lkml.org 
[lkml]   [2021]   [Dec]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] ksmbd: use F_SETLK to force vfs_file_lock() to return asynchronously
From
Date
On 22.12.2021 18:17, Vasily Averin wrote:
> On 22.12.2021 11:58, Namjae Jeon wrote:
>> 2021-12-22 15:51 GMT+09:00, Vasily Averin <vvs@virtuozzo.com>:
>>> On 22.12.2021 08:25, Namjae Jeon wrote:
>>>> 2021-12-22 13:32 GMT+09:00, Vasily Averin <vvs@virtuozzo.com>:
>>>>> On 22.12.2021 05:50, Namjae Jeon wrote:
>>>>>> 2021-12-21 22:08 GMT+09:00, Vasily Averin <vvs@virtuozzo.com>:
>>>>>>> On 21.12.2021 15:02, Namjae Jeon wrote:
>>>>>>>> 2021-12-19 18:34 GMT+09:00, Vasily Averin <vvs@virtuozzo.com>:
>>>>>>>>> To avoid possible deadlock ksmbd should process locks
>>>>>>>>> asynchronously.
>>>>>>>>> Callers expecting vfs_file_locks() to return asynchronously should
>>>>>>>>> only
>>>>>>>>> use F_SETLK, not F_SETLKW.
>>>>>>>> Should I check this patch instead of
>>>>>>>> [PATCH] ksmbd: force "fail immediately" flag on fs with its own
>>>>>>>> ->lock
>>>>>>>> ?
>>>>>>>
>>>>>>> no, these patches are independent and both ones are required.
>>>>>>> current patch fixes incorrect kernel thread behaviour:
>>>>>>> kernel threads should not use F_SETLKW for locking requests.
>>>>>> How does this patch work? posix_lock_file in vfs_lock_file() does not
>>>>>> use
>>>>>> cmd.
>>>>>> And your patch still leaves FL_SLEEP.
>>>>>
>>>>> "use F_SETLK, not F_SETLKW" was copy-pasted from requirement described
>>>>> in
>>>>> comment above vfs_lock_file().
>>>>>
>>>>> posix_lock_file() is not used in all ->lock() functions, and use
>>>>> F_SETLKW
>>>>> forces some of affected filesystem use blocking locks:
>>>> What I'm saying is that when we apply "ksmbd: force "fail immediately"
>>>> flag on fs with its own ->lock ", this patch is meaningless. How is
>>>> ->lock() with F_SETLKW called?
>>>
>>> I got your point finally,
>>> yes, you are right, now this cannot happen.
>>> However I'm going to fix all affected filesystems and then revert
>>> "ksmbd: force "fail immediately" flag on fs with its own ->lock"
>>> When this happen and ksmbd will still use IS_SETLKW it will trigger the
>>> problems described below.
>> If so, You can include one patch(this patch + revert patch) in patch
>> series for fixing ->lock of all filesystem.

I've checked how smb2_lock() handles FILE_LOCK_DEFERRED returned by vfs_lock_file() call.
It seems for me, working thread will be blocked in ksmbd_vfs_posix_lock_wait(),
so whole ksmbd server still can deadlock. Am I wrong probably?

Thank you,
Vasily Averin

\
 
 \ /
  Last update: 2021-12-24 13:31    [W:0.147 / U:1.912 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site