Skip to contentSkip to conversation

ocfs2/dlm: queue ast in lockres->spinlock

PATCH New Wengang Wang 2 participants 6 messages Nov 30, 2010 – Dec 1, 2010 OCFS2

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

There is bug hitting the BUG() in ocfs2_prepare_downconvert(), in 1.4 code though.

static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
int new_level)

{               
        assert_spin_locked(&lockres->l_lock);
                
        BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
                
        if (lockres->l_level <= new_level) {
                mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
                     "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
                     "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
                     new_level, list_empty(&lockres->l_blocked_list),
                     list_empty(&lockres->l_mask_waiters), lockres->l_type,
                     lockres->l_flags, lockres->l_ro_holders,
                     lockres->l_ex_holders, lockres->l_action,
                     lockres->l_unlock_action, lockres->l_requested,
                     lockres->l_blocking, lockres->l_pending_gen);
                BUG();  
        } 

....

By checking vmcore, lockres->l_level == new_level == 0.

Though no tcpdump shows that, I think it's caused by a BAST before AST. It can happen in the following case(steps in time order):

1) Node A send a "up convert" request(for a EX) to the master node, node B.

2) Node C send a "create lock" request(for a PR) to the master node, node B against the same lockres.

3) The "convert request" was added to dlm_thread because the lock can't be granted immediately. So that the "create lock" thread and the "up convert" thread can run in parallel.

4) the "create lock" thread get the lockres->spinlock and found no conflict for the lock request, so it put the lock to granted list. then lockres->spinlock is released. (see dlmlock_master()).

5) the dlm_thread get the spinlcok and found the conflict lock, which was granted in 4), and queue a BAST that will send to node C. (see dlm_shuffle_lists()).

6) the "create lock" thread then continue to queue the AST for the granted lock to node C. So the BAST and AST come to Node C in order.

7) node C crashes when responding the BAST.

This patch tries to fix it by moving the queue_ast into protection of lockres->spinlock.

This patch is based on
http://oss.oracle.com/pipermail/ocfs2-devel/2010-November/007523.html
Otherwise, it could cause deadlock.(by different order of lockres->spinlock and dlm->ast_lock)

---
 fs/ocfs2/dlm/dlmconvert.c |   14 ++++++--------
 fs/ocfs2/dlm/dlmlock.c    |    6 ++----
 2 files changed, 8 insertions(+), 12 deletions(-)
fs/ocfs2/dlm/dlmconvert.c +6 -8
index 9f30491..ff0b84f 100644
@@ -90,15 +90,14 @@ enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
 				     &call_ast, &kick_thread);
 
 	res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
+	if (call_ast)
+		dlm_queue_ast(dlm, lock);
 	spin_unlock(&res->spinlock);
 	wake_up(&res->wq);
 	if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
 		dlm_error(status);
 
-	/* either queue the ast or release it */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else
+	if (!call_ast)
 		dlm_lockres_release_ast(dlm, res);
 
 	if (kick_thread)
@@ -514,6 +513,8 @@ int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
 					     cnv->requested_type,
 					     &call_ast, &kick_thread);
 		res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
+		if (call_ast)
+			dlm_queue_ast(dlm, lock);
 		wake = 1;
 	}
 	spin_unlock(&res->spinlock);
@@ -530,10 +531,7 @@ leave:
 	if (lock)
 		dlm_lock_put(lock);
 
-	/* either queue the ast or release it, if reserved */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else if (ast_reserved)
+	if ((!call_ast) && ast_reserved)
 		dlm_lockres_release_ast(dlm, res);
 
 	if (kick_thread)
fs/ocfs2/dlm/dlmlock.c +2 -4
index 69cf369..f84d1fb 100644
@@ -157,6 +157,7 @@ static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
 		if (!dlm_is_recovery_lock(res->lockname.name,
 					  res->lockname.len)) {
 			kick_thread = 1;
+			dlm_queue_ast(dlm, lock);
 			call_ast = 1;
 		} else {
 			mlog(0, "%s: returning DLM_NORMAL to "
@@ -188,10 +189,7 @@ static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
 	spin_unlock(&res->spinlock);
 	wake_up(&res->wq);
 
-	/* either queue the ast or release it */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else
+	if (!call_ast)
 		dlm_lockres_release_ast(dlm, res);
 
 	dlm_lockres_calc_usage(dlm, res);
Signature · 1 line
1.7.2.3
#1 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

http://oss.oracle.com/pipermail/ocfs2-devel/2010-January/005801.html

The above bug/fix fits better.

Having said that, I am not discounting your point. But I will need more to analyse it. Questions inlined.

Quoted 42 lines · Wengang Wang

There is bug hitting the BUG() in ocfs2_prepare_downconvert(), in 1.4 code though.

static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
int new_level)

{
         assert_spin_locked(&lockres->l_lock);

         BUG_ON(lockres->l_blocking<= DLM_LOCK_NL);

         if (lockres->l_level<= new_level) {
                 mlog(ML_ERROR, "lockres %s, lvl %d<= %d, blcklst %d, mask %d, "
                      "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
                      "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
                      new_level, list_empty(&lockres->l_blocked_list),
                      list_empty(&lockres->l_mask_waiters), lockres->l_type,
                      lockres->l_flags, lockres->l_ro_holders,
                      lockres->l_ex_holders, lockres->l_action,
                      lockres->l_unlock_action, lockres->l_requested,
                      lockres->l_blocking, lockres->l_pending_gen);
                 BUG();
         }

....

By checking vmcore, lockres->l_level == new_level == 0.

Though no tcpdump shows that, I think it's caused by a BAST before AST. It can happen in the following case(steps in time order):

1) Node A send a "up convert" request(for a EX) to the master node, node B.

2) Node C send a "create lock" request(for a PR) to the master node, node B against the same lockres.

3) The "convert request" was added to dlm_thread because the lock can't be granted immediately. So that the "create lock" thread and the "up convert" thread can run in parallel.

4) the "create lock" thread get the lockres->spinlock and found no conflict for the lock request, so it put the lock to granted list. then lockres->spinlock is released. (see dlmlock_master()).

Quoted 8 lines

5) the dlm_thread get the spinlcok and found the conflict lock, which was granted in 4), and queue a BAST that will send to node C. (see dlm_shuffle_lists()).

6) the "create lock" thread then continue to queue the AST for the granted lock to node C. So the BAST and AST come to Node C in order.

7) node C crashes when responding the BAST.

So in your scenario, the upconvert request is before create lock. If that is so, then even if the new level cannot be granted, the lock will be moved to the convert queue. The create lock request is always added to the blocked queue which always has the lowest priority.

Note that the dlm always processes convert queues before blocked queues. Also, o2dlm, always processes all ASTs before BASTs.

That is not to say what you have said cannot happen. But I will need more information. Like, Node A was up converting to EX from what level? Were there any other locks on the resource?

Quoted 83 lines

This patch tries to fix it by moving the queue_ast into protection of lockres->spinlock.

This patch is based on
http://oss.oracle.com/pipermail/ocfs2-devel/2010-November/007523.html
Otherwise, it could cause deadlock.(by different order of lockres->spinlock and dlm->ast_lock)

---
  fs/ocfs2/dlm/dlmconvert.c |   14 ++++++--------
  fs/ocfs2/dlm/dlmlock.c    |    6 ++----
  2 files changed, 8 insertions(+), 12 deletions(-)
fs/ocfs2/dlm/dlmconvert.c +6 -8
index 9f30491..ff0b84f 100644
@@ -90,15 +90,14 @@ enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
  				&call_ast,&kick_thread);
 
  	res->state&= ~DLM_LOCK_RES_IN_PROGRESS;
+	if (call_ast)
+		dlm_queue_ast(dlm, lock);
  	spin_unlock(&res->spinlock);
  	wake_up(&res->wq);
  	if (status != DLM_NORMAL&&  status != DLM_NOTQUEUED)
  		dlm_error(status);
 
-	/* either queue the ast or release it */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else
+	if (!call_ast)
  		dlm_lockres_release_ast(dlm, res);
 
  	if (kick_thread)
@@ -514,6 +513,8 @@ int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
  					     cnv->requested_type,
  					&call_ast,&kick_thread);
  		res->state&= ~DLM_LOCK_RES_IN_PROGRESS;
+		if (call_ast)
+			dlm_queue_ast(dlm, lock);
  		wake = 1;
  	}
  	spin_unlock(&res->spinlock);
@@ -530,10 +531,7 @@ leave:
  	if (lock)
  		dlm_lock_put(lock);
 
-	/* either queue the ast or release it, if reserved */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else if (ast_reserved)
+	if ((!call_ast)&&  ast_reserved)
  		dlm_lockres_release_ast(dlm, res);
 
  	if (kick_thread)
fs/ocfs2/dlm/dlmlock.c +2 -4
index 69cf369..f84d1fb 100644
@@ -157,6 +157,7 @@ static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
  		if (!dlm_is_recovery_lock(res->lockname.name,
  					  res->lockname.len)) {
  			kick_thread = 1;
+			dlm_queue_ast(dlm, lock);
  			call_ast = 1;
  		} else {
  			mlog(0, "%s: returning DLM_NORMAL to "
@@ -188,10 +189,7 @@ static enum dlm_status dlmlock_master(struct dlm_ctxt *dlm,
  	spin_unlock(&res->spinlock);
  	wake_up(&res->wq);
 
-	/* either queue the ast or release it */
-	if (call_ast)
-		dlm_queue_ast(dlm, lock);
-	else
+	if (!call_ast)
  		dlm_lockres_release_ast(dlm, res);
 
  	dlm_lockres_calc_usage(dlm, res);
#2 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

Quoted 64 lines · Sunil Mushran

http://oss.oracle.com/pipermail/ocfs2-devel/2010-January/005801.html

The above bug/fix fits better.

Having said that, I am not discounting your point. But I will need more to analyse it. Questions inlined.

Quoted 42 lines · Wengang Wang

There is bug hitting the BUG() in ocfs2_prepare_downconvert(), in 1.4 code though.

static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
int new_level)

{
        assert_spin_locked(&lockres->l_lock);

        BUG_ON(lockres->l_blocking<= DLM_LOCK_NL);

        if (lockres->l_level<= new_level) {
                mlog(ML_ERROR, "lockres %s, lvl %d<= %d, blcklst %d, mask %d, "
                     "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
                     "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
                     new_level, list_empty(&lockres->l_blocked_list),
                     list_empty(&lockres->l_mask_waiters), lockres->l_type,
                     lockres->l_flags, lockres->l_ro_holders,
                     lockres->l_ex_holders, lockres->l_action,
                     lockres->l_unlock_action, lockres->l_requested,
                     lockres->l_blocking, lockres->l_pending_gen);
                BUG();
        }

....

By checking vmcore, lockres->l_level == new_level == 0.

Though no tcpdump shows that, I think it's caused by a BAST before AST. It can happen in the following case(steps in time order):

1) Node A send a "up convert" request(for a EX) to the master node, node B.

2) Node C send a "create lock" request(for a PR) to the master node, node B against the same lockres.

3) The "convert request" was added to dlm_thread because the lock can't be granted immediately. So that the "create lock" thread and the "up convert" thread can run in parallel.

4) the "create lock" thread get the lockres->spinlock and found no conflict for the lock request, so it put the lock to granted list. then lockres->spinlock is released. (see dlmlock_master()).

Quoted 8 lines

5) the dlm_thread get the spinlcok and found the conflict lock, which was granted in 4), and queue a BAST that will send to node C. (see dlm_shuffle_lists()).

6) the "create lock" thread then continue to queue the AST for the granted lock to node C. So the BAST and AST come to Node C in order.

7) node C crashes when responding the BAST.

So in your scenario, the upconvert request is before create lock. If that is so, then even if the new level cannot be granted, the lock will be moved to the convert queue. The create lock request is always added to the blocked queue which always has the lowest priority.

Quoted 7 lines

Note that the dlm always processes convert queues before blocked queues. Also, o2dlm, always processes all ASTs before BASTs.

That is not to say what you have said cannot happen. But I will need more information. Like, Node A was up converting to EX from what level? Were there any other locks on the resource?

First I have to say that the scenario is in my thought only, no tcpdump shows that. Only the node C wanting a PR is the truth from the vmcore.

Let's consider this situation:

The create lock(CL) is for a PR;
the up convert(UC) is PR->EX;
there is an existing PR granted for another node(say node D) before the UC comes

The UC comes. Because of the existing PR, the UC lock is moved to converting list. dlm_thread takes care it later.
The CL comes. no conflict found(see dlm_can_grant_new_lock, convert_type not checked there). so it's not added to blocked list.

Seems the above scenario can trigger a BAST before AST.

regards,
wengang.

#3 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

Quoted 9 lines · Wengang Wang

First I have to say that the scenario is in my thought only, no tcpdump shows that. Only the node C wanting a PR is the truth from the vmcore.

Let's consider this situation:

The create lock(CL) is for a PR;
the up convert(UC) is PR->EX;
there is an existing PR granted for another node(say node D) before the UC comes

Quoted 6 lines

The UC comes. Because of the existing PR, the UC lock is moved to converting list. dlm_thread takes care it later.
The CL comes. no conflict found(see dlm_can_grant_new_lock, convert_type not checked there). so it's not added to blocked list.

Seems the above scenario can trigger a BAST before AST.

dlm_can_grant_new_lock() should be looking at convert_type also and add the lock to the blocked list in this case.

This is not just to plug this hole but is as per the spec.

#4 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

Quoted 15 lines · Sunil Mushran
Quoted 14 lines · Wengang Wang

First I have to say that the scenario is in my thought only, no tcpdump shows that. Only the node C wanting a PR is the truth from the vmcore.

Let's consider this situation:

The create lock(CL) is for a PR;
the up convert(UC) is PR->EX;
there is an existing PR granted for another node(say node D) before the UC comes

The UC comes. Because of the existing PR, the UC lock is moved to converting list. dlm_thread takes care it later.
The CL comes. no conflict found(see dlm_can_grant_new_lock, convert_type not checked there). so it's not added to blocked list.

Quoted 6 lines
Quoted 1 line

Seems the above scenario can trigger a BAST before AST.

dlm_can_grant_new_lock() should be looking at convert_type also and add the lock to the blocked list in this case.

This is not just to plug this hole but is as per the spec.

So you meant o2dlm is (should be) implementing fair locking? Anywhere I can take a look at the spec?

regards,
wengang.

#5 lore raw

PATCH 1/1 ocfs2/dlm: queue ast in lockres->spinlock

Quoted 16 lines · Wengang Wang
Quoted 15 lines · Sunil Mushran
Quoted 14 lines · Wengang Wang

First I have to say that the scenario is in my thought only, no tcpdump shows that. Only the node C wanting a PR is the truth from the vmcore.

Let's consider this situation:

The create lock(CL) is for a PR;
the up convert(UC) is PR->EX;
there is an existing PR granted for another node(say node D) before the UC comes

The UC comes. Because of the existing PR, the UC lock is moved to converting list. dlm_thread takes care it later.
The CL comes. no conflict found(see dlm_can_grant_new_lock, convert_type not checked there). so it's not added to blocked list.

Quoted 9 lines
Quoted 6 lines
Quoted 1 line

Seems the above scenario can trigger a BAST before AST.

dlm_can_grant_new_lock() should be looking at convert_type also and add the lock to the blocked list in this case.

This is not just to plug this hole but is as per the spec.

So you meant o2dlm is (should be) implementing fair locking? Anywhere I can take a look at the spec?

Google for Programming Locking Applications by Kristin Thomas. Or, grep the same in the source. The link is there.