Ocfs2 devel - Jul 2022 - [PATCH 2/3] ocfs2: Remove a useless spinlock

Le 20/07/2022 ? 03:59, Joseph Qi a ?crit?:
>
> On 7/19/22 9:25 PM, Christophe JAILLET wrote:
>> Le 19/07/2022 ? 12:24, David Laight a ?crit?:
>>> From: Christophe JAILLET
>>>> Sent: 19 July 2022 11:02
>>>>
>>>> 'node_map_lock' is a spinlock only used to protect
calls to set_bit(),
>>>> clear_bit() and test_bit().
>>>>
>>>> {set|clear}_bit() are already atomic and don't need this
extra spinlock.
>>>> test_bit() only reads the bitmap for a given bit.
>>>>
>>>> Remove this useless spinlock.
>>> It looks to me like the calling code is racy
>>> unless there is another lock in the callers.
>> The call chains are:
>>  ? ocfs2_recover_orphans()
>>  ??? ocfs2_mark_recovering_orphan_dir()
>>  ????? spin_lock(&osb->osb_lock);??????? <-- osb_lock
spinlock
>>  ????? ocfs2_node_map_set_bit()??????????? <-- uses node_map_lock
>>  ????? ...
>>  ????? spin_unlock(&osb->osb_lock);
>>  ??? ...
>>  ??? ocfs2_clear_recovering_orphan_dir()
>>  ????? ocfs2_node_map_clear_bit()??????? <-- uses node_map_lock
>>  ??????????????????????????? osb_lock is NOT taken
>>
>>
>>  ? ocfs2_check_orphan_recovery_state()
>>  ??? spin_lock(&osb->osb_lock);??????????? <-- osb_lock
spinlock
>>  ??? ...
>>  ??? ocfs2_node_map_test_bit()??????????? <-- uses node_map_lock
>>  ??? ...
>>  ??? spin_unlock(&osb->osb_lock);
>>
>>
>> So the code looks already protected by the 'osb_lock' spinlock,
but I don't know this code and ocfs2_mark_recovering_orphan_dir() looks
tricky to me. (so some other eyes are much welcome)
>   
> osb_lock is to protect osb filed such as 'osb_orphan_wipes', while
> node_map_lock is to protect the node map
'osb_recovering_orphan_dirs'
> specifically.
Thanks for this explanation.

But does "node_map_lock" really protects anything?
It is just around some atomic function calls which shouldn't need any, 
right?

test_bit() is not documented as atomic, but {clear|set}_bit() could be 
executed just before or just after it with the current locking 
mechanism, so I don't really see how it would make a difference.

I don't understand the logic of this lock here.

Can you elaborate?

CJ

> Thanks,
> Joseph
>

On 7/20/22 4:26 PM, Marion & Christophe JAILLET
wrote:
> 
> Le 20/07/2022 ? 03:59, Joseph Qi a ?crit?:
>>
>> On 7/19/22 9:25 PM, Christophe JAILLET wrote:
>>> Le 19/07/2022 ? 12:24, David Laight a ?crit?:
>>>> From: Christophe JAILLET
>>>>> Sent: 19 July 2022 11:02
>>>>>
>>>>> 'node_map_lock' is a spinlock only used to protect
calls to set_bit(),
>>>>> clear_bit() and test_bit().
>>>>>
>>>>> {set|clear}_bit() are already atomic and don't need
this extra spinlock.
>>>>> test_bit() only reads the bitmap for a given bit.
>>>>>
>>>>> Remove this useless spinlock.
>>>> It looks to me like the calling code is racy
>>>> unless there is another lock in the callers.
>>> The call chains are:
>>> ?? ocfs2_recover_orphans()
>>> ???? ocfs2_mark_recovering_orphan_dir()
>>> ?????? spin_lock(&osb->osb_lock);??????? <-- osb_lock
spinlock
>>> ?????? ocfs2_node_map_set_bit()??????????? <-- uses
node_map_lock
>>> ?????? ...
>>> ?????? spin_unlock(&osb->osb_lock);
>>> ???? ...
>>> ???? ocfs2_clear_recovering_orphan_dir()
>>> ?????? ocfs2_node_map_clear_bit()??????? <-- uses node_map_lock
>>> ???????????????????????????? osb_lock is NOT taken
>>>
>>>
>>> ?? ocfs2_check_orphan_recovery_state()
>>> ???? spin_lock(&osb->osb_lock);??????????? <-- osb_lock
spinlock
>>> ???? ...
>>> ???? ocfs2_node_map_test_bit()??????????? <-- uses node_map_lock
>>> ???? ...
>>> ???? spin_unlock(&osb->osb_lock);
>>>
>>>
>>> So the code looks already protected by the 'osb_lock'
spinlock, but I don't know this code and ocfs2_mark_recovering_orphan_dir()
looks tricky to me. (so some other eyes are much welcome)
>> ? osb_lock is to protect osb filed such as 'osb_orphan_wipes',
while
>> node_map_lock is to protect the node map
'osb_recovering_orphan_dirs'
>> specifically.
> 
> Thanks for this explanation.
> 
> But does "node_map_lock" really protects anything?
> It is just around some atomic function calls which shouldn't need any,
right?
> 
> test_bit() is not documented as atomic, but {clear|set}_bit() could be
executed just before or just after it with the current locking mechanism, so I
don't really see how it would make a difference.
> 
> I don't understand the logic of this lock here.
> 
> Can you elaborate?
These code are introduced long time ago...
Refer to commit b4df6ed8db0c "[PATCH] ocfs2: fix orphan recovery
deadlock", I guess it plays a role 'barrier' and make sure test
node map
is executed prior than signal orphan recovery thread. In other words, to
serialize evict inode and orphan recovery.

Thanks,
Joseph