| Message ID | 1461805789-3632-3-git-send-email-ming.lei@canonical.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, 28 Apr 2016, Ming Lei wrote: > There were reports about heavy stack use by recursive calling > .bi_end_io()([1][2][3]). For example, more than 16K stack is > consumed in a single bio complete path[3], and in [2] stack > overflow can be triggered if 20 nested dm-crypt is used. > > Also patches[1] [2] [3] were posted for addressing the issue, > but never be merged. And the idea in these patches is basically > similar, all serializes the recursive calling of .bi_end_io() by > percpu list. > > This patch still takes the same idea, but uses bio_list to > implement it, which turns out more simple and the code becomes > more readable meantime. > > One corner case which wasn't covered before is that > bi_endio() may be scheduled to run in process context(such > as btrfs), and this patch just bypasses the optimizing for > that case because one new context should have enough stack space, > and this approach isn't capable of optimizing it too because > there isn't easy way to get a per-task linked list head. Hi You could use preempt_disable() and then you could use per-cpu list even in the process context. Mikulas > xfstests(-g auto) is run with this patch and no regression is > found on ext4, xfs and btrfs. > > [1] http://marc.info/?t=121428502000004&r=1&w=2 > [2] http://marc.info/?l=dm-devel&m=139595190620008&w=2 > [3] http://marc.info/?t=145974644100001&r=1&w=2 > > Cc: Shaun Tancheff <shaun.tancheff@seagate.com> > Cc: Christoph Hellwig <hch@infradead.org> > Cc: Mikulas Patocka <mpatocka@redhat.com> > Cc: Alan Cox <alan@linux.intel.com> > Cc: Neil Brown <neilb@suse.de> > Cc: Liu Bo <bo.li.liu@oracle.com> > Signed-off-by: Ming Lei <ming.lei@canonical.com> > --- > block/bio.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++++++++-- > 1 file changed, 53 insertions(+), 2 deletions(-) > > diff --git a/block/bio.c b/block/bio.c > index 807d25e..e20a83c 100644 > --- a/block/bio.c > +++ b/block/bio.c > @@ -68,6 +68,8 @@ static DEFINE_MUTEX(bio_slab_lock); > static struct bio_slab *bio_slabs; > static unsigned int bio_slab_nr, bio_slab_max; > > +static DEFINE_PER_CPU(struct bio_list *, bio_end_list) = { NULL }; > + > static struct kmem_cache *bio_find_or_create_slab(unsigned int extra_size) > { > unsigned int sz = sizeof(struct bio) + extra_size; > @@ -1737,6 +1739,56 @@ static inline bool bio_remaining_done(struct bio *bio) > return false; > } > > +static void __bio_endio(struct bio *bio) > +{ > + if (bio->bi_end_io) > + bio->bi_end_io(bio); > +} > + > +/* disable local irq when manipulating the percpu bio_list */ > +static void unwind_bio_endio(struct bio *bio) > +{ > + struct bio_list *bl; > + unsigned long flags; > + > + /* > + * We can't optimize if bi_endio() is scheduled to run from > + * process context because there isn't easy way to get a > + * per-task bio list head or allocate a per-task variable. > + */ > + if (!in_interrupt()) { > + /* > + * It has to be a top calling when it is run from > + * process context. > + */ > + WARN_ON(this_cpu_read(bio_end_list)); > + __bio_endio(bio); > + return; > + } > + > + local_irq_save(flags); > + bl = __this_cpu_read(bio_end_list); > + if (bl) { > + bio_list_add(bl, bio); > + } else { > + struct bio_list bl_in_stack; > + > + bl = &bl_in_stack; > + bio_list_init(bl); > + > + __this_cpu_write(bio_end_list, bl); > + while (bio) { > + local_irq_restore(flags); > + __bio_endio(bio); > + local_irq_save(flags); > + > + bio = bio_list_pop(bl); > + } > + __this_cpu_write(bio_end_list, NULL); > + } > + local_irq_restore(flags); > +} > + > /** > * bio_endio - end I/O on a bio > * @bio: bio > @@ -1765,8 +1817,7 @@ again: > goto again; > } > > - if (bio->bi_end_io) > - bio->bi_end_io(bio); > + unwind_bio_endio(bio); > } > EXPORT_SYMBOL(bio_endio); > > -- > 1.9.1 > -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Mikulas, On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Thu, 28 Apr 2016, Ming Lei wrote: > >> There were reports about heavy stack use by recursive calling >> .bi_end_io()([1][2][3]). For example, more than 16K stack is >> consumed in a single bio complete path[3], and in [2] stack >> overflow can be triggered if 20 nested dm-crypt is used. >> >> Also patches[1] [2] [3] were posted for addressing the issue, >> but never be merged. And the idea in these patches is basically >> similar, all serializes the recursive calling of .bi_end_io() by >> percpu list. >> >> This patch still takes the same idea, but uses bio_list to >> implement it, which turns out more simple and the code becomes >> more readable meantime. >> >> One corner case which wasn't covered before is that >> bi_endio() may be scheduled to run in process context(such >> as btrfs), and this patch just bypasses the optimizing for >> that case because one new context should have enough stack space, >> and this approach isn't capable of optimizing it too because >> there isn't easy way to get a per-task linked list head. > > Hi > > You could use preempt_disable() and then you could use per-cpu list even > in the process context. Image why the .bi_end_io() is scheduled to process context, and the only workable/simple way I thought of is to use per-task list because it may sleep. Given new context should have enough stack and only btrfs has this kind of usage as far as I see, so don't think that is worth of the optimization. Thanks, Ming > > Mikulas > >> xfstests(-g auto) is run with this patch and no regression is >> found on ext4, xfs and btrfs. >> >> [1] http://marc.info/?t=121428502000004&r=1&w=2 >> [2] http://marc.info/?l=dm-devel&m=139595190620008&w=2 >> [3] http://marc.info/?t=145974644100001&r=1&w=2 >> >> Cc: Shaun Tancheff <shaun.tancheff@seagate.com> >> Cc: Christoph Hellwig <hch@infradead.org> >> Cc: Mikulas Patocka <mpatocka@redhat.com> >> Cc: Alan Cox <alan@linux.intel.com> >> Cc: Neil Brown <neilb@suse.de> >> Cc: Liu Bo <bo.li.liu@oracle.com> >> Signed-off-by: Ming Lei <ming.lei@canonical.com> >> --- >> block/bio.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++++++++-- >> 1 file changed, 53 insertions(+), 2 deletions(-) >> >> diff --git a/block/bio.c b/block/bio.c >> index 807d25e..e20a83c 100644 >> --- a/block/bio.c >> +++ b/block/bio.c >> @@ -68,6 +68,8 @@ static DEFINE_MUTEX(bio_slab_lock); >> static struct bio_slab *bio_slabs; >> static unsigned int bio_slab_nr, bio_slab_max; >> >> +static DEFINE_PER_CPU(struct bio_list *, bio_end_list) = { NULL }; >> + >> static struct kmem_cache *bio_find_or_create_slab(unsigned int extra_size) >> { >> unsigned int sz = sizeof(struct bio) + extra_size; >> @@ -1737,6 +1739,56 @@ static inline bool bio_remaining_done(struct bio *bio) >> return false; >> } >> >> +static void __bio_endio(struct bio *bio) >> +{ >> + if (bio->bi_end_io) >> + bio->bi_end_io(bio); >> +} >> + >> +/* disable local irq when manipulating the percpu bio_list */ >> +static void unwind_bio_endio(struct bio *bio) >> +{ >> + struct bio_list *bl; >> + unsigned long flags; >> + >> + /* >> + * We can't optimize if bi_endio() is scheduled to run from >> + * process context because there isn't easy way to get a >> + * per-task bio list head or allocate a per-task variable. >> + */ >> + if (!in_interrupt()) { >> + /* >> + * It has to be a top calling when it is run from >> + * process context. >> + */ >> + WARN_ON(this_cpu_read(bio_end_list)); >> + __bio_endio(bio); >> + return; >> + } >> + >> + local_irq_save(flags); >> + bl = __this_cpu_read(bio_end_list); >> + if (bl) { >> + bio_list_add(bl, bio); >> + } else { >> + struct bio_list bl_in_stack; >> + >> + bl = &bl_in_stack; >> + bio_list_init(bl); >> + >> + __this_cpu_write(bio_end_list, bl); >> + while (bio) { >> + local_irq_restore(flags); >> + __bio_endio(bio); >> + local_irq_save(flags); >> + >> + bio = bio_list_pop(bl); >> + } >> + __this_cpu_write(bio_end_list, NULL); >> + } >> + local_irq_restore(flags); >> +} >> + >> /** >> * bio_endio - end I/O on a bio >> * @bio: bio >> @@ -1765,8 +1817,7 @@ again: >> goto again; >> } >> >> - if (bio->bi_end_io) >> - bio->bi_end_io(bio); >> + unwind_bio_endio(bio); >> } >> EXPORT_SYMBOL(bio_endio); >> >> -- >> 1.9.1 >> > -- > To unsubscribe from this list: send the line "unsubscribe linux-block" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, 28 Apr 2016, Ming Lei wrote: > Hi Mikulas, > > On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > On Thu, 28 Apr 2016, Ming Lei wrote: > > > >> There were reports about heavy stack use by recursive calling > >> .bi_end_io()([1][2][3]). For example, more than 16K stack is > >> consumed in a single bio complete path[3], and in [2] stack > >> overflow can be triggered if 20 nested dm-crypt is used. > >> > >> Also patches[1] [2] [3] were posted for addressing the issue, > >> but never be merged. And the idea in these patches is basically > >> similar, all serializes the recursive calling of .bi_end_io() by > >> percpu list. > >> > >> This patch still takes the same idea, but uses bio_list to > >> implement it, which turns out more simple and the code becomes > >> more readable meantime. > >> > >> One corner case which wasn't covered before is that > >> bi_endio() may be scheduled to run in process context(such > >> as btrfs), and this patch just bypasses the optimizing for > >> that case because one new context should have enough stack space, > >> and this approach isn't capable of optimizing it too because > >> there isn't easy way to get a per-task linked list head. > > > > Hi > > > > You could use preempt_disable() and then you could use per-cpu list even > > in the process context. > > Image why the .bi_end_io() is scheduled to process context, and the only > workable/simple way I thought of is to use per-task list because it may sleep. The bi_end_io callback should not sleep, even if it is called from the process context. > Given new context should have enough stack and only btrfs has this kind of > usage as far as I see, so don't think that is worth of the optimization. > > Thanks, > Ming Mikulas -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Apr 28, 2016 at 11:58 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Thu, 28 Apr 2016, Ming Lei wrote: > >> Hi Mikulas, >> >> On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> > >> > >> > On Thu, 28 Apr 2016, Ming Lei wrote: >> > >> >> There were reports about heavy stack use by recursive calling >> >> .bi_end_io()([1][2][3]). For example, more than 16K stack is >> >> consumed in a single bio complete path[3], and in [2] stack >> >> overflow can be triggered if 20 nested dm-crypt is used. >> >> >> >> Also patches[1] [2] [3] were posted for addressing the issue, >> >> but never be merged. And the idea in these patches is basically >> >> similar, all serializes the recursive calling of .bi_end_io() by >> >> percpu list. >> >> >> >> This patch still takes the same idea, but uses bio_list to >> >> implement it, which turns out more simple and the code becomes >> >> more readable meantime. >> >> >> >> One corner case which wasn't covered before is that >> >> bi_endio() may be scheduled to run in process context(such >> >> as btrfs), and this patch just bypasses the optimizing for >> >> that case because one new context should have enough stack space, >> >> and this approach isn't capable of optimizing it too because >> >> there isn't easy way to get a per-task linked list head. >> > >> > Hi >> > >> > You could use preempt_disable() and then you could use per-cpu list even >> > in the process context. >> >> Image why the .bi_end_io() is scheduled to process context, and the only >> workable/simple way I thought of is to use per-task list because it may sleep. > > The bi_end_io callback should not sleep, even if it is called from the > process context. If it shouldn't sleep, why is it scheduled to run in process context by paying extra context switch cost? And you can find that btrfs_subio_endio_read() does sleep for checksum stuff. Thanks, Ming > >> Given new context should have enough stack and only btrfs has this kind of >> usage as far as I see, so don't think that is worth of the optimization. >> >> Thanks, >> Ming > > Mikulas > -- > To unsubscribe from this list: send the line "unsubscribe linux-block" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 29 Apr 2016, Ming Lei wrote: > On Thu, Apr 28, 2016 at 11:58 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > On Thu, 28 Apr 2016, Ming Lei wrote: > > > >> Hi Mikulas, > >> > >> On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > >> > > >> > > >> > On Thu, 28 Apr 2016, Ming Lei wrote: > >> > > >> >> There were reports about heavy stack use by recursive calling > >> >> .bi_end_io()([1][2][3]). For example, more than 16K stack is > >> >> consumed in a single bio complete path[3], and in [2] stack > >> >> overflow can be triggered if 20 nested dm-crypt is used. > >> >> > >> >> Also patches[1] [2] [3] were posted for addressing the issue, > >> >> but never be merged. And the idea in these patches is basically > >> >> similar, all serializes the recursive calling of .bi_end_io() by > >> >> percpu list. > >> >> > >> >> This patch still takes the same idea, but uses bio_list to > >> >> implement it, which turns out more simple and the code becomes > >> >> more readable meantime. > >> >> > >> >> One corner case which wasn't covered before is that > >> >> bi_endio() may be scheduled to run in process context(such > >> >> as btrfs), and this patch just bypasses the optimizing for > >> >> that case because one new context should have enough stack space, > >> >> and this approach isn't capable of optimizing it too because > >> >> there isn't easy way to get a per-task linked list head. > >> > > >> > Hi > >> > > >> > You could use preempt_disable() and then you could use per-cpu list even > >> > in the process context. > >> > >> Image why the .bi_end_io() is scheduled to process context, and the only > >> workable/simple way I thought of is to use per-task list because it may sleep. > > > > The bi_end_io callback should not sleep, even if it is called from the > > process context. > > If it shouldn't sleep, why is it scheduled to run in process context by paying > extra context switch cost? Some device mapper (and other) drivers use a worker thread to process bios. So the bio may be finished from the worker thread. It would be advantageous to prevent stack overflow even in this case. > And you can find that btrfs_subio_endio_read() does sleep for checksum stuff. I'm not an expert on btrfs. What happens if it is called from an interrupt? Do you have an actual stracktrace when this function is called from bio_endio and when it sleeps? > Thanks, > Ming Mikulas -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Apr 29, 2016 at 12:59 AM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Fri, 29 Apr 2016, Ming Lei wrote: > >> On Thu, Apr 28, 2016 at 11:58 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> > >> > >> > On Thu, 28 Apr 2016, Ming Lei wrote: >> > >> >> Hi Mikulas, >> >> >> >> On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> >> > >> >> > >> >> > On Thu, 28 Apr 2016, Ming Lei wrote: >> >> > >> >> >> There were reports about heavy stack use by recursive calling >> >> >> .bi_end_io()([1][2][3]). For example, more than 16K stack is >> >> >> consumed in a single bio complete path[3], and in [2] stack >> >> >> overflow can be triggered if 20 nested dm-crypt is used. >> >> >> >> >> >> Also patches[1] [2] [3] were posted for addressing the issue, >> >> >> but never be merged. And the idea in these patches is basically >> >> >> similar, all serializes the recursive calling of .bi_end_io() by >> >> >> percpu list. >> >> >> >> >> >> This patch still takes the same idea, but uses bio_list to >> >> >> implement it, which turns out more simple and the code becomes >> >> >> more readable meantime. >> >> >> >> >> >> One corner case which wasn't covered before is that >> >> >> bi_endio() may be scheduled to run in process context(such >> >> >> as btrfs), and this patch just bypasses the optimizing for >> >> >> that case because one new context should have enough stack space, >> >> >> and this approach isn't capable of optimizing it too because >> >> >> there isn't easy way to get a per-task linked list head. >> >> > >> >> > Hi >> >> > >> >> > You could use preempt_disable() and then you could use per-cpu list even >> >> > in the process context. >> >> >> >> Image why the .bi_end_io() is scheduled to process context, and the only >> >> workable/simple way I thought of is to use per-task list because it may sleep. >> > >> > The bi_end_io callback should not sleep, even if it is called from the >> > process context. >> >> If it shouldn't sleep, why is it scheduled to run in process context by paying >> extra context switch cost? > > Some device mapper (and other) drivers use a worker thread to process > bios. So the bio may be finished from the worker thread. It would be > advantageous to prevent stack overflow even in this case. If the .bi_end_io wouldn't sleep, it can be put back into interrupt context for the sake of performance when this patch is merged. The cost of context switch in high IOPS case isn't cheap. It isn't easy to avoid the recursive calling in process context except you can add something 'task_struct' or introduce 'alloca()' in kernel. Or do you have better ideas? > >> And you can find that btrfs_subio_endio_read() does sleep for checksum stuff. > > I'm not an expert on btrfs. What happens if it is called from an > interrupt? Do you have an actual stracktrace when this function is called What do you expect if sleepable function is called in softirq or hardirq handler? :-) > from bio_endio and when it sleeps? The problem is observed in xfstests generic/323 by this patch v1. Sometimes the test hangs, and sometimes kernel oops is triggered. and the issue is fixed by introducing 'if (!in_interrupt())' block for handling running .bi_end_io() from process context. If the block of 'if (!in_interrupt())' is removed and preempt_disable()/preempt_enable() is added around bio->bi_end_io(), the following kernel warning can be seen easily: [ 51.086303] BUG: sleeping function called from invalid context at mm/slab.h:388 [ 51.087442] in_atomic(): 1, irqs_disabled(): 0, pid: 633, name: kworker/u8:4 [ 51.088575] CPU: 3 PID: 633 Comm: kworker/u8:4 Not tainted 4.6.0-rc3+ #2017 [ 51.088578] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.9.0-0-g01a84be-prebuilt.qemu-project.org 04/01/2014 [ 51.088637] Workqueue: btrfs-endio btrfs_endio_helper [btrfs] [ 51.088640] 0000000000000000 ffff88007bbebc00 ffffffff813d92d3 ffff88007ba6ce00 [ 51.088643] 0000000000000184 ffff88007bbebc18 ffffffff810a38bb ffffffff81a35310 [ 51.088645] ffff88007bbebc40 ffffffff810a3949 0000000002400040 0000000002400040 [ 51.088648] Call Trace: [ 51.088651] [<ffffffff813d92d3>] dump_stack+0x63/0x90 [ 51.088655] [<ffffffff810a38bb>] ___might_sleep+0xdb/0x120 [ 51.088657] [<ffffffff810a3949>] __might_sleep+0x49/0x80 [ 51.088659] [<ffffffff811e98e7>] kmem_cache_alloc+0x1a7/0x210 [ 51.088670] [<ffffffffc0102741>] ? alloc_extent_state+0x21/0xe0 [btrfs] [ 51.088680] [<ffffffffc0102741>] alloc_extent_state+0x21/0xe0 [btrfs] [ 51.088689] [<ffffffffc01046ce>] __clear_extent_bit+0x2ae/0x3d0 [btrfs] [ 51.088698] [<ffffffffc0104e6a>] clear_extent_bit+0x2a/0x30 [btrfs] [ 51.088708] [<ffffffffc00e96d0>] btrfs_endio_direct_read+0x70/0xf0 [btrfs] [ 51.088711] [<ffffffff8139f1e7>] bio_endio+0xf7/0x140 [ 51.088718] [<ffffffffc00dbb9c>] end_workqueue_fn+0x3c/0x40 [btrfs] [ 51.088728] [<ffffffffc01184d7>] normal_work_helper+0xc7/0x310 [btrfs] [ 51.088737] [<ffffffffc01187f2>] btrfs_endio_helper+0x12/0x20 [btrfs] [ 51.088740] [<ffffffff81097b77>] process_one_work+0x157/0x420 [ 51.088742] [<ffffffff810985ab>] worker_thread+0x12b/0x4d0 [ 51.088744] [<ffffffff817171a8>] ? __schedule+0x368/0x950 [ 51.088746] [<ffffffff81098480>] ? rescuer_thread+0x380/0x380 [ 51.088748] [<ffffffff8109de84>] kthread+0xd4/0xf0 [ 51.088750] [<ffffffff8171b7a2>] ret_from_fork+0x22/0x40 [ 51.088752] [<ffffffff8109ddb0>] ? kthread_park+0x60/0x60 Not mention wait_for_completion() in both __btrfs_correct_data_nocsum() and __btrfs_subio_endio_read(), which are called by btrfs_subio_endio_read() in btrfs_endio_direct_read(). Thanks, Ming > >> Thanks, >> Ming > > Mikulas > -- > To unsubscribe from this list: send the line "unsubscribe linux-block" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 29 Apr 2016, Ming Lei wrote: > On Fri, Apr 29, 2016 at 12:59 AM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > > > > On Fri, 29 Apr 2016, Ming Lei wrote: > > > >> On Thu, Apr 28, 2016 at 11:58 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > >> > > >> > > >> > On Thu, 28 Apr 2016, Ming Lei wrote: > >> > > >> >> Hi Mikulas, > >> >> > >> >> On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > >> >> > > >> >> > > >> >> > On Thu, 28 Apr 2016, Ming Lei wrote: > >> >> > > >> >> >> There were reports about heavy stack use by recursive calling > >> >> >> .bi_end_io()([1][2][3]). For example, more than 16K stack is > >> >> >> consumed in a single bio complete path[3], and in [2] stack > >> >> >> overflow can be triggered if 20 nested dm-crypt is used. > >> >> >> > >> >> >> Also patches[1] [2] [3] were posted for addressing the issue, > >> >> >> but never be merged. And the idea in these patches is basically > >> >> >> similar, all serializes the recursive calling of .bi_end_io() by > >> >> >> percpu list. > >> >> >> > >> >> >> This patch still takes the same idea, but uses bio_list to > >> >> >> implement it, which turns out more simple and the code becomes > >> >> >> more readable meantime. > >> >> >> > >> >> >> One corner case which wasn't covered before is that > >> >> >> bi_endio() may be scheduled to run in process context(such > >> >> >> as btrfs), and this patch just bypasses the optimizing for > >> >> >> that case because one new context should have enough stack space, > >> >> >> and this approach isn't capable of optimizing it too because > >> >> >> there isn't easy way to get a per-task linked list head. > >> >> > > >> >> > Hi > >> >> > > >> >> > You could use preempt_disable() and then you could use per-cpu list even > >> >> > in the process context. > >> >> > >> >> Image why the .bi_end_io() is scheduled to process context, and the only > >> >> workable/simple way I thought of is to use per-task list because it may sleep. > >> > > >> > The bi_end_io callback should not sleep, even if it is called from the > >> > process context. > >> > >> If it shouldn't sleep, why is it scheduled to run in process context by paying > >> extra context switch cost? > > > > Some device mapper (and other) drivers use a worker thread to process > > bios. So the bio may be finished from the worker thread. It would be > > advantageous to prevent stack overflow even in this case. > > If the .bi_end_io wouldn't sleep, it can be put back into interrupt context > for the sake of performance when this patch is merged. The cost of context > switch in high IOPS case isn't cheap. If some block device driver in a process context finds out that it needs to terminate a bio, it calls bio_endio in a process context. Why would it need to trigger an interrupt just to call bio_endio? (and how could it trigger an interrupt if that driver perhaps doesn't use interrupts at all?) I don't know what are you trying to suggest. > It isn't easy to avoid the recursive calling in process context except you > can add something 'task_struct' or introduce 'alloca()' in kernel. Or do you > have better ideas? preempt_disable around the bi_endio callback should be sufficient. > > > >> And you can find that btrfs_subio_endio_read() does sleep for checksum stuff. > > > > I'm not an expert on btrfs. What happens if it is called from an > > interrupt? Do you have an actual stracktrace when this function is called > > What do you expect if sleepable function is called in softirq or > hardirq handler? :-) > > > from bio_endio and when it sleeps? > > The problem is observed in xfstests generic/323 by this patch v1. Sometimes the > test hangs, and sometimes kernel oops is triggered. and the issue is > fixed by introducing 'if (!in_interrupt())' block for handling running > .bi_end_io() from > process context. > > If the block of 'if (!in_interrupt())' is removed and > preempt_disable()/preempt_enable() is added around bio->bi_end_io(), > the following kernel warning can be seen easily: > > [ 51.086303] BUG: sleeping function called from invalid context at > mm/slab.h:388 > [ 51.087442] in_atomic(): 1, irqs_disabled(): 0, pid: 633, name: kworker/u8:4 > [ 51.088575] CPU: 3 PID: 633 Comm: kworker/u8:4 Not tainted 4.6.0-rc3+ #2017 > [ 51.088578] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), > BIOS rel-1.9.0-0-g01a84be-prebuilt.qemu-project.org 04/01/2014 > [ 51.088637] Workqueue: btrfs-endio btrfs_endio_helper [btrfs] > [ 51.088640] 0000000000000000 ffff88007bbebc00 ffffffff813d92d3 > ffff88007ba6ce00 > [ 51.088643] 0000000000000184 ffff88007bbebc18 ffffffff810a38bb > ffffffff81a35310 > [ 51.088645] ffff88007bbebc40 ffffffff810a3949 0000000002400040 > 0000000002400040 > [ 51.088648] Call Trace: > [ 51.088651] [<ffffffff813d92d3>] dump_stack+0x63/0x90 > [ 51.088655] [<ffffffff810a38bb>] ___might_sleep+0xdb/0x120 > [ 51.088657] [<ffffffff810a3949>] __might_sleep+0x49/0x80 > [ 51.088659] [<ffffffff811e98e7>] kmem_cache_alloc+0x1a7/0x210 > [ 51.088670] [<ffffffffc0102741>] ? alloc_extent_state+0x21/0xe0 [btrfs] > [ 51.088680] [<ffffffffc0102741>] alloc_extent_state+0x21/0xe0 [btrfs] > [ 51.088689] [<ffffffffc01046ce>] __clear_extent_bit+0x2ae/0x3d0 [btrfs] > [ 51.088698] [<ffffffffc0104e6a>] clear_extent_bit+0x2a/0x30 [btrfs] > [ 51.088708] [<ffffffffc00e96d0>] btrfs_endio_direct_read+0x70/0xf0 [btrfs] > [ 51.088711] [<ffffffff8139f1e7>] bio_endio+0xf7/0x140 > [ 51.088718] [<ffffffffc00dbb9c>] end_workqueue_fn+0x3c/0x40 [btrfs] > [ 51.088728] [<ffffffffc01184d7>] normal_work_helper+0xc7/0x310 [btrfs] > [ 51.088737] [<ffffffffc01187f2>] btrfs_endio_helper+0x12/0x20 [btrfs] > [ 51.088740] [<ffffffff81097b77>] process_one_work+0x157/0x420 > [ 51.088742] [<ffffffff810985ab>] worker_thread+0x12b/0x4d0 > [ 51.088744] [<ffffffff817171a8>] ? __schedule+0x368/0x950 > [ 51.088746] [<ffffffff81098480>] ? rescuer_thread+0x380/0x380 > [ 51.088748] [<ffffffff8109de84>] kthread+0xd4/0xf0 > [ 51.088750] [<ffffffff8171b7a2>] ret_from_fork+0x22/0x40 > [ 51.088752] [<ffffffff8109ddb0>] ? kthread_park+0x60/0x60 > > > Not mention wait_for_completion() in both __btrfs_correct_data_nocsum() > and __btrfs_subio_endio_read(), which are called by btrfs_subio_endio_read() > in btrfs_endio_direct_read(). btrfs is calling bio_endio on bio that it created. So, bio_endio in btrfs can be replaced with: if (bio->bi_end_io == btrfs_endio_direct_read) btrfs_endio_direct_read(bio); else bio_endio(bio); ... or just maybe just with this: bio->bi_end_io(bio); This could be fixed easily. Mikulas > Thanks, > Ming > > > > >> Thanks, > >> Ming > > > > Mikulas > > -- > > To unsubscribe from this list: send the line "unsubscribe linux-block" in > > the body of a message to majordomo@vger.kernel.org > > More majordomo info at http://vger.kernel.org/majordomo-info.html > -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Apr 29, 2016 at 4:39 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: > > > On Fri, 29 Apr 2016, Ming Lei wrote: > >> On Fri, Apr 29, 2016 at 12:59 AM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> > >> > >> > On Fri, 29 Apr 2016, Ming Lei wrote: >> > >> >> On Thu, Apr 28, 2016 at 11:58 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> >> > >> >> > >> >> > On Thu, 28 Apr 2016, Ming Lei wrote: >> >> > >> >> >> Hi Mikulas, >> >> >> >> >> >> On Thu, Apr 28, 2016 at 11:29 PM, Mikulas Patocka <mpatocka@redhat.com> wrote: >> >> >> > >> >> >> > >> >> >> > On Thu, 28 Apr 2016, Ming Lei wrote: >> >> >> > >> >> >> >> There were reports about heavy stack use by recursive calling >> >> >> >> .bi_end_io()([1][2][3]). For example, more than 16K stack is >> >> >> >> consumed in a single bio complete path[3], and in [2] stack >> >> >> >> overflow can be triggered if 20 nested dm-crypt is used. >> >> >> >> >> >> >> >> Also patches[1] [2] [3] were posted for addressing the issue, >> >> >> >> but never be merged. And the idea in these patches is basically >> >> >> >> similar, all serializes the recursive calling of .bi_end_io() by >> >> >> >> percpu list. >> >> >> >> >> >> >> >> This patch still takes the same idea, but uses bio_list to >> >> >> >> implement it, which turns out more simple and the code becomes >> >> >> >> more readable meantime. >> >> >> >> >> >> >> >> One corner case which wasn't covered before is that >> >> >> >> bi_endio() may be scheduled to run in process context(such >> >> >> >> as btrfs), and this patch just bypasses the optimizing for >> >> >> >> that case because one new context should have enough stack space, >> >> >> >> and this approach isn't capable of optimizing it too because >> >> >> >> there isn't easy way to get a per-task linked list head. >> >> >> > >> >> >> > Hi >> >> >> > >> >> >> > You could use preempt_disable() and then you could use per-cpu list even >> >> >> > in the process context. >> >> >> >> >> >> Image why the .bi_end_io() is scheduled to process context, and the only >> >> >> workable/simple way I thought of is to use per-task list because it may sleep. >> >> > >> >> > The bi_end_io callback should not sleep, even if it is called from the >> >> > process context. >> >> >> >> If it shouldn't sleep, why is it scheduled to run in process context by paying >> >> extra context switch cost? >> > >> > Some device mapper (and other) drivers use a worker thread to process >> > bios. So the bio may be finished from the worker thread. It would be >> > advantageous to prevent stack overflow even in this case. >> >> If the .bi_end_io wouldn't sleep, it can be put back into interrupt context >> for the sake of performance when this patch is merged. The cost of context >> switch in high IOPS case isn't cheap. > > If some block device driver in a process context finds out that it needs > to terminate a bio, it calls bio_endio in a process context. Why would it > need to trigger an interrupt just to call bio_endio? (and how could it > trigger an interrupt if that driver perhaps doesn't use interrupts at > all?) I don't know what are you trying to suggest. That should be the failure path, and this patch just aims at the normal bio complete path which is run at 99.999% time. And irq handler often borrows current process's stack and cause huge stack uses. > >> It isn't easy to avoid the recursive calling in process context except you >> can add something 'task_struct' or introduce 'alloca()' in kernel. Or do you >> have better ideas? > > preempt_disable around the bi_endio callback should be sufficient. How can a sleepable function be run in preempt disabled context? Looks you still don't believe the bi_endio scheduled to process context can sleep, do you? > >> > >> >> And you can find that btrfs_subio_endio_read() does sleep for checksum stuff. >> > >> > I'm not an expert on btrfs. What happens if it is called from an >> > interrupt? Do you have an actual stracktrace when this function is called >> >> What do you expect if sleepable function is called in softirq or >> hardirq handler? :-) >> >> > from bio_endio and when it sleeps? >> >> The problem is observed in xfstests generic/323 by this patch v1. Sometimes the >> test hangs, and sometimes kernel oops is triggered. and the issue is >> fixed by introducing 'if (!in_interrupt())' block for handling running >> .bi_end_io() from >> process context. >> >> If the block of 'if (!in_interrupt())' is removed and >> preempt_disable()/preempt_enable() is added around bio->bi_end_io(), >> the following kernel warning can be seen easily: >> >> [ 51.086303] BUG: sleeping function called from invalid context at >> mm/slab.h:388 >> [ 51.087442] in_atomic(): 1, irqs_disabled(): 0, pid: 633, name: kworker/u8:4 >> [ 51.088575] CPU: 3 PID: 633 Comm: kworker/u8:4 Not tainted 4.6.0-rc3+ #2017 >> [ 51.088578] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), >> BIOS rel-1.9.0-0-g01a84be-prebuilt.qemu-project.org 04/01/2014 >> [ 51.088637] Workqueue: btrfs-endio btrfs_endio_helper [btrfs] >> [ 51.088640] 0000000000000000 ffff88007bbebc00 ffffffff813d92d3 >> ffff88007ba6ce00 >> [ 51.088643] 0000000000000184 ffff88007bbebc18 ffffffff810a38bb >> ffffffff81a35310 >> [ 51.088645] ffff88007bbebc40 ffffffff810a3949 0000000002400040 >> 0000000002400040 >> [ 51.088648] Call Trace: >> [ 51.088651] [<ffffffff813d92d3>] dump_stack+0x63/0x90 >> [ 51.088655] [<ffffffff810a38bb>] ___might_sleep+0xdb/0x120 >> [ 51.088657] [<ffffffff810a3949>] __might_sleep+0x49/0x80 >> [ 51.088659] [<ffffffff811e98e7>] kmem_cache_alloc+0x1a7/0x210 >> [ 51.088670] [<ffffffffc0102741>] ? alloc_extent_state+0x21/0xe0 [btrfs] >> [ 51.088680] [<ffffffffc0102741>] alloc_extent_state+0x21/0xe0 [btrfs] >> [ 51.088689] [<ffffffffc01046ce>] __clear_extent_bit+0x2ae/0x3d0 [btrfs] >> [ 51.088698] [<ffffffffc0104e6a>] clear_extent_bit+0x2a/0x30 [btrfs] >> [ 51.088708] [<ffffffffc00e96d0>] btrfs_endio_direct_read+0x70/0xf0 [btrfs] >> [ 51.088711] [<ffffffff8139f1e7>] bio_endio+0xf7/0x140 >> [ 51.088718] [<ffffffffc00dbb9c>] end_workqueue_fn+0x3c/0x40 [btrfs] >> [ 51.088728] [<ffffffffc01184d7>] normal_work_helper+0xc7/0x310 [btrfs] >> [ 51.088737] [<ffffffffc01187f2>] btrfs_endio_helper+0x12/0x20 [btrfs] >> [ 51.088740] [<ffffffff81097b77>] process_one_work+0x157/0x420 >> [ 51.088742] [<ffffffff810985ab>] worker_thread+0x12b/0x4d0 >> [ 51.088744] [<ffffffff817171a8>] ? __schedule+0x368/0x950 >> [ 51.088746] [<ffffffff81098480>] ? rescuer_thread+0x380/0x380 >> [ 51.088748] [<ffffffff8109de84>] kthread+0xd4/0xf0 >> [ 51.088750] [<ffffffff8171b7a2>] ret_from_fork+0x22/0x40 >> [ 51.088752] [<ffffffff8109ddb0>] ? kthread_park+0x60/0x60 >> >> >> Not mention wait_for_completion() in both __btrfs_correct_data_nocsum() >> and __btrfs_subio_endio_read(), which are called by btrfs_subio_endio_read() >> in btrfs_endio_direct_read(). > > btrfs is calling bio_endio on bio that it created. So, bio_endio in btrfs > can be replaced with: > if (bio->bi_end_io == btrfs_endio_direct_read) > btrfs_endio_direct_read(bio); > else > bio_endio(bio); > > ... or just maybe just with this: > bio->bi_end_io(bio); It is really ugly and should be avoided most of times. > > This could be fixed easily. Suppose it might be done in this way, and there may be other .bi_end_io which can sleep too, you need to make sure there isn't any such usage first before running it with preempt disabled. Anyway looks you worry about the seldom-run failure path, which isn't addressed by this patch. And you can optimize that in future and that isn't contradictory with this patch. Thanks, Ming > > Mikulas > >> Thanks, >> Ming >> >> > >> >> Thanks, >> >> Ming >> > >> > Mikulas >> > -- >> > To unsubscribe from this list: send the line "unsubscribe linux-block" in >> > the body of a message to majordomo@vger.kernel.org >> > More majordomo info at http://vger.kernel.org/majordomo-info.html >> > -- > To unsubscribe from this list: send the line "unsubscribe linux-block" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 29 Apr 2016, Ming Lei wrote: > > If some block device driver in a process context finds out that it needs > > to terminate a bio, it calls bio_endio in a process context. Why would it > > need to trigger an interrupt just to call bio_endio? (and how could it > > trigger an interrupt if that driver perhaps doesn't use interrupts at > > all?) I don't know what are you trying to suggest. > > That should be the failure path, and this patch just aims at the normal bio > complete path which is run at 99.999% time. And irq handler often > borrows current process's stack and cause huge stack uses. For some block devices it is common to call bi_endio from process context, for example dm-crypt or dm-verity do it for all read bios. loop driver calls bi_endio from process context on every bio it processes. > >> It isn't easy to avoid the recursive calling in process context except you > >> can add something 'task_struct' or introduce 'alloca()' in kernel. Or do you > >> have better ideas? > > > > preempt_disable around the bi_endio callback should be sufficient. > > How can a sleepable function be run in preempt disabled context? > > Looks you still don't believe the bi_endio scheduled to process context > can sleep, do you? I don't believe bi_endio function can sleep. bi_endio function doesn't know if it will be called from process or interrupt context - so it can't sleep. If bi_endio function slept, bug would happen if it were called from interrupt context. btrfs is a special case where btrfs calls bio_endio on bios that it created on it own (note that this trick only works when the same driver creates a bio and terminates the bio with bio_endio - if some other driver terminated the bio, it wouldn't work because the other driver could terminate the bio from interrupt context). That could be simply fixed by calling bio->bi_end_io or btrfs_endio_direct_read directly or making a new function btrfs_endio. > >> Not mention wait_for_completion() in both __btrfs_correct_data_nocsum() > >> and __btrfs_subio_endio_read(), which are called by btrfs_subio_endio_read() > >> in btrfs_endio_direct_read(). > > > > btrfs is calling bio_endio on bio that it created. So, bio_endio in btrfs > > can be replaced with: > > if (bio->bi_end_io == btrfs_endio_direct_read) > > btrfs_endio_direct_read(bio); > > else > > bio_endio(bio); > > > > ... or just maybe just with this: > > bio->bi_end_io(bio); > > It is really ugly and should be avoided most of times. > > > > > This could be fixed easily. > > Suppose it might be done in this way, and there may be other .bi_end_io > which can sleep too, you need to make sure there isn't any such usage > first before running it with preempt disabled. > > Anyway looks you worry about the seldom-run failure path, which isn't > addressed by this patch. And you can optimize that in future and that isn't > contradictory with this patch. Worrying about seldom-run failrure paths is nothing wrong. These paths are not any less important than the common code paths. Mikulas > Thanks, > Ming -- To unsubscribe from this list: send the line "unsubscribe linux-block" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 04/27/2016 07:09 PM, Ming Lei wrote: > There were reports about heavy stack use by recursive calling > .bi_end_io()([1][2][3]). For example, more than 16K stack is > consumed in a single bio complete path[3], and in [2] stack > overflow can be triggered if 20 nested dm-crypt is used. > > Also patches[1] [2] [3] were posted for addressing the issue, > but never be merged. And the idea in these patches is basically > similar, all serializes the recursive calling of .bi_end_io() by > percpu list. > > This patch still takes the same idea, but uses bio_list to > implement it, which turns out more simple and the code becomes > more readable meantime. > > One corner case which wasn't covered before is that > bi_endio() may be scheduled to run in process context(such > as btrfs), and this patch just bypasses the optimizing for > that case because one new context should have enough stack space, > and this approach isn't capable of optimizing it too because > there isn't easy way to get a per-task linked list head. > > xfstests(-g auto) is run with this patch and no regression is > found on ext4, xfs and btrfs. I don't like this at all, it's going about fixing this the completely wrong way. For the fast path, single device, we don't have an issue. Yet this patch slows down that part unnecessarily. Fix this for the fat dm/md stacks, and don't add unwarranted fat in the normal fast path. The dm/md stacks won't notice a bit of extra overhead, whereas the core path is pretty well optimized.
diff --git a/block/bio.c b/block/bio.c index 807d25e..e20a83c 100644 --- a/block/bio.c +++ b/block/bio.c @@ -68,6 +68,8 @@ static DEFINE_MUTEX(bio_slab_lock); static struct bio_slab *bio_slabs; static unsigned int bio_slab_nr, bio_slab_max; +static DEFINE_PER_CPU(struct bio_list *, bio_end_list) = { NULL }; + static struct kmem_cache *bio_find_or_create_slab(unsigned int extra_size) { unsigned int sz = sizeof(struct bio) + extra_size; @@ -1737,6 +1739,56 @@ static inline bool bio_remaining_done(struct bio *bio) return false; } +static void __bio_endio(struct bio *bio) +{ + if (bio->bi_end_io) + bio->bi_end_io(bio); +} + +/* disable local irq when manipulating the percpu bio_list */ +static void unwind_bio_endio(struct bio *bio) +{ + struct bio_list *bl; + unsigned long flags; + + /* + * We can't optimize if bi_endio() is scheduled to run from + * process context because there isn't easy way to get a + * per-task bio list head or allocate a per-task variable. + */ + if (!in_interrupt()) { + /* + * It has to be a top calling when it is run from + * process context. + */ + WARN_ON(this_cpu_read(bio_end_list)); + __bio_endio(bio); + return; + } + + local_irq_save(flags); + bl = __this_cpu_read(bio_end_list); + if (bl) { + bio_list_add(bl, bio); + } else { + struct bio_list bl_in_stack; + + bl = &bl_in_stack; + bio_list_init(bl); + + __this_cpu_write(bio_end_list, bl); + while (bio) { + local_irq_restore(flags); + __bio_endio(bio); + local_irq_save(flags); + + bio = bio_list_pop(bl); + } + __this_cpu_write(bio_end_list, NULL); + } + local_irq_restore(flags); +} + /** * bio_endio - end I/O on a bio * @bio: bio @@ -1765,8 +1817,7 @@ again: goto again; } - if (bio->bi_end_io) - bio->bi_end_io(bio); + unwind_bio_endio(bio); } EXPORT_SYMBOL(bio_endio);
There were reports about heavy stack use by recursive calling .bi_end_io()([1][2][3]). For example, more than 16K stack is consumed in a single bio complete path[3], and in [2] stack overflow can be triggered if 20 nested dm-crypt is used. Also patches[1] [2] [3] were posted for addressing the issue, but never be merged. And the idea in these patches is basically similar, all serializes the recursive calling of .bi_end_io() by percpu list. This patch still takes the same idea, but uses bio_list to implement it, which turns out more simple and the code becomes more readable meantime. One corner case which wasn't covered before is that bi_endio() may be scheduled to run in process context(such as btrfs), and this patch just bypasses the optimizing for that case because one new context should have enough stack space, and this approach isn't capable of optimizing it too because there isn't easy way to get a per-task linked list head. xfstests(-g auto) is run with this patch and no regression is found on ext4, xfs and btrfs. [1] http://marc.info/?t=121428502000004&r=1&w=2 [2] http://marc.info/?l=dm-devel&m=139595190620008&w=2 [3] http://marc.info/?t=145974644100001&r=1&w=2 Cc: Shaun Tancheff <shaun.tancheff@seagate.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: Mikulas Patocka <mpatocka@redhat.com> Cc: Alan Cox <alan@linux.intel.com> Cc: Neil Brown <neilb@suse.de> Cc: Liu Bo <bo.li.liu@oracle.com> Signed-off-by: Ming Lei <ming.lei@canonical.com> --- block/bio.c | 55 +++++++++++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 53 insertions(+), 2 deletions(-)