On Wed 11-05-16 10:35:12, Peter Zijlstra wrote:
> On Wed, May 11, 2016 at 09:23:57AM +0200, Michal Hocko wrote:
> > On Tue 10-05-16 14:38:06, Peter Zijlstra wrote:
>
> > > Also, looking at it again; I think we're forgetting to re-adjust the
> > > BIAS for the cancelled writer.
> >
> > Hmm, __rwsem_down_write_failed_common does
> >
> > /* undo write bias from down_write operation, stop active locking */
> > count = rwsem_atomic_update(-RWSEM_ACTIVE_WRITE_BIAS, sem);
> >
> > which should remove the bias AFAIU.
>
> Right; at this point we're neutral wrt bias.
>
> > Later we do
> >
> > if (waiting) {
> > count = READ_ONCE(sem->count);
> >
> > /*
> > * If there were already threads queued before us and there are
> > * no active writers, the lock must be read owned; so we try to
> > * wake any read locks that were queued ahead of us.
> > */
> > if (count > RWSEM_WAITING_BIAS)
> > sem = __rwsem_do_wake(sem, RWSEM_WAKE_READERS);
> >
> > } else
> > count = rwsem_atomic_update(RWSEM_WAITING_BIAS, sem);
> >
> > and that might set RWSEM_WAITING_BIAS but the current holder of the lock
> > should handle that correctly and wake the waiting tasks IIUC. I will go
> > and check the code closer. It is quite easy to get this subtle code
> > wrong..
>
> Subtle; yes.
>
> So if you look at rwsem_try_write_lock() -- traditionally the only way
> to exit this wait loop, you see it does:
>
> if (count == RWSEM_WAITING_BIAS &&
> cmpxchg_acquire(&sem->count, RWSEM_WAITING_BIAS,
> RWSEM_ACTIVE_WRITE_BIAS) == RWSEM_WAITING_BIAS) {
> if (!list_is_singular(&sem->wait_list))
> rwsem_atomic_update(RWSEM_WAITING_BIAS, sem);
> rwsem_set_owner(sem);
> return true;
> }
>
> Which ends up clearing RWSEM_WAITING_BIAS is we were the only waiter --
> or rather, it always clear WAITING, but then tests the list and re-sets
> it if there's more than one waiters on.
>
> Now, the signal break doesn't clear WAITING if we were the only waiter
> on the list; which means any further down_read() will block (I didn't
> look at what a subsequent down_write() would do).
I was staring at this part as well but then I convinced myself that this
is OK because rwsem_down_read_failed does:
if (list_empty(&sem->wait_list))
adjustment += RWSEM_WAITING_BIAS;
list_add_tail(&waiter.list, &sem->wait_list);
/* we're now waiting on the lock, but no longer actively locking */
count = rwsem_atomic_update(adjustment, sem);
/* If there are no active locks, wake the front queued process(es).
*
* If there are no writers and we are first in the queue,
* wake our own waiter to join the existing active readers !
*/
if (count == RWSEM_WAITING_BIAS ||
(count > RWSEM_WAITING_BIAS &&
adjustment != -RWSEM_ACTIVE_READ_BIAS))
sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
__rwsem_do_wake should then see all the readers (including the current
one) and wake them up and set waiter->task to NULL to allow the current
one to break out of the loop as well.
> So I think we needs something like this, to clear WAITING if we leave
> the list empty.
But maybe this is the correct way to go.
> Does that make sense?
I do not see an immediate problem with this. Maybe Tetsuo can try this
out and see if it really makes a difference.
> diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
> index df4dcb883b50..7011dd1c286c 100644
> --- a/kernel/locking/rwsem-xadd.c
> +++ b/kernel/locking/rwsem-xadd.c
> @@ -489,6 +489,8 @@ __rwsem_down_write_failed_common(struct rw_semaphore
> *sem, int state)
> do {
> if (signal_pending_state(state, current)) {
> raw_spin_lock_irq(&sem->wait_lock);
> + if (list_singular(&sem->wait_list))
> +
> rwsem_atomic_update(-RWSEM_WAITING_BIAS, sem);
> ret = ERR_PTR(-EINTR);
> goto out;
> }
--
Michal Hocko
SUSE Labs
