On Thu, Feb 04, 2016 at 09:29:57PM +0100, Jan Kara wrote:
On Thu 04-02-16 12:56:19, Ross Zwisler wrote:
> On Wed, Feb 03, 2016 at 11:46:11AM +0100, Jan Kara wrote:
> > Let's clear this up a bit: The problem with using
->fsync() method is that
> > it doesn't get called for sync(2). We could use ->sync_fs() to flush
> > in case of sync(2) (that's what's happening for normal storage) but
> > problem with PMEM is that "flush all cached data" operation
> > means iterate through all modified pages and we didn't want to implement
> > this for DAX fsync code.
> > So we have decided to do cache flushing for DAX at a different point - mark
> > inodes which may have writes cached as dirty and use writeback code for the
> > cache flushing. But looking at it now, we have actually chosen a wrong
> > place to do the flushing in the writeback path - note that sync(2) writes
> > data via __writeback_single_inode() -> do_writepages() and thus doesn't
> > even get to filemap_write_and_wait().
> > So revisiting the decision I see two options:
> > 1) Move the DAX flushing code from filemap_write_and_wait() into
> > ->writepages() fs callback. There the filesystem can provide all the
> > information it needs including bdev, get_block callback, or whatever.
> > 2) Back out even further and implement own tracking and iteration of inodes
> > to write.
> > So far I still think 2) is not worth the complexity (although it would
> > bring DAX code closer to how things behave with standard storage) so I
> > would go for 1).
> Jan, just to clarify, are you proposing this change for v4.5 in the remaining
> RCs as an alternative to the get_bdev() patch?
Yes, because I don't think anything like ->get_bdev() is needed at all.
Look: dax_do_io(), __dax_fault(), __dax_pmd_fault(), dax_zero_page_range()
don't really need bdev - we have agreed that get_block() fills that in just
dax_clear_blocks() has IMO just the wrong signature - it should take bdev
and not inode as an argument. Because combination inode + bdev sector
doesn't really make much sense.
dax_writeback_mapping_range() is the only remaining offender and it can
easily take bdev as an argument when called from ->writepages().
> Or can we move forward with get_bdev(), and try and figure out this new way of
> calling fsync/msync for v4.6? My main concern here is that changing how the
> DAX sync code gets called will affect all three filesystems as well as MM, and
> that it might be too much for RC inclusion...
I think changes aren't very intrusive so we can feed them in during RC
phase and frankly, you have to move to using ->writepages() anyway to make
sync(2) work reliably.
Okay, sounds good. I'll send it out once I've got it working & tested.