On Thu, Feb 14, 2019 at 11:38:16AM -0800, Darrick J. Wong wrote: > > + xfs_ilock(ip, XFS_ILOCK_EXCL); > > + xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL); > > Aha, this is the problem -- this operation involves deferred rmap > updates, which means that we have to retain the ILOCK until we've > finished processing the rmap updates so that another thread cannot jump > in and start modifying the file's bmap while we're still trying to > finish the rmap updates. Hmm, if our automatic unlock doesn't work with deferred operations we should probably just get rid of that mechanism entirely, as it is highly unsafe.. > > The patch below fixes generic/127 for me, for this configuration: > > FSTYP=xfs > MKFS_OPTIONS="-m reflink=1,rmapbt=1 -i sparse=1" > MOUNT_OPTIONS="-o usrquota,grpquota,prjquota" I didn't have -i sparc in my test config, but I doubt it makes difference. Let me check if I could reproduce the original issue with it.
