On Wed, May 30, 2018 at 1:13 AM, Jan Kara [off-list ref] wrote:

On Tue 29-05-18 18:38:41, Dan Williams wrote:

quoted

On Wed, May 23, 2018 at 1:40 AM, Jan Kara [off-list ref] wrote:

quoted

On Tue 22-05-18 07:39:57, Dan Williams wrote:

quoted

In support of enabling memory_failure() handling for filesystem-dax
mappings, set ->index to the pgoff of the page. The rmap implementation
requires ->index to bound the search through the vma interval tree. The
index is set and cleared at dax_associate_entry() and
dax_disassociate_entry() time respectively.

Cc: Jan Kara <jack@suse.cz>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Matthew Wilcox <redacted>
Cc: Ross Zwisler <redacted>
Signed-off-by: Dan Williams <redacted>
---
 fs/dax.c |   11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/fs/dax.c b/fs/dax.c
index aaec72ded1b6..2e4682cd7c69 100644
--- a/fs/dax.c
+++ b/fs/dax.c

@@ -319,18 +319,22 @@ static unsigned long dax_radix_end_pfn(void *entry)
      for (pfn = dax_radix_pfn(entry); \
                      pfn < dax_radix_end_pfn(entry); pfn++)

-static void dax_associate_entry(void *entry, struct address_space *mapping)
+static void dax_associate_entry(void *entry, struct address_space *mapping,
+             struct vm_area_struct *vma, unsigned long address)
 {
-     unsigned long pfn;
+     unsigned long size = dax_entry_size(entry), pfn, index;
+     int i = 0;

      if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
              return;

+     index = linear_page_index(vma, address & ~(size - 1));
      for_each_mapped_pfn(entry, pfn) {
              struct page *page = pfn_to_page(pfn);

              WARN_ON_ONCE(page->mapping);
              page->mapping = mapping;
+             page->index = index + i++;
      }
 }

Hum, this just made me think: How is this going to work with XFS reflink?
In fact is not the page->mapping association already broken by XFS reflink?
Because with reflink we can have two or more mappings pointing to the same
physical blocks (i.e., pages in DAX case)...

Good question. I assume we are ok in the non-DAX reflink case because
rmap of failing / poison pages is only relative to the specific page
cache page for a given inode in the reflink. However, DAX would seem
to break this because we only get one shared 'struct page' for all
possible mappings of the physical file block. I think this means for
iterating over the rmap of "where is this page mapped" would require
iterating over the other "sibling" inodes that know about the given
physical file block.

As far as I can see reflink+dax would require teaching kernel code
paths that ->mapping may not be a singular relationship. Something
along the line's of what Jerome was presenting at LSF to create a
special value to indicate, "call back into the filesystem (or the page
owner)" to perform this operation.

In the meantime the kernel crashes when userspace accesses poisoned
pmem via DAX. I assume that reworking rmap for the dax+reflink case
should not block dax poison handling? Yell if you disagree.

The thing is, up until get_user_pages() vs truncate series ("fs, dax: use
page->mapping to warn if truncate collides with a busy page" in
particular), DAX was perfectly fine with reflinks since we never needed
page->mapping.

Sure, but if this rmap series had come first I still would have needed
to implement ->mapping. So unless we invent a general ->mapping
replacement and switch all mapping users, it was always going to
collide with DAX eventually.

Now this series adds even page->index dependency which makes
life for rmap with reflinks even harder. So if nothing else we should at
least make sure reflinked filesystems cannot be mounted with dax mount
option for now and seriously start looking into how to implement rmap with
reflinked files for DAX because this noticeably reduces its usefulness.

This restriction is already in place. In xfs_reflink_remap_range() we have:

        /* Don't share DAX file data for now. */
        if (IS_DAX(inode_in) || IS_DAX(inode_out))
                goto out_unlock;

All this said, perhaps we don't need to set ->link, it would just mean
a wider search through the rmap tree to find if the given page is
mapped. So, I think I can forgo setting ->link if I teach the rmap
code to search the entire ->mapping.