On Sun, Jan 10, 2021 at 04:14:55PM -0500, Mikulas Patocka wrote: > That's a good point. I split nvfs_rw_iter to separate functions > nvfs_read_iter and nvfs_write_iter - and inlined nvfs_rw_iter_locked into > both of them. It improved performance by 1.3%. > > > Not that it had been more useful on the write side, really, > > but that's another story (nvfs_write_pages() handling of > > copyin is... interesting). Let's figure out what's going > > on with the read overhead first... > > > > lib/iov_iter.c primitives certainly could use massage for > > better code generation, but let's find out how much of the > > PITA is due to those and how much comes from you fighing > > the damn thing instead of using it sanely... > > The results are: > > read: 6.744s > read_iter: 7.417s > read_iter - separate read and write path: 7.321s > Al's read_iter: 7.182s > Al's read_iter with _copy_to_iter: 7.181s So * overhead of hardening stuff is noise here * switching to more straightforward ->read_iter() cuts the overhead by about 1/3. Interesting... I wonder how much of that is spent in iterate_and_advance() glue inside copy_to_iter() here. There's certainly quite a bit of optimizations possible in those primitives and your usecase makes a decent test for that... Could you profile that and see where is it spending the time, on instruction level?
