Re: [fuse-devel] Semantics of fuse_notify_delete()

[Nikolaus Rath <Nikolaus@rath.org>]

On Jul 27 2023, Nikolaus Rath <Nikolaus@rath.org> wrote:
> On Jul 27 2023, Miklos Szeredi via fuse-devel <fuse-devel@lists.sourceforge.net> wrote:
>> On Wed, 26 Jul 2023 at 20:09, Nikolaus Rath <Nikolaus@rath.org> wrote:
>>>
>>> Hello,
>>>
>>> It seems to me that fuse_notify_delete
>>> (https://elixir.bootlin.com/linux/v6.1/source/fs/fuse/dev.c#L1512) fails
>>> with ENOTEMPTY if there is a pending FORGET request for a directory
>>> entry within. Is that correct?
>>
>> It's bug if it does that.
>>
>> The code related to NOTIFY_DELETE in fuse_reverse_inval_entry() seems
>> historic.  It's supposed to be careful about mountpoints and
>> referenced dentries, but d_invalidate() should have already gotten all
>> that out of the way and left an unhashed dentry without any submounts
>> or children. The checks just seem redundant, but not harmful.
>>
>> If you are managing to trigger the ENOTEMPTY case, then something
>> strange is going on, and we need to investigate.
>
> I can trigger this reliable on kernel 6.1.0-10-amd64 (Debian stable)
> with this sequence of operations:
>
> $ mkdir test
> $ echo foo > test/bar
> $ Trigger removal of test/bar and then test within the filesystem (not
> through unlink()/rmdir() but out-of-band)
>
>
> What can I do to help with the investigation?


I've pushed an instrumented snapshot to
https://github.com/s3ql/s3ql/tree/notify_delete_bug. For me, this
reliably reproduces the problem:

$ python3 setup.py build_cython build_ext --inplace
$ md bucket
$ bin/mkfs.s3ql --plain local://bucket
[...]
$ bin/mount.s3ql --fg local://bucket mnt &
[...]
$ md mnt/test; echo foo > mnt/test/bar
$ bin/s3qlrm mnt/test
fuse: writing device: Directory not empty
ERROR: Failed to submit invalidate_entry request for parent inode 1, name b'test'
Traceback (most recent call last):
  File "src/internal.pxi", line 125, in pyfuse3._notify_loop
  File "src/pyfuse3.pyx", line 915, in pyfuse3.invalidate_entry
OSError: [Errno 39] fuse_lowlevel_notify_delete returned: Directory not
empty


I've looked into reproducing this with e.g. example/passthrough_ll.c,
but it's non-trivial because of the need to run notify_delete in a
separate thread and giving it all the right arguments. I can look into
it more if that's needed though.

Best,
-Nikolaus