Message ID | pull.1772.git.1723641747309.gitgitgadget@gmail.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | rebase: apply and cleanup autostash when rebase fails to start | expand |
On Wed, Aug 14, 2024 at 01:22:27PM +0000, Phillip Wood via GitGitGadget wrote: > From: Phillip Wood <phillip.wood@dunelm.org.uk> > > If "git rebase" fails to start after stashing the user's uncommitted > changes then it forgets to restore the stashed changes and remove state s/remove/& the/ > directory. To make matters worse running "git rebase --abort" to apply s/worse/&,/ > the stashed changes and cleanup the state directory fails because the s/cleanup/& of/ > diff --git a/builtin/rebase.c b/builtin/rebase.c > index e3a8e74cfc2..ac23c4ddbb0 100644 > --- a/builtin/rebase.c > +++ b/builtin/rebase.c > @@ -526,6 +526,23 @@ static int rebase_write_basic_state(struct rebase_options *opts) > return 0; > } > > +static int cleanup_autostash(struct rebase_options *opts) > +{ > + int ret; > + struct strbuf dir = STRBUF_INIT; > + const char *path = state_dir_path("autostash", opts); > + > + if (!file_exists(path)) > + return 0; > + ret = apply_autostash(path); > + strbuf_addstr(&dir, opts->state_dir); > + if (remove_dir_recursively(&dir, 0)) > + ret = error_errno(_("could not remove '%s'"), opts->state_dir); This seems somewhat dangerous to me though. Should we really delete the autostash directory in case applying the autostash has failed? > @@ -1851,9 +1871,14 @@ run_rebase: > > cleanup: > strbuf_release(&buf); > + strbuf_release(&msg); > strbuf_release(&revisions); > rebase_options_release(&options); > free(squash_onto_name); > free(keep_base_onto_name); > return !!ret; > + > +cleanup_autostash: > + ret |= !!cleanup_autostash(&options); > + goto cleanup; > } It's somewhat curious of a code flow to jump backwards. It might be easier to reason about the flow if we kept track of a variable `autostash_needs_cleanup` that we set such that all jumps can go to the `cleanup` label instead. Patrick
Hi Patrick On 14/08/2024 15:36, Patrick Steinhardt wrote: > On Wed, Aug 14, 2024 at 01:22:27PM +0000, Phillip Wood via GitGitGadget wrote: >> From: Phillip Wood <phillip.wood@dunelm.org.uk> >> >> If "git rebase" fails to start after stashing the user's uncommitted >> changes then it forgets to restore the stashed changes and remove state > > s/remove/& the/ > >> directory. To make matters worse running "git rebase --abort" to apply > > s/worse/&,/ > >> the stashed changes and cleanup the state directory fails because the > > s/cleanup/& of/ Thanks for catching those typos >> diff --git a/builtin/rebase.c b/builtin/rebase.c >> index e3a8e74cfc2..ac23c4ddbb0 100644 >> --- a/builtin/rebase.c >> +++ b/builtin/rebase.c >> @@ -526,6 +526,23 @@ static int rebase_write_basic_state(struct rebase_options *opts) >> return 0; >> } >> >> +static int cleanup_autostash(struct rebase_options *opts) >> +{ >> + int ret; >> + struct strbuf dir = STRBUF_INIT; >> + const char *path = state_dir_path("autostash", opts); >> + >> + if (!file_exists(path)) >> + return 0; >> + ret = apply_autostash(path); >> + strbuf_addstr(&dir, opts->state_dir); >> + if (remove_dir_recursively(&dir, 0)) >> + ret = error_errno(_("could not remove '%s'"), opts->state_dir); > > This seems somewhat dangerous to me though. Should we really delete the > autostash directory in case applying the autostash has failed? Applying the stash should not fail because the rebase has not started and so HEAD, the index and the worktree are unchanged since the stash was created. If it does fail for some reason then apply_autostash() creates a new entry under refs/stash. We definitely do want to remove the directory otherwise we're left with the inconsistent state we're tying to fix. >> @@ -1851,9 +1871,14 @@ run_rebase: >> >> cleanup: >> strbuf_release(&buf); >> + strbuf_release(&msg); >> strbuf_release(&revisions); >> rebase_options_release(&options); >> free(squash_onto_name); >> free(keep_base_onto_name); >> return !!ret; >> + >> +cleanup_autostash: >> + ret |= !!cleanup_autostash(&options); >> + goto cleanup; >> } > > It's somewhat curious of a code flow to jump backwards. It might be > easier to reason about the flow if we kept track of a variable > `autostash_needs_cleanup` that we set such that all jumps can go to the > `cleanup` label instead. It is slightly odd, but in the end I decided it was simpler to say "goto cleanup_autostash" than try and keep track of what needs cleaning up when saying "goto cleanup". There are a couple of similar examples in builtin/stash.c:show_stash() and config.c:git_config_set_multivar_in_file_gently() Thanks for the review Phillip > Patrick >
Phillip Wood <phillip.wood123@gmail.com> writes: > Applying the stash should not fail because the rebase has not started > and so HEAD, the index and the worktree are unchanged since the stash > was created. If it does fail for some reason then apply_autostash() > creates a new entry under refs/stash. We definitely do want to remove > the directory otherwise we're left with the inconsistent state we're > tying to fix. If it is not expected to fail 99% of times, it feels more prudent to abort loudly without making further damage to lose information and ask the user to check what happened in the working tree, rather than blindly removing the clue to understand what went wrong. For example, could the reason why applying the stash failed be because the user forgot that the working tree was being used for rebasing and mucked with its contents from say another terminal? Thanks.
Hi Junio On 14/08/2024 18:27, Junio C Hamano wrote: > Phillip Wood <phillip.wood123@gmail.com> writes: > >> Applying the stash should not fail because the rebase has not started >> and so HEAD, the index and the worktree are unchanged since the stash >> was created. If it does fail for some reason then apply_autostash() >> creates a new entry under refs/stash. We definitely do want to remove >> the directory otherwise we're left with the inconsistent state we're >> tying to fix. > > If it is not expected to fail 99% of times, it feels more prudent to > abort loudly without making further damage to lose information and > ask the user to check what happened in the working tree, rather than > blindly removing the clue to understand what went wrong. For > example, could the reason why applying the stash failed be because > the user forgot that the working tree was being used for rebasing > and mucked with its contents from say another terminal? If the working tree has changed then the stash will still apply but possibly with conflicts - the same thing can happen when the branch has been successfully rebased. If there are conflicts then the stash is saved in refs/stash as well. This code is just doing what we do at the end of a successful rebase so I'm don't really understand what the issue is. Looking at finish_rebase() we don't even check the return value of apply_autostash() when applying the stash at the end of a successful rebase. Best Wishes Phillip > Thanks. >
Phillip Wood <phillip.wood123@gmail.com> writes: > ... This code is just doing what we do at > the end of a successful rebase so I'm don't really understand what the > issue is. Looking at finish_rebase() we don't even check the return > value of apply_autostash() when applying the stash at the end of a > successful rebase. At that point we give control back the user, so if things are left in conflicted or any other "unexpected" funny state, the user kill keep the both halves. As long as the user clearly understands why the working tree is in such a funny state, we should be OK (and I would imagine that we are giving messages like "applying preexisting changes stashed away before rebasing" or something). Thanks.
Phillip Wood <phillip.wood123@gmail.com> writes: > ... This code is just doing what we do at > the end of a successful rebase so I'm don't really understand what the > issue is. Looking at finish_rebase() we don't even check the return > value of apply_autostash() when applying the stash at the end of a > successful rebase. At that point we give control back the user, so if things are left in conflicted or any other "unexpected" funny state, the user kill keep the both halves. As long as the user clearly understands why the working tree is in such a funny state, we should be OK (and I would imagine that we are giving messages like "applying preexisting changes stashed away before rebasing" or something). Thanks.
diff --git a/builtin/rebase.c b/builtin/rebase.c index e3a8e74cfc2..ac23c4ddbb0 100644 --- a/builtin/rebase.c +++ b/builtin/rebase.c @@ -526,6 +526,23 @@ static int rebase_write_basic_state(struct rebase_options *opts) return 0; } +static int cleanup_autostash(struct rebase_options *opts) +{ + int ret; + struct strbuf dir = STRBUF_INIT; + const char *path = state_dir_path("autostash", opts); + + if (!file_exists(path)) + return 0; + ret = apply_autostash(path); + strbuf_addstr(&dir, opts->state_dir); + if (remove_dir_recursively(&dir, 0)) + ret = error_errno(_("could not remove '%s'"), opts->state_dir); + strbuf_release(&dir); + + return ret; +} + static int finish_rebase(struct rebase_options *opts) { struct strbuf dir = STRBUF_INIT; @@ -1726,7 +1743,7 @@ int cmd_rebase(int argc, const char **argv, const char *prefix) if (require_clean_work_tree(the_repository, "rebase", _("Please commit or stash them."), 1, 1)) { ret = -1; - goto cleanup; + goto cleanup_autostash; } /* @@ -1749,7 +1766,7 @@ int cmd_rebase(int argc, const char **argv, const char *prefix) if (options.switch_to) { ret = checkout_up_to_date(&options); if (ret) - goto cleanup; + goto cleanup_autostash; } if (!(options.flags & REBASE_NO_QUIET)) @@ -1775,8 +1792,10 @@ int cmd_rebase(int argc, const char **argv, const char *prefix) /* If a hook exists, give it a chance to interrupt*/ if (!ok_to_skip_pre_rebase && run_hooks_l("pre-rebase", options.upstream_arg, - argc ? argv[0] : NULL, NULL)) - die(_("The pre-rebase hook refused to rebase.")); + argc ? argv[0] : NULL, NULL)) { + ret = error(_("The pre-rebase hook refused to rebase.")); + goto cleanup_autostash; + } if (options.flags & REBASE_DIFFSTAT) { struct diff_options opts; @@ -1821,9 +1840,10 @@ int cmd_rebase(int argc, const char **argv, const char *prefix) RESET_HEAD_RUN_POST_CHECKOUT_HOOK; ropts.head_msg = msg.buf; ropts.default_reflog_action = options.reflog_action; - if (reset_head(the_repository, &ropts)) - die(_("Could not detach HEAD")); - strbuf_release(&msg); + if (reset_head(the_repository, &ropts)) { + ret = error(_("Could not detach HEAD")); + goto cleanup_autostash; + } /* * If the onto is a proper descendant of the tip of the branch, then @@ -1851,9 +1871,14 @@ run_rebase: cleanup: strbuf_release(&buf); + strbuf_release(&msg); strbuf_release(&revisions); rebase_options_release(&options); free(squash_onto_name); free(keep_base_onto_name); return !!ret; + +cleanup_autostash: + ret |= !!cleanup_autostash(&options); + goto cleanup; } diff --git a/t/t3400-rebase.sh b/t/t3400-rebase.sh index ae34bfad607..972ea85febc 100755 --- a/t/t3400-rebase.sh +++ b/t/t3400-rebase.sh @@ -145,7 +145,9 @@ test_expect_success 'Show verbose error when HEAD could not be detached' ' test_when_finished "rm -f B" && test_must_fail git rebase topic 2>output.err >output.out && test_grep "The following untracked working tree files would be overwritten by checkout:" output.err && - test_grep B output.err + test_grep B output.err && + test_must_fail git rebase --quit 2>err && + test_grep "no rebase in progress" err ' test_expect_success 'fail when upstream arg is missing and not on branch' ' @@ -422,7 +424,9 @@ test_expect_success 'refuse to switch to branch checked out elsewhere' ' git checkout main && git worktree add wt && test_must_fail git -C wt rebase main main 2>err && - test_grep "already used by worktree at" err + test_grep "already used by worktree at" err && + test_must_fail git -C wt rebase --quit 2>err && + test_grep "no rebase in progress" err ' test_expect_success 'rebase when inside worktree subdirectory' ' diff --git a/t/t3413-rebase-hook.sh b/t/t3413-rebase-hook.sh index e8456831e8b..426ff098e1d 100755 --- a/t/t3413-rebase-hook.sh +++ b/t/t3413-rebase-hook.sh @@ -110,7 +110,9 @@ test_expect_success 'pre-rebase hook stops rebase (1)' ' git reset --hard side && test_must_fail git rebase main && test "z$(git symbolic-ref HEAD)" = zrefs/heads/test && - test 0 = $(git rev-list HEAD...side | wc -l) + test 0 = $(git rev-list HEAD...side | wc -l) && + test_must_fail git rebase --quit 2>err && + test_grep "no rebase in progress" err ' test_expect_success 'pre-rebase hook stops rebase (2)' ' diff --git a/t/t3420-rebase-autostash.sh b/t/t3420-rebase-autostash.sh index 63e400b89f2..b43046b3b0d 100755 --- a/t/t3420-rebase-autostash.sh +++ b/t/t3420-rebase-autostash.sh @@ -82,6 +82,46 @@ testrebase () { type=$1 dotest=$2 + test_expect_success "rebase$type: restore autostash when pre-rebase hook fails" ' + git checkout -f feature-branch && + test_hook pre-rebase <<-\EOF && + exit 1 + EOF + + echo changed >file0 && + test_must_fail git rebase $type --autostash -f HEAD^ && + test_must_fail git rebase --quit 2>err && + test_grep "no rebase in progress" err && + echo changed >expect && + test_cmp expect file0 + ' + + test_expect_success "rebase$type: restore autostash when checkout onto fails" ' + git checkout -f --detach feature-branch && + echo uncommitted-content >file0 && + echo untracked >file4 && + test_when_finished "rm file4" && + test_must_fail git rebase $type --autostash \ + unrelated-onto-branch && + test_must_fail git rebase --quit 2>err && + test_grep "no rebase in progress" err && + echo uncommitted-content >expect && + test_cmp expect file0 + ' + + test_expect_success "rebase$type: restore autostash when branch checkout fails" ' + git checkout -f unrelated-onto-branch^ && + echo uncommitted-content >file0 && + echo untracked >file4 && + test_when_finished "rm file4" && + test_must_fail git rebase $type --autostash HEAD \ + unrelated-onto-branch && + test_must_fail git rebase --quit 2>err && + test_grep "no rebase in progress" err && + echo uncommitted-content >expect && + test_cmp expect file0 + ' + test_expect_success "rebase$type: dirty worktree, --no-autostash" ' test_config rebase.autostash true && git reset --hard &&