On Mon, Jul 09, 2018 at 02:33:47PM +0200, Jan Kara wrote: > On Thu 05-07-18 10:53:10, Ross Zwisler wrote: > > On Wed, Jul 04, 2018 at 08:59:52PM -0700, Darrick J. Wong wrote: > > > On Thu, Jul 05, 2018 at 09:54:14AM +1000, Dave Chinner wrote: > > > > On Wed, Jul 04, 2018 at 02:27:23PM +0200, Jan Kara wrote: > > > > > On Wed 04-07-18 10:49:23, Dave Chinner wrote: > > > > > > On Mon, Jul 02, 2018 at 11:29:12AM -0600, Ross Zwisler wrote: > > > > > > > Follow the lead of xfs_break_dax_layouts() and add synchronization between > > > > > > > operations in ext4 which remove blocks from an inode (hole punch, truncate > > > > > > > down, etc.) and pages which are pinned due to DAX DMA operations. > > > > > > > > > > > > > > Signed-off-by: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx> > > > > > > > Reviewed-by: Jan Kara <jack@xxxxxxx> > > > > > > > Reviewed-by: Lukas Czerner <lczerner@xxxxxxxxxx> > > > > > > > --- > > > > > > > > > > > > > > Changes since v2: > > > > > > > * Added a comment to ext4_insert_range() explaining why we don't call > > > > > > > ext4_break_layouts(). (Jan) > > > > > > > > > > > > Which I think is wrong and will cause data corruption. > > > > > > > > > > > > > @@ -5651,6 +5663,11 @@ int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len) > > > > > > > LLONG_MAX); > > > > > > > if (ret) > > > > > > > goto out_mmap; > > > > > > > + /* > > > > > > > + * We don't need to call ext4_break_layouts() because we aren't > > > > > > > + * removing any blocks from the inode. We are just changing their > > > > > > > + * offset by inserting a hole. > > > > > > > + */ > > > > > > Does calling ext4_break_layouts from insert range not work? > > > > > > It's my understanding that file leases work are a mechanism for the > > > filesystem to delegate some of its authority over physical space > > > mappings to "client" software. AFAICT it's used for mmap'ing pmem > > > directly into userspace and exporting space on shared storage over > > > pNFS. Some day we might use the same mechanism for the similar things > > > that RDMA does, or the swapfile code since that's essentially how it > > > works today. > > > > > > The other part of these file leases is that the filesystem revokes them > > > any time it wants to perform a mapping operation on a file. This breaks > > > my mental model of how leases work, and if you commit to this for ext4 > > > then I'll have to remember that leases are different between xfs and > > > ext4. Worse, since the reason for skipping ext4_break_layouts seems to > > > be the implementation detail that "DAX won't care", then someone else > > > wiring up pNFS/future RDMA/whatever will also have to remember to put it > > > back into ext4 or else kaboom. > > > > > > Granted, Dave said all these things already, but I actually feel > > > strongly enough to reiterate. > > > > Jan, would you like me to call ext4_break_layouts() in ext4_insert_range() to > > keep the lease mechanism consistent between ext4 and XFS, or would you prefer > > the s/ext4_break_layouts()/ext4_dax_unmap_pages()/ rename? > > Let's just call it from ext4_insert_range(). I think the simple semantics > Dave and Darrick defend is more maintainable and insert range isn't really > performance critical operation. > > The question remains whether equivalent of BREAK_UNMAP is really required > also for allocation of new blocks using fallocate. Because that doesn't > really seem fundamentally different from normal write which uses > BREAK_WRITE for xfs_break_layouts(). And that it more often used operation > so bothering with GUP synchronization when not needed could hurt there. > Dave, Darrick? Hmm, IIRC BREAK_UNMAP is supposed to be for callers who are going to remove (or move) mappings that already exist, so that the caller blocks until the lessee acknowledges that they've forgotten all the mappings they knew about. So I /think/ for fallocate mode 0 I think this could be BREAK_WRITE instead of _UNMAP, though (at least for xfs) the other modes all need _UNMAP. Side question: in xfs_file_aio_write_checks, do we need to do BREAK_UNMAP if is possible that writeback will end up performing a copy write? Granted, the pnfs export and dax stuff don't support reflink or cow so I guess this is an academic question for now... --D > Honza > -- > Jan Kara <jack@xxxxxxxx> > SUSE Labs, CR
