ipc/sem.c: fix race in sem_lock()
authorManfred Spraul <manfred@colorfullife.com>
Mon, 30 Sep 2013 20:45:04 +0000 (13:45 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 30 Sep 2013 21:31:01 +0000 (14:31 -0700)
The exclusion of complex operations in sem_lock() is insufficient: after
acquiring the per-semaphore lock, a simple op must first check that
sem_perm.lock is not locked and only after that test check
complex_count.  The current code does it the other way around - and that
creates a race.  Details are below.

The patch is a complete rewrite of sem_lock(), based in part on the code
from Mike Galbraith.  It removes all gotos and all loops and thus the
risk of livelocks.

I have tested the patch (together with the next one) on my i3 laptop and
it didn't cause any problems.

The bug is probably also present in 3.10 and 3.11, but for these kernels
it might be simpler just to move the test of sma->complex_count after
the spin_is_locked() test.

Details of the bug:

Assume:
 - sma->complex_count = 0.
 - Thread 1: semtimedop(complex op that must sleep)
 - Thread 2: semtimedop(simple op).

Pseudo-Trace:

Thread 1: sem_lock(): acquire sem_perm.lock
Thread 1: sem_lock(): check for ongoing simple ops
Nothing ongoing, thread 2 is still before sem_lock().
Thread 1: try_atomic_semop()
<<< preempted.

Thread 2: sem_lock():
        static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
                                      int nsops)
        {
                int locknum;
         again:
                if (nsops == 1 && !sma->complex_count) {
                        struct sem *sem = sma->sem_base + sops->sem_num;

                        /* Lock just the semaphore we are interested in. */
                        spin_lock(&sem->lock);

                        /*
                         * If sma->complex_count was set while we were spinning,
                         * we may need to look at things we did not lock here.
                         */
                        if (unlikely(sma->complex_count)) {
                                spin_unlock(&sem->lock);
                                goto lock_array;
                        }
        <<<<<<<<<
<<< complex_count is still 0.
<<<
        <<< Here it is preempted
        <<<<<<<<<

Thread 1: try_atomic_semop() returns, notices that it must sleep.
Thread 1: increases sma->complex_count.
Thread 1: drops sem_perm.lock
Thread 2:
                /*
                 * Another process is holding the global lock on the
                 * sem_array; we cannot enter our critical section,
                 * but have to wait for the global lock to be released.
                 */
                if (unlikely(spin_is_locked(&sma->sem_perm.lock))) {
                        spin_unlock(&sem->lock);
                        spin_unlock_wait(&sma->sem_perm.lock);
                        goto again;
                }
<<< sem_perm.lock already dropped, thus no "goto again;"

                locknum = sops->sem_num;

Signed-off-by: Manfred Spraul <manfred@colorfullife.com>
Cc: Mike Galbraith <bitbucket@online.de>
Cc: Rik van Riel <riel@redhat.com>
Cc: Davidlohr Bueso <davidlohr.bueso@hp.com>
Cc: <stable@vger.kernel.org> [3.10+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ipc/sem.c

index 19c8b980d1feff430d82297a53be528cea801c42..4a92c0447ad66d4a351a3988dbda69df641573e1 100644 (file)
--- a/ipc/sem.c
+++ b/ipc/sem.c
@@ -252,71 +252,105 @@ static void sem_rcu_free(struct rcu_head *head)
        ipc_rcu_free(head);
 }
 
+/*
+ * Wait until all currently ongoing simple ops have completed.
+ * Caller must own sem_perm.lock.
+ * New simple ops cannot start, because simple ops first check
+ * that sem_perm.lock is free.
+ */
+static void sem_wait_array(struct sem_array *sma)
+{
+       int i;
+       struct sem *sem;
+
+       for (i = 0; i < sma->sem_nsems; i++) {
+               sem = sma->sem_base + i;
+               spin_unlock_wait(&sem->lock);
+       }
+}
+
 /*
  * If the request contains only one semaphore operation, and there are
  * no complex transactions pending, lock only the semaphore involved.
  * Otherwise, lock the entire semaphore array, since we either have
  * multiple semaphores in our own semops, or we need to look at
  * semaphores from other pending complex operations.
- *
- * Carefully guard against sma->complex_count changing between zero
- * and non-zero while we are spinning for the lock. The value of
- * sma->complex_count cannot change while we are holding the lock,
- * so sem_unlock should be fine.
- *
- * The global lock path checks that all the local locks have been released,
- * checking each local lock once. This means that the local lock paths
- * cannot start their critical sections while the global lock is held.
  */
 static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
                              int nsops)
 {
-       int locknum;
- again:
-       if (nsops == 1 && !sma->complex_count) {
-               struct sem *sem = sma->sem_base + sops->sem_num;
+       struct sem *sem;
 
-               /* Lock just the semaphore we are interested in. */
-               spin_lock(&sem->lock);
+       if (nsops != 1) {
+               /* Complex operation - acquire a full lock */
+               ipc_lock_object(&sma->sem_perm);
 
-               /*
-                * If sma->complex_count was set while we were spinning,
-                * we may need to look at things we did not lock here.
+               /* And wait until all simple ops that are processed
+                * right now have dropped their locks.
                 */
-               if (unlikely(sma->complex_count)) {
-                       spin_unlock(&sem->lock);
-                       goto lock_array;
-               }
+               sem_wait_array(sma);
+               return -1;
+       }
+
+       /*
+        * Only one semaphore affected - try to optimize locking.
+        * The rules are:
+        * - optimized locking is possible if no complex operation
+        *   is either enqueued or processed right now.
+        * - The test for enqueued complex ops is simple:
+        *      sma->complex_count != 0
+        * - Testing for complex ops that are processed right now is
+        *   a bit more difficult. Complex ops acquire the full lock
+        *   and first wait that the running simple ops have completed.
+        *   (see above)
+        *   Thus: If we own a simple lock and the global lock is free
+        *      and complex_count is now 0, then it will stay 0 and
+        *      thus just locking sem->lock is sufficient.
+        */
+       sem = sma->sem_base + sops->sem_num;
 
+       if (sma->complex_count == 0) {
                /*
-                * Another process is holding the global lock on the
-                * sem_array; we cannot enter our critical section,
-                * but have to wait for the global lock to be released.
+                * It appears that no complex operation is around.
+                * Acquire the per-semaphore lock.
                 */
-               if (unlikely(spin_is_locked(&sma->sem_perm.lock))) {
-                       spin_unlock(&sem->lock);
-                       spin_unlock_wait(&sma->sem_perm.lock);
-                       goto again;
+               spin_lock(&sem->lock);
+
+               /* Then check that the global lock is free */
+               if (!spin_is_locked(&sma->sem_perm.lock)) {
+                       /* spin_is_locked() is not a memory barrier */
+                       smp_mb();
+
+                       /* Now repeat the test of complex_count:
+                        * It can't change anymore until we drop sem->lock.
+                        * Thus: if is now 0, then it will stay 0.
+                        */
+                       if (sma->complex_count == 0) {
+                               /* fast path successful! */
+                               return sops->sem_num;
+                       }
                }
+               spin_unlock(&sem->lock);
+       }
 
-               locknum = sops->sem_num;
+       /* slow path: acquire the full lock */
+       ipc_lock_object(&sma->sem_perm);
+
+       if (sma->complex_count == 0) {
+               /* False alarm:
+                * There is no complex operation, thus we can switch
+                * back to the fast path.
+                */
+               spin_lock(&sem->lock);
+               ipc_unlock_object(&sma->sem_perm);
+               return sops->sem_num;
        } else {
-               int i;
-               /*
-                * Lock the semaphore array, and wait for all of the
-                * individual semaphore locks to go away.  The code
-                * above ensures no new single-lock holders will enter
-                * their critical section while the array lock is held.
+               /* Not a false alarm, thus complete the sequence for a
+                * full lock.
                 */
- lock_array:
-               ipc_lock_object(&sma->sem_perm);
-               for (i = 0; i < sma->sem_nsems; i++) {
-                       struct sem *sem = sma->sem_base + i;
-                       spin_unlock_wait(&sem->lock);
-               }
-               locknum = -1;
+               sem_wait_array(sma);
+               return -1;
        }
-       return locknum;
 }
 
 static inline void sem_unlock(struct sem_array *sma, int locknum)