Re: [PATCH v6 12/15] xfs: require mmap lock for xfs_break_layouts()

[Dan Williams <dan.j.williams@intel.com>]

On Mon, Mar 19, 2018 at 11:19 AM, Darrick J. Wong
<darrick.wong@oracle.com> wrote:
> On Mon, Mar 19, 2018 at 10:57:55AM -0700, Dan Williams wrote:
>> On Mon, Mar 19, 2018 at 10:33 AM, Darrick J. Wong
>> <darrick.wong@oracle.com> wrote:
>> > On Thu, Mar 15, 2018 at 08:52:29AM -0700, Dan Williams wrote:
>> >> In preparation for adding coordination between truncate operations and
>> >> busy dax-pages, extend xfs_break_layouts() to assume it must be called
>> >> with the mmap lock held. This locking scheme will be required for
>> >> coordinating the break of 'dax layouts' (non-idle dax (ZONE_DEVICE)
>> >> pages mapped into the file's address space).
>> >
>> > If I'm reading this right, you've added a requirement (for xfs anyway)
>> > that we have to have grabbed MMAPLOCK_EXCL before calling break_layout()
>> > so that the layout breaking process will block until active dmas have
>> > finished.
>> >
>> > In 4.16 we added xfs_iolock_two_inodes_and_break_layout (in
>> > xfs_reflink.c) to break pnfs leases for files that are about to be
>> > reflinked (since pnfs and reflink aren't compatible either).  I think
>> > that function will have to be adapted to take the appropriate mmap locks
>> > too -- definitely the exclusive mmap lock for the destination file
>> > because we anticipate punching out blocks.  I'm not sure about the
>> > source file; I think taking the shared mmap lock is fine for that?
>>
>> I don't see anything to adapt with respect to mmap locks since reflink
>> and dax are mutually exclusive. The new lock coordination and layout
>> breaking in xfs_break_dax_layouts() only applies to dax vs truncate,
>> so xfs_iolock_two_inodes_and_break_layout() looks ok to me as is.
>
> I was aiming for DAX + reflink someday working together.  :)

Of course, but then fixing up xfs_iolock_two_inodes_and_break_layout()
is a follow on consideration from this patch series.

> As far as that goes, I think the only thing we really have to do is
> figure out how to make iomap_begin do the cow operation before letting
> userspace write to (or MAP_SYNC a writable region) pmem, right?

Yes, I think so, trigger xfs_break_dax_layouts() on the destination
inode on any write fault.