On Tue, May 03 2022 at 11:06, Peter Zijlstra wrote:
> On Mon, May 02, 2022 at 05:58:40PM +0200, Thomas Gleixner wrote:
>> Right, though currently it's guaranteed that softirq processing context
>> can use the FPU. Quite some of the network crypto work runs in softirq
>> context, so this might cause a regression. If so, then this needs to be
>> an explicit commit on top which is easy to revert. Let me stare at it
>> some more.
>
> Right, so with the:
>
> preempt_disable();
> this_cpu_write(fpu_in_use, true);
> barrier();
>
> sequence it is safe against both softirq and hardirq fpu usage. The only
> concern is performance not correctness when dropping that
> local_bh_disable() thing.
>
> So what Thomas proposes makes sense to me.
Now I was looking at it the other way round too; i.e. to use
local_bh_disable() for both fpregs_lock() and kernel_fpu_begin().
Using local_bh_disable() for both fpregs_lock() and kernel_fpu_begin()
is not possible with the current constraints, because kernel_fpu_begin()
can be called from hard interrupt context.
But the only use case which utilizes FPU from hard interrupt context is
the random generator via add_randomness_...().
I did a benchmark of these functions, which invoke blake2s_update()
three times in a row, on a SKL-X and a ZEN3. The generic code and the
FPU accelerated code are pretty much on par vs. execution time of the
algorithm itself plus/minus noise.
But in case that the interrupt hits a userspace task the FPU needs to be
saved and if the interrupt does not result in rescheduling then the
return to user space has to restore it. That's _expensive_ and the
actual cost depends on the FPU state, but 200-300 cycles for save and
200-700 cycles for restore are due.
Even if we ignore the save/restore part and assume that it averages out
vs. schedule() having to save FPU state anyway, then there is another
aspect to this: power consumption which affects also thermal budget and
capacity.
Though that made me more curious and I did the same comparison for crc32
which is heavily used by ext4. crc32c_pcl_intel_update() already
contains a switch to software when the buffer length is less than 512
bytes. But even on larger buffers, typically ~4k, FPU is not necessarily
a win. It's consistently slower by a factor of ~1.4x. And that's not due
to xsave/rstor overhead because these computations run on a worker
thread which does not do that dance at all.
IOW, using the FPU blindly for this kind of computations is not
necessarily a good plan. I have no idea how these things are analyzed
and evaluated if at all. Maybe the crypto people can shed some light on
this.
Thanks,
tglx
