Re: pw keeps setting /etc/group to 0600

On Sat, Nov 17, 2012 at 11:20:21AM -0500, Ryan Stone wrote:
> My original complaint that /etc/group gets permissions of 0600 is a result
> of a bug in libutil, which bapt_at_ ported pw to use in r242349.  The new
> group manipulation API using mktemp to create a temporary file, writes the
> new group database to the temp file and then renames the temp file to
> /etc/group.  The problem here is that mktemp creates a file with a mode of
> 600, and libutil never chmods it.  That should be pretty trivial to fix.

My additional 0,03$:

I took closer look to this and I think that problems are much broader
than this. I don't know if similar problems were present before.

First, pw should not fail if other instance is running, it should wait
instead (think of parallel batch scripts adding some users/groups).

Second, current code has a race:
lockfd = open(group_file, O_RDONLY, 0);
if (lockfd < 0 || fcntl(lockfd, F_SETFD, 1) == -1)
	err(1, "%s", group_file);
if (flock(lockfd, LOCK_EX|LOCK_NB) == -1) {
[..]
gr_copy(pfd, tfd, gr, old_gr); /* copy from groupfile to tempfile */
[..]
rename(tempfile,groupfile);

Now let's consider threads A and B:

A: open()
A: lock();
A: gr_copy
B: open()

Now B has file descriptor to /etc/group that is about to be removed.

A: rename()
A: unlock()
B: lock()

Now B has a lock on unlinked file.

B: gr_copy()
B: rename()

... and stores new content losing modifications done by A

Third, I don't like current api.
gr_lock and gr_tmp have no arguments (that matter anyway)
gr_copy operates on two descriptors given as arguments
gr_mkdb takes nothing and is expected to do The Right Thing

I think descriptos should be hidden. Also I think that passing around
struct gr_something (sorry, never had talent for names) that would
contain all necessary data would be nice.

-- 
Mateusz Guzik <mjguzik gmail.com>