* [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio @ 2013-03-04 16:45 Paul Taysom 2013-03-04 17:19 ` Paul Menzel 2013-03-04 17:42 ` Alasdair G Kergon 0 siblings, 2 replies; 9+ messages in thread From: Paul Taysom @ 2013-03-04 16:45 UTC (permalink / raw) To: agk Cc: dm-devel, neilb, linux-raid, linux-kernel, msb, mpatocka, olofj, Paul Taysom Changed the dm-verity prefetching to use a worker thread to avoid a deadlock in dm-bufio. If generic_make_request is called recursively, it queues the I/O request on the current->bio_list without making the I/O request and returns. The routine making the recursive call cannot wait for the I/O to complete. The deadlock occurred when one thread grabbed the bufio_client mutex and waited for an I/O to complete but the I/O was queued on another thread’s current->bio_list and it was waiting to get the mutex held by the first thread. The fix allows only one I/O request from dm-verity to dm-bufio per thread. To do this, the prefetch requests were queued on worker threads. In addition to avoiding the deadlock, this fix made a slight improvement in performance. seconds_kernel_to_login: with prefetch: 8.43s without prefetch: 9.2s worker prefetch: 8.28s Signed-off-by: Paul Taysom <taysom@chromium.org> --- drivers/md/dm-verity.c | 29 +++++++++++++++++++++++++++-- 1 file changed, 27 insertions(+), 2 deletions(-) diff --git a/drivers/md/dm-verity.c b/drivers/md/dm-verity.c index 52cde98..7313498 100644 --- a/drivers/md/dm-verity.c +++ b/drivers/md/dm-verity.c @@ -93,6 +93,13 @@ struct dm_verity_io { */ }; +struct dm_verity_prefetch_work { + struct work_struct work; + struct dm_bufio_client *bufio; + sector_t block; + unsigned n_blocks; +}; + static struct shash_desc *io_hash_desc(struct dm_verity *v, struct dm_verity_io *io) { return (struct shash_desc *)(io + 1); @@ -419,6 +426,17 @@ static void verity_end_io(struct bio *bio, int error) queue_work(io->v->verify_wq, &io->work); } + +static void do_verity_prefetch_work(struct work_struct *work) +{ + struct dm_verity_prefetch_work *vw = + container_of(work, struct dm_verity_prefetch_work, work); + + dm_bufio_prefetch(vw->bufio, vw->block, vw->n_blocks); + + kfree(vw); +} + /* * Prefetch buffers for the specified io. * The root buffer is not prefetched, it is assumed that it will be cached @@ -427,6 +445,7 @@ static void verity_end_io(struct bio *bio, int error) static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) { int i; + struct dm_verity_prefetch_work *vw; for (i = v->levels - 2; i >= 0; i--) { sector_t hash_block_start; @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) hash_block_end = v->hash_blocks - 1; } no_prefetch_cluster: - dm_bufio_prefetch(v->bufio, hash_block_start, - hash_block_end - hash_block_start + 1); + vw = kmalloc(sizeof(*vw), GFP_KERNEL); + if (!vw) /* Just prefetching, ignore errors */ + return; + vw->bufio = v->bufio; + vw->block = hash_block_start; + vw->n_blocks = hash_block_end - hash_block_start + 1; + INIT_WORK(&vw->work, do_verity_prefetch_work); + queue_work(v->verify_wq, &vw->work); } } -- 1.8.1.3 -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply related [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 16:45 [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio Paul Taysom @ 2013-03-04 17:19 ` Paul Menzel 2013-03-04 21:21 ` Paul Taysom 2013-03-04 17:42 ` Alasdair G Kergon 1 sibling, 1 reply; 9+ messages in thread From: Paul Menzel @ 2013-03-04 17:19 UTC (permalink / raw) To: Paul Taysom Cc: agk, dm-devel, neilb, linux-raid, linux-kernel, msb, mpatocka, olofj [-- Attachment #1: Type: text/plain, Size: 1496 bytes --] Am Montag, den 04.03.2013, 08:45 -0800 schrieb Paul Taysom: > Changed the dm-verity prefetching to use a worker thread to avoid > a deadlock in dm-bufio. Do you know of any bug reports about this? Searching for »dm-verity dead lock« I found [1], which seems to be the issue, right? If yes, please reference it. > If generic_make_request is called recursively, it queues the I/O > request on the current->bio_list without making the I/O request > and returns. The routine making the recursive call cannot wait > for the I/O to complete. > > The deadlock occurred when one thread grabbed the bufio_client > mutex and waited for an I/O to complete but the I/O was queued > on another thread’s current->bio_list and it was waiting to get > the mutex held by the first thread. > > The fix allows only one I/O request from dm-verity to dm-bufio > per thread. To do this, the prefetch requests were queued on worker > threads. > > In addition to avoiding the deadlock, this fix made a slight > improvement in performance. > > seconds_kernel_to_login: > with prefetch: 8.43s > without prefetch: 9.2s > worker prefetch: 8.28s > > Signed-off-by: Paul Taysom <taysom@chromium.org> > --- > drivers/md/dm-verity.c | 29 +++++++++++++++++++++++++++-- > 1 file changed, 27 insertions(+), 2 deletions(-) Acked-by: Paul Menzel <paulepanter@users.sourceforge.net> Thanks, Paul [1] https://code.google.com/p/chromium-os/issues/detail?id=39148 [-- Attachment #2: This is a digitally signed message part --] [-- Type: application/pgp-signature, Size: 198 bytes --] ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 17:19 ` Paul Menzel @ 2013-03-04 21:21 ` Paul Taysom 0 siblings, 0 replies; 9+ messages in thread From: Paul Taysom @ 2013-03-04 21:21 UTC (permalink / raw) To: Paul Menzel Cc: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Mikulas Patocka, Olof Johansson On Mon, Mar 4, 2013 at 9:19 AM, Paul Menzel <pm.debian@googlemail.com> wrote: > > Am Montag, den 04.03.2013, 08:45 -0800 schrieb Paul Taysom: > > Changed the dm-verity prefetching to use a worker thread to avoid > > a deadlock in dm-bufio. > > Do you know of any bug reports about this? Searching for »dm-verity dead > lock« I found [1], which seems to be the issue, right? If yes, please > reference it. > > > If generic_make_request is called recursively, it queues the I/O > > request on the current->bio_list without making the I/O request > > and returns. The routine making the recursive call cannot wait > > for the I/O to complete. > > > > The deadlock occurred when one thread grabbed the bufio_client > > mutex and waited for an I/O to complete but the I/O was queued > > on another thread’s current->bio_list and it was waiting to get > > the mutex held by the first thread. > > > > The fix allows only one I/O request from dm-verity to dm-bufio > > per thread. To do this, the prefetch requests were queued on worker > > threads. > > > > In addition to avoiding the deadlock, this fix made a slight > > improvement in performance. > > > > seconds_kernel_to_login: > > with prefetch: 8.43s > > without prefetch: 9.2s > > worker prefetch: 8.28s > > > > Signed-off-by: Paul Taysom <taysom@chromium.org> > > --- > > drivers/md/dm-verity.c | 29 +++++++++++++++++++++++++++-- > > 1 file changed, 27 insertions(+), 2 deletions(-) > > Acked-by: Paul Menzel <paulepanter@users.sourceforge.net> > > > Thanks, > > Paul > > > [1] https://code.google.com/p/chromium-os/issues/detail?id=39148 Yes, I should have included that this fix addresses http://crosbug.com/39148 Thanks, Paul ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 16:45 [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio Paul Taysom 2013-03-04 17:19 ` Paul Menzel @ 2013-03-04 17:42 ` Alasdair G Kergon 2013-03-04 21:23 ` Paul Taysom 1 sibling, 1 reply; 9+ messages in thread From: Alasdair G Kergon @ 2013-03-04 17:42 UTC (permalink / raw) To: Paul Taysom Cc: agk, dm-devel, neilb, linux-raid, linux-kernel, msb, mpatocka, olofj On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: > @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) > hash_block_end = v->hash_blocks - 1; > } > no_prefetch_cluster: > - dm_bufio_prefetch(v->bufio, hash_block_start, > - hash_block_end - hash_block_start + 1); > + vw = kmalloc(sizeof(*vw), GFP_KERNEL); kmalloc? mempool? ... Alasdair ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 17:42 ` Alasdair G Kergon @ 2013-03-04 21:23 ` Paul Taysom 2013-03-04 22:15 ` Mikulas Patocka 0 siblings, 1 reply; 9+ messages in thread From: Paul Taysom @ 2013-03-04 21:23 UTC (permalink / raw) To: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Mikulas Patocka, Olof Johansson On Mon, Mar 4, 2013 at 9:42 AM, Alasdair G Kergon <agk@redhat.com> wrote: > On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: >> @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >> hash_block_end = v->hash_blocks - 1; >> } >> no_prefetch_cluster: >> - dm_bufio_prefetch(v->bufio, hash_block_start, >> - hash_block_end - hash_block_start + 1); >> + vw = kmalloc(sizeof(*vw), GFP_KERNEL); > > kmalloc? mempool? ... > > Alasdair > The use of mempool would be a separate patch that would have to be measured for performance impact. -Paul ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 21:23 ` Paul Taysom @ 2013-03-04 22:15 ` Mikulas Patocka 2013-03-05 23:30 ` Paul Taysom 0 siblings, 1 reply; 9+ messages in thread From: Mikulas Patocka @ 2013-03-04 22:15 UTC (permalink / raw) To: Paul Taysom Cc: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Olof Johansson On Mon, 4 Mar 2013, Paul Taysom wrote: > On Mon, Mar 4, 2013 at 9:42 AM, Alasdair G Kergon <agk@redhat.com> wrote: > > On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: > >> @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) > >> hash_block_end = v->hash_blocks - 1; > >> } > >> no_prefetch_cluster: > >> - dm_bufio_prefetch(v->bufio, hash_block_start, > >> - hash_block_end - hash_block_start + 1); > >> + vw = kmalloc(sizeof(*vw), GFP_KERNEL); > > > > kmalloc? mempool? ... > > > > Alasdair > > > The use of mempool would be a separate patch that would have to be > measured for performance impact. > -Paul > You don't have to use mempool. Just avoid prefetching if there is not enough memory for the prefetch structure. I reworked the patch, is uses an allocation that can fail and it generates just one workqueue entry for one request (the original patch generated one workqueue entry for hash tree level). Please test the patch and if it works and performs well, let's submit it. Mikulas --- Changed the dm-verity prefetching to use a worker thread to avoid a deadlock in dm-bufio. If generic_make_request is called recursively, it queues the I/O request on the current->bio_list without making the I/O request and returns. The routine making the recursive call cannot wait for the I/O to complete. The deadlock occurred when one thread grabbed the bufio_client mutex and waited for an I/O to complete but the I/O was queued on another thread's current->bio_list and it was waiting to get the mutex held by the first thread. The fix allows only one I/O request from dm-verity to dm-bufio per thread. To do this, the prefetch requests were queued on worker threads. In addition to avoiding the deadlock, this fix made a slight improvement in performance. seconds_kernel_to_login: with prefetch: 8.43s without prefetch: 9.2s worker prefetch: 8.28s Signed-off-by: Paul Taysom <taysom@chromium.org> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> Cc: stable@kernel.org --- drivers/md/dm-bufio.c | 2 ++ drivers/md/dm-verity.c | 37 +++++++++++++++++++++++++++++++++---- 2 files changed, 35 insertions(+), 4 deletions(-) Index: linux-3.8-fast/drivers/md/dm-verity.c =================================================================== --- linux-3.8-fast.orig/drivers/md/dm-verity.c 2013-03-04 22:49:20.000000000 +0100 +++ linux-3.8-fast/drivers/md/dm-verity.c 2013-03-04 23:10:45.000000000 +0100 @@ -93,6 +93,13 @@ struct dm_verity_io { */ }; +struct dm_verity_prefetch_work { + struct work_struct work; + struct dm_verity *v; + sector_t block; + unsigned n_blocks; +}; + static struct shash_desc *io_hash_desc(struct dm_verity *v, struct dm_verity_io *io) { return (struct shash_desc *)(io + 1); @@ -424,15 +431,18 @@ static void verity_end_io(struct bio *bi * The root buffer is not prefetched, it is assumed that it will be cached * all the time. */ -static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) +static void verity_prefetch_io(struct work_struct *work) { + struct dm_verity_prefetch_work *pw = + container_of(work, struct dm_verity_prefetch_work, work); + struct dm_verity *v = pw->v; int i; for (i = v->levels - 2; i >= 0; i--) { sector_t hash_block_start; sector_t hash_block_end; - verity_hash_at_level(v, io->block, i, &hash_block_start, NULL); - verity_hash_at_level(v, io->block + io->n_blocks - 1, i, &hash_block_end, NULL); + verity_hash_at_level(v, pw->block, i, &hash_block_start, NULL); + verity_hash_at_level(v, pw->block + pw->n_blocks - 1, i, &hash_block_end, NULL); if (!i) { unsigned cluster = ACCESS_ONCE(dm_verity_prefetch_cluster); @@ -452,6 +462,25 @@ no_prefetch_cluster: dm_bufio_prefetch(v->bufio, hash_block_start, hash_block_end - hash_block_start + 1); } + + kfree(pw); +} + +static void verity_submit_prefetch(struct dm_verity *v, struct dm_verity_io *io) +{ + struct dm_verity_prefetch_work *pw; + + pw = kmalloc(sizeof(struct dm_verity_prefetch_work), + GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); + + if (!pw) + return; + + INIT_WORK(&pw->work, verity_prefetch_io); + pw->v = v; + pw->block = io->block; + pw->n_blocks = io->n_blocks; + queue_work(v->verify_wq, &pw->work); } /* @@ -498,7 +527,7 @@ static int verity_map(struct dm_target * memcpy(io->io_vec, bio_iovec(bio), io->io_vec_size * sizeof(struct bio_vec)); - verity_prefetch_io(v, io); + verity_submit_prefetch(v, io); generic_make_request(bio); Index: linux-3.8-fast/drivers/md/dm-bufio.c =================================================================== --- linux-3.8-fast.orig/drivers/md/dm-bufio.c 2013-03-04 23:03:14.000000000 +0100 +++ linux-3.8-fast/drivers/md/dm-bufio.c 2013-03-04 23:04:19.000000000 +0100 @@ -1026,6 +1026,8 @@ void dm_bufio_prefetch(struct dm_bufio_c { struct blk_plug plug; + BUG_ON(dm_bufio_in_request()); + blk_start_plug(&plug); dm_bufio_lock(c); ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-04 22:15 ` Mikulas Patocka @ 2013-03-05 23:30 ` Paul Taysom 2013-03-12 14:46 ` Paul Taysom 0 siblings, 1 reply; 9+ messages in thread From: Paul Taysom @ 2013-03-05 23:30 UTC (permalink / raw) To: Mikulas Patocka Cc: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Olof Johansson seconds_kernel_to_login: 8.17s Passed all my other tests On Mon, Mar 4, 2013 at 2:15 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Mon, 4 Mar 2013, Paul Taysom wrote: > >> On Mon, Mar 4, 2013 at 9:42 AM, Alasdair G Kergon <agk@redhat.com> wrote: >> > On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: >> >> @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >> >> hash_block_end = v->hash_blocks - 1; >> >> } >> >> no_prefetch_cluster: >> >> - dm_bufio_prefetch(v->bufio, hash_block_start, >> >> - hash_block_end - hash_block_start + 1); >> >> + vw = kmalloc(sizeof(*vw), GFP_KERNEL); >> > >> > kmalloc? mempool? ... >> > >> > Alasdair >> > >> The use of mempool would be a separate patch that would have to be >> measured for performance impact. >> -Paul >> > > You don't have to use mempool. Just avoid prefetching if there is not > enough memory for the prefetch structure. > > I reworked the patch, is uses an allocation that can fail and it generates > just one workqueue entry for one request (the original patch generated one > workqueue entry for hash tree level). > > Please test the patch and if it works and performs well, let's submit it. > > Mikulas > > --- > > Changed the dm-verity prefetching to use a worker thread to avoid > a deadlock in dm-bufio. > > If generic_make_request is called recursively, it queues the I/O > request on the current->bio_list without making the I/O request > and returns. The routine making the recursive call cannot wait > for the I/O to complete. > > The deadlock occurred when one thread grabbed the bufio_client > mutex and waited for an I/O to complete but the I/O was queued > on another thread's current->bio_list and it was waiting to get > the mutex held by the first thread. > > The fix allows only one I/O request from dm-verity to dm-bufio > per thread. To do this, the prefetch requests were queued on worker > threads. > > In addition to avoiding the deadlock, this fix made a slight > improvement in performance. > > seconds_kernel_to_login: > with prefetch: 8.43s > without prefetch: 9.2s > worker prefetch: 8.28s > > Signed-off-by: Paul Taysom <taysom@chromium.org> > Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> > Cc: stable@kernel.org > --- > drivers/md/dm-bufio.c | 2 ++ > drivers/md/dm-verity.c | 37 +++++++++++++++++++++++++++++++++---- > 2 files changed, 35 insertions(+), 4 deletions(-) > > Index: linux-3.8-fast/drivers/md/dm-verity.c > =================================================================== > --- linux-3.8-fast.orig/drivers/md/dm-verity.c 2013-03-04 22:49:20.000000000 +0100 > +++ linux-3.8-fast/drivers/md/dm-verity.c 2013-03-04 23:10:45.000000000 +0100 > @@ -93,6 +93,13 @@ struct dm_verity_io { > */ > }; > > +struct dm_verity_prefetch_work { > + struct work_struct work; > + struct dm_verity *v; > + sector_t block; > + unsigned n_blocks; > +}; > + > static struct shash_desc *io_hash_desc(struct dm_verity *v, struct dm_verity_io *io) > { > return (struct shash_desc *)(io + 1); > @@ -424,15 +431,18 @@ static void verity_end_io(struct bio *bi > * The root buffer is not prefetched, it is assumed that it will be cached > * all the time. > */ > -static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) > +static void verity_prefetch_io(struct work_struct *work) > { > + struct dm_verity_prefetch_work *pw = > + container_of(work, struct dm_verity_prefetch_work, work); > + struct dm_verity *v = pw->v; > int i; > > for (i = v->levels - 2; i >= 0; i--) { > sector_t hash_block_start; > sector_t hash_block_end; > - verity_hash_at_level(v, io->block, i, &hash_block_start, NULL); > - verity_hash_at_level(v, io->block + io->n_blocks - 1, i, &hash_block_end, NULL); > + verity_hash_at_level(v, pw->block, i, &hash_block_start, NULL); > + verity_hash_at_level(v, pw->block + pw->n_blocks - 1, i, &hash_block_end, NULL); > if (!i) { > unsigned cluster = ACCESS_ONCE(dm_verity_prefetch_cluster); > > @@ -452,6 +462,25 @@ no_prefetch_cluster: > dm_bufio_prefetch(v->bufio, hash_block_start, > hash_block_end - hash_block_start + 1); > } > + > + kfree(pw); > +} > + > +static void verity_submit_prefetch(struct dm_verity *v, struct dm_verity_io *io) > +{ > + struct dm_verity_prefetch_work *pw; > + > + pw = kmalloc(sizeof(struct dm_verity_prefetch_work), > + GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); > + > + if (!pw) > + return; > + > + INIT_WORK(&pw->work, verity_prefetch_io); > + pw->v = v; > + pw->block = io->block; > + pw->n_blocks = io->n_blocks; > + queue_work(v->verify_wq, &pw->work); > } > > /* > @@ -498,7 +527,7 @@ static int verity_map(struct dm_target * > memcpy(io->io_vec, bio_iovec(bio), > io->io_vec_size * sizeof(struct bio_vec)); > > - verity_prefetch_io(v, io); > + verity_submit_prefetch(v, io); > > generic_make_request(bio); > > Index: linux-3.8-fast/drivers/md/dm-bufio.c > =================================================================== > --- linux-3.8-fast.orig/drivers/md/dm-bufio.c 2013-03-04 23:03:14.000000000 +0100 > +++ linux-3.8-fast/drivers/md/dm-bufio.c 2013-03-04 23:04:19.000000000 +0100 > @@ -1026,6 +1026,8 @@ void dm_bufio_prefetch(struct dm_bufio_c > { > struct blk_plug plug; > > + BUG_ON(dm_bufio_in_request()); > + > blk_start_plug(&plug); > dm_bufio_lock(c); > ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-05 23:30 ` Paul Taysom @ 2013-03-12 14:46 ` Paul Taysom 2013-03-13 20:18 ` Paul Taysom 0 siblings, 1 reply; 9+ messages in thread From: Paul Taysom @ 2013-03-12 14:46 UTC (permalink / raw) To: Mikulas Patocka Cc: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Olof Johansson Looks good to me. On Tue, Mar 5, 2013 at 3:30 PM, Paul Taysom <taysom@google.com> wrote: > seconds_kernel_to_login: 8.17s > Passed all my other tests > > On Mon, Mar 4, 2013 at 2:15 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> >> >> On Mon, 4 Mar 2013, Paul Taysom wrote: >> >>> On Mon, Mar 4, 2013 at 9:42 AM, Alasdair G Kergon <agk@redhat.com> wrote: >>> > On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: >>> >> @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >>> >> hash_block_end = v->hash_blocks - 1; >>> >> } >>> >> no_prefetch_cluster: >>> >> - dm_bufio_prefetch(v->bufio, hash_block_start, >>> >> - hash_block_end - hash_block_start + 1); >>> >> + vw = kmalloc(sizeof(*vw), GFP_KERNEL); >>> > >>> > kmalloc? mempool? ... >>> > >>> > Alasdair >>> > >>> The use of mempool would be a separate patch that would have to be >>> measured for performance impact. >>> -Paul >>> >> >> You don't have to use mempool. Just avoid prefetching if there is not >> enough memory for the prefetch structure. >> >> I reworked the patch, is uses an allocation that can fail and it generates >> just one workqueue entry for one request (the original patch generated one >> workqueue entry for hash tree level). >> >> Please test the patch and if it works and performs well, let's submit it. >> >> Mikulas >> >> --- >> >> Changed the dm-verity prefetching to use a worker thread to avoid >> a deadlock in dm-bufio. >> >> If generic_make_request is called recursively, it queues the I/O >> request on the current->bio_list without making the I/O request >> and returns. The routine making the recursive call cannot wait >> for the I/O to complete. >> >> The deadlock occurred when one thread grabbed the bufio_client >> mutex and waited for an I/O to complete but the I/O was queued >> on another thread's current->bio_list and it was waiting to get >> the mutex held by the first thread. >> >> The fix allows only one I/O request from dm-verity to dm-bufio >> per thread. To do this, the prefetch requests were queued on worker >> threads. >> >> In addition to avoiding the deadlock, this fix made a slight >> improvement in performance. >> >> seconds_kernel_to_login: >> with prefetch: 8.43s >> without prefetch: 9.2s >> worker prefetch: 8.28s >> >> Signed-off-by: Paul Taysom <taysom@chromium.org> >> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> >> Cc: stable@kernel.org >> --- >> drivers/md/dm-bufio.c | 2 ++ >> drivers/md/dm-verity.c | 37 +++++++++++++++++++++++++++++++++---- >> 2 files changed, 35 insertions(+), 4 deletions(-) >> >> Index: linux-3.8-fast/drivers/md/dm-verity.c >> =================================================================== >> --- linux-3.8-fast.orig/drivers/md/dm-verity.c 2013-03-04 22:49:20.000000000 +0100 >> +++ linux-3.8-fast/drivers/md/dm-verity.c 2013-03-04 23:10:45.000000000 +0100 >> @@ -93,6 +93,13 @@ struct dm_verity_io { >> */ >> }; >> >> +struct dm_verity_prefetch_work { >> + struct work_struct work; >> + struct dm_verity *v; >> + sector_t block; >> + unsigned n_blocks; >> +}; >> + >> static struct shash_desc *io_hash_desc(struct dm_verity *v, struct dm_verity_io *io) >> { >> return (struct shash_desc *)(io + 1); >> @@ -424,15 +431,18 @@ static void verity_end_io(struct bio *bi >> * The root buffer is not prefetched, it is assumed that it will be cached >> * all the time. >> */ >> -static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >> +static void verity_prefetch_io(struct work_struct *work) >> { >> + struct dm_verity_prefetch_work *pw = >> + container_of(work, struct dm_verity_prefetch_work, work); >> + struct dm_verity *v = pw->v; >> int i; >> >> for (i = v->levels - 2; i >= 0; i--) { >> sector_t hash_block_start; >> sector_t hash_block_end; >> - verity_hash_at_level(v, io->block, i, &hash_block_start, NULL); >> - verity_hash_at_level(v, io->block + io->n_blocks - 1, i, &hash_block_end, NULL); >> + verity_hash_at_level(v, pw->block, i, &hash_block_start, NULL); >> + verity_hash_at_level(v, pw->block + pw->n_blocks - 1, i, &hash_block_end, NULL); >> if (!i) { >> unsigned cluster = ACCESS_ONCE(dm_verity_prefetch_cluster); >> >> @@ -452,6 +462,25 @@ no_prefetch_cluster: >> dm_bufio_prefetch(v->bufio, hash_block_start, >> hash_block_end - hash_block_start + 1); >> } >> + >> + kfree(pw); >> +} >> + >> +static void verity_submit_prefetch(struct dm_verity *v, struct dm_verity_io *io) >> +{ >> + struct dm_verity_prefetch_work *pw; >> + >> + pw = kmalloc(sizeof(struct dm_verity_prefetch_work), >> + GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); >> + >> + if (!pw) >> + return; >> + >> + INIT_WORK(&pw->work, verity_prefetch_io); >> + pw->v = v; >> + pw->block = io->block; >> + pw->n_blocks = io->n_blocks; >> + queue_work(v->verify_wq, &pw->work); >> } >> >> /* >> @@ -498,7 +527,7 @@ static int verity_map(struct dm_target * >> memcpy(io->io_vec, bio_iovec(bio), >> io->io_vec_size * sizeof(struct bio_vec)); >> >> - verity_prefetch_io(v, io); >> + verity_submit_prefetch(v, io); >> >> generic_make_request(bio); >> >> Index: linux-3.8-fast/drivers/md/dm-bufio.c >> =================================================================== >> --- linux-3.8-fast.orig/drivers/md/dm-bufio.c 2013-03-04 23:03:14.000000000 +0100 >> +++ linux-3.8-fast/drivers/md/dm-bufio.c 2013-03-04 23:04:19.000000000 +0100 >> @@ -1026,6 +1026,8 @@ void dm_bufio_prefetch(struct dm_bufio_c >> { >> struct blk_plug plug; >> >> + BUG_ON(dm_bufio_in_request()); >> + >> blk_start_plug(&plug); >> dm_bufio_lock(c); >> ^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio 2013-03-12 14:46 ` Paul Taysom @ 2013-03-13 20:18 ` Paul Taysom 0 siblings, 0 replies; 9+ messages in thread From: Paul Taysom @ 2013-03-13 20:18 UTC (permalink / raw) To: Mikulas Patocka Cc: Paul Taysom, agk, dm-devel, neilb, linux-raid, linux-kernel, Mandeep Singh Baines, Olof Johansson On Tue, Mar 12, 2013 at 7:46 AM, Paul Taysom <taysom@google.com> wrote: > Looks good to me. > > On Tue, Mar 5, 2013 at 3:30 PM, Paul Taysom <taysom@google.com> wrote: >> seconds_kernel_to_login: 8.17s >> Passed all my other tests >> >> On Mon, Mar 4, 2013 at 2:15 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >>> >>> >>> On Mon, 4 Mar 2013, Paul Taysom wrote: >>> >>>> On Mon, Mar 4, 2013 at 9:42 AM, Alasdair G Kergon <agk@redhat.com> wrote: >>>> > On Mon, Mar 04, 2013 at 08:45:48AM -0800, Paul Taysom wrote: >>>> >> @@ -449,8 +468,14 @@ static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >>>> >> hash_block_end = v->hash_blocks - 1; >>>> >> } >>>> >> no_prefetch_cluster: >>>> >> - dm_bufio_prefetch(v->bufio, hash_block_start, >>>> >> - hash_block_end - hash_block_start + 1); >>>> >> + vw = kmalloc(sizeof(*vw), GFP_KERNEL); >>>> > >>>> > kmalloc? mempool? ... >>>> > >>>> > Alasdair >>>> > >>>> The use of mempool would be a separate patch that would have to be >>>> measured for performance impact. >>>> -Paul >>>> >>> >>> You don't have to use mempool. Just avoid prefetching if there is not >>> enough memory for the prefetch structure. >>> >>> I reworked the patch, is uses an allocation that can fail and it generates >>> just one workqueue entry for one request (the original patch generated one >>> workqueue entry for hash tree level). >>> >>> Please test the patch and if it works and performs well, let's submit it. >>> >>> Mikulas >>> >>> --- >>> >>> Changed the dm-verity prefetching to use a worker thread to avoid >>> a deadlock in dm-bufio. >>> >>> If generic_make_request is called recursively, it queues the I/O >>> request on the current->bio_list without making the I/O request >>> and returns. The routine making the recursive call cannot wait >>> for the I/O to complete. >>> >>> The deadlock occurred when one thread grabbed the bufio_client >>> mutex and waited for an I/O to complete but the I/O was queued >>> on another thread's current->bio_list and it was waiting to get >>> the mutex held by the first thread. >>> >>> The fix allows only one I/O request from dm-verity to dm-bufio >>> per thread. To do this, the prefetch requests were queued on worker >>> threads. >>> >>> In addition to avoiding the deadlock, this fix made a slight >>> improvement in performance. >>> >>> seconds_kernel_to_login: >>> with prefetch: 8.43s >>> without prefetch: 9.2s >>> worker prefetch: 8.28s >>> >>> Signed-off-by: Paul Taysom <taysom@chromium.org> >>> Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> >>> Cc: stable@kernel.org >>> --- >>> drivers/md/dm-bufio.c | 2 ++ >>> drivers/md/dm-verity.c | 37 +++++++++++++++++++++++++++++++++---- >>> 2 files changed, 35 insertions(+), 4 deletions(-) >>> >>> Index: linux-3.8-fast/drivers/md/dm-verity.c >>> =================================================================== >>> --- linux-3.8-fast.orig/drivers/md/dm-verity.c 2013-03-04 22:49:20.000000000 +0100 >>> +++ linux-3.8-fast/drivers/md/dm-verity.c 2013-03-04 23:10:45.000000000 +0100 >>> @@ -93,6 +93,13 @@ struct dm_verity_io { >>> */ >>> }; >>> >>> +struct dm_verity_prefetch_work { >>> + struct work_struct work; >>> + struct dm_verity *v; >>> + sector_t block; >>> + unsigned n_blocks; >>> +}; >>> + >>> static struct shash_desc *io_hash_desc(struct dm_verity *v, struct dm_verity_io *io) >>> { >>> return (struct shash_desc *)(io + 1); >>> @@ -424,15 +431,18 @@ static void verity_end_io(struct bio *bi >>> * The root buffer is not prefetched, it is assumed that it will be cached >>> * all the time. >>> */ >>> -static void verity_prefetch_io(struct dm_verity *v, struct dm_verity_io *io) >>> +static void verity_prefetch_io(struct work_struct *work) >>> { >>> + struct dm_verity_prefetch_work *pw = >>> + container_of(work, struct dm_verity_prefetch_work, work); >>> + struct dm_verity *v = pw->v; >>> int i; >>> >>> for (i = v->levels - 2; i >= 0; i--) { >>> sector_t hash_block_start; >>> sector_t hash_block_end; >>> - verity_hash_at_level(v, io->block, i, &hash_block_start, NULL); >>> - verity_hash_at_level(v, io->block + io->n_blocks - 1, i, &hash_block_end, NULL); >>> + verity_hash_at_level(v, pw->block, i, &hash_block_start, NULL); >>> + verity_hash_at_level(v, pw->block + pw->n_blocks - 1, i, &hash_block_end, NULL); >>> if (!i) { >>> unsigned cluster = ACCESS_ONCE(dm_verity_prefetch_cluster); >>> >>> @@ -452,6 +462,25 @@ no_prefetch_cluster: >>> dm_bufio_prefetch(v->bufio, hash_block_start, >>> hash_block_end - hash_block_start + 1); >>> } >>> + >>> + kfree(pw); >>> +} >>> + >>> +static void verity_submit_prefetch(struct dm_verity *v, struct dm_verity_io *io) >>> +{ >>> + struct dm_verity_prefetch_work *pw; >>> + >>> + pw = kmalloc(sizeof(struct dm_verity_prefetch_work), >>> + GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); >>> + >>> + if (!pw) >>> + return; >>> + >>> + INIT_WORK(&pw->work, verity_prefetch_io); >>> + pw->v = v; >>> + pw->block = io->block; >>> + pw->n_blocks = io->n_blocks; >>> + queue_work(v->verify_wq, &pw->work); >>> } >>> >>> /* >>> @@ -498,7 +527,7 @@ static int verity_map(struct dm_target * >>> memcpy(io->io_vec, bio_iovec(bio), >>> io->io_vec_size * sizeof(struct bio_vec)); >>> >>> - verity_prefetch_io(v, io); >>> + verity_submit_prefetch(v, io); >>> >>> generic_make_request(bio); >>> >>> Index: linux-3.8-fast/drivers/md/dm-bufio.c >>> =================================================================== >>> --- linux-3.8-fast.orig/drivers/md/dm-bufio.c 2013-03-04 23:03:14.000000000 +0100 >>> +++ linux-3.8-fast/drivers/md/dm-bufio.c 2013-03-04 23:04:19.000000000 +0100 >>> @@ -1026,6 +1026,8 @@ void dm_bufio_prefetch(struct dm_bufio_c >>> { >>> struct blk_plug plug; >>> >>> + BUG_ON(dm_bufio_in_request()); >>> + >>> blk_start_plug(&plug); >>> dm_bufio_lock(c); >>> Signed-off-by: Paul Taysom <taysom@chromium.org> ^ permalink raw reply [flat|nested] 9+ messages in thread
end of thread, other threads:[~2013-03-13 20:18 UTC | newest] Thread overview: 9+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2013-03-04 16:45 [PATCH] md: dm-verity: Fix to avoid a deadlock in dm-bufio Paul Taysom 2013-03-04 17:19 ` Paul Menzel 2013-03-04 21:21 ` Paul Taysom 2013-03-04 17:42 ` Alasdair G Kergon 2013-03-04 21:23 ` Paul Taysom 2013-03-04 22:15 ` Mikulas Patocka 2013-03-05 23:30 ` Paul Taysom 2013-03-12 14:46 ` Paul Taysom 2013-03-13 20:18 ` Paul Taysom
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).