[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-block] [PATCH 14/17] block: optimize access to reqs_lock

From: Stefan Hajnoczi
Subject: Re: [Qemu-block] [PATCH 14/17] block: optimize access to reqs_lock
Date: Fri, 5 May 2017 11:25:50 +0100
User-agent: Mutt/1.8.0 (2017-02-23)

On Thu, May 04, 2017 at 06:06:39PM +0200, Paolo Bonzini wrote:
> On 04/05/2017 16:59, Stefan Hajnoczi wrote:
> > On Thu, Apr 20, 2017 at 02:00:55PM +0200, Paolo Bonzini wrote:
> >> Hot path reqs_lock critical sections are very small; the only large 
> >> critical
> >> sections happen when a request waits for serialising requests, and these
> >> should never happen in usual circumstances.
> >>
> >> We do not want these small critical sections to yield in any case,
> >> which calls for using a spinlock while writing the list.
> > 
> > Is this patch purely an optimization?
> Yes, it is, and pretty much a no-op until we have true multiqueue.  But
> I expect it to have a significant effect for multiqueue.
> > I'm hesitant about using spinlocks in userspace.  There are cases where
> > the thread is descheduled that are beyond our control.  Nested virt will
> > probably make things worse.  People have been optimizing and trying
> > paravirt approaches to kernel spinlocks for these reasons for years.
> This is true, but here we're talking about a 5-10 instruction window for
> preemption; it matches the usage of spinlocks in other parts of QEMU.

Only util/qht.c uses spinlocks, it's not a widely used primitive.

> The long critical sections, which only happen with combination with
> copy-on-read or RMW (large logical block sizes on the host), take the
> CoMutex.
> On one hand it's true that the more you nest, the more things get worse.
>  On the other hand there can only ever be contention with multiqueue,
> and the multiqueue scenarios are going to use pinning.
> > Isn't a futex-based lock efficient enough?  That way we don't hog the
> > CPU when there is contention.
> It is efficient when there is no contention, but when there is, the
> latency goes up by several orders of magnitude.

Doesn't glibc spin for a while before waiting on the futex?  i.e. the
best of both worlds.


Attachment: signature.asc
Description: PGP signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]