Messages in this thread | | | Subject | Re: [PATCH 2/2] mm/slub: Fix slab_mutex circular locking problem in slab_attr_store() | From | Waiman Long <> | Date | Mon, 27 Apr 2020 15:00:05 -0400 |
| |
On 4/24/20 3:44 PM, Qian Cai wrote: > >> On Apr 24, 2020, at 11:12 AM, Waiman Long <longman@redhat.com> wrote: >> >> The following lockdep splat was reported: >> >> [ 176.241923] ====================================================== >> [ 176.241924] WARNING: possible circular locking dependency detected >> [ 176.241926] 4.18.0-172.rt13.29.el8.x86_64+debug #1 Not tainted >> [ 176.241927] ------------------------------------------------------ >> [ 176.241929] slub_cpu_partia/5371 is trying to acquire lock: >> [ 176.241930] ffffffffa0b83718 (slab_mutex){+.+.}, at: slab_attr_store+0x6b/0xe0 >> [ 176.241941] >> but task is already holding lock: >> [ 176.241942] ffff88bb6d8b83c8 (kn->count#103){++++}, at: kernfs_fop_write+0x1cc/0x400 >> [ 176.241947] >> which lock already depends on the new lock. >> >> [ 176.241949] >> the existing dependency chain (in reverse order) is: >> [ 176.241949] >> -> #1 (kn->count#103){++++}: >> [ 176.241955] __kernfs_remove+0x616/0x800 >> [ 176.241957] kernfs_remove_by_name_ns+0x3e/0x80 >> [ 176.241959] sysfs_slab_add+0x1c6/0x330 >> [ 176.241961] __kmem_cache_create+0x15f/0x1b0 >> [ 176.241964] create_cache+0xe1/0x220 >> [ 176.241966] kmem_cache_create_usercopy+0x1a3/0x260 >> [ 176.241967] kmem_cache_create+0x12/0x20 >> [ 176.242076] mlx5_init_fs+0x18d/0x1a00 [mlx5_core] >> [ 176.242100] mlx5_load_one+0x3b4/0x1730 [mlx5_core] >> [ 176.242124] init_one+0x901/0x11b0 [mlx5_core] >> [ 176.242127] local_pci_probe+0xd4/0x180 >> [ 176.242131] work_for_cpu_fn+0x51/0xa0 >> [ 176.242133] process_one_work+0x91a/0x1ac0 >> [ 176.242134] worker_thread+0x536/0xb40 >> [ 176.242136] kthread+0x30c/0x3d0 >> [ 176.242140] ret_from_fork+0x27/0x50 >> [ 176.242140] >> -> #0 (slab_mutex){+.+.}: >> [ 176.242145] __lock_acquire+0x22cb/0x48c0 >> [ 176.242146] lock_acquire+0x134/0x4c0 >> [ 176.242148] _mutex_lock+0x28/0x40 >> [ 176.242150] slab_attr_store+0x6b/0xe0 >> [ 176.242151] kernfs_fop_write+0x251/0x400 >> [ 176.242154] vfs_write+0x157/0x460 >> [ 176.242155] ksys_write+0xb8/0x170 >> [ 176.242158] do_syscall_64+0x13c/0x710 >> [ 176.242160] entry_SYSCALL_64_after_hwframe+0x6a/0xdf >> [ 176.242161] >> other info that might help us debug this: >> >> [ 176.242161] Possible unsafe locking scenario: >> >> [ 176.242162] CPU0 CPU1 >> [ 176.242163] ---- ---- >> [ 176.242163] lock(kn->count#103); >> [ 176.242165] lock(slab_mutex); >> [ 176.242166] lock(kn->count#103); >> [ 176.242167] lock(slab_mutex); >> [ 176.242169] >> *** DEADLOCK *** >> >> [ 176.242170] 3 locks held by slub_cpu_partia/5371: >> [ 176.242170] #0: ffff888705e3a800 (sb_writers#4){.+.+}, at: vfs_write+0x31c/0x460 >> [ 176.242174] #1: ffff889aeec4d658 (&of->mutex){+.+.}, at: kernfs_fop_write+0x1a9/0x400 >> [ 176.242177] #2: ffff88bb6d8b83c8 (kn->count#103){++++}, at: kernfs_fop_write+0x1cc/0x400 >> [ 176.242180] >> stack backtrace: >> [ 176.242183] CPU: 36 PID: 5371 Comm: slub_cpu_partia Not tainted 4.18.0-172.rt13.29.el8.x86_64+debug #1 >> [ 176.242184] Hardware name: AMD Corporation DAYTONA_X/DAYTONA_X, BIOS RDY1005C 11/22/2019 >> [ 176.242185] Call Trace: >> [ 176.242190] dump_stack+0x9a/0xf0 >> [ 176.242193] check_noncircular+0x317/0x3c0 >> [ 176.242195] ? print_circular_bug+0x1e0/0x1e0 >> [ 176.242199] ? native_sched_clock+0x32/0x1e0 >> [ 176.242202] ? sched_clock+0x5/0x10 >> [ 176.242205] ? sched_clock_cpu+0x238/0x340 >> [ 176.242208] __lock_acquire+0x22cb/0x48c0 >> [ 176.242213] ? trace_hardirqs_on+0x10/0x10 >> [ 176.242215] ? trace_hardirqs_on+0x10/0x10 >> [ 176.242218] lock_acquire+0x134/0x4c0 >> [ 176.242220] ? slab_attr_store+0x6b/0xe0 >> [ 176.242223] _mutex_lock+0x28/0x40 >> [ 176.242225] ? slab_attr_store+0x6b/0xe0 >> [ 176.242227] slab_attr_store+0x6b/0xe0 >> [ 176.242229] ? sysfs_file_ops+0x160/0x160 >> [ 176.242230] kernfs_fop_write+0x251/0x400 >> [ 176.242232] ? __sb_start_write+0x26a/0x3f0 >> [ 176.242234] vfs_write+0x157/0x460 >> [ 176.242237] ksys_write+0xb8/0x170 >> [ 176.242239] ? __ia32_sys_read+0xb0/0xb0 >> [ 176.242242] ? do_syscall_64+0xb9/0x710 >> [ 176.242245] do_syscall_64+0x13c/0x710 >> [ 176.242247] entry_SYSCALL_64_after_hwframe+0x6a/0xdf >> >> There was another lockdep splat generated by echoing "1" to >> "/sys/kernel/slab/fs_cache/shrink": >> >> [ 445.231443] Chain exists of: >> cpu_hotplug_lock --> mem_hotplug_lock --> slab_mutex >> >> [ 445.242025] Possible unsafe locking scenario: >> >> [ 445.247977] CPU0 CPU1 >> [ 445.252529] ---- ---- >> [ 445.257082] lock(slab_mutex); >> [ 445.260239] lock(mem_hotplug_lock); >> [ 445.266452] lock(slab_mutex); >> [ 445.272141] lock(cpu_hotplug_lock); >> >> So it is problematic to use slab_mutex to iterate the list of >> child memcgs with for_each_memcg_cache(). Fortunately, there is >> another way to do child memcg iteration by going through the array >> entries in memcg_params.memcg_caches while holding a read lock on >> memcg_cache_ids_sem. >> >> To avoid other possible circular locking problems, we only take a >> reference to the child memcgs and store their addresses while holding >> memcg_cache_ids_sem. The actual store method is called for each of the >> child memcgs after releasing the lock. > Even on x86 where it compiles, this patch is insufficient to prevent a lockdep splat, > because there are still cpu_hotplug_lock and mem_hotplug_lock in the way. > Yes, I was aware of that. Will include additional patch in v2 to address that.
Thanks, Longman
| |