On Tue, Jul 30, 2019 at 03:44:47PM +0800, Jason Wang wrote:
>
> On 2019/7/29 下午10:44, Michael S. Tsirkin wrote:
> > On Mon, Jul 29, 2019 at 10:24:43PM +0800, Jason Wang wrote:
> > > On 2019/7/29 下午4:59, Michael S. Tsirkin wrote:
> > > > On Mon, Jul 29, 2019 at 01:54:49PM +0800, Jason Wang wrote:
> > > > > On 2019/7/26 下午9:49, Michael S. Tsirkin wrote:
> > > > > > > > Ok, let me retry if necessary (but I do remember I end up with deadlocks
> > > > > > > > last try).
> > > > > > > Ok, I play a little with this. And it works so far. Will do more testing
> > > > > > > tomorrow.
> > > > > > >
> > > > > > > One reason could be I switch to use get_user_pages_fast() to
> > > > > > > __get_user_pages_fast() which doesn't need mmap_sem.
> > > > > > >
> > > > > > > Thanks
> > > > > > OK that sounds good. If we also set a flag to make
> > > > > > vhost_exceeds_weight exit, then I think it will be all good.
> > > > > After some experiments, I came up two methods:
> > > > >
> > > > > 1) switch to use vq->mutex, then we must take the vq lock during range
> > > > > checking (but I don't see obvious slowdown for 16vcpus + 16queues). Setting
> > > > > flags during weight check should work but it still can't address the worst
> > > > > case: wait for the page to be swapped in. Is this acceptable?
> > > > >
> > > > > 2) using current RCU but replace synchronize_rcu() with vhost_work_flush().
> > > > > The worst case is the same as 1) but we can check range without holding any
> > > > > locks.
> > > > >
> > > > > Which one did you prefer?
> > > > >
> > > > > Thanks
> > > > I would rather we start with 1 and switch to 2 after we
> > > > can show some gain.
> > > >
> > > > But the worst case needs to be addressed.
> > >
> > > Yes.
> > >
> > >
> > > > How about sending a signal to
> > > > the vhost thread? We will need to fix up error handling (I think that
> > > > at the moment it will error out in that case, handling this as EFAULT -
> > > > and we don't want to drop packets if we can help it, and surely not
> > > > enter any error states. In particular it might be especially tricky if
> > > > we wrote into userspace memory and are now trying to log the write.
> > > > I guess we can disable the optimization if log is enabled?).
> > >
> > > This may work but requires a lot of changes.
> > I agree.
> >
> > > And actually it's the price of
> > > using vq mutex.
> > Not sure what's meant here.
>
>
> I mean if we use vq mutex, it means the critical section was increased and
> we need to deal with swapping then.
>
>
> >
> > > Actually, the critical section should be rather small, e.g
> > > just inside memory accessors.
> > Also true.
> >
> > > I wonder whether or not just do synchronize our self like:
> > >
> > > static void inline vhost_inc_vq_ref(struct vhost_virtqueue *vq)
> > > {
> > > int ref = READ_ONCE(vq->ref);
> > >
> > > WRITE_ONCE(vq->ref, ref + 1);
> > > smp_rmb();
> > > }
> > >
> > > static void inline vhost_dec_vq_ref(struct vhost_virtqueue *vq)
> > > {
> > > int ref = READ_ONCE(vq->ref);
> > >
> > > smp_wmb();
> > > WRITE_ONCE(vq->ref, ref - 1);
> > > }
> > >
> > > static void inline vhost_wait_for_ref(struct vhost_virtqueue *vq)
> > > {
> > > while (READ_ONCE(vq->ref));
> > > mb();
> > > }
> > Looks good but I'd like to think of a strategy/existing lock that let us
> > block properly as opposed to spinning, that would be more friendly to
> > e.g. the realtime patch.
>
>
> Does it make sense to disable preemption in the critical section? Then we
> don't need to block and we have a deterministic time spent on memory
> accssors?
Hmm maybe. I'm getting really nervious at this point - we
seem to be using every trick in the book.
>
> >
> > > Or using smp_load_acquire()/smp_store_release() instead?
> > >
> > > Thanks
> > These are cheaper on x86, yes.
>
>
> Will use this.
>
> Thanks
>
>
This looks suspiciously like a seqlock though.
Can that be used somehow?
