btrfs/058 deadlock with lseek

btrfs/058 deadlock with lseek

[Qu Wenruo]

[-- Attachment #1.1: Type: text/plain, Size: 1502 bytes --]

Hi guys,

Just got a reproducible error in btrfs/058.
The backtrace is completely in VFS territory, not btrfs related lock at all:

BTRFS info (device dm-5): checking UUID tree
sysrq: Show Blocked State
  task                        PC stack   pid father
rm              D    0 560678 560445 0x00000000
Call Trace:
 __schedule+0x5c7/0xea0
 ? __sched_text_start+0x8/0x8
 ? lock_downgrade+0x380/0x380
 ? lock_contended+0x730/0x730
 ? debug_check_no_locks_held+0x60/0x60
 schedule+0x7b/0x170
 schedule_preempt_disabled+0x18/0x30
 __mutex_lock+0x481/0xc70
 ? __fdget_pos+0x7e/0x80
 ? mutex_trylock+0x190/0x190
 ? debug_lockdep_rcu_enabled+0x26/0x40
 ? kmem_cache_free+0x157/0x3b0
 ? putname+0x73/0x80
 ? __ia32_sys_rmdir+0x30/0x30
 ? __check_object_size+0x134/0x1e6
 mutex_lock_nested+0x1b/0x20
 ? mutex_lock_nested+0x1b/0x20
 __fdget_pos+0x7e/0x80
 ksys_lseek+0x1d/0xf0
 __x64_sys_lseek+0x43/0x50
 do_syscall_64+0x79/0xe0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
RIP: 0033:0x7f7518e5652b
Code: Bad RIP value.
RSP: 002b:00007ffead7508e8 EFLAGS: 00000246 ORIG_RAX: 0000000000000008
RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f7518e5652b
RDX: 0000000000000001 RSI: 0000000000000000 RDI: 0000000000000000
RBP: 00007f7518f267e0 R08: 0000000000000001 R09: 0000000000000001
R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
R13: 0000000000000002 R14: 00007f7518f2be68 R15: 00007f7518f287e0

Is this a known bug in VFS layer?

Thanks,
Qu


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: btrfs/058 deadlock with lseek

[Qu Wenruo]

[-- Attachment #1.1: Type: text/plain, Size: 2054 bytes --]



On 2019/11/26 下午4:17, Qu Wenruo wrote:
> Hi guys,
> 
> Just got a reproducible error in btrfs/058.
> The backtrace is completely in VFS territory, not btrfs related lock at all:

With the help of Nikolay and Johannes, the offending commit is pinned
down to 0be0ee71816b ("vfs: properly and reliably lock f_pos in
fdget_pos()"), and Linus will soon revert it.

Not a big deal, but testers would have a much easier life using David's
misc-5.5 (still based on v5.4-rc).

And to David, would you please keep your misc-5.5 branch until the
offending patch get reverted?

Thanks,
Qu

> BTRFS info (device dm-5): checking UUID tree
> sysrq: Show Blocked State
>   task                        PC stack   pid father
> rm              D    0 560678 560445 0x00000000
> Call Trace:
>  __schedule+0x5c7/0xea0
>  ? __sched_text_start+0x8/0x8
>  ? lock_downgrade+0x380/0x380
>  ? lock_contended+0x730/0x730
>  ? debug_check_no_locks_held+0x60/0x60
>  schedule+0x7b/0x170
>  schedule_preempt_disabled+0x18/0x30
>  __mutex_lock+0x481/0xc70
>  ? __fdget_pos+0x7e/0x80
>  ? mutex_trylock+0x190/0x190
>  ? debug_lockdep_rcu_enabled+0x26/0x40
>  ? kmem_cache_free+0x157/0x3b0
>  ? putname+0x73/0x80
>  ? __ia32_sys_rmdir+0x30/0x30
>  ? __check_object_size+0x134/0x1e6
>  mutex_lock_nested+0x1b/0x20
>  ? mutex_lock_nested+0x1b/0x20
>  __fdget_pos+0x7e/0x80
>  ksys_lseek+0x1d/0xf0
>  __x64_sys_lseek+0x43/0x50
>  do_syscall_64+0x79/0xe0
>  entry_SYSCALL_64_after_hwframe+0x49/0xbe
> RIP: 0033:0x7f7518e5652b
> Code: Bad RIP value.
> RSP: 002b:00007ffead7508e8 EFLAGS: 00000246 ORIG_RAX: 0000000000000008
> RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f7518e5652b
> RDX: 0000000000000001 RSI: 0000000000000000 RDI: 0000000000000000
> RBP: 00007f7518f267e0 R08: 0000000000000001 R09: 0000000000000001
> R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
> R13: 0000000000000002 R14: 00007f7518f2be68 R15: 00007f7518f287e0
> 
> Is this a known bug in VFS layer?
> 
> Thanks,
> Qu
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: btrfs/058 deadlock with lseek

[David Sterba]

On Tue, Nov 26, 2019 at 05:56:27PM +0800, Qu Wenruo wrote:
> On 2019/11/26 下午4:17, Qu Wenruo wrote:
> > Just got a reproducible error in btrfs/058.
> > The backtrace is completely in VFS territory, not btrfs related lock at all:
> 
> With the help of Nikolay and Johannes, the offending commit is pinned
> down to 0be0ee71816b ("vfs: properly and reliably lock f_pos in
> fdget_pos()"), and Linus will soon revert it.
> 
> Not a big deal, but testers would have a much easier life using David's
> misc-5.5 (still based on v5.4-rc).
> 
> And to David, would you please keep your misc-5.5 branch until the
> offending patch get reverted?

misc-5.5 will not rebase, it's the branch that gets pulled to master.
I won't rebase misc-next until rc1 at least and will rebase only after I
see that tests pass. During merge window lots of things are in flux so
breakage is expected so we need a development base for testing.