Re: [PATCH 04/10] mm, fsdax: Refactor memory-failure handler for dax mapping

On Thu, Jan 14, 2021 at 05:38:33PM +0800, zhong jiang wrote:

On 2021/1/14 11:52 上午, Ruan Shiyang wrote:

quoted

On 2021/1/14 上午11:26, zhong jiang wrote:

quoted

On 2021/1/14 9:44 上午, Ruan Shiyang wrote:

quoted

On 2021/1/13 下午6:04, zhong jiang wrote:

quoted

On 2021/1/12 10:55 上午, Ruan Shiyang wrote:

quoted

On 2021/1/6 下午11:41, Jan Kara wrote:

quoted

On Thu 31-12-20 00:55:55, Shiyang Ruan wrote:

quoted

The current memory_failure_dev_pagemap() can
only handle single-mapped
dax page for fsdax mode.  The dax page could be
mapped by multiple files
and offsets if we let reflink feature & fsdax
mode work together. So,
we refactor current implementation to support
handle memory failure on
each file and offset.

Signed-off-by: Shiyang Ruan <redacted>

Overall this looks OK to me, a few comments below.

quoted

---
  fs/dax.c            | 21 +++++++++++
  include/linux/dax.h |  1 +
  include/linux/mm.h  |  9 +++++
  mm/memory-failure.c | 91
++++++++++++++++++++++++++++++++++-----------
  4 files changed, 100 insertions(+), 22 deletions(-)

...

quoted

quoted

  @@ -345,9 +348,12 @@ static void
add_to_kill(struct task_struct *tsk, struct page
*p,
      }
        tk->addr = page_address_in_vma(p, vma);
-    if (is_zone_device_page(p))
-        tk->size_shift = dev_pagemap_mapping_shift(p, vma);
-    else
+    if (is_zone_device_page(p)) {
+        if (is_device_fsdax_page(p))
+            tk->addr = vma->vm_start +
+                    ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);

It seems strange to use 'pgoff' for dax pages and
not for any other page.
Why? I'd rather pass correct pgoff from all callers
of add_to_kill() and
avoid this special casing...

Because one fsdax page can be shared by multiple pgoffs.
I have to pass each pgoff in each iteration to calculate
the address in vma (for tk->addr).  Other kinds of pages
don't need this. They can get their unique address by
calling "page_address_in_vma()".

IMO,   an fsdax page can be shared by multiple files rather
than multiple pgoffs if fs query support reflink.   Because
an page only located in an mapping(page->mapping is
exclusive), hence it  only has an pgoff or index pointing at
the node.

  or  I miss something for the feature ?  thanks,

Yes, a fsdax page is shared by multiple files because of
reflink. I think my description of 'pgoff' here is not correct. 
This 'pgoff' means the offset within the a file. (We use rmap to
find out all the sharing files and their offsets.)  So, I said
that "can be shared by multiple pgoffs".  It's my bad.

I think I should name it another word to avoid misunderstandings.

IMO,  All the sharing files should be the same offset to share the
fsdax page.  why not that ?

The dedupe operation can let different files share their same data
extent, though offsets are not same.  So, files can share one fsdax page
at different offset.

Ok,  Get it.

quoted

quoted

As you has said,  a shared fadax page should be inserted to
different mapping files.  but page->index and page->mapping is
exclusive.  hence an page only should be placed in an mapping tree.

We can't use page->mapping and page->index here for reflink & fsdax. And
that's this patchset aims to solve.  I introduced a series of
->corrupted_range(), from mm to pmem driver to block device and finally
to filesystem, to use rmap feature of filesystem to find out all files
sharing same data extent (fsdax page).

From this patch,  each file has mapping tree,  the shared page will be
inserted into multiple file mapping tree.  then filesystem use file and
offset to get the killed process.   Is it correct?

FWIW I thought the purpose of this patchset is to remove the (dax)
memory poison code's reliance on the pagecache mapping structure by
pushing poison notifications directly into the filesystem and letting
the filesystem perform reverse lookup operations to figure out which
file(s) have gone bad, and using the file list to call back into the mm
to kill processes.

Once that's done, I think(?) that puts us significantly closer to being
able to share pmem between files in dax mode without having to rewrite
the entire memory manager's mapping and rmapping code to support
sharing.

--D

Thanks,

quoted

-- 
Thanks,
Ruan Shiyang.

quoted

And In the current patch,  we failed to found out that all process
use the fsdax page shared by multiple files and kill them.


Thanks,

quoted

-- 
Thanks,
Ruan Shiyang.

quoted

quoted

So, I added this fsdax case here. This patchset only
implemented the fsdax case, other cases also need to be
added here if to be implemented.


-- 
Thanks,
Ruan Shiyang.

quoted

quoted

+        tk->size_shift =
dev_pagemap_mapping_shift(p, vma, tk->addr);
+    } else
          tk->size_shift = page_shift(compound_head(p));
        /*

@@ -495,7 +501,7 @@ static void

collect_procs_anon(struct page *page, struct
list_head *to_kill,
              if (!page_mapped_in_vma(page, vma))
                  continue;
              if (vma->vm_mm == t->mm)
-                add_to_kill(t, page, vma, to_kill);
+                add_to_kill(t, page, NULL, 0, vma, to_kill);
          }
      }
      read_unlock(&tasklist_lock);

@@ -505,24 +511,19 @@ static void

collect_procs_anon(struct page *page, struct
list_head *to_kill,
  /*
   * Collect processes when the error hit a file mapped page.
   */
-static void collect_procs_file(struct page
*page, struct list_head *to_kill,
-                int force_early)
+static void collect_procs_file(struct page
*page, struct address_space *mapping,
+        pgoff_t pgoff, struct list_head *to_kill, int force_early)
  {
      struct vm_area_struct *vma;
      struct task_struct *tsk;
-    struct address_space *mapping = page->mapping;
-    pgoff_t pgoff;
        i_mmap_lock_read(mapping);
      read_lock(&tasklist_lock);
-    pgoff = page_to_pgoff(page);
      for_each_process(tsk) {
          struct task_struct *t =
task_early_kill(tsk, force_early);
-
          if (!t)
              continue;
-        vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
-                      pgoff) {
+        vma_interval_tree_foreach(vma,
&mapping->i_mmap, pgoff, pgoff) {
              /*
               * Send early kill signal to tasks where a vma covers
               * the page but the corrupted page is not necessarily

@@ -531,7 +532,7 @@ static void

collect_procs_file(struct page *page, struct
list_head *to_kill,
               * to be informed of all such data corruptions.
               */
              if (vma->vm_mm == t->mm)
-                add_to_kill(t, page, vma, to_kill);
+                add_to_kill(t, page, mapping,
pgoff, vma, to_kill);
          }
      }
      read_unlock(&tasklist_lock);

@@ -550,7 +551,8 @@ static void

collect_procs(struct page *page, struct
list_head *tokill,
      if (PageAnon(page))
          collect_procs_anon(page, tokill, force_early);
      else
-        collect_procs_file(page, tokill, force_early);
+        collect_procs_file(page, page->mapping,
page_to_pgoff(page),

Why not use page_mapping() helper here? It would be
safer for THPs if they
ever get here...

                                Honza