Messages in this thread |  | | Date | Wed, 6 Mar 2024 18:54:51 +0000 | | From | Catalin Marinas <> | | Subject | Re: [PATCH v4] arm64/fpsimd: Suppress SVE access traps when loading FPSIMD state |
| |
On Mon, Jan 22, 2024 at 07:42:14PM +0000, Mark Brown wrote: > This indicates that there should be some useful benefit from reducing the > number of SVE access traps for blocking system calls like we did for non > blocking system calls in commit 8c845e273104 ("arm64/sve: Leave SVE enabled > on syscall if we don't context switch"). Let's do this by counting the > number of times we have loaded FPSIMD only register state for SVE tasks > and only disabling traps after some number of times, otherwise leaving > traps disabled and flushing the non-shared register state like we would on > trap.
It looks like some people complain about SVE being disabled, though I assume this is for kernels prior to 6.2 and commit 8c845e273104 ("arm64/sve: Leave SVE enabled on syscall if we don't context switch"):
https://bugs.launchpad.net/ubuntu/+source/glibc/+bug/1999551/comments/52
I assume we may see the other camp complaining about the additional state saving on context switch.
Anyway, I don't see why we should treat blocking syscalls differently from non-blocking ones (addressed by the commit above). I guess the difference is that one goes through a context switch but, from a user perspective, it's still a syscall. The SVE state is expected to be discarded and there may be a preference for avoiding the subsequent fault.
> I pulled 64 out of thin air for the number of flushes to do, there is > doubtless room for tuning here. Ideally we would be able to tell if the > task is actually using SVE but without using performance counters (which > would be substantial work) we can't currently tell. I picked the number > because so many of the tasks using SVE used it so frequently.
So I wonder whether we should make the timeout disabling behaviour the same for both blocking and non-blocking syscalls. IOW, ignore the context switching aspect. Every X number of returns, disable SVE irrespective of whether it was context switched or not. Or, if the number of returns has a variation in time, just use a jiffy (or other time based figure), it would time out in the same way for all types of workloads.
-- Catalin
|  |