Re: [PATCH V2 7/9] vhost: do not use RCU to synchronize MMU notifier with worker

From: Jason Wang
Date: Thu Aug 01 2019 - 01:02:41 EST



On 2019/8/1 äå3:30, Jason Gunthorpe wrote:
On Wed, Jul 31, 2019 at 09:28:20PM +0800, Jason Wang wrote:
On 2019/7/31 äå8:39, 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.


Yes, so we have two mappings:

[1] vring address to VA
[2] VA to PA

And have several readers and writers

1) set_vring_num_addr(): writer of both [1] and [2]
2) MMU notifier: reader of [1] writer of [2]
3) GUP: reader of [1] writer of [2]
4) memory accessors: reader of [1] and [2]

Fortunately, 1) 3) and 4) have already synchronized through vq->mutex. We only need to deal with synchronization between 2) and each of the reset:
Sync between 1) and 2): For mapping [1], I do mmu_notifier_unregister/register. This help to avoid holding any lock to do overlap check. Anyway we only care about one or three pages , but the whole guest memory could be several TBs. For mapping [2], both 1) and 2) are writers, so use spinlock (mmu_lock) to synchronize.
Sync between 2) and 3): For mapping [1], both are readers, no need any synchronization. For mapping [2], both 2) and 3) are writers, so synchronize through spinlock (mmu_lock);
Sync between 2) and 4): For mapping [1], both are readers, no need any synchronization. For mapping [2], synchronize through RCU (or something simliar to seqlock).

You suggestion is about the synchronization of [1] which may make sense, but it could be done on top as an optimization. What this path tries to do is to not use RCU for [2]. Of course, the simplest way is to use vq mutex in 2) but it means:
- we must hold vq lock to check range overlap
- since the critical section was increased, the worst case is to wait guest memory to be swapped in, this could be even slower than synchronize_rcu().



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.


Technically yes, but we probably can't afford that for vhost fast path, the atomics eliminate almost all the performance improvement brought by this patch on a machine without SMAP.



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.


Yes. Having a hard thought of the code, it looks to me write_seqcount_begin()/end() is sufficient here:

- Notifier will only assign NULL to map, so it doesn't harm to read map before seq, then we will fallback to normal copy_from/to_user() slow path earlier
- if we write through map->addr, it should be done before increasing the seqcount because of the smp_wmb() in write_seqcount_end()
- if we read through map->addr which also contain a store to a pointer, we have a good data dependency so smp_wmb() also work here.



The requirement I see is you need invalidate_range_start to block
until another thread exits its critical section (ie stops accessing
the SPTEs).


Yes.



That is a spinlock/mutex.


Or a semantics similar to RCU.



You just can't invent a faster spinlock by open coding something with
barriers, it doesn't work.

Jason


If write_seqlock() works here, we can simply wait for seqcount to move advance in MMU notifier. The original idea is to use RCU which solves this perfectly. But as pointed out it could be slow.

Thanks