Re: [PATCH 0/5] Add parsing for Zimop ISA extension

[Andrew Jones]

On Mon, Apr 08, 2024 at 01:19:39PM +0200, Clément Léger wrote:
> 
> 
> On 08/04/2024 13:03, Andrew Jones wrote:
> > On Mon, Apr 08, 2024 at 10:01:12AM +0200, Clément Léger wrote:
> >>
> >>
> >> On 05/04/2024 19:33, Deepak Gupta wrote:
> >>> On Fri, Apr 5, 2024 at 8:26 AM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
> >>>>
> >>>> On Thu, Apr 04, 2024 at 12:32:46PM +0200, Clément Léger wrote:
> >>>>> The Zimop ISA extension was ratified recently. This series adds support
> >>>>> for parsing it from riscv,isa, hwprobe export and kvm support for
> >>>>> Guest/VM.
> >>>>
> >>>> I'm not sure we need this. Zimop by itself isn't useful, so I don't know
> >>>> if we need to advertise it at all. When an extension comes along that
> >>>> redefines some MOPs, then we'll advertise that extension, but the fact
> >>>> Zimop is used for that extension is really just an implementation detail.
> >>>
> >>> Only situation I see this can be useful is this:--
> >>>
> >>> An implementer, implemented Zimops in CPU solely for the purpose that they can
> >>> run mainline distro & packages on their hardware and don't want to leverage any
> >>> feature which are built on top of Zimop.
> >>
> >> Yes, the rationale was that some binaries using extensions that overload
> >> MOPs could still be run. With Zimop exposed, the loader could determine
> >> if the binary can be executed without potentially crashing. We could
> >> also let the program run anyway but the execution could potentially
> >> crash unexpectedly, which IMHO is not really good for the user
> >> experience nor for debugging. I already think that the segfaults which
> >> happens when executing binaries that need some missing extension are not
> >> so easy to debug, so better add more guards.
> > 
> > OK. It's only one more extension out of dozens, so I won't complain more,
> 
> No worries, your point *is* valid since I'm not sure yet that the loader
> will actually do that one day.
> 
> BTW, are you aware of any effort to make the elf dynamic loader
> "smarter" and actually check for needed extensions to be present rather
> than blindly running the elf and potentially catching SIGILL ?

Jeff Law told me a bit about FMV (function multiversioning). I don't know
much about this, but, from what he's told me, it sounds like there will be
an ifunc resolver which invokes hwprobe to determine which variants are
possible/best to use, so it should be possible to avoid SIGILL by always
having a basic variant.

Thanks,
drew

> 
> Thanks,
> 
> Clément
> 
> > but I was thinking that binaries that use particular extensions would
> > check for those particular extensions (step 2), rather than Zimop.
> > 
> > Thanks,
> > drew
> > 
> >>
> >>>
> >>> As an example zicfilp and zicfiss are dependent on zimops. glibc can
> >>> do following
> >>>
> >>> 1) check elf header if binary was compiled with zicfiss and zicfilp,
> >>> if yes goto step 2, else goto step 6.
> >>> 2) check if zicfiss/zicfilp is available in hw via hwprobe, if yes
> >>> goto step 5. else goto step 3
> >>> 3) check if zimop is available via hwprobe, if yes goto step 6, else goto step 4
> >>
> >> I think you meant step 5 rather than step 6.
> >>
> >> Clément
> >>
> >>> 4) This binary won't be able to run successfully on this platform,
> >>> issue exit syscall. <-- termination
> >>> 5) issue prctl to enable shadow stack and landing pad for current task
> >>> <-- enable feature
> >>> 6) let the binary run <-- let the binary run because no harm can be done