I'll see what I can do; possibly retrying the merge against a copy of
the target from just before we actually committed it would work. (I
didn't think I could come up with a small test case either, the
branches in question have been touched with various 1.5.x versions...
I wonder if it's worth trying to cook up a mergeinfo-fsck or
mergeinfo-lint to try and massage an existing repository into
more-correct mergeinfo...)
On Tue, Mar 17, 2009 at 1:28 PM, Paul Burba <ptburba_at_gmail.com> wrote:
> On Mon, Mar 16, 2009 at 8:24 PM, Paul Burba <ptburba_at_gmail.com> wrote:
>> On Thu, Mar 12, 2009 at 11:59 PM, Mark Eichin <eichin_at_gmail.com> wrote:
>>> A coworker of mine (Ken Baker) got a segfault merging a branch back
>>> into trunk, under 1.5.5, reproduced with 1.5.6. Diving into the code
>>> turns up a straightforard invariant issue - some code assumes that all
>>> elements of the children_with_mergeinfo array have values, but the use
>>> of remove_children_with_deleted_mergeinfo makes that not universally
>>> true. Bug (as described below) exists on trunk as well.
>>>
>>> (I relate this to issue 2821 because some of the places that *do*
>>> defend against NULL were added in r30464 and r26803; issue 3067 might
>>> also apply. That combined with the fact that it's mergeinfo related
>>> suggests that Paul Burba should have a look at it...)
>>
>> Hi Mark,
>>
>> Thanks for the detailed bug report and analysis. I fixed this in
>> r36613, but instead of adding more checks for NULL
>> children_with_mergeinfo elements I stopped setting these elements to
>> NULL in the first place, and just deleted them.
>>
>> I'll nominate this change for backport to 1.5.x and 1.6.x.
>
> Made a mistake in r36613, fixed in r36631.
>
> Mark - I should have asked this before; I was unable to work up a test
> case that provoked the segfault, so could you try the offending merge
> again with trunk?
>
> Paul
>
>> Paul
>>
>>> The particular segfault was do_directory_merge calling
>>> drive_merge_report_editor which calls find_nearest_ancestor, which
>>> does
>>>
>>> svn_client__merge_path_t *child =
>>> APR_ARRAY_IDX(children_with_mergeinfo, i, svn_client__merge_path_t *);
>>>
>>> and immediately uses child without checking it. Many places in
>>> merge.c *do* check that randomly selected values in
>>> children_with_mergeinfo are not NULL first, because
>>> drive_merge_report_editor and remove_children_with_deleted_mergeinfo
>>> are called in a loop, so most things *can* be called with a
>>> children_with_mergeinfo that has cleared slots in it; the patch is
>>> based on replicating that existing pattern.
>>>
>>> Fixing the first one exposes a second: do_directory_merge calls
>>> process_children_with_new_mergeinfo which calls
>>> find_child_with_mergeinfo which calls find_child_or_parent, which also
>>> lacks protection (of potential_child_or_parent, same thing.)
>>>
>>> The following patch fixes both, and causes our merge to complete
>>> successfully (and by inspection, correctly.)
>>>
>>> Rule 3 suggests a scan of the rest of merge.c - the only suspicious
>>> case remaining is near the top of drive_merge_report_editor, in the
>>> "if (honor_mergeinfo)" clause, but since we haven't seen that one fail
>>> I don't know if conditions allow it to - it's also not an obvious
>>> "search loop" pattern like the others where NULL is an obvious "not
>>> matched" case, so we didn't touch that one.
>>>
>>> _Mark_ <eichin_at_metacarta.com>
>>>
>>> ps. Rule 3 of bugs: once you've fixed the bug, look for other places
>>> you made the same mistake...
>>
>
--
_Mark_ <eichin_at_thok.org> <eichin_at_gmail.com>
------------------------------------------------------
http://subversion.tigris.org/ds/viewMessage.do?dsForumId=462&dsMessageId=1342499
Received on 2009-03-17 18:44:26 CET