On Fri, Jul 26, 2019 at 10:00:20PM +0800, Jason Wang wrote:
>
> On 2019/7/26 下午9:47, Michael S. Tsirkin wrote:
> > On Fri, Jul 26, 2019 at 08:53:18PM +0800, Jason Wang wrote:
> > > On 2019/7/26 下午8:38, Michael S. Tsirkin wrote:
> > > > On Fri, Jul 26, 2019 at 08:00:58PM +0800, Jason Wang wrote:
> > > > > On 2019/7/26 下午7:49, Michael S. Tsirkin wrote:
> > > > > > On Thu, Jul 25, 2019 at 10:25:25PM +0800, Jason Wang wrote:
> > > > > > > On 2019/7/25 下午9:26, Michael S. Tsirkin wrote:
> > > > > > > > > Exactly, and that's the reason actually I use synchronize_rcu() there.
> > > > > > > > >
> > > > > > > > > So the concern is still the possible synchronize_expedited()?
> > > > > > > > I think synchronize_srcu_expedited.
> > > > > > > >
> > > > > > > > synchronize_expedited sends lots of IPI and is bad for realtime VMs.
> > > > > > > >
> > > > > > > > > Can I do this
> > > > > > > > > on through another series on top of the incoming V2?
> > > > > > > > >
> > > > > > > > > Thanks
> > > > > > > > >
> > > > > > > > The question is this: is this still a gain if we switch to the
> > > > > > > > more expensive srcu? If yes then we can keep the feature on,
> > > > > > > I think we only care about the cost on srcu_read_lock() which looks pretty
> > > > > > > tiny form my point of view. Which is basically a READ_ONCE() + WRITE_ONCE().
> > > > > > >
> > > > > > > Of course I can benchmark to see the difference.
> > > > > > >
> > > > > > >
> > > > > > > > if not we'll put it off until next release and think
> > > > > > > > of better solutions. rcu->srcu is just a find and replace,
> > > > > > > > don't see why we need to defer that. can be a separate patch
> > > > > > > > for sure, but we need to know how well it works.
> > > > > > > I think I get here, let me try to do that in V2 and let's see the numbers.
> > > > > > >
> > > > > > > Thanks
> > > > > It looks to me for tree rcu, its srcu_read_lock() have a mb() which is too
> > > > > expensive for us.
> > > > I will try to ponder using vq lock in some way.
> > > > Maybe with trylock somehow ...
> > >
> > > Ok, let me retry if necessary (but I do remember I end up with deadlocks
> > > last try).
> > >
> > >
> > > >
> > > > > If we just worry about the IPI,
> > > > With synchronize_rcu what I would worry about is that guest is stalled
> > >
> > > Can this synchronize_rcu() be triggered by guest? If yes, there are several
> > > other MMU notifiers that can block. Is vhost something special here?
> > Sorry, let me explain: guests (and tasks in general)
> > can trigger activity that will
> > make synchronize_rcu take a long time.
>
>
> Yes, I get this.
>
>
> > Thus blocking
> > an mmu notifier until synchronize_rcu finishes
> > is a bad idea.
>
>
> The question is, MMU notifier are allowed to be blocked on
> invalidate_range_start() which could be much slower than synchronize_rcu()
> to finish.
>
> Looking at amdgpu_mn_invalidate_range_start_gfx() which calls
> amdgpu_mn_invalidate_node() which did:
>
> r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
> true, false, MAX_SCHEDULE_TIMEOUT);
>
> ...
>
Right. And the result will probably be VMs freezing/timing out, too.
It's just that we care about VMs more than the GPU guys :)
> > > > because system is busy because of other guests.
> > > > With expedited it's the IPIs...
> > > >
> > > The current synchronize_rcu() can force a expedited grace period:
> > >
> > > void synchronize_rcu(void)
> > > {
> > > ...
> > > if (rcu_blocking_is_gp())
> > > return;
> > > if (rcu_gp_is_expedited())
> > > synchronize_rcu_expedited();
> > > else
> > > wait_rcu_gp(call_rcu);
> > > }
> > > EXPORT_SYMBOL_GPL(synchronize_rcu);
> >
> > An admin can force rcu to finish faster, trading
> > interrupts for responsiveness.
>
>
> Yes, so when set, all each synchronize_rcu() will go for
> synchronize_rcu_expedited().
And that's bad for realtime things. I understand what you are saying,
host admin can set this and VMs won't time-out. What I'm saying is we
should not make admins choose between two types of bugs. Tuning for
performance is fine.
>
> >
> > > > > can we do something like in
> > > > > vhost_invalidate_vq_start()?
> > > > >
> > > > > if (map) {
> > > > > /* In order to avoid possible IPIs with
> > > > > * synchronize_rcu_expedited() we use call_rcu() +
> > > > > * completion.
> > > > > */
> > > > > init_completion(&c.completion);
> > > > > call_rcu(&c.rcu_head, vhost_finish_vq_invalidation);
> > > > > wait_for_completion(&c.completion);
> > > > > vhost_set_map_dirty(vq, map, index);
> > > > > vhost_map_unprefetch(map);
> > > > > }
> > > > >
> > > > > ?
> > > > Why would that be faster than synchronize_rcu?
> > >
> > > No faster but no IPI.
> > >
> > Sorry I still don't see the point.
> > synchronize_rcu doesn't normally do an IPI either.
> >
>
> Not the case of when rcu_expedited is set. This can just 100% make sure
> there's no IPI.
Right but then the latency can be pretty big.
>
> > > >
> > > > > > There's one other thing that bothers me, and that is that
> > > > > > for large rings which are not physically contiguous
> > > > > > we don't implement the optimization.
> > > > > >
> > > > > > For sure, that can wait, but I think eventually we should
> > > > > > vmap large rings.
> > > > > Yes, worth to try. But using direct map has its own advantage: it can use
> > > > > hugepage that vmap can't
> > > > >
> > > > > Thanks
> > > > Sure, so we can do that for small rings.
> > >
> > > Yes, that's possible but should be done on top.
> > >
> > > Thanks
> > Absolutely. Need to fix up the bugs first.
> >
>
> Yes.
>
> Thanks
