| Message ID | 20220725073855.76049-15-eesposit@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | job: replace AioContext lock with job_mutex | expand |
On 7/25/22 10:38, Emanuele Giuseppe Esposito wrote: > In order to make it thread safe, implement a "fake rwlock", > where we allow reads under BQL *or* job_mutex held, but > writes only under BQL *and* job_mutex. > > The only write we have is in child_job_set_aio_ctx, which always > happens under drain (so the job is paused). > For this reason, introduce job_set_aio_context and make sure that > the context is set under BQL, job_mutex and drain. actually, we only make sure that pause was scheduled > Also make sure all other places where the aiocontext is read > are protected. > > Note: at this stage, job_{lock/unlock} and job lock guard macros > are *nop*. > > Suggested-by: Paolo Bonzini <pbonzini@redhat.com> > Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> > Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> > --- > block/replication.c | 6 ++++-- > blockjob.c | 3 ++- > include/qemu/job.h | 19 ++++++++++++++++++- > job.c | 12 ++++++++++++ > 4 files changed, 36 insertions(+), 4 deletions(-) > > diff --git a/block/replication.c b/block/replication.c > index 55c8f894aa..2189863df1 100644 > --- a/block/replication.c > +++ b/block/replication.c > @@ -148,8 +148,10 @@ static void replication_close(BlockDriverState *bs) > } > if (s->stage == BLOCK_REPLICATION_FAILOVER) { > commit_job = &s->commit_job->job; > - assert(commit_job->aio_context == qemu_get_current_aio_context()); > - job_cancel_sync(commit_job, false); > + WITH_JOB_LOCK_GUARD() { > + assert(commit_job->aio_context == qemu_get_current_aio_context()); > + job_cancel_sync_locked(commit_job, false); > + } > } > > if (s->mode == REPLICATION_MODE_SECONDARY) { > diff --git a/blockjob.c b/blockjob.c > index 96fb9d9f73..9ff2727025 100644 > --- a/blockjob.c > +++ b/blockjob.c > @@ -162,12 +162,13 @@ static void child_job_set_aio_ctx(BdrvChild *c, AioContext *ctx, > bdrv_set_aio_context_ignore(sibling->bs, ctx, ignore); > } > > - job->job.aio_context = ctx; > + job_set_aio_context(&job->job, ctx); > } > > static AioContext *child_job_get_parent_aio_context(BdrvChild *c) > { > BlockJob *job = c->opaque; > + assert(qemu_in_main_thread()); > > return job->job.aio_context; > } > diff --git a/include/qemu/job.h b/include/qemu/job.h > index 5709e8d4a8..c144aabefc 100644 > --- a/include/qemu/job.h > +++ b/include/qemu/job.h > @@ -77,7 +77,12 @@ typedef struct Job { > > /** Protected by AioContext lock */ > > - /** AioContext to run the job coroutine in */ > + /** > + * AioContext to run the job coroutine in. > + * This field can be read when holding either the BQL (so we are in > + * the main loop) or the job_mutex. > + * It can be only written when we hold *both* BQL and job_mutex. > + */ > AioContext *aio_context; > > /** Reference count of the block job */ > @@ -741,4 +746,16 @@ int job_finish_sync(Job *job, void (*finish)(Job *, Error **errp), > int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), > Error **errp); > > +/** > + * Sets the @job->aio_context. > + * Called with job_mutex *not* held. > + * > + * This function must run in the main thread to protect against > + * concurrent read in job_finish_sync_locked(), > + * takes the job_mutex lock to protect against the read in > + * job_do_yield_locked(), and must be called when the coroutine > + * is quiescent. What means "coroutine is quescent"? Could we just swap it by "job is paused"? > + */ > +void job_set_aio_context(Job *job, AioContext *ctx); > + > #endif > diff --git a/job.c b/job.c > index ecec66b44e..0a857b1468 100644 > --- a/job.c > +++ b/job.c > @@ -394,6 +394,17 @@ Job *job_get(const char *id) > return job_get_locked(id); > } > > +void job_set_aio_context(Job *job, AioContext *ctx) > +{ > + /* protect against read in job_finish_sync_locked and job_start */ > + assert(qemu_in_main_thread()); > + /* protect against read in job_do_yield_locked */ > + JOB_LOCK_GUARD(); > + /* ensure the coroutine is quiescent while the AioContext is changed */ pause_count means pause was scheduled. Job may be paused already, or may be not. I'm not against the assertion, it helps. I just think that we don't have the guarantee that comment claims. (And I still don't understand what means coroutine is quiescent. And not that there are may be several job related coroutines: the main one and some worker coroutines). > + assert(job->pause_count > 0); > + job->aio_context = ctx; > +} > + > /* Called with job_mutex *not* held. */ > static void job_sleep_timer_cb(void *opaque) > { > @@ -1376,6 +1387,7 @@ int job_finish_sync_locked(Job *job, > { > Error *local_err = NULL; > int ret; > + assert(qemu_in_main_thread()); > > job_ref_locked(job); > without "quiescent coroutine" concept: Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Am 25.07.2022 um 09:38 hat Emanuele Giuseppe Esposito geschrieben: > In order to make it thread safe, implement a "fake rwlock", > where we allow reads under BQL *or* job_mutex held, but > writes only under BQL *and* job_mutex. Oh, so the "or BQL" part is only for job.aio_context? Okay. > The only write we have is in child_job_set_aio_ctx, which always > happens under drain (so the job is paused). > For this reason, introduce job_set_aio_context and make sure that > the context is set under BQL, job_mutex and drain. > Also make sure all other places where the aiocontext is read > are protected. > > Note: at this stage, job_{lock/unlock} and job lock guard macros > are *nop*. > > Suggested-by: Paolo Bonzini <pbonzini@redhat.com> > Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> > Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> > --- > block/replication.c | 6 ++++-- > blockjob.c | 3 ++- > include/qemu/job.h | 19 ++++++++++++++++++- > job.c | 12 ++++++++++++ > 4 files changed, 36 insertions(+), 4 deletions(-) > > diff --git a/block/replication.c b/block/replication.c > index 55c8f894aa..2189863df1 100644 > --- a/block/replication.c > +++ b/block/replication.c > @@ -148,8 +148,10 @@ static void replication_close(BlockDriverState *bs) > } > if (s->stage == BLOCK_REPLICATION_FAILOVER) { > commit_job = &s->commit_job->job; > - assert(commit_job->aio_context == qemu_get_current_aio_context()); > - job_cancel_sync(commit_job, false); > + WITH_JOB_LOCK_GUARD() { > + assert(commit_job->aio_context == qemu_get_current_aio_context()); > + job_cancel_sync_locked(commit_job, false); > + } > } .bdrv_close runs under the BQL, so why is this needed? Maybe a GLOBAL_STATE_CODE() annotation would be helpful, though. > if (s->mode == REPLICATION_MODE_SECONDARY) { > diff --git a/blockjob.c b/blockjob.c > index 96fb9d9f73..9ff2727025 100644 > --- a/blockjob.c > +++ b/blockjob.c > @@ -162,12 +162,13 @@ static void child_job_set_aio_ctx(BdrvChild *c, AioContext *ctx, > bdrv_set_aio_context_ignore(sibling->bs, ctx, ignore); > } > > - job->job.aio_context = ctx; > + job_set_aio_context(&job->job, ctx); > } > > static AioContext *child_job_get_parent_aio_context(BdrvChild *c) > { > BlockJob *job = c->opaque; > + assert(qemu_in_main_thread()); Any reason not to use GLOBAL_STATE_CODE()? > return job->job.aio_context; > } > diff --git a/include/qemu/job.h b/include/qemu/job.h > index 5709e8d4a8..c144aabefc 100644 > --- a/include/qemu/job.h > +++ b/include/qemu/job.h > @@ -77,7 +77,12 @@ typedef struct Job { > > /** Protected by AioContext lock */ I think this section comment should move down below aio_context now. > - /** AioContext to run the job coroutine in */ > + /** > + * AioContext to run the job coroutine in. > + * This field can be read when holding either the BQL (so we are in > + * the main loop) or the job_mutex. > + * It can be only written when we hold *both* BQL and job_mutex. > + */ > AioContext *aio_context; > > /** Reference count of the block job */ > @@ -741,4 +746,16 @@ int job_finish_sync(Job *job, void (*finish)(Job *, Error **errp), > int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), > Error **errp); > > +/** > + * Sets the @job->aio_context. > + * Called with job_mutex *not* held. > + * > + * This function must run in the main thread to protect against > + * concurrent read in job_finish_sync_locked(), Odd line break here in the middle of a sentence. > + * takes the job_mutex lock to protect against the read in > + * job_do_yield_locked(), and must be called when the coroutine > + * is quiescent. > + */ > +void job_set_aio_context(Job *job, AioContext *ctx); > + > #endif > diff --git a/job.c b/job.c > index ecec66b44e..0a857b1468 100644 > --- a/job.c > +++ b/job.c > @@ -394,6 +394,17 @@ Job *job_get(const char *id) > return job_get_locked(id); > } > > +void job_set_aio_context(Job *job, AioContext *ctx) > +{ > + /* protect against read in job_finish_sync_locked and job_start */ > + assert(qemu_in_main_thread()); Same question about GLOBAL_STATE_CODE(). > + /* protect against read in job_do_yield_locked */ > + JOB_LOCK_GUARD(); > + /* ensure the coroutine is quiescent while the AioContext is changed */ > + assert(job->pause_count > 0); job->pause_count only shows that pausing was requested. The coroutine is only really quiescent if job->busy == false, too. Or maybe job->paused is actually the one you want here. > + job->aio_context = ctx; > +} > + > /* Called with job_mutex *not* held. */ > static void job_sleep_timer_cb(void *opaque) > { > @@ -1376,6 +1387,7 @@ int job_finish_sync_locked(Job *job, > { > Error *local_err = NULL; > int ret; > + assert(qemu_in_main_thread()); > > job_ref_locked(job); Another GLOBAL_STATE_CODE()? Kevin
Am 05/08/2022 um 11:12 schrieb Kevin Wolf: > Am 25.07.2022 um 09:38 hat Emanuele Giuseppe Esposito geschrieben: >> In order to make it thread safe, implement a "fake rwlock", >> where we allow reads under BQL *or* job_mutex held, but >> writes only under BQL *and* job_mutex. > > Oh, so the "or BQL" part is only for job.aio_context? Okay. > >> The only write we have is in child_job_set_aio_ctx, which always >> happens under drain (so the job is paused). >> For this reason, introduce job_set_aio_context and make sure that >> the context is set under BQL, job_mutex and drain. >> Also make sure all other places where the aiocontext is read >> are protected. >> >> Note: at this stage, job_{lock/unlock} and job lock guard macros >> are *nop*. >> >> Suggested-by: Paolo Bonzini <pbonzini@redhat.com> >> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> >> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> >> --- >> block/replication.c | 6 ++++-- >> blockjob.c | 3 ++- >> include/qemu/job.h | 19 ++++++++++++++++++- >> job.c | 12 ++++++++++++ >> 4 files changed, 36 insertions(+), 4 deletions(-) >> >> diff --git a/block/replication.c b/block/replication.c >> index 55c8f894aa..2189863df1 100644 >> --- a/block/replication.c >> +++ b/block/replication.c >> @@ -148,8 +148,10 @@ static void replication_close(BlockDriverState *bs) >> } >> if (s->stage == BLOCK_REPLICATION_FAILOVER) { >> commit_job = &s->commit_job->job; >> - assert(commit_job->aio_context == qemu_get_current_aio_context()); >> - job_cancel_sync(commit_job, false); >> + WITH_JOB_LOCK_GUARD() { >> + assert(commit_job->aio_context == qemu_get_current_aio_context()); >> + job_cancel_sync_locked(commit_job, false); >> + } >> } > > .bdrv_close runs under the BQL, so why is this needed? Maybe a > GLOBAL_STATE_CODE() annotation would be helpful, though. I think I left it because it would be confusing to leave a _locked function without the job lock. I'll add the GLOBAL_STATE_CODE anyways. > >> if (s->mode == REPLICATION_MODE_SECONDARY) { >> diff --git a/blockjob.c b/blockjob.c >> index 96fb9d9f73..9ff2727025 100644 >> --- a/blockjob.c >> +++ b/blockjob.c >> @@ -162,12 +162,13 @@ static void child_job_set_aio_ctx(BdrvChild *c, AioContext *ctx, >> bdrv_set_aio_context_ignore(sibling->bs, ctx, ignore); >> } >> >> - job->job.aio_context = ctx; >> + job_set_aio_context(&job->job, ctx); >> } >> >> static AioContext *child_job_get_parent_aio_context(BdrvChild *c) >> { >> BlockJob *job = c->opaque; >> + assert(qemu_in_main_thread()); > > Any reason not to use GLOBAL_STATE_CODE()? 4 months ago GLOBAL_STATE_CODE did not exist yet, and I didn't think about updating it :) > >> return job->job.aio_context; >> } >> + /* protect against read in job_do_yield_locked */ >> + JOB_LOCK_GUARD(); >> + /* ensure the coroutine is quiescent while the AioContext is changed */ >> + assert(job->pause_count > 0); > > job->pause_count only shows that pausing was requested. The coroutine is > only really quiescent if job->busy == false, too. > > Or maybe job->paused is actually the one you want here. I think job->paused works too. Emanuele
Am 17/08/2022 um 10:04 schrieb Emanuele Giuseppe Esposito: >>> + /* protect against read in job_do_yield_locked */ >>> + JOB_LOCK_GUARD(); >>> + /* ensure the coroutine is quiescent while the AioContext is changed */ >>> + assert(job->pause_count > 0); >> job->pause_count only shows that pausing was requested. The coroutine is >> only really quiescent if job->busy == false, too. >> >> Or maybe job->paused is actually the one you want here. > I think job->paused works too. > Actually it doesn't. At least test-block-iothread test_propagate_mirror() fails, for both job->paused and !job->busy. I think the reason is that if we wait for the flag to be set, we need to actually wait that the job gets to the next pause point, because job_pause() doesn't really pause the job, it just kind of "schedules" the pause on next pause point. So, either we leave pause_count > 0, or somehow wait (I was thinking AIO_WAIT_WHILE(job->paused) but that's probably a very bad idea). Do you have any suggestion for that? Maybe Paolo has a better idea on how to do it? Emanuele
Am 17/08/2022 um 15:10 schrieb Emanuele Giuseppe Esposito: > > > Am 17/08/2022 um 10:04 schrieb Emanuele Giuseppe Esposito: >>>> + /* protect against read in job_do_yield_locked */ >>>> + JOB_LOCK_GUARD(); >>>> + /* ensure the coroutine is quiescent while the AioContext is changed */ >>>> + assert(job->pause_count > 0); >>> job->pause_count only shows that pausing was requested. The coroutine is >>> only really quiescent if job->busy == false, too. >>> >>> Or maybe job->paused is actually the one you want here. >> I think job->paused works too. >> > > Actually it doesn't. At least test-block-iothread > test_propagate_mirror() fails, for both job->paused and !job->busy. I > think the reason is that if we wait for the flag to be set, we need to > actually wait that the job gets to the next pause point, because > job_pause() doesn't really pause the job, it just kind of "schedules" > the pause on next pause point. > > So, either we leave pause_count > 0, or somehow wait (I was thinking > AIO_WAIT_WHILE(job->paused) but that's probably a very bad idea). > > Do you have any suggestion for that? > Maybe Paolo has a better idea on how to do it? > Additional update: after debugging it a little bit, I figured that there are places where busy=false and paused=false too. This is clear in job_do_dismiss and not-so-clear in job_exit (at least for me). I don't know if this is expected or a bug, because it just means that if we are draining and the job is completing, there is no way to stop it. Is it a bug? Or am I missing something? Emanuele
diff --git a/block/replication.c b/block/replication.c index 55c8f894aa..2189863df1 100644 --- a/block/replication.c +++ b/block/replication.c @@ -148,8 +148,10 @@ static void replication_close(BlockDriverState *bs) } if (s->stage == BLOCK_REPLICATION_FAILOVER) { commit_job = &s->commit_job->job; - assert(commit_job->aio_context == qemu_get_current_aio_context()); - job_cancel_sync(commit_job, false); + WITH_JOB_LOCK_GUARD() { + assert(commit_job->aio_context == qemu_get_current_aio_context()); + job_cancel_sync_locked(commit_job, false); + } } if (s->mode == REPLICATION_MODE_SECONDARY) { diff --git a/blockjob.c b/blockjob.c index 96fb9d9f73..9ff2727025 100644 --- a/blockjob.c +++ b/blockjob.c @@ -162,12 +162,13 @@ static void child_job_set_aio_ctx(BdrvChild *c, AioContext *ctx, bdrv_set_aio_context_ignore(sibling->bs, ctx, ignore); } - job->job.aio_context = ctx; + job_set_aio_context(&job->job, ctx); } static AioContext *child_job_get_parent_aio_context(BdrvChild *c) { BlockJob *job = c->opaque; + assert(qemu_in_main_thread()); return job->job.aio_context; } diff --git a/include/qemu/job.h b/include/qemu/job.h index 5709e8d4a8..c144aabefc 100644 --- a/include/qemu/job.h +++ b/include/qemu/job.h @@ -77,7 +77,12 @@ typedef struct Job { /** Protected by AioContext lock */ - /** AioContext to run the job coroutine in */ + /** + * AioContext to run the job coroutine in. + * This field can be read when holding either the BQL (so we are in + * the main loop) or the job_mutex. + * It can be only written when we hold *both* BQL and job_mutex. + */ AioContext *aio_context; /** Reference count of the block job */ @@ -741,4 +746,16 @@ int job_finish_sync(Job *job, void (*finish)(Job *, Error **errp), int job_finish_sync_locked(Job *job, void (*finish)(Job *, Error **errp), Error **errp); +/** + * Sets the @job->aio_context. + * Called with job_mutex *not* held. + * + * This function must run in the main thread to protect against + * concurrent read in job_finish_sync_locked(), + * takes the job_mutex lock to protect against the read in + * job_do_yield_locked(), and must be called when the coroutine + * is quiescent. + */ +void job_set_aio_context(Job *job, AioContext *ctx); + #endif diff --git a/job.c b/job.c index ecec66b44e..0a857b1468 100644 --- a/job.c +++ b/job.c @@ -394,6 +394,17 @@ Job *job_get(const char *id) return job_get_locked(id); } +void job_set_aio_context(Job *job, AioContext *ctx) +{ + /* protect against read in job_finish_sync_locked and job_start */ + assert(qemu_in_main_thread()); + /* protect against read in job_do_yield_locked */ + JOB_LOCK_GUARD(); + /* ensure the coroutine is quiescent while the AioContext is changed */ + assert(job->pause_count > 0); + job->aio_context = ctx; +} + /* Called with job_mutex *not* held. */ static void job_sleep_timer_cb(void *opaque) { @@ -1376,6 +1387,7 @@ int job_finish_sync_locked(Job *job, { Error *local_err = NULL; int ret; + assert(qemu_in_main_thread()); job_ref_locked(job);
In order to make it thread safe, implement a "fake rwlock", where we allow reads under BQL *or* job_mutex held, but writes only under BQL *and* job_mutex. The only write we have is in child_job_set_aio_ctx, which always happens under drain (so the job is paused). For this reason, introduce job_set_aio_context and make sure that the context is set under BQL, job_mutex and drain. Also make sure all other places where the aiocontext is read are protected. Note: at this stage, job_{lock/unlock} and job lock guard macros are *nop*. Suggested-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> --- block/replication.c | 6 ++++-- blockjob.c | 3 ++- include/qemu/job.h | 19 ++++++++++++++++++- job.c | 12 ++++++++++++ 4 files changed, 36 insertions(+), 4 deletions(-)