bfc58bf915b4 introduced "clever" reuse of p2 but did that convert could fail
with
abort: f1@f73e02ae52c5: not found in manifest!
when it tried to reuse a file from p2 but the file didn't exist there, for
example because filemap changes.
b5e7ec5dace8 fixed that (using changes from 64a7de6e3aa1), but with a quite
different reasoning and test case.
Add another test that makes sure this case is covered too.
Recently we fixed converting merges to correctly sync changes from p2. We missed
the case of deletes though (so p2 deleted a file that p1 had not yet deleted,
and the file does not belong to the source).
The fix is to detect when p2 doesn't have the file, so we just sync it as a
delete to p1 in the merge.
Updated the test, and verified it failed before the fix.
When converting a merge commit using a filemap convert (i.e. when moving
contents from the root of the repo into subdir1/), convert would silently drop
the entire contents of the target repo's p2. This was because when it built the
target commit, it did so by taking the target p1 and adding only the files that
changed in the source repo's merge commit.
This breaks in the case where the target repo has files that are unrelated to
the source repo (like in the case where you use convert to import a repo as a
subdirectory of another).
The fix is to use Mercurial's merge logic to detect which files in p2 we should
carry over to the merge. It follows three rules:
1) if the file belongs to the source, don't try to merge it. Rely on the list of
files provided to putcommit to be correct.
2) if the file requires merging or user input (change vs deleted), throw an
exception. We don't have enough info to do this.
3) if p2 has the newest, non-merge-requiring version of the file, take it
I've also added a test to cover this issue.
If the conversion doesn't change the hash, and the cset is public in the source,
it should be public in the destination. (This can happen if file remapping is
done that doesn't affect the initial commits.) This also propagates the secret
phase from the source, regardless of the hash, because presumably the content is
what is secret. Otherwise, the destination commit stays in the draft phase.
Maybe any draft cset with an unchanged hash should be changed to public, because
it has effectively been shared, but convert pretty strongly implies throwing
away (or at least readonly archiving) the source repo.
The change in the rollback output is because the name of the outer transaction
is now 'convert', which seems more accurate. Unfortunately, the memctx won't
indicate the hash prior to committing, so the proper phase can't be applied with
the commit.
The repo is already write locked in mercurial_sink.before().
When the progress extension is not enabled, each call to 'ui.progress' used to
issue a debug message. This results is a very verbose output and often redundant
in tests. Dropping it makes tests less volatile to factor they do not meant to
test.
We had to alter the sed trick in 'test-rename-merge2.t'. Sed is used to drop all
output from a certain point and hidding the progress output remove its anchor.
So we anchor on something else.
In plain `hg log` there is no indication that a commit closes a
branch. You can use hg log --debug, but this is too verbose. A simple
idea copied from thg and other graphical viewers is to display the
node for a closing-branch commit as a horizontal line.
I think this technically is a BC if we consider the graphlog to be
part of the stdout API, but I really can't imagine who the hell is
parsing the graphlog to determine information about commits.
Previously, a backup bundle could overwrite an existing bundle and cause user
data loss. For instance, if you have A<-B<-C and strip B, it produces backup
bundle B-backup.hg. If you then hg pull -r B B-backup.hg and strip it again, it
overwrites the existing B-backup.hg and C is lost.
The fix is to add a hash of all the nodes inside that bundle to the filename.
Fixed up existing tests and added a new test in test-strip.t
These slashes are a hangover from issue3612, fixed in d5787cfaa7cf.
Although the bugfix in that commit is correct, the test it adds
does not replicate the conditions for the bug correctly.
convert doesn't normalise double slashes in paths. Path normalization
is applied when a path is loaded into filemap and when a file lookup
request is issued to filemap.
When the source repository had a revision renaming "$new -> $old",
but the filemap a "$old -> $new" rename, the converted revision could
use either $new (deleting the file) or $old (keeping the file) when
getting the file data, depending on the lexicographical order of
those names. So the resulting revision would leave some files
untouched (as expected), but delete others arbitrarely.
When running convert with a filemap, merge parents which are ancestors
of other parents are ignored. This is hardly a problem when parents
belong to the same branch, but the result could be confusing when named
branches are involved. With:
-o-a1-a2-a3... <- A
\ \
b1-b2-b3...-m- <- B
If all b* revisions are discarded, it is useful to preserve 'm' even if
it is empty after filtering to record the branch switch.
This patch makes filemap preserve "ancestor parents" if there is no
"non-ancestor parent" on the same branch than the merge revision.
Remarks:
- I am not completely convinced by the reasons given above and those
detailed by Matt in this thread:
http://selenic.com/pipermail/mercurial-devel/2012-May/040627.html
The properties we try to preserve are not clearly defined. That said,
I know this patch already helped someone on IRC and the tests output
look reasonable.
- This is a new version of the original "convert: filemap must preserve
fast-forward merges" patch. It has exactly the same output for 2
parents merges, the additional complexity is here to handle more than
two parents.
The previous behaviour was almost as if convert.hg.ignoreerrors was always set
for revisions without parents, except that errors were silently ignored. Revlog
errors are handled as a side effect of getcopies(), but getcopies() was only
called when convert.hg.ignoreerrors was set.
Now we always call self.getcopies for root revisions, not only when
convert.hg.ignoreerrors is set, just like we do on all other revisions.
The extra call might be a bit expensive, but the proper fix for that would be
to catch these errors in another way.