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

[Vineet Gupta]

On 12/22/22 10:33, Andy Chiu wrote:
> On Thu, Dec 22, 2022 at 1:32 PM Richard Henderson
> <richard.henderson at linaro.org> wrote:
>> E.g.
>>
>>       reserved[0] -> magic
>>       reserved[1] -> displacement to "extension area"
>>       reserved[2] -> size of "extension area"
>>
>> Thus the area can be located anywhere within 4GB and expand to 4GB.
>> Not that I'd hope any signal frame would require 4GB.  :-)
>>
> By encoding the extension magic into fp reserved space, and attaching
> actual Vector states underneath, it is possible to make no size
> changes to the sigcontext itself. In fact the comment section of
> __riscv_q_ext_state specifies those bytes were purposely reserved for
> sigcontext expansion. If this is the case then maybe we should just
> use those reserved spaces anyway.
>
> struct __riscv_q_ext_state {
>          __u64 f[64] __attribute__((aligned(16)));
>          __u32 fcsr;
>          /*
>           * Reserved for expansion of sigcontext structure.  Currently zeroed
>           * upon signal, and must be zero upon sigreturn.
>           */
>          __u32 reserved[3];
> };
>
> Here is a way that keeps the size and layout of sigcontext, while it
> also manages to let the kernel write Vector state into an user's
> signal stack. This approach also lets the user space leverage existing
> reserved space to get context from new extensions. We introduce a new
> struct, __riscv_extra_ext_header, unioning with __riscv_fp_state in
> sigcontext. __riscv_extra_ext_header is the same size as
> __riscv_fp_state. The only purpose of the struct is to point to the
> magic header of a following extension, e.g. Vector, located at the
> reserved space. If there is no more extension to come, then all of
> those bytes should be zeros.
>
>   struct sigcontext {
>          struct user_regs_struct sc_regs;
> -       union __riscv_fp_state sc_fpregs;
> +       union {
> +               union __riscv_fp_state sc_fpregs;
> +               struct __riscv_extra_ext_header sc_extdesc;
> +       };
>   };
>
> I wrote a PoC patch for this and it has been pushed into the following git tree:
> https://github.com/sifive/riscv-linux/tree/dev/andyc/for-next-v13
> I tested it on a rv32 QEMU virt machine and the user space can get/set
> Vector registers normally. I haven't tested it on rv64 yet but it
> should be no difference. The patch is not the final version and maybe
> I missed some basic ideas. But if everyone agrees with this approach
> then I would like to start formalizing and submit the series.

This approach looks perfect. Lets productize it to fold this patch into 
the respective patch(es).
We would then need fixups to not unconditionally enable V on fork/execve 
and hook that up to a prctl.
Let me work on that and provide something on top of your series.

-Vineet