Re: [fuse-devel] Violation of path[0] == '/' invariant on rm -rf + threads

[Lucas C. V. R. <lu...@go...>]

On Wed, Oct 13, 2010 at 6:38 AM, Goswin von Brederlow <gos...@we...> wrote:
> Hraban Luyat <hr...@0b...> writes:
>
>> Hello all,
>>
>> While working on a FUSE application in C I decided to add an
>> assert(path[0] == '/') to every operation handler from fusexmp.c to
>> check the invariant. Apparently, releasedir() is sometimes passed path
>> '-'. Only when threading is enabled: the -s flag solves this issue.
>> This can be seen simply when doing rm -rf on a big directory with lots
>> of small files (I use a git repository to test it with; 9/10
>> reproducible, different spot in the tree every time). Interestingly,
>> this does not happen when populating the target with a cp -r, which,
>> presumably, also uses a lot of releasedir. Only rm -rf, so far.
>
> I believe this happens when a directory is released that is already been
> deleted.

That's also what I've observed on my side.

To ensure the provided path is valid we keep a "deleted" member in the
dentry structure that we attach to the fuse_file_info structure. If a
dentry has the "deleted" bit set then we don't trust the "path" that's
provided to the function.

Thanks,

-- 
Lucas
"If you're looking for a reason I've a reason to give: pleasure,
little treasure"