[PATCH -next v18 00/20] riscv: Add vector ISA support

Björn Töpel bjorn at kernel.org
Wed Apr 19 00:43:22 PDT 2023


Andy Chiu <andy.chiu at sifive.com> writes:

> This patchset is implemented based on vector 1.0 spec to add vector support
> in riscv Linux kernel. There are some assumptions for this implementations.
>
> 1. We assume all harts has the same ISA in the system.
> 2. We disable vector in both kernel and user space [1] by default. Only
>    enable an user's vector after an illegal instruction trap where it
>    actually starts executing vector (the first-use trap [2]).
> 3. We detect "riscv,isa" to determine whether vector is support or not.
>
> We defined a new structure __riscv_v_ext_state in struct thread_struct to
> save/restore the vector related registers. It is used for both kernel space
> and user space.
>  - In kernel space, the datap pointer in __riscv_v_ext_state will be
>    allocated to save vector registers.
>  - In user space,
> 	- In signal handler of user space, the structure is placed
> 	  right after __riscv_ctx_hdr, which is embedded in fp reserved
> 	  aera. This is required to avoid ABI break [2]. And datap points
> 	  to the end of __riscv_v_ext_state.
> 	- In ptrace, the data will be put in ubuf in which we use
> 	  riscv_vr_get()/riscv_vr_set() to get or set the
> 	  __riscv_v_ext_state data structure from/to it, datap pointer
> 	  would be zeroed and vector registers will be copied to the
> 	  address right after the __riscv_v_ext_state structure in ubuf.
>
> This patchset is rebased to v6.3-rc1 and it is tested by running several
> vector programs simultaneously. It delivers signals correctly in a test
> where we can see a valid ucontext_t in a signal handler, and a correct V
> context returing back from it. And the ptrace interface is tested by
> PTRACE_{GET,SET}REGSET. Lastly, KVM is tested by running above tests in
> a guest using the same kernel image. All tests are done on an rv64gcv
> virt QEMU.
>
> Note: please apply the patch at [4] due to a regression introduced by
> commit 596ff4a09b89 ("cpumask: re-introduce constant-sized cpumask
> optimizations") before testing the series.
>
> Source tree:
> https://github.com/sifive/riscv-linux/tree/riscv/for-next/vector-v18

After some offlist discussions, we might have a identified a
potential libc->application ABI break.

Given an application that does custom task scheduling via a signal
handler. The application binary is not vector aware, but libc is. Libc
is using vector registers for memcpy. It's an "old application, new
library, new kernel"-scenario.

 | ...
 | struct context *p1_ctx;
 | struct context *p2_ctx;
 | 
 | void sighandler(int sig, siginfo_t *info, void *ucontext)
 | {
 |   if (p1_running)
 |     switch_to(p1_ctx, p2_ctx);
 |   if (p2_running)
 |     switch_to(p2_ctx, p1_ctx);
 | }
 | 
 | void p1(void)
 | {
 |   memcpy(foo, bar, 17);
 | }
 | 
 | void p2(void)
 | {
 |   ...
 | }
 | ...

The switch_to() function schedules p1() and p2(). E.g., the
application (assumes that it) saves the complete task state from
sigcontext (ucontext) to p1_ctx, and restores sigcontext to p2_ctx, so
when sigreturn is called, p2() is running, and p1() has been
interrupted.

The "old application" which is not aware of vector, is now run on a
vector enabled kernel/glibc.

Assume that the sighandler is hit, and p1() is in the middle of the
vector memcpy. The switch_to() function will not save the vector
state, and next time p2() is scheduled to run it will have incorrect
machine state.

Now:

Is this an actual or theoretical problem (i.e. are there any
applications in the wild)? I'd be surprised if it would not be the
latter...

Regardless, a kernel knob for disabling vector (sysctl/prctl) to avoid
these kind of breaks is needed (right?). Could this knob be a
follow-up patch to the existing v18 series?

Note that arm64 does not suffer from this with SVE, because the default
vector length (vl==0/128b*32) fits in the "legacy" sigcontext.


Björn



More information about the linux-riscv mailing list