From: Jason Gunthorpe <jgg@ziepe.ca>
To: Jason Wang <jasowang@redhat.com>
Cc: mst@redhat.com, kvm@vger.kernel.org,
virtualization@lists.linux-foundation.org,
netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
linux-mm@kvack.org
Subject: Re: [PATCH V2 7/9] vhost: do not use RCU to synchronize MMU notifier with worker
Date: Wed, 31 Jul 2019 16:30:57 -0300 [thread overview]
Message-ID: <20190731193057.GG3946@ziepe.ca> (raw)
In-Reply-To: <7555c949-ae6f-f105-6e1d-df21ddae9e4e@redhat.com>
On Wed, Jul 31, 2019 at 09:28:20PM +0800, Jason Wang wrote:
>
> On 2019/7/31 下午8:39, Jason Gunthorpe wrote:
> > On Wed, Jul 31, 2019 at 04:46:53AM -0400, Jason Wang wrote:
> > > We used to use RCU to synchronize MMU notifier with worker. This leads
> > > calling synchronize_rcu() in invalidate_range_start(). But on a busy
> > > system, there would be many factors that may slow down the
> > > synchronize_rcu() which makes it unsuitable to be called in MMU
> > > notifier.
> > >
> > > A solution is SRCU but its overhead is obvious with the expensive full
> > > memory barrier. Another choice is to use seqlock, but it doesn't
> > > provide a synchronization method between readers and writers. The last
> > > choice is to use vq mutex, but it need to deal with the worst case
> > > that MMU notifier must be blocked and wait for the finish of swap in.
> > >
> > > So this patch switches use a counter to track whether or not the map
> > > was used. The counter was increased when vq try to start or finish
> > > uses the map. This means, when it was even, we're sure there's no
> > > readers and MMU notifier is synchronized. When it was odd, it means
> > > there's a reader we need to wait it to be even again then we are
> > > synchronized.
> > You just described a seqlock.
>
>
> Kind of, see my explanation below.
>
>
> >
> > We've been talking about providing this as some core service from mmu
> > notifiers because nearly every use of this API needs it.
>
>
> That would be very helpful.
>
>
> >
> > IMHO this gets the whole thing backwards, the common pattern is to
> > protect the 'shadow pte' data with a seqlock (usually open coded),
> > such that the mmu notififer side has the write side of that lock and
> > the read side is consumed by the thread accessing or updating the SPTE.
>
>
> Yes, I've considered something like that. But the problem is, mmu notifier
> (writer) need to wait for the vhost worker to finish the read before it can
> do things like setting dirty pages and unmapping page. It looks to me
> seqlock doesn't provide things like this.
The seqlock is usually used to prevent a 2nd thread from accessing the
VA while it is being changed by the mm. ie you use something seqlocky
instead of the ugly mmu_notifier_unregister/register cycle.
You are supposed to use something simple like a spinlock or mutex
inside the invalidate_range_start to serialized tear down of the SPTEs
with their accessors.
> write_seqcount_begin()
>
> map = vq->map[X]
>
> write or read through map->addr directly
>
> write_seqcount_end()
>
>
> There's no rmb() in write_seqcount_begin(), so map could be read before
> write_seqcount_begin(), but it looks to me now that this doesn't harm at
> all, maybe we can try this way.
That is because it is a write side lock, not a read lock. IIRC
seqlocks have weaker barriers because the write side needs to be
serialized in some other way.
The requirement I see is you need invalidate_range_start to block
until another thread exits its critical section (ie stops accessing
the SPTEs).
That is a spinlock/mutex.
You just can't invent a faster spinlock by open coding something with
barriers, it doesn't work.
Jason
next prev parent reply other threads:[~2019-07-31 19:31 UTC|newest]
Thread overview: 51+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-07-31 8:46 [PATCH V2 0/9] Fixes for metadata accelreation Jason Wang
2019-07-31 8:46 ` [PATCH V2 1/9] vhost: don't set uaddr for invalid address Jason Wang
2019-07-31 8:46 ` [PATCH V2 2/9] vhost: validate MMU notifier registration Jason Wang
2019-07-31 8:46 ` [PATCH V2 3/9] vhost: fix vhost map leak Jason Wang
2019-07-31 8:46 ` [PATCH V2 4/9] vhost: reset invalidate_count in vhost_set_vring_num_addr() Jason Wang
2019-07-31 12:41 ` Jason Gunthorpe
2019-07-31 13:29 ` Jason Wang
2019-07-31 19:32 ` Jason Gunthorpe
2019-07-31 19:37 ` Michael S. Tsirkin
2019-08-01 5:03 ` Jason Wang
2019-07-31 8:46 ` [PATCH V2 5/9] vhost: mark dirty pages during map uninit Jason Wang
2019-07-31 8:46 ` [PATCH V2 6/9] vhost: don't do synchronize_rcu() in vhost_uninit_vq_maps() Jason Wang
2019-07-31 8:46 ` [PATCH V2 7/9] vhost: do not use RCU to synchronize MMU notifier with worker Jason Wang
2019-07-31 8:50 ` Jason Wang
2019-07-31 12:39 ` Jason Gunthorpe
2019-07-31 13:28 ` Jason Wang
2019-07-31 19:30 ` Jason Gunthorpe [this message]
2019-08-01 5:02 ` Jason Wang
2019-08-01 14:15 ` Jason Gunthorpe
2019-08-02 9:40 ` Jason Wang
2019-08-02 12:46 ` Jason Gunthorpe
2019-08-02 14:27 ` Michael S. Tsirkin
2019-08-02 17:24 ` Jason Gunthorpe
2019-08-03 21:36 ` Michael S. Tsirkin
2019-08-04 0:14 ` Jason Gunthorpe
2019-08-04 8:07 ` Michael S. Tsirkin
2019-08-05 4:39 ` Jason Wang
2019-08-06 11:53 ` Jason Gunthorpe
2019-08-06 13:36 ` Michael S. Tsirkin
2019-08-06 13:40 ` Jason Gunthorpe
2019-08-05 4:36 ` Jason Wang
2019-08-05 4:41 ` Jason Wang
2019-08-05 6:40 ` Michael S. Tsirkin
2019-08-05 8:24 ` Jason Wang
2019-08-05 6:30 ` Michael S. Tsirkin
2019-08-05 8:22 ` Jason Wang
2019-08-05 4:20 ` Jason Wang
2019-08-06 12:04 ` Jason Gunthorpe
2019-08-07 6:49 ` Jason Wang
2019-08-02 14:03 ` Michael S. Tsirkin
2019-08-05 4:33 ` Jason Wang
2019-08-05 6:28 ` Michael S. Tsirkin
2019-08-05 8:21 ` Jason Wang
2019-07-31 18:29 ` Michael S. Tsirkin
2019-08-01 8:06 ` Jason Wang
2019-08-03 21:54 ` Michael S. Tsirkin
2019-08-05 8:18 ` Jason Wang
2019-07-31 8:46 ` [PATCH V2 8/9] vhost: correctly set dirty pages in MMU notifiers callback Jason Wang
2019-07-31 8:46 ` [PATCH V2 9/9] vhost: do not return -EAGIAN for non blocking invalidation too early Jason Wang
2019-07-31 9:59 ` Stefano Garzarella
2019-07-31 10:05 ` Jason Wang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20190731193057.GG3946@ziepe.ca \
--to=jgg@ziepe.ca \
--cc=jasowang@redhat.com \
--cc=kvm@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=mst@redhat.com \
--cc=netdev@vger.kernel.org \
--cc=virtualization@lists.linux-foundation.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).