On Tue, Jun 07, 2022 at 04:06:53PM +0000, Al Viro wrote: > > Also the FUA check below needs to check IOMAP_DIO_NOSYNC as > > well. > > Does it? AFAICS, we don't really care about REQ_FUA on any requests - what > btrfs hack tries to avoid is stepping into > if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC)) > ret = generic_write_sync(iocb, ret); > with generic_write_sync() called by btrfs_do_write_iter() after it has > dropped the lock held through btrfs_direct_write(). Do we want to > suppress REQ_FUA on the requests generated by __iomap_dio_rw() in > that case (DSYNC, !SYNC)? Confused... Yes. FUA is an used an an optimization so that we can avid the generic_write_sync or similar call if a O_DSYNC write doesn't need any metadata update. If the caller already does the generic_write_sync equivalent we don't also need FUA. It is not actually harmful in that it gives worse results, but it will kill performance.
