[PATCH] "Fix uidhash_lock <-> RXU deadlock" fix
Andrew Morton [Wed, 1 Feb 2006 00:34:26 +0000 (16:34 -0800)]
I get storms of warnings from local_bh_enable().  Better-tested patches,
please.

Cc: Ingo Molnar <mingo@elte.hu>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

kernel/user.c

index d1ae234..d9deae4 100644 (file)
@@ -33,6 +33,10 @@ static struct list_head uidhash_table[UIDHASH_SZ];
  * The uidhash_lock is mostly taken from process context, but it is
  * occasionally also taken from softirq/tasklet context, when
  * task-structs get RCU-freed. Hence all locking must be softirq-safe.
+ * But free_uid() is also called with local interrupts disabled, and running
+ * local_bh_enable() with local interrupts disabled is an error - we'll run
+ * softirq callbacks, and they can unconditionally enable interrupts, and
+ * the caller of free_uid() didn't expect that..
  */
 static DEFINE_SPINLOCK(uidhash_lock);
 
@@ -89,16 +93,19 @@ static inline struct user_struct *uid_hash_find(uid_t uid, struct list_head *has
 struct user_struct *find_user(uid_t uid)
 {
        struct user_struct *ret;
+       unsigned long flags;
 
-       spin_lock_bh(&uidhash_lock);
+       spin_lock_irqsave(&uidhash_lock, flags);
        ret = uid_hash_find(uid, uidhashentry(uid));
-       spin_unlock_bh(&uidhash_lock);
+       spin_unlock_irqrestore(&uidhash_lock, flags);
        return ret;
 }
 
 void free_uid(struct user_struct *up)
 {
-       local_bh_disable();
+       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);
@@ -106,7 +113,7 @@ void free_uid(struct user_struct *up)
                kmem_cache_free(uid_cachep, up);
                spin_unlock(&uidhash_lock);
        }
-       local_bh_enable();
+       local_irq_restore(flags);
 }
 
 struct user_struct * alloc_uid(uid_t uid)
@@ -114,9 +121,9 @@ struct user_struct * alloc_uid(uid_t uid)
        struct list_head *hashent = uidhashentry(uid);
        struct user_struct *up;
 
-       spin_lock_bh(&uidhash_lock);
+       spin_lock_irq(&uidhash_lock);
        up = uid_hash_find(uid, hashent);
-       spin_unlock_bh(&uidhash_lock);
+       spin_unlock_irq(&uidhash_lock);
 
        if (!up) {
                struct user_struct *new;
@@ -146,7 +153,7 @@ struct user_struct * alloc_uid(uid_t uid)
                 * Before adding this, check whether we raced
                 * on adding the same user already..
                 */
-               spin_lock_bh(&uidhash_lock);
+               spin_lock_irq(&uidhash_lock);
                up = uid_hash_find(uid, hashent);
                if (up) {
                        key_put(new->uid_keyring);
@@ -156,7 +163,7 @@ struct user_struct * alloc_uid(uid_t uid)
                        uid_hash_insert(new, hashent);
                        up = new;
                }
-               spin_unlock_bh(&uidhash_lock);
+               spin_unlock_irq(&uidhash_lock);
 
        }
        return up;
@@ -192,9 +199,9 @@ static int __init uid_cache_init(void)
                INIT_LIST_HEAD(uidhash_table + n);
 
        /* Insert the root user immediately (init already runs as root) */
-       spin_lock_bh(&uidhash_lock);
+       spin_lock_irq(&uidhash_lock);
        uid_hash_insert(&root_user, uidhashentry(0));
-       spin_unlock_bh(&uidhash_lock);
+       spin_unlock_irq(&uidhash_lock);
 
        return 0;
 }