ocfs2/dlm: ignore cleaning the migration mle that is inuse
diff mbox

Message ID 56849B8A.80004@oracle.com
State New
Headers show

Commit Message

Junxiao Bi Dec. 31, 2015, 3:05 a.m. UTC
On 12/30/2015 05:56 PM, xuejiufei wrote:
> Hi Junxiao,
> You are right. But it may happen that mle->woken is set to 1 in
> dlm_clean_migration_mle() just after atomic_read() in
> dlm_migrate_lockres(). Actually we trigger this BUG when dlm_send_one_lockres()
> return error.
Yes, that's possible because of that 5s timeout wakeup, I think this
timeout is useless and can be removed?
> 
> And I think dlm_migrate_lockres() should not set owner to target and return 0
> when mle->woken is set to 1 in dlm_clean_migration_mle(). This is another problem?
Yes, it should be fixed, or the lockres owner will be set to a down node
wrongly. Can the following fix these two issue?


 	wake_up(&mle->wq);

Thanks,
Junxiao.


> 
> Thanks
> Jiufei.
> 
> On 2015/12/30 10:52, Junxiao Bi wrote:
>> Hi Jiufei,
>>
>> When target node down, mle is cleared from
>> dlm_do_local_recovery_cleanup()->dlm_clean_master_list()->dlm_clean_migration_mle()?
>> mle->woken is set to 1 in dlm_clean_migration_mle(), so the code to
>> detect target node down(if (dlm_is_node_dead(dlm, target))) will never
>> be run in dlm_migrate_lockres()?
>>
>>
>> 2621         while (1) {
>> 2622                 ret = wait_event_interruptible_timeout(mle->wq,
>> 2623                                         (atomic_read(&mle->woken)
>> == 1),
>> 2624                                         msecs_to_jiffies(5000));
>> 2625
>> 2626                 if (ret >= 0) {
>> 2627                         if (atomic_read(&mle->woken) == 1 ||
>> 2628                             res->owner == target)
>> 2629                                 break;
>> 2630
>> 2631                         mlog(0, "%s:%.*s: timed out during
>> migration\n",
>> 2632                              dlm->name, res->lockname.len,
>> res->lockname.name);
>> 2633                         /* avoid hang during shutdown when
>> migrating lockres
>> 2634                          * to a node which also goes down */
>> 2635                         if (dlm_is_node_dead(dlm, target)) {
>> 2636                                 mlog(0, "%s:%.*s: expected migration "
>> 2637                                      "target %u is no longer up,
>> restarting\n",
>> 2638                                      dlm->name, res->lockname.len,
>> 2639                                      res->lockname.name, target);
>> 2640                                 ret = -EINVAL;
>> 2641                                 /* migration failed, detach and
>> clean up mle */
>> 2642                                 dlm_mle_detach_hb_events(dlm, mle);
>> 2643                                 dlm_put_mle(mle);
>> 2644                                 dlm_put_mle_inuse(mle);
>> 2645                                 spin_lock(&res->spinlock);
>> 2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
>> 2647                                 wake = 1;
>> 2648                                 spin_unlock(&res->spinlock);
>> 2649                                 goto leave;
>> 2650                         }
>> 2651                 } else
>> 2652                         mlog(0, "%s:%.*s: caught signal during
>> migration\n",
>> 2653                              dlm->name, res->lockname.len,
>> res->lockname.name);
>> 2654         }
>>
>>
>> Thanks,
>> Junxiao.
>> On 12/28/2015 03:44 PM, xuejiufei wrote:
>>> We have found that migration source will trigger a BUG that the
>>> refcount of mle is already zero before put when the target is
>>> down during migration. The situation is as follows:
>>>
>>> dlm_migrate_lockres
>>>   dlm_add_migration_mle
>>>   dlm_mark_lockres_migrating
>>>   dlm_get_mle_inuse
>>>   <<<<<< Now the refcount of the mle is 2.
>>>   dlm_send_one_lockres and wait for the target to become the
>>>   new master.
>>>   <<<<<< o2hb detect the target down and clean the migration
>>>   mle. Now the refcount is 1.
>>>
>>> dlm_migrate_lockres woken, and put the mle twice when found
>>> the target goes down which trigger the BUG with the following
>>> message:
>>> "ERROR: bad mle: ".
>>>
>>> Signed-off-by: Jiufei Xue <xuejiufei@huawei.com>
>>> Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
>>> ---
>>>  fs/ocfs2/dlm/dlmmaster.c | 26 +++++++++++++++-----------
>>>  1 file changed, 15 insertions(+), 11 deletions(-)
>>>
>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>> index 936e11b..b713140 100644
>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>> @@ -2519,6 +2519,11 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
>>>  	spin_lock(&dlm->master_lock);
>>>  	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
>>>  				    namelen, target, dlm->node_num);
>>> +	/* get an extra reference on the mle.
>>> +	 * otherwise the assert_master from the new
>>> +	 * master will destroy this.
>>> +	 */
>>> +	dlm_get_mle_inuse(mle);
>>>  	spin_unlock(&dlm->master_lock);
>>>  	spin_unlock(&dlm->spinlock);
>>>
>>> @@ -2554,6 +2559,7 @@ fail:
>>>  		if (mle_added) {
>>>  			dlm_mle_detach_hb_events(dlm, mle);
>>>  			dlm_put_mle(mle);
>>> +			dlm_put_mle_inuse(mle);
>>>  		} else if (mle) {
>>>  			kmem_cache_free(dlm_mle_cache, mle);
>>>  			mle = NULL;
>>> @@ -2571,17 +2577,6 @@ fail:
>>>  	 * ensure that all assert_master work is flushed. */
>>>  	flush_workqueue(dlm->dlm_worker);
>>>
>>> -	/* get an extra reference on the mle.
>>> -	 * otherwise the assert_master from the new
>>> -	 * master will destroy this.
>>> -	 * also, make sure that all callers of dlm_get_mle
>>> -	 * take both dlm->spinlock and dlm->master_lock */
>>> -	spin_lock(&dlm->spinlock);
>>> -	spin_lock(&dlm->master_lock);
>>> -	dlm_get_mle_inuse(mle);
>>> -	spin_unlock(&dlm->master_lock);
>>> -	spin_unlock(&dlm->spinlock);
>>> -
>>>  	/* notify new node and send all lock state */
>>>  	/* call send_one_lockres with migration flag.
>>>  	 * this serves as notice to the target node that a
>>> @@ -3312,6 +3307,15 @@ top:
>>>  			    mle->new_master != dead_node)
>>>  				continue;
>>>
>>> +			if (mle->new_master == dead_node && mle->inuse) {
>>> +				mlog(ML_NOTICE, "%s: target %u died during "
>>> +						"migration from %u, the MLE is "
>>> +						"still keep used, ignore it!\n",
>>> +						dlm->name, dead_node,
>>> +						mle->master);
>>> +				continue;
>>> +			}
>>> +
>>>  			/* If we have reached this point, this mle needs to be
>>>  			 * removed from the list and freed. */
>>>  			dlm_clean_migration_mle(dlm, mle);
>>>
>>
>>
>> .
>>
>

Comments

Xue jiufei Dec. 31, 2015, 7:15 a.m. UTC | #1
Hi Junxiao,
On 2015/12/31 11:05, Junxiao Bi wrote:
> On 12/30/2015 05:56 PM, xuejiufei wrote:
>> Hi Junxiao,
>> You are right. But it may happen that mle->woken is set to 1 in
>> dlm_clean_migration_mle() just after atomic_read() in
>> dlm_migrate_lockres(). Actually we trigger this BUG when dlm_send_one_lockres()
>> return error.
> Yes, that's possible because of that 5s timeout wakeup, I think this
> timeout is useless and can be removed?
>>
>> And I think dlm_migrate_lockres() should not set owner to target and return 0
>> when mle->woken is set to 1 in dlm_clean_migration_mle(). This is another problem?
> Yes, it should be fixed, or the lockres owner will be set to a down node
> wrongly. Can the following fix these two issue?
> 
It can not fix the first issue when dlm_send_one_lockres() or
dlm_mark_lockres_migrating() return error, right?

> 
> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
> index 84f2f8079466..d0380ea62340 100644
> --- a/fs/ocfs2/dlm/dlmmaster.c
> +++ b/fs/ocfs2/dlm/dlmmaster.c
> @@ -2618,39 +2618,22 @@ fail:
> 
> 
>  	/* wait for new node to assert master */
> -	while (1) {
> -		ret = wait_event_interruptible_timeout(mle->wq,
> -					(atomic_read(&mle->woken) == 1),
> -					msecs_to_jiffies(5000));
> -
> -		if (ret >= 0) {
> -		       	if (atomic_read(&mle->woken) == 1 ||
> -			    res->owner == target)
> -				break;
> -
> -			mlog(0, "%s:%.*s: timed out during migration\n",
> -			     dlm->name, res->lockname.len, res->lockname.name);
> -			/* avoid hang during shutdown when migrating lockres
> -			 * to a node which also goes down */
> -			if (dlm_is_node_dead(dlm, target)) {
> -				mlog(0, "%s:%.*s: expected migration "
> -				     "target %u is no longer up, restarting\n",
> -				     dlm->name, res->lockname.len,
> -				     res->lockname.name, target);
> -				ret = -EINVAL;
> -				/* migration failed, detach and clean up mle */
> -				dlm_mle_detach_hb_events(dlm, mle);
> -				dlm_put_mle(mle);
> -				dlm_put_mle_inuse(mle);
> -				spin_lock(&res->spinlock);
> -				res->state &= ~DLM_LOCK_RES_MIGRATING;
> -				wake = 1;
> -				spin_unlock(&res->spinlock);
> -				goto leave;
> -			}
> -		} else
> -			mlog(0, "%s:%.*s: caught signal during migration\n",
> -			     dlm->name, res->lockname.len, res->lockname.name);
> +	wait_event(mle->wq, ((atomic_read(&mle->woken) == 1) ||
> +				(atomic_read(&mle->woken) == 2)));
> +
> +	/* migrate target down */
> +	if (atomic_read(&mle->woken) == 2) {
> +		mlog(0, "%s:%.*s: expected migration "
> +				"target %u is no longer up, restarting\n",
> +				dlm->name, res->lockname.len,
> +				res->lockname.name, target);
> +		ret = -EINVAL;
> +		dlm_put_mle_inuse(mle);
> +		spin_lock(&res->spinlock);
> +		res->state &= ~DLM_LOCK_RES_MIGRATING;
> +		wake = 1;
> +		spin_unlock(&res->spinlock);
> +		goto leave;
>  	}
> 
>  	/* all done, set the owner, clear the flag */
> @@ -3227,7 +3210,7 @@ static void dlm_clean_migration_mle(struct
> dlm_ctxt *dlm,
> 
>  	spin_lock(&mle->spinlock);
>  	__dlm_unlink_mle(dlm, mle);
> -	atomic_set(&mle->woken, 1);
> +	atomic_set(&mle->woken, 2);
>  	spin_unlock(&mle->spinlock);
> 
>  	wake_up(&mle->wq);
> 
> Thanks,
> Junxiao.
> 
> 
>>
>> Thanks
>> Jiufei.
>>
>> On 2015/12/30 10:52, Junxiao Bi wrote:
>>> Hi Jiufei,
>>>
>>> When target node down, mle is cleared from
>>> dlm_do_local_recovery_cleanup()->dlm_clean_master_list()->dlm_clean_migration_mle()?
>>> mle->woken is set to 1 in dlm_clean_migration_mle(), so the code to
>>> detect target node down(if (dlm_is_node_dead(dlm, target))) will never
>>> be run in dlm_migrate_lockres()?
>>>
>>>
>>> 2621         while (1) {
>>> 2622                 ret = wait_event_interruptible_timeout(mle->wq,
>>> 2623                                         (atomic_read(&mle->woken)
>>> == 1),
>>> 2624                                         msecs_to_jiffies(5000));
>>> 2625
>>> 2626                 if (ret >= 0) {
>>> 2627                         if (atomic_read(&mle->woken) == 1 ||
>>> 2628                             res->owner == target)
>>> 2629                                 break;
>>> 2630
>>> 2631                         mlog(0, "%s:%.*s: timed out during
>>> migration\n",
>>> 2632                              dlm->name, res->lockname.len,
>>> res->lockname.name);
>>> 2633                         /* avoid hang during shutdown when
>>> migrating lockres
>>> 2634                          * to a node which also goes down */
>>> 2635                         if (dlm_is_node_dead(dlm, target)) {
>>> 2636                                 mlog(0, "%s:%.*s: expected migration "
>>> 2637                                      "target %u is no longer up,
>>> restarting\n",
>>> 2638                                      dlm->name, res->lockname.len,
>>> 2639                                      res->lockname.name, target);
>>> 2640                                 ret = -EINVAL;
>>> 2641                                 /* migration failed, detach and
>>> clean up mle */
>>> 2642                                 dlm_mle_detach_hb_events(dlm, mle);
>>> 2643                                 dlm_put_mle(mle);
>>> 2644                                 dlm_put_mle_inuse(mle);
>>> 2645                                 spin_lock(&res->spinlock);
>>> 2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
>>> 2647                                 wake = 1;
>>> 2648                                 spin_unlock(&res->spinlock);
>>> 2649                                 goto leave;
>>> 2650                         }
>>> 2651                 } else
>>> 2652                         mlog(0, "%s:%.*s: caught signal during
>>> migration\n",
>>> 2653                              dlm->name, res->lockname.len,
>>> res->lockname.name);
>>> 2654         }
>>>
>>>
>>> Thanks,
>>> Junxiao.
>>> On 12/28/2015 03:44 PM, xuejiufei wrote:
>>>> We have found that migration source will trigger a BUG that the
>>>> refcount of mle is already zero before put when the target is
>>>> down during migration. The situation is as follows:
>>>>
>>>> dlm_migrate_lockres
>>>>   dlm_add_migration_mle
>>>>   dlm_mark_lockres_migrating
>>>>   dlm_get_mle_inuse
>>>>   <<<<<< Now the refcount of the mle is 2.
>>>>   dlm_send_one_lockres and wait for the target to become the
>>>>   new master.
>>>>   <<<<<< o2hb detect the target down and clean the migration
>>>>   mle. Now the refcount is 1.
>>>>
>>>> dlm_migrate_lockres woken, and put the mle twice when found
>>>> the target goes down which trigger the BUG with the following
>>>> message:
>>>> "ERROR: bad mle: ".
>>>>
>>>> Signed-off-by: Jiufei Xue <xuejiufei@huawei.com>
>>>> Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
>>>> ---
>>>>  fs/ocfs2/dlm/dlmmaster.c | 26 +++++++++++++++-----------
>>>>  1 file changed, 15 insertions(+), 11 deletions(-)
>>>>
>>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>>> index 936e11b..b713140 100644
>>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>>> @@ -2519,6 +2519,11 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
>>>>  	spin_lock(&dlm->master_lock);
>>>>  	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
>>>>  				    namelen, target, dlm->node_num);
>>>> +	/* get an extra reference on the mle.
>>>> +	 * otherwise the assert_master from the new
>>>> +	 * master will destroy this.
>>>> +	 */
>>>> +	dlm_get_mle_inuse(mle);
>>>>  	spin_unlock(&dlm->master_lock);
>>>>  	spin_unlock(&dlm->spinlock);
>>>>
>>>> @@ -2554,6 +2559,7 @@ fail:
>>>>  		if (mle_added) {
>>>>  			dlm_mle_detach_hb_events(dlm, mle);
>>>>  			dlm_put_mle(mle);
>>>> +			dlm_put_mle_inuse(mle);
>>>>  		} else if (mle) {
>>>>  			kmem_cache_free(dlm_mle_cache, mle);
>>>>  			mle = NULL;
>>>> @@ -2571,17 +2577,6 @@ fail:
>>>>  	 * ensure that all assert_master work is flushed. */
>>>>  	flush_workqueue(dlm->dlm_worker);
>>>>
>>>> -	/* get an extra reference on the mle.
>>>> -	 * otherwise the assert_master from the new
>>>> -	 * master will destroy this.
>>>> -	 * also, make sure that all callers of dlm_get_mle
>>>> -	 * take both dlm->spinlock and dlm->master_lock */
>>>> -	spin_lock(&dlm->spinlock);
>>>> -	spin_lock(&dlm->master_lock);
>>>> -	dlm_get_mle_inuse(mle);
>>>> -	spin_unlock(&dlm->master_lock);
>>>> -	spin_unlock(&dlm->spinlock);
>>>> -
>>>>  	/* notify new node and send all lock state */
>>>>  	/* call send_one_lockres with migration flag.
>>>>  	 * this serves as notice to the target node that a
>>>> @@ -3312,6 +3307,15 @@ top:
>>>>  			    mle->new_master != dead_node)
>>>>  				continue;
>>>>
>>>> +			if (mle->new_master == dead_node && mle->inuse) {
>>>> +				mlog(ML_NOTICE, "%s: target %u died during "
>>>> +						"migration from %u, the MLE is "
>>>> +						"still keep used, ignore it!\n",
>>>> +						dlm->name, dead_node,
>>>> +						mle->master);
>>>> +				continue;
>>>> +			}
>>>> +
>>>>  			/* If we have reached this point, this mle needs to be
>>>>  			 * removed from the list and freed. */
>>>>  			dlm_clean_migration_mle(dlm, mle);
>>>>
>>>
>>>
>>> .
>>>
>>
> 
> 
> .
>
Junxiao Bi Jan. 4, 2016, 8:20 a.m. UTC | #2
Hi Jiufei,

On 12/31/2015 03:15 PM, xuejiufei wrote:
> Hi Junxiao,
> On 2015/12/31 11:05, Junxiao Bi wrote:
>> On 12/30/2015 05:56 PM, xuejiufei wrote:
>>> Hi Junxiao,
>>> You are right. But it may happen that mle->woken is set to 1 in
>>> dlm_clean_migration_mle() just after atomic_read() in
>>> dlm_migrate_lockres(). Actually we trigger this BUG when dlm_send_one_lockres()
>>> return error.
>> Yes, that's possible because of that 5s timeout wakeup, I think this
>> timeout is useless and can be removed?
>>>
>>> And I think dlm_migrate_lockres() should not set owner to target and return 0
>>> when mle->woken is set to 1 in dlm_clean_migration_mle(). This is another problem?
>> Yes, it should be fixed, or the lockres owner will be set to a down node
>> wrongly. Can the following fix these two issue?
>>
> It can not fix the first issue when dlm_send_one_lockres() or
> dlm_mark_lockres_migrating() return error, right?
Right. This is for second issue. Any way please consider these two
issue, I think they can be fixed in one patch since both happened when
target node down during migrate. You can make your own one or merge this
patch to yours if you like.

Thanks,
Junxiao.
> 
>>
>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>> index 84f2f8079466..d0380ea62340 100644
>> --- a/fs/ocfs2/dlm/dlmmaster.c
>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>> @@ -2618,39 +2618,22 @@ fail:
>>
>>
>>  	/* wait for new node to assert master */
>> -	while (1) {
>> -		ret = wait_event_interruptible_timeout(mle->wq,
>> -					(atomic_read(&mle->woken) == 1),
>> -					msecs_to_jiffies(5000));
>> -
>> -		if (ret >= 0) {
>> -		       	if (atomic_read(&mle->woken) == 1 ||
>> -			    res->owner == target)
>> -				break;
>> -
>> -			mlog(0, "%s:%.*s: timed out during migration\n",
>> -			     dlm->name, res->lockname.len, res->lockname.name);
>> -			/* avoid hang during shutdown when migrating lockres
>> -			 * to a node which also goes down */
>> -			if (dlm_is_node_dead(dlm, target)) {
>> -				mlog(0, "%s:%.*s: expected migration "
>> -				     "target %u is no longer up, restarting\n",
>> -				     dlm->name, res->lockname.len,
>> -				     res->lockname.name, target);
>> -				ret = -EINVAL;
>> -				/* migration failed, detach and clean up mle */
>> -				dlm_mle_detach_hb_events(dlm, mle);
>> -				dlm_put_mle(mle);
>> -				dlm_put_mle_inuse(mle);
>> -				spin_lock(&res->spinlock);
>> -				res->state &= ~DLM_LOCK_RES_MIGRATING;
>> -				wake = 1;
>> -				spin_unlock(&res->spinlock);
>> -				goto leave;
>> -			}
>> -		} else
>> -			mlog(0, "%s:%.*s: caught signal during migration\n",
>> -			     dlm->name, res->lockname.len, res->lockname.name);
>> +	wait_event(mle->wq, ((atomic_read(&mle->woken) == 1) ||
>> +				(atomic_read(&mle->woken) == 2)));
>> +
>> +	/* migrate target down */
>> +	if (atomic_read(&mle->woken) == 2) {
>> +		mlog(0, "%s:%.*s: expected migration "
>> +				"target %u is no longer up, restarting\n",
>> +				dlm->name, res->lockname.len,
>> +				res->lockname.name, target);
>> +		ret = -EINVAL;
>> +		dlm_put_mle_inuse(mle);
>> +		spin_lock(&res->spinlock);
>> +		res->state &= ~DLM_LOCK_RES_MIGRATING;
>> +		wake = 1;
>> +		spin_unlock(&res->spinlock);
>> +		goto leave;
>>  	}
>>
>>  	/* all done, set the owner, clear the flag */
>> @@ -3227,7 +3210,7 @@ static void dlm_clean_migration_mle(struct
>> dlm_ctxt *dlm,
>>
>>  	spin_lock(&mle->spinlock);
>>  	__dlm_unlink_mle(dlm, mle);
>> -	atomic_set(&mle->woken, 1);
>> +	atomic_set(&mle->woken, 2);
>>  	spin_unlock(&mle->spinlock);
>>
>>  	wake_up(&mle->wq);
>>
>> Thanks,
>> Junxiao.
>>
>>
>>>
>>> Thanks
>>> Jiufei.
>>>
>>> On 2015/12/30 10:52, Junxiao Bi wrote:
>>>> Hi Jiufei,
>>>>
>>>> When target node down, mle is cleared from
>>>> dlm_do_local_recovery_cleanup()->dlm_clean_master_list()->dlm_clean_migration_mle()?
>>>> mle->woken is set to 1 in dlm_clean_migration_mle(), so the code to
>>>> detect target node down(if (dlm_is_node_dead(dlm, target))) will never
>>>> be run in dlm_migrate_lockres()?
>>>>
>>>>
>>>> 2621         while (1) {
>>>> 2622                 ret = wait_event_interruptible_timeout(mle->wq,
>>>> 2623                                         (atomic_read(&mle->woken)
>>>> == 1),
>>>> 2624                                         msecs_to_jiffies(5000));
>>>> 2625
>>>> 2626                 if (ret >= 0) {
>>>> 2627                         if (atomic_read(&mle->woken) == 1 ||
>>>> 2628                             res->owner == target)
>>>> 2629                                 break;
>>>> 2630
>>>> 2631                         mlog(0, "%s:%.*s: timed out during
>>>> migration\n",
>>>> 2632                              dlm->name, res->lockname.len,
>>>> res->lockname.name);
>>>> 2633                         /* avoid hang during shutdown when
>>>> migrating lockres
>>>> 2634                          * to a node which also goes down */
>>>> 2635                         if (dlm_is_node_dead(dlm, target)) {
>>>> 2636                                 mlog(0, "%s:%.*s: expected migration "
>>>> 2637                                      "target %u is no longer up,
>>>> restarting\n",
>>>> 2638                                      dlm->name, res->lockname.len,
>>>> 2639                                      res->lockname.name, target);
>>>> 2640                                 ret = -EINVAL;
>>>> 2641                                 /* migration failed, detach and
>>>> clean up mle */
>>>> 2642                                 dlm_mle_detach_hb_events(dlm, mle);
>>>> 2643                                 dlm_put_mle(mle);
>>>> 2644                                 dlm_put_mle_inuse(mle);
>>>> 2645                                 spin_lock(&res->spinlock);
>>>> 2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
>>>> 2647                                 wake = 1;
>>>> 2648                                 spin_unlock(&res->spinlock);
>>>> 2649                                 goto leave;
>>>> 2650                         }
>>>> 2651                 } else
>>>> 2652                         mlog(0, "%s:%.*s: caught signal during
>>>> migration\n",
>>>> 2653                              dlm->name, res->lockname.len,
>>>> res->lockname.name);
>>>> 2654         }
>>>>
>>>>
>>>> Thanks,
>>>> Junxiao.
>>>> On 12/28/2015 03:44 PM, xuejiufei wrote:
>>>>> We have found that migration source will trigger a BUG that the
>>>>> refcount of mle is already zero before put when the target is
>>>>> down during migration. The situation is as follows:
>>>>>
>>>>> dlm_migrate_lockres
>>>>>   dlm_add_migration_mle
>>>>>   dlm_mark_lockres_migrating
>>>>>   dlm_get_mle_inuse
>>>>>   <<<<<< Now the refcount of the mle is 2.
>>>>>   dlm_send_one_lockres and wait for the target to become the
>>>>>   new master.
>>>>>   <<<<<< o2hb detect the target down and clean the migration
>>>>>   mle. Now the refcount is 1.
>>>>>
>>>>> dlm_migrate_lockres woken, and put the mle twice when found
>>>>> the target goes down which trigger the BUG with the following
>>>>> message:
>>>>> "ERROR: bad mle: ".
>>>>>
>>>>> Signed-off-by: Jiufei Xue <xuejiufei@huawei.com>
>>>>> Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
>>>>> ---
>>>>>  fs/ocfs2/dlm/dlmmaster.c | 26 +++++++++++++++-----------
>>>>>  1 file changed, 15 insertions(+), 11 deletions(-)
>>>>>
>>>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>>>> index 936e11b..b713140 100644
>>>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>>>> @@ -2519,6 +2519,11 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
>>>>>  	spin_lock(&dlm->master_lock);
>>>>>  	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
>>>>>  				    namelen, target, dlm->node_num);
>>>>> +	/* get an extra reference on the mle.
>>>>> +	 * otherwise the assert_master from the new
>>>>> +	 * master will destroy this.
>>>>> +	 */
>>>>> +	dlm_get_mle_inuse(mle);
>>>>>  	spin_unlock(&dlm->master_lock);
>>>>>  	spin_unlock(&dlm->spinlock);
>>>>>
>>>>> @@ -2554,6 +2559,7 @@ fail:
>>>>>  		if (mle_added) {
>>>>>  			dlm_mle_detach_hb_events(dlm, mle);
>>>>>  			dlm_put_mle(mle);
>>>>> +			dlm_put_mle_inuse(mle);
>>>>>  		} else if (mle) {
>>>>>  			kmem_cache_free(dlm_mle_cache, mle);
>>>>>  			mle = NULL;
>>>>> @@ -2571,17 +2577,6 @@ fail:
>>>>>  	 * ensure that all assert_master work is flushed. */
>>>>>  	flush_workqueue(dlm->dlm_worker);
>>>>>
>>>>> -	/* get an extra reference on the mle.
>>>>> -	 * otherwise the assert_master from the new
>>>>> -	 * master will destroy this.
>>>>> -	 * also, make sure that all callers of dlm_get_mle
>>>>> -	 * take both dlm->spinlock and dlm->master_lock */
>>>>> -	spin_lock(&dlm->spinlock);
>>>>> -	spin_lock(&dlm->master_lock);
>>>>> -	dlm_get_mle_inuse(mle);
>>>>> -	spin_unlock(&dlm->master_lock);
>>>>> -	spin_unlock(&dlm->spinlock);
>>>>> -
>>>>>  	/* notify new node and send all lock state */
>>>>>  	/* call send_one_lockres with migration flag.
>>>>>  	 * this serves as notice to the target node that a
>>>>> @@ -3312,6 +3307,15 @@ top:
>>>>>  			    mle->new_master != dead_node)
>>>>>  				continue;
>>>>>
>>>>> +			if (mle->new_master == dead_node && mle->inuse) {
>>>>> +				mlog(ML_NOTICE, "%s: target %u died during "
>>>>> +						"migration from %u, the MLE is "
>>>>> +						"still keep used, ignore it!\n",
>>>>> +						dlm->name, dead_node,
>>>>> +						mle->master);
>>>>> +				continue;
>>>>> +			}
>>>>> +
>>>>>  			/* If we have reached this point, this mle needs to be
>>>>>  			 * removed from the list and freed. */
>>>>>  			dlm_clean_migration_mle(dlm, mle);
>>>>>
>>>>
>>>>
>>>> .
>>>>
>>>
>>
>>
>> .
>>
>
Xue jiufei Jan. 6, 2016, 12:57 a.m. UTC | #3
Hi Junxiao,
Function dlm_clean_migration_mle() will not be called after the first
patch applied. So the issue that the owner of lockres will be set to
a down node is not exist.

Thanks,
Jiufei

On 2016/1/4 16:20, Junxiao Bi wrote:
> Hi Jiufei,
> 
> On 12/31/2015 03:15 PM, xuejiufei wrote:
>> Hi Junxiao,
>> On 2015/12/31 11:05, Junxiao Bi wrote:
>>> On 12/30/2015 05:56 PM, xuejiufei wrote:
>>>> Hi Junxiao,
>>>> You are right. But it may happen that mle->woken is set to 1 in
>>>> dlm_clean_migration_mle() just after atomic_read() in
>>>> dlm_migrate_lockres(). Actually we trigger this BUG when dlm_send_one_lockres()
>>>> return error.
>>> Yes, that's possible because of that 5s timeout wakeup, I think this
>>> timeout is useless and can be removed?
>>>>
>>>> And I think dlm_migrate_lockres() should not set owner to target and return 0
>>>> when mle->woken is set to 1 in dlm_clean_migration_mle(). This is another problem?
>>> Yes, it should be fixed, or the lockres owner will be set to a down node
>>> wrongly. Can the following fix these two issue?
>>>
>> It can not fix the first issue when dlm_send_one_lockres() or
>> dlm_mark_lockres_migrating() return error, right?
> Right. This is for second issue. Any way please consider these two
> issue, I think they can be fixed in one patch since both happened when
> target node down during migrate. You can make your own one or merge this
> patch to yours if you like.
> 
> Thanks,
> Junxiao.
>>
>>>
>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>> index 84f2f8079466..d0380ea62340 100644
>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>> @@ -2618,39 +2618,22 @@ fail:
>>>
>>>
>>>  	/* wait for new node to assert master */
>>> -	while (1) {
>>> -		ret = wait_event_interruptible_timeout(mle->wq,
>>> -					(atomic_read(&mle->woken) == 1),
>>> -					msecs_to_jiffies(5000));
>>> -
>>> -		if (ret >= 0) {
>>> -		       	if (atomic_read(&mle->woken) == 1 ||
>>> -			    res->owner == target)
>>> -				break;
>>> -
>>> -			mlog(0, "%s:%.*s: timed out during migration\n",
>>> -			     dlm->name, res->lockname.len, res->lockname.name);
>>> -			/* avoid hang during shutdown when migrating lockres
>>> -			 * to a node which also goes down */
>>> -			if (dlm_is_node_dead(dlm, target)) {
>>> -				mlog(0, "%s:%.*s: expected migration "
>>> -				     "target %u is no longer up, restarting\n",
>>> -				     dlm->name, res->lockname.len,
>>> -				     res->lockname.name, target);
>>> -				ret = -EINVAL;
>>> -				/* migration failed, detach and clean up mle */
>>> -				dlm_mle_detach_hb_events(dlm, mle);
>>> -				dlm_put_mle(mle);
>>> -				dlm_put_mle_inuse(mle);
>>> -				spin_lock(&res->spinlock);
>>> -				res->state &= ~DLM_LOCK_RES_MIGRATING;
>>> -				wake = 1;
>>> -				spin_unlock(&res->spinlock);
>>> -				goto leave;
>>> -			}
>>> -		} else
>>> -			mlog(0, "%s:%.*s: caught signal during migration\n",
>>> -			     dlm->name, res->lockname.len, res->lockname.name);
>>> +	wait_event(mle->wq, ((atomic_read(&mle->woken) == 1) ||
>>> +				(atomic_read(&mle->woken) == 2)));
>>> +
>>> +	/* migrate target down */
>>> +	if (atomic_read(&mle->woken) == 2) {
>>> +		mlog(0, "%s:%.*s: expected migration "
>>> +				"target %u is no longer up, restarting\n",
>>> +				dlm->name, res->lockname.len,
>>> +				res->lockname.name, target);
>>> +		ret = -EINVAL;
>>> +		dlm_put_mle_inuse(mle);
>>> +		spin_lock(&res->spinlock);
>>> +		res->state &= ~DLM_LOCK_RES_MIGRATING;
>>> +		wake = 1;
>>> +		spin_unlock(&res->spinlock);
>>> +		goto leave;
>>>  	}
>>>
>>>  	/* all done, set the owner, clear the flag */
>>> @@ -3227,7 +3210,7 @@ static void dlm_clean_migration_mle(struct
>>> dlm_ctxt *dlm,
>>>
>>>  	spin_lock(&mle->spinlock);
>>>  	__dlm_unlink_mle(dlm, mle);
>>> -	atomic_set(&mle->woken, 1);
>>> +	atomic_set(&mle->woken, 2);
>>>  	spin_unlock(&mle->spinlock);
>>>
>>>  	wake_up(&mle->wq);
>>>
>>> Thanks,
>>> Junxiao.
>>>
>>>
>>>>
>>>> Thanks
>>>> Jiufei.
>>>>
>>>> On 2015/12/30 10:52, Junxiao Bi wrote:
>>>>> Hi Jiufei,
>>>>>
>>>>> When target node down, mle is cleared from
>>>>> dlm_do_local_recovery_cleanup()->dlm_clean_master_list()->dlm_clean_migration_mle()?
>>>>> mle->woken is set to 1 in dlm_clean_migration_mle(), so the code to
>>>>> detect target node down(if (dlm_is_node_dead(dlm, target))) will never
>>>>> be run in dlm_migrate_lockres()?
>>>>>
>>>>>
>>>>> 2621         while (1) {
>>>>> 2622                 ret = wait_event_interruptible_timeout(mle->wq,
>>>>> 2623                                         (atomic_read(&mle->woken)
>>>>> == 1),
>>>>> 2624                                         msecs_to_jiffies(5000));
>>>>> 2625
>>>>> 2626                 if (ret >= 0) {
>>>>> 2627                         if (atomic_read(&mle->woken) == 1 ||
>>>>> 2628                             res->owner == target)
>>>>> 2629                                 break;
>>>>> 2630
>>>>> 2631                         mlog(0, "%s:%.*s: timed out during
>>>>> migration\n",
>>>>> 2632                              dlm->name, res->lockname.len,
>>>>> res->lockname.name);
>>>>> 2633                         /* avoid hang during shutdown when
>>>>> migrating lockres
>>>>> 2634                          * to a node which also goes down */
>>>>> 2635                         if (dlm_is_node_dead(dlm, target)) {
>>>>> 2636                                 mlog(0, "%s:%.*s: expected migration "
>>>>> 2637                                      "target %u is no longer up,
>>>>> restarting\n",
>>>>> 2638                                      dlm->name, res->lockname.len,
>>>>> 2639                                      res->lockname.name, target);
>>>>> 2640                                 ret = -EINVAL;
>>>>> 2641                                 /* migration failed, detach and
>>>>> clean up mle */
>>>>> 2642                                 dlm_mle_detach_hb_events(dlm, mle);
>>>>> 2643                                 dlm_put_mle(mle);
>>>>> 2644                                 dlm_put_mle_inuse(mle);
>>>>> 2645                                 spin_lock(&res->spinlock);
>>>>> 2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
>>>>> 2647                                 wake = 1;
>>>>> 2648                                 spin_unlock(&res->spinlock);
>>>>> 2649                                 goto leave;
>>>>> 2650                         }
>>>>> 2651                 } else
>>>>> 2652                         mlog(0, "%s:%.*s: caught signal during
>>>>> migration\n",
>>>>> 2653                              dlm->name, res->lockname.len,
>>>>> res->lockname.name);
>>>>> 2654         }
>>>>>
>>>>>
>>>>> Thanks,
>>>>> Junxiao.
>>>>> On 12/28/2015 03:44 PM, xuejiufei wrote:
>>>>>> We have found that migration source will trigger a BUG that the
>>>>>> refcount of mle is already zero before put when the target is
>>>>>> down during migration. The situation is as follows:
>>>>>>
>>>>>> dlm_migrate_lockres
>>>>>>   dlm_add_migration_mle
>>>>>>   dlm_mark_lockres_migrating
>>>>>>   dlm_get_mle_inuse
>>>>>>   <<<<<< Now the refcount of the mle is 2.
>>>>>>   dlm_send_one_lockres and wait for the target to become the
>>>>>>   new master.
>>>>>>   <<<<<< o2hb detect the target down and clean the migration
>>>>>>   mle. Now the refcount is 1.
>>>>>>
>>>>>> dlm_migrate_lockres woken, and put the mle twice when found
>>>>>> the target goes down which trigger the BUG with the following
>>>>>> message:
>>>>>> "ERROR: bad mle: ".
>>>>>>
>>>>>> Signed-off-by: Jiufei Xue <xuejiufei@huawei.com>
>>>>>> Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
>>>>>> ---
>>>>>>  fs/ocfs2/dlm/dlmmaster.c | 26 +++++++++++++++-----------
>>>>>>  1 file changed, 15 insertions(+), 11 deletions(-)
>>>>>>
>>>>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>>>>> index 936e11b..b713140 100644
>>>>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>>>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>>>>> @@ -2519,6 +2519,11 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
>>>>>>  	spin_lock(&dlm->master_lock);
>>>>>>  	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
>>>>>>  				    namelen, target, dlm->node_num);
>>>>>> +	/* get an extra reference on the mle.
>>>>>> +	 * otherwise the assert_master from the new
>>>>>> +	 * master will destroy this.
>>>>>> +	 */
>>>>>> +	dlm_get_mle_inuse(mle);
>>>>>>  	spin_unlock(&dlm->master_lock);
>>>>>>  	spin_unlock(&dlm->spinlock);
>>>>>>
>>>>>> @@ -2554,6 +2559,7 @@ fail:
>>>>>>  		if (mle_added) {
>>>>>>  			dlm_mle_detach_hb_events(dlm, mle);
>>>>>>  			dlm_put_mle(mle);
>>>>>> +			dlm_put_mle_inuse(mle);
>>>>>>  		} else if (mle) {
>>>>>>  			kmem_cache_free(dlm_mle_cache, mle);
>>>>>>  			mle = NULL;
>>>>>> @@ -2571,17 +2577,6 @@ fail:
>>>>>>  	 * ensure that all assert_master work is flushed. */
>>>>>>  	flush_workqueue(dlm->dlm_worker);
>>>>>>
>>>>>> -	/* get an extra reference on the mle.
>>>>>> -	 * otherwise the assert_master from the new
>>>>>> -	 * master will destroy this.
>>>>>> -	 * also, make sure that all callers of dlm_get_mle
>>>>>> -	 * take both dlm->spinlock and dlm->master_lock */
>>>>>> -	spin_lock(&dlm->spinlock);
>>>>>> -	spin_lock(&dlm->master_lock);
>>>>>> -	dlm_get_mle_inuse(mle);
>>>>>> -	spin_unlock(&dlm->master_lock);
>>>>>> -	spin_unlock(&dlm->spinlock);
>>>>>> -
>>>>>>  	/* notify new node and send all lock state */
>>>>>>  	/* call send_one_lockres with migration flag.
>>>>>>  	 * this serves as notice to the target node that a
>>>>>> @@ -3312,6 +3307,15 @@ top:
>>>>>>  			    mle->new_master != dead_node)
>>>>>>  				continue;
>>>>>>
>>>>>> +			if (mle->new_master == dead_node && mle->inuse) {
>>>>>> +				mlog(ML_NOTICE, "%s: target %u died during "
>>>>>> +						"migration from %u, the MLE is "
>>>>>> +						"still keep used, ignore it!\n",
>>>>>> +						dlm->name, dead_node,
>>>>>> +						mle->master);
>>>>>> +				continue;
>>>>>> +			}
>>>>>> +
>>>>>>  			/* If we have reached this point, this mle needs to be
>>>>>>  			 * removed from the list and freed. */
>>>>>>  			dlm_clean_migration_mle(dlm, mle);
>>>>>>
>>>>>
>>>>>
>>>>> .
>>>>>
>>>>
>>>
>>>
>>> .
>>>
>>
> 
> 
> .
>
Junxiao Bi Jan. 7, 2016, 1:46 a.m. UTC | #4
On 01/06/2016 08:57 AM, xuejiufei wrote:
> Hi Junxiao,
> Function dlm_clean_migration_mle() will not be called after the first
> patch applied. So the issue that the owner of lockres will be set to
> a down node is not exist.
Right. Thank you for the explanation.

Thanks,
Junxiao.
> 
> Thanks,
> Jiufei
> 
> On 2016/1/4 16:20, Junxiao Bi wrote:
>> Hi Jiufei,
>>
>> On 12/31/2015 03:15 PM, xuejiufei wrote:
>>> Hi Junxiao,
>>> On 2015/12/31 11:05, Junxiao Bi wrote:
>>>> On 12/30/2015 05:56 PM, xuejiufei wrote:
>>>>> Hi Junxiao,
>>>>> You are right. But it may happen that mle->woken is set to 1 in
>>>>> dlm_clean_migration_mle() just after atomic_read() in
>>>>> dlm_migrate_lockres(). Actually we trigger this BUG when dlm_send_one_lockres()
>>>>> return error.
>>>> Yes, that's possible because of that 5s timeout wakeup, I think this
>>>> timeout is useless and can be removed?
>>>>>
>>>>> And I think dlm_migrate_lockres() should not set owner to target and return 0
>>>>> when mle->woken is set to 1 in dlm_clean_migration_mle(). This is another problem?
>>>> Yes, it should be fixed, or the lockres owner will be set to a down node
>>>> wrongly. Can the following fix these two issue?
>>>>
>>> It can not fix the first issue when dlm_send_one_lockres() or
>>> dlm_mark_lockres_migrating() return error, right?
>> Right. This is for second issue. Any way please consider these two
>> issue, I think they can be fixed in one patch since both happened when
>> target node down during migrate. You can make your own one or merge this
>> patch to yours if you like.
>>
>> Thanks,
>> Junxiao.
>>>
>>>>
>>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>>> index 84f2f8079466..d0380ea62340 100644
>>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>>> @@ -2618,39 +2618,22 @@ fail:
>>>>
>>>>
>>>>  	/* wait for new node to assert master */
>>>> -	while (1) {
>>>> -		ret = wait_event_interruptible_timeout(mle->wq,
>>>> -					(atomic_read(&mle->woken) == 1),
>>>> -					msecs_to_jiffies(5000));
>>>> -
>>>> -		if (ret >= 0) {
>>>> -		       	if (atomic_read(&mle->woken) == 1 ||
>>>> -			    res->owner == target)
>>>> -				break;
>>>> -
>>>> -			mlog(0, "%s:%.*s: timed out during migration\n",
>>>> -			     dlm->name, res->lockname.len, res->lockname.name);
>>>> -			/* avoid hang during shutdown when migrating lockres
>>>> -			 * to a node which also goes down */
>>>> -			if (dlm_is_node_dead(dlm, target)) {
>>>> -				mlog(0, "%s:%.*s: expected migration "
>>>> -				     "target %u is no longer up, restarting\n",
>>>> -				     dlm->name, res->lockname.len,
>>>> -				     res->lockname.name, target);
>>>> -				ret = -EINVAL;
>>>> -				/* migration failed, detach and clean up mle */
>>>> -				dlm_mle_detach_hb_events(dlm, mle);
>>>> -				dlm_put_mle(mle);
>>>> -				dlm_put_mle_inuse(mle);
>>>> -				spin_lock(&res->spinlock);
>>>> -				res->state &= ~DLM_LOCK_RES_MIGRATING;
>>>> -				wake = 1;
>>>> -				spin_unlock(&res->spinlock);
>>>> -				goto leave;
>>>> -			}
>>>> -		} else
>>>> -			mlog(0, "%s:%.*s: caught signal during migration\n",
>>>> -			     dlm->name, res->lockname.len, res->lockname.name);
>>>> +	wait_event(mle->wq, ((atomic_read(&mle->woken) == 1) ||
>>>> +				(atomic_read(&mle->woken) == 2)));
>>>> +
>>>> +	/* migrate target down */
>>>> +	if (atomic_read(&mle->woken) == 2) {
>>>> +		mlog(0, "%s:%.*s: expected migration "
>>>> +				"target %u is no longer up, restarting\n",
>>>> +				dlm->name, res->lockname.len,
>>>> +				res->lockname.name, target);
>>>> +		ret = -EINVAL;
>>>> +		dlm_put_mle_inuse(mle);
>>>> +		spin_lock(&res->spinlock);
>>>> +		res->state &= ~DLM_LOCK_RES_MIGRATING;
>>>> +		wake = 1;
>>>> +		spin_unlock(&res->spinlock);
>>>> +		goto leave;
>>>>  	}
>>>>
>>>>  	/* all done, set the owner, clear the flag */
>>>> @@ -3227,7 +3210,7 @@ static void dlm_clean_migration_mle(struct
>>>> dlm_ctxt *dlm,
>>>>
>>>>  	spin_lock(&mle->spinlock);
>>>>  	__dlm_unlink_mle(dlm, mle);
>>>> -	atomic_set(&mle->woken, 1);
>>>> +	atomic_set(&mle->woken, 2);
>>>>  	spin_unlock(&mle->spinlock);
>>>>
>>>>  	wake_up(&mle->wq);
>>>>
>>>> Thanks,
>>>> Junxiao.
>>>>
>>>>
>>>>>
>>>>> Thanks
>>>>> Jiufei.
>>>>>
>>>>> On 2015/12/30 10:52, Junxiao Bi wrote:
>>>>>> Hi Jiufei,
>>>>>>
>>>>>> When target node down, mle is cleared from
>>>>>> dlm_do_local_recovery_cleanup()->dlm_clean_master_list()->dlm_clean_migration_mle()?
>>>>>> mle->woken is set to 1 in dlm_clean_migration_mle(), so the code to
>>>>>> detect target node down(if (dlm_is_node_dead(dlm, target))) will never
>>>>>> be run in dlm_migrate_lockres()?
>>>>>>
>>>>>>
>>>>>> 2621         while (1) {
>>>>>> 2622                 ret = wait_event_interruptible_timeout(mle->wq,
>>>>>> 2623                                         (atomic_read(&mle->woken)
>>>>>> == 1),
>>>>>> 2624                                         msecs_to_jiffies(5000));
>>>>>> 2625
>>>>>> 2626                 if (ret >= 0) {
>>>>>> 2627                         if (atomic_read(&mle->woken) == 1 ||
>>>>>> 2628                             res->owner == target)
>>>>>> 2629                                 break;
>>>>>> 2630
>>>>>> 2631                         mlog(0, "%s:%.*s: timed out during
>>>>>> migration\n",
>>>>>> 2632                              dlm->name, res->lockname.len,
>>>>>> res->lockname.name);
>>>>>> 2633                         /* avoid hang during shutdown when
>>>>>> migrating lockres
>>>>>> 2634                          * to a node which also goes down */
>>>>>> 2635                         if (dlm_is_node_dead(dlm, target)) {
>>>>>> 2636                                 mlog(0, "%s:%.*s: expected migration "
>>>>>> 2637                                      "target %u is no longer up,
>>>>>> restarting\n",
>>>>>> 2638                                      dlm->name, res->lockname.len,
>>>>>> 2639                                      res->lockname.name, target);
>>>>>> 2640                                 ret = -EINVAL;
>>>>>> 2641                                 /* migration failed, detach and
>>>>>> clean up mle */
>>>>>> 2642                                 dlm_mle_detach_hb_events(dlm, mle);
>>>>>> 2643                                 dlm_put_mle(mle);
>>>>>> 2644                                 dlm_put_mle_inuse(mle);
>>>>>> 2645                                 spin_lock(&res->spinlock);
>>>>>> 2646                                 res->state &= ~DLM_LOCK_RES_MIGRATING;
>>>>>> 2647                                 wake = 1;
>>>>>> 2648                                 spin_unlock(&res->spinlock);
>>>>>> 2649                                 goto leave;
>>>>>> 2650                         }
>>>>>> 2651                 } else
>>>>>> 2652                         mlog(0, "%s:%.*s: caught signal during
>>>>>> migration\n",
>>>>>> 2653                              dlm->name, res->lockname.len,
>>>>>> res->lockname.name);
>>>>>> 2654         }
>>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Junxiao.
>>>>>> On 12/28/2015 03:44 PM, xuejiufei wrote:
>>>>>>> We have found that migration source will trigger a BUG that the
>>>>>>> refcount of mle is already zero before put when the target is
>>>>>>> down during migration. The situation is as follows:
>>>>>>>
>>>>>>> dlm_migrate_lockres
>>>>>>>   dlm_add_migration_mle
>>>>>>>   dlm_mark_lockres_migrating
>>>>>>>   dlm_get_mle_inuse
>>>>>>>   <<<<<< Now the refcount of the mle is 2.
>>>>>>>   dlm_send_one_lockres and wait for the target to become the
>>>>>>>   new master.
>>>>>>>   <<<<<< o2hb detect the target down and clean the migration
>>>>>>>   mle. Now the refcount is 1.
>>>>>>>
>>>>>>> dlm_migrate_lockres woken, and put the mle twice when found
>>>>>>> the target goes down which trigger the BUG with the following
>>>>>>> message:
>>>>>>> "ERROR: bad mle: ".
>>>>>>>
>>>>>>> Signed-off-by: Jiufei Xue <xuejiufei@huawei.com>
>>>>>>> Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
>>>>>>> ---
>>>>>>>  fs/ocfs2/dlm/dlmmaster.c | 26 +++++++++++++++-----------
>>>>>>>  1 file changed, 15 insertions(+), 11 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
>>>>>>> index 936e11b..b713140 100644
>>>>>>> --- a/fs/ocfs2/dlm/dlmmaster.c
>>>>>>> +++ b/fs/ocfs2/dlm/dlmmaster.c
>>>>>>> @@ -2519,6 +2519,11 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
>>>>>>>  	spin_lock(&dlm->master_lock);
>>>>>>>  	ret = dlm_add_migration_mle(dlm, res, mle, &oldmle, name,
>>>>>>>  				    namelen, target, dlm->node_num);
>>>>>>> +	/* get an extra reference on the mle.
>>>>>>> +	 * otherwise the assert_master from the new
>>>>>>> +	 * master will destroy this.
>>>>>>> +	 */
>>>>>>> +	dlm_get_mle_inuse(mle);
>>>>>>>  	spin_unlock(&dlm->master_lock);
>>>>>>>  	spin_unlock(&dlm->spinlock);
>>>>>>>
>>>>>>> @@ -2554,6 +2559,7 @@ fail:
>>>>>>>  		if (mle_added) {
>>>>>>>  			dlm_mle_detach_hb_events(dlm, mle);
>>>>>>>  			dlm_put_mle(mle);
>>>>>>> +			dlm_put_mle_inuse(mle);
>>>>>>>  		} else if (mle) {
>>>>>>>  			kmem_cache_free(dlm_mle_cache, mle);
>>>>>>>  			mle = NULL;
>>>>>>> @@ -2571,17 +2577,6 @@ fail:
>>>>>>>  	 * ensure that all assert_master work is flushed. */
>>>>>>>  	flush_workqueue(dlm->dlm_worker);
>>>>>>>
>>>>>>> -	/* get an extra reference on the mle.
>>>>>>> -	 * otherwise the assert_master from the new
>>>>>>> -	 * master will destroy this.
>>>>>>> -	 * also, make sure that all callers of dlm_get_mle
>>>>>>> -	 * take both dlm->spinlock and dlm->master_lock */
>>>>>>> -	spin_lock(&dlm->spinlock);
>>>>>>> -	spin_lock(&dlm->master_lock);
>>>>>>> -	dlm_get_mle_inuse(mle);
>>>>>>> -	spin_unlock(&dlm->master_lock);
>>>>>>> -	spin_unlock(&dlm->spinlock);
>>>>>>> -
>>>>>>>  	/* notify new node and send all lock state */
>>>>>>>  	/* call send_one_lockres with migration flag.
>>>>>>>  	 * this serves as notice to the target node that a
>>>>>>> @@ -3312,6 +3307,15 @@ top:
>>>>>>>  			    mle->new_master != dead_node)
>>>>>>>  				continue;
>>>>>>>
>>>>>>> +			if (mle->new_master == dead_node && mle->inuse) {
>>>>>>> +				mlog(ML_NOTICE, "%s: target %u died during "
>>>>>>> +						"migration from %u, the MLE is "
>>>>>>> +						"still keep used, ignore it!\n",
>>>>>>> +						dlm->name, dead_node,
>>>>>>> +						mle->master);
>>>>>>> +				continue;
>>>>>>> +			}
>>>>>>> +
>>>>>>>  			/* If we have reached this point, this mle needs to be
>>>>>>>  			 * removed from the list and freed. */
>>>>>>>  			dlm_clean_migration_mle(dlm, mle);
>>>>>>>
>>>>>>
>>>>>>
>>>>>> .
>>>>>>
>>>>>
>>>>
>>>>
>>>> .
>>>>
>>>
>>
>>
>> .
>>
>

Patch
diff mbox

diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
index 84f2f8079466..d0380ea62340 100644
--- a/fs/ocfs2/dlm/dlmmaster.c
+++ b/fs/ocfs2/dlm/dlmmaster.c
@@ -2618,39 +2618,22 @@  fail:


 	/* wait for new node to assert master */
-	while (1) {
-		ret = wait_event_interruptible_timeout(mle->wq,
-					(atomic_read(&mle->woken) == 1),
-					msecs_to_jiffies(5000));
-
-		if (ret >= 0) {
-		       	if (atomic_read(&mle->woken) == 1 ||
-			    res->owner == target)
-				break;
-
-			mlog(0, "%s:%.*s: timed out during migration\n",
-			     dlm->name, res->lockname.len, res->lockname.name);
-			/* avoid hang during shutdown when migrating lockres
-			 * to a node which also goes down */
-			if (dlm_is_node_dead(dlm, target)) {
-				mlog(0, "%s:%.*s: expected migration "
-				     "target %u is no longer up, restarting\n",
-				     dlm->name, res->lockname.len,
-				     res->lockname.name, target);
-				ret = -EINVAL;
-				/* migration failed, detach and clean up mle */
-				dlm_mle_detach_hb_events(dlm, mle);
-				dlm_put_mle(mle);
-				dlm_put_mle_inuse(mle);
-				spin_lock(&res->spinlock);
-				res->state &= ~DLM_LOCK_RES_MIGRATING;
-				wake = 1;
-				spin_unlock(&res->spinlock);
-				goto leave;
-			}
-		} else
-			mlog(0, "%s:%.*s: caught signal during migration\n",
-			     dlm->name, res->lockname.len, res->lockname.name);
+	wait_event(mle->wq, ((atomic_read(&mle->woken) == 1) ||
+				(atomic_read(&mle->woken) == 2)));
+
+	/* migrate target down */
+	if (atomic_read(&mle->woken) == 2) {
+		mlog(0, "%s:%.*s: expected migration "
+				"target %u is no longer up, restarting\n",
+				dlm->name, res->lockname.len,
+				res->lockname.name, target);
+		ret = -EINVAL;
+		dlm_put_mle_inuse(mle);
+		spin_lock(&res->spinlock);
+		res->state &= ~DLM_LOCK_RES_MIGRATING;
+		wake = 1;
+		spin_unlock(&res->spinlock);
+		goto leave;
 	}

 	/* all done, set the owner, clear the flag */
@@ -3227,7 +3210,7 @@  static void dlm_clean_migration_mle(struct
dlm_ctxt *dlm,

 	spin_lock(&mle->spinlock);
 	__dlm_unlink_mle(dlm, mle);
-	atomic_set(&mle->woken, 1);
+	atomic_set(&mle->woken, 2);
 	spin_unlock(&mle->spinlock);