This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "".
The branch, api-next has been updated via 0c7741d067d9b96e14e420b53a256a1b93096e16 (commit) via 614773850d4d572b793ce2db6cb0cd6c94dada3b (commit) via 6a2d72b399388d7d29cba4f60bf3aea3f1ff555c (commit) via e8ec57b012fea43ae501adb66f5c8ee2f03d0a6c (commit) from f6dab2bfa90ef5d3146cb76f86bc8782666f0f3e (commit)
Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below.
- Log ----------------------------------------------------------------- commit 0c7741d067d9b96e14e420b53a256a1b93096e16 Author: Petri Savolainen petri.savolainen@linaro.org Date: Tue Dec 12 15:12:37 2017 +0200
linux-gen: queue: lock-free implementation
Simple implementation of non-blocking, lock-free plain queues. Enqueues are done freely to any free ring node with an atomically increasing counter value. Dequeue operation finds the node with lowest counter value. Implementation requires lockfree 128 bit atomics. Lock-free queues are not supported when those are not available.
Signed-off-by: Petri Savolainen petri.savolainen@linaro.org Reviewed-by: Bill Fischofer bill.fischofer@linaro.org Signed-off-by: Maxim Uvarov maxim.uvarov@linaro.org
diff --git a/platform/linux-generic/Makefile.am b/platform/linux-generic/Makefile.am index e25970b0..b701d807 100644 --- a/platform/linux-generic/Makefile.am +++ b/platform/linux-generic/Makefile.am @@ -160,6 +160,7 @@ noinst_HEADERS = \ include/odp_queue_scalable_internal.h \ include/odp_ring_internal.h \ include/odp_queue_if.h \ + include/odp_queue_lf.h \ include/odp_schedule_if.h \ include/odp_schedule_scalable.h \ include/odp_schedule_scalable_config.h \ @@ -220,6 +221,7 @@ __LIB__libodp_linux_la_SOURCES = \ odp_pool.c \ odp_queue.c \ odp_queue_if.c \ + odp_queue_lf.c \ odp_queue_scalable.c \ odp_rwlock.c \ odp_rwlock_recursive.c \ diff --git a/platform/linux-generic/include/odp_queue_internal.h b/platform/linux-generic/include/odp_queue_internal.h index dd846d59..26410ac7 100644 --- a/platform/linux-generic/include/odp_queue_internal.h +++ b/platform/linux-generic/include/odp_queue_internal.h @@ -53,6 +53,7 @@ struct queue_entry_s { odp_queue_param_t param; odp_pktin_queue_t pktin; odp_pktout_queue_t pktout; + void *queue_lf; char name[ODP_QUEUE_NAME_LEN]; };
diff --git a/platform/linux-generic/include/odp_queue_lf.h b/platform/linux-generic/include/odp_queue_lf.h new file mode 100644 index 00000000..9bd61e42 --- /dev/null +++ b/platform/linux-generic/include/odp_queue_lf.h @@ -0,0 +1,36 @@ +/* Copyright (c) 2018, Linaro Limited + * All rights reserved. + * + * SPDX-License-Identifier: BSD-3-Clause + */ + +#ifndef ODP_QUEUE_LF_H_ +#define ODP_QUEUE_LF_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include <odp_queue_if.h> +#include <odp_queue_internal.h> + +/* Lock-free queue functions */ +typedef struct { + queue_enq_fn_t enq; + queue_enq_multi_fn_t enq_multi; + queue_deq_fn_t deq; + queue_deq_multi_fn_t deq_multi; + +} queue_lf_func_t; + +uint32_t queue_lf_init_global(uint32_t *queue_lf_size, + queue_lf_func_t *lf_func); +void queue_lf_term_global(void); +void *queue_lf_create(queue_entry_t *queue); +void queue_lf_destroy(void *queue_lf); + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/platform/linux-generic/odp_queue.c b/platform/linux-generic/odp_queue.c index 6a69eb84..7cb60562 100644 --- a/platform/linux-generic/odp_queue.c +++ b/platform/linux-generic/odp_queue.c @@ -8,6 +8,7 @@
#include <odp/api/queue.h> #include <odp_queue_internal.h> +#include <odp_queue_lf.h> #include <odp_queue_if.h> #include <odp/api/std_types.h> #include <odp/api/align.h> @@ -39,11 +40,16 @@ static int queue_init(queue_entry_t *queue, const char *name, const odp_queue_param_t *param);
-typedef struct queue_table_t { - queue_entry_t queue[ODP_CONFIG_QUEUES]; -} queue_table_t; +typedef struct queue_global_t { + queue_entry_t queue[ODP_CONFIG_QUEUES];
-static queue_table_t *queue_tbl; + uint32_t queue_lf_num; + uint32_t queue_lf_size; + queue_lf_func_t queue_lf_func; + +} queue_global_t; + +static queue_global_t *queue_glb;
static queue_entry_t *get_qentry(uint32_t queue_id); @@ -64,26 +70,28 @@ static inline odp_queue_t queue_from_id(uint32_t queue_id) static queue_entry_t *get_qentry(uint32_t queue_id) { - return &queue_tbl->queue[queue_id]; + return &queue_glb->queue[queue_id]; }
static int queue_init_global(void) { uint32_t i; odp_shm_t shm; + uint32_t lf_size = 0; + queue_lf_func_t *lf_func;
ODP_DBG("Queue init ... ");
shm = odp_shm_reserve("odp_queues", - sizeof(queue_table_t), + sizeof(queue_global_t), sizeof(queue_entry_t), 0);
- queue_tbl = odp_shm_addr(shm); + queue_glb = odp_shm_addr(shm);
- if (queue_tbl == NULL) + if (queue_glb == NULL) return -1;
- memset(queue_tbl, 0, sizeof(queue_table_t)); + memset(queue_glb, 0, sizeof(queue_global_t));
for (i = 0; i < ODP_CONFIG_QUEUES; i++) { /* init locks */ @@ -93,6 +101,10 @@ static int queue_init_global(void) queue->s.handle = queue_from_id(i); }
+ lf_func = &queue_glb->queue_lf_func; + queue_glb->queue_lf_num = queue_lf_init_global(&lf_size, lf_func); + queue_glb->queue_lf_size = lf_size; + ODP_DBG("done\n"); ODP_DBG("Queue init global\n"); ODP_DBG(" struct queue_entry_s size %zu\n", @@ -122,7 +134,7 @@ static int queue_term_global(void) int i;
for (i = 0; i < ODP_CONFIG_QUEUES; i++) { - queue = &queue_tbl->queue[i]; + queue = &queue_glb->queue[i]; LOCK(&queue->s.lock); if (queue->s.status != QUEUE_STATUS_FREE) { ODP_ERR("Not destroyed queue: %s\n", queue->s.name); @@ -131,6 +143,8 @@ static int queue_term_global(void) UNLOCK(&queue->s.lock); }
+ queue_lf_term_global(); + ret = odp_shm_free(odp_shm_lookup("odp_queues")); if (ret < 0) { ODP_ERR("shm free failed for odp_queues"); @@ -151,6 +165,8 @@ static int queue_capability(odp_queue_capability_t *capa) capa->sched_prios = odp_schedule_num_prio(); capa->plain.max_num = capa->max_queues; capa->sched.max_num = capa->max_queues; + capa->plain.lockfree.max_num = queue_glb->queue_lf_num; + capa->plain.lockfree.max_size = queue_glb->queue_lf_size;
return 0; } @@ -188,6 +204,7 @@ static odp_queue_t queue_create(const char *name, { uint32_t i; queue_entry_t *queue; + void *queue_lf; odp_queue_t handle = ODP_QUEUE_INVALID; odp_queue_type_t type = ODP_QUEUE_TYPE_PLAIN; odp_queue_param_t default_param; @@ -198,7 +215,7 @@ static odp_queue_t queue_create(const char *name, }
for (i = 0; i < ODP_CONFIG_QUEUES; i++) { - queue = &queue_tbl->queue[i]; + queue = &queue_glb->queue[i];
if (queue->s.status != QUEUE_STATUS_FREE) continue; @@ -207,7 +224,26 @@ static odp_queue_t queue_create(const char *name, if (queue->s.status == QUEUE_STATUS_FREE) { if (queue_init(queue, name, param)) { UNLOCK(&queue->s.lock); - return handle; + return ODP_QUEUE_INVALID; + } + + if (param->nonblocking == ODP_NONBLOCKING_LF) { + queue_lf_func_t *lf_func; + + lf_func = &queue_glb->queue_lf_func; + + queue_lf = queue_lf_create(queue); + + if (queue_lf == NULL) { + UNLOCK(&queue->s.lock); + return ODP_QUEUE_INVALID; + } + queue->s.queue_lf = queue_lf; + + queue->s.enqueue = lf_func->enq; + queue->s.enqueue_multi = lf_func->enq_multi; + queue->s.dequeue = lf_func->deq; + queue->s.dequeue_multi = lf_func->deq_multi; }
type = queue->s.type; @@ -224,7 +260,10 @@ static odp_queue_t queue_create(const char *name, UNLOCK(&queue->s.lock); }
- if (handle != ODP_QUEUE_INVALID && type == ODP_QUEUE_TYPE_SCHED) { + if (handle == ODP_QUEUE_INVALID) + return ODP_QUEUE_INVALID; + + if (type == ODP_QUEUE_TYPE_SCHED) { if (sched_fn->init_queue(queue->s.index, &queue->s.param.sched)) { queue->s.status = QUEUE_STATUS_FREE; @@ -289,6 +328,10 @@ static int queue_destroy(odp_queue_t handle) default: ODP_ABORT("Unexpected queue status\n"); } + + if (queue->s.param.nonblocking == ODP_NONBLOCKING_LF) + queue_lf_destroy(queue->s.queue_lf); + UNLOCK(&queue->s.lock);
return 0; @@ -313,7 +356,7 @@ static odp_queue_t queue_lookup(const char *name) uint32_t i;
for (i = 0; i < ODP_CONFIG_QUEUES; i++) { - queue_entry_t *queue = &queue_tbl->queue[i]; + queue_entry_t *queue = &queue_glb->queue[i];
if (queue->s.status == QUEUE_STATUS_FREE || queue->s.status == QUEUE_STATUS_DESTROYED) diff --git a/platform/linux-generic/odp_queue_lf.c b/platform/linux-generic/odp_queue_lf.c new file mode 100644 index 00000000..9f509082 --- /dev/null +++ b/platform/linux-generic/odp_queue_lf.c @@ -0,0 +1,346 @@ +/* Copyright (c) 2018, Linaro Limited + * All rights reserved. + * + * SPDX-License-Identifier: BSD-3-Clause + */ + +#include <odp/api/queue.h> +#include <odp/api/atomic.h> +#include <odp/api/shared_memory.h> +#include <odp_queue_lf.h> +#include <string.h> +#include <stdio.h> + +#include "config.h" +#include <odp_debug_internal.h> + +#define RING_LF_SIZE 32 +#define QUEUE_LF_NUM 128 +#define ENQ_RETRIES (RING_LF_SIZE / 4) +#define DEQ_RETRIES (RING_LF_SIZE / 8) + +#ifdef __SIZEOF_INT128__ + +typedef unsigned __int128 u128_t; + +static inline u128_t atomic_load_u128(u128_t *atomic) +{ + return __atomic_load_n(atomic, __ATOMIC_RELAXED); +} + +static inline void atomic_zero_u128(u128_t *atomic) +{ + __atomic_store_n(atomic, 0, __ATOMIC_RELAXED); +} + +static inline int atomic_cas_rel_u128(u128_t *atomic, u128_t old_val, + u128_t new_val) +{ + return __atomic_compare_exchange_n(atomic, &old_val, new_val, + 0 /* strong */, + __ATOMIC_RELEASE, + __ATOMIC_RELAXED); +} + +static inline int atomic_cas_acq_u128(u128_t *atomic, u128_t old_val, + u128_t new_val) +{ + return __atomic_compare_exchange_n(atomic, &old_val, new_val, + 0 /* strong */, + __ATOMIC_ACQUIRE, + __ATOMIC_RELAXED); +} + +#else + +/* These definitions enable build in non 128 bit compatible systems. + * Implementation is active only when 128 bit lockfree atomics are available. + * So, these are never actually used. */ +typedef struct { + uint64_t u64[2]; +} u128_t ODP_ALIGNED(16); + +static inline u128_t atomic_load_u128(u128_t *atomic) +{ + return *atomic; +} + +static inline void atomic_zero_u128(u128_t *atomic) +{ + atomic->u64[0] = 0; + atomic->u64[1] = 0; +} + +static inline int atomic_cas_rel_u128(u128_t *atomic, u128_t old_val, + u128_t new_val) +{ + (void)old_val; + *atomic = new_val; + return 1; +} + +static inline int atomic_cas_acq_u128(u128_t *atomic, u128_t old_val, + u128_t new_val) +{ + return atomic_cas_rel_u128(atomic, old_val, new_val); +} + +#endif + +/* Node in lock-free ring */ +typedef union { + u128_t u128; + + struct { + /* 0: empty, 1: data */ + uint64_t mark: 1; + + /* A cache aligned pointer fits into 63 bits, since the least + * significant bits are zero. */ + uint64_t ptr: 63; + + /* Data with lowest counter value is the head. */ + uint64_t count; + } s; + +} ring_lf_node_t; + +/* Lock-free ring */ +typedef struct { + ring_lf_node_t node[RING_LF_SIZE]; + int used; + odp_atomic_u64_t enq_counter; + +} queue_lf_t ODP_ALIGNED_CACHE; + +/* Lock-free queue globals */ +typedef struct { + queue_lf_t queue_lf[QUEUE_LF_NUM]; + odp_shm_t shm; + +} queue_lf_global_t ODP_ALIGNED_CACHE; + +static queue_lf_global_t *queue_lf_glb; + +static inline int next_idx(int idx) +{ + int next = idx + 1; + + if (next == RING_LF_SIZE) + next = 0; + + return next; +} + +static int queue_lf_enq(queue_t q_int, odp_buffer_hdr_t *buf_hdr) +{ + queue_entry_t *queue; + queue_lf_t *queue_lf; + int i, j, i_node; + int found; + ring_lf_node_t node_val; + ring_lf_node_t new_val; + ring_lf_node_t *node; + uint64_t counter; + + queue = qentry_from_int(q_int); + queue_lf = queue->s.queue_lf; + + i_node = 0; + + counter = odp_atomic_fetch_inc_u64(&queue_lf->enq_counter); + + for (j = 0; j < ENQ_RETRIES; j++) { + found = 0; + + /* Find empty node */ + for (i = 0; i < RING_LF_SIZE; i++) { + i_node = next_idx(i_node); + node = &queue_lf->node[i_node]; + node_val.u128 = atomic_load_u128(&node->u128); + + if (node_val.s.mark == 0) { + found = 1; + break; + } + } + + /* Queue is full */ + if (found == 0) + return -1; + + /* Try to insert data */ + new_val.s.mark = 1; + new_val.s.count = counter; + new_val.s.ptr = ((uintptr_t)buf_hdr) >> 1; + + if (atomic_cas_rel_u128(&node->u128, node_val.u128, + new_val.u128)) + return 0; + } + + return -1; +} + +static int queue_lf_enq_multi(queue_t q_int, odp_buffer_hdr_t **buf_hdr, + int num) +{ + (void)num; + + if (queue_lf_enq(q_int, buf_hdr[0]) == 0) + return 1; + + return 0; +} + +static odp_buffer_hdr_t *queue_lf_deq(queue_t q_int) +{ + queue_entry_t *queue; + queue_lf_t *queue_lf; + int i, j, i_node; + int found; + ring_lf_node_t node_val, old_val, new_val; + ring_lf_node_t *node, *old; + uint64_t lowest; + + queue = qentry_from_int(q_int); + queue_lf = queue->s.queue_lf; + i_node = 0; + + for (j = 0; j < DEQ_RETRIES; j++) { + found = 0; + lowest = -1; + + /* Find the head node. The one with data and + * the lowest counter. */ + for (i = 0; i < RING_LF_SIZE; i++) { + i_node = next_idx(i_node); + node = &queue_lf->node[i_node]; + node_val.u128 = atomic_load_u128(&node->u128); + + if (node_val.s.mark == 1 && node_val.s.count < lowest) { + old = node; + old_val.u128 = node_val.u128; + lowest = node_val.s.count; + found = 1; + } + } + + /* Queue is empty */ + if (found == 0) + return NULL; + + /* Try to remove data */ + new_val.u128 = old_val.u128; + new_val.s.mark = 0; + + if (atomic_cas_acq_u128(&old->u128, old_val.u128, + new_val.u128)) + return (void *)(((uintptr_t)old_val.s.ptr) << 1); + } + + return NULL; +} + +static int queue_lf_deq_multi(queue_t q_int, odp_buffer_hdr_t **buf_hdr, + int num) +{ + odp_buffer_hdr_t *buf; + + (void)num; + + buf = queue_lf_deq(q_int); + + if (buf == NULL) + return 0; + + buf_hdr[0] = buf; + return 1; +} + +uint32_t queue_lf_init_global(uint32_t *queue_lf_size, + queue_lf_func_t *lf_func) +{ + odp_shm_t shm; + bool lockfree = 0; + + /* 16 byte lockfree CAS operation is needed. */ +#ifdef __SIZEOF_INT128__ + lockfree = __atomic_is_lock_free(16, NULL); +#endif + + ODP_DBG("\nLock-free queue init\n"); + ODP_DBG(" u128 lock-free: %i\n\n", lockfree); + + if (!lockfree) + return 0; + + shm = odp_shm_reserve("odp_queues_lf", sizeof(queue_lf_global_t), + ODP_CACHE_LINE_SIZE, 0); + + queue_lf_glb = odp_shm_addr(shm); + memset(queue_lf_glb, 0, sizeof(queue_lf_global_t)); + + queue_lf_glb->shm = shm; + + memset(lf_func, 0, sizeof(queue_lf_func_t)); + lf_func->enq = queue_lf_enq; + lf_func->enq_multi = queue_lf_enq_multi; + lf_func->deq = queue_lf_deq; + lf_func->deq_multi = queue_lf_deq_multi; + + *queue_lf_size = RING_LF_SIZE; + + return QUEUE_LF_NUM; +} + +void queue_lf_term_global(void) +{ + odp_shm_t shm; + + if (queue_lf_glb == NULL) + return; + + shm = queue_lf_glb->shm; + + if (odp_shm_free(shm) < 0) + ODP_ERR("shm free failed"); +} + +static void init_queue(queue_lf_t *queue_lf) +{ + int i; + + odp_atomic_init_u64(&queue_lf->enq_counter, 0); + + for (i = 0; i < RING_LF_SIZE; i++) + atomic_zero_u128(&queue_lf->node[i].u128); +} + +void *queue_lf_create(queue_entry_t *queue) +{ + int i; + queue_lf_t *queue_lf = NULL; + + if (queue->s.type != ODP_QUEUE_TYPE_PLAIN) + return NULL; + + for (i = 0; i < QUEUE_LF_NUM; i++) { + if (queue_lf_glb->queue_lf[i].used == 0) { + queue_lf = &queue_lf_glb->queue_lf[i]; + memset(queue_lf, 0, sizeof(queue_lf_t)); + init_queue(queue_lf); + queue_lf->used = 1; + break; + } + } + + return queue_lf; +} + +void queue_lf_destroy(void *queue_lf_ptr) +{ + queue_lf_t *queue_lf = queue_lf_ptr; + + queue_lf->used = 0; +}
commit 614773850d4d572b793ce2db6cb0cd6c94dada3b Author: Petri Savolainen petri.savolainen@linaro.org Date: Wed Jan 10 15:44:00 2018 +0200
validation: queue: multi-thread plain queue test
Test plain queue enqueue and dequeue with multiple concurrent threads. Test blocking and non-blocking lock-free implementations.
Signed-off-by: Petri Savolainen petri.savolainen@linaro.org Reviewed-by: Bill Fischofer bill.fischofer@linaro.org Signed-off-by: Maxim Uvarov maxim.uvarov@linaro.org
diff --git a/test/validation/api/queue/queue.c b/test/validation/api/queue/queue.c index 1ff02917..59a917c0 100644 --- a/test/validation/api/queue/queue.c +++ b/test/validation/api/queue/queue.c @@ -14,6 +14,22 @@ #define MAX_NUM_EVENT (1 * 1024) #define MAX_ITERATION (100) #define MAX_QUEUES (64 * 1024) +#define GLOBALS_NAME "queue_test_globals" +#define DEQ_RETRIES 100 +#define ENQ_RETRIES 100 + +typedef struct { + pthrd_arg cu_thr; + int num_workers; + odp_barrier_t barrier; + odp_queue_t queue; + odp_atomic_u32_t num_event; + + struct { + uint32_t num_event; + } thread[ODP_THREAD_COUNT_MAX]; + +} test_globals_t;
static int queue_context = 0xff; static odp_pool_t pool; @@ -31,7 +47,30 @@ static void generate_name(char *name, uint32_t index)
int queue_suite_init(void) { + odp_shm_t shm; + test_globals_t *globals; odp_pool_param_t params; + int num_workers; + odp_cpumask_t mask; + + shm = odp_shm_reserve(GLOBALS_NAME, sizeof(test_globals_t), + ODP_CACHE_LINE_SIZE, 0); + + if (shm == ODP_SHM_INVALID) { + printf("Shared memory reserve failed\n"); + return -1; + } + + globals = odp_shm_addr(shm); + memset(globals, 0, sizeof(test_globals_t)); + + num_workers = odp_cpumask_default_worker(&mask, 0); + + if (num_workers > MAX_WORKERS) + num_workers = MAX_WORKERS; + + globals->num_workers = num_workers; + odp_barrier_init(&globals->barrier, num_workers);
odp_pool_param_init(¶ms);
@@ -51,7 +90,25 @@ int queue_suite_init(void)
int queue_suite_term(void) { - return odp_pool_destroy(pool); + odp_shm_t shm; + + shm = odp_shm_lookup(GLOBALS_NAME); + if (shm == ODP_SHM_INVALID) { + printf("SHM lookup failed.\n"); + return -1; + } + + if (odp_shm_free(shm)) { + printf("SHM free failed.\n"); + return -1; + } + + if (odp_pool_destroy(pool)) { + printf("Pool destroy failed.\n"); + return -1; + } + + return 0; }
void queue_test_capa(void) @@ -411,12 +468,211 @@ void queue_test_info(void) CU_ASSERT(odp_queue_destroy(q_order) == 0); }
+static uint32_t alloc_and_enqueue(odp_queue_t queue, odp_pool_t pool, + uint32_t num) +{ + uint32_t i, ret; + odp_buffer_t buf; + odp_event_t ev; + + for (i = 0; i < num; i++) { + buf = odp_buffer_alloc(pool); + + CU_ASSERT(buf != ODP_BUFFER_INVALID); + + ev = odp_buffer_to_event(buf); + + ret = odp_queue_enq(queue, ev); + + CU_ASSERT(ret == 0); + + if (ret) + break; + } + + return i; +} + +static uint32_t dequeue_and_free_all(odp_queue_t queue) +{ + odp_event_t ev; + uint32_t num, retries; + + num = 0; + retries = 0; + + while (1) { + ev = odp_queue_deq(queue); + + if (ev == ODP_EVENT_INVALID) { + if (retries >= DEQ_RETRIES) + return num; + + retries++; + continue; + } + + retries = 0; + num++; + + odp_event_free(ev); + } + + return num; +} + +static int enqueue_with_retry(odp_queue_t queue, odp_event_t ev) +{ + int i; + + for (i = 0; i < ENQ_RETRIES; i++) + if (odp_queue_enq(queue, ev) == 0) + return 0; + + return -1; +} + +static int queue_test_worker(void *arg) +{ + uint32_t num, retries, num_workers; + int thr_id, ret; + odp_event_t ev; + odp_queue_t queue; + test_globals_t *globals = arg; + + thr_id = odp_thread_id(); + queue = globals->queue; + num_workers = globals->num_workers; + + if (num_workers > 1) + odp_barrier_wait(&globals->barrier); + + retries = 0; + num = odp_atomic_fetch_inc_u32(&globals->num_event); + + /* On average, each worker deq-enq each event once */ + while (num < (num_workers * MAX_NUM_EVENT)) { + ev = odp_queue_deq(queue); + + if (ev == ODP_EVENT_INVALID) { + if (retries < DEQ_RETRIES) { + retries++; + continue; + } + + /* Prevent thread to starve */ + num = odp_atomic_fetch_inc_u32(&globals->num_event); + retries = 0; + continue; + } + + globals->thread[thr_id].num_event++; + + ret = enqueue_with_retry(queue, ev); + + CU_ASSERT(ret == 0); + + num = odp_atomic_fetch_inc_u32(&globals->num_event); + } + + return 0; +} + +static void reset_thread_stat(test_globals_t *globals) +{ + int i; + + odp_atomic_init_u32(&globals->num_event, 0); + + for (i = 0; i < ODP_THREAD_COUNT_MAX; i++) + globals->thread[i].num_event = 0; +} + +static void multithread_test(odp_nonblocking_t nonblocking) +{ + odp_shm_t shm; + test_globals_t *globals; + odp_queue_t queue; + odp_queue_param_t qparams; + odp_queue_capability_t capa; + uint32_t queue_size, max_size; + uint32_t num, sum, num_free, i; + + CU_ASSERT(odp_queue_capability(&capa) == 0); + + queue_size = 2 * MAX_NUM_EVENT; + + max_size = capa.plain.max_size; + + if (nonblocking == ODP_NONBLOCKING_LF) { + if (capa.plain.lockfree.max_num == 0) + return; + + max_size = capa.plain.lockfree.max_size; + } + + if (max_size && queue_size > max_size) + queue_size = max_size; + + num = MAX_NUM_EVENT; + + if (num > queue_size) + num = queue_size / 2; + + shm = odp_shm_lookup(GLOBALS_NAME); + CU_ASSERT_FATAL(shm != ODP_SHM_INVALID); + + globals = odp_shm_addr(shm); + globals->cu_thr.numthrds = globals->num_workers; + + odp_queue_param_init(&qparams); + qparams.type = ODP_QUEUE_TYPE_PLAIN; + qparams.size = queue_size; + qparams.nonblocking = nonblocking; + + queue = odp_queue_create("queue_test_mt", &qparams); + CU_ASSERT_FATAL(queue != ODP_QUEUE_INVALID); + + globals->queue = queue; + reset_thread_stat(globals); + + CU_ASSERT(alloc_and_enqueue(queue, pool, num) == num); + + odp_cunit_thread_create(queue_test_worker, (pthrd_arg *)globals); + + /* Wait for worker threads to terminate */ + odp_cunit_thread_exit((pthrd_arg *)globals); + + sum = 0; + for (i = 0; i < ODP_THREAD_COUNT_MAX; i++) + sum += globals->thread[i].num_event; + + CU_ASSERT(sum != 0); + + num_free = dequeue_and_free_all(queue); + + CU_ASSERT(num_free == num); + CU_ASSERT(odp_queue_destroy(queue) == 0); +} + +static void queue_test_mt_plain_block(void) +{ + multithread_test(ODP_BLOCKING); +} + +static void queue_test_mt_plain_nonblock_lf(void) +{ + multithread_test(ODP_NONBLOCKING_LF); +} + odp_testinfo_t queue_suite[] = { ODP_TEST_INFO(queue_test_capa), ODP_TEST_INFO(queue_test_mode), ODP_TEST_INFO(queue_test_lockfree), ODP_TEST_INFO(queue_test_param), ODP_TEST_INFO(queue_test_info), + ODP_TEST_INFO(queue_test_mt_plain_block), + ODP_TEST_INFO(queue_test_mt_plain_nonblock_lf), ODP_TEST_INFO_NULL, };
commit 6a2d72b399388d7d29cba4f60bf3aea3f1ff555c Author: Petri Savolainen petri.savolainen@linaro.org Date: Wed Dec 20 13:10:55 2017 +0200
validation: queue: test lock-free queue
Added simple, single threaded enqueue/dequeue test for lock-free queues.
Signed-off-by: Petri Savolainen petri.savolainen@linaro.org Reviewed-by: Bill Fischofer bill.fischofer@linaro.org Signed-off-by: Maxim Uvarov maxim.uvarov@linaro.org
diff --git a/test/validation/api/queue/queue.c b/test/validation/api/queue/queue.c index f674b081..1ff02917 100644 --- a/test/validation/api/queue/queue.c +++ b/test/validation/api/queue/queue.c @@ -10,9 +10,9 @@ #include <odp_cunit_common.h> #include "queue.h"
-#define MAX_BUFFER_QUEUE (8) -#define MSG_POOL_SIZE (4 * 1024 * 1024) -#define CONFIG_MAX_ITERATION (100) +#define BURST_SIZE (8) +#define MAX_NUM_EVENT (1 * 1024) +#define MAX_ITERATION (100) #define MAX_QUEUES (64 * 1024)
static int queue_context = 0xff; @@ -33,9 +33,11 @@ int queue_suite_init(void) { odp_pool_param_t params;
- params.buf.size = 0; + odp_pool_param_init(¶ms); + + params.buf.size = 4; params.buf.align = ODP_CACHE_LINE_SIZE; - params.buf.num = 1024 * 10; + params.buf.num = MAX_NUM_EVENT; params.type = ODP_POOL_BUFFER;
pool = odp_pool_create("msg_pool", ¶ms); @@ -153,18 +155,105 @@ void queue_test_mode(void) } }
+static odp_event_t dequeue_event(odp_queue_t queue) +{ + odp_event_t ev; + int i; + + for (i = 0; i < MAX_ITERATION; i++) { + ev = odp_queue_deq(queue); + if (ev != ODP_EVENT_INVALID) + break; + } + + return ev; +} + +void queue_test_lockfree(void) +{ + odp_queue_param_t param; + odp_queue_t queue; + odp_queue_capability_t capa; + uint32_t max_burst, burst, i, j; + odp_pool_t pool; + odp_buffer_t buf; + odp_event_t ev; + uint32_t *data; + + CU_ASSERT_FATAL(odp_queue_capability(&capa) == 0); + + if (capa.plain.lockfree.max_num == 0) + return; + + max_burst = capa.plain.lockfree.max_size; + + if (max_burst == 0 || max_burst > MAX_NUM_EVENT) + max_burst = MAX_NUM_EVENT; + + pool = odp_pool_lookup("msg_pool"); + CU_ASSERT_FATAL(pool != ODP_POOL_INVALID); + + odp_queue_param_init(¶m); + param.type = ODP_QUEUE_TYPE_PLAIN; + param.nonblocking = ODP_NONBLOCKING_LF; + param.size = max_burst; + + queue = odp_queue_create("lockfree_queue", ¶m); + CU_ASSERT_FATAL(queue != ODP_QUEUE_INVALID); + + CU_ASSERT(odp_queue_deq(queue) == ODP_EVENT_INVALID); + + buf = odp_buffer_alloc(pool); + CU_ASSERT_FATAL(buf != ODP_BUFFER_INVALID); + ev = odp_buffer_to_event(buf); + CU_ASSERT(odp_queue_enq(queue, ev) == 0); + ev = dequeue_event(queue); + CU_ASSERT_FATAL(ev != ODP_EVENT_INVALID); + if (ev != ODP_EVENT_INVALID) + odp_event_free(ev); + + for (j = 0; j < 2; j++) { + if (j == 0) + burst = max_burst / 4; + else + burst = max_burst; + + for (i = 0; i < burst; i++) { + buf = odp_buffer_alloc(pool); + CU_ASSERT_FATAL(buf != ODP_BUFFER_INVALID); + data = odp_buffer_addr(buf); + *data = i; + ev = odp_buffer_to_event(buf); + CU_ASSERT(odp_queue_enq(queue, ev) == 0); + } + + for (i = 0; i < burst; i++) { + ev = dequeue_event(queue); + CU_ASSERT(ev != ODP_EVENT_INVALID); + if (ev != ODP_EVENT_INVALID) { + buf = odp_buffer_from_event(ev); + data = odp_buffer_addr(buf); + CU_ASSERT(*data == i); + odp_event_free(ev); + } + } + } + + CU_ASSERT(odp_queue_destroy(queue) == 0); +} + void queue_test_param(void) { odp_queue_t queue, null_queue; - odp_event_t enev[MAX_BUFFER_QUEUE]; - odp_event_t deev[MAX_BUFFER_QUEUE]; + odp_event_t enev[BURST_SIZE]; + odp_event_t deev[BURST_SIZE]; odp_buffer_t buf; odp_event_t ev; odp_pool_t msg_pool; odp_event_t *pev_tmp; int i, deq_ret, ret; int nr_deq_entries = 0; - int max_iteration = CONFIG_MAX_ITERATION; + int max_iteration = MAX_ITERATION; odp_queue_param_t qparams; odp_buffer_t enbuf;
@@ -223,7 +312,7 @@ void queue_test_param(void) odp_buffer_free(buf); }
- for (i = 0; i < MAX_BUFFER_QUEUE; i++) { + for (i = 0; i < BURST_SIZE; i++) { buf = odp_buffer_alloc(msg_pool); enev[i] = odp_buffer_to_event(buf); } @@ -233,23 +322,22 @@ void queue_test_param(void) * constraints in the implementation at that given point of time. * But here we assume that we succeed in enqueuing all buffers. */ - ret = odp_queue_enq_multi(queue, enev, MAX_BUFFER_QUEUE); - CU_ASSERT(MAX_BUFFER_QUEUE == ret); + ret = odp_queue_enq_multi(queue, enev, BURST_SIZE); + CU_ASSERT(BURST_SIZE == ret); i = ret < 0 ? 0 : ret; - for ( ; i < MAX_BUFFER_QUEUE; i++) + for ( ; i < BURST_SIZE; i++) odp_event_free(enev[i]);
pev_tmp = deev; do { - deq_ret = odp_queue_deq_multi(queue, pev_tmp, - MAX_BUFFER_QUEUE); + deq_ret = odp_queue_deq_multi(queue, pev_tmp, BURST_SIZE); nr_deq_entries += deq_ret; max_iteration--; pev_tmp += deq_ret; CU_ASSERT(max_iteration >= 0); - } while (nr_deq_entries < MAX_BUFFER_QUEUE); + } while (nr_deq_entries < BURST_SIZE);
- for (i = 0; i < MAX_BUFFER_QUEUE; i++) { + for (i = 0; i < BURST_SIZE; i++) { enbuf = odp_buffer_from_event(enev[i]); CU_ASSERT(enev[i] == deev[i]); odp_buffer_free(enbuf); @@ -326,6 +414,7 @@ void queue_test_info(void) odp_testinfo_t queue_suite[] = { ODP_TEST_INFO(queue_test_capa), ODP_TEST_INFO(queue_test_mode), + ODP_TEST_INFO(queue_test_lockfree), ODP_TEST_INFO(queue_test_param), ODP_TEST_INFO(queue_test_info), ODP_TEST_INFO_NULL, diff --git a/test/validation/api/queue/queue.h b/test/validation/api/queue/queue.h index 6b787b1d..b8509070 100644 --- a/test/validation/api/queue/queue.h +++ b/test/validation/api/queue/queue.h @@ -12,6 +12,7 @@ /* test functions: */ void queue_test_capa(void); void queue_test_mode(void); +void queue_test_lockfree(void); void queue_test_param(void); void queue_test_info(void);
commit e8ec57b012fea43ae501adb66f5c8ee2f03d0a6c Author: Petri Savolainen petri.savolainen@linaro.org Date: Tue Dec 12 15:52:38 2017 +0200
api: queue: block-free capabilities
Lock-free and wait-free implementations may differ a lot from the default (blocking) implementation. Thus the maximum number of queues and queue sizes may be more limited. Non-blocking enum is not needed anymore as capability, since number of queues may be zero for LF/WF when not implemented.
Signed-off-by: Petri Savolainen petri.savolainen@linaro.org Reviewed-by: Bill Fischofer bill.fischofer@linaro.org Signed-off-by: Maxim Uvarov maxim.uvarov@linaro.org
diff --git a/include/odp/api/spec/queue.h b/include/odp/api/spec/queue.h index 79a26df3..107ef3f7 100644 --- a/include/odp/api/spec/queue.h +++ b/include/odp/api/spec/queue.h @@ -164,35 +164,81 @@ typedef struct odp_queue_capability_t {
/** Plain queue capabilities */ struct { - /** Maximum number of plain queues of the default size. */ + /** Maximum number of plain (ODP_BLOCKING) queues of the + * default size. */ uint32_t max_num;
- /** Maximum number of events a plain queue can store - * simultaneously. The value of zero means that plain + /** Maximum number of events a plain (ODP_BLOCKING) queue can + * store simultaneously. The value of zero means that plain * queues do not have a size limit, but a single queue can * store all available events. */ uint32_t max_size;
- /** The strongest guarantee of block freedom that is supported - * for plain queues. */ - odp_nonblocking_t nonblocking; + /** Lock-free (ODP_NONBLOCKING_LF) implementation capabilities. + * The specification is the same as for the blocking + * implementation. */ + struct { + /** Maximum number of queues. Lock-free queues are not + * supported when zero. */ + uint32_t max_num; + + /** Maximum queue size */ + uint32_t max_size; + + } lockfree; + + /** Wait-free (ODP_NONBLOCKING_WF) implementation capabilities. + * The specification is the same as for the blocking + * implementation. */ + struct { + /** Maximum number of queues. Wait-free queues are not + * supported when zero. */ + uint32_t max_num; + + /** Maximum queue size */ + uint32_t max_size; + + } waitfree;
} plain;
/** Scheduled queue capabilities */ struct { - /** Maximum number of scheduled queues of the default size. */ + /** Maximum number of scheduled (ODP_BLOCKING) queues of the + * default size. */ uint32_t max_num;
- /** Maximum number of events a scheduled queue can store - * simultaneously. The value of zero means that scheduled - * queues do not have a size limit, but a single queue can - * store all available events. */ + /** Maximum number of events a scheduled (ODP_BLOCKING) queue + * can store simultaneously. The value of zero means that + * scheduled queues do not have a size limit, but a single + * queue can store all available events. */ uint32_t max_size;
- /** The strongest guarantee of block freedom that is supported - * for scheduled queues. */ - odp_nonblocking_t nonblocking; + /** Lock-free (ODP_NONBLOCKING_LF) implementation capabilities. + * The specification is the same as for the blocking + * implementation. */ + struct { + /** Maximum number of queues. Lock-free queues are not + * supported when zero. */ + uint32_t max_num; + + /** Maximum queue size */ + uint32_t max_size; + + } lockfree; + + /** Wait-free (ODP_NONBLOCKING_WF) implementation capabilities. + * The specification is the same as for the blocking + * implementation. */ + struct { + /** Maximum number of queues. Wait-free queues are not + * supported when zero. */ + uint32_t max_num; + + /** Maximum queue size */ + uint32_t max_size; + + } waitfree;
} sched;
diff --git a/platform/linux-generic/odp_queue.c b/platform/linux-generic/odp_queue.c index 58103930..6a69eb84 100644 --- a/platform/linux-generic/odp_queue.c +++ b/platform/linux-generic/odp_queue.c @@ -150,9 +150,7 @@ static int queue_capability(odp_queue_capability_t *capa) capa->max_sched_groups = sched_fn->num_grps(); capa->sched_prios = odp_schedule_num_prio(); capa->plain.max_num = capa->max_queues; - capa->plain.nonblocking = ODP_BLOCKING; capa->sched.max_num = capa->max_queues; - capa->sched.nonblocking = ODP_BLOCKING;
return 0; } diff --git a/platform/linux-generic/odp_queue_scalable.c b/platform/linux-generic/odp_queue_scalable.c index 88a5a8c2..72dec15b 100644 --- a/platform/linux-generic/odp_queue_scalable.c +++ b/platform/linux-generic/odp_queue_scalable.c @@ -315,10 +315,8 @@ static int queue_capability(odp_queue_capability_t *capa) capa->sched_prios = odp_schedule_num_prio(); capa->plain.max_num = ODP_CONFIG_QUEUES - NUM_INTERNAL_QUEUES; capa->plain.max_size = 0; - capa->plain.nonblocking = ODP_BLOCKING; capa->sched.max_num = ODP_CONFIG_QUEUES - NUM_INTERNAL_QUEUES; capa->sched.max_size = 0; - capa->sched.nonblocking = ODP_BLOCKING;
return 0; } diff --git a/test/validation/api/queue/queue.c b/test/validation/api/queue/queue.c index 3c6db33a..f674b081 100644 --- a/test/validation/api/queue/queue.c +++ b/test/validation/api/queue/queue.c @@ -69,8 +69,6 @@ void queue_test_capa(void) CU_ASSERT(capa.sched_prios != 0); CU_ASSERT(capa.plain.max_num != 0); CU_ASSERT(capa.sched.max_num != 0); - CU_ASSERT(capa.plain.nonblocking >= ODP_BLOCKING); - CU_ASSERT(capa.sched.nonblocking >= ODP_BLOCKING);
min = capa.plain.max_num; if (min > capa.sched.max_num)
-----------------------------------------------------------------------
Summary of changes: include/odp/api/spec/queue.h | 74 +++- platform/linux-generic/Makefile.am | 2 + .../linux-generic/include/odp_queue_internal.h | 1 + platform/linux-generic/include/odp_queue_lf.h | 36 ++ platform/linux-generic/odp_queue.c | 73 +++- platform/linux-generic/odp_queue_lf.c | 346 +++++++++++++++++++ platform/linux-generic/odp_queue_scalable.c | 2 - test/validation/api/queue/queue.c | 381 ++++++++++++++++++++- test/validation/api/queue/queue.h | 1 + 9 files changed, 865 insertions(+), 51 deletions(-) create mode 100644 platform/linux-generic/include/odp_queue_lf.h create mode 100644 platform/linux-generic/odp_queue_lf.c
hooks/post-receive
-
git＠git-us.linaro.org