On Sat, Mar 03, 2018 at 10:15:17AM +1100, Dave Chinner wrote: > On Sat, Mar 03, 2018 at 12:00:42AM +0100, Christoph Hellwig wrote: > > Oh, and another thing: I think you want to make this new code dependent > > on the block devie actually supporting REQ_FUA natively. Otherwise > > you'll cause a flush for every emulated FUA write, which is only going > > make things worse, especially for ATA where FLUSH is not queued. And > > last time I check libata still disabled FUA by default. > > Yup, but the issue we have right now is that for pure RWF_DSYNC data > overwrites we are already doing a post-flush on every IO. It's being > issued as a separate zero-length IO, which is why REQ_FUA is faster > and results in lower overall IOPS. The flush comes from this path: That is only the case if your device actually supports FUA. If the device does notit is emulated by the block/flk-flush.c code by issuing a FLUSH once the write has returned. So for e.g. a direct I/O write() call with O_DSYNC that turns into e.g. four write calls on the wire you currently have: WRITE WRITE WRITE WRITE FLUSH with your patch and a device that supports FUA you get WRITE (FUA) WRITE (FUA) WRITE (FUA) WRITE (FUA) but with a device that does not support FUA you get WRITE FLUSH WRITE FLUSH WRITE FLUSH WRITE FLUSH with the additional pain point that on ATA FLUSH is not a queueable command, so it will have to wait for the completion of every other non-related command first, and no other command can be started. So we should absolutely use your new approach IFF the device actually supports FUA (aka QUEUE_FLAG_FUA is set), but it will not help much or even be harmful if the device does not actually support the FUA bit.
