On Sun, Feb 16, 2025 at 10:27:02AM +0800, Herbert Xu wrote: > On Sat, Feb 15, 2025 at 09:04:12AM -0800, Jakub Kicinski wrote: > > > > Can confirm, FWIW. I don't know as much about IPsec, but for TLS > > lightweight SW-only crypto would be ideal. > > Please note that while CPU-only crypto is the best for networking, > it actually operates in asynchronous mode on x86. This is because > RX occurs in softirq context, which may not be able to use SIMD on > x86. Well, the async fallback (using cryptd) occurs only when a kernel-mode FPU section in process context is interrupted by a hardirq and at the end of it a softirq also tries to use kernel-mode FPU. It's generally a rare case but also a terrible implementation that is really bad for performance; this should never have been implemented this way. I am planning to fix it so that softirqs on x86 will always be able to use the FPU, like they can on some of the other arches like arm64 and riscv. - Eric
