On Mon, Jun 22, 2015 at 11:48 AM, Jeff Moyer <jmoyer@xxxxxxxxxx> wrote: > Christoph Hellwig <hch@xxxxxx> writes: > >> On Mon, Jun 22, 2015 at 12:42:44PM -0400, Jeff Moyer wrote: >>> OK, add torn sector detection/recovery to that statement, then. More >>> importantly, do you agree with the sentiment or not? >> >> I think we're getting on a very slipper slope if we think about >> application here. Buffered I/O application must deal with torn >> writes at any granulairty anyway, e.g. fsync + rename is the >> only thing they can rely on right now (I actually have software O_ATOMIC >> code to avoid this, but that's another story). > > OK, so you think applications using buffered I/O will Just Work(TM)? My > guess is that things will start to break that hadn't broken in the > past. Sure, the application isn't designed properly, and that should be > fixed, but we shouldn't foist this on users as the default. > >> Direct I/O using application can make assumption if they know the sector >> size, and we must have a way for them to be able to see our new >> "subsector sector size". > > You need to let them determine that when NOT using the btt, yes. Right > now, I don't think there's a way to determine what the underlying atomic > write unit is. That's something the NFIT spec probably should have > defined. There are no atomic write units for NFIT to advertise beyond cpu register width. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in