On Mon, Aug 10, 2020 at 09:31:48AM -0700, Linus Torvalds wrote: > On Mon, Aug 10, 2020 at 4:47 AM Willy Tarreau <w...@1wt.eu> wrote: > > > > Doing testing on real hardware showed that retrieving the TSC on every > > call had a non negligible cost, causing a loss of 2.5% on the accept() > > rate and 4% on packet rate when using iptables -m statistics. > > And by "real hardware" I assume you mean x86, with a fairly fast and > high-performance TSC for get_random_entropy().
Yep. > Reading the TSC takes on the order of 20-50 cycles, iirc. > > But it can actually be *much* more expensive. On non-x86, it can be an > IO cycle to external chips. I took what we were already using in add_interrupt_randomness() since I considered that if it was acceptable there, it probably was elsewhere. > And on older hardware VM's in x86, it can be a vm exit etc, so > thousands of cycles. I hope nobody uses those VM's any more, but it > would be a reasonable test-case for some non-x86 implementations, so.. Yes, I remember these ones, they were not fun at all. > IOW, no. You guys are - once again - ignoring reality. I'm not ignoring reality, quite the opposite, trying to take all knowledge into account. If I'm missing some points, fine. But if we were already calling that in the interrupt handler I expected that this would be OK. The alternative Florian talked about is quite interesting as well, which is to collect some cheap noise in the network rx/tx paths since these are the areas we care about. Willy
