From: Andrew Jones Date: Mon, 8 Apr 2024 15:56:29 +0200 Subject: [PATCH 0/5] Add parsing for Zimop ISA extension In-Reply-To: <89d4a24c-db24-487b-8c5c-bdc1fa2d42b4@rivosinc.com> References: <20240404103254.1752834-1-cleger@rivosinc.com> <20240405-091c6c174f023d74b434059d@orel> <20240408-6c93f3f50b55234f3825ca33@orel> <89d4a24c-db24-487b-8c5c-bdc1fa2d42b4@rivosinc.com> Message-ID: <20240408-8d245caaaeddc6a23e39d6dd@orel> List-Id: To: kvm-riscv@lists.infradead.org MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit On Mon, Apr 08, 2024 at 01:19:39PM +0200, Cl?ment L?ger wrote: > > > On 08/04/2024 13:03, Andrew Jones wrote: > > On Mon, Apr 08, 2024 at 10:01:12AM +0200, Cl?ment L?ger wrote: > >> > >> > >> On 05/04/2024 19:33, Deepak Gupta wrote: > >>> On Fri, Apr 5, 2024 at 8:26?AM Andrew Jones wrote: > >>>> > >>>> On Thu, Apr 04, 2024 at 12:32:46PM +0200, Cl?ment L?ger 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, > > Cl?ment > > > 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. > >> > >> Cl?ment > >> > >>> 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