Skip to contentSkip to conversation

atomic: Fix _atomic_dec_and_lock() deadlock on UP

PATCH In review Valerie Aurora 3 participants 5 messages Jun 15, 2009 Library code

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 atomic: Fix _atomic_dec_and_lock() deadlock on UP

From: Jan Blunck <[email protected]>

_atomic_dec_and_lock() can deadlock on UP with spinlock debugging enabled. Currently, on UP we unconditionally spin_lock() first, which calls __spin_lock_debug(), which takes the lock unconditionally even on UP. This will deadlock in situations in which we call atomic_dec_and_lock() knowing that the counter won't go to zero (because we hold another reference) and that we already hold the lock. Instead, we should use the SMP code path which only takes the lock if necessary.

---
 lib/dec_and_lock.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
lib/dec_and_lock.c +1 -2
index a65c314..e73822a 100644
@@ -19,11 +19,10 @@
  */
 int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock)
 {
-#ifdef CONFIG_SMP
 	/* Subtract 1 from counter unless that drops it to 0 (ie. it was 1) */
 	if (atomic_add_unless(atomic, -1, 1))
 		return 0;
-#endif
+
 	/* Otherwise do it the slow way */
 	spin_lock(lock);
 	if (atomic_dec_and_test(atomic))
Signature · 1 line
1.6.0.6
maintainer Acked-by Paul E. McKenney Acked-by Nick Piggin on v1 1/1
#1 lore raw
Quoted 8 lines · On Mon, 15 Jun 2009 14:11:13 -0400 Valerie Aurora <[email protected]> wrote:

_atomic_dec_and_lock() can deadlock on UP with spinlock debugging enabled. Currently, on UP we unconditionally spin_lock() first, which calls __spin_lock_debug(), which takes the lock unconditionally even on UP. This will deadlock in situations in which we call atomic_dec_and_lock() knowing that the counter won't go to zero (because we hold another reference) and that we already hold the lock. Instead, we should use the SMP code path which only takes the lock if necessary.

Yup, I have this queued for 2.6.31 as atomic-only-take-lock-when-the-counter-drops-to-zero-on-up-as-well.patch, with a different changelog:

_atomic_dec_and_lock() should not unconditionally take the lock before calling atomic_dec_and_test() in the UP case. For consistency reasons it should behave exactly like in the SMP case.

Besides that this works around the problem that with CONFIG_DEBUG_SPINLOCK this spins in __spin_lock_debug() if the lock is already taken even if the counter doesn't drop to 0.

I can't remember why we decided that 2.6.30 doesn't need this.

Reviewed-by on v1 1/1
#4 lore raw
In reply to Valerie Aurora
Quoted 10 lines · Valerie Aurora

From: Jan Blunck <[email protected]>

_atomic_dec_and_lock() can deadlock on UP with spinlock debugging enabled. Currently, on UP we unconditionally spin_lock() first, which calls __spin_lock_debug(), which takes the lock unconditionally even on UP. This will deadlock in situations in which we call atomic_dec_and_lock() knowing that the counter won't go to zero (because we hold another reference) and that we already hold the lock. Instead, we should use the SMP code path which only takes the lock if necessary.

Quoted 26 lines
---
 lib/dec_and_lock.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)
lib/dec_and_lock.c +1 -2
index a65c314..e73822a 100644
@@ -19,11 +19,10 @@
  */
 int _atomic_dec_and_lock(atomic_t *atomic, spinlock_t *lock)
 {
-#ifdef CONFIG_SMP
 	/* Subtract 1 from counter unless that drops it to 0 (ie. it was 1) */
 	if (atomic_add_unless(atomic, -1, 1))
 		return 0;
-#endif
+
 	/* Otherwise do it the slow way */
 	spin_lock(lock);
 	if (atomic_dec_and_test(atomic))
Signature · 1 line
1.6.0.6
#2 lore raw
In reply to Andrew Morton
Quoted 24 lines · Andrew Morton
Quoted 8 lines · On Mon, 15 Jun 2009 14:11:13 -0400 Valerie Aurora <[email protected]> wrote:

_atomic_dec_and_lock() can deadlock on UP with spinlock debugging enabled. Currently, on UP we unconditionally spin_lock() first, which calls __spin_lock_debug(), which takes the lock unconditionally even on UP. This will deadlock in situations in which we call atomic_dec_and_lock() knowing that the counter won't go to zero (because we hold another reference) and that we already hold the lock. Instead, we should use the SMP code path which only takes the lock if necessary.

Yup, I have this queued for 2.6.31 as atomic-only-take-lock-when-the-counter-drops-to-zero-on-up-as-well.patch, with a different changelog:

_atomic_dec_and_lock() should not unconditionally take the lock before calling atomic_dec_and_test() in the UP case. For consistency reasons it should behave exactly like in the SMP case.

Besides that this works around the problem that with CONFIG_DEBUG_SPINLOCK this spins in __spin_lock_debug() if the lock is already taken even if the counter doesn't drop to 0.

Quoted 7 lines

I can't remember why we decided that 2.6.30 doesn't need this.

Great, last I heard the changelog was still a problem. Thanks,

-VAL

maintainer
#3 lore raw
Quoted 31 lines · On Mon, 15 Jun 2009 15:12:23 -0400 Valerie Aurora <[email protected]> wrote:
Quoted 30 lines · Andrew Morton
Quoted 8 lines · On Mon, 15 Jun 2009 14:11:13 -0400 Valerie Aurora <[email protected]> wrote:

_atomic_dec_and_lock() can deadlock on UP with spinlock debugging enabled. Currently, on UP we unconditionally spin_lock() first, which calls __spin_lock_debug(), which takes the lock unconditionally even on UP. This will deadlock in situations in which we call atomic_dec_and_lock() knowing that the counter won't go to zero (because we hold another reference) and that we already hold the lock. Instead, we should use the SMP code path which only takes the lock if necessary.

Yup, I have this queued for 2.6.31 as atomic-only-take-lock-when-the-counter-drops-to-zero-on-up-as-well.patch, with a different changelog:

_atomic_dec_and_lock() should not unconditionally take the lock before calling atomic_dec_and_test() in the UP case. For consistency reasons it should behave exactly like in the SMP case.

Besides that this works around the problem that with CONFIG_DEBUG_SPINLOCK this spins in __spin_lock_debug() if the lock is already taken even if the counter doesn't drop to 0.

Quoted 4 lines
Quoted 1 line

I can't remember why we decided that 2.6.30 doesn't need this.

Great, last I heard the changelog was still a problem. Thanks,

<goes back and checks>

OK, I decided that we didn't need this in 2.6.30 or earlier because Jan's union mount code is the only known triggerer of the problem.

However the patch is clearly a suitable thing for -stable. Opinions are sought..