[PATCH 0/2] fuse: fix races related to fuse writeback

[Maxim Patlasov <MPatlasov@parallels.com>]

Hi,

The patchset fixes a few subtle races stemmed from incorrect expectation
of what fuse_set_nowrite() guarantees. The fact that it makes fi->writectr
negative and waits for fi->writectr == FUSE_NOWRITE ensures only two things:

1) If there are any in-flight writeback requests right now, let's wait for
them being completed.
2) Suspend processing new writeback requests until fuse_release_nowrite().

Both are related to communication between in-kernel fuse and userspace
fuse daemon. But fuse_set_nowrite() does not prevent generic kernel code
from sending dirty pages to writeback resulting in fuse_writepage being
called. I.e. fi->queued_writes may grow independently on fuse_set_nowrite()
machinery.

As soon as fuse_writepage_locked() called end_page_writeback() generic
kernel code may do with the page virtually anything w/o notifying fuse. See
per-patch descriptions for details of some races.

Thanks,
Maxim

---

Maxim Patlasov (2):
      fuse: postpone end_page_writeback() in fuse_writepage_locked()
      fuse: wait for writeback in fuse_file_fallocate()


 fs/fuse/file.c |   56 ++++++++++++++++++++++++++++++++++++++++++++++----------
 1 files changed, 46 insertions(+), 10 deletions(-)

-- 
Signature