| Message ID | 754d1973-31cb-d3ca-1f6f-2d35b96364db@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v3] dm-io: don't warn if flush takes too long time | expand |
On 4/17/24 3:05 AM, Mikulas Patocka wrote: > There was reported hang warning when using dm-integrity on the top of loop > device on XFS on a rotational disk. The warning was triggered because > flush on the loop device was too slow. > > There's no easy way to reduce the latency, so I made a patch that shuts > the warning up. > > There's already a function blk_wait_io that avoids the hung task warning. > This commit moves this function from block/blk.h to > include/linux/completion.h, renames it to wait_for_completion_long_io > (because it is not dependent on the block layer at all) and uses it in > dm-io instead of wait_for_completion_io. Change looks fine to me, but while at it, let's just move it into blk-core.c and make it public, no need for this function to be a static inline.
> --- linux-2.6.orig/include/linux/completion.h 2024-04-15 15:54:22.000000000 +0200 > +++ linux-2.6/include/linux/completion.h 2024-04-15 15:57:14.000000000 +0200 > @@ -10,6 +10,7 @@ > */ > > #include <linux/swait.h> > +#include <linux/sched/sysctl.h> If you're touching completion.h you need to CC lkml and the people who wrote/maintain it even if we don't have a proper maintainer. I don't think adding yet another include into it is a good idea. As is this whole hack here. Pleas just add the proper TASK_STATE for task that can legitimately sleep for very long times instead of extending this hack again and again, just like I told Kent when he messed with the timeout. > > /* > * struct completion - structure used to maintain state for a "completion" > @@ -119,4 +120,20 @@ extern void complete(struct completion * > extern void complete_on_current_cpu(struct completion *x); > extern void complete_all(struct completion *); > > +/** > + * wait_for_completion_long_io - this is like wait_for_completion_io, > + * but it doesn't warn if the wait takes too long. > + */ > +static inline void wait_for_completion_long_io(struct completion *done) > +{ > + /* Prevent hang_check timer from firing at us during very long I/O */ > + unsigned long timeout = sysctl_hung_task_timeout_secs * HZ / 2; > + > + if (timeout) > + while (!wait_for_completion_io_timeout(done, timeout)) > + ; > + else > + wait_for_completion_io(done); > +} > + > #endif > Index: linux-2.6/block/bio.c > =================================================================== > --- linux-2.6.orig/block/bio.c 2024-03-30 20:07:03.000000000 +0100 > +++ linux-2.6/block/bio.c 2024-04-15 15:55:13.000000000 +0200 > @@ -1378,7 +1378,7 @@ int submit_bio_wait(struct bio *bio) > bio->bi_end_io = submit_bio_wait_endio; > bio->bi_opf |= REQ_SYNC; > submit_bio(bio); > - blk_wait_io(&done); > + wait_for_completion_long_io(&done); > > return blk_status_to_errno(bio->bi_status); > } > Index: linux-2.6/block/blk-mq.c > =================================================================== > --- linux-2.6.orig/block/blk-mq.c 2024-03-30 20:07:03.000000000 +0100 > +++ linux-2.6/block/blk-mq.c 2024-04-15 15:55:05.000000000 +0200 > @@ -1407,7 +1407,7 @@ blk_status_t blk_execute_rq(struct reque > if (blk_rq_is_poll(rq)) > blk_rq_poll_completion(rq, &wait.done); > else > - blk_wait_io(&wait.done); > + wait_for_completion_long_io(&wait.done); > > return wait.ret; > } > > ---end quoted text---
On Wed, 17 Apr 2024, Jens Axboe wrote: > On 4/17/24 3:05 AM, Mikulas Patocka wrote: > > There was reported hang warning when using dm-integrity on the top of loop > > device on XFS on a rotational disk. The warning was triggered because > > flush on the loop device was too slow. > > > > There's no easy way to reduce the latency, so I made a patch that shuts > > the warning up. > > > > There's already a function blk_wait_io that avoids the hung task warning. > > This commit moves this function from block/blk.h to > > include/linux/completion.h, renames it to wait_for_completion_long_io > > (because it is not dependent on the block layer at all) and uses it in > > dm-io instead of wait_for_completion_io. > > Change looks fine to me, but while at it, let's just move it into > blk-core.c and make it public, no need for this function to be a static > inline. > > -- > Jens Axboe I think we should move it to ./kernel/sched/completion.c. Because the function has no dependency on the block layer. I'll send a patch that does it. Mikulas
Index: linux-2.6/block/blk.h =================================================================== --- linux-2.6.orig/block/blk.h 2024-04-15 15:54:22.000000000 +0200 +++ linux-2.6/block/blk.h 2024-04-15 15:54:21.000000000 +0200 @@ -72,18 +72,6 @@ static inline int bio_queue_enter(struct return __bio_queue_enter(q, bio); } -static inline void blk_wait_io(struct completion *done) -{ - /* Prevent hang_check timer from firing at us during very long I/O */ - unsigned long timeout = sysctl_hung_task_timeout_secs * HZ / 2; - - if (timeout) - while (!wait_for_completion_io_timeout(done, timeout)) - ; - else - wait_for_completion_io(done); -} - #define BIO_INLINE_VECS 4 struct bio_vec *bvec_alloc(mempool_t *pool, unsigned short *nr_vecs, gfp_t gfp_mask); Index: linux-2.6/drivers/md/dm-io.c =================================================================== --- linux-2.6.orig/drivers/md/dm-io.c 2024-04-15 15:54:22.000000000 +0200 +++ linux-2.6/drivers/md/dm-io.c 2024-04-15 15:54:21.000000000 +0200 @@ -450,7 +450,7 @@ static int sync_io(struct dm_io_client * dispatch_io(opf, num_regions, where, dp, io, 1, ioprio); - wait_for_completion_io(&sio.wait); + wait_for_completion_long_io(&sio.wait); if (error_bits) *error_bits = sio.error_bits; Index: linux-2.6/include/linux/completion.h =================================================================== --- linux-2.6.orig/include/linux/completion.h 2024-04-15 15:54:22.000000000 +0200 +++ linux-2.6/include/linux/completion.h 2024-04-15 15:57:14.000000000 +0200 @@ -10,6 +10,7 @@ */ #include <linux/swait.h> +#include <linux/sched/sysctl.h> /* * struct completion - structure used to maintain state for a "completion" @@ -119,4 +120,20 @@ extern void complete(struct completion * extern void complete_on_current_cpu(struct completion *x); extern void complete_all(struct completion *); +/** + * wait_for_completion_long_io - this is like wait_for_completion_io, + * but it doesn't warn if the wait takes too long. + */ +static inline void wait_for_completion_long_io(struct completion *done) +{ + /* Prevent hang_check timer from firing at us during very long I/O */ + unsigned long timeout = sysctl_hung_task_timeout_secs * HZ / 2; + + if (timeout) + while (!wait_for_completion_io_timeout(done, timeout)) + ; + else + wait_for_completion_io(done); +} + #endif Index: linux-2.6/block/bio.c =================================================================== --- linux-2.6.orig/block/bio.c 2024-03-30 20:07:03.000000000 +0100 +++ linux-2.6/block/bio.c 2024-04-15 15:55:13.000000000 +0200 @@ -1378,7 +1378,7 @@ int submit_bio_wait(struct bio *bio) bio->bi_end_io = submit_bio_wait_endio; bio->bi_opf |= REQ_SYNC; submit_bio(bio); - blk_wait_io(&done); + wait_for_completion_long_io(&done); return blk_status_to_errno(bio->bi_status); } Index: linux-2.6/block/blk-mq.c =================================================================== --- linux-2.6.orig/block/blk-mq.c 2024-03-30 20:07:03.000000000 +0100 +++ linux-2.6/block/blk-mq.c 2024-04-15 15:55:05.000000000 +0200 @@ -1407,7 +1407,7 @@ blk_status_t blk_execute_rq(struct reque if (blk_rq_is_poll(rq)) blk_rq_poll_completion(rq, &wait.done); else - blk_wait_io(&wait.done); + wait_for_completion_long_io(&wait.done); return wait.ret; }
There was reported hang warning when using dm-integrity on the top of loop device on XFS on a rotational disk. The warning was triggered because flush on the loop device was too slow. There's no easy way to reduce the latency, so I made a patch that shuts the warning up. There's already a function blk_wait_io that avoids the hung task warning. This commit moves this function from block/blk.h to include/linux/completion.h, renames it to wait_for_completion_long_io (because it is not dependent on the block layer at all) and uses it in dm-io instead of wait_for_completion_io. [ 1352.586981] INFO: task kworker/1:2:14820 blocked for more than 120 seconds. [ 1352.593951] Not tainted 4.18.0-552.el8_10.x86_64 #1 [ 1352.599358] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. [ 1352.607202] Call Trace: [ 1352.609670] __schedule+0x2d1/0x870 [ 1352.613173] ? update_load_avg+0x7e/0x710 [ 1352.617193] ? update_load_avg+0x7e/0x710 [ 1352.621214] schedule+0x55/0xf0 [ 1352.624371] schedule_timeout+0x281/0x320 [ 1352.628393] ? __schedule+0x2d9/0x870 [ 1352.632065] io_schedule_timeout+0x19/0x40 [ 1352.636176] wait_for_completion_io+0x96/0x100 [ 1352.640639] sync_io+0xcc/0x120 [dm_mod] [ 1352.644592] dm_io+0x209/0x230 [dm_mod] [ 1352.648436] ? bit_wait_timeout+0xa0/0xa0 [ 1352.652461] ? vm_next_page+0x20/0x20 [dm_mod] [ 1352.656924] ? km_get_page+0x60/0x60 [dm_mod] [ 1352.661298] dm_bufio_issue_flush+0xa0/0xd0 [dm_bufio] [ 1352.666448] dm_bufio_write_dirty_buffers+0x1a0/0x1e0 [dm_bufio] [ 1352.672462] dm_integrity_flush_buffers+0x32/0x140 [dm_integrity] [ 1352.678567] ? lock_timer_base+0x67/0x90 [ 1352.682505] ? __timer_delete.part.36+0x5c/0x90 [ 1352.687050] integrity_commit+0x31a/0x330 [dm_integrity] [ 1352.692368] ? __switch_to+0x10c/0x430 [ 1352.696131] process_one_work+0x1d3/0x390 [ 1352.700152] ? process_one_work+0x390/0x390 [ 1352.704348] worker_thread+0x30/0x390 [ 1352.708019] ? process_one_work+0x390/0x390 [ 1352.712214] kthread+0x134/0x150 [ 1352.715459] ? set_kthread_struct+0x50/0x50 [ 1352.719659] ret_from_fork+0x1f/0x40 Signed-off-by: Mikulas Patocka <mpatocka@redhat.com> --- block/bio.c | 2 +- block/blk-mq.c | 2 +- block/blk.h | 12 ------------ drivers/md/dm-io.c | 2 +- include/linux/completion.h | 17 +++++++++++++++++ 5 files changed, 20 insertions(+), 15 deletions(-)