On Wed, Oct 20, 2021 at 08:19:40PM -1000, Linus Torvalds wrote: > On Wed, Oct 20, 2021 at 12:44 PM Catalin Marinas > <catalin.marinas@xxxxxxx> wrote: > > > > However, with MTE doing both get_user() every 16 bytes and > > gup can get pretty expensive. > > So I really think that anything that is performance-critical had > better only do the "fault_in_write()" code path in the cold error path > where you took a page fault. [...] > So I wouldn't worry too much about the performance concerns. It simply > shouldn't be a common or hot path. > > And yes, I've seen code that does that "fault_in_xyz()" before the > critical operation that cannot take page faults, and does it > unconditionally. > > But then it isn't the "fault_in_xyz()" that should be blamed if it is > slow, but the caller that does things the wrong way around. Some more thinking out loud. I did some unscientific benchmarks on a Raspberry Pi 4 with the filesystem in a RAM block device and a "dd if=/dev/zero of=/mnt/test" writing 512MB in 1MB blocks. I changed fault_in_readable() in linux-next to probe every 16 bytes: - ext4 drops from around 261MB/s to 246MB/s: 5.7% penalty - btrfs drops from around 360MB/s to 337MB/s: 6.4% penalty For generic_perform_write() Dave Hansen attempted to move the fault-in after the uaccess in commit 998ef75ddb57 ("fs: do not prefault sys_write() user buffer pages"). This was reverted as it was exposing an ext4 bug. I don't whether it was fixed but re-applying Dave's commit avoids the performance drop. btrfs_buffered_write() has a comment about faulting pages in before locking them in prepare_pages(). I suspect it's a similar problem and the fault_in() could be moved, though I can't say I understand this code well enough. Probing only the first byte(s) in fault_in() would be ideal, no need to go through all filesystems and try to change the uaccess/probing order. -- Catalin