[PATCH RESEND] Take over futex of dead task only if FUTEX_WAITERS isnot set

From: Siddhesh Poyarekar
Date: Wed Oct 17 2012 - 03:15:53 EST

In futex_lock_pi_atomic, we consider that if the value in the futex
variable is 0 with additional flags, then it is safe for takeover
since the owner of the futex is dead. However, when FUTEX_WAITERS is
set in the futex value, handle_futex_death calls futex_wake to wake up
one task. Hence the assumption in futex_lock_pi_atomic is not correct.
The correct assumption is that a futex may be considered safe for a
takeover if The FUTEX_OWNER_DIED bit is set, the TID bits are 0 and
the FUTEX_WAITERS bit is not set.

The race described above can be seen in the reproducer in the
following glibc bug report:


Signed-off-by: Siddhesh Poyarekar <siddhesh.poyarekar@xxxxxxxxx>
kernel/futex.c | 7 ++++++-
1 files changed, 6 insertions(+), 1 deletions(-)

diff --git a/kernel/futex.c b/kernel/futex.c
index 3717e7b..9aa2d5a 100644
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -760,9 +760,14 @@ retry:
* case. We also do an unconditional take over, when the owner
* of the futex died.
+ * We do not take over the futex if FUTEX_WAITERS is set because we
+ * could end up waking two tasks, the current one and the one that the
+ * futex death event wakes in handle_futex_death.
+ *
* This is safe as we are protected by the hash bucket lock !
- if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
+ if (unlikely(ownerdied ||
+ !(curval & (FUTEX_TID_MASK | FUTEX_WAITERS)))) {
/* Keep the OWNER_DIED bit */
newval = (curval & ~FUTEX_TID_MASK) | vpid;
ownerdied = 0;

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/