On Sat, Mar 05, 2022 at 06:19:29AM +0100, Christoph Hellwig wrote: > On Sat, Mar 05, 2022 at 09:12:55AM +1100, Dave Chinner wrote: > > AFAICT, all the filesystem/IO path passthrough plumbing for hints is > > now gone, and no hardware will ever receive hints. Doesn't this > > mean that file_write_hint(), file->f_write_hint and iocb->ki_hint > > are now completely unused, too? > > No, for the reason tha you state below. f2fs still uses it. My point is that f2fs uses i_write_hint, not f_write_hint or ki_hint. IOWs, nothing in the IO path use the iocb or file write hints anymore because they only ever got used to set the hint for bios. It's now unused information. According to the io_uring ppl, setup of unnecessary fields in the iocb has a measurable cost and they've done work to minimise it in the past. So if these fields are not actually used by anyone in the IO path, why should we still pay the cost calling ki_hint_validate(file_write_hint(file)) when setting up an iocb? > > AFAICT, this patch leaves just the f2fs allocator usage of > > inode->i_rw_hint to select a segment to allocate from as the > > remaining consumer of this entire plumbing and user API. Is that > > used by applications anywhere, or can that be removed and so the > > rest of the infrastructure get removed and the fcntl()s no-op'd or > > -EOPNOTSUPP? > > I was told it is used quite heavily in android. So it's primarily used by out of tree code? And that after this patch, there's really no way to test that this API does anything useful at all? Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
