numam-dpdk/app/test/test_hash_readwrite.c
Dharmik Thakkar ee8cee83fe test/hash: use compiler atomics for sync
Convert rte_atomic usages to compiler atomic built-ins
for stats sync.

Signed-off-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Reviewed-by: Joyce Kong <joyce.kong@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Acked-by: Yipeng Wang <yipeng1.wang@intel.com>
Tested-by: David Christensen <drc@linux.vnet.ibm.com>
2021-10-19 17:15:10 +02:00

761 lines
19 KiB
C

/* SPDX-License-Identifier: BSD-3-Clause
* Copyright(c) 2018 Intel Corporation
*/
#include <inttypes.h>
#include <locale.h>
#include <rte_cycles.h>
#include <rte_hash.h>
#include <rte_hash_crc.h>
#include <rte_jhash.h>
#include <rte_launch.h>
#include <rte_malloc.h>
#include <rte_random.h>
#include <rte_spinlock.h>
#include "test.h"
#define RTE_RWTEST_FAIL 0
#define TOTAL_ENTRY (5*1024*1024)
#define TOTAL_INSERT (4.5*1024*1024)
#define TOTAL_INSERT_EXT (5*1024*1024)
#define NUM_TEST 3
unsigned int core_cnt[NUM_TEST] = {2, 4, 8};
unsigned int worker_core_ids[RTE_MAX_LCORE];
struct perf {
uint32_t single_read;
uint32_t single_write;
uint32_t read_only[NUM_TEST];
uint32_t write_only[NUM_TEST];
uint32_t read_write_r[NUM_TEST];
uint32_t read_write_w[NUM_TEST];
};
static struct perf htm_results, non_htm_results;
struct {
uint32_t *keys;
uint8_t *found;
uint32_t num_insert;
uint32_t rounded_tot_insert;
struct rte_hash *h;
} tbl_rw_test_param;
static uint64_t gcycles;
static uint64_t ginsertions;
static uint64_t gread_cycles;
static uint64_t gwrite_cycles;
static uint64_t greads;
static uint64_t gwrites;
static int
test_hash_readwrite_worker(__rte_unused void *arg)
{
uint64_t i, offset;
uint32_t lcore_id = rte_lcore_id();
uint64_t begin, cycles;
int *ret;
ret = rte_malloc(NULL, sizeof(int) *
tbl_rw_test_param.num_insert, 0);
for (i = 0; i < rte_lcore_count(); i++) {
if (worker_core_ids[i] == lcore_id)
break;
}
offset = tbl_rw_test_param.num_insert * i;
printf("Core #%d inserting and reading %d: %'"PRId64" - %'"PRId64"\n",
lcore_id, tbl_rw_test_param.num_insert,
offset, offset + tbl_rw_test_param.num_insert - 1);
begin = rte_rdtsc_precise();
for (i = offset; i < offset + tbl_rw_test_param.num_insert; i++) {
if (rte_hash_lookup(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i) > 0)
break;
ret[i - offset] = rte_hash_add_key(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i);
if (ret[i - offset] < 0)
break;
/* lookup a random key */
uint32_t rand = rte_rand() % (i + 1 - offset);
if (rte_hash_lookup(tbl_rw_test_param.h,
tbl_rw_test_param.keys + rand) != ret[rand])
break;
if (rte_hash_del_key(tbl_rw_test_param.h,
tbl_rw_test_param.keys + rand) != ret[rand])
break;
ret[rand] = rte_hash_add_key(tbl_rw_test_param.h,
tbl_rw_test_param.keys + rand);
if (ret[rand] < 0)
break;
if (rte_hash_lookup(tbl_rw_test_param.h,
tbl_rw_test_param.keys + rand) != ret[rand])
break;
}
cycles = rte_rdtsc_precise() - begin;
__atomic_fetch_add(&gcycles, cycles, __ATOMIC_RELAXED);
__atomic_fetch_add(&ginsertions, i - offset, __ATOMIC_RELAXED);
for (; i < offset + tbl_rw_test_param.num_insert; i++)
tbl_rw_test_param.keys[i] = RTE_RWTEST_FAIL;
rte_free(ret);
return 0;
}
static int
init_params(int use_ext, int use_htm, int rw_lf, int use_jhash)
{
unsigned int i;
uint32_t *keys = NULL;
uint8_t *found = NULL;
struct rte_hash *handle;
struct rte_hash_parameters hash_params = {
.entries = TOTAL_ENTRY,
.key_len = sizeof(uint32_t),
.hash_func_init_val = 0,
.socket_id = rte_socket_id(),
};
if (use_jhash)
hash_params.hash_func = rte_jhash;
else
hash_params.hash_func = rte_hash_crc;
hash_params.extra_flag = RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD;
if (use_htm)
hash_params.extra_flag |=
RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT;
if (rw_lf)
hash_params.extra_flag |=
RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF;
else
hash_params.extra_flag |=
RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY;
if (use_ext)
hash_params.extra_flag |=
RTE_HASH_EXTRA_FLAGS_EXT_TABLE;
else
hash_params.extra_flag &=
~RTE_HASH_EXTRA_FLAGS_EXT_TABLE;
hash_params.name = "tests";
handle = rte_hash_create(&hash_params);
if (handle == NULL) {
printf("hash creation failed");
return -1;
}
tbl_rw_test_param.h = handle;
keys = rte_malloc(NULL, sizeof(uint32_t) * TOTAL_ENTRY, 0);
if (keys == NULL) {
printf("RTE_MALLOC failed\n");
goto err;
}
found = rte_zmalloc(NULL, sizeof(uint8_t) * TOTAL_ENTRY, 0);
if (found == NULL) {
printf("RTE_ZMALLOC failed\n");
goto err;
}
tbl_rw_test_param.keys = keys;
tbl_rw_test_param.found = found;
for (i = 0; i < TOTAL_ENTRY; i++)
keys[i] = i;
return 0;
err:
rte_free(keys);
rte_hash_free(handle);
return -1;
}
static int
test_hash_readwrite_functional(int use_htm, int use_rw_lf, int use_ext)
{
unsigned int i;
const void *next_key;
void *next_data;
uint32_t iter = 0;
uint32_t duplicated_keys = 0;
uint32_t lost_keys = 0;
int use_jhash = 1;
int worker_cnt = rte_lcore_count() - 1;
uint32_t tot_insert = 0;
__atomic_store_n(&gcycles, 0, __ATOMIC_RELAXED);
__atomic_store_n(&ginsertions, 0, __ATOMIC_RELAXED);
if (init_params(use_ext, use_htm, use_rw_lf, use_jhash) != 0)
goto err;
if (use_ext)
tot_insert = TOTAL_INSERT_EXT;
else
tot_insert = TOTAL_INSERT;
tbl_rw_test_param.num_insert =
tot_insert / worker_cnt;
tbl_rw_test_param.rounded_tot_insert =
tbl_rw_test_param.num_insert * worker_cnt;
printf("\nHTM = %d, RW-LF = %d, EXT-Table = %d\n",
use_htm, use_rw_lf, use_ext);
printf("++++++++Start function tests:+++++++++\n");
/* Fire all threads. */
rte_eal_mp_remote_launch(test_hash_readwrite_worker,
NULL, SKIP_MAIN);
rte_eal_mp_wait_lcore();
while (rte_hash_iterate(tbl_rw_test_param.h, &next_key,
&next_data, &iter) >= 0) {
/* Search for the key in the list of keys added .*/
i = *(const uint32_t *)next_key;
tbl_rw_test_param.found[i]++;
}
for (i = 0; i < tbl_rw_test_param.rounded_tot_insert; i++) {
if (tbl_rw_test_param.keys[i] != RTE_RWTEST_FAIL) {
if (tbl_rw_test_param.found[i] > 1) {
duplicated_keys++;
break;
}
if (tbl_rw_test_param.found[i] == 0) {
lost_keys++;
printf("key %d is lost\n", i);
break;
}
}
}
if (duplicated_keys > 0) {
printf("%d key duplicated\n", duplicated_keys);
goto err_free;
}
if (lost_keys > 0) {
printf("%d key lost\n", lost_keys);
goto err_free;
}
printf("No key corrupted during read-write test.\n");
unsigned long long int cycles_per_insertion =
__atomic_load_n(&gcycles, __ATOMIC_RELAXED) /
__atomic_load_n(&ginsertions, __ATOMIC_RELAXED);
printf("cycles per insertion and lookup: %llu\n", cycles_per_insertion);
rte_free(tbl_rw_test_param.found);
rte_free(tbl_rw_test_param.keys);
rte_hash_free(tbl_rw_test_param.h);
printf("+++++++++Complete function tests+++++++++\n");
return 0;
err_free:
rte_free(tbl_rw_test_param.found);
rte_free(tbl_rw_test_param.keys);
rte_hash_free(tbl_rw_test_param.h);
err:
return -1;
}
static int
test_rw_reader(void *arg)
{
uint64_t i;
uint64_t begin, cycles;
uint64_t read_cnt = (uint64_t)((uintptr_t)arg);
begin = rte_rdtsc_precise();
for (i = 0; i < read_cnt; i++) {
void *data = arg;
rte_hash_lookup_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
&data);
if (i != (uint64_t)(uintptr_t)data) {
printf("lookup find wrong value %"PRIu64","
"%"PRIu64"\n", i,
(uint64_t)(uintptr_t)data);
break;
}
}
cycles = rte_rdtsc_precise() - begin;
__atomic_fetch_add(&gread_cycles, cycles, __ATOMIC_RELAXED);
__atomic_fetch_add(&greads, i, __ATOMIC_RELAXED);
return 0;
}
static int
test_rw_writer(void *arg)
{
uint64_t i;
uint32_t lcore_id = rte_lcore_id();
uint64_t begin, cycles;
int ret;
uint64_t start_coreid = (uint64_t)(uintptr_t)arg;
uint64_t offset;
for (i = 0; i < rte_lcore_count(); i++) {
if (worker_core_ids[i] == lcore_id)
break;
}
offset = TOTAL_INSERT / 2 + (i - (start_coreid)) *
tbl_rw_test_param.num_insert;
begin = rte_rdtsc_precise();
for (i = offset; i < offset + tbl_rw_test_param.num_insert; i++) {
ret = rte_hash_add_key_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
(void *)((uintptr_t)i));
if (ret < 0) {
printf("writer failed %"PRIu64"\n", i);
break;
}
}
cycles = rte_rdtsc_precise() - begin;
__atomic_fetch_add(&gwrite_cycles, cycles, __ATOMIC_RELAXED);
__atomic_fetch_add(&gwrites, tbl_rw_test_param.num_insert,
__ATOMIC_RELAXED);
return 0;
}
static int
test_hash_readwrite_perf(struct perf *perf_results, int use_htm,
int reader_faster)
{
unsigned int n;
int ret;
int start_coreid;
uint64_t i, read_cnt;
const void *next_key;
void *next_data;
uint32_t iter;
int use_jhash = 0;
uint32_t duplicated_keys = 0;
uint32_t lost_keys = 0;
uint64_t start = 0, end = 0;
__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
if (init_params(0, use_htm, 0, use_jhash) != 0)
goto err;
/*
* Do a readers finish faster or writers finish faster test.
* When readers finish faster, we timing the readers, and when writers
* finish faster, we timing the writers.
* Divided by 10 or 2 is just experimental values to vary the workload
* of readers.
*/
if (reader_faster) {
printf("++++++Start perf test: reader++++++++\n");
read_cnt = TOTAL_INSERT / 10;
} else {
printf("++++++Start perf test: writer++++++++\n");
read_cnt = TOTAL_INSERT / 2;
}
/* We first test single thread performance */
start = rte_rdtsc_precise();
/* Insert half of the keys */
for (i = 0; i < TOTAL_INSERT / 2; i++) {
ret = rte_hash_add_key_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
(void *)((uintptr_t)i));
if (ret < 0) {
printf("Failed to insert half of keys\n");
goto err_free;
}
}
end = rte_rdtsc_precise() - start;
perf_results->single_write = end / i;
start = rte_rdtsc_precise();
for (i = 0; i < read_cnt; i++) {
void *data;
rte_hash_lookup_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
&data);
if (i != (uint64_t)(uintptr_t)data) {
printf("lookup find wrong value"
" %"PRIu64",%"PRIu64"\n", i,
(uint64_t)(uintptr_t)data);
break;
}
}
end = rte_rdtsc_precise() - start;
perf_results->single_read = end / i;
for (n = 0; n < NUM_TEST; n++) {
unsigned int tot_worker_lcore = rte_lcore_count() - 1;
if (tot_worker_lcore < core_cnt[n] * 2)
goto finish;
__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
rte_hash_reset(tbl_rw_test_param.h);
tbl_rw_test_param.num_insert = TOTAL_INSERT / 2 / core_cnt[n];
tbl_rw_test_param.rounded_tot_insert = TOTAL_INSERT / 2 +
tbl_rw_test_param.num_insert *
core_cnt[n];
for (i = 0; i < TOTAL_INSERT / 2; i++) {
ret = rte_hash_add_key_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
(void *)((uintptr_t)i));
if (ret < 0) {
printf("Failed to insert half of keys\n");
goto err_free;
}
}
/* Then test multiple thread case but only all reads or
* all writes
*/
/* Test only reader cases */
for (i = 0; i < core_cnt[n]; i++)
rte_eal_remote_launch(test_rw_reader,
(void *)(uintptr_t)read_cnt,
worker_core_ids[i]);
rte_eal_mp_wait_lcore();
start_coreid = i;
/* Test only writer cases */
for (; i < core_cnt[n] * 2; i++)
rte_eal_remote_launch(test_rw_writer,
(void *)((uintptr_t)start_coreid),
worker_core_ids[i]);
rte_eal_mp_wait_lcore();
if (reader_faster) {
unsigned long long int cycles_per_insertion =
__atomic_load_n(&gread_cycles, __ATOMIC_RELAXED) /
__atomic_load_n(&greads, __ATOMIC_RELAXED);
perf_results->read_only[n] = cycles_per_insertion;
printf("Reader only: cycles per lookup: %llu\n",
cycles_per_insertion);
}
else {
unsigned long long int cycles_per_insertion =
__atomic_load_n(&gwrite_cycles, __ATOMIC_RELAXED) /
__atomic_load_n(&gwrites, __ATOMIC_RELAXED);
perf_results->write_only[n] = cycles_per_insertion;
printf("Writer only: cycles per writes: %llu\n",
cycles_per_insertion);
}
__atomic_store_n(&greads, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gread_cycles, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gwrites, 0, __ATOMIC_RELAXED);
__atomic_store_n(&gwrite_cycles, 0, __ATOMIC_RELAXED);
rte_hash_reset(tbl_rw_test_param.h);
for (i = 0; i < TOTAL_INSERT / 2; i++) {
ret = rte_hash_add_key_data(tbl_rw_test_param.h,
tbl_rw_test_param.keys + i,
(void *)((uintptr_t)i));
if (ret < 0) {
printf("Failed to insert half of keys\n");
goto err_free;
}
}
start_coreid = core_cnt[n];
if (reader_faster) {
for (i = core_cnt[n]; i < core_cnt[n] * 2; i++)
rte_eal_remote_launch(test_rw_writer,
(void *)((uintptr_t)start_coreid),
worker_core_ids[i]);
for (i = 0; i < core_cnt[n]; i++)
rte_eal_remote_launch(test_rw_reader,
(void *)(uintptr_t)read_cnt,
worker_core_ids[i]);
} else {
for (i = 0; i < core_cnt[n]; i++)
rte_eal_remote_launch(test_rw_reader,
(void *)(uintptr_t)read_cnt,
worker_core_ids[i]);
for (; i < core_cnt[n] * 2; i++)
rte_eal_remote_launch(test_rw_writer,
(void *)((uintptr_t)start_coreid),
worker_core_ids[i]);
}
rte_eal_mp_wait_lcore();
iter = 0;
memset(tbl_rw_test_param.found, 0, TOTAL_ENTRY);
while (rte_hash_iterate(tbl_rw_test_param.h,
&next_key, &next_data, &iter) >= 0) {
/* Search for the key in the list of keys added .*/
i = *(const uint32_t *)next_key;
tbl_rw_test_param.found[i]++;
}
for (i = 0; i < tbl_rw_test_param.rounded_tot_insert; i++) {
if (tbl_rw_test_param.keys[i] != RTE_RWTEST_FAIL) {
if (tbl_rw_test_param.found[i] > 1) {
duplicated_keys++;
break;
}
if (tbl_rw_test_param.found[i] == 0) {
lost_keys++;
printf("key %"PRIu64" is lost\n", i);
break;
}
}
}
if (duplicated_keys > 0) {
printf("%d key duplicated\n", duplicated_keys);
goto err_free;
}
if (lost_keys > 0) {
printf("%d key lost\n", lost_keys);
goto err_free;
}
printf("No key corrupted during read-write test.\n");
if (reader_faster) {
unsigned long long int cycles_per_insertion =
__atomic_load_n(&gread_cycles, __ATOMIC_RELAXED) /
__atomic_load_n(&greads, __ATOMIC_RELAXED);
perf_results->read_write_r[n] = cycles_per_insertion;
printf("Read-write cycles per lookup: %llu\n",
cycles_per_insertion);
}
else {
unsigned long long int cycles_per_insertion =
__atomic_load_n(&gwrite_cycles, __ATOMIC_RELAXED) /
__atomic_load_n(&gwrites, __ATOMIC_RELAXED);
perf_results->read_write_w[n] = cycles_per_insertion;
printf("Read-write cycles per writes: %llu\n",
cycles_per_insertion);
}
}
finish:
rte_free(tbl_rw_test_param.found);
rte_free(tbl_rw_test_param.keys);
rte_hash_free(tbl_rw_test_param.h);
return 0;
err_free:
rte_free(tbl_rw_test_param.found);
rte_free(tbl_rw_test_param.keys);
rte_hash_free(tbl_rw_test_param.h);
err:
return -1;
}
static int
test_hash_rw_perf_main(void)
{
/*
* Variables used to choose different tests.
* use_htm indicates if hardware transactional memory should be used.
* reader_faster indicates if the reader threads should finish earlier
* than writer threads. This is to timing either reader threads or
* writer threads for performance numbers.
*/
int use_htm, reader_faster;
unsigned int i = 0, core_id = 0;
if (rte_lcore_count() < 3) {
printf("Not enough cores for hash_readwrite_autotest, expecting at least 3\n");
return TEST_SKIPPED;
}
RTE_LCORE_FOREACH_WORKER(core_id) {
worker_core_ids[i] = core_id;
i++;
}
setlocale(LC_NUMERIC, "");
if (rte_tm_supported()) {
printf("Hardware transactional memory (lock elision) "
"is supported\n");
printf("Test read-write with Hardware transactional memory\n");
use_htm = 1;
reader_faster = 1;
if (test_hash_readwrite_perf(&htm_results, use_htm,
reader_faster) < 0)
return -1;
reader_faster = 0;
if (test_hash_readwrite_perf(&htm_results, use_htm,
reader_faster) < 0)
return -1;
} else {
printf("Hardware transactional memory (lock elision) "
"is NOT supported\n");
}
printf("Test read-write without Hardware transactional memory\n");
use_htm = 0;
reader_faster = 1;
if (test_hash_readwrite_perf(&non_htm_results, use_htm,
reader_faster) < 0)
return -1;
reader_faster = 0;
if (test_hash_readwrite_perf(&non_htm_results, use_htm,
reader_faster) < 0)
return -1;
printf("================\n");
printf("Results summary:\n");
printf("================\n");
printf("single read: %u\n", htm_results.single_read);
printf("single write: %u\n", htm_results.single_write);
for (i = 0; i < NUM_TEST; i++) {
printf("+++ core_cnt: %u +++\n", core_cnt[i]);
printf("HTM:\n");
printf(" read only: %u\n", htm_results.read_only[i]);
printf(" write only: %u\n", htm_results.write_only[i]);
printf(" read-write read: %u\n", htm_results.read_write_r[i]);
printf(" read-write write: %u\n", htm_results.read_write_w[i]);
printf("non HTM:\n");
printf(" read only: %u\n", non_htm_results.read_only[i]);
printf(" write only: %u\n", non_htm_results.write_only[i]);
printf(" read-write read: %u\n",
non_htm_results.read_write_r[i]);
printf(" read-write write: %u\n",
non_htm_results.read_write_w[i]);
}
return 0;
}
static int
test_hash_rw_func_main(void)
{
/*
* Variables used to choose different tests.
* use_htm indicates if hardware transactional memory should be used.
* reader_faster indicates if the reader threads should finish earlier
* than writer threads. This is to timing either reader threads or
* writer threads for performance numbers.
*/
unsigned int i = 0, core_id = 0;
if (rte_lcore_count() < 3) {
printf("Not enough cores for hash_readwrite_autotest, expecting at least 3\n");
return TEST_SKIPPED;
}
RTE_LCORE_FOREACH_WORKER(core_id) {
worker_core_ids[i] = core_id;
i++;
}
setlocale(LC_NUMERIC, "");
if (rte_tm_supported()) {
printf("Hardware transactional memory (lock elision) "
"is supported\n");
printf("Test read-write with Hardware transactional memory\n");
/* htm = 1, rw_lf = 0, ext = 0 */
if (test_hash_readwrite_functional(1, 0, 0) < 0)
return -1;
/* htm = 1, rw_lf = 1, ext = 0 */
if (test_hash_readwrite_functional(1, 1, 0) < 0)
return -1;
/* htm = 1, rw_lf = 0, ext = 1 */
if (test_hash_readwrite_functional(1, 0, 1) < 0)
return -1;
/* htm = 1, rw_lf = 1, ext = 1 */
if (test_hash_readwrite_functional(1, 1, 1) < 0)
return -1;
} else {
printf("Hardware transactional memory (lock elision) "
"is NOT supported\n");
}
printf("Test read-write without Hardware transactional memory\n");
/* htm = 0, rw_lf = 0, ext = 0 */
if (test_hash_readwrite_functional(0, 0, 0) < 0)
return -1;
/* htm = 0, rw_lf = 1, ext = 0 */
if (test_hash_readwrite_functional(0, 1, 0) < 0)
return -1;
/* htm = 0, rw_lf = 0, ext = 1 */
if (test_hash_readwrite_functional(0, 0, 1) < 0)
return -1;
/* htm = 0, rw_lf = 1, ext = 1 */
if (test_hash_readwrite_functional(0, 1, 1) < 0)
return -1;
return 0;
}
REGISTER_TEST_COMMAND(hash_readwrite_func_autotest, test_hash_rw_func_main);
REGISTER_TEST_COMMAND(hash_readwrite_perf_autotest, test_hash_rw_perf_main);