Re: [PATCH v2 3/6] crypto: x86/chacha20 - limit the preemption-disabled section

[Eric Biggers]

On Mon, Dec 03, 2018 at 03:13:37PM +0100, Ard Biesheuvel wrote:
> On Sun, 2 Dec 2018 at 11:47, Martin Willi <martin@xxxxxxxxxxxxxx> wrote:
> >
> >
> > > To improve responsiveness, disable preemption for each step of the
> > > walk (which is at most PAGE_SIZE) rather than for the entire
> > > encryption/decryption operation.
> >
> > It seems that it is not that uncommon for IPsec to get small inputs
> > scattered over multiple blocks. Doing FPU context saving for each walk
> > step then can slow down things.
> >
> > An alternative approach could be to re-enable preemption not based on
> > the walk steps, but on the amount of bytes processed. This would
> > satisfy both users, I guess.
> >
> > In the long run we probably need a better approach for FPU context
> > saving, as this really hurts performance-wise. For IPsec we should find
> > a way to avoid the (multiple) per-packet FPU save/restores in softirq
> > context, but I guess this requires support from process context
> > switching.
> >
> 
> At Jason's Zinc talk at plumbers, this came up, and apparently someone
> is working on this, i.e., to ensure that on x86, the FPU restore only
> occurs lazily, when returning to userland rather than every time you
> call kernel_fpu_end() [like we do on arm64 as well]
> 
> Not sure what the ETA for that work is, though, nor did I get the name
> of the guy working on it.

Thanks for the suggestion; I'll replace this with a patch that re-enables
preemption every 4 KiB encrypted.  That also avoids having to do a
kernel_fpu_begin(), kernel_fpu_end() pair just for hchacha_block_ssse3().  But
yes, I'd definitely like repeated kernel_fpu_begin(), kernel_fpu_end() to not be
incredibly slow.  That would help in a lot of other places too.

- Eric