        
        rcu_call(&barrier->rcu_item, barrier_callback);
        rcu_barrier();
        
        if (1 == atomic_get(&barrier->done)) {
                free(barrier);
                return true;
        } else {
                TPRINTF("rcu_barrier() exited prematurely.\n");
                /* Leak some mem. */
                return false;
        }
}

/*-------------------------------------------------------------------*/

typedef struct {
        size_t iters;
        bool master;
} stress_t;


static void stress_reader(void *arg)
{
        bool *done = (bool*) arg;
        
        while (!*done) {
                rcu_read_lock();
                rcu_read_unlock();
                
                /* 
                 * Do some work outside of the reader section so we are not always
                 * preempted in the reader section.
                 */
                delay(5);
        }
}

static void stress_cb(rcu_item_t *item)
{
        /* 5 us * 1000 * 1000 iters == 5 sec per updater thread */
        delay(5);
        free(item);
}

static void stress_updater(void *arg)
{
        stress_t *s = (stress_t *)arg;
        
        for (size_t i = 0; i < s->iters; ++i) {
                rcu_item_t *item = malloc(sizeof(rcu_item_t), FRAME_ATOMIC);
                
                if (item) {
                        rcu_call(item, stress_cb);
                } else {
                        TPRINTF("[out-of-mem]\n");
                        return;
                }
                
                /* Print a dot if we make a progress of 1% */
                if (s->master && 0 == (i % (s->iters/100)))
                        TPRINTF(".");
        }
}

static bool do_stress(void)
{
        size_t cb_per_thread = 1000 * 1000;
        bool done = false;
        stress_t master = { .iters = cb_per_thread, .master = true }; 
        stress_t worker = { .iters = cb_per_thread, .master = false }; 
        
        size_t thread_cnt = min(MAX_THREADS / 2, config.cpu_active);
        /* Each cpu has one reader and one updater. */
        size_t reader_cnt = thread_cnt;
        size_t updater_cnt = thread_cnt;
        
        size_t exp_upd_calls = updater_cnt * cb_per_thread;
        size_t max_used_mem = exp_upd_calls * sizeof(rcu_item_t);
        
        const char *mem_suffix;
        uint64_t mem_units;
        bin_order_suffix(max_used_mem, &mem_units, &mem_suffix, false);

        TPRINTF("\nStress: Run %zu nop-readers and %zu updaters. %zu callbacks"
                " total (max %" PRIu64 " %s used). Be very patient.\n", 
                reader_cnt, updater_cnt, exp_upd_calls, mem_units, mem_suffix);
        
        for (size_t k = 0; k < reader_cnt; ++k) {
                run_one(stress_reader, &done);
        }

        for (size_t k = 0; k < updater_cnt; ++k) {
                run_one(stress_updater, k > 0 ? &worker : &master);
        }
        
        TPRINTF("\nJoining %zu stress updaters.\n", updater_cnt);
        
        for (size_t k = 0; k < updater_cnt; ++k) {
                join_one();
        }
        
        done = true;

        TPRINTF("\nJoining %zu stress nop-readers.\n", reader_cnt);
        
        join_all();
        return true;
}
/*-------------------------------------------------------------------*/

typedef struct {
        rcu_item_t r;
        size_t total_cnt;
        size_t count_down;
        bool expedite;
} expedite_t;

static void expedite_cb(rcu_item_t *arg)
{
        expedite_t *e = (expedite_t *)arg;
        
        if (1 < e->count_down) {
                --e->count_down;
                
                if (0 == (e->count_down % (e->total_cnt/100))) {
                        TPRINTF("*");
                }
                
                _rcu_call(e->expedite, &e->r, expedite_cb);
        } else {
                /* Do not touch any of e's mem after we declare we're done with it. */
                memory_barrier();
                e->count_down = 0;
        }
}

static void run_expedite(bool exp, size_t cnt)
{
        expedite_t e;
        e.total_cnt = cnt;
        e.count_down = cnt;
        e.expedite = exp;
        
        _rcu_call(e.expedite, &e.r, expedite_cb);
        
        while (0 < e.count_down) {
                thread_sleep(1);
                TPRINTF(".");
        }
}

static bool do_expedite(void)
{
        size_t exp_cnt = 1000 * 1000;
        size_t normal_cnt = 1 * 1000;
        
        TPRINTF("Expedited: sequence of %zu rcu_calls\n", exp_cnt);
        run_expedite(true, exp_cnt);
        TPRINTF("Normal/non-expedited: sequence of %zu rcu_calls\n", normal_cnt);
        run_expedite(false, normal_cnt);
        return true;
}
/*-------------------------------------------------------------------*/

struct test_func {
        bool include;
        bool (*func)(void);
        const char *desc;
};


const char *test_rcu1(void)
{
        struct test_func test_func[] = {
                { 1, do_one_cb, "do_one_cb" },
                { 1, do_reader_preempt, "do_reader_preempt" },
                { 1, do_synch, "do_synch" },
                { 1, do_barrier, "do_barrier" },
                { 1, do_reader_exit, "do_reader_exit" },
                { 1, do_nop_readers, "do_nop_readers" },
                { 1, do_seq_check, "do_seq_check" },
                { 0, do_long_readers, "do_long_readers" },
                { 1, do_nop_callbacks, "do_nop_callbacks" },
                { 0, do_expedite, "do_expedite" },
                { 1, do_stress, "do_stress" },
                { 0, NULL, NULL }
        };
        
        bool success = true;
        bool ok = true;
        uint64_t completed_gps = rcu_completed_gps();
        uint64_t delta_gps = 0;
        
        for (int i = 0; test_func[i].func; ++i) {
                if (!test_func[i].include) {
                        TPRINTF("\nSubtest %s() skipped.\n", test_func[i].desc);
                        continue;
                } else {
                        TPRINTF("\nRunning subtest %s.\n", test_func[i].desc);
                }
                
                ok = test_func[i].func();
                success = success && ok;
                
                delta_gps = rcu_completed_gps() - completed_gps;
                completed_gps += delta_gps;

                if (ok) {  
                        TPRINTF("\nSubtest %s() ok (GPs: %" PRIu64 ").\n", 
                                test_func[i].desc, delta_gps);
                } else {
                        TPRINTF("\nFailed: %s(). Pausing for 5 secs.\n", test_func[i].desc);
                        thread_sleep(5);
                } 
        }

        if (success)
                return NULL;
        else
                return "One of the tests failed.";
}