powerpc: Fix write-after-destroy in lock elision [BZ #20822]

The update of *adapt_count after the release of the lock causes a race
condition when thread A unlocks, thread B continues and destroys the
mutex, and thread A writes to *adapt_count.

3 files changed, 20 insertions, 12 deletions

diff --git a/sysdeps/unix/sysv/linux/powerpc/elision-lock.c b/sysdeps/unix/sysv/linux/powerpc/elision-lock.c
index 4589491..f7a5cbc 100644
--- a/sysdeps/unix/sysv/linux/powerpc/elision-lock.c
+++ b/sysdeps/unix/sysv/linux/powerpc/elision-lock.c
@@ -45,7 +45,9 @@
 int
 __lll_lock_elision (int *lock, short *adapt_count, EXTRAARG int pshared)
 {
-  if (*adapt_count > 0)
+  /* adapt_count is accessed concurrently but is just a hint.  Thus,
+     use atomic accesses but relaxed MO is sufficient.  */
+  if (atomic_load_relaxed (adapt_count) > 0)
     {
       goto use_lock;
     }
@@ -67,7 +69,8 @@ __lll_lock_elision (int *lock, short *adapt_count, EXTRAARG int pshared)
 	  if (_TEXASRU_FAILURE_PERSISTENT (__builtin_get_texasru ()))
 	    {
 	      if (aconf.skip_lock_internal_abort > 0)
-		*adapt_count = aconf.skip_lock_internal_abort;
+		atomic_store_relaxed (adapt_count,
+				      aconf.skip_lock_internal_abort);
 	      goto use_lock;
 	    }
 	}
@@ -75,7 +78,8 @@ __lll_lock_elision (int *lock, short *adapt_count, EXTRAARG int pshared)
 
   /* Fall back to locks for a bit if retries have been exhausted */
   if (aconf.try_tbegin > 0 && aconf.skip_lock_out_of_tbegin_retries > 0)
-    *adapt_count = aconf.skip_lock_out_of_tbegin_retries;
+    atomic_store_relaxed (adapt_count,
+			  aconf.skip_lock_out_of_tbegin_retries);
 
 use_lock:
   return LLL_LOCK ((*lock), pshared);
diff --git a/sysdeps/unix/sysv/linux/powerpc/elision-trylock.c b/sysdeps/unix/sysv/linux/powerpc/elision-trylock.c
index 1e5cbe8..ed244d3 100644
--- a/sysdeps/unix/sysv/linux/powerpc/elision-trylock.c
+++ b/sysdeps/unix/sysv/linux/powerpc/elision-trylock.c
@@ -34,7 +34,7 @@ __lll_trylock_elision (int *futex, short *adapt_count)
   __libc_tabort (_ABORT_NESTED_TRYLOCK);
 
   /* Only try a transaction if it's worth it.  */
-  if (*adapt_count > 0)
+  if (atomic_load_relaxed (adapt_count) > 0)
     {
       goto use_lock;
     }
@@ -49,7 +49,7 @@ __lll_trylock_elision (int *futex, short *adapt_count)
       __libc_tend (0);
 
       if (aconf.skip_lock_busy > 0)
-	*adapt_count = aconf.skip_lock_busy;
+	atomic_store_relaxed (adapt_count, aconf.skip_lock_busy);
     }
   else
     {
@@ -59,7 +59,8 @@ __lll_trylock_elision (int *futex, short *adapt_count)
 	     result in another failure.  Use normal locking now and
 	     for the next couple of calls.  */
 	  if (aconf.skip_trylock_internal_abort > 0)
-	    *adapt_count = aconf.skip_trylock_internal_abort;
+	    atomic_store_relaxed (adapt_count,
+				aconf.skip_trylock_internal_abort);
 	}
     }
 
diff --git a/sysdeps/unix/sysv/linux/powerpc/elision-unlock.c b/sysdeps/unix/sysv/linux/powerpc/elision-unlock.c
index 6f45a9c..759c146 100644
--- a/sysdeps/unix/sysv/linux/powerpc/elision-unlock.c
+++ b/sysdeps/unix/sysv/linux/powerpc/elision-unlock.c
@@ -28,13 +28,16 @@ __lll_unlock_elision (int *lock, short *adapt_count, int pshared)
     __libc_tend (0);
   else
     {
-      lll_unlock ((*lock), pshared);
+      /* Update adapt_count in the critical section to prevent a
+	 write-after-destroy error as mentioned in BZ 20822.  The
+	 following update of adapt_count has to be contained within
+	 the critical region of the fall-back lock in order to not violate
+	 the mutex destruction requirements.  */
+      short __tmp = atomic_load_relaxed (adapt_count);
+      if (__tmp > 0)
+	atomic_store_relaxed (adapt_count, __tmp--);
 
-      /* Update the adapt count AFTER completing the critical section.
-         Doing this here prevents unneeded stalling when entering
-         a critical section.  Saving about 8% runtime on P8.  */
-      if (*adapt_count > 0)
-	(*adapt_count)--;
+      lll_unlock ((*lock), pshared);
     }
   return 0;
 }