> On 19 Jun 2016, at 17:04, Lars Schneider <larsxschneider@xxxxxxxxx> wrote: > > >> On 19 Jun 2016, at 09:59, Michael Haggerty <mhagger@xxxxxxxxxxxx> wrote: >> >> On 06/18/2016 12:05 AM, Lars Schneider wrote: >>> >>>> On 17 Jun 2016, at 05:20, Junio C Hamano <gitster@xxxxxxxxx> wrote: >>>> >>>> ... >>>> >>>> * mh/split-under-lock (2016-05-13) 33 commits >>>> ... >>>> >>>> Further preparatory work on the refs API before the pluggable >>>> backend series can land. >>>> >>>> Will merge to 'master'. >>> >>> This topic seems break two git-p4 tests (t9801 and t9803) on next: >>> https://travis-ci.org/git/git/jobs/137333785 >>> >>> According to git bisect the commit "ref_transaction_update(): >>> check refname_is_safe() at a minimum" (3da1f3) introduces the problem: >>> https://s3.amazonaws.com/archive.travis-ci.org/jobs/138457628/log.txt >>> (scroll all the way down to see the bisecting) >>> >>> - Lars >>> >> >> Lars, >> >> According to [1], something in that test seems to have been trying to run >> >> git update-ref -d git-p4-tmp/6 >> >> Similarly in the other failed test. >> >> Because `update-ref` doesn't do DWIM for reference names, this is *not* >> expanded to `refs/heads/git-p4-tmp/6` or something. Previously this >> command would have quietly failed to do anything. But after >> "ref_transaction_update(): check refname_is_safe() at a minimum", `git >> update-ref` notices that `git/p4/tmp/6` is not a safe refname (according >> to `refname_is_safe()` [2]), and correctly fails with an error message. > > All errors seem to be related to the Git-P4 branch import. I am no expert > in that area because the branch import never worked for me (and I am puzzled > to some extend how it is supposed to work given the differences how branches > work in Git and P4). > > This is the offending call: > https://github.com/git/git/blob/05219a1276341e72d8082d76b7f5ed394b7437a4/git-p4.py#L3464 > > This is only a cleanup call and we could make all tests work if we remove the > cleanup and also the "cleanup successful check": > https://github.com/git/git/blob/05219a1276341e72d8082d76b7f5ed394b7437a4/t/t9801-git-p4-branch.sh#L303 > https://github.com/git/git/blob/05219a1276341e72d8082d76b7f5ed394b7437a4/t/t9801-git-p4-branch.sh#L355 > > I am a bit surprised that we do not see other errors given the fact > that the branch name is clearly invalid: > https://github.com/git/git/blob/05219a1276341e72d8082d76b7f5ed394b7437a4/t/t9803-git-p4-shell-metachars.sh#L102 > > I see two ways to proceed: > > (1) We remove the cleanup. > > (2) We sanitize the branch names (e.g. by removing invalid characters). > @Michael: Is there a function to "sanitize" a given branch name already? > > Option 1 is trivial and option 2 (my preference) shouldn't be too hard. > But maybe Luke has some insights since he added the "branch with shell char" > test in 52a4880. > > >> Even before this change, Git didn't allow such references to be created >> or updated. So I think this test failure is revealing an error in `git >> p4 clone` that went undetected before this change. >> >> Please let me know whether you agree. If so, it is realistic to fix >> `git-p4` promptly? This failure is currently blocking >> mh/split-under-lock, so if `git-p4` can't be fixed, then I'd have to >> either disable t9801 and t9803 in this patch series, or omit the >> `refname_is_safe()` check. > I am looking into option 2. After looking more into it I realized that the character "\$" in the branch name is not even the problem. The git-p4 temp refs are just not located under refs/heads. This seems to fix the issue: --- a/git-p4.py +++ b/git-p4.py @@ -2274,7 +2274,7 @@ class P4Sync(Command, P4UserMap): self.useClientSpec_from_options = False self.clientSpecDirs = None self.tempBranches = [] - self.tempBranchLocation = "git-p4-tmp" + self.tempBranchLocation = "refs/heads/git-p4-tmp" self.largeFileSystem = None if gitConfig('git-p4.largeFileSystem'): -- @Luke: Would that be an acceptable solution? Thanks, Lars > >> >> In the interest of backwards compatibility, I considered making `git >> update-ref -d` continue to fail silently for NOOP operations with unsafe >> refnames (one of the requirements being that no old_oid is specified). >> But I think that would be giving the wrong signal to scripts that are >> doing something that is invalid but pausible, like trying to delete the >> reference `../$(basename $PWD)/refs/heads/foo`. Such scripts would be >> misled into thinking the deletion was successful. And yet treating >> plausibly-sensible requests differently than obviously bogus requests >> seems like a path to madness. > Agreed! > > Cheers, > Lars -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html