waitqueue lockdep annotation
Hi all,
this series adds a strategic lockdep_assert_held to __wake_up_common to ensure callers really do hold the wait_queue_head lock when calling the unlocked wake_up variants. It turns out epoll did not do this for a fairly common path (hit all the time by systemd during bootup), so the second patch fixed this instance as well.
PATCH 1/2 sched/wait: assert the wait_queue_head lock is held in __wake_up_common
Better ensure we actually hold the lock using lockdep than just commenting on it. Due to the various exported _locked interfaces it is far too easy to get the locking wrong.
--- kernel/sched/wait.c | 2 ++ 1 file changed, 2 insertions(+)
kernel/sched/wait.c +2 -0
@@ -76,6 +76,8 @@ static int __wake_up_common(struct wait_queue_head *wq_head, unsigned int mode,wait_queue_entry_t *curr, *next;int cnt =0 ; +lockdep_assert_held (&wq_head->lock); +if (bookmark && (bookmark->flags &WQ_FLAG_BOOKMARK )) { curr =list_next_entry (bookmark, entry);
Signature · 1 line
2.14.2
PATCH 2/2 epoll: use proper wake_up variant in ep_poll_callback
In reply to Christoph HellwigWe do not hold ep->wq->lock, so we should not use wake_up_locked on it.
--- fs/eventpoll.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/eventpoll.c +1 -1
@@ -1190,7 +1190,7 @@ static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, vbreak ; } } -wake_up_locked (&ep->wq); +wake_up (&ep->wq); }if (waitqueue_active (&ep->poll_wait)) pwake++;
Signature · 1 line
2.14.2
Quoted 7 lines · On Thu, 30 Nov 2017 06:20:35 -0800 Christoph Hellwig <[email protected]> wrote:
Hi all,
this series adds a strategic
lockdep_assert_heldto__wake_up_commonto ensure callers really do hold thewait_queue_headlock when calling the unlockedwake_upvariants. It turns out epoll did not do this for a fairly common path (hit all the time by systemd during bootup), so the second patch fixed this instance as well.
What are the runtime effects of the epoll bug?
Quoted 4 lines · Andrew Morton
Quoted 2 lines · On Thu, 30 Nov 2017 06:20:35 -0800 Christoph Hellwig <[email protected]> wrote:
Hi all,
Quoted 8 lines
Quoted 5 lines
this series adds a strategic
lockdep_assert_heldto__wake_up_commonto ensure callers really do hold thewait_queue_headlock when calling the unlockedwake_upvariants. It turns out epoll did not do this for a fairly common path (hit all the time by systemd during bootup), so the second patch fixed this instance as well.What are the runtime effects of the epoll bug?
I don't think there is a bug here. The 'wake_up_locked()' calls in epoll are being protected by the ep->lock, not the wait_queue_head lock. So arguably the 'annotation' is wrong, but I don't think there is a bug beyond that.
Thanks,
-Jason
Quoted 4 lines · Jason Baron
I don't think there is a bug here. The '
wake_up_locked()' calls in epoll are being protected by theep->lock, not thewait_queue_headlock. So arguably the 'annotation' is wrong, but I don't think there is a bug beyond that.
They can't be protected by ep->lock. The file might as well be watched for using poll or select as well, or just using epoll using another epoll fd.
Quoted 10 lines · Christoph Hellwig
Quoted 4 lines · Jason Baron
I don't think there is a bug here. The '
wake_up_locked()' calls in epoll are being protected by theep->lock, not thewait_queue_headlock. So arguably the 'annotation' is wrong, but I don't think there is a bug beyond that.They can't be protected by
ep->lock. The file might as well be watched for using poll or select as well, or just using epoll using another epoll fd.
Yes, but for those cases it uses the ep->poll_wait waitqueue not the ep->wq, which is guarded by the ep->wq->lock.
See the comments in 'struct eventpoll':
/* Wait queue used by sys_epoll_wait() */
wait_queue_head_t wq;
/* Wait queue used by file->poll() */
wait_queue_head_t poll_wait;Thanks,
-Jason
Quoted 2 lines · Jason Baron
Yes, but for those cases it uses the
ep->poll_waitwaitqueue not theep->wq, which is guarded by theep->wq->lock.
True. So it looks like we have one waitqueue in the system that is special in providing its own synchronization for waitqueues while entirely ignoring the waitqueue code documentation that states that waitqueues are internally synchronized.
We could drop the lockdep annotation, updated the documentation and not add any validation of the assumptions, or just make epoll fit the scheme used by everyone else. So either we can drop these patches, or I need to fix up more of the epoll code.
Quoted 9 lines · Christoph Hellwig
Quoted 2 lines · Jason Baron
Yes, but for those cases it uses the
ep->poll_waitwaitqueue not theep->wq, which is guarded by theep->wq->lock.True. So it looks like we have one waitqueue in the system that is special in providing its own synchronization for waitqueues while entirely ignoring the waitqueue code documentation that states that waitqueues are internally synchronized.
Quoted 5 lines
We could drop the lockdep annotation, updated the documentation and not add any validation of the assumptions, or just make epoll fit the scheme used by everyone else. So either we can drop these patches, or I need to fix up more of the epoll code.
You could leave the annotation and do something like:
s/ep->lock/ep->wq->lock. And then that would remove the ep->lock saving a bit of space.
Thanks,
-Jason
Quoted 3 lines · Jason Baron
You could leave the annotation and do something like:
s/ep->lock/ep->wq->lock. And then that would remove theep->locksaving a bit of space.
Looks like this isn't going to work due to ep_poll_safewake taking another waitqueue lock. If we had a strict lock order it might work, but the mess in ep_call_nested makes me fear it doesn't.
Quoted 9 lines · Christoph Hellwig
Quoted 3 lines · Jason Baron
You could leave the annotation and do something like:
s/ep->lock/ep->wq->lock. And then that would remove theep->locksaving a bit of space.Looks like this isn't going to work due to
ep_poll_safewaketaking another waitqueue lock. If we had a strict lock order it might work, but the mess inep_call_nestedmakes me fear it doesn't.
hmmm...I'm not sure how this suggestion would change the locking rules from what we currently have. Right now, we use ep->lock, if we remove that and use ep->wq->lock instead, there is just a 1-to-1 mapping there that has not changed, since ep->wq->lock currently is completely not being used.
Thanks,
-Jason
Quoted 5 lines · Jason Baron
hmmm...I'm not sure how this suggestion would change the locking rules from what we currently have. Right now, we use
ep->lock, if we remove that and useep->wq->lockinstead, there is just a 1-to-1 mapping there that has not changed, sinceep->wq->lockcurrently is completely not being used.
True. The patch below survives the amazing complex booting and starting systemd with lockdep enabled test. Do we have something resembling a epoll test suite?
fs/eventpoll.c +22 -24
@@ -182,11 +182,10 @@ struct epitem {* This structure is stored inside the "private_data" member of the file * structure and represents the main data structure for the eventpoll * interface. +* +* Access to it is protected by the lock inside wq. */ struct eventpoll { -/* Protect the access to this structure */ -spinlock_t lock; -/* * This mutex is used to ensure that files are not removed * while epoll is using them. This is held during the event
@@ -210,7 +209,7 @@ struct eventpoll {/* * This is a single linked list that chains all the "struct epitem" that * happened while transferring ready events to userspace w/out -* holding ->lock. +* holding ->wq.lock. */ struct epitem *ovflist;
@@ -686,17 +685,17 @@ static int ep_scan_ready_list(struct eventpoll *ep,* because we want the "sproc" callback to be able to do it * in a lockless way. */ -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);list_splice_init (&ep->rdllist, &txlist); ep->ovflist =NULL ; -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);/* * Now call the callback function. */ error = (*sproc)(ep, &txlist, priv); -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);/* * During the time we spent inside the "sproc" callback, some * other events might have been queued by the poll callback.
@@ -738,7 +737,7 @@ static int ep_scan_ready_list(struct eventpoll *ep,if (waitqueue_active (&ep->poll_wait)) pwake++; } -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);if (!ep_locked)mutex_unlock (&ep->mtx);
@@ -782,10 +781,10 @@ static int ep_remove(struct eventpoll *ep, struct epitem *epi)rb_erase_cached (&epi->rbn, &ep->rbr); -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);if (ep_is_linked (&epi->rdllink))list_del_init (&epi->rdllink); -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);wakeup_source_unregister (ep_wakeup_source (epi));/*
@@ -1015,7 +1014,6 @@ static int ep_alloc(struct eventpoll **pep)if (unlikely (!ep)) goto free_uid; -spin_lock_init (&ep->lock);mutex_init (&ep->mtx);init_waitqueue_head (&ep->wq);init_waitqueue_head (&ep->poll_wait);
@@ -1119,7 +1117,7 @@ static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, vstruct eventpoll *ep = epi->ep;int ewake =0 ; -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);ep_set_busy_poll_napi_id (epi);
@@ -1196,7 +1194,7 @@ static int ep_poll_callback(wait_queue_entry_t *wait, unsigned mode, int sync, v pwake++; out_unlock: -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);/* We have to call this outside the lock */ if (pwake)
@@ -1480,7 +1478,7 @@ static int ep_insert(struct eventpoll *ep, struct epoll_event *event, goto error_remove_epi;/* We have to drop the new item inside our item list to keep track of it */ -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);/* record NAPI ID of new item if present */ ep_set_busy_poll_napi_id (epi);
@@ -1497,7 +1495,7 @@ static int ep_insert(struct eventpoll *ep, struct epoll_event *event, pwake++; } -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);atomic_long_inc (&ep->user->epoll_watches);
@@ -1523,10 +1521,10 @@ static int ep_insert(struct eventpoll *ep, struct epoll_event *event,* list, since that is used/cleaned only inside a section bound by "mtx". * And ep_insert() is called with "mtx" held. */ -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);if (ep_is_linked (&epi->rdllink))list_del_init (&epi->rdllink); -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);wakeup_source_unregister (ep_wakeup_source (epi));
@@ -1593,7 +1591,7 @@ static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_even* list, push it inside. */ if (revents & event->events) { -spin_lock_irq (&ep->lock); +spin_lock_irq (&ep->wq.lock);if (!ep_is_linked (&epi->rdllink)) {list_add_tail (&epi->rdllink, &ep->rdllist);ep_pm_stay_awake (epi);
@@ -1604,7 +1602,7 @@ static int ep_modify(struct eventpoll *ep, struct epitem *epi, struct epoll_evenif (waitqueue_active (&ep->poll_wait)) pwake++; } -spin_unlock_irq (&ep->lock); +spin_unlock_irq (&ep->wq.lock); }/* We have to call this outside the lock */
@@ -1754,7 +1752,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,* caller specified a non blocking operation. */ timed_out =1 ; -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags); goto check_events; }
@@ -1763,7 +1761,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,if (!ep_events_available (ep))ep_busy_loop (ep, timed_out); -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags);if (!ep_events_available (ep)) {/*
@@ -1805,11 +1803,11 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,break ; } -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);if (!schedule_hrtimeout_range (to, slack,HRTIMER_MODE_ABS )) timed_out =1 ; -spin_lock_irqsave (&ep->lock, flags); +spin_lock_irqsave (&ep->wq.lock, flags); }__remove_wait_queue (&ep->wq, &wait);
@@ -1819,7 +1817,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,/* Is it worth to try to dig for events ? */ eavail =ep_events_available (ep); -spin_unlock_irqrestore (&ep->lock, flags); +spin_unlock_irqrestore (&ep->wq.lock, flags);/* * Try to transfer events to user space. In case we get 0 events and
Quoted 11 lines · Christoph Hellwig
Quoted 5 lines · Jason Baron
hmmm...I'm not sure how this suggestion would change the locking rules from what we currently have. Right now, we use
ep->lock, if we remove that and useep->wq->lockinstead, there is just a 1-to-1 mapping there that has not changed, sinceep->wq->lockcurrently is completely not being used.True. The patch below survives the amazing complex booting and starting systemd with lockdep enabled test. Do we have something resembling a epoll test suite?
I don't think we have any in the kernel tree proper (other than some selftests using epoll) but there are tests in ltp and some performance tests such as:
http://linux-scalability.org/epoll/epoll-test.c
http://www.xmailserver.org/linux-patches/pipetest.c
Thanks,
-Jason
Quoted 7 lines · Jason Baron
Quoted 4 lines · Christoph Hellwig
True. The patch below survives the amazing complex booting and starting systemd with lockdep enabled test. Do we have something resembling a epoll test suite?
Quoted 6 lines
I don't think we have any in the kernel tree proper (other than some selftests using epoll) but there are tests in ltp and some performance tests such as:
http://linux-scalability.org/epoll/epoll-test.c
http://www.xmailserver.org/linux-patches/pipetest.c
fyi I'm working on adding epoll to perf-bench, hope to have patches soon.
Thanks,
Davidlohr
Quoted 13 lines · Jason Baron
Quoted 11 lines · Christoph Hellwig
Quoted 5 lines · Jason Baron
hmmm...I'm not sure how this suggestion would change the locking rules from what we currently have. Right now, we use
ep->lock, if we remove that and useep->wq->lockinstead, there is just a 1-to-1 mapping there that has not changed, sinceep->wq->lockcurrently is completely not being used.True. The patch below survives the amazing complex booting and starting systemd with lockdep enabled test. Do we have something resembling a epoll test suite?
Quoted 5 lines
I don't think we have any in the kernel tree proper (other than some selftests using epoll) but there are tests in ltp and some performance tests such as:
That one just seems to keep running until interrupted. I've run it for a while and it seems fine, but I doesn't seem to get any ok/failed kind of status.
Quoted 1 line
Seems to work fine as well, so I'm going to resend the updated patch.