Cord Amfmgm <[email protected]> writes:
> On Thu, May 30, 2024 at 3:33 AM Alex Bennée <[email protected]> wrote:
>
> Cord Amfmgm <[email protected]> writes:
>
> > On Tue, May 28, 2024 at 11:32 AM Peter Maydell <[email protected]>
> wrote:
> >
> > On Tue, 28 May 2024 at 16:37, Cord Amfmgm <[email protected]> wrote:
> > >
> > > On Tue, May 28, 2024 at 9:03 AM Peter Maydell
> <[email protected]> wrote:
> > >>
> > >> On Mon, 20 May 2024 at 23:24, Cord Amfmgm <[email protected]> wrote:
> > >> > On Mon, May 20, 2024 at 12:05 PM Peter Maydell
> <[email protected]> wrote:
> <snip>
> > >> > And here's an example buffer of length 0 -- you probably already
> know what I'm going to do here:
> > >> >
> > >> > char buf[0];
> > >> > char * CurrentBufferPointer = &buf[0];
> > >> > char * BufferEnd = &buf[-1]; // "address of the last byte in the
> buffer"
> > >> > // The OHCI Host Controller than advances CurrentBufferPointer like
> this: CurrentBufferPointer += 0
> > >> > // After the transfer:
> > >> > // CurrentBufferPointer = &buf[0];
> > >> > // BufferEnd = &buf[-1];
> > >>
> > >> Right, but why do you think this is valid, rather than
> > >> being a guest software bug? My reading of the spec is that it's
> > >> pretty clear about how to say "zero length buffer", and this
> > >> isn't it.
> > >>
> > >> Is there some real-world guest OS that programs the OHCI
> > >> controller this way that we're trying to accommodate?
> > >
> > >
> > > qemu versions 4.2 and before allowed this behavior.
> >
> > So? That might just mean we had a bug and we fixed it.
> > 4.2 is a very old version of QEMU and nobody seems to have
> > complained in the four years since we released 5.0 about this,
> > which suggests that generally guest OS drivers don't try
> > to send zero-length buffers in this way.
> >
> > > I don't think it's valid to ask for a *popular* guest OS as a
> proof-of-concept because I'm not an expert on those.
> >
> > I didn't ask for "popular"; I asked for "real-world".
> > What is the actual guest code you're running that falls over
> > because of the behaviour change?
> >
> > More generally, why do you want this behaviour to be
> > changed? Reasonable reasons might include:
> > * we're out of spec based on reading the documentation
> > * you're trying to run some old Windows VM/QNX/etc image,
> > and it doesn't work any more
> > * all the real hardware we tested behaves this way
> >
> > But don't necessarily include:
> > * something somebody wrote and only tested on QEMU happens to
> > assume the old behaviour rather than following the hw spec
> >
> > QEMU occasionally works around guest OS bugs, but only as
> > when we really have to. It's usually better to fix the
> > bug in the guest.
> >
> > It's not, and I've already demonstrated that real hardware is consistent
> with the fix in this patch.
> >
> > Please check your tone.
>
> I don't think that is a particularly helpful comment for someone who is
> taking the time to review your patches. Reading through the thread I
> didn't see anything that said this is how real HW behaves but I may well
> have missed it. However you have a number of review comments to address
> so I suggest you spin a v2 of the series to address them and outline the
> reason to accept an out of spec transaction.
>
> I did a rework of the patch -- see my email from May 20, quoted below -- and
> I was under the impression it addressed all the
> review comments. Did I miss something? I apologize if I did.
Ahh I see - I'd only seen this thread continue so wasn't aware a new
version had been posted. For future patches consider using -vN when
sending them so we can clearly see a new revision is available.
>
>> index acd6016980..71b54914d3 100644
>> --- a/hw/usb/hcd-ohci.c
>> +++ b/hw/usb/hcd-ohci.c
>> @@ -941,8 +941,8 @@ static int ohci_service_td(OHCIState *ohci, struct
>> ohci_ed *ed)
>> if ((td.cbp & 0xfffff000) != (td.be & 0xfffff000)) {
>> len = (td.be & 0xfff) + 0x1001 - (td.cbp & 0xfff);
>> } else {
>> - if (td.cbp > td.be) {
>> - trace_usb_ohci_iso_td_bad_cc_overrun(td.cbp, td.be);
>> + if (td.cbp - 1 > td.be) { /* rely on td.cbp != 0 */
>
>> Reading through the thread I didn't see anything that said this is how real
>> HW behaves but I may well have missed it.
>
> This is what I wrote regarding real HW:
>
> Results are:
>
> qemu 4.2 | qemu HEAD | actual HW
> ------------+------------+------------
> works fine | ohci_die() | works fine
>
> Would additional verification of the actual HW be useful?
>
> Peter posted the following which is more specific than "qemu 4.2" -- I agree
> this is most likely the qemu commit where this
> thread is focused:
>
>> Almost certainly this was commit 1328fe0c32d54 ("hw: usb: hcd-ohci:
>> check len and frame_number variables"), which added these bounds
>> checks. Prior to that we did no bounds checking at all, which
>> meant that we permitted cbp=be+1 to mean a zero length, but also
>> that we permitted the guest to overrun host-side buffers by
>> specifying completely bogus cbp and be values. The timeframe is
>> more or less right (2020), at least.
>>
>> -- PMM
>
> Where does the conversation go from here? I'm under the impression I have
> provided objective answers to all the questions
> and resolved all review comments on the code. I receive the feedback
> that I missed something - please restate the question?
I can see patch 1/2 has been queued and 2/2 is still outstanding. I'm
having trouble finding the referenced entry in the OHCI spec. The only
one I can see is Release 1.1, January 6th, 2000 and that doesn't have a
section 4.3.1.2.
I think discussion should continue on that thread.
--
Alex Bennée
Virtualisation Tech Lead @ Linaro
