iothread: create gcontext unconditionally
v2 cover iothread: create gcontext unconditionally
v2:
- add comment in patch 4
- add another patch to comment why we need explicit aio_poll() in iothread_run loop
When I first read the iothread code, the gcontext confused me for quite a while. Meanwhile, I've been tackling with some races due to this complexity as well. How much we'll pay for creating the gcontext unconditionally? Do we really need this flexibitily (or is it really a flexibility after all)? I don't see much gain of existing code, but I might be wrong. Anyway, I wrote this patchset to see how the list would think about it.
This series directly originates from previous discussion with Marc-Andre where there's a slightly hacky way to try to acquire the gcontext:
https://lists.gnu.org/archive/html/qemu-devel/2019-02/msg05460.html
Now with this series logically above patch is not needed any more. Please read patch 4 for more information.
And if this patchset can survive... how about running gcontext directly in iothread_run()? I believe there could be a bit more things to clean but I'll see.
Make check passes for me.
Comments welcomed. Thanks,
Peter Xu (5):
iothread: replace init_done_cond with a semaphore
iothread: create the gcontext unconditionally
iothread: create main loop unconditionally
iothread: push gcontext earlier in the thread_fn
iothread: document about why we need explicit aio_poll()
include/sysemu/iothread.h | 5 +-- iothread.c | 91 +++++++++++++++++++-------------------- 2 files changed, 46 insertions(+), 50 deletions(-)
Signature · 1 line
2.17.1
PATCH v2 1/5 iothread: replace init_done_cond with a semaphore
Only sending an init-done message using lock+cond seems an overkill to me. Replacing it with a simpler semaphore.
Meanwhile, init the semaphore unconditionally, then we can destroy it unconditionally too in finalize which seems cleaner.
--- include/sysemu/iothread.h | 3 +-- iothread.c | 17 ++++------------- 2 files changed, 5 insertions(+), 15 deletions(-)
include/sysemu/iothread.h +1 -2
@@ -27,8 +27,7 @@ typedef struct {GMainContext *worker_context;GMainLoop *main_loop;GOnce once; -QemuMutex init_done_lock; -QemuCond init_done_cond;/* is thread initialization done? */ +QemuSemaphore init_done_sem;/* is thread init done? */ bool stopping;/* has iothread_stop() been called? */ bool running;/* should iothread_run() continue? */ int thread_id;
iothread.c +4 -13
@@ -55,10 +55,8 @@ static void *iothread_run(void *opaque)rcu_register_thread (); my_iothread = iothread; -qemu_mutex_lock (&iothread->init_done_lock); iothread->thread_id =qemu_get_thread_id (); -qemu_cond_signal (&iothread->init_done_cond); -qemu_mutex_unlock (&iothread->init_done_lock); +qemu_sem_post (&iothread->init_done_sem);while (iothread->running) {aio_poll (iothread->ctx, true);
@@ -115,6 +113,7 @@ static void iothread_instance_init(Object *obj) iothread->poll_max_ns =IOTHREAD_POLL_MAX_NS_DEFAULT ; iothread->thread_id =-1 ; +qemu_sem_init (&iothread->init_done_sem,0 ); }static void iothread_instance_finalize (Object *obj)
@@ -123,10 +122,6 @@ static void iothread_instance_finalize(Object *obj)iothread_stop (iothread); -if (iothread->thread_id !=-1 ) { -qemu_cond_destroy (&iothread->init_done_cond); -qemu_mutex_destroy (&iothread->init_done_lock); - }/* * Before glib2 2.33.10, there is a glib2 bug that GSource context * pointer may not be cleared even if the context has already been
@@ -145,6 +140,7 @@ static void iothread_instance_finalize(Object *obj)g_main_context_unref (iothread->worker_context); iothread->worker_context =NULL ; } +qemu_sem_destroy (&iothread->init_done_sem); }static void iothread_complete (UserCreatable *obj,Error **errp)
@@ -173,8 +169,6 @@ static void iothread_complete(UserCreatable *obj, Error **errp)return ; } -qemu_mutex_init (&iothread->init_done_lock); -qemu_cond_init (&iothread->init_done_cond); iothread->once = (GOnce )G_ONCE_INIT ;/* This assumes we are called from a thread with useful CPU affinity for us
@@ -188,12 +182,9 @@ static void iothread_complete(UserCreatable *obj, Error **errp)g_free (name);/* Wait for initialization to complete */ -qemu_mutex_lock (&iothread->init_done_lock);while (iothread->thread_id ==-1 ) { -qemu_cond_wait (&iothread->init_done_cond, - &iothread->init_done_lock); +qemu_sem_wait (&iothread->init_done_sem); } -qemu_mutex_unlock (&iothread->init_done_lock); } typedef struct {
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH v2 1/5] iothread: replace init_done_cond with a semaphore
Quoted 7 lines · Peter Xu
Only sending an init-done message using lock+cond seems an overkill to me. Replacing it with a simpler semaphore.
Meanwhile, init the semaphore unconditionally, then we can destroy it unconditionally too in finalize which seems cleaner.
Signed-off-by Peter Xu <[email protected]>
Sorry I've definitely lost Stefan's r-b for this patch. It's the same patch as the one I posted in v1.
Quoted 88 lines
--- include/sysemu/iothread.h | 3 +-- iothread.c | 17 ++++------------- 2 files changed, 5 insertions(+), 15 deletions(-)include/sysemu/iothread.h +1
-2@@ -27,8 +27,7 @@ typedef struct {GMainContext *worker_context;GMainLoop *main_loop;GOnce once; -QemuMutex init_done_lock; -QemuCond init_done_cond;/* is thread initialization done? */ +QemuSemaphore init_done_sem;/* is thread init done? */ bool stopping;/* has iothread_stop() been called? */ bool running;/* should iothread_run() continue? */ int thread_id;iothread.c +4
-13@@ -55,10 +55,8 @@ static void *iothread_run(void *opaque)rcu_register_thread (); my_iothread = iothread; -qemu_mutex_lock (&iothread->init_done_lock); iothread->thread_id =qemu_get_thread_id (); -qemu_cond_signal (&iothread->init_done_cond); -qemu_mutex_unlock (&iothread->init_done_lock); +qemu_sem_post (&iothread->init_done_sem);while (iothread->running) {aio_poll (iothread->ctx, true);@@ -115,6 +113,7 @@ static void iothread_instance_init(Object *obj) iothread->poll_max_ns =IOTHREAD_POLL_MAX_NS_DEFAULT ; iothread->thread_id =-1 ; +qemu_sem_init (&iothread->init_done_sem,0 ); }static void iothread_instance_finalize (Object *obj)@@ -123,10 +122,6 @@ static void iothread_instance_finalize(Object *obj)iothread_stop (iothread); -if (iothread->thread_id !=-1 ) { -qemu_cond_destroy (&iothread->init_done_cond); -qemu_mutex_destroy (&iothread->init_done_lock); - }/* * Before glib2 2.33.10, there is a glib2 bug that GSource context * pointer may not be cleared even if the context has already been @@ -145,6 +140,7 @@ static void iothread_instance_finalize(Object *obj)g_main_context_unref (iothread->worker_context); iothread->worker_context =NULL ; } +qemu_sem_destroy (&iothread->init_done_sem); }static void iothread_complete (UserCreatable *obj,Error **errp)@@ -173,8 +169,6 @@ static void iothread_complete(UserCreatable *obj, Error **errp)return ; } -qemu_mutex_init (&iothread->init_done_lock); -qemu_cond_init (&iothread->init_done_cond); iothread->once = (GOnce )G_ONCE_INIT ;/* This assumes we are called from a thread with useful CPU affinity for us @@ -188,12 +182,9 @@ static void iothread_complete(UserCreatable *obj, Error **errp)g_free (name);/* Wait for initialization to complete */ -qemu_mutex_lock (&iothread->init_done_lock);while (iothread->thread_id ==-1 ) { -qemu_cond_wait (&iothread->init_done_cond, - &iothread->init_done_lock); +qemu_sem_wait (&iothread->init_done_sem); } -qemu_mutex_unlock (&iothread->init_done_lock); } typedef struct {Signature · 1 line
2.17.1
Regards,
Signature · 1 line
Peter Xu
PATCH v2 2/5 iothread: create the gcontext unconditionally
In reply to Peter XuIn existing code we create the gcontext dynamically at the first access of the gcontext from caller. That can bring some complexity and potential races during using iothread. Since the context itself is not that big a resource, and we won't have millions of iothread, let's simply create the gcontext unconditionally.
This will also be a preparation work further to move the thread context push operation earlier than before (now it's only pushed right before we want to start running the gmainloop).
Removing the g_once since it's not necessary, while introducing a new run_gcontext boolean to show whether we want to run the gcontext.
--- include/sysemu/iothread.h | 2 +- iothread.c | 43 +++++++++++++++++++-------------------- 2 files changed, 22 insertions(+), 23 deletions(-)
include/sysemu/iothread.h +1 -1
@@ -24,9 +24,9 @@ typedef struct {QemuThread thread;AioContext *ctx; +bool run_gcontext;/* whether we should run gcontext */ GMainContext *worker_context;GMainLoop *main_loop; -GOnce once;QemuSemaphore init_done_sem;/* is thread init done? */ bool stopping;/* has iothread_stop() been called? */ bool running;/* should iothread_run() continue? */
iothread.c +21 -22
81 lines not on this page · raw message
Signature · 1 line
2.17.1
PATCH v2 3/5 iothread: create main loop unconditionally
In reply to Peter XuSince we've have the gcontext always there, create the main loop altogether. The iothread_run() is even cleaner.
--- iothread.c | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-)
iothread.c +3 -9
@@ -66,17 +66,8 @@ static void *iothread_run(void *opaque)* changed in previous aio_poll() */ if (iothread->running &&atomic_read (&iothread->run_gcontext)) { -GMainLoop *loop; -g_main_context_push_thread_default (iothread->worker_context); - iothread->main_loop = -g_main_loop_new (iothread->worker_context, TRUE); - loop = iothread->main_loop; -g_main_loop_run (iothread->main_loop); - iothread->main_loop =NULL ; -g_main_loop_unref (loop); -g_main_context_pop_thread_default (iothread->worker_context); } }
@@ -141,6 +132,8 @@ static void iothread_instance_finalize(Object *obj)if (iothread->worker_context) {g_main_context_unref (iothread->worker_context); iothread->worker_context =NULL ; +g_main_loop_unref (iothread->main_loop); + iothread->main_loop =NULL ; }qemu_sem_destroy (&iothread->init_done_sem); }
@@ -153,6 +146,7 @@ static void iothread_init_gcontext(IOThread *iothread) source =aio_get_g_source (iothread_get_aio_context (iothread));g_source_attach (source, iothread->worker_context);g_source_unref (source); + iothread->main_loop =g_main_loop_new (iothread->worker_context, TRUE); }static void iothread_complete (UserCreatable *obj,Error **errp)
Signature · 1 line
2.17.1
PATCH v2 4/5 iothread: push gcontext earlier in the thread_fn
In reply to Peter XuWe were pushing the context until right before running the gmainloop. Now since we have everything unconditionally, we can move this earlier.
One benefit is that now it's done even before init_done_sem, so as long as the iothread user calls iothread_create() and completes, we know that the thread stack is ready.
--- iothread.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
iothread.c +7 -3
@@ -53,7 +53,12 @@ static void *iothread_run(void *opaque)IOThread *iothread = opaque;rcu_register_thread (); - +/* +* We should do this as soon as we enter the thread, because the +* function will silently fail if it fails to acquire the +* gcontext. +*/ +g_main_context_push_thread_default (iothread->worker_context); my_iothread = iothread; iothread->thread_id =qemu_get_thread_id ();qemu_sem_post (&iothread->init_done_sem);
@@ -66,12 +71,11 @@ static void *iothread_run(void *opaque)* changed in previous aio_poll() */ if (iothread->running &&atomic_read (&iothread->run_gcontext)) { -g_main_context_push_thread_default (iothread->worker_context);g_main_loop_run (iothread->main_loop); -g_main_context_pop_thread_default (iothread->worker_context); } } +g_main_context_pop_thread_default (iothread->worker_context);rcu_unregister_thread ();return NULL ; }
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH v2 4/5] iothread: push gcontext earlier in the thread_fn
iothread.c static void *iothread_run(void *opaque) · line 61
Quoted 6 lines · Peter Xu
iothread.c +6
-0+/* +* We should do this as soon as we enter the thread, because the +* function will silently fail if it fails to acquire the +* gcontext. +*/ +g_main_context_push_thread_default (iothread->worker_context);
I have a hard time understanding this comment. The mention of how it fails makes me think "we'll never find out about failures anyway, so how does it help to call this early?".
I suggest sticking to the point that this function must always be called first:
/*
* g_main_context_push_thread_default() must be called before anything
* in this new thread uses glib.
*/
Now people will think before moving this function call.
Stefan
Re: [Qemu-devel] [PATCH v2 4/5] iothread: push gcontext earlier in the thread_fn
Quoted 15 lines · Stefan Hajnoczi
iothread.c static void *iothread_run(void *opaque) · line 61
Quoted 6 lines · Peter Xu
iothread.c +6
-0+/* +* We should do this as soon as we enter the thread, because the +* function will silently fail if it fails to acquire the +* gcontext. +*/ +g_main_context_push_thread_default (iothread->worker_context);I have a hard time understanding this comment. The mention of how it fails makes me think "we'll never find out about failures anyway, so how does it help to call this early?".
I suggest sticking to the point that this function must always be called first:
Quoted 6 lines
/*
*g_main_context_push_thread_default()must be called before anything
* in this new thread uses glib.
*/Now people will think before moving this function call.
Sorry to be confusing; this looks good to me. Please let me know if you want me to repost this patch alone or the patchset with the change.
Thanks,
Signature · 1 line
Peter Xu
PATCH v2 5/5 iothread: document about why we need explicit aio_poll()
In reply to Peter XuAfter consulting Paolo I know why we'd better keep the explicit aio_poll() in iothread_run(). Document it directly into the code so that future readers will know the answer from day one.
--- iothread.c | 9 +++++++++ 1 file changed, 9 insertions(+)
iothread.c +9 -0
@@ -64,6 +64,15 @@ static void *iothread_run(void *opaque)qemu_sem_post (&iothread->init_done_sem);while (iothread->running) { +/* +* Note: from functional-wise the g_main_loop_run() below can +* already cover the aio_poll() events, but we can't run the +* main loop unconditionally because explicit aio_poll() here +* is faster than g_main_loop_run() when we do not need the +* gcontext at all (e.g., pure block layer iothreads). In +* other words, when we want to run the gcontext with the +* iothread we need to pay some performance for functionality. +*/ aio_poll (iothread->ctx, true);/*
Signature · 1 line
2.17.1
Re: [Qemu-devel] [PATCH v2 5/5] iothread: document about why we need explicit aio_poll()
Quoted 6 lines · Peter Xu
After consulting Paolo I know why we'd better keep the explicit
aio_poll()iniothread_run(). Document it directly into the code so that future readers will know the answer from day one.Signed-off-by Peter Xu <[email protected]>
Quoted 28 lines
--- iothread.c | 9 +++++++++ 1 file changed, 9 insertions(+)iothread.c +9
-1@@ -64,6 +64,15 @@ static void *iothread_run(void *opaque)qemu_sem_post (&iothread->init_done_sem);while (iothread->running) { +/* +* Note: from functional-wise the g_main_loop_run() below can +* already cover the aio_poll() events, but we can't run the +* main loop unconditionally because explicit aio_poll() here +* is faster than g_main_loop_run() when we do not need the +* gcontext at all (e.g., pure block layer iothreads). In +* other words, when we want to run the gcontext with the +* iothread we need to pay some performance for functionality. +*/ aio_poll (iothread->ctx, true);/* -- 2.17.1
Signature · 1 line
Marc-André Lureau
Stefan Hajnoczi
AppliedQuoted 37 lines · Peter Xu
v2:
- add comment in patch 4
- add another patch to comment why we need explicitaio_poll()iniothread_runloopWhen I first read the iothread code, the gcontext confused me for quite a while. Meanwhile, I've been tackling with some races due to this complexity as well. How much we'll pay for creating the gcontext unconditionally? Do we really need this flexibitily (or is it really a flexibility after all)? I don't see much gain of existing code, but I might be wrong. Anyway, I wrote this patchset to see how the list would think about it.
This series directly originates from previous discussion with Marc-Andre where there's a slightly hacky way to try to acquire the gcontext:
https://lists.gnu.org/archive/html/qemu-devel/2019-02/msg05460.html
Now with this series logically above patch is not needed any more. Please read patch 4 for more information.
And if this patchset can survive... how about running gcontext directly in
iothread_run()? I believe there could be a bit more things to clean but I'll see.Make check passes for me.
Comments welcomed. Thanks,
Peter Xu (5):
iothread: replaceinit_done_condwith a semaphore
iothread: create the gcontext unconditionally
iothread: create main loop unconditionally
iothread: push gcontext earlier in thethread_fn
iothread: document about why we need explicitaio_poll()
Quoted 7 lines
include/sysemu/iothread.h | 5 +-- iothread.c | 91 +++++++++++++++++++-------------------- 2 files changed, 46 insertions(+), 50 deletions(-)Signature · 1 line
2.17.1
I made the comment tweak discussed in this thread.
Thanks, applied to my block tree:
https://github.com/stefanha/qemu/commits/block
Stefan