On Fri, 31 Jul 2020 11:06:25 +0200
Janosch Frank <frankja@xxxxxxxxxxxxx> wrote:
> On 7/31/20 10:42 AM, Cornelia Huck wrote:
> > On Fri, 31 Jul 2020 09:34:41 +0200
> > Janosch Frank <frankja@xxxxxxxxxxxxx> wrote:
> >
> >> On 7/30/20 5:58 PM, Thomas Huth wrote:
> >>> On 30/07/2020 13.16, Cornelia Huck wrote:
> >>>> On Mon, 27 Jul 2020 05:54:15 -0400
> >>>> Janosch Frank <frankja@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>>> Test the error conditions of guest 2 Ultravisor calls, namely:
> >>>>> * Query Ultravisor information
> >>>>> * Set shared access
> >>>>> * Remove shared access
> >>>>>
> >>>>> Signed-off-by: Janosch Frank <frankja@xxxxxxxxxxxxx>
> >>>>> Reviewed-by: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx>
> >>>>> ---
> >>>>> lib/s390x/asm/uv.h | 68 +++++++++++++++++++
> >>>>> s390x/Makefile | 1 +
> >>>>> s390x/unittests.cfg | 3 +
> >>>>> s390x/uv-guest.c | 159 ++++++++++++++++++++++++++++++++++++++++++++
> >>>>> 4 files changed, 231 insertions(+)
> >>>>> create mode 100644 lib/s390x/asm/uv.h
> >>>>> create mode 100644 s390x/uv-guest.c
> >>>>>
> >>>>
> >>>> (...)
> >>>>
> >>>>> +static inline int uv_call(unsigned long r1, unsigned long r2)
> >>>>> +{
> >>>>> + int cc;
> >>>>> +
> >>>>> + asm volatile(
> >>>>> + "0: .insn rrf,0xB9A40000,%[r1],%[r2],0,0\n"
> >>>>> + " brc 3,0b\n"
> >>>>> + " ipm %[cc]\n"
> >>>>> + " srl %[cc],28\n"
> >>>>> + : [cc] "=d" (cc)
> >>>>> + : [r1] "a" (r1), [r2] "a" (r2)
> >>>>> + : "memory", "cc");
> >>>>> + return cc;
> >>>>> +}
> >>>>
> >>>> This returns the condition code, but no caller seems to check it
> >>>> (instead, they look at header.rc, which is presumably only set if the
> >>>> instruction executed successfully in some way?)
> >>>>
> >>>> Looking at the kernel, it retries for cc > 1 (presumably busy
> >>>> conditions), and cc != 0 seems to be considered a failure. Do we want
> >>>> to look at the cc here as well?
> >>>
> >>> It's there - but here it's in the assembly code, the "brc 3,0b".
> >
> > Ah yes, I missed that.
> >
> >>
> >> Yes, we needed to factor that out in KVM because we sometimes need to
> >> schedule and then it looks nicer handling that in C code. The branch on
> >> condition will jump back for cc 2 and 3. cc 0 and 1 are success and
> >> error respectively and only then the rc and rrc in the UV header are set.
> >
> > Yeah, it's a bit surprising that rc/rrc are also set with cc 1.
>
> Is it?
> The (r)rc *only* contain meaningful information on CC 1.
> On CC 0 they will simply say everything is fine which CC 0 states
> already anyway.
I would consider "things worked" to actually be meaningful :)
(I've seen other instructions indicating different kinds of success.)
>
> >
> > (Can you add a comment? Just so that it is clear that callers never
> > need to check the cc, as rc/rrc already contain more information than
> > that.)
>
> I'd rather fix my test code and also check the CC.
> I did check it for my other UV tests so I've no idea why I didn't do it
> here...
>
>
> How about adding a comment for the cc 2/3 case?
> "The brc instruction will take care of the cc 2/3 case where we need to
> continue the execution because we were interrupted.
> The inline assembly will only return on success/error i.e. cc 0/1."
Sounds good.
Attachment:
pgpWK_94Fvwhg.pgp
Description: OpenPGP digital signature
