Skip to contentSkip to conversation

AB-BA deadlock between uidhash_lock and tasklist_lock.

PATCH In review Robin Holt 2 participants 2 messages Mar 10, 2005

lore
Share & tools
lore

Robin Holt

author
#0 lore raw

PATCH 1/1 AB-BA deadlock between uidhash_lock and tasklist_lock.

We have uncovered a very difficult to trip AB-BA deadlock between the uidhash_lock and tasklist_lock.

reparent_to_init() does write_lock_irq(&tasklist_lock) then calls switch_uid() which calls free_uid() which grabs the uidhash_lock.

Independent of that, we have seen a different cpu call free_uid as a result of sys_wait4 and, immediately after acquiring the uidhash_lock, receive a timer interrupt which eventually leads to an attempt to grab the tasklist_lock.

kernel/user.c +12 -6
Index: linux/kernel/user.c
===================================================================
@@ -90,6 +90,9 @@
 
 void free_uid(struct user_struct *up)
 {
+	unsigned long flags;
+
+	local_irq_save(flags);
 	if (up && atomic_dec_and_lock(&up->__count, &uidhash_lock)) {
 		uid_hash_remove(up);
 		key_put(up->uid_keyring);
@@ -97,16 +100,18 @@
 		kmem_cache_free(uid_cachep, up);
 		spin_unlock(&uidhash_lock);
 	}
+	local_irq_restore(flags);
 }
 
 struct user_struct * alloc_uid(uid_t uid)
 {
 	struct list_head *hashent = uidhashentry(uid);
 	struct user_struct *up;
+	unsigned long flags;
 
-	spin_lock(&uidhash_lock);
+	spin_lock_irqsave(&uidhash_lock, flags);
 	up = uid_hash_find(uid, hashent);
-	spin_unlock(&uidhash_lock);
+	spin_unlock_irqrestore(&uidhash_lock, flags);
 
 	if (!up) {
 		struct user_struct *new;
@@ -132,7 +137,7 @@
 		 * Before adding this, check whether we raced
 		 * on adding the same user already..
 		 */
-		spin_lock(&uidhash_lock);
+		spin_lock_irqsave(&uidhash_lock, flags);
 		up = uid_hash_find(uid, hashent);
 		if (up) {
 			key_put(new->uid_keyring);
@@ -142,7 +147,7 @@
 			uid_hash_insert(new, hashent);
 			up = new;
 		}
-		spin_unlock(&uidhash_lock);
+		spin_unlock_irqrestore(&uidhash_lock, flags);
 
 	}
 	return up;
@@ -170,6 +175,7 @@
 static int __init uid_cache_init(void)
 {
 	int n;
+	unsigned long flags;
 
 	uid_cachep = kmem_cache_create("uid_cache", sizeof(struct user_struct),
 			0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
@@ -178,9 +184,9 @@
 		INIT_LIST_HEAD(uidhash_table + n);
 
 	/* Insert the root user immediately (init already runs as root) */
-	spin_lock(&uidhash_lock);
+	spin_lock_irqsave(&uidhash_lock, flags);
 	uid_hash_insert(&root_user, uidhashentry(0));
-	spin_unlock(&uidhash_lock);
+	spin_unlock_irqrestore(&uidhash_lock, flags);
 
 	return 0;
 }
#1 lore raw
Quoted 8 lines · Robin Holt

reparent_to_init() does write_lock_irq(&tasklist_lock) then calls switch_uid() which calls free_uid() which grabs the uidhash_lock.

Independent of that, we have seen a different cpu call free_uid as a result of sys_wait4 and, immediately after acquiring the uidhash_lock, receive a timer interrupt which eventually leads to an attempt to grab the tasklist_lock.

Hmm.. We fixed this already, and the current tree doesn't have this problem (and the fix was much simpler: just move "switch_uid()" to outside the tasklist_lock.

That fix was done late december last year (current BK revision:
1.1938.446.38), and I really think your patch is stale. Please double-check,

Linus