On Tue, 21 Apr 2020 at 06:15, Jason A. Donenfeld Jason@zx2c4.com wrote:
Hi David,
On Mon, Apr 20, 2020 at 2:32 AM David Laight David.Laight@aculab.com wrote:
Maybe kernel_fp_begin() should be passed the address of somewhere the address of an fpu save area buffer can be written to. Then the pre-emption code can allocate the buffer and save the state into it.
Interesting idea. It looks like `struct xregs_state` is only 576 bytes. That's not exactly small, but it's not insanely huge either, and maybe we could justifiably stick that on the stack, or even reserve part of the stack allocation for that that the function would know about, without needing to specify any address.
kernel_fpu_begin() ought also be passed a parameter saying which fpu features are required, and return which are allocated. On x86 this could be used to check for AVX512 (etc) which may be available in an ISR unless it interrupted inside a kernel_fpu_begin() section (etc). It would also allow optimisations if only 1 or 2 fpu registers are needed (eg for some of the crypto functions) rather than the whole fpu register set.
For AVX512 this probably makes sense, I suppose. But I'm not sure if there are too many bits of crypto code that only use a few registers. There are those accelerated memcpy routines in i915 though -- ever see drivers/gpu/drm/i915/i915_memcpy.c? sort of wild. But if we did go this way, I wonder if it'd make sense to totally overengineer it and write a gcc/as plugin to create the register mask for us. Or, maybe some checker inside of objtool could help here.
Actually, though, the thing I've been wondering about is actually moving in the complete opposite direction: is there some efficient-enough way that we could allow FPU registers in all contexts always, without the need for kernel_fpu_begin/end? I was reversing ntoskrnl.exe and was kind of impressed (maybe not the right word?) by their judicious use of vectorisation everywhere. I assume a lot of that is being generated by their compiler, which of course gcc could do for us if we let it. Is that an interesting avenue to consider? Or are you pretty certain that it'd be a huge mistake, with an irreversible speed hit?
When I added support for kernel mode SIMD to arm64 originally, there was a kernel_neon_begin_partial() that took an int to specify how many registers were being used, the reason being that NEON preserve/store was fully eager at this point, and arm64 has 32 SIMD registers, many of which weren't really used, e.g., in the basic implementation of AES based on special instructions.
With the introduction of lazy restore, and SVE handling for userspace, we decided to remove this because it didn't really help anymore, and made the code more difficult to manage.
However, I think it would make sense to have something like this in the general case. I.e., NEON registers 0-3 are always preserved when an exception or interrupt (or syscall) is taken, and so they can be used anywhere in the kernel. If you want the whole set, you will have to use begin/end as before. This would already unlock a few interesting case, like memcpy, xor, and sequences that can easily be implemented with only a few registers such as instructio based AES.
Unfortunately, the compiler needs to be taught about this to be completely useful, which means lots of prototyping and benchmarking upfront, as the contract will be set in stone once the compilers get on board.