From: Matthew Wilcox > Sent: 10 January 2021 06:13 ... > nvfs_rw_iter_locked() looks very complicated. I suspect it can > be simplified. Of course new_sync_read() needs to be improved too, > as do the other functions here, but fully a third of the difference > between read() and read_iter() is the difference between nvfs_read() > and nvfs_rw_iter_locked(). There is also the non-zero cost of import_iovec(). I've got some slight speedups, but haven't measured an old kernel yet to see how much slower 5.11-rc1 made it. Basic test is: fd = open("/dev/null", O_RDWR); for (1 = 0; 1 < 10000; i++) { start = rdtsc(); writev(fd, iovec, count); histogram[rdtsc() - start]++; } This doesn't actually copy any data - the iovec isn't iterated. I'm seeing pretty stable counts for most of the 10000 iterations. But different program runs can give massively different timings. I'm quessing that depends on cache collisions due to the addresses (virtual of physical?) selected for some items. For 5.11-rc2 -mx32 is slightly faster than 64bit. Whereas -m32 has a much slower syscall entry/exit path, but the difference between gettid() and writev() is lower. The compat code for import_iovec() is actually faster. This isn't really surprising since copy_from_user() is absolutely horrid these days - especially with userspace hardening. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)