obj_pmalloc_mt.c

// SPDX-License-Identifier: BSD-3-Clause
/* Copyright 2015-2023, Intel Corporation */

/*
 * obj_pmalloc_mt.c -- multithreaded test of allocator
 */
#include <stdint.h>

#include "file.h"
#include "obj.h"
#include "pmalloc.h"
#include "sys_util.h"
#include "unittest.h"

#define MAX_THREADS 32
#define MAX_OPS_PER_THREAD 1000
#define ALLOC_SIZE 104
#define REALLOC_SIZE (ALLOC_SIZE * 3)
#define MIX_RERUNS 2

#define CHUNKSIZE (1 << 18)
#define CHUNKS_PER_THREAD 3

static unsigned Threads;
static unsigned Ops_per_thread;
static unsigned Tx_per_thread;

struct action {
	struct pobj_action pact;
	os_mutex_t lock;
	os_cond_t cond;
};

struct root {
	uint64_t offs[MAX_THREADS][MAX_OPS_PER_THREAD];
	struct action actions[MAX_THREADS][MAX_OPS_PER_THREAD];
};

struct worker_args {
	PMEMobjpool *pop;
	struct root *r;
	unsigned idx;
};

static void *
alloc_worker(void *arg)
{
	struct worker_args *a = arg;

	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		pmalloc(a->pop, &a->r->offs[a->idx][i], ALLOC_SIZE, 0, 0);
		UT_ASSERTne(a->r->offs[a->idx][i], 0);
	}

	return NULL;
}

static void *
realloc_worker(void *arg)
{
	struct worker_args *a = arg;
	int ret;

	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		ret = prealloc(a->pop, &a->r->offs[a->idx][i], REALLOC_SIZE,
				0, 0);
		UT_ASSERTne(a->r->offs[a->idx][i], 0);
		UT_ASSERTeq(ret, 0);
	}

	return NULL;
}

static void *
free_worker(void *arg)
{
	struct worker_args *a = arg;

	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		pfree(a->pop, &a->r->offs[a->idx][i]);
		UT_ASSERTeq(a->r->offs[a->idx][i], 0);
	}

	return NULL;
}

static void *
mix_worker(void *arg)
{
	struct worker_args *a = arg;

	/*
	 * The mix scenario is ran twice to increase the chances of run
	 * contention.
	 */
	for (unsigned j = 0; j < MIX_RERUNS; ++j) {
		for (unsigned i = 0; i < Ops_per_thread; ++i) {
			pmalloc(a->pop, &a->r->offs[a->idx][i],
				ALLOC_SIZE, 0, 0);
			UT_ASSERTne(a->r->offs[a->idx][i], 0);
		}

		for (unsigned i = 0; i < Ops_per_thread; ++i) {
			pfree(a->pop, &a->r->offs[a->idx][i]);
			UT_ASSERTeq(a->r->offs[a->idx][i], 0);
		}
	}

	return NULL;
}

static void *
tx_worker(void *arg)
{
	struct worker_args *a = arg;
	PMEMoid oid;

	/*
	 * Allocate objects until exhaustion, once that happens the transaction
	 * will automatically abort and all of the objects will be freed.
	 */
	TX_BEGIN(a->pop) {
		for (unsigned n = 0; ; ++n) { /* this is NOT an infinite loop */
			oid = pmemobj_tx_alloc(ALLOC_SIZE, a->idx);
			UT_ASSERT(!OID_IS_NULL(oid));
			if (Ops_per_thread != MAX_OPS_PER_THREAD &&
			    n == Ops_per_thread) {
				pmemobj_tx_abort(0);
			}
		}
	} TX_END

	return NULL;
}

static void *
tx3_worker(void *arg)
{
	struct worker_args *a = arg;
	PMEMoid oid;

	/*
	 * Allocate N objects, abort, repeat M times. Should reveal issues in
	 * transaction abort handling.
	 */
	for (unsigned n = 0; n < Tx_per_thread; ++n) {
		TX_BEGIN(a->pop) {
			for (unsigned i = 0; i < Ops_per_thread; ++i) {
				oid = pmemobj_tx_alloc(ALLOC_SIZE, a->idx);
				UT_ASSERT(!OID_IS_NULL(oid));
			}
			pmemobj_tx_abort(EINVAL);
		} TX_END
	}

	return NULL;
}

static void *
alloc_free_worker(void *arg)
{
	struct worker_args *a = arg;

	PMEMoid oid;
	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		int err = pmemobj_alloc(a->pop, &oid, ALLOC_SIZE,
				0, NULL, NULL);
		UT_ASSERTeq(err, 0);
		pmemobj_free(&oid);
	}

	return NULL;
}

#define OPS_PER_TX 10
#define STEP 8
#define TEST_LANES 4

static void *
tx2_worker(void *arg)
{
	struct worker_args *a = arg;

	for (unsigned n = 0; n < Tx_per_thread; ++n) {
		PMEMoid oids[OPS_PER_TX];
		TX_BEGIN(a->pop) {
			for (int i = 0; i < OPS_PER_TX; ++i) {
				oids[i] = pmemobj_tx_alloc(ALLOC_SIZE, a->idx);
				for (unsigned j = 0; j < ALLOC_SIZE;
						j += STEP) {
					pmemobj_tx_add_range(oids[i], j, STEP);
				}
			}
		} TX_END

		TX_BEGIN(a->pop) {
			for (int i = 0; i < OPS_PER_TX; ++i)
				pmemobj_tx_free(oids[i]);
		} TX_ONABORT {
			UT_ASSERT(0);
		} TX_END
	}

	return NULL;
}

static void *
action_cancel_worker(void *arg)
{
	struct worker_args *a = arg;

	PMEMoid oid;
	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		unsigned arr_id = a->idx / 2;
		struct action *act = &a->r->actions[arr_id][i];
		if (a->idx % 2 == 0) {
			os_mutex_lock(&act->lock);
			oid = pmemobj_reserve(a->pop,
				&act->pact, ALLOC_SIZE, 0);
			UT_ASSERT(!OID_IS_NULL(oid));
			os_cond_signal(&act->cond);
			os_mutex_unlock(&act->lock);
		} else {
			os_mutex_lock(&act->lock);
			while (act->pact.heap.offset == 0)
				os_cond_wait(&act->cond, &act->lock);
			pmemobj_cancel(a->pop, &act->pact, 1);
			os_mutex_unlock(&act->lock);
		}
	}

	return NULL;
}

static void *
action_publish_worker(void *arg)
{
	struct worker_args *a = arg;

	PMEMoid oid;
	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		unsigned arr_id = a->idx / 2;
		struct action *act = &a->r->actions[arr_id][i];
		if (a->idx % 2 == 0) {
			os_mutex_lock(&act->lock);
			oid = pmemobj_reserve(a->pop,
				&act->pact, ALLOC_SIZE, 0);
			UT_ASSERT(!OID_IS_NULL(oid));
			os_cond_signal(&act->cond);
			os_mutex_unlock(&act->lock);
		} else {
			os_mutex_lock(&act->lock);
			while (act->pact.heap.offset == 0)
				os_cond_wait(&act->cond, &act->lock);
			pmemobj_publish(a->pop, &act->pact, 1);
			os_mutex_unlock(&act->lock);
		}
	}

	return NULL;
}

static void *
action_mix_worker(void *arg)
{
	struct worker_args *a = arg;

	PMEMoid oid;
	for (unsigned i = 0; i < Ops_per_thread; ++i) {
		unsigned arr_id = a->idx / 2;
		unsigned publish = i % 2;
		struct action *act = &a->r->actions[arr_id][i];
		if (a->idx % 2 == 0) {
			os_mutex_lock(&act->lock);
			oid = pmemobj_reserve(a->pop,
				&act->pact, ALLOC_SIZE, 0);
			UT_ASSERT(!OID_IS_NULL(oid));
			os_cond_signal(&act->cond);
			os_mutex_unlock(&act->lock);
		} else {
			os_mutex_lock(&act->lock);
			while (act->pact.heap.offset == 0)
				os_cond_wait(&act->cond, &act->lock);
			if (publish)
				pmemobj_publish(a->pop, &act->pact, 1);
			else
				pmemobj_cancel(a->pop, &act->pact, 1);
			os_mutex_unlock(&act->lock);
		}
		pmemobj_persist(a->pop, act, sizeof(*act));
	}

	return NULL;
}

static void
actions_clear(PMEMobjpool *pop, struct root *r)
{
	for (unsigned i = 0; i < Threads; ++i) {
		for (unsigned j = 0; j < Ops_per_thread; ++j) {
			struct action *a = &r->actions[i][j];
			util_mutex_destroy(&a->lock);
			util_mutex_init(&a->lock);
			util_cond_destroy(&a->cond);
			util_cond_init(&a->cond);
			memset(&a->pact, 0, sizeof(a->pact));
			pmemobj_persist(pop, a, sizeof(*a));
		}
	}
}

static void
run_worker(void *(worker_func)(void *arg), struct worker_args args[])
{
	os_thread_t t[MAX_THREADS];

	for (unsigned i = 0; i < Threads; ++i)
		THREAD_CREATE(&t[i], NULL, worker_func, &args[i]);

	for (unsigned i = 0; i < Threads; ++i)
		THREAD_JOIN(&t[i], NULL);
}

int
main(int argc, char *argv[])
{
	START(argc, argv, "obj_pmalloc_mt");

	if (argc < 5)
		UT_FATAL(
		"usage: %s <threads> <ops/t> <tx/t> <file> [enable stats]",
		argv[0]);

	PMEMobjpool *pop;
	unsigned enable_stats = 0;
	size_t allocPre, alloc, allocPost;

	Threads = ATOU(argv[1]);
	if (Threads > MAX_THREADS)
		UT_FATAL("Threads %d > %d", Threads, MAX_THREADS);
	Ops_per_thread = ATOU(argv[2]);
	if (Ops_per_thread > MAX_OPS_PER_THREAD)
		UT_FATAL("Ops per thread %d > %d", Threads, MAX_THREADS);
	Tx_per_thread = ATOU(argv[3]);

	int exists = util_file_exists(argv[4]);
	if (exists < 0)
		UT_FATAL("!util_file_exists");

	if (!exists) {
		pop = pmemobj_create(argv[4], "TEST", (PMEMOBJ_MIN_POOL) +
			(MAX_THREADS * CHUNKSIZE * CHUNKS_PER_THREAD),
		0666);

		if (pop == NULL)
			UT_FATAL("!pmemobj_create");
	} else {
		pop = pmemobj_open(argv[4], "TEST");

		if (pop == NULL)
			UT_FATAL("!pmemobj_open");
	}
	if (argc > 5)
		enable_stats = ATOU(argv[5]);

	if (enable_stats) {
		int ret = pmemobj_ctl_set(pop, "stats.enabled", &enable_stats);
		UT_ASSERTeq(ret, 0);
	}

	PMEMoid oid = pmemobj_root(pop, sizeof(struct root));
	struct root *r = pmemobj_direct(oid);
	UT_ASSERTne(r, NULL);

	int ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPre);
	UT_ASSERTeq(ret, 0);

	struct worker_args args[MAX_THREADS];

	for (unsigned i = 0; i < Threads; ++i) {
		args[i].pop = pop;
		args[i].r = r;
		args[i].idx = i;
		for (unsigned j = 0; j < Ops_per_thread; ++j) {
			struct action *a = &r->actions[i][j];
			util_mutex_init(&a->lock);
			util_cond_init(&a->cond);
		}
	}

	alloc = allocPre;
	if (enable_stats)
		alloc += Ops_per_thread * Threads * ((ALLOC_SIZE / 128) + 1)
									* 128;
	run_worker(alloc_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	if (enable_stats) {
		alloc -= Ops_per_thread * Threads * ((ALLOC_SIZE / 128) + 1)
									* 128;
		alloc += Ops_per_thread * Threads * ((REALLOC_SIZE / 128) + 1)
									* 128;
	}
	run_worker(realloc_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	alloc = allocPre;
	run_worker(free_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	run_worker(mix_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	run_worker(alloc_free_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	run_worker(action_cancel_worker, args);
	actions_clear(pop, r);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);
	if (enable_stats && Threads > 1)
		alloc += Ops_per_thread / 2 * Threads
					* ((ALLOC_SIZE / 128) + 1) * 128;
	run_worker(action_publish_worker, args);
	actions_clear(pop, r);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated",	&allocPost);
	UT_ASSERTeq(alloc, allocPost);

	if (enable_stats && Threads > 1)
		alloc += Ops_per_thread / 4 * Threads
					 * ((ALLOC_SIZE / 128) + 1) * 128;
	run_worker(action_mix_worker, args);
	ret = pmemobj_ctl_get(pop, "stats.heap.curr_allocated", &allocPost);
	UT_ASSERTeq(alloc, allocPost);

	if (enable_stats) {
		enable_stats = 0;
		ret = pmemobj_ctl_set(pop, "stats.enabled", &enable_stats);
		UT_ASSERTeq(ret, 0);
	}

	/*
	 * Reduce the number of lanes to a value smaller than the number of
	 * threads. This will ensure that at least some of the state of the lane
	 * will be shared between threads. Doing this might reveal bugs related
	 * to runtime race detection instrumentation.
	 */
	unsigned old_nlanes = pop->lanes_desc.runtime_nlanes;
	pop->lanes_desc.runtime_nlanes = TEST_LANES;
	run_worker(tx2_worker, args);
	pop->lanes_desc.runtime_nlanes = old_nlanes;

	/*
	 * This workload might create many allocation classes due to pvector,
	 * keep it last.
	 */
	if (Threads == MAX_THREADS) /* don't run for short tests */
		run_worker(tx_worker, args);
	run_worker(tx3_worker, args);

	pmemobj_close(pop);

	DONE(NULL);
}

#ifdef _MSC_VER
/*
 * Since libpmemobj is linked statically, we need to invoke its ctor/dtor.
 */
MSVC_CONSTR(libpmemobj_init)
MSVC_DESTR(libpmemobj_fini)
#endif