Re: [PATCH v2 4/4] proc: avoid ->f_pos overflows inproc_task_readdir() paths

From: Al Viro
Date: Tue Jun 04 2013 - 17:06:48 EST

On Tue, Jun 04, 2013 at 09:57:00PM +0200, Oleg Nesterov wrote:
> Will do... but so far I am confused.
> I do not see how they could race (I mean /proc/pid/task only). OK, OK,
> the usage of ->f_pos in sys_getdents() looks "obviously wrong", but this
> is another story? And "put f_pos in a local variable" can't help.

For one thing, a bunch of directories use generic_file_llseek(), which
does *not* use ->i_mutex. For another, there's a very unpleasant problem
with read(2) (failing) attempt racing with ->f_pos modifications in
->readdir(). Take a look at sys_read() and note that it is done with no
serialization at all (not in the top level, that is) and that it puts the
(unmodified by generic_read_dir()) value of pos back into file->f_pos as
soon as vfs_read() passes -EISDIR (returned by generic_read_dir()) back to

I.e. ->f_pos is silently reset back to the value it used to have on the
entry to read(2). Despite foo_readdir() assumptions that it won't be
changed behind its back.

Reset itself wouldn't be a problem - if several threads mess with read()
on the same opened file in parallel, you are not promised anything good
about the resulting IO pointer position. The same applies here. However,
many ->readdir() instances use file->f_pos as a variable they can use for
internal needs and _that_ leads to very unpleasant races.

The sane solution is to do what ->read()/->write()/etc. are doing - pass
an address of local copy of ->f_pos, so they are able to use it without
worrying about concurrent modifications of that value. That obviously
solves all problems with generic_file_lseek(), etc., as well as this
sys_read() shite.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at