On Fri, Feb 07, 2020 at 12:49:54PM +0100, SZEDER Gábor wrote:
> On Fri, Feb 07, 2020 at 12:10:08PM +0100, SZEDER Gábor wrote:
> > That's a good subject, isn't it? :)
> >
> > So, to clarify: apparently it is possible to abort an ongoing 'git
> > rebase' process with ctrl-C in just the right moment that a subsequent
> > 'git rebase --abort' will refuse to clear it up.
> >
> > I somehow messed up the upstream and branch parameters of 'git
> > rebase', and ended up trying to rebase a fairly recent (post v2.24.0)
> > branch on top of v2.22.0. Upon seeing the unexpectedly large number
> > of patches I realized that something is wrong, hit ctrl-C right away,
> > and this is what happened:
> >
> > $ git rebase v2.22.0 <a-branch-on-top-of-2.24.0>
> > First, rewinding head to replay your work on top of it...
> > Generating patches: 100% (1108/1108), done.
> > Applying: send-email: move the read_config() function above getopts
> > Applying: send-email: rename the @bcclist variable for consistency
> > Applying: send-email: do defaults -> config -> getopt in that order
> > Using index info to reconstruct a base tree...
> > M git-send-email.perl
> > M t/t9001-send-email.sh
> > Falling back to patching base and 3-way merge...
> > Auto-merging t/t9001-send-email.sh
> > Auto-merging git-send-email.perl
> > ^C
> > ((5f07da12ac...) *|REBASE 3/1108)$ git rebase --abort
> > error: could not read '/home/szeder/src/git/.git/worktrees/WT/rebase-apply/head-name': No such file or directory
> >
> > "Fortunately" it was in a separate worktree, so I could easily get out
> > of the situation by forcibly deleting that worktree. Unfortunately,
> > that was exactly what I did, instead of securing the failed state for
> > later analysis... sorry.
>
> > All this is with a git built from current 'next', with a bunch of
> > unrelated (none of them touches rebase or the sequencer) patches on
> > top.
>
> Trying to reproduce it is a hit or miss... well, mostly miss :)
>
> There is a rather short window when 'git rebase' applies patches
> before hitting a first merge conflict. If the ctrl-C arrives before
> 'git rebase' starts applying patches, then it cleans everything up,
> and we are not rebasing, so there is no need for 'git rebase --abort'.
> Once 'git rebase' stops because of the merge conflict we get our shell
> back, and 'git rebase --abort' works as it should. But after a good
> couple of tries I managed to hit ctrl-C while 'git rebase' was
> applying patches:
This patch below increases the size of the window where a ctrl-C gets
us into that problematic situation:
diff --git a/builtin/am.c b/builtin/am.c
index 8181c2aef3..5d62766000 100644
--- a/builtin/am.c
+++ b/builtin/am.c
@@ -1737,6 +1737,14 @@ static void am_run(struct am_state *state, int resume)
exit(1);
say(state, stdout, _("Applying: %.*s"), linelen(state->msg), state->msg);
+ {
+ int i;
+ for (i = 0; i < 60; i++) {
+ fprintf(stderr, "sleeping... %d\r", i);
+ sleep(1);
+ }
+ fprintf(stderr, "\n");
+ }
apply_status = run_apply(state, NULL);
And then we can reliably reproduce the issue even when rebasing only a
single commit:
$ ./bin-wrappers/git rebase v2.25.0 9c8a294
First, rewinding head to replay your work on top of it...
Applying: sha1-file: remove OBJECT_INFO_SKIP_CACHED
^Ceeping... 3
((v2.25.0)|REBASE 1/1)$ ./bin-wrappers/git rebase --abort
error: could not read '.git/rebase-apply/head-name': No such file or directory
Note that the sleep() calls were added to 'builtin/am.c', i.e. this
issue is present in the 'am' rebase backend. I tried to break 'git
rebase -m ...' by adding sleep()s to various places in pick_commits()
in 'sequencer.c', but, FWIW, the subsequent 'git rebase --abort'
always worked as expected.
So we may have yet another reason to switch the default rebase backend
from 'am' to 'merge'.
> Finally, note the 'v2.24.0^{commit}' parameter, in particular the
> '^{commit}' part. That's important, because without it we stumble
> upon _another_ bug:
>
> $ git rebase v2.22.0 v2.24.0
> error: Object 1cc4bc0fcd93f816d514d77c29f2cc9ffdd8ae09 not a commit
> First, rewinding head to replay your work on top of it...
> Generating patches: 100% (1049/1049), done.
> Applying: send-email: move the read_config() function above getopts
> < ... applying further patches and hitting a merge conflict ... >
> Resolve all conflicts manually, mark them as resolved with
> "git add/rm <conflicted_files>", then run "git rebase --continue".
> You can instead skip this commit: run "git rebase --skip".
> To abort and get back to the state before "git rebase", run "git rebase --abort".
> (detached HEAD *+|REBASE 7/1049)$ git rebase --abort
> error: update_ref failed for ref 'HEAD': cannot update ref 'HEAD': trying to write non-commit object 1cc4bc0fcd93f816d514d77c29f2cc9ffdd8ae09 to branch 'HEAD'
>
> So 'git rebase' shows an error right at the beginning when rebasing a
> tag, but then continues anyway. However, 'git rebase --abort' can't
> restore the original state.
Now let's rebase a tag with the 'merge' backend, and then try to abort
it:
$ ./BUILDS/v2.25.0/bin/git rebase -m v2.22.0 v2.24.0
error: Object 1cc4bc0fcd93f816d514d77c29f2cc9ffdd8ae09 not a commit
Auto-merging t/t9001-send-email.sh
Auto-merging git-send-email.perl
CONFLICT (content): Merge conflict in git-send-email.perl
Auto-merging Documentation/git-send-email.txt
error: could not apply 3ff15040e2... send-email: fix regression in sendemail.identity parsing
Resolve all conflicts manually, mark them as resolved with
"git add/rm <conflicted_files>", then run "git rebase --continue".
You can instead skip this commit: run "git rebase --skip".
To abort and get back to the state before "git rebase", run "git rebase --abort".
Recorded preimage for 'git-send-email.perl'
Could not apply 3ff15040e2... send-email: fix regression in sendemail.identity parsing
(detached HEAD *+|REBASE 7/1049)$ ./BUILDS/v2.25.0/bin/git rebase --abort
((v2.24.0))$
So it prints the same error as the 'am' backend (or perhaps that error
comes from the common, backend-independent parts of rebase? I didn't
look), and it continues all the same, but in the end '--abort' is
capable to abort the operation. So we do have yet another reason to
switch the default backend.
