Message ID | 1596634551-27526-1-git-send-email-xianting_tian@126.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | aio: use wait_for_completion_io() when waiting for completion of io | expand |
On Wed 05-08-20 09:35:51, Xianting Tian wrote: > When waiting for the completion of io, we need account iowait time. As > wait_for_completion() calls schedule_timeout(), which doesn't account > iowait time. While wait_for_completion_io() calls io_schedule_timeout(), > which will account iowait time. > > So using wait_for_completion_io() instead of wait_for_completion() > when waiting for completion of io before exit_aio and io_destroy. > > Signed-off-by: Xianting Tian <xianting_tian@126.com> Thanks for the patch! It looks good to me but IMO this is just scratching the surface. E.g. for AIO we are mostly going to wait in read_events() by wait_event_interruptible_hrtimeout() and *that* doesn't account as IO wait either? Which is IMO far bigger misaccounting... The two case you fix seem to be just rare cornercases so what they do isn't a big deal either way. So I agree it may be worth it to properly account waiting for AIO but if you want to do that, then please handle mainly the common cases in AIO code. Honza > --- > fs/aio.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/fs/aio.c b/fs/aio.c > index 91e7cc4..498b8a0 100644 > --- a/fs/aio.c > +++ b/fs/aio.c > @@ -892,7 +892,7 @@ void exit_aio(struct mm_struct *mm) > > if (!atomic_sub_and_test(skipped, &wait.count)) { > /* Wait until all IO for the context are done. */ > - wait_for_completion(&wait.comp); > + wait_for_completion_io(&wait.comp); > } > > RCU_INIT_POINTER(mm->ioctx_table, NULL); > @@ -1400,7 +1400,7 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr, > * is destroyed. > */ > if (!ret) > - wait_for_completion(&wait.comp); > + wait_for_completion_io(&wait.comp); > > return ret; > } > -- > 1.8.3.1 >
Hello! On Wed 26-08-20 23:44:11, 田 wrote: > thanks for your kindly reply, > the normal wait path read_events()->wait_event_interruptible_hrtimeout(), > which will call schedule(), it does not account IO wait time. Not sure if there isn't some misunderstanding so I'll repeat what I've said: Yes, above path will not account as IO wait time and IMO that is much more common path which should be accounted as IO wait time. So I think that without fixing that path, fixing cornercases like you did in your patch is rather pointless. Honza > On 08/26/2020 21:23, Jan Kara wrote: > On Wed 05-08-20 09:35:51, Xianting Tian wrote: > > When waiting for the completion of io, we need account iowait time. As > > wait_for_completion() calls schedule_timeout(), which doesn't account > > iowait time. While wait_for_completion_io() calls io_schedule_timeout(), > > which will account iowait time. > > > > So using wait_for_completion_io() instead of wait_for_completion() > > when waiting for completion of io before exit_aio and io_destroy. > > > > Signed-off-by: Xianting Tian <xianting_tian@126.com> > > Thanks for the patch! It looks good to me but IMO this is just scratching > the surface. E.g. for AIO we are mostly going to wait in read_events() by > wait_event_interruptible_hrtimeout() and *that* doesn't account as IO wait > either? Which is IMO far bigger misaccounting... The two case you fix seem > to be just rare cornercases so what they do isn't a big deal either way. > > So I agree it may be worth it to properly account waiting for AIO but if > you want to do that, then please handle mainly the common cases in AIO > code. > > Honza > > > --- > > fs/aio.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/fs/aio.c b/fs/aio.c > > index 91e7cc4..498b8a0 100644 > > --- a/fs/aio.c > > +++ b/fs/aio.c > > @@ -892,7 +892,7 @@ void exit_aio(struct mm_struct *mm) > > > > if (!atomic_sub_and_test(skipped, &wait.count)) { > > /* Wait until all IO for the context are done. */ > > - wait_for_completion(&wait.comp); > > + wait_for_completion_io(&wait.comp); > > } > > > > RCU_INIT_POINTER(mm->ioctx_table, NULL); > > @@ -1400,7 +1400,7 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr, > > * is destroyed. > > */ > > if (!ret) > > - wait_for_completion(&wait.comp); > > + wait_for_completion_io(&wait.comp); > > > > return ret; > > } > > -- > > 1.8.3.1 > > > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR
Hi! On Thu 27-08-20 16:28:37, 田 wrote: > I understood what you said before:) Good :) > Totally agree with you, that we should fix the common path to make it to > account IO wait time. Currently kernel only has io_wait_event(), which > does not support timeout, maybe we need develop new interface like > io_wait_event_hrtimeout(), then we can use it instead of > wait_event_interruptible_hrtimeout()? Yes, that's what I'd do. Honza > > On 08/27/2020 15:55, Jan Kara wrote: > Hello! > > On Wed 26-08-20 23:44:11, 田 wrote: > > thanks for your kindly reply, > > the normal wait path read_events()->wait_event_interruptible_hrtimeout(), > > which will call schedule(), it does not account IO wait time. > > Not sure if there isn't some misunderstanding so I'll repeat what I've > said: Yes, above path will not account as IO wait time and IMO that is much > more common path which should be accounted as IO wait time. So I think that > without fixing that path, fixing cornercases like you did in your patch is > rather pointless. > > Honza > > > On 08/26/2020 21:23, Jan Kara wrote: > > On Wed 05-08-20 09:35:51, Xianting Tian wrote: > > > When waiting for the completion of io, we need account iowait time. As > > > wait_for_completion() calls schedule_timeout(), which doesn't account > > > iowait time. While wait_for_completion_io() calls io_schedule_timeout(), > > > which will account iowait time. > > > > > > So using wait_for_completion_io() instead of wait_for_completion() > > > when waiting for completion of io before exit_aio and io_destroy. > > > > > > Signed-off-by: Xianting Tian <xianting_tian@126.com> > > > > Thanks for the patch! It looks good to me but IMO this is just scratching > > the surface. E.g. for AIO we are mostly going to wait in read_events() by > > wait_event_interruptible_hrtimeout() and *that* doesn't account as IO wait > > either? Which is IMO far bigger misaccounting... The two case you fix seem > > to be just rare cornercases so what they do isn't a big deal either way. > > > > So I agree it may be worth it to properly account waiting for AIO but if > > you want to do that, then please handle mainly the common cases in AIO > > code. > > > > Honza > > > > > --- > > > fs/aio.c | 4 ++-- > > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > > > diff --git a/fs/aio.c b/fs/aio.c > > > index 91e7cc4..498b8a0 100644 > > > --- a/fs/aio.c > > > +++ b/fs/aio.c > > > @@ -892,7 +892,7 @@ void exit_aio(struct mm_struct *mm) > > > > > > if (!atomic_sub_and_test(skipped, &wait.count)) { > > > /* Wait until all IO for the context are done. */ > > > - wait_for_completion(&wait.comp); > > > + wait_for_completion_io(&wait.comp); > > > } > > > > > > RCU_INIT_POINTER(mm->ioctx_table, NULL); > > > @@ -1400,7 +1400,7 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr, > > > * is destroyed. > > > */ > > > if (!ret) > > > - wait_for_completion(&wait.comp); > > > + wait_for_completion_io(&wait.comp); > > > > > > return ret; > > > } > > > -- > > > 1.8.3.1 > > > > > -- > > Jan Kara <jack@suse.com> > > SUSE Labs, CR > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR
diff --git a/fs/aio.c b/fs/aio.c index 91e7cc4..498b8a0 100644 --- a/fs/aio.c +++ b/fs/aio.c @@ -892,7 +892,7 @@ void exit_aio(struct mm_struct *mm) if (!atomic_sub_and_test(skipped, &wait.count)) { /* Wait until all IO for the context are done. */ - wait_for_completion(&wait.comp); + wait_for_completion_io(&wait.comp); } RCU_INIT_POINTER(mm->ioctx_table, NULL); @@ -1400,7 +1400,7 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr, * is destroyed. */ if (!ret) - wait_for_completion(&wait.comp); + wait_for_completion_io(&wait.comp); return ret; }
When waiting for the completion of io, we need account iowait time. As wait_for_completion() calls schedule_timeout(), which doesn't account iowait time. While wait_for_completion_io() calls io_schedule_timeout(), which will account iowait time. So using wait_for_completion_io() instead of wait_for_completion() when waiting for completion of io before exit_aio and io_destroy. Signed-off-by: Xianting Tian <xianting_tian@126.com> --- fs/aio.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)