Skip to content

sched: take the lock before checking the wait queue in condvar wake_one/wake_all #1480

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
gburd wants to merge 1 commit into
base: master
Choose a base branch
from
+18 −11
Open

sched: take the lock before checking the wait queue in condvar wake_one/wake_all #1480

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 18 additions & 11 deletions core/condvar.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,13 +101,11 @@ int condvar::wait(mutex* user_mutex, sched::timer* tmr)
void condvar::wake_one()
{
trace_condvar_wake_one(this);
// To make wake with no waiters faster, and avoid unnecessary contention
// in that case, first check the queue head outside the lock. If it is not
// empty, we still need to take the lock, and re-read the head.
if (!_waiters_fifo.oldest) {
return;
}

// Do NOT early-out on an unlocked read of _waiters_fifo.oldest: the waiter
// links its wait_record under _m then releases _m, and this reader does not
// take _m for the pre-check, so there is no acquire pairing that release --
// the just-linked record can read as a stale nullptr and the wakeup is
// dropped (see the fuller note in wake_all()). Take _m first, then check.
_m.lock();
wait_record *wr = _waiters_fifo.oldest;
if (wr) {
Expand All @@ -131,12 +129,21 @@ void condvar::wake_one()
void condvar::wake_all()
{
trace_condvar_wake_all(this);
if (!_waiters_fifo.oldest) {
return;
}

// Do NOT early-out on an unlocked read of _waiters_fifo.oldest: a waiter
// links its wait_record under _m and then releases _m, but this reader does
// not take _m for the check, so there is no acquire to pair with that
// release and the just-linked record can be read as a stale nullptr --
// dropping the wakeup. Under CONF_fork the waiter may be a forked child
// whose wait_record lives in another address space (coherent_wait_record),
// widening the window. A sparse cross-address-space broadcast (e.g. a ZFS
// zio completion signalling a forked backend blocked in zio_wait) then
// wakes nobody and the backend hangs forever. Take _m first, then check.
_m.lock();
wait_record *wr = _waiters_fifo.oldest;
if (!wr) {
_m.unlock();
return;
}

// To help the assert() in condvar_wait(), we need to zero saved
// user_mutex when all concurrent condvar_wait()s are done.
Expand Down