Skip to contentSkip to conversation

OCFS2: don't leave free'd mle attached to hb events

PATCH New wengang wang 2 participants 5 messages Dec 7, 2009 – Feb 2, 2010 OCFS2

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 OCFS2: don't leave free'd mle attached to hb events

don't leave free'd mle attached to hb events.

in dlm_add_migration_mle() the mle is attched to "heartbeat events" anyway no matter there is an existing mle with same name(returns -EEXIST). dlm_migrate_lockres() calls dlm_add_migration_mle(). in case the later function returning -EEXIST, dlm_migrate_lockres() frees the (new) mle without detaching it from "hb events". so that later "hb events" related operations could improperly operate against wrong mle objects or against an invalid memory address.

the patch fixes above problem. it marks the mle as "added" just after dlm_add_migration_mle() is called.

fs/ocfs2/dlm/dlmmaster.c +2 -1
index 83bcaf2..0df80e9 100644
@@ -2498,11 +2498,12 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
 	spin_unlock(&dlm->master_lock);
 	spin_unlock(&dlm->spinlock);
 
+	mle_added = 1;
+
 	if (ret == -EEXIST) {
 		mlog(0, "another process is already migrating it\n");
 		goto fail;
 	}
-	mle_added = 1;
 
 	/*
 	 * set the MIGRATING flag and flush asts
#1 lore raw

PATCH 1/1 OCFS2: don't leave free'd mle attached to hb events

NAK

Quoted 8 lines · wengang wang

don't leave free'd mle attached to hb events.

in dlm_add_migration_mle() the mle is attched to "heartbeat events" anyway no matter there is an existing mle with same name(returns -EEXIST). dlm_migrate_lockres() calls dlm_add_migration_mle(). in case the later function returning -EEXIST, dlm_migrate_lockres() frees the (new) mle without detaching it from "hb events". so that later "hb events" related operations could improperly operate against wrong mle objects or against an invalid memory address.

The mle is attached to hb events in dlm_init_mle() which is not called if it returns -EEXIST. When it returns -EEXIST, oldmle is set to the existing mle and its refcounting is handled correctly. mle is not touched and thus only needs to be freed.

Quoted 24 lines

the patch fixes above problem. it marks the mle as "added" just after dlm_add_migration_mle() is called.

fs/ocfs2/dlm/dlmmaster.c +2 -1
index 83bcaf2..0df80e9 100644
@@ -2498,11 +2498,12 @@ static int dlm_migrate_lockres(struct dlm_ctxt *dlm,
 	spin_unlock(&dlm->master_lock);
 	spin_unlock(&dlm->spinlock);
 
+	mle_added = 1;
+
 	if (ret == -EEXIST) {
 		mlog(0, "another process is already migrating it\n");
 		goto fail;
 	}
-	mle_added = 1;
 
 	/*
 	 * set the MIGRATING flag and flush asts
#2 lore raw

PATCH 1/1 OCFS2: don't leave free'd mle attached to hb events

Hi Sunil,

Quoted 15 lines · Sunil Mushran

NAK

Quoted 11 lines · wengang wang

don't leave free'd mle attached to hb events.
in dlm_add_migration_mle() the mle is attched to "heartbeat
events" anyway no
matter there is an existing mle with same name(returns -EEXIST). dlm_migrate_lockres() calls dlm_add_migration_mle(). in case the later function
returning -EEXIST, dlm_migrate_lockres() frees the (new) mle without detaching
it from "hb events". so that later "hb events" related operations could improperly
operate against wrong mle objects or against an invalid memory address.

Quoted 5 lines

The mle is attached to hb events in dlm_init_mle() which is not called if it returns -EEXIST. When it returns -EEXIST, oldmle is set to the existing mle and its refcounting is handled correctly. mle is not touched and thus only needs to be freed.

Maybe I am wrong. but are you sure dlm_init_mle() is not called when it returns -EEXIST?
in code, it doesn't return immediately after setting ret with -EEXIST, but continue to call
dlm_init_mle();
and then
mle->new_master = new_master;

Maybe I am wrong somewhere?
simplified code pasted here:

3098 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,

3099                                  struct dlm_lock_resource *res,
3100                                  struct dlm_master_list_entry *mle,
3101                                  struct dlm_master_list_entry **oldmle,
3102                                  const char *name, unsigned int 

namelen,
3103 u8 new_master, u8 master)
3104 {
3116 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3117 if (found) {
3118 struct dlm_master_list_entry *tmp = *oldmle;
3120 if (tmp->type == DLM_MLE_MIGRATION) {
3121 if (master == dlm->node_num) {
3126 ret = -EEXIST;
3127 } else {
3136 BUG();
3137 }

3138                 } else {
			     ....

3151 }
3153 }
3154
3156 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);

3157         mle->new_master = new_master;
	     ......

3163 __dlm_insert_mle(dlm, mle);
3164
3165 return ret;
3166 }

regards,
wengang.

#3 lore raw

PATCH 1/1 OCFS2: don't leave free'd mle attached to hb events

oops... I read it incorrectly. But that does not sound right. Let me think about it.

Quoted 72 lines · Wengang Wang

Hi Sunil,

Quoted 15 lines · Sunil Mushran

NAK

Quoted 11 lines · wengang wang

don't leave free'd mle attached to hb events.
in dlm_add_migration_mle() the mle is attched to "heartbeat
events" anyway no
matter there is an existing mle with same name(returns -EEXIST). dlm_migrate_lockres() calls dlm_add_migration_mle(). in case the later function
returning -EEXIST, dlm_migrate_lockres() frees the (new) mle without detaching
it from "hb events". so that later "hb events" related operations could improperly
operate against wrong mle objects or against an invalid memory address.

Quoted 6 lines

The mle is attached to hb events in dlm_init_mle() which is not called if it returns -EEXIST. When it returns -EEXIST, oldmle is set to the existing mle and its refcounting is handled correctly. mle is not touched
and thus only needs to be freed.

Maybe I am wrong. but are you sure dlm_init_mle() is not called when it returns -EEXIST?
in code, it doesn't return immediately after setting ret with -EEXIST, but continue to call
dlm_init_mle();
and then
mle->new_master = new_master;

Maybe I am wrong somewhere?
simplified code pasted here:

3098 static int dlm_add_migration_mle(struct dlm_ctxt *dlm,

3099                                  struct dlm_lock_resource *res,
3100                                  struct dlm_master_list_entry *mle,
3101                                  struct dlm_master_list_entry 

**oldmle,
3102 const char *name, unsigned int
namelen,
3103 u8 new_master, u8 master)
3104 {
3116 found = dlm_find_mle(dlm, oldmle, (char *)name, namelen);
3117 if (found) {
3118 struct dlm_master_list_entry *tmp = *oldmle;
3120 if (tmp->type == DLM_MLE_MIGRATION) {
3121 if (master == dlm->node_num) {
3126 ret = -EEXIST;
3127 } else {
3136 BUG();
3137 }

3138                 } else {
                 ....

3151 }
3153 }
3154
3156 dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name,
namelen);

3157         mle->new_master = new_master;
         ......

3163 __dlm_insert_mle(dlm, mle);
3164
3165 return ret;
3166 }

regards,
wengang.

#4 lore raw

PATCH 1/1 OCFS2: don't leave free'd mle attached to hb events

Just a reminder:)

Quoted 22 lines · Sunil Mushran

oops... I read it incorrectly. But that does not sound right. Let me think about it.

Quoted 18 lines · Wengang Wang

Hi Sunil,

Quoted 15 lines · Sunil Mushran

NAK

Quoted 11 lines · wengang wang

don't leave free'd mle attached to hb events.
in dlm_add_migration_mle() the mle is attched to "heartbeat
events" anyway no
matter there is an existing mle with same name(returns -EEXIST). dlm_migrate_lockres() calls dlm_add_migration_mle(). in case the later function
returning -EEXIST, dlm_migrate_lockres() frees the (new) mle without detaching
it from "hb events". so that later "hb events" related operations could improperly
operate against wrong mle objects or against an invalid memory address.