On Wed, 19 Apr 2023 07:54:23 PDT (-0700), bjorn@xxxxxxxxxx wrote:
> Björn Töpel <bjorn@xxxxxxxxxx> writes:
>
>> Andy Chiu <andy.chiu@xxxxxxxxxx> 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.
Thanks for writing this up, and sorry I've dropped the ball a few times on
describing it.
>> 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...
I also have no idea. It's kind of odd to say "nobody cares about the
ABI break" when we can manifest it with some fairly simple example, but
I'd bet that nobody cares.
>> 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.
>
> Andy, to clarify from the patchwork call; In
> Documentation/arm64/sve.rst:
>
> There's a per-process prctl (section 6), and a system runtime conf
> (section 9).
