Hey Ard,
On Fri, 8 Dec 2023 12:32:19 +0100, Ard Biesheuvel wrote:
> From: Ard Biesheuvel <ardb@xxxxxxxxxx>
>
> Currently, kernel mode NEON (SIMD) support is implemented in a way that
> requires preemption to be disabled while the SIMD registers are live.
> The reason for this is that those registers are not in the set that is
> preserved/restored on exception entry/exit and context switch, as this
> would impact performance generally, even for workloads where kernel mode
> SIMD is not the bottleneck.
>
> [...]
I applied the first three patches to for-next/fpsimd:
[1/4] arm64: fpsimd: Drop unneeded 'busy' flag
https://git.kernel.org/arm64/c/e109130b0e5e
[2/4] arm64: fpsimd: Preserve/restore kernel mode NEON at context switch
https://git.kernel.org/arm64/c/1e3a3de1ff6c
[3/4] arm64: fpsimd: Implement lazy restore for kernel mode FPSIMD
https://git.kernel.org/arm64/c/035262623959
It would be nice to have an Ack from Herbert on the last one so that
he's aware of the possible conflicts.
The other thing I tangentially wondered about is what happens now if code
calls uaccess routines (e.g. get_user()) within a kernel_neon_{begin,end}
section? I think previously the fact that preemption had to be disabled
would've caused the might_fault() to explode, but now I suppose the BUG_ON()
in kernel_neon_begin() will save us. Is that right?
Cheers,
--
Will
https://fixes.arm64.dev
https://next.arm64.dev
https://will.arm64.dev
