lkml.org 
[lkml]   [2021]   [Jan]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] locking/lockdep: Use local_irq_save() with call_rcu()
On Tue, Dec 22, 2020 at 05:55:53PM -0500, Waiman Long wrote:
> The following lockdep splat was hit:
>
> [ 560.638354] WARNING: CPU: 79 PID: 27458 at kernel/rcu/tree_plugin.h:1749 call_rcu+0x6dc/0xf00
> :
> [ 560.647761] RIP: 0010:call_rcu+0x6dc/0xf00
> [ 560.647763] Code: 0f 8f 29 04 00 00 e8 93 da 1c 00 48 8b 3c 24 57 9d 0f 1f 44 00 00 e9 19 fa ff ff 65 8b 05 38 83 c4 49 85 c0 0f 84 cd fb ff ff <0f> 0b e9 c6 fb ff ff e8 b8 45 51 00 4c 89 f2 48 b8 00 00 00 00 00
> [ 560.647764] RSP: 0018:ff11001050097b58 EFLAGS: 00010002
> [ 560.647766] RAX: 0000000000000001 RBX: ffffffffbb1f3360 RCX: 0000000000000001
> [ 560.647766] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffffffffb99bac9c
> [ 560.647767] RBP: 1fe220020a012f73 R08: 000000010004005c R09: dffffc0000000000
> [ 560.647768] R10: dffffc0000000000 R11: 0000000000000003 R12: ff1100105b7f70e1
> [ 560.647769] R13: ffffffffb635d8a0 R14: ff1100105b7f72d8 R15: ff1100105b7f7040
> [ 560.647770] FS: 00007fd9b3437080(0000) GS:ff1100105b600000(0000) knlGS:0000000000000000
> [ 560.647771] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [ 560.647772] CR2: 00007fd9b30112bc CR3: 000000105e898006 CR4: 0000000000761ee0
> [ 560.647773] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [ 560.647773] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [ 560.647774] PKRU: 55555554
> [ 560.647774] Call Trace:
> [ 560.647778] ? invoke_rcu_core+0x180/0x180
> [ 560.647782] ? __is_module_percpu_address+0xed/0x440
> [ 560.647787] lockdep_unregister_key+0x2ab/0x5b0
> [ 560.647791] destroy_workqueue+0x40b/0x610
> [ 560.647862] xlog_dealloc_log+0x216/0x2b0 [xfs]
> :
>
> This splat is caused by the fact that lockdep_unregister_key() uses
> raw_local_irq_save() which doesn't update the hardirqs_enabled
> percpu flag. The call_rcu() function, however, will call
> lockdep_assert_irqs_disabled() to check the hardirqs_enabled flag which
> remained set in this case.
>
> Fix this problem by using local_irq_save()/local_irq_restore() pairs
> whenever call_rcu() is being called.

I'm not sure I much like all this,.. :/

> I think raw_local_irq_save() function can be used if no external
> function is being called except maybe printk() as it means another
> lockdep problem exists.

The reason lockdep is using raw_local_irq_save() is to avoid calling
into itself again, notably local_irq_restore() will end up in
mark_held_locks().

> Fixes: a0b0fd53e1e67 ("locking/lockdep: Free lock classes that are no longer in use")

Seems dubious, as the lockdep_assert_irqs_disabled() that triggered was
added after that patch.

I'm thinking another solution would be to increment the lockdep
recursion count before calling RCU, because then we'll fail
__lockdep_enabled and the assertion gets killed. Hmm?

\
 
 \ /
  Last update: 2021-01-04 16:41    [W:0.089 / U:0.192 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site