On Thu, Sep 21, 2017 at 09:43:41AM +0300, Amir Goldstein wrote: > On Thu, Sep 21, 2017 at 1:22 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > [cc lkml, PeterZ and Byungchul] > ... > > The thing is, this IO completion has nothing to do with the lower > > filesystem - it's the IO completion for the filesystem on the loop > > device (the upper filesystem) and is not in any way related to the > > IO completion from the dax device the lower filesystem is waiting > > on. > > > > IOWs, this is a false positive. > > > > Peter, this is the sort of false positive I mentioned were likely to > > occur without some serious work to annotate the IO stack to prevent > > them. We can nest multiple layers of IO completions and locking in > > the IO stack via things like loop and RAID devices. They can be > > nested to arbitrary depths, too (e.g. loop on fs on loop on fs on > > dm-raid on n * (loop on fs) on bdev) so this new completion lockdep > > checking is going to be a source of false positives until there is > > an effective (and simple!) way of providing context based completion > > annotations to avoid them... > > > > IMO, the way to handle this is to add 'nesting_depth' information > on blockdev (or bdi?). 'nesting' in the sense of blockdev->fs->blockdev->fs. > AFAIK, the only blockdev drivers that need to bump nesting_depth > are loop and nbd?? You're assumming that this sort of "completion inversion" can only happen with bdev->fs->bdev, and that submit_bio_wait() is the only place where completions are used in stackable block devices. AFAICT, this could happen on with any block device that can be stacked multiple times that uses completions. e.g. MD has a function sync_page_io() that calls submit_bio_wait(), and that is called from places in the raid 5, raid 10, raid 1 and bitmap layers (plus others in DM). These can get stacked anywhere - even on top of loop devices - and so I think the issue has a much wider scope than just loop and nbd devices. > Not sure if the kernel should limit loop blockdev nesting depth?? There's no way we should do that just because new lockdep functionality is unable to express such constructs. -Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
