Skip to contentSkip to conversation

ocfs2/dlm: resend deref to new master if recovery occures

PATCH New Srinivas Eeda 2 participants 5 messages May 24, 2010 – Jun 3, 2010 OCFS2

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ocfs2/dlm: resend deref to new master if recovery occures

thanks for doing this patch. I have a little comment, wondering if there could be a window where node B sent the lock info to node C as part of recovery and removed flag DLM_LOCK_RES_RECOVERING while dlm_thread was still purging it. In that case dlm_thread will still continue to remove it from hash list.

Also, this patch puts dlm_thread to sleep ... may be it's ok, but wondering if we can avoid that.

delay deref message if DLM_LOCK_RES_RECOVERING is set (which means recovery got to the lockres before dlm_thread could), move the lockres to the end of the purgelist and retry later.

do not inform recovery master if DLM_LOCK_RES_DROPPING_REF is set (which means dlm_thread got to the lockres before recovery). So in the case you described, node C will not know about node B dropping the dereference and node B will just go ahead and remove it from hash list and free it.

Quoted 110 lines · Wengang Wang

When purge a lockres, we unhash the lockres ignore the result of deref request and ignore the lockres state.
There is a problem that rarely happen. It can happen when recovery take places. Say node A is the master of the lockres with node B wants to deref and there is a node C. If things happen in the following order, the bug is triggered.

1) node B send DEREF to node A for lockres A and waiting for result.
2) node A crashed, node C become the recovery master.
3) node C mastered lockres A with node B has a ref on it.
4) node B goes to unhashes the lockres A with a ref on node C.

After step 4), if a umount comes on node C, it will hang at migrate lockres A since node B has a ref on it.

The fix is that we check if recovery happened on lockres A after sending DEREF request. If that happened, we keep lockres A in hash and in purge list for another try to send DEREF to the new master(node C). So that node C can clear the incorrect refbit.

---
 fs/ocfs2/dlm/dlmthread.c |   44 +++++++++++++++++++++++++++++++++++++++-----
 1 files changed, 39 insertions(+), 5 deletions(-)
fs/ocfs2/dlm/dlmthread.c +39 -5
index d4f73ca..946ff1a 100644
@@ -158,6 +158,8 @@ static int dlm_purge_lockres(struct dlm_ctxt *dlm,
 	int master;
 	int ret = 0;
 
+	assert_spin_locked(&dlm->spinlock);
+
 	spin_lock(&res->spinlock);
 	if (!__dlm_lockres_unused(res)) {
 		mlog(0, "%s:%.*s: tried to purge but not unused\n",
@@ -211,18 +213,30 @@ static int dlm_purge_lockres(struct dlm_ctxt *dlm,
 	}
 
 	spin_lock(&res->spinlock);
+	/*
+	 * we dropped dlm->spinlock and res->spinlock when sending the DEREF
+	 * request, there is a chance that a recovery happend on this lockres.
+	 * in that case, we have to DEREF to the new master when recovery
+	 * finished. otherwise, there can be an incorrect ref on the lockres
+	 * on the new master  on behalf of this node.
+	 */
+	if (unlikely(res->state & DLM_LOCK_RES_RECOVERING)) {
+		spin_unlock(&res->spinlock);
+		return -EAGAIN;
+	}
+
 	if (!list_empty(&res->purge)) {
 		mlog(0, "removing lockres %.*s:%p from purgelist, "
 		     "master = %d\n", res->lockname.len, res->lockname.name,
 		     res, master);
 		list_del_init(&res->purge);
-		spin_unlock(&res->spinlock);
+		/* not the last ref */
 		dlm_lockres_put(res);
 		dlm->purge_count--;
-	} else
-		spin_unlock(&res->spinlock);
+	}
 
 	__dlm_unhash_lockres(res);
+	spin_unlock(&res->spinlock);
 
 	/* lockres is not in the hash now.  drop the flag and wake up
 	 * any processes waiting in dlm_get_lock_resource. */
@@ -241,6 +255,9 @@ static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 	unsigned int run_max, unused;
 	unsigned long purge_jiffies;
 	struct dlm_lock_resource *lockres;
+	int ret;
+
+#define DLM_WAIT_RECOVERY_FINISH_MS 500
 
 	spin_lock(&dlm->spinlock);
 	run_max = dlm->purge_count;
@@ -280,8 +297,25 @@ static void dlm_run_purge_list(struct dlm_ctxt *dlm,
 
 		/* This may drop and reacquire the dlm spinlock if it
 		 * has to do migration. */
-		if (dlm_purge_lockres(dlm, lockres))
-			BUG();
+		ret = dlm_purge_lockres(dlm, lockres);
+		if (ret) {
+			if (ret == -EAGAIN) {
+				/*
+				 * recovery occured on this lockres. try to
+				 * DEREF to the new master.
+				 */
+				dlm_lockres_put(lockres);
+				spin_unlock(&dlm->spinlock);
+				mlog(ML_ERROR, "retry to purge %.*s after %d"
+				     "ms\n", lockres->lockname.len,
+				     lockres->lockname.name,
+				     DLM_WAIT_RECOVERY_FINISH_MS);
+				msleep(DLM_WAIT_RECOVERY_FINISH_MS);
+				spin_lock(&dlm->spinlock);
+				continue;
+			} else
+				BUG();
+		}
 
 		dlm_lockres_put(lockres);
 
#1 lore raw

PATCH 1/1 ocfs2/dlm: resend deref to new master if recovery occures

Quoted 5 lines · Srinivas Eeda

thanks for doing this patch. I have a little comment, wondering if there could be a window where node B sent the lock info to node C as part of recovery and removed flag DLM_LOCK_RES_RECOVERING while dlm_thread was still purging it. In that case dlm_thread will still continue to remove it from hash list.

Yes, you are right. There do is such a window. I missed that.

Quoted 3 lines

Also, this patch puts dlm_thread to sleep ... may be it's ok, but wondering if we can avoid that.

Yes. I considered about that too but failed at finding a simple way to avoid that.

Quoted 3 lines

delay deref message if DLM_LOCK_RES_RECOVERING is set (which means recovery got to the lockres before dlm_thread could), move the lockres to the end of the purgelist and retry later.

Good point! I meant that but the patch deosn't prove that :P.

Quoted 5 lines

do not inform recovery master if DLM_LOCK_RES_DROPPING_REF is set (which means dlm_thread got to the lockres before recovery). So in the case you described, node C will not know about node B dropping the dereference and node B will just go ahead and remove it from hash list and free it.

Cool idea! let me try and re-create the patch!

thanks much Srini.
wengang.

Quoted 19 lines
Quoted 18 lines · Wengang Wang

When purge a lockres, we unhash the lockres ignore the result of deref request and ignore the lockres state.
There is a problem that rarely happen. It can happen when recovery take places. Say node A is the master of the lockres with node B wants to deref and there is a node C. If things happen in the following order, the bug is triggered.

1) node B send DEREF to node A for lockres A and waiting for result.
2) node A crashed, node C become the recovery master.
3) node C mastered lockres A with node B has a ref on it.
4) node B goes to unhashes the lockres A with a ref on node C. After step 4), if a umount comes on node C, it will hang at
migrate lockres A since node B has a ref on it.

The fix is that we check if recovery happened on lockres A after sending DEREF request. If that happened, we keep lockres A in hash and in purge list for another try to send DEREF to the new master(node C). So that node C can clear the incorrect refbit.

#2 lore raw

PATCH 1/1 ocfs2/dlm: resend deref to new master if recovery occures

Quoted 3 lines
Quoted 3 lines

delay deref message if DLM_LOCK_RES_RECOVERING is set (which means recovery got to the lockres before dlm_thread could), move the lockres to the end of the purgelist and retry later.

If you meant checking before sending DEREF, it could cause a high cpu useage when there are only the ones in recovery left in the purge list until they are recovered. --Is that acceptable?

regards,
wengang.

#3 lore raw

PATCH 1/1 ocfs2/dlm: resend deref to new master if recovery occures

Quoted 9 lines · Wengang Wang
Quoted 4 lines
Quoted 4 lines

delay deref message if DLM_LOCK_RES_RECOVERING is set (which means recovery got to the lockres before dlm_thread could), move the lockres to the end of the purgelist and retry later.

If you meant checking before sending DEREF, it could cause a high cpu useage when there are only the ones in recovery left in the purge list until they are recovered. --Is that acceptable?

Yea, you are right about the cpu usage. I think the right thing would be to exit the loop after all locks are traversed once.

Quoted 3 lines

regards,
wengang.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20100524/d283fc49/attachment.html

#4 lore raw

PATCH 1/1 ocfs2/dlm: cancel the migration or redo deref to recovery master

Changes to V1:
1 move the msleep to the second runs when the lockres is in recovery so the
purging work on other lockres' can go.
2 do not inform recovery master if DLM_LOCK_RES_DROPPING_REF is set and don't
resend deref in this case.

---
 fs/ocfs2/dlm/dlmcommon.h   |    1 +
 fs/ocfs2/dlm/dlmrecovery.c |   25 +++++++++++++++
 fs/ocfs2/dlm/dlmthread.c   |   73 ++++++++++++++++++++++++++++++++++++++-----
 3 files changed, 90 insertions(+), 9 deletions(-)
fs/ocfs2/dlm/dlmcommon.h +1 -0
index 4b6ae2c..4194087 100644
@@ -280,6 +280,7 @@ static inline void __dlm_set_joining_node(struct dlm_ctxt *dlm,
 #define DLM_LOCK_RES_IN_PROGRESS          0x00000010
 #define DLM_LOCK_RES_MIGRATING            0x00000020
 #define DLM_LOCK_RES_DROPPING_REF         0x00000040
+#define DLM_LOCK_RES_DE_DROP_REF          0x00000080
 #define DLM_LOCK_RES_BLOCK_DIRTY          0x00001000
 #define DLM_LOCK_RES_SETREF_INPROG        0x00002000
 
fs/ocfs2/dlm/dlmrecovery.c +25 -0
index f8b75ce..7241070 100644
@@ -913,6 +913,27 @@ static void dlm_request_all_locks_worker(struct dlm_work_item *item, void *data)
 	/* any errors returned will be due to the new_master dying,
 	 * the dlm_reco_thread should detect this */
 	list_for_each_entry(res, &resources, recovering) {
+		int ignore_mig = 0;
+		spin_lock(&res->spinlock);
+		/*
+		 * if we are dropping the lockres, no need to let the new master
+		 * know the reference of this node. That is don't migrate the
+		 * lockres to the new master. Also make sure we don't send DEREF
+		 * request for the same lockres to the new master either.
+		 */
+		if (unlikely(res->state & DLM_LOCK_RES_DROPPING_REF)) {
+			ignore_mig = 1;
+			res->state |= DLM_LOCK_RES_DE_DROP_REF;
+		}
+		spin_unlock(&res->spinlock);
+		if (ignore_mig) {
+			mlog(ML_NOTICE, "ignore migrating %.*s to recovery "
+			     "master %u as we are dropping it\n",
+			     res->lockname.len, res->lockname.name,
+			     reco_master);
+			continue;
+		}
+
 		ret = dlm_send_one_lockres(dlm, res, mres, reco_master,
 				   	DLM_MRES_RECOVERY);
 		if (ret < 0) {
@@ -1997,7 +2018,11 @@ void dlm_move_lockres_to_recovery_list(struct dlm_ctxt *dlm,
 	struct list_head *queue;
 	struct dlm_lock *lock, *next;
 
+	assert_spin_locked(&res->spinlock);
+
 	res->state |= DLM_LOCK_RES_RECOVERING;
+	res->state &= ~DLM_LOCK_RES_DE_DROP_REF;
+
 	if (!list_empty(&res->recovering)) {
 		mlog(0,
 		     "Recovering res %s:%.*s, is already on recovery list!\n",
fs/ocfs2/dlm/dlmthread.c +64 -9

124 lines not on this page · raw message

Signature · 1 line
1.6.6.1