From: Stefano Stabellini <sstabellini@kernel.org>
To: Julien Grall <julien@xen.org>
Cc: Stefano Stabellini <sstabellini@kernel.org>,
xen-devel@lists.xenproject.org, bertrand.marquis@arm.com,
ash.j.wilding@gmail.com, Julien Grall <jgrall@amazon.com>,
Volodymyr Babchuk <Volodymyr_Babchuk@epam.com>,
Dario Faggioli <dfaggioli@suse.com>,
George Dunlap <george.dunlap@citrix.com>,
Jan Beulich <jbeulich@suse.com>,
Andrew Cooper <andrew.cooper3@citrix.com>
Subject: Re: [PATCH] xen/arm: Ensure the vCPU context is seen before clearing the _VPF_down
Date: Tue, 20 Apr 2021 17:38:31 -0700 (PDT) [thread overview]
Message-ID: <alpine.DEB.2.21.2104201726061.26180@sstabellini-ThinkPad-T480s> (raw)
In-Reply-To: <37631386-a53f-d99d-d71b-0b871b5dd9b0@xen.org>
On Tue, 20 Apr 2021, Julien Grall wrote:
> (+ Andrew and Jan)
>
> Hi Stefano,
>
> On 20/04/2021 20:38, Stefano Stabellini wrote:
> > On Fri, 16 Apr 2021, Julien Grall wrote:
> > > > I think your explanation is correct. However, don't we need a smp_rmb()
> > > > barrier after reading v->is_initialised in xen/common/domctl.c:do_domctl
> > > > ? That would be the barrier that pairs with smp_wmb in regards to
> > > > v->is_initialised.
> > >
> > > There is already a smp_mb() in sync_vcpu_exec_state() which is called from
> > > vcpu_pause() -> vcpu_sleep_sync().
> >
> > But that's too late, isn't? The v->is_initialized check is done before
> > the vcpu_pause() call. We might end up taking the wrong code path:
> >
> > https://gitlab.com/xen-project/xen/-/blob/staging/xen/common/domctl.c#L587
> > https://gitlab.com/xen-project/xen/-/blob/staging/xen/common/domctl.c#L598
>
> I am a bit confused what you mean by "wrong path" here. There is no timing
> guarantee with a memory barrier. What the barrier will guarantee you is
> v->is_initialized will be read *before* anything after the barrier.
>
> Are you worried that some variables in vcpu_pause() may be read before
> v->is_initialized?
>
> >
> > > I don't think we can ever remove the memory barrier in
> > > sync_vcpu_exec_state()
> > > because the vCPU paused may have run (or initialized) on a different pCPU.
> > > So
> > > I would like to rely on the barrier rather than adding an extra one (even
> > > thought it is not a fast path).
> > >
> > > I am thinking to add a comment on top of vcpu_pause() to clarify that
> > > after
> > > the call, the vCPU context will be observed without extra synchronization
> > > required.
> >
> > Given that an if.. goto is involved, even if both code paths lead to
> > good results, I think it would be best to have the smp_rmb() barrier
> > above after the first v->is_initialised read to make sure we take the
> > right path.
>
> I don't understand what you are referring by "wrong" and "right" path. The
> processor will always execute/commit the path based on the value of
> v->is_initialized. What may happen is the processor may start to speculate the
> wrong path and then backtrack if it discovers the value is not the expected
> one. But, AFAIK, smp_rmb() is not going to protect you against speculation...
> smp_rmb() is only going to enforce a memory ordering between read.
>
> > Instead of having to make sure that even the "wrong" path
> > behaves correctly.
> Just for clarification, I think you meant writing the following code:
>
> if ( !v->is_initialized )
> goto getvcpucontext_out;
>
> smp_rmb();
No, sorry, I'll try to be clearer, see below
> ...
>
> vcpu_pause();
>
> AFAICT, there is nothing in the implementation of XEN_DOMCTL_getvcpucontext
> that justify the extra barrier (assuming we consider vcpu_pause() as a full
> memory barrier).
>
> From your e-mail, I also could not infer what is your exact concern regarding
> the memory ordering. If you have something in mind, then would you mind to
> provide a sketch showing what could go wrong?
Let me start with what I had in mind:
initialized = v->is_initialized;
smp_rmb();
if ( !initialized )
goto getvcpucontext_out;
Without smp_rmb there are no guarantees that we see an up-to-date value
of v->is_initialized, right? This READ of v->is_initialized is supposed
to pair with the WRITE in arch_set_info_guest.
Relying on the the barrier in vcpu_pause is OK only if is_initialised
was already read as "1".
I think it might be OK in this case, because probably nothing wrong
happens if we read the old value of v->is_initialized as "0" but it
doesn't seem to be a good idea. Theoretically it could pass a very long
time before we see v->is_initialized as "1" without the smp_rmb.
Please let me know if I am missing something.
next prev parent reply other threads:[~2021-04-21 0:39 UTC|newest]
Thread overview: 13+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-02-26 20:51 [PATCH] xen/arm: Ensure the vCPU context is seen before clearing the _VPF_down Julien Grall
2021-02-27 1:58 ` Stefano Stabellini
2021-02-27 14:30 ` Julien Grall
2021-03-20 0:01 ` Stefano Stabellini
2021-03-20 11:47 ` Julien Grall
2021-04-01 15:09 ` Julien Grall
2021-04-13 22:43 ` Stefano Stabellini
2021-04-16 18:21 ` Julien Grall
2021-04-20 19:38 ` Stefano Stabellini
2021-04-20 20:47 ` Julien Grall
2021-04-21 0:38 ` Stefano Stabellini [this message]
2021-04-21 12:33 ` Julien Grall
2021-04-22 20:33 ` Stefano Stabellini
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=alpine.DEB.2.21.2104201726061.26180@sstabellini-ThinkPad-T480s \
--to=sstabellini@kernel.org \
--cc=Volodymyr_Babchuk@epam.com \
--cc=andrew.cooper3@citrix.com \
--cc=ash.j.wilding@gmail.com \
--cc=bertrand.marquis@arm.com \
--cc=dfaggioli@suse.com \
--cc=george.dunlap@citrix.com \
--cc=jbeulich@suse.com \
--cc=jgrall@amazon.com \
--cc=julien@xen.org \
--cc=xen-devel@lists.xenproject.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).