On Tue, Feb 02, 2016 at 12:17:23PM +0100, Jan Kara wrote:
On Tue 02-02-16 08:47:30, Dave Chinner wrote:
> On Mon, Feb 01, 2016 at 03:51:47PM +0100, Jan Kara wrote:
> > On Sat 30-01-16 00:28:33, Matthew Wilcox wrote:
> > > On Fri, Jan 29, 2016 at 11:28:15AM -0700, Ross Zwisler wrote:
> > > > I guess I need to go off and understand if we can have DAX mappings
on such a
> > > > device. If we can, we may have a problem - we can get the
> > > > get_block() in I/O path and the various fault paths, but we don't
> > > > to get_block() when flushing via dax_writeback_mapping_range(). We
> > > > needing it the normal case by storing the sector results from
> > > > the radix tree.
> > >
> > > I think we're doing it wrong by storing the sector in the radix tree;
> > > really need to store both the sector and the bdev which is too much data.
> > >
> > > If we store the PFN of the underlying page instead, we don't have
> > > problem. Instead, we have a different problem; of the device going
> > > away under us. I'm trying to find the code which tears down PTEs
> > > the device goes away, and I'm not seeing it. What do we do about
> > > mappings of the device?
> > So I don't have a strong opinion whether storing PFN or sector is better.
> > Maybe PFN is somewhat more generic but OTOH turning DAX off for special
> > cases like inodes on XFS RT devices would be IMHO fine.
> We need to support alternate devices.
> There is a strong case for using the XFS RT device with DAX,
> especially for applications that know they are going to always use
> large/huge/giant pages to access their data files. The XFS RT device
> can guarantee allocation is always aligned to large/huge/giant page
> constraints right up to ENOSPC and throughout the production life of
> the filesystem. We have no other filesystem capable of providing
> such guarantees, which means the XFS RT device is uniquely suited to
> certain aplications with DAX...
I see, thanks for explanation. So I'm OK with changing what is stored in
the radix tree to accommodate this use case but my reservation that we IHMO
have other more pressing things to fix remains...
IMO this is pretty pressing - without it neither XFS RT devices nor DAX raw
block devices work. The case has been made above for XFS RT devices, and with
DAX raw block devices we really need a fix because the current code will cause
a kernel BUG when a user tries to fsync/msync a raw block device mmap(). This
is especially bad because, unlike with filesystems where you mount with the
dax mount option, there is no opt-in step for raw block devices.
This has to be fixed - it seems like we either figure out how to fix DAX
fsync, or we have to disable DAX on raw block devices for a kernel cycle. I'm
hoping for the former. :)