On Thu, May 17, 2018 at 07:19:43PM +0100, Dave Martin wrote: [...]
> kernel_neon_begin() could then do > > local_fpsimd_context_lock(); > > /* ... */ > > preempt_disable(); > local_unlock(fpsimd_context_lock); > > ... with the following in kernel_neon_end(): > > local_unlock(fpsimd_lock); > preempt_enable(); > > > If kernel-mode NEON was considered harmful to RT due to the context > switch overheads, then the above might be overkill. SVE will be worse > in that regard, and also needs thinking about at some point -- I've not > looked at if from the RT angle at all. Hmmm, !KERNEL_MODE_NEON breaks EFI, so this probably does want looking at. Ard's recent rework to enable voluntary preemption the crypto backends for arm64 [1] should reduce the fpsimd_lock blackouts, but it still depends on the backends playing nice. Cheers ---Dave [1] [PATCH resend 00/10] crypto: arm64 - play nice with CONFIG_PREEMPT lists.infradead.org/pipermail/linux-arm-kernel/2018-April/574819.html
