Re: x86/vMSI-X emulation issue

["Jan Beulich" <JBeulich@suse.com>]

>>> On 24.03.16 at 10:39, <Paul.Durrant@citrix.com> wrote:
>>  -----Original Message-----
>> From: Jan Beulich [mailto:JBeulich@suse.com]
>> Sent: 24 March 2016 09:35
>> To: Paul Durrant
>> Cc: Andrew Cooper; xen-devel
>> Subject: RE: [Xen-devel] x86/vMSI-X emulation issue
>> 
>> >>> On 24.03.16 at 10:09, <Paul.Durrant@citrix.com> wrote:
>> >> From: Xen-devel [mailto:xen-devel-bounces@lists.xen.org] On Behalf Of
>> Jan
>> >> Beulich
>> >> Sent: 24 March 2016 07:52
>> >> > 2) Do aforementioned chopping automatically on seeing
>> >> >     X86EMUL_UNHANDLEABLE, on the basis that the .check
>> >> >     handler had indicated that the full range was acceptable. That
>> >> >     would at once cover other similarly undesirable cases like the
>> >> >     vLAPIC code returning this error. However, any stdvga like
>> >> >     emulated device would clearly not want such to happen, and
>> >> >     would instead prefer the entire batch to get forwarded in one
>> >> >     go (stdvga itself sits on a different path). Otoh, with the
>> >> >     devices we have currently, this would seem to be the least
>> >> >     intrusive solution.
>> >>
>> >> Having thought about it more over night, I think this indeed is
>> >> the most reasonable route, not just because it's least intrusive:
>> >> For non-buffered internally handled I/O requests, no good can
>> >> come from forwarding full batches to qemu, when the respective
>> >> range checking function has indicated that this is an acceptable
>> >> request. And in fact neither vHPET not vIO-APIC code generate
>> >> X86EMUL_UNHANDLEABLE. And vLAPIC code doing so is also
>> >> just apparently so - I'll submit a patch to make this obvious once
>> >> tested.
>> >>
>> >> Otoh stdvga_intercept_pio() uses X86EMUL_UNHANDLEABLE in
>> >> a manner similar to the vMSI-X code - for internal caching and
>> >> then forwarding to qemu. Clearly that is also broken for
>> >> REP OUTS, and hence a similar rep count reduction is going to
>> >> be needed for the port I/O case.
>> >
>> > It suggests that such cache-and/or-forward models should probably sit
>> > somewhere else in the flow, possibly being invoked from
>> hvm_send_ioreq()
>> > since there should indeed be a selected ioreq server for these cases.
>> 
>> I don't really think so. As I have gone through and carried out
>> what I had described above, I think I managed to address at
>> least one more issue with not properly handled rep counts, and
>> hence I think doing it that way is correct. I'll have to test the
>> thing before I can send it out, for you to take a look.
>> 
> 
> Ok. I never particularly liked using X86EMUL_UNHANDLEABLE to invoke the 
> forwarding behaviour though as it's only legitimate to do it on the first 
> rep.

Well, that's explicitly one of the wrong assumptions that patch
addresses: It is perfectly fine for an individual handler to return
this on other than the first iteration. It's only the generic
infrastructure which doesn't currently permit this (for no
apparent reason).

> I always had the feeling there had to be a nicer way of doing it. 
> Possibly just too intrusive a change at this point though.

I'm of course up for alternatives, if you're willing to work on such.
Yet I think backporting would become even more of a problem when
going such an alternative route.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel