On Fri, 2016-03-25 at 14:20 -0700, Dan Williams wrote:

On Fri, Mar 25, 2016 at 2:03 PM, Verma, Vishal L
[off-list ref] wrote:

quoted

On Fri, 2016-03-25 at 11:47 -0700, Dan Williams wrote:

quoted

On Thu, Mar 24, 2016 at 4:17 PM, Vishal Verma <vishal.l.verma@int
el.c
om> wrote:

quoted

From: Matthew Wilcox <redacted>

dax_clear_sectors() cannot handle poisoned blocks.  These must
be
zeroed using the BIO interface instead.  Convert ext2 and XFS
to
use
only sb_issue_zerout().

Signed-off-by: Matthew Wilcox <redacted>
[vishal: Also remove the dax_clear_sectors function entirely]
Signed-off-by: Vishal Verma <vishal.l.verma@intel.com>
---
 fs/dax.c               | 32 --------------------------------
 fs/ext2/inode.c        |  7 +++----
 fs/xfs/xfs_bmap_util.c |  9 ---------
 include/linux/dax.h    |  1 -
 4 files changed, 3 insertions(+), 46 deletions(-)

diff --git a/fs/dax.c b/fs/dax.c
index bb7e9f8..a30481e 100644
--- a/fs/dax.c
+++ b/fs/dax.c

@@ -78,38 +78,6 @@ struct page *read_dax_sector(struct

block_device
*bdev, sector_t n)
        return page;
 }

-/*
- * dax_clear_sectors() is called from within transaction
context
from XFS,
- * and hence this means the stack from this point must follow
GFP_NOFS
- * semantics for all operations.
- */
-int dax_clear_sectors(struct block_device *bdev, sector_t
_sector,
long _size)
-{
-       struct blk_dax_ctl dax = {
-               .sector = _sector,
-               .size = _size,
-       };
-
-       might_sleep();
-       do {
-               long count, sz;
-
-               count = dax_map_atomic(bdev, &dax);
-               if (count < 0)
-                       return count;
-               sz = min_t(long, count, SZ_128K);
-               clear_pmem(dax.addr, sz);
-               dax.size -= sz;
-               dax.sector += sz / 512;
-               dax_unmap_atomic(bdev, &dax);
-               cond_resched();
-       } while (dax.size);
-
-       wmb_pmem();
-       return 0;
-}
-EXPORT_SYMBOL_GPL(dax_clear_sectors);

What about the other unwritten extent conversions in the dax
path?
Shouldn't those be converted to block-layer zero-outs as well?

Could you point me to where these might be? I thought once we've
converted all the zeroout type callers (by removing
dax_clear_sectors),
and fixed up dax_do_io to try a driver fallback, we've handled all
the
media error cases in dax..

grep for usages of clear_pmem()... which I was hoping to eliminate
after this change to push zeroing down to the driver.

Ok, so I looked at these, and it looks like the majority of callers of
clear_pmem are from the fault path (either pmd or regular), and in
those cases we should be 'protected', as we would have failed at a
prior step (dax_map_atomic).

The two cases that may not be well handled are the calls to
dax_zero_page_range and dax_truncate_page which are called from file
systems. I think we may need to do a fallback to the driver for those
cases just like we do for dax_direct_io.. Thoughts?