Commit e0f43fa5 authored by Yang Shi's avatar Yang Shi Committed by Linus Torvalds

mm: filemap: coding style cleanup for filemap_map_pmd()

Patch series "Solve silent data loss caused by poisoned page cache (shmem/tmpfs)", v5.

When discussing the patch that splits page cache THP in order to offline
the poisoned page, Noaya mentioned there is a bigger problem [1] that
prevents this from working since the page cache page will be truncated
if uncorrectable errors happen.  By looking this deeper it turns out
this approach (truncating poisoned page) may incur silent data loss for
all non-readonly filesystems if the page is dirty.  It may be worse for
in-memory filesystem, e.g.  shmem/tmpfs since the data blocks are
actually gone.

To solve this problem we could keep the poisoned dirty page in page
cache then notify the users on any later access, e.g.  page fault,
read/write, etc.  The clean page could be truncated as is since they can
be reread from disk later on.

The consequence is the filesystems may find poisoned page and manipulate
it as healthy page since all the filesystems actually don't check if the
page is poisoned or not in all the relevant paths except page fault.  In
general, we need make the filesystems be aware of poisoned page before
we could keep the poisoned page in page cache in order to solve the data
loss problem.

To make filesystems be aware of poisoned page we should consider:

 - The page should be not written back: clearing dirty flag could
   prevent from writeback.

 - The page should not be dropped (it shows as a clean page) by drop
   caches or other callers: the refcount pin from hwpoison could prevent
   from invalidating (called by cache drop, inode cache shrinking, etc),
   but it doesn't avoid invalidation in DIO path.

 - The page should be able to get truncated/hole punched/unlinked: it
   works as it is.

 - Notify users when the page is accessed, e.g. read/write, page fault
   and other paths (compression, encryption, etc).

The scope of the last one is huge since almost all filesystems need do
it once a page is returned from page cache lookup.  There are a couple
of options to do it:

 1. Check hwpoison flag for every path, the most straightforward way.

 2. Return NULL for poisoned page from page cache lookup, the most
    callsites check if NULL is returned, this should have least work I
    think. But the error handling in filesystems just return -ENOMEM,
    the error code will incur confusion to the users obviously.

 3. To improve #2, we could return error pointer, e.g. ERR_PTR(-EIO),
    but this will involve significant amount of code change as well
    since all the paths need check if the pointer is ERR or not just
    like option #1.

I did prototypes for both #1 and #3, but it seems #3 may require more
changes than #1.  For #3 ERR_PTR will be returned so all the callers
need to check the return value otherwise invalid pointer may be
dereferenced, but not all callers really care about the content of the
page, for example, partial truncate which just sets the truncated range
in one page to 0.  So for such paths it needs additional modification if
ERR_PTR is returned.  And if the callers have their own way to handle
the problematic pages we need to add a new FGP flag to tell FGP
functions to return the pointer to the page.

It may happen very rarely, but once it happens the consequence (data
corruption) could be very bad and it is very hard to debug.  It seems
this problem had been slightly discussed before, but seems no action was
taken at that time.  [2]

As the aforementioned investigation, it needs huge amount of work to
solve the potential data loss for all filesystems.  But it is much
easier for in-memory filesystems and such filesystems actually suffer
more than others since even the data blocks are gone due to truncating.
So this patchset starts from shmem/tmpfs by taking option #1.

TODO:
* The unpoison has been broken since commit 0ed950d1 ("mm,hwpoison: make
  get_hwpoison_page() call get_any_page()"), and this patch series make
  refcount check for unpoisoning shmem page fail.
* Expand to other filesystems.  But I haven't heard feedback from filesystem
  developers yet.

Patch breakdown:
Patch #1: cleanup, depended by patch #2
Patch #2: fix THP with hwpoisoned subpage(s) PMD map bug
Patch #3: coding style cleanup
Patch #4: refactor and preparation.
Patch #5: keep the poisoned page in page cache and handle such case for all
          the paths.
Patch #6: the previous patches unblock page cache THP split, so this patch
          add page cache THP split support.

This patch (of 4):

A minor cleanup to the indent.

Link: https://lkml.kernel.org/r/20211020210755.23964-1-shy828301@gmail.com
Link: https://lkml.kernel.org/r/20211020210755.23964-4-shy828301@gmail.comSigned-off-by: default avatarYang Shi <shy828301@gmail.com>
Reviewed-by: default avatarNaoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Xu <peterx@redhat.com>
Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
parent ba9eb3ce
......@@ -3203,12 +3203,12 @@ static bool filemap_map_pmd(struct vm_fault *vmf, struct page *page)
}
if (pmd_none(*vmf->pmd) && PageTransHuge(page)) {
vm_fault_t ret = do_set_pmd(vmf, page);
if (!ret) {
/* The page is mapped successfully, reference consumed. */
unlock_page(page);
return true;
}
vm_fault_t ret = do_set_pmd(vmf, page);
if (!ret) {
/* The page is mapped successfully, reference consumed. */
unlock_page(page);
return true;
}
}
if (pmd_none(*vmf->pmd))
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment