futex_unlock_pi() gets uval before taking the hb lock. Now imagine
someone in futex_lock_pi() took the lock. While futex_unlock_pi() waits
for the hb lock, the LOCK_PI sets FUTEX_WAITERS and drops the lock.
Now, futex_unlock_pi() figures out that there is waiter and invokes
wake_futex_pi() with the old uval which does not yet have FUTEX_WAITERS
set. This flaw lets cmpxchg_futex_value_locked() fail and return -EINVAL.
To address this race we could either use get_futex_value_locked() after
we obtained the lock but then we would need to handle the EFAULT case
(which I think means get_user() without the lock).
To avoid this I let wake_futex_pi() detect this and return -EAGAIN so it
can read the new value and try again.
As bad as this may sound: FUTEX_OWNER_DIED is not set while the owner is
unlocking the lock. FUTEX_WAITERS gets set by another task *but* this
means that the owner invoked the futex() syscall rather than doing a
cmpxchg() in userland (and libc does not do that).
Fixes: ccf9e6a80d9e ("futex: Make unlock_pi more robust")
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
---
kernel/futex.c | 16 ++++++++++++++--
1 file changed, 14 insertions(+), 2 deletions(-)
diff --git a/kernel/futex.c b/kernel/futex.c
index 0d557e13823a..2911f54a6dde 100644
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -1496,8 +1496,12 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
ret = -EFAULT;
- else if (curval != uval)
- ret = -EINVAL;
+ else if (curval != uval) {
+ if ((FUTEX_TID_MASK & curval) == uval)
+ ret = -EAGAIN;
+ else
+ ret = -EINVAL;
+ }
if (ret) {
raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
return ret;
@@ -2852,6 +2856,12 @@ static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
if (ret == -EFAULT)
goto pi_faulted;
/*
+ * Between get_user() and obtaining the hb->lock the uval
+ * gained a flag. Retry with the new value.
+ */
+ if (ret == -EAGAIN)
+ goto pi_faulted;
+ /*
* wake_futex_pi has detected invalid state. Tell user
* space.
*/
@@ -2883,6 +2893,8 @@ static int futex_unlock_pi(u32 __user *uaddr, unsigned int flags)
spin_unlock(&hb->lock);
put_futex_key(&key);
+ if (ret == -EAGAIN)
+ goto retry;
ret = fault_in_user_writeable(uaddr);
if (!ret)
goto retry;
--
2.8.0.rc3
On Fri, 15 Apr 2016, Sebastian Andrzej Siewior wrote:
>futex_unlock_pi() gets uval before taking the hb lock. Now imagine
>someone in futex_lock_pi() took the lock. While futex_unlock_pi() waits
>for the hb lock, the LOCK_PI sets FUTEX_WAITERS and drops the lock.
>Now, futex_unlock_pi() figures out that there is waiter and invokes
>wake_futex_pi() with the old uval which does not yet have FUTEX_WAITERS
>set. This flaw lets cmpxchg_futex_value_locked() fail and return -EINVAL.
Hmm but if we're calling futex_unlock_pi() in the first place, doesn't that
indicate that the uval already has FUTEX_WAITERS and therefore failed the
TID->0 transition in userland? That or the thread is bogusly unlocking a
lock that it doesn't own.
This is of course different than the requeue_pi case which can specify
set_waiters but also gets the value via get_futex_value_locked().
Is this a real issue or did you find it by code inspection?
Thanks,
Davidlohr
On 04/20/2016 12:27 AM, Davidlohr Bueso wrote:
> On Fri, 15 Apr 2016, Sebastian Andrzej Siewior wrote:
>
>> futex_unlock_pi() gets uval before taking the hb lock. Now imagine
>> someone in futex_lock_pi() took the lock. While futex_unlock_pi() waits
>> for the hb lock, the LOCK_PI sets FUTEX_WAITERS and drops the lock.
>> Now, futex_unlock_pi() figures out that there is waiter and invokes
>> wake_futex_pi() with the old uval which does not yet have FUTEX_WAITERS
>> set. This flaw lets cmpxchg_futex_value_locked() fail and return -EINVAL.
>
> Hmm but if we're calling futex_unlock_pi() in the first place, doesn't that
> indicate that the uval already has FUTEX_WAITERS and therefore failed the
> TID->0 transition in userland? That or the thread is bogusly unlocking a
> lock that it doesn't own.
I hacked a testing tool which always did the syscall for LOCK_PI /
UNLOCK_PI and this popped up.
>
> This is of course different than the requeue_pi case which can specify
> set_waiters but also gets the value via get_futex_value_locked().
>
> Is this a real issue or did you find it by code inspection?
real issue.
https://breakpoint.cc/mass-futex2-rl.c
> Thanks,
> Davidlohr
Sebastian
On Tue, 19 Apr 2016, Davidlohr Bueso wrote:
> On Fri, 15 Apr 2016, Sebastian Andrzej Siewior wrote:
>
> > futex_unlock_pi() gets uval before taking the hb lock. Now imagine
> > someone in futex_lock_pi() took the lock. While futex_unlock_pi() waits
> > for the hb lock, the LOCK_PI sets FUTEX_WAITERS and drops the lock.
> > Now, futex_unlock_pi() figures out that there is waiter and invokes
> > wake_futex_pi() with the old uval which does not yet have FUTEX_WAITERS
> > set. This flaw lets cmpxchg_futex_value_locked() fail and return -EINVAL.
>
> Hmm but if we're calling futex_unlock_pi() in the first place, doesn't that
> indicate that the uval already has FUTEX_WAITERS and therefore failed the
> TID->0 transition in userland? That or the thread is bogusly unlocking a
> lock that it doesn't own.
It can be called unconditionally w/o trying the TID->0 transition in user
space first and we should handle that case.
Thanks,
tglx
Commit-ID: 89e9e66ba1b3bde9d8ea90566c2aee20697ad681
Gitweb: http://git.kernel.org/tip/89e9e66ba1b3bde9d8ea90566c2aee20697ad681
Author: Sebastian Andrzej Siewior <[email protected]>
AuthorDate: Fri, 15 Apr 2016 14:35:39 +0200
Committer: Thomas Gleixner <[email protected]>
CommitDate: Wed, 20 Apr 2016 12:33:13 +0200
futex: Handle unlock_pi race gracefully
If userspace calls UNLOCK_PI unconditionally without trying the TID -> 0
transition in user space first then the user space value might not have the
waiters bit set. This opens the following race:
CPU0 CPU1
uval = get_user(futex)
lock(hb)
lock(hb)
futex |= FUTEX_WAITERS
....
unlock(hb)
cmpxchg(futex, uval, newval)
So the cmpxchg fails and returns -EINVAL to user space, which is wrong because
the futex value is valid.
To handle this (yes, yet another) corner case gracefully, check for a flag
change and retry.
[ tglx: Massaged changelog and slightly reworked implementation ]
Fixes: ccf9e6a80d9e ("futex: Make unlock_pi more robust")
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
Cc: [email protected]
Cc: Davidlohr Bueso <[email protected]>
Cc: Darren Hart <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Thomas Gleixner <[email protected]>
---
kernel/futex.c | 25 ++++++++++++++++++++++---
1 file changed, 22 insertions(+), 3 deletions(-)
diff --git a/kernel/futex.c b/kernel/futex.c
index a5d2e74..fd204e1 100644
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -1295,10 +1295,20 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval, struct futex_q *this,
if (unlikely(should_fail_futex(true)))
ret = -EFAULT;
- if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
+ if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval)) {
ret = -EFAULT;
- else if (curval != uval)
- ret = -EINVAL;
+ } else if (curval != uval) {
+ /*
+ * If a unconditional UNLOCK_PI operation (user space did not
+ * try the TID->0 transition) raced with a waiter setting the
+ * FUTEX_WAITERS flag between get_user() and locking the hash
+ * bucket lock, retry the operation.
+ */
+ if ((FUTEX_TID_MASK & curval) == uval)
+ ret = -EAGAIN;
+ else
+ ret = -EINVAL;
+ }
if (ret) {
raw_spin_unlock_irq(&pi_state->pi_mutex.wait_lock);
return ret;
@@ -2623,6 +2633,15 @@ retry:
if (ret == -EFAULT)
goto pi_faulted;
/*
+ * A unconditional UNLOCK_PI op raced against a waiter
+ * setting the FUTEX_WAITERS bit. Try again.
+ */
+ if (ret == -EAGAIN) {
+ spin_unlock(&hb->lock);
+ put_futex_key(&key);
+ goto retry;
+ }
+ /*
* wake_futex_pi has detected invalid state. Tell user
* space.
*/