Adding V-ext regs to signal context w/o expanding kernel struct sigcontext to avoid glibc ABI break

[Vincent Chen]

On Fri, Dec 23, 2022 at 3:25 AM Vineet Gupta <vineetg at rivosinc.com> wrote:
>
>
> On 12/21/22 19:37, Vincent Chen wrote:
> > On Thu, Dec 22, 2022 at 3:52 AM Vineet Gupta <vineetg at rivosinc.com> wrote:
> >>
> >>
> >> On 12/21/22 11:45, Vineet Gupta wrote:
> >>> 4. Kernel with RVV support + user program using original Glibc sigcontext
> >>> In this case, the kernel needs to save vector registers context to
> >>> memory. The user program may encounter memory issues if the user space
> >>> does not reserve enough memory size for the kernel to create the
> >>> sigcontext. However, we can't seem to avoid this case since there is
> >>> no flexible memory area in struct sigcontext for future expansion.
> >> This is not an issue, if we don't change sigcontext (in kernel and
> >> glibc) - it is essentially the case of existing binaries.
> >> kernel still saves regs on user stack, in rt_sigframe, its just that
> >> userspace is not able to access them in SA_SIGINFO signal handers.
> >> aarch64 have this implemented this approach and it is likely they can't
> >> do that either for SVE regs.
> > Sorry, I don't clearly describe the case. As you mentioned, the kernel
> > will save the vector registers to the user stack or user-specified
> > memory region by struct rt_sigframe in your patch. But, if there is an
> > existing binary compiled with the original sigcontext.h, it will
> > assume that the kernel only occupies the sizeof(struct sigcontext) to
> > save these registers. It will not aware the RVV extension is supported
> > and not expect that the kernel with RVV support needs an extra huge
> > memory region on its stack or specified memory region to save vector
> > registers context. In this case, the user program will encounter
> > memory corruption issues if the size of the memory region specified by
> > the user program is not enough to store these vector registers'
> > context.
>
> No, it will not. In this scheme struct sigcontext remains same as
> before. Kernel is copying the RVV context not in sigcontext, but beyond
> the canonical sigcontext, in layout specified in the rt_sigframe. Please
> take a look at my patch again. It works.

If I understand correctly, in your patch, the kernel uses rt_sigframe
to back up all register contexts in the user space, including RVV
registers. Therefore, the user program needs to reserve enough memory
space for the kernel, which enough size of this memory space is the
sizeof(rt_sigframe) plus rvv_sc_size. However, the rvv_sc_size is
unexpected to existing RISC-V programs. Therefore, some memory of the
existing program may be corrupted by the kernel when the kernel backs
up the RVV registers context.

>
> Again I don't care what scheme we follow, I just want o make forward
> progress.
>

 I understand your thoughts and I sincerely appreciate everything you do.

> -Vineet
>