Re: [Qemu-devel] [Xen-devel] new barrier type for paravirt (was Re: [PATCH] virtio_ring: use smp_store_mb)

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 20/12/15 09:25, Michael S. Tsirkin wrote:
> On Thu, Dec 17, 2015 at 03:39:10PM +0100, Peter Zijlstra wrote:
>> On Thu, Dec 17, 2015 at 04:33:44PM +0200, Michael S. Tsirkin wrote:
>>> On Thu, Dec 17, 2015 at 02:57:26PM +0100, Peter Zijlstra wrote:
>>>> You could of course go fix that instead of mutilating things into
>>>> sort-of functional state.
>>> Yes, we'd just need to touch all architectures, all for
>>> the sake of UP which almost no one uses.
>>> Basically, we need APIs that explicitly are
>>> for talking to another kernel on a different CPU on
>>> the same SMP system, and implemented identically
>>> between CONFIG_SMP and !CONFIG_SMP on all architectures.
>>>
>>> Do you think this is something of general usefulness,
>>> outside virtio?
>> I'm not aware of any other case, but if there are more parts of virt
>> that need this then I see no problem adding it.
> When I wrote this, I assumed there are no other users, and I'm still not
> sure there are other users at the moment. Do you see a problem then?
>
>> That is, virt in general is the only use-case that I can think of,
>> because this really is an artifact of interfacing with an SMP host while
>> running an UP kernel.
> Or another guest kernel on an SMP host.
>
>> But I'm really not familiar with virt, so I do not know if there's more
>> sites outside of virtio that could use this.
>> Touching all archs is a tad tedious, but its fairly straight forward.
> So I looked and I was only able to find other another possible user in Xen.
>
> Cc Xen folks.
>
> I noticed that drivers/xen/xenbus/xenbus_comms.c uses
> full memory barriers to communicate with the other side.
> For example:
>
>                 /* Must write data /after/ reading the consumer index.  * */
>                 mb();
>
>                 memcpy(dst, data, avail);
>                 data += avail;
>                 len -= avail;
>         
>                 /* Other side must not see new producer until data is * there. */
>                 wmb();
>                 intf->req_prod += avail;
>                 
>                 /* Implies mb(): other side will see the updated producer. */
>                 notify_remote_via_evtchn(xen_store_evtchn);
>
> To me, it looks like for guests compiled with CONFIG_SMP, smp_wmb and smp_mb
> would be sufficient, so mb() and wmb() here are only needed if
> a non-SMP guest runs on an SMP host.
>
> Is my analysis correct?

Correct.  The reason full barriers are used is so non-SMP guests still
function correctly.  It is certainly inefficient.

>
> So what I'm suggesting is something like the below patch,
> except instead of using virtio directly, a new set of barriers
> that behaves identically for SMP and non-SMP guests will be introduced.
>
> And of course the weak barriers flag is not needed for Xen -
> that's a virtio only thing.
>
> For example:
>
> smp_pv_wmb()
> smp_pv_rmb()
> smp_pv_mb()
>
> I'd like to get confirmation from Xen folks before posting
> this patchset.
>
> Comments/suggestions?

Very much +1 for fixing this.

Those names would be fine, but they do add yet another set of options in
an already-complicated area.

An alternative might be to have the regular smp_{w,r,}mb() not revert
back to nops if CONFIG_PARAVIRT, or perhaps if pvops have detected a
non-native environment.  (I don't know how feasible this suggestion is,
however.)

~Andrew