Hello,
On Fri, May 06, 2022 at 09:50:30PM +0200, Alexander Bluhm wrote:
> Hi,
>
> When creating network load by forwarding packets, SSH gets unusable
> and ping time gets above 10 seconds.
>
> Problem is that while multiple forwarding threads are running with
> shared net lock, the exclusive lock cannot be acquired. This is
> unfair.
>
> Diff below prevents that a read lock is granted when another thread
> is waiting for the exclusive lock. With that ping time stays under
> 300 ms.
>
> Does this read write lock prio change make sense?
I can confirm this diff helps. And if I understand things right,
then we basically let all readers to take a slow path via a sleep,
if there is a writer waiting (think of like putting readers
behind the waiting writer in queue).
I was afraid the change can trade writer starvation for reader
starvation. However I think it is not the case after seeing
rw_do_exit(), where we zero out lock after writer does rw_exit().
this basically gives equal chance to all readers/writers to acquire
the lock.
not my department, however I would say change looks good to me.
OK sashan@ (given my OK counts in this area)
>
> bluhm
>
> Index: kern/kern_rwlock.c
> ===================================================================
> RCS file: /data/mirror/openbsd/cvs/src/sys/kern/kern_rwlock.c,v
> retrieving revision 1.47
> diff -u -p -r1.47 kern_rwlock.c
> --- kern/kern_rwlock.c 8 Feb 2021 08:18:45 -0000 1.47
> +++ kern/kern_rwlock.c 6 May 2022 12:08:01 -0000
> @@ -81,7 +81,7 @@ static const struct rwlock_op {
> },
> { /* RW_READ */
> RWLOCK_READ_INCR,
> - RWLOCK_WRLOCK,
> + RWLOCK_WRLOCK | RWLOCK_WRWANT,
> RWLOCK_WAIT,
> 0,
> PLOCK
> @@ -103,7 +103,7 @@ rw_enter_read(struct rwlock *rwl)
> {
> unsigned long owner = rwl->rwl_owner;
>
> - if (__predict_false((owner & RWLOCK_WRLOCK) ||
> + if (__predict_false((owner & (RWLOCK_WRLOCK | RWLOCK_WRWANT)) ||
> rw_cas(&rwl->rwl_owner, owner, owner + RWLOCK_READ_INCR)))
> rw_enter(rwl, RW_READ);
> else {
>
