rsync - Oct 2007 - Rsync hard-links devices with different mtimes despite -t: expected?

I noticed that rsync is happy to hard-link a device node from a
--link-dest dir even if its mtime differs from that of the source device
node and --times is given.  Is this behavior expected?  It seems to
break the rule that a difference in preserved attributes disqualifies a
hard link.

To see the behavior, run the following as root:

mkdir src dest basis
mknod src/null c 1 3
sleep 1
mknod basis/null c 1 3
rsync -a -ii --link-dest=../basis/ src/ dest/
find . -ls
stat src/null dest/null

Rsync produces the following very bogus-looking itemize output:

.d          ./
hD..t...... null

Matt

On Tue, 2007-10-30 at 17:20 -0400, Matt McCutchen wrote:
> I noticed that rsync is happy to hard-link a device node from a
> --link-dest dir even if its mtime differs from that of the source device
> node and --times is given.  Is this behavior expected?  It seems to
> break the rule that a difference in preserved attributes disqualifies a
> hard link.
The same strange behavior happens for regular files if --size-only is
passed.  The underlying issue is that unchanged_attrs compares all
preserved attributes except mtimes.  Why are mtimes omitted?

Matt

On Tue, Oct 30, 2007 at 05:20:30PM -0400, Matt McCutchen
wrote:
> I noticed that rsync is happy to hard-link a device node from a
> --link-dest dir even if its mtime differs from that of the source
> device node and --times is given.  Is this behavior expected?
Not really.  It doesn't hurt anything with devices, but we should obey
the general rule only linking things don't differ in attributes that we
are preserving.  The code in unchanged_attrs() was expecting that the
mtime would have been compared in the unchanged_file() routine, but
that assumption is not true for some of the code paths that ended up
using the routine.

The latest CVS makes unchanged_attrs() ensure that the mtime is the
same.

..wayne..

On Sat, 2007-11-03 at 14:19 -0700, Wayne Davison wrote:
> The latest CVS makes unchanged_attrs() ensure that the mtime is the
> same.
Good.  Now the mtime check in try_dests_reg at generator.c:906 is
redundant; I suggest that you remove it.

Matt