lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Date:   Wed, 3 Jan 2018 11:04:30 +0100
From:   Jan Kara <jack@...e.cz>
To:     linux-mm@...ck.org
Cc:     linux-fsdevel@...r.kernel.org, linux-xfs@...r.kernel.org,
        linux-ext4@...r.kernel.org, Dan Williams <dan.j.williams@...el.com>
Subject: Filesystem crashes due to pages without buffers

Hello,

Over the years I have seen so far unexplained crashed in filesystem's
(ext4, xfs) writeback path due to dirty pages without buffers attached to
them (see [1] and [2] for relatively recent reports). This was confusing as
reclaim takes care not to strip buffers from a dirty page and both
filesystems do add buffers to a page when it is first written to - in
->page_mkwrite() and ->write_begin callbacks.

Recently I have come across a code path that is probably leading to this
inconsistent state and I'd like to discuss how to best fix the problem
because it's not obvious to me. Consider the following race:

CPU1					CPU2

addr = mmap(file1, MAP_SHARED, ...);
fd2 = open(file2, O_DIRECT | O_RDONLY);
read(fd2, addr, len)
  do_direct_IO()
    page = dio_get_page()
      dio_refill_pages()
        iov_iter_get_pages()
	  get_user_pages_fast()
            - page fault
              ->page_mkwrite()
                block_page_mkwrite()
                  lock_page(page);
                  - attaches buffers to page
                  - makes sure blocks are allocated
                  set_page_dirty(page)
              - install writeable PTE
              unlock_page(page);
    submit_page_section(page)
      - submits bio with 'page' as a buffer
					kswapd reclaims pages:
					...
					shrink_page_list()
					  trylock_page(page) - this is the
					    page CPU1 has just faulted in
					  try_to_unmap(page)
					  pageout(page);
					    clear_page_dirty_for_io(page);
					    ->writepage()
					  - let's assume page got written
					    out fast enough, alternatively
					    we could get to the same path as
					    soon as the page IO completes
					  if (page_has_private(page)) {
					    try_to_release_page(page)
					      - reclaims buffers from the
					        page
					   __remove_mapping(page)
					     - fails as DIO code still
					       holds page reference
...

eventually read completes
  dio_bio_complete(bio)
    set_page_dirty_lock(page)
      Bummer, we've just marked the page as dirty without having buffers.
      Eventually writeback will find it and filesystem will complain...

Am I missing something?
 
The problem here is that filesystems fundamentally assume that a page can
be written to only between ->write_begin - ->write_end (in this interval
the page is locked), or between ->page_mkwrite - ->writepage and above is
an example where this does not hold because when a page reference is
acquired through get_user_pages(), page can get written to by the holder of
the reference and dirtied even after it has been unmapped from page tables
and ->writepage has been called. This is not only a cosmetic issue leading
to assertion failure but it can also lead to data loss, data corruption, or
other unpleasant surprises as filesystems assume page contents cannot be
modified until either ->write_begin() or ->page_mkwrite gets called and
those calls are serialized by proper locking with problematic operations
such as hole punching etc.

I'm not sure how to fix this problem. We could 'simulate' a writeable page
fault in set_page_dirty_lock(). It is a bit ugly since we don't have a
virtual address of the fault, don't hold mmap_sem, etc., possibly
expensive, but it would make filesystems happy. Data stored by GUP user
(e.g. read by DIO in the above case) could still get lost if someone e.g.
punched hole under the buffer or otherwise messed with the underlying
storage of the page while DIO was running but arguably users could expect
such outcome.

Another possible solution would be to make sure page is writeably mapped
until GUP user drops its reference. That would be arguably cleaner but
probably that would mean we have to track number of writeable GUP page
references separately (no space space in struct page is a problem here) and
block page_mkclean() until they are dropped. Also for long term GUP users
like Infiniband or V4L we'd have to come up with some solution as we should
not block page_mkclean() for so long.

As a side note DAX needs some solution for GUP users as well. The problems
are similar there in nature, just much easier to hit. So at least a
solution for long-term GUP users can (and I strongly believe should) be
shared between standard and DAX paths.

Anybody has other ideas how to fix the problem or opinions on which
solution would be better to use or some complications I have missed?

								Honza

[1] https://www.spinics.net/lists/linux-xfs/msg10090.html
[2] https://www.spinics.net/lists/linux-ext4/msg54377.html

-- 
Jan Kara <jack@...e.com>
SUSE Labs, CR

Powered by blists - more mailing lists