> On Aug 7, 2020, at 12:57 PM, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Fri, Aug 7, 2020 at 12:33 PM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote: >> >> No one said we have to do only one ChaCha20 block per slow path hit. > > Sure, doing more might be better for amortizing the cost. > > But you have to be very careful about latency spikes. I would be > *really* nervous about doing a whole page at a time, when this is > called from routines that literally expect it to be less than 50 > cycles. > > So I would seriously suggest you look at a much smaller buffer. Maybe > not a single block, but definitely not multiple kB either. > > Maybe something like 2 cachelines might be ok, but there's a reason > the current code only works with 16 bytes (or whatever) and only does > simple operations with no looping. > > That's why I think you might look at a single double-round ChaCha20 > instead. Maybe do it for two blocks - by the time you wrap around, > you'll have done more than a full ChaCaa20. > > That would imnsho *much* better than doing some big block, and have > huge latency spikes and flush a large portion of your L1 when they > happen. Nasty nasty behavior. > > I really think the whole "we can amortize it with bigger blocks" is > complete and utter garbage. It's classic "benchmarketing" crap. > I think this will come down to actual measurements :). If the cost of one block of cache-cold ChaCha20 is 100 cycles of actual computation and 200 cycles of various cache misses, then let’s do more than one block. I’ll get something working and we’ll see.
