summaryrefslogtreecommitdiffstats
path: root/sys
diff options
context:
space:
mode:
authormtm <mtm@FreeBSD.org>2003-07-17 11:06:40 +0000
committermtm <mtm@FreeBSD.org>2003-07-17 11:06:40 +0000
commited2409013ed9cb392695694f321405fc7b0f89b8 (patch)
treeaa1d125272406cc517371df61e50075c3b91fe39 /sys
parent003334ca8fd4b95ebd29f7cc4862495ea88067f4 (diff)
downloadFreeBSD-src-ed2409013ed9cb392695694f321405fc7b0f89b8.zip
FreeBSD-src-ed2409013ed9cb392695694f321405fc7b0f89b8.tar.gz
Fix umtx locking, for libthr, in the kernel.
1. There was a race condition between a thread unlocking a umtx and the thread contesting it. If the unlocking thread won the race it may try to wakeup a thread that was not yet in msleep(). The contesting thread would then go to sleep to await a wakeup that would never come. It's not possible to close the race by using a lock because calls to casuptr() may have to fault a page in from swap. Instead, the race was closed by introducing a flag that the unlocking thread will set when waking up a thread. The contesting thread will check for this flag before going to sleep. For now the flag is kept in td_flags, but it may be better to use some other member or create a new one because of the possible performance/contention issues of having to own sched_lock. Thanks to jhb for pointing me in the right direction on this one. 2. Once a umtx was contested all future locks and unlocks were happening in the kernel, regardless of whether it was contested or not. To prevent this from happening, when a thread locks a umtx it checks the queue for that umtx and unsets the contested bit if there are no other threads waiting on it. Again, this is slightly more complicated than it needs to be because we can't hold a lock across casuptr(). So, the thread has to check the queue again after unseting the bit, and reset the contested bit if it finds that another thread has put itself on the queue in the mean time. 3. Remove the if... block for unlocking an uncontested umtx, and replace it with a KASSERT. The _only_ time a thread should be unlocking a umtx in the kernel is if it is contested.
Diffstat (limited to 'sys')
-rw-r--r--sys/kern/kern_umtx.c71
-rw-r--r--sys/sys/proc.h1
2 files changed, 48 insertions, 24 deletions
diff --git a/sys/kern/kern_umtx.c b/sys/kern/kern_umtx.c
index d974b1d..09a419f 100644
--- a/sys/kern/kern_umtx.c
+++ b/sys/kern/kern_umtx.c
@@ -181,7 +181,7 @@ _umtx_lock(struct thread *td, struct _umtx_lock_args *uap)
return (EFAULT);
if (owner == UMTX_CONTESTED)
- return (0);
+ goto out;
/* If this failed the lock has changed, restart. */
continue;
@@ -211,14 +211,20 @@ _umtx_lock(struct thread *td, struct _umtx_lock_args *uap)
/*
* We set the contested bit, sleep. Otherwise the lock changed
- * and we need to retry.
+ * and we need to retry or we lost a race to the thread
+ * unlocking the umtx.
*/
UMTX_LOCK();
- if (old == owner)
+ mtx_lock_spin(&sched_lock);
+ if (old == owner && (td->td_flags & TDF_UMTXWAKEUP) == 0) {
+ mtx_unlock_spin(&sched_lock);
error = msleep(td, &umtx_lock,
td->td_priority | PCATCH, "umtx", 0);
- else
+ mtx_lock_spin(&sched_lock);
+ } else
error = 0;
+ td->td_flags &= ~TDF_UMTXWAKEUP;
+ mtx_unlock_spin(&sched_lock);
umtx_remove(uq, td);
UMTX_UNLOCK();
@@ -230,8 +236,37 @@ _umtx_lock(struct thread *td, struct _umtx_lock_args *uap)
if (error)
return (error);
}
-
- return (0);
+out:
+ /*
+ * We reach here only if we just acquired a contested umtx.
+ *
+ * If there are no other threads on this umtx's queue
+ * clear the contested bit. However, we cannot hold
+ * a lock across casuptr(). So after we unset it we
+ * have to recheck, and set it again if another thread has
+ * put itself on the queue in the mean time.
+ */
+ error = 0;
+ UMTX_LOCK();
+ uq = umtx_lookup(td, umtx);
+ UMTX_UNLOCK();
+ if (uq == NULL)
+ old = casuptr((intptr_t *)&umtx->u_owner,
+ ((intptr_t)td | UMTX_CONTESTED), (intptr_t)td);
+ if (uq == NULL && old == ((intptr_t)td | UMTX_CONTESTED)) {
+ UMTX_LOCK();
+ uq = umtx_lookup(td, umtx);
+ UMTX_UNLOCK();
+ if (uq != NULL) {
+ old = casuptr((intptr_t *)&umtx->u_owner,
+ (intptr_t)td, ((intptr_t)td | UMTX_CONTESTED));
+ if (old == -1)
+ error = EFAULT;
+ else if (old != (intptr_t)td)
+ error = EINVAL;
+ }
+ }
+ return (error);
}
int
@@ -257,25 +292,9 @@ _umtx_unlock(struct thread *td, struct _umtx_unlock_args *uap)
if ((struct thread *)(owner & ~UMTX_CONTESTED) != td)
return (EPERM);
- /*
- * If we own it but it isn't contested then we can just release and
- * return.
- */
- if ((owner & UMTX_CONTESTED) == 0) {
- owner = casuptr((intptr_t *)&umtx->u_owner,
- (intptr_t)td, UMTX_UNOWNED);
- if (owner == -1)
- return (EFAULT);
- /*
- * If this failed someone modified the memory without going
- * through this api.
- */
- if (owner != (intptr_t)td)
- return (EINVAL);
-
- return (0);
- }
+ /* We should only ever be in here for contested locks */
+ KASSERT((owner & UMTX_CONTESTED) != 0, ("contested umtx is not."));
old = casuptr((intptr_t *)&umtx->u_owner, owner, UMTX_CONTESTED);
@@ -296,6 +315,10 @@ _umtx_unlock(struct thread *td, struct _umtx_unlock_args *uap)
uq = umtx_lookup(td, umtx);
if (uq != NULL) {
blocked = TAILQ_FIRST(&uq->uq_tdq);
+ KASSERT(blocked != NULL, ("umtx_q with no waiting threads."));
+ mtx_lock_spin(&sched_lock);
+ blocked->td_flags |= TDF_UMTXWAKEUP;
+ mtx_unlock_spin(&sched_lock);
wakeup(blocked);
}
diff --git a/sys/sys/proc.h b/sys/sys/proc.h
index 966432e..c907e29 100644
--- a/sys/sys/proc.h
+++ b/sys/sys/proc.h
@@ -350,6 +350,7 @@ struct thread {
#define TDF_NEEDRESCHED 0x010000 /* Thread needs to yield. */
#define TDF_NEEDSIGCHK 0x020000 /* Thread may need signal delivery. */
#define TDF_SA 0x040000 /* A scheduler activation based thread */
+#define TDF_UMTXWAKEUP 0x080000 /* Libthr thread must not sleep on a umtx. */
#define TDF_DEADLKTREAT 0x800000 /* Lock aquisition - deadlock treatment. */
/* "private" flags kept in td_pflags */
OpenPOWER on IntegriCloud