[Ocfs2-devel] [PATCH] ocfs2/dlm: fix race between purge and get lock resource

[Joseph Qi <joseph.qi@huawei.com>]

Hi Mark,

On 2015/4/30 5:44, Mark Fasheh wrote:
> Hi Joseph, thanks for finding and trying to fix this bug.
> 
> On Sat, Apr 25, 2015 at 03:05:15PM +0800, Joseph Qi wrote:
>> There is a race between purge and get lock resource, which will lead to
>> ast unfinished and system hung. The case is described below:
>>
>> mkdir                                  dlm_thread
>> -----------------------------------------------------------------------
>> o2cb_dlm_lock                        |
>> -> dlmlock                           |
>>   -> dlm_get_lock_resource           |
>>     -> __dlm_lookup_lockres_full     |
>>       -> spin_unlock(&dlm->spinlock) |
>>                                      | dlm_run_purge_list
>>                                      | -> dlm_purge_lockres
>>                                      |   -> dlm_drop_lockres_ref
>>                                      |   -> spin_lock(&dlm->spinlock)
>>                                      |   -> spin_lock(&res->spinlock)
>>                                      |   -> ~DLM_LOCK_RES_DROPPING_REF
>>                                      |   -> spin_unlock(&res->spinlock)
>>                                      |   -> spin_unlock(&dlm->spinlock)
>>       -> spin_lock(&tmpres->spinlock)|
>>       DLM_LOCK_RES_DROPPING_REF cleared |
>>       -> spin_unlock(&tmpres->spinlock) |
>>       return the purged lockres         |
>>
>> So after this, once ast comes, it will ingore the ast because the
>> lockres cannot be found anymore. Thus the OCFS2_LOCK_BUSY won't be
>> cleared and corresponding thread hangs.
>> The &dlm->spinlock was hold when checking DLM_LOCK_RES_DROPPING_REF at
>> the very begining. And commit 7b791d6856 (ocfs2/dlm: Fix race during
>> lockres mastery) moved it up because of the possible wait.
>> So take the &dlm->spinlock and introduce a new wait function to fix the
>> race.
> 
> Ok, I _think_ I understand the deadlock. But we can't say for sure who will
> come and wake up the sleeping process? If that's the case I don't think we
> want this for -stable right now.
> 
For DLM_LOCK_RES_DROPPING_REF, only dlm_thread will do it during purging
lockres.
For others three (DLM_LOCK_RES_IN_PROGRESS|DLM_LOCK_RES_RECOVERING|
DLM_LOCK_RES_MIGRATING), many flows are involved.

This bug has been found in our test environment when running race testcase.
We can reproduce it in the following steps:
1) dreate a dir/file in N1, then N1 is the owner (take open lockres for example);
2) drop cache in N1 and then ls the dir/file in N2, so N2 is the owner now;
3) rm the dir/file in N1, the lockres is scheduled to be purged;
4) create the dir/file again in N1.
To 100% reproduce it, we can inject some delay right after lookup lockres and drop
the &dlm->spinlock.

> 
>>
>> Signed-off-by: Joseph Qi <joseph.qi@huawei.com>
>> Reviewed-by: joyce.xue <xuejiufei@huawei.com>
>> ---
>>  fs/ocfs2/dlm/dlmcommon.h |  2 ++
>>  fs/ocfs2/dlm/dlmmaster.c | 13 +++++++++----
>>  fs/ocfs2/dlm/dlmthread.c | 23 +++++++++++++++++++++++
>>  3 files changed, 34 insertions(+), 4 deletions(-)
>>
>> diff --git a/fs/ocfs2/dlm/dlmcommon.h b/fs/ocfs2/dlm/dlmcommon.h
>> index e88ccf8..c6b76f4 100644
>> --- a/fs/ocfs2/dlm/dlmcommon.h
>> +++ b/fs/ocfs2/dlm/dlmcommon.h
>> @@ -1014,6 +1014,8 @@ void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
>>
>>  /* will exit holding res->spinlock, but may drop in function */
>>  void __dlm_wait_on_lockres_flags(struct dlm_lock_resource *res, int flags);
>> +void __dlm_wait_on_lockres_flags_new(struct dlm_ctxt *dlm,
>> +		struct dlm_lock_resource *res, int flags);
> 
> If we retain this function, it should have a more descriptive name than
> '_new'.
> 
> 
>>  /* will exit holding res->spinlock, but may drop in function */
>>  static inline void __dlm_wait_on_lockres(struct dlm_lock_resource *res)
>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>> index a6944b2..9a5f45d 100644
>> --- a/fs/ocfs2/dlm/dlmmaster.c
>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>> @@ -755,13 +755,16 @@ lookup:
>>  	spin_lock(&dlm->spinlock);
>>  	tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
>>  	if (tmpres) {
>> -		spin_unlock(&dlm->spinlock);
>>  		spin_lock(&tmpres->spinlock);
>>  		/* Wait on the thread that is mastering the resource */
>>  		if (tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN) {
>> -			__dlm_wait_on_lockres(tmpres);
>> +			__dlm_wait_on_lockres_flags_new(dlm, tmpres,
>> +					(DLM_LOCK_RES_IN_PROGRESS|
>> +					DLM_LOCK_RES_RECOVERING|
>> +					DLM_LOCK_RES_MIGRATING));
>>  			BUG_ON(tmpres->owner == DLM_LOCK_RES_OWNER_UNKNOWN);
>>  			spin_unlock(&tmpres->spinlock);
>> +			spin_unlock(&dlm->spinlock);
>>  			dlm_lockres_put(tmpres);
>>  			tmpres = NULL;
>>  			goto lookup;
>> @@ -770,9 +773,10 @@ lookup:
>>  		/* Wait on the resource purge to complete before continuing */
>>  		if (tmpres->state & DLM_LOCK_RES_DROPPING_REF) {
>>  			BUG_ON(tmpres->owner == dlm->node_num);
>> -			__dlm_wait_on_lockres_flags(tmpres,
>> -						    DLM_LOCK_RES_DROPPING_REF);
>> +			__dlm_wait_on_lockres_flags_new(dlm, tmpres,
>> +				DLM_LOCK_RES_DROPPING_REF);
>>  			spin_unlock(&tmpres->spinlock);
>> +			spin_unlock(&dlm->spinlock);
>>  			dlm_lockres_put(tmpres);
>>  			tmpres = NULL;
>>  			goto lookup;
>> @@ -782,6 +786,7 @@ lookup:
>>  		dlm_lockres_grab_inflight_ref(dlm, tmpres);
>>
>>  		spin_unlock(&tmpres->spinlock);
>> +		spin_unlock(&dlm->spinlock);
>>  		if (res)
>>  			dlm_lockres_put(res);
>>  		res = tmpres;
>> diff --git a/fs/ocfs2/dlm/dlmthread.c b/fs/ocfs2/dlm/dlmthread.c
>> index 69aac6f..505730a 100644
>> --- a/fs/ocfs2/dlm/dlmthread.c
>> +++ b/fs/ocfs2/dlm/dlmthread.c
>> @@ -77,6 +77,29 @@ repeat:
>>  	__set_current_state(TASK_RUNNING);
>>  }
>>
>> +void __dlm_wait_on_lockres_flags_new(struct dlm_ctxt *dlm,
>> +		struct dlm_lock_resource *res, int flags)
>> +{
>> +	DECLARE_WAITQUEUE(wait, current);
>> +
>> +	assert_spin_locked(&dlm->spinlock);
>> +	assert_spin_locked(&res->spinlock);
>> +
>> +	add_wait_queue(&res->wq, &wait);
>> +repeat:
>> +	set_current_state(TASK_UNINTERRUPTIBLE);
>> +	if (res->state & flags) {
>> +		spin_unlock(&res->spinlock);
>> +		spin_unlock(&dlm->spinlock);
>> +		schedule();
>> +		spin_lock(&dlm->spinlock);
>> +		spin_lock(&res->spinlock);
>> +		goto repeat;
>> +	}
>> +	remove_wait_queue(&res->wq, &wait);
>> +	__set_current_state(TASK_RUNNING);
>> +}
> 
> Is it possible to rework this using wait_event()? The code you copied from
> __dlm_wait_on_lockres_flags() is seriously ugly :(
> 	--Mark
> 
Andrew has the same comments:)
Since __dlm_wait_on_lockres_flags is used in many other places, so I just
copied the logic and add mine to limit the impact.
Reworking and then testing it may take much time:)

> --
> Mark Fasheh
> 
> .
>