bug#29038: df hangs on fifos/named pipes

[Pádraig Brady]

On 28/10/17 00:18, Stephane Chazelas wrote:
> test case:
> 
>    mkfifo p
>    df p
> 
> That hangs, unless you make "p" non-readable or some other process
> has the fifo open in write mode.
> 
> The reason is that df tries to open the fifo in read-only mode,
> according to comments in the source code so as to trigger a
> potential automout.
> 
> That goes back to this commit:
> 
>> commit dbd17157d7e693b8de9737f802db0e235ff5a3e6
>> Author: Tomas Smetana <address@hidden>
>> Date:   Tue Apr 28 11:21:49 2009 +0200
>>
>>     df: use open(2), not stat, to trigger automounting
>>
>>     * src/df.c (main): When iterating over command-line arguments,
>>     attempting to ensure each backing file system is mounted, use
>>     open, not stat.  stat is no longer sufficient to trigger
>>     automounting, in some cases.  Based on a suggestion from Ian Kent.
>>     More details in http://bugzilla.redhat.com/497830
> 
> More info at the bugzilla link.
> 
> It's arguable whether df, a reporting tool, should have such a
> side effect as automounting a file system.
> 
> The fifo issue though is a bug IMO, especially considering that
> POSIX explicitely says that df should work on fifos.
> 
> Here, it may be enough to add the O_DIRECTORY flag to open()
> where available if we only care about automounting files of type
> directory (or portably use opendir()).
> 
> Or use O_PATH  on Linux 3.6+ followed by openat() on non-fifos if
> open(O_PATH) is not enough to trigger the automount in the
> unlikely event we care about automounting non-directory files
> (and report their disk usage).
> 
> Or not open() at all, and not automount file systems.
> 
> Note that busybox, heirloom or ast-open df implementations on
> Linux don't have the problem (and presumably don't automount
> file systems). Nor does FreeBSD.
> 
> Reproduced with:
> 
> $ df --version
> df (GNU coreutils) 8.25
> 
> and:
> 
> $ df --version
> df (GNU coreutils) 8.27.46-e13fe
> 
> That was discovered by Martijn Dekker, CCed, when looking for a
> portable way to identify the file system of an arbitrary file.
> 

Yes we shouldn't hang.

RE side effects, open() is a fairly innocuous operation,
and we expect stat() to have side effects to auto mount.
(Note we avoid stat() with non specified arguments if possible due to this):
https://git.sv.gnu.org/gitweb/?p=coreutils.git;a=commitdiff;h=v8.27-92-ga19ff5d

I suppose we could stat() and if that succeeds && !fifo, only then call open() ?
Patch to do that is attached.

cheers,
Pádraig

df-fifo-hang.patch
Description: Text Data