> 2023年6月29日 08:46，Eric Wheeler <bcache@xxxxxxxxxxxxxxxxxx> 写道：
>
> On Wed, 28 Jun 2023, Eric Wheeler wrote:
>> On Tue, 27 Jun 2023, Coly Li wrote:
>>>> 2023年6月27日 20:54，Coly Li <colyli@xxxxxxx> 写道：
>>>>> 2023年6月27日 11:19，Eric Wheeler <bcache@xxxxxxxxxxxxxxxxxx> 写道：
>>>>> We have a system running a 5.15 kernel with the following errors in dmesg
>>>>> displaying repeatedly. Ultimately the system crashed so I'll reset
>>>>> it---but I was able to get some good information out of it before it died
>>>>> so maybe we can pin it down. This happened under high CPU and disk load:
>>>>>
>>>>> [Jun26 19:24] BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=0 stuck for 1382s!
>>>>> [ +0.001211] BUG: workqueue lockup - pool cpus=0 node=0 flags=0x0 nice=-20 stuck for 1381s!
>>>>> [ +0.001117] Showing busy workqueues and worker pools:
>>>>> [ +0.001081] workqueue events: flags=0x0
>>>>> [ +0.000007] pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=9/256 refcnt=10
>>>>> [ +0.000005] pending: vmstat_shepherd, kfree_rcu_monitor, drm_fb_helper_damage_work [drm_kms_helper], kfree_rcu_monitor, mlx5_timestamp_overflow [mlx5_core], mlx5_timestamp_overflow [mlx5_core], kfree_rcu_monitor, mlx5e_tx_dim_work [mlx5_core], mlx5e_rx_dim_work [mlx5_core]
>>>>> [ +0.000301] workqueue events_highpri: flags=0x10
>>>>> [ +0.000002] pwq 59: cpus=29 node=6 flags=0x0 nice=-20 active=1/256 refcnt=2
>>>>> [ +0.000004] pending: mix_interrupt_randomness
>>>>> [ +0.000006] pwq 29: cpus=14 node=7 flags=0x0 nice=-20 active=1/256 refcnt=2
>>>>> [ +0.000003] pending: mix_interrupt_randomness
>>>>> [ +0.000005] pwq 1: cpus=0 node=0 flags=0x0 nice=-20 active=1/256 refcnt=2
>>>>> [ +0.000002] pending: mix_interrupt_randomness
>>>>> [ +0.000007] workqueue events_long: flags=0x0
>>>>> [ +0.000003] pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000002] pending: br_fdb_cleanup
>>>>> [ +0.000014] workqueue events_power_efficient: flags=0x80
>>>>> [ +0.000004] pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000002] pending: fb_flashcursor
>>>>> [ +0.000008] workqueue events_freezable_power_: flags=0x84
>>>>> [ +0.000003] pwq 8: cpus=4 node=2 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000002] pending: disk_events_workfn
>>>>> [ +0.000019] workqueue mm_percpu_wq: flags=0x8
>>>>> [ +0.000004] pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=1/256 refcnt=3
>>>>> [ +0.000002] pending: lru_add_drain_per_cpu BAR(212)
>>>>> [ +0.000036] workqueue kblockd: flags=0x18
>>>>> [ +0.000005] pwq 9: cpus=4 node=2 flags=0x0 nice=-20 active=1/256 refcnt=2
>>>>> [ +0.000003] pending: blk_mq_timeout_work
>>>>> [ +0.000005] pwq 1: cpus=0 node=0 flags=0x0 nice=-20 active=3/256 refcnt=4
>>>>> [ +0.000003] pending: blk_mq_timeout_work, blk_mq_run_work_fn, blk_mq_run_work_fn
>>>>> [ +0.000044] workqueue bch_btree_io: flags=0x8
>>>>> [ +0.000001] pwq 62: cpus=31 node=7 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000004] in-flight: 3331:btree_node_write_work
>>>>> [ +0.000006] pwq 38: cpus=19 node=1 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000003] in-flight: 4829:btree_node_write_work
>>>>> [ +0.000005] pwq 0: cpus=0 node=0 flags=0x0 nice=0 active=7/256 refcnt=8
>>>>> [ +0.000002] pending: btree_node_write_work, btree_node_write_work, btree_node_write_work, btree_node_write_work, btree_node_write_work, btree_node_write_work, btree_node_write_work
>>>>> [ +0.000011] workqueue bcache: flags=0x8
>>>>> [ +0.000002] pwq 6: cpus=3 node=1 flags=0x0 nice=0 active=1/256 refcnt=2
>>>>> [ +0.000003] in-flight: 6295:bch_data_insert_keys
>>>>> [ +0.000073] workqueue bcache_writeback_wq: flags=0x8
>>>>> [ +0.000002] pwq 62: cpus=31 node=7 flags=0x0 nice=0 active=64/256 refcnt=65
>>>>> [ +0.000002] in-flight: 10178:write_dirty_finish, 10067:write_dirty_finish, 3302:write_dirty_finish, 10184:write_dirty_finish, 10181:write_dirty_finish, 10066:write_dirty_finish, 10105:write_dirty_finish, 10195:write_dirty_finish, 980:write_dirty_finish, 10141:write_dirty_finish, 10139:write_dirty_finish, 10098:write_dirty_finish, 10008:write_dirty_finish, 10180:write_dirty_finish, 20178:write_dirty_finish, 3252:write_dirty_finish, 10007:write_dirty_finish, 10279:write_dirty_finish, 10142:write_dirty_finish, 10223:write_dirty_finish, 10097:write_dirty_finish, 7311:write_dirty_finish, 10234:write_dirty_finish, 10196:write_dirty_finish, 10280:write_dirty_finish, 10063:write_dirty_finish, 10064:write_dirty_finish, 10188:write_dirty_finish, 10043:write_dirty_finish, 10101:write_dirty_finish, 10185:write_dirty_finish, 10103:write_dirty_finish, 10102:write_dirty_finish, 10224:write_dirty_finish, 10186:write_dirty_finish, 10114:write_dirty_finish, 10011:write_dirty_finish
>>>>> [ +0.000083] , 3253:write_dirty_finish, 10112:write_dirty_finish, 10187:write_dirty_finish, 10009:write_dirty_finish, 10138:write_dirty_finish, 10104:write_dirty_finish, 10140:write_dirty_finish, 10065:write_dirty_finish, 10193:write_dirty_finish, 10095:write_dirty_finish, 10041:write_dirty_finish, 10010:write_dirty_finish, 10115:write_dirty_finish, 10094:write_dirty_finish, 10113:write_dirty_finish, 10194:write_dirty_finish, 10177:write_dirty_finish, 10042:write_dirty_finish, 10226:write_dirty_finish, 10179:write_dirty_finish, 10096:write_dirty_finish, 10192:write_dirty_finish, 10222:write_dirty_finish, 10045:write_dirty_finish, 10116:write_dirty_finish, 10044:write_dirty_finish, 10225:write_dirty_finish
>>>>> [ +0.000133] workqueue kcopyd: flags=0x8
>>>>> [ +0.000004] pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=2/256 refcnt=3
>>>>> [ +0.000003] in-flight: 3424:do_work [dm_mod] do_work [dm_mod]
>>>>> [ +0.000520] pool 2: cpus=1 node=0 flags=0x0 nice=0 hung=14s workers=3 idle: 21748 21749
>>>>> [ +0.000006] pool 6: cpus=3 node=1 flags=0x0 nice=0 hung=43s workers=3 idle: 11428 25181
>>>>> [ +0.000006] pool 38: cpus=19 node=1 flags=0x0 nice=0 hung=34s workers=3 idle: 22403 25014
>>>>> [ +0.000006] pool 62: cpus=31 node=7 flags=0x0 nice=0 hung=2s workers=68 idle: 10281 23228 10286
>>>>>
>>>>> Clearly there are many write_dirty_finish() calls stuck, here are some
>>>>> traces:
>>>>>
>>>>> I did `cat /proc/<pid>/stack` for each bcache workqueue PID and these
>>>>> are the unique stacks:
>>>>>
>>>>> dmesg | grep in-flight | \
>>>>> perl -lne 'while(/(\d+):((write_dirty|bch|btree)\S+)/g) { print "$1 $2" }' | \
>>>>> sort -u | \
>>>>> while read a b; do echo === $a $b ; cat /proc/$a/stack; done
>>>>>
>>>>> Which prints lots of these:
>>>>>
>>>>> === 3253: write_dirty_finish
>>>>> [<0>] rwsem_down_write_slowpath+0x27b/0x4bd
>>>>> [<0>] bch_btree_node_get.part.0+0x7e/0x2d7 <<<, _probably_ called with write=true
>>>>> [<0>] bch_btree_map_nodes_recurse+0xed/0x1a7 | since this is an insert
>>>>> [<0>] __bch_btree_map_nodes+0x17c/0x1c4
>>>>> [<0>] bch_btree_insert+0x102/0x188 <<<<< race?
>>>>> [<0>] write_dirty_finish+0x122/0x1d3 <<<<< entry
>>>>> [<0>] process_one_work+0x1f1/0x3c6
>>>>> [<0>] worker_thread+0x53/0x3e4
>>>>> [<0>] kthread+0x127/0x144
>>>>> [<0>] ret_from_fork+0x22/0x2d
>>>>>
>>>>> and one of these:
>>>>> === 6295 bch_data_insert_keys
>>>>> [<0>] bch_btree_insert_node+0x6b/0x287
>>>>> [<0>] btree_insert_fn+0x20/0x48
>>>>> [<0>] bch_btree_map_nodes_recurse+0x111/0x1a7
>>>>> [<0>] __bch_btree_map_nodes+0x17c/0x1c4
>>>>> [<0>] bch_btree_insert+0x102/0x188 <<<<< race?
>>>>> [<0>] bch_data_insert_keys+0x30/0xba <<<<< entry
>>>>> [<0>] process_one_work+0x1f1/0x3c6
>>>>> [<0>] worker_thread+0x53/0x3e4
>>>>> [<0>] kthread+0x127/0x144
>>>>> [<0>] ret_from_fork+0x22/0x2d
>>>>>
>>>>> Note that above, both threads (workqueues) are similar until they call
>>>>> bch_btree_map_nodes_recurse(), then they diverge where one is doing
>>>>> bch_btree_insert_node(), which holds b->write_lock:
>>>>>
>>>>> bch_btree_insert_node
>>>>> https://elixir.bootlin.com/linux/latest/source/drivers/md/bcache/btree.c#L2322
>>>>>
>>>>> and the other is trying bch_btree_node_get(). While I don't have debug
>>>>> data about the arguments, I am guessing that bch_btree_node_get is
>>>>> called with `write=true` since the caller is bch_btree_insert:
>>>>>
>>>>> /* bch_btree_node_get - find a btree node in the cache and lock it, reading it
>>>>> * in from disk if necessary. */
>>>>> https://elixir.bootlin.com/linux/v6.4/source/drivers/md/bcache/btree.c#L969
>>>>>
>>>>> The call to bch_btree_node_get() does quite a bit of rw_lock/rw_unlock/mutex work.
>>>>>
>>>>> There are also two of the traces below which are waiting on a down():
>>>>> https://elixir.bootlin.com/linux/latest/source/drivers/md/bcache/btree.c#L420
>>>>>
>>>>> These could be relevant since __bch_btree_node_write() does call
>>>>> `lockdep_assert_held(&b->write_lock)` and b->write_lock is held above by
>>>>> bch_btree_insert_node:
>>>>>
>>>>> === 3331 btree_node_write_work
>>>>> [<0>] down+0x43/0x54
>>>>> [<0>] __bch_btree_node_write+0xa3/0x220
>>>>> [<0>] btree_node_write_work+0x43/0x4f
>>>>> [<0>] process_one_work+0x1f1/0x3c6
>>>>> [<0>] worker_thread+0x53/0x3e4
>>>>> [<0>] kthread+0x127/0x144
>>>>> [<0>] ret_from_fork+0x22/0x2d
>>>>>
>>>>> === 4829 btree_node_write_work
>>>>> [<0>] down+0x43/0x54
>>>>> [<0>] __bch_btree_node_write+0xa3/0x220
>>>>> [<0>] btree_node_write_work+0x43/0x4f
>>>>> [<0>] process_one_work+0x1f1/0x3c6
>>>>> [<0>] worker_thread+0x53/0x3e4
>>>>> [<0>] kthread+0x127/0x144
>>>>> [<0>] ret_from_fork+0x22/0x2d
>>>>>
>>>>> Thanks for your help!
>>>>
>>>> When does this lockup happen? Is it in initialization or bootup time ?
>>
>> After its been running for a few days and always under heavy CPU and disk
>> IO.
>>
>> I'll check the patches below and see if we are missing any.
>
>
> These were missing, so I'm cherry-picking and rebuilding:
>
> bcache: fixup btree_cache_wait list damage
> bcache: remove unnecessary flush_workqueue
> bcache: avoid unnecessary soft lockup in kworker update_writeback_rate()
> bcache: fixup bcache_dev_sectors_dirty_add() multithreaded CPU false sharing
> bcache: fix NULL pointer reference in cached_dev_detach_finish
> bcache: move calc_cached_dev_sectors to proper place on backing device detach
>
> This one fails to build, unsafe_memcpy is missing, but I don't think we
> need it:
> bcache: Silence memcpy() run-time false positive warnings
>
Yes the above patch can be ignored.
Coly Li
