* [PATCH] eal/unix: optimize thread creation with glibc
@ 2024-11-02 10:08 David Marchand
2024-11-02 11:32 ` [PATCH v2] " David Marchand
0 siblings, 1 reply; 3+ messages in thread
From: David Marchand @ 2024-11-02 10:08 UTC (permalink / raw)
To: dev; +Cc: thomas, stephen, luca.boccassi, stable, Tyler Retzlaff
Setting the cpu affinity of the child thread from the parent thread is
racy when using pthread_setaffinity_np, as the child thread may start
running and initialize before affinity is set.
On the other hand, setting the cpu affinity from the child thread itself
may fail, so the parent thread waits for the child thread to report
whether this call succeeded.
This synchronisation point resulted in a significant slow down of
rte_thread_create() (as seen in the lcores_autotest unit tests, in OBS
for some ARM systems).
Another option for setting cpu affinity is to use the not portable
pthread_attr_setaffinity_np, but it is not available with musl.
Assume availability by relying on __USE_GNU that is not set with musl.
Fixes: b28c6196b132 ("eal/unix: fix thread creation")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
---
lib/eal/unix/rte_thread.c | 26 ++++++++++++++++++++++++++
1 file changed, 26 insertions(+)
diff --git a/lib/eal/unix/rte_thread.c b/lib/eal/unix/rte_thread.c
index 1b4c73f58e..e42b6c37a2 100644
--- a/lib/eal/unix/rte_thread.c
+++ b/lib/eal/unix/rte_thread.c
@@ -4,6 +4,7 @@
*/
#include <errno.h>
+#include <features.h>
#include <pthread.h>
#include <stdbool.h>
#include <stdlib.h>
@@ -19,6 +20,7 @@ struct eal_tls_key {
pthread_key_t thread_index;
};
+#ifndef __USE_GNU
struct thread_start_context {
rte_thread_func thread_func;
void *thread_args;
@@ -28,6 +30,7 @@ struct thread_start_context {
int wrapper_ret;
bool wrapper_done;
};
+#endif
static int
thread_map_priority_to_os_value(enum rte_thread_priority eal_pri, int *os_pri,
@@ -88,6 +91,7 @@ thread_map_os_priority_to_eal_priority(int policy, int os_pri,
return 0;
}
+#ifndef __USE_GNU
static void *
thread_start_wrapper(void *arg)
{
@@ -113,6 +117,7 @@ thread_start_wrapper(void *arg)
return (void *)(uintptr_t)thread_func(thread_args);
}
+#endif
int
rte_thread_create(rte_thread_t *thread_id,
@@ -126,6 +131,7 @@ rte_thread_create(rte_thread_t *thread_id,
.sched_priority = 0,
};
int policy = SCHED_OTHER;
+#ifndef __USE_GNU
struct thread_start_context ctx = {
.thread_func = thread_func,
.thread_args = args,
@@ -134,6 +140,7 @@ rte_thread_create(rte_thread_t *thread_id,
.wrapper_mutex = PTHREAD_MUTEX_INITIALIZER,
.wrapper_cond = PTHREAD_COND_INITIALIZER,
};
+#endif
if (thread_attr != NULL) {
ret = pthread_attr_init(&attr);
@@ -144,6 +151,16 @@ rte_thread_create(rte_thread_t *thread_id,
attrp = &attr;
+#ifdef __USE_GNU
+ if (CPU_COUNT(&thread_attr->cpuset) > 0) {
+ ret = pthread_attr_setaffinity_np(attrp, sizeof(thread_attr->cpuset),
+ &thread_attr->cpuset);
+ if (ret != 0) {
+ EAL_LOG(DEBUG, "pthread_attr_setaffinity_np failed");
+ goto cleanup;
+ }
+ }
+#endif
/*
* Set the inherit scheduler parameter to explicit,
* otherwise the priority attribute is ignored.
@@ -178,6 +195,14 @@ rte_thread_create(rte_thread_t *thread_id,
}
}
+#ifdef __USE_GNU
+ ret = pthread_create((pthread_t *)&thread_id->opaque_id, attrp,
+ (void *)(void *)thread_func, args);
+ if (ret != 0) {
+ EAL_LOG(DEBUG, "pthread_create failed");
+ goto cleanup;
+ }
+#else /* !__USE_GNU */
ret = pthread_create((pthread_t *)&thread_id->opaque_id, attrp,
thread_start_wrapper, &ctx);
if (ret != 0) {
@@ -193,6 +218,7 @@ rte_thread_create(rte_thread_t *thread_id,
if (ret != 0)
rte_thread_join(*thread_id, NULL);
+#endif /* __USE_GNU */
cleanup:
if (attrp != NULL)
--
2.46.2
^ permalink raw reply [flat|nested] 3+ messages in thread
* [PATCH v2] eal/unix: optimize thread creation with glibc
2024-11-02 10:08 [PATCH] eal/unix: optimize thread creation with glibc David Marchand
@ 2024-11-02 11:32 ` David Marchand
2024-11-02 12:36 ` Luca Boccassi
0 siblings, 1 reply; 3+ messages in thread
From: David Marchand @ 2024-11-02 11:32 UTC (permalink / raw)
To: dev; +Cc: thomas, stephen, luca.boccassi, stable, Tyler Retzlaff
Setting the cpu affinity of the child thread from the parent thread is
racy when using pthread_setaffinity_np, as the child thread may start
running and initialize before affinity is set.
On the other hand, setting the cpu affinity from the child thread itself
may fail, so the parent thread waits for the child thread to report
whether this call succeeded.
This synchronisation point resulted in a significant slow down of
rte_thread_create() (as seen in the lcores_autotest unit tests, in OBS
for some ARM systems).
Another option for setting cpu affinity is to use the not portable
pthread_attr_setaffinity_np, but it is not available with musl.
Assume availability by relying on __USE_GNU that is set with glibc.
Fixes: b28c6196b132 ("eal/unix: fix thread creation")
Cc: stable@dpdk.org
Signed-off-by: David Marchand <david.marchand@redhat.com>
---
Changes since v1:
- fixed build with FreeBSD,
---
lib/eal/unix/rte_thread.c | 25 +++++++++++++++++++++++++
1 file changed, 25 insertions(+)
diff --git a/lib/eal/unix/rte_thread.c b/lib/eal/unix/rte_thread.c
index 1b4c73f58e..03c4164059 100644
--- a/lib/eal/unix/rte_thread.c
+++ b/lib/eal/unix/rte_thread.c
@@ -19,6 +19,7 @@ struct eal_tls_key {
pthread_key_t thread_index;
};
+#ifndef __USE_GNU
struct thread_start_context {
rte_thread_func thread_func;
void *thread_args;
@@ -28,6 +29,7 @@ struct thread_start_context {
int wrapper_ret;
bool wrapper_done;
};
+#endif
static int
thread_map_priority_to_os_value(enum rte_thread_priority eal_pri, int *os_pri,
@@ -88,6 +90,7 @@ thread_map_os_priority_to_eal_priority(int policy, int os_pri,
return 0;
}
+#ifndef __USE_GNU
static void *
thread_start_wrapper(void *arg)
{
@@ -113,6 +116,7 @@ thread_start_wrapper(void *arg)
return (void *)(uintptr_t)thread_func(thread_args);
}
+#endif
int
rte_thread_create(rte_thread_t *thread_id,
@@ -126,6 +130,7 @@ rte_thread_create(rte_thread_t *thread_id,
.sched_priority = 0,
};
int policy = SCHED_OTHER;
+#ifndef __USE_GNU
struct thread_start_context ctx = {
.thread_func = thread_func,
.thread_args = args,
@@ -134,6 +139,7 @@ rte_thread_create(rte_thread_t *thread_id,
.wrapper_mutex = PTHREAD_MUTEX_INITIALIZER,
.wrapper_cond = PTHREAD_COND_INITIALIZER,
};
+#endif
if (thread_attr != NULL) {
ret = pthread_attr_init(&attr);
@@ -144,6 +150,16 @@ rte_thread_create(rte_thread_t *thread_id,
attrp = &attr;
+#ifdef __USE_GNU
+ if (CPU_COUNT(&thread_attr->cpuset) > 0) {
+ ret = pthread_attr_setaffinity_np(attrp, sizeof(thread_attr->cpuset),
+ &thread_attr->cpuset);
+ if (ret != 0) {
+ EAL_LOG(DEBUG, "pthread_attr_setaffinity_np failed");
+ goto cleanup;
+ }
+ }
+#endif
/*
* Set the inherit scheduler parameter to explicit,
* otherwise the priority attribute is ignored.
@@ -178,6 +194,14 @@ rte_thread_create(rte_thread_t *thread_id,
}
}
+#ifdef __USE_GNU
+ ret = pthread_create((pthread_t *)&thread_id->opaque_id, attrp,
+ (void *)(void *)thread_func, args);
+ if (ret != 0) {
+ EAL_LOG(DEBUG, "pthread_create failed");
+ goto cleanup;
+ }
+#else /* !__USE_GNU */
ret = pthread_create((pthread_t *)&thread_id->opaque_id, attrp,
thread_start_wrapper, &ctx);
if (ret != 0) {
@@ -193,6 +217,7 @@ rte_thread_create(rte_thread_t *thread_id,
if (ret != 0)
rte_thread_join(*thread_id, NULL);
+#endif /* __USE_GNU */
cleanup:
if (attrp != NULL)
--
2.46.2
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH v2] eal/unix: optimize thread creation with glibc
2024-11-02 11:32 ` [PATCH v2] " David Marchand
@ 2024-11-02 12:36 ` Luca Boccassi
0 siblings, 0 replies; 3+ messages in thread
From: Luca Boccassi @ 2024-11-02 12:36 UTC (permalink / raw)
To: David Marchand; +Cc: dev, thomas, stephen, stable, Tyler Retzlaff
On Sat, 2 Nov 2024 at 11:32, David Marchand <david.marchand@redhat.com> wrote:
>
> Setting the cpu affinity of the child thread from the parent thread is
> racy when using pthread_setaffinity_np, as the child thread may start
> running and initialize before affinity is set.
>
> On the other hand, setting the cpu affinity from the child thread itself
> may fail, so the parent thread waits for the child thread to report
> whether this call succeeded.
>
> This synchronisation point resulted in a significant slow down of
> rte_thread_create() (as seen in the lcores_autotest unit tests, in OBS
> for some ARM systems).
>
> Another option for setting cpu affinity is to use the not portable
> pthread_attr_setaffinity_np, but it is not available with musl.
> Assume availability by relying on __USE_GNU that is set with glibc.
>
> Fixes: b28c6196b132 ("eal/unix: fix thread creation")
> Cc: stable@dpdk.org
>
> Signed-off-by: David Marchand <david.marchand@redhat.com>
> ---
> Changes since v1:
> - fixed build with FreeBSD,
>
> ---
> lib/eal/unix/rte_thread.c | 25 +++++++++++++++++++++++++
> 1 file changed, 25 insertions(+)
The test now completes in 1.19 seconds, so this fixes the issue with glibc:
[ 438s] 36/82 DPDK:fast-tests / lcores_autotest OK
1.19s
I do not use musl, so this is good enough for me. Thanks!
Acked-by: Luca Boccassi <bluca@debian.org>
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2024-11-02 12:37 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-11-02 10:08 [PATCH] eal/unix: optimize thread creation with glibc David Marchand
2024-11-02 11:32 ` [PATCH v2] " David Marchand
2024-11-02 12:36 ` Luca Boccassi
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).