From patchwork Thu Jul 1 12:26:40 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Yunsheng Lin X-Patchwork-Id: 12354045 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 8BE91C11F64 for ; Thu, 1 Jul 2021 12:27:23 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 749E5613FE for ; Thu, 1 Jul 2021 12:27:23 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S236426AbhGAM3w (ORCPT ); Thu, 1 Jul 2021 08:29:52 -0400 Received: from szxga01-in.huawei.com ([45.249.212.187]:13049 "EHLO szxga01-in.huawei.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S236407AbhGAM3w (ORCPT ); Thu, 1 Jul 2021 08:29:52 -0400 Received: from dggemv703-chm.china.huawei.com (unknown [172.30.72.57]) by szxga01-in.huawei.com (SkyGuard) with ESMTP id 4GFy7W34K8zZjZl; Thu, 1 Jul 2021 20:24:07 +0800 (CST) Received: from dggpemm500005.china.huawei.com (7.185.36.74) by dggemv703-chm.china.huawei.com (10.3.19.46) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 Received: from localhost.localdomain (10.69.192.56) by dggpemm500005.china.huawei.com (7.185.36.74) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 From: Yunsheng Lin To: , , , CC: , , , , , , , , Subject: [PATCH net-next v3 1/3] selftests/ptr_ring: add benchmark application for ptr_ring Date: Thu, 1 Jul 2021 20:26:40 +0800 Message-ID: <1625142402-64945-2-git-send-email-linyunsheng@huawei.com> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> References: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> MIME-Version: 1.0 X-Originating-IP: [10.69.192.56] X-ClientProxiedBy: dggems701-chm.china.huawei.com (10.3.19.178) To dggpemm500005.china.huawei.com (7.185.36.74) X-CFilter-Loop: Reflected Precedence: bulk List-ID: X-Mailing-List: linux-kselftest@vger.kernel.org Currently ptr_ring selftest is embedded within the virtio selftest, which involves some specific virtio operation, such as notifying and kicking. As ptr_ring has been used by various subsystems, it deserves it's owner selftest in order to benchmark different usecase of ptr_ring, such as page pool and pfifo_fast qdisc. So add a simple application to benchmark ptr_ring performance. Currently two test mode is supported: Mode 0: Both producing and consuming is done in a single thread, it is called simple test mode in the test app. Mode 1: Producing and consuming is done in different thread concurrently, also known as SPSC(single-producer/ single-consumer) test. The multi-producer/single-consumer test for pfifo_fast case is not added yet, which can be added if using CAS atomic operation to enable lockless multi-producer is proved to be better than using r->producer_lock. Signed-off-by: Yunsheng Lin --- V3: Remove timestamp sampling, use standard C library as much as possible. --- MAINTAINERS | 5 + tools/testing/selftests/ptr_ring/Makefile | 6 + tools/testing/selftests/ptr_ring/ptr_ring_test.c | 224 +++++++++++++++++++++++ tools/testing/selftests/ptr_ring/ptr_ring_test.h | 130 +++++++++++++ 4 files changed, 365 insertions(+) create mode 100644 tools/testing/selftests/ptr_ring/Makefile create mode 100644 tools/testing/selftests/ptr_ring/ptr_ring_test.c create mode 100644 tools/testing/selftests/ptr_ring/ptr_ring_test.h diff --git a/MAINTAINERS b/MAINTAINERS index cc375fd..1227022 100644 --- a/MAINTAINERS +++ b/MAINTAINERS @@ -14847,6 +14847,11 @@ F: drivers/net/phy/dp83640* F: drivers/ptp/* F: include/linux/ptp_cl* +PTR RING BENCHMARK +M: Yunsheng Lin +L: netdev@vger.kernel.org +F: tools/testing/selftests/ptr_ring/ + PTRACE SUPPORT M: Oleg Nesterov S: Maintained diff --git a/tools/testing/selftests/ptr_ring/Makefile b/tools/testing/selftests/ptr_ring/Makefile new file mode 100644 index 0000000..346dea9 --- /dev/null +++ b/tools/testing/selftests/ptr_ring/Makefile @@ -0,0 +1,6 @@ +# SPDX-License-Identifier: GPL-2.0 +LDLIBS = -lpthread + +TEST_GEN_PROGS := ptr_ring_test + +include ../lib.mk diff --git a/tools/testing/selftests/ptr_ring/ptr_ring_test.c b/tools/testing/selftests/ptr_ring/ptr_ring_test.c new file mode 100644 index 0000000..4a5312f --- /dev/null +++ b/tools/testing/selftests/ptr_ring/ptr_ring_test.c @@ -0,0 +1,224 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * Copyright (C) 2021 HiSilicon Limited. + */ + +#include +#include +#include +#include +#include +#include +#include + +#include "ptr_ring_test.h" +#include "../../../../include/linux/ptr_ring.h" + +#define MIN_RING_SIZE 2 +#define MAX_RING_SIZE 10000000 + +static struct ptr_ring ring ____cacheline_aligned_in_smp; + +struct worker_info { + pthread_t tid; + int test_count; + bool error; +}; + +static void *produce_worker(void *arg) +{ + struct worker_info *info = arg; + unsigned long i = 0; + int ret; + + while (++i <= info->test_count) { + while (__ptr_ring_full(&ring)) + cpu_relax(); + + ret = __ptr_ring_produce(&ring, (void *)i); + if (ret) { + fprintf(stderr, "produce failed: %d\n", ret); + info->error = true; + return NULL; + } + } + + info->error = false; + + return NULL; +} + +static void *consume_worker(void *arg) +{ + struct worker_info *info = arg; + unsigned long i = 0; + int *ptr; + + while (++i <= info->test_count) { + while (__ptr_ring_empty(&ring)) + cpu_relax(); + + ptr = __ptr_ring_consume(&ring); + if ((unsigned long)ptr != i) { + fprintf(stderr, "consumer failed, ptr: %lu, i: %lu\n", + (unsigned long)ptr, i); + info->error = true; + return NULL; + } + } + + if (!__ptr_ring_empty(&ring)) { + fprintf(stderr, "ring should be empty, test failed\n"); + info->error = true; + return NULL; + } + + info->error = false; + return NULL; +} + +/* test case for single producer single consumer */ +static void spsc_test(int size, int count) +{ + struct worker_info producer, consumer; + pthread_attr_t attr; + void *res; + int ret; + + ret = ptr_ring_init(&ring, size, 0); + if (ret) { + fprintf(stderr, "init failed: %d\n", ret); + return; + } + + producer.test_count = count; + consumer.test_count = count; + + ret = pthread_attr_init(&attr); + if (ret) { + fprintf(stderr, "pthread attr init failed: %d\n", ret); + goto out; + } + + ret = pthread_create(&producer.tid, &attr, + produce_worker, &producer); + if (ret) { + fprintf(stderr, "create producer thread failed: %d\n", ret); + goto out; + } + + ret = pthread_create(&consumer.tid, &attr, + consume_worker, &consumer); + if (ret) { + fprintf(stderr, "create consumer thread failed: %d\n", ret); + goto out; + } + + ret = pthread_join(producer.tid, &res); + if (ret) { + fprintf(stderr, "join producer thread failed: %d\n", ret); + goto out; + } + + ret = pthread_join(consumer.tid, &res); + if (ret) { + fprintf(stderr, "join consumer thread failed: %d\n", ret); + goto out; + } + + if (producer.error || consumer.error) { + fprintf(stderr, "spsc test failed\n"); + goto out; + } + + printf("ptr_ring(size:%d) perf spsc test produced/comsumed %d items, finished\n", + size, count); +out: + ptr_ring_cleanup(&ring, NULL); +} + +static void simple_test(int size, int count) +{ + struct timeval start, end; + int i = 0; + int *ptr; + int ret; + + ret = ptr_ring_init(&ring, size, 0); + if (ret) { + fprintf(stderr, "init failed: %d\n", ret); + return; + } + + while (++i <= count) { + ret = __ptr_ring_produce(&ring, &count); + if (ret) { + fprintf(stderr, "produce failed: %d\n", ret); + goto out; + } + + ptr = __ptr_ring_consume(&ring); + if (ptr != &count) { + fprintf(stderr, "consume failed: %p\n", ptr); + goto out; + } + } + + printf("ptr_ring(size:%d) perf simple test produced/consumed %d items, finished\n", + size, count); + +out: + ptr_ring_cleanup(&ring, NULL); +} + +int main(int argc, char *argv[]) +{ + int count = 1000000; + int size = 1000; + int mode = 0; + int opt; + + while ((opt = getopt(argc, argv, "N:s:m:h")) != -1) { + switch (opt) { + case 'N': + count = atoi(optarg); + break; + case 's': + size = atoi(optarg); + break; + case 'm': + mode = atoi(optarg); + break; + case 'h': + printf("usage: ptr_ring_test [-N COUNT] [-s RING_SIZE] [-m TEST_MODE]\n"); + return 0; + default: + return -1; + } + } + + if (count <= 0) { + fprintf(stderr, "invalid test count, must be > 0\n"); + return -1; + } + + if (size < MIN_RING_SIZE || size > MAX_RING_SIZE) { + fprintf(stderr, "invalid ring size, must be in %d-%d\n", + MIN_RING_SIZE, MAX_RING_SIZE); + return -1; + } + + switch (mode) { + case 0: + simple_test(size, count); + break; + case 1: + spsc_test(size, count); + break; + default: + fprintf(stderr, "invalid test mode\n"); + return -1; + } + + return 0; +} diff --git a/tools/testing/selftests/ptr_ring/ptr_ring_test.h b/tools/testing/selftests/ptr_ring/ptr_ring_test.h new file mode 100644 index 0000000..32bfefb --- /dev/null +++ b/tools/testing/selftests/ptr_ring/ptr_ring_test.h @@ -0,0 +1,130 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ + +#ifndef _TEST_PTR_RING_TEST_H +#define _TEST_PTR_RING_TEST_H + +#include +#include +#include + +/* Assuming the cache line size is 64 for most cpu, + * change it accordingly if the running cpu has different + * cache line size in order to get more accurate result. + */ +#define SMP_CACHE_BYTES 64 + +#define cpu_relax() sched_yield() +#define smp_release() atomic_thread_fence(memory_order_release) +#define smp_acquire() atomic_thread_fence(memory_order_acquire) +#define smp_wmb() smp_release() +#define smp_store_release(p, v) \ + atomic_store_explicit(p, v, memory_order_release) + +#define READ_ONCE(x) (*(volatile typeof(x) *)&(x)) +#define WRITE_ONCE(x, val) ((*(volatile typeof(x) *)&(x)) = (val)) +#define cache_line_size SMP_CACHE_BYTES +#define unlikely(x) (__builtin_expect(!!(x), 0)) +#define likely(x) (__builtin_expect(!!(x), 1)) +#define ALIGN(x, a) (((x) + (a) - 1) / (a) * (a)) +#define SIZE_MAX (~(size_t)0) +#define KMALLOC_MAX_SIZE SIZE_MAX +#define spinlock_t pthread_spinlock_t +#define gfp_t int +#define __GFP_ZERO 0x1 + +#define ____cacheline_aligned_in_smp \ + __attribute__((aligned(SMP_CACHE_BYTES))) + +static inline void *kmalloc(unsigned int size, gfp_t gfp) +{ + void *p; + + p = memalign(64, size); + if (!p) + return p; + + if (gfp & __GFP_ZERO) + memset(p, 0, size); + + return p; +} + +static inline void *kzalloc(unsigned int size, gfp_t flags) +{ + return kmalloc(size, flags | __GFP_ZERO); +} + +static inline void *kmalloc_array(size_t n, size_t size, gfp_t flags) +{ + if (size != 0 && n > SIZE_MAX / size) + return NULL; + return kmalloc(n * size, flags); +} + +static inline void *kcalloc(size_t n, size_t size, gfp_t flags) +{ + return kmalloc_array(n, size, flags | __GFP_ZERO); +} + +static inline void kfree(void *p) +{ + free(p); +} + +#define kvmalloc_array kmalloc_array +#define kvfree kfree + +static inline void spin_lock_init(spinlock_t *lock) +{ + int r = pthread_spin_init(lock, 0); + + assert(!r); +} + +static inline void spin_lock(spinlock_t *lock) +{ + int ret = pthread_spin_lock(lock); + + assert(!ret); +} + +static inline void spin_unlock(spinlock_t *lock) +{ + int ret = pthread_spin_unlock(lock); + + assert(!ret); +} + +static inline void spin_lock_bh(spinlock_t *lock) +{ + spin_lock(lock); +} + +static inline void spin_unlock_bh(spinlock_t *lock) +{ + spin_unlock(lock); +} + +static inline void spin_lock_irq(spinlock_t *lock) +{ + spin_lock(lock); +} + +static inline void spin_unlock_irq(spinlock_t *lock) +{ + spin_unlock(lock); +} + +static inline void spin_lock_irqsave(spinlock_t *lock, + unsigned long f) +{ + spin_lock(lock); +} + +static inline void spin_unlock_irqrestore(spinlock_t *lock, + unsigned long f) +{ + spin_unlock(lock); +} + +#endif From patchwork Thu Jul 1 12:26:41 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Yunsheng Lin X-Patchwork-Id: 12354051 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,USER_AGENT_GIT autolearn=unavailable autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 82976C11F67 for ; Thu, 1 Jul 2021 12:27:33 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id 6A29061407 for ; Thu, 1 Jul 2021 12:27:33 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S236437AbhGAM3x (ORCPT ); Thu, 1 Jul 2021 08:29:53 -0400 Received: from szxga02-in.huawei.com ([45.249.212.188]:9442 "EHLO szxga02-in.huawei.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S236332AbhGAM3w (ORCPT ); Thu, 1 Jul 2021 08:29:52 -0400 Received: from dggemv711-chm.china.huawei.com (unknown [172.30.72.55]) by szxga02-in.huawei.com (SkyGuard) with ESMTP id 4GFy7W35R8zZpSZ; Thu, 1 Jul 2021 20:24:07 +0800 (CST) Received: from dggpemm500005.china.huawei.com (7.185.36.74) by dggemv711-chm.china.huawei.com (10.1.198.66) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 Received: from localhost.localdomain (10.69.192.56) by dggpemm500005.china.huawei.com (7.185.36.74) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 From: Yunsheng Lin To: , , , CC: , , , , , , , , Subject: [PATCH net-next v3 2/3] ptr_ring: move r->queue[] clearing after r->consumer_head updating Date: Thu, 1 Jul 2021 20:26:41 +0800 Message-ID: <1625142402-64945-3-git-send-email-linyunsheng@huawei.com> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> References: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> MIME-Version: 1.0 X-Originating-IP: [10.69.192.56] X-ClientProxiedBy: dggems701-chm.china.huawei.com (10.3.19.178) To dggpemm500005.china.huawei.com (7.185.36.74) X-CFilter-Loop: Reflected Precedence: bulk List-ID: X-Mailing-List: linux-kselftest@vger.kernel.org Currently r->queue[] clearing is done before r->consumer_head updating, which makes the __ptr_ring_empty() returning false positive result(the ring is non-empty, but __ptr_ring_empty() suggest that it is empty) if the checking is done after the r->queue clearing and before the consumer_head moving forward. Move the r->queue[] clearing after consumer_head moving forward to avoid the above case. As a side effect of above change, a consumer_head checking is avoided for the likely case, and it has noticeable performance improvement when it is tested using the ptr_ring_test selftest added in the previous patch. Tested using the "perf stat -r 1000 ./ptr_ring_test -s 1000 -m 1 -N 100000000", comparing the elapsed time: arch unpatched patched improvement arm64 2.087205 sec 1.888224 sec +9.5% X86 2.6538 sec 2.5422 sec +4.2% Signed-off-by: Yunsheng Lin --- V3: adjust the title and comment log according to disscusion in V2, and update performance data using "perf stat -r". V2: Add performance data. --- include/linux/ptr_ring.h | 25 ++++++++++++++++--------- 1 file changed, 16 insertions(+), 9 deletions(-) diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h index 808f9d3..db9c282 100644 --- a/include/linux/ptr_ring.h +++ b/include/linux/ptr_ring.h @@ -261,8 +261,7 @@ static inline void __ptr_ring_discard_one(struct ptr_ring *r) /* Note: we must keep consumer_head valid at all times for __ptr_ring_empty * to work correctly. */ - int consumer_head = r->consumer_head; - int head = consumer_head++; + int consumer_head = r->consumer_head + 1; /* Once we have processed enough entries invalidate them in * the ring all at once so producer can reuse their space in the ring. @@ -271,19 +270,27 @@ static inline void __ptr_ring_discard_one(struct ptr_ring *r) */ if (unlikely(consumer_head - r->consumer_tail >= r->batch || consumer_head >= r->size)) { + int tail = r->consumer_tail; + + if (unlikely(consumer_head >= r->size)) { + r->consumer_tail = 0; + WRITE_ONCE(r->consumer_head, 0); + } else { + r->consumer_tail = consumer_head; + WRITE_ONCE(r->consumer_head, consumer_head); + } + /* Zero out entries in the reverse order: this way we touch the * cache line that producer might currently be reading the last; * producer won't make progress and touch other cache lines * besides the first one until we write out all entries. */ - while (likely(head >= r->consumer_tail)) - r->queue[head--] = NULL; - r->consumer_tail = consumer_head; - } - if (unlikely(consumer_head >= r->size)) { - consumer_head = 0; - r->consumer_tail = 0; + while (likely(--consumer_head >= tail)) + r->queue[consumer_head] = NULL; + + return; } + /* matching READ_ONCE in __ptr_ring_empty for lockless tests */ WRITE_ONCE(r->consumer_head, consumer_head); } From patchwork Thu Jul 1 12:26:42 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Yunsheng Lin X-Patchwork-Id: 12354047 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-16.8 required=3.0 tests=BAYES_00, HEADER_FROM_DIFFERENT_DOMAINS,INCLUDES_CR_TRAILER,INCLUDES_PATCH, MAILING_LIST_MULTI,SPF_HELO_NONE,SPF_PASS,USER_AGENT_GIT autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id CCCAEC11F69 for ; Thu, 1 Jul 2021 12:27:24 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id B3E0161407 for ; Thu, 1 Jul 2021 12:27:24 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S236447AbhGAM3x (ORCPT ); Thu, 1 Jul 2021 08:29:53 -0400 Received: from szxga03-in.huawei.com ([45.249.212.189]:9334 "EHLO szxga03-in.huawei.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S236423AbhGAM3w (ORCPT ); Thu, 1 Jul 2021 08:29:52 -0400 Received: from dggemv704-chm.china.huawei.com (unknown [172.30.72.54]) by szxga03-in.huawei.com (SkyGuard) with ESMTP id 4GFy6F2RBTz74M0; Thu, 1 Jul 2021 20:23:01 +0800 (CST) Received: from dggpemm500005.china.huawei.com (7.185.36.74) by dggemv704-chm.china.huawei.com (10.3.19.47) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 Received: from localhost.localdomain (10.69.192.56) by dggpemm500005.china.huawei.com (7.185.36.74) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256) id 15.1.2176.2; Thu, 1 Jul 2021 20:27:14 +0800 From: Yunsheng Lin To: , , , CC: , , , , , , , , Subject: [PATCH net-next v3 3/3] ptr_ring: add barrier to ensure the visiblity of r->queue[] Date: Thu, 1 Jul 2021 20:26:42 +0800 Message-ID: <1625142402-64945-4-git-send-email-linyunsheng@huawei.com> X-Mailer: git-send-email 2.7.4 In-Reply-To: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> References: <1625142402-64945-1-git-send-email-linyunsheng@huawei.com> MIME-Version: 1.0 X-Originating-IP: [10.69.192.56] X-ClientProxiedBy: dggems701-chm.china.huawei.com (10.3.19.178) To dggpemm500005.china.huawei.com (7.185.36.74) X-CFilter-Loop: Reflected Precedence: bulk List-ID: X-Mailing-List: linux-kselftest@vger.kernel.org After r->consumer_head is updated in __ptr_ring_discard_one(), r->queue[r->consumer_head] is already cleared in the previous round of __ptr_ring_discard_one(). But there is no guarantee other thread will see the r->queue[r->consumer_head] being NULL because there is no explicit barrier between r->queue[] clearing and r->consumer_head updating. So add two explicit barrier to make sure r->queue[] cleared in __ptr_ring_discard_one() to be visible to other cpu, mainly to make sure the cpu calling the __ptr_ring_empty() will see the correct r->queue[r->consumer_head]. Hopefully the previous and this patch have ensured the correct visibility of r->queue[], so update the comment accordingly about __ptr_ring_empty(). Tested using the "perf stat -r 1000 ./ptr_ring_test -s 1000 -m 1 -N 100000000", comparing the elapsed time: arch unpatched patched improvement arm64 1.888224 sec 1.893673 sec -0.2% X86 2.5422 sec 2.5587 sec -0.6% Reported-by: Michael S. Tsirkin Signed-off-by: Yunsheng Lin --- include/linux/ptr_ring.h | 29 +++++++++++++++++++---------- 1 file changed, 19 insertions(+), 10 deletions(-) diff --git a/include/linux/ptr_ring.h b/include/linux/ptr_ring.h index db9c282..d78aab8 100644 --- a/include/linux/ptr_ring.h +++ b/include/linux/ptr_ring.h @@ -178,15 +178,11 @@ static inline void *__ptr_ring_peek(struct ptr_ring *r) * * NB: This is only safe to call if ring is never resized. * - * However, if some other CPU consumes ring entries at the same time, the value - * returned is not guaranteed to be correct. - * - * In this case - to avoid incorrectly detecting the ring - * as empty - the CPU consuming the ring entries is responsible - * for either consuming all ring entries until the ring is empty, - * or synchronizing with some other CPU and causing it to - * re-test __ptr_ring_empty and/or consume the ring enteries - * after the synchronization point. + * caller might need to use the smp_rmb() to pair with smp_wmb() + * or smp_store_release() in __ptr_ring_discard_one() and smp_wmb() + * in __ptr_ring_produce() to ensure correct ordering between + * __ptr_ring_empty() checking and subsequent operation after + * __ptr_ring_empty() checking. * * Note: callers invoking this in a loop must use a compiler barrier, * for example cpu_relax(). @@ -274,7 +270,12 @@ static inline void __ptr_ring_discard_one(struct ptr_ring *r) if (unlikely(consumer_head >= r->size)) { r->consumer_tail = 0; - WRITE_ONCE(r->consumer_head, 0); + + /* Make sure r->queue[0] ~ r->queue[r->consumer_tail] + * cleared in previous __ptr_ring_discard_one() is + * visible to other cpu. + */ + smp_store_release(&r->consumer_head, 0); } else { r->consumer_tail = consumer_head; WRITE_ONCE(r->consumer_head, consumer_head); @@ -288,6 +289,14 @@ static inline void __ptr_ring_discard_one(struct ptr_ring *r) while (likely(--consumer_head >= tail)) r->queue[consumer_head] = NULL; + if (unlikely(!r->consumer_head)) { + /* Make sure r->queue[r->consumer_tail] ~ + * r->queue[r->size - 1] cleared above is visible to + * other cpu. + */ + smp_wmb(); + } + return; }