pá 14. 1. 2022 v 15:23 odesílatel Xin Long <lucien.xin@xxxxxxxxx> napsal:
>
> cpus_read_lock() is introduced into kmem_cache_destroy() by
> commit 5a836bf6b09f ("mm: slub: move flush_cpu_slab() invocations
> __free_slab() invocations out of IRQ context"), and it could cause
> a deadlock.
FYI,
I received a bug report from one of our customers, he complains that
his system (with nvmefc boot from SAN) hangs when rebooting.
He runs a RHEL-9 kernel based on version 5.14.0.
What is interesting is that, according to him, after reverting commit
5a836bf6b09f
("mm: slub: move flush_cpu_slab() invocations __free_slab()
invocations out of IRQ context")
the reboot operation doesn't hang anymore.
The call trace seems to point to a possible problem due to the fact that
nvme_delete_ctrl_work is allocated with the WQ_MEM_RECLAIM bit set.
[ 453.012078] ------------[ cut here ]------------
[ 453.016744] workqueue: WQ_MEM_RECLAIM
nvme-delete-wq:nvme_delete_ctrl_work [nvme_core] is flushing
!WQ_MEM_RECLAIM events:flush_cpu_slab
[ 453.016789] WARNING: CPU: 37 PID: 410 at kernel/workqueue.c:2637
check_flush_dependency+0x10a/0x120
[...]
[ 453.262125] Call Trace:
[ 453.264582] __flush_work.isra.0+0xbf/0x220
[ 453.268775] ? __queue_work+0x1dc/0x420
[ 453.272623] flush_all_cpus_locked+0xfb/0x120
[ 453.276992] __kmem_cache_shutdown+0x2b/0x320
[ 453.281361] kmem_cache_destroy+0x49/0x100
[ 453.285465] bioset_exit+0x143/0x190
[ 453.289052] blk_release_queue+0xb9/0x100
[ 453.293075] kobject_cleanup+0x37/0x130
[ 453.296922] nvme_fc_ctrl_free+0xc6/0x150 [nvme_fc]
[ 453.302397] nvme_free_ctrl+0x1ac/0x2b0 [nvme_core]
[ 453.307818] device_release+0x31/0x90
[ 453.312005] kobject_cleanup+0x37/0x130
[ 453.316369] process_one_work+0x1e5/0x3c0
[ 453.320895] worker_thread+0x50/0x3b0
[ 453.325074] ? rescuer_thread+0x370/0x370
[ 453.329592] kthread+0x146/0x170
[ 453.333322] ? set_kthread_struct+0x40/0x40
[ 453.338027] ret_from_fork+0x1f/0x30
[ 453.342082] ---[ end trace 8c9cdd85adbbfc4f ]---
Maurizio Lombardi
