On Tue, May 29 2018 at 3:51pm -0400,
Ross Zwisler <ross.zwisler(a)linux.intel.com> wrote:
Currently the code in dm_dax_direct_access() only checks whether the
type has a direct_access() operation defined, not whether the underlying
block devices all support DAX. This latter property can be seen by looking
at whether we set the QUEUE_FLAG_DAX request queue flag when creating the
Wait... I thought DAX support was all or nothing?
This is problematic if we have, for example, a dm-linear device made
a PMEM namespace in fsdax mode followed by a ramdisk from BRD.
QUEUE_FLAG_DAX won't be set on the dm-linear device's request queue, but
we have a working direct_access() entry point and the first member of the
dm-linear set *does* support DAX.
If you don't have a uniformly capable device then it is very dangerous
to advertise that the entire device has a certain capability. That
completely bit me in the past with discard (because for every IO I
wasn't then checking if the destination device supported discards).
It is all well and good that you're adding that check here. But what I
don't like is how you're saying QUEUE_FLAG_DAX implies direct_access()
operation exists.. yet for raw PMEM namespaces we just discussed how
that is a lie.
SO this type of change showcases how the QUEUE_FLAG_DAX doesn't _really_
imply direct_access() exists.
This allows the user to create a filesystem on the dm-linear device,
then mount it with DAX. The filesystem's bdev_dax_supported() test will
pass because it'll operate on the first member of the dm-linear device,
which happens to be a fsdax PMEM namespace.
All DAX I/O will then fail to that dm-linear device because the lack of
QUEUE_FLAG_DAX prevents fs_dax_get_by_bdev() from working. This means that
the struct dax_device isn't ever set in the filesystem, so
dax_direct_access() will always return -EOPNOTSUPP.
Now you've lost me... these past 2 paragraphs. Why can a user mount it
is DAX mode? Because bdev_dax_supported() only accesses the first
portion (which happens to have DAX capabilities?)
Isn't this exactly why you should be checking for QUEUE_FLAG_DAX in the
caller (bdev_dax_supported)? Why not use bdev_get_queue() and verify
QUEUE_FLAG_DAX is set in there?
By failing out of dm_dax_direct_access() if QUEUE_FLAG_DAX isn't
set we let
the filesystem know we don't support DAX at mount time. The filesystem
will then silently fall back and remove the dax mount option, causing it to
This shouldn't be needed. Again, QUEUE_FLAG_DAX wasn't set.. so don't
allow code to falsely try operations that should've been gated by the
fact it wasn't set.
SO Nack on this patch.. until/unless I'm corrected ;)
Signed-off-by: Ross Zwisler <ross.zwisler(a)linux.intel.com>
Fixes: commit 545ed20e6df6 ("dm: add infrastructure for DAX support")
drivers/md/dm.c | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 0a7b0107ca78..9728433362d1 100644
@@ -1050,14 +1050,13 @@ static long dm_dax_direct_access(struct dax_device *dax_dev,
- if (!ti->type->direct_access)
+ if (!blk_queue_dax(md->queue))
len = max_io_len(sector, ti) / PAGE_SECTORS;
if (len < 1)
nr_pages = min(len, nr_pages);
- if (ti->type->direct_access)
- ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);
+ ret = ti->type->direct_access(ti, pgoff, nr_pages, kaddr, pfn);