Re: [PATCH v2 6/6] merge: do not exit restore_state() prematurely

From: Elijah Newren <newren@gmail.com>
Date: Jul 21, 2022, 03:27 UTC
Message-ID: <CABPp-BEywagXQtVQVxZgTCaH+TxrA9VBrMr2Lyv_3uADoqYY+Q@mail.gmail.com>
In-Reply-To: <xmqqzgh466ia.fsf@gitster.g>

On Tue, Jul 19, 2022 at 4:13 PM Junio C Hamano <gitster@pobox.com> wrote:

Show 52 quoted lines

>
> "Elijah Newren via GitGitGadget" <gitgitgadget@gmail.com> writes:
>
> > Fix the main problem by making sure that restore_state() only skips the
> > stash application if the stash is null rather than skipping the whole
> > function.
>
> OK.
>
>
> > However, there is a secondary problem -- since merge.c forks
> > subprocesses to do the cleanup, the in-memory index is left out-of-sync.
> > While there was a refresh_cache(REFRESH_QUIET) call that attempted to
> > correct that, that function would not handle cases where the previous
> > merge strategy added conflicted entries.  We need to drop the index and
> > re-read it to handle such cases.
>
> Absolutely right.
>
> > diff --git a/builtin/merge.c b/builtin/merge.c
> > index aaee8f6a553..a21dece1b55 100644
> > --- a/builtin/merge.c
> > +++ b/builtin/merge.c
> > @@ -385,11 +385,11 @@ static void restore_state(const struct object_id *head,
> >  {
> >       const char *args[] = { "stash", "apply", "--index", NULL, NULL };
> >
> > -     if (is_null_oid(stash))
> > -             return;
> > -
> >       reset_hard(head, 1);
> >
> > +     if (is_null_oid(stash))
> > +             goto refresh_cache;
> > +
> >       args[3] = oid_to_hex(stash);
> >
> >       /*
> > @@ -398,7 +398,9 @@ static void restore_state(const struct object_id *head,
> >        */
> >       run_command_v_opt(args, RUN_GIT_CMD);
> >
> > -     refresh_cache(REFRESH_QUIET);
> > +refresh_cache:
> > +     if (discard_cache() < 0 || read_cache() < 0)
> > +             die(_("could not read index"));
>
> Don't we need refresh_cache() after re-reading the on-disk index, or
> do we have nothing to do further after restore_state() returns and
> the stat-info being stale does not matter?  Given that [3/6] exists,
> I suspect that we do want to make sure the in-core index is refreshed
> before we go ahead and run the next merge, no?

I don't think so; the situation for [3/6] is different.  The basic
timeline is as follows:
    1. <User does lots of stuff over weeks and months>
    2. User decides to merge one or more branches
    3. merge does save_state() [i.e. "git stash create"]
    4. The first strategy fails
    5. We restore the state before trying the next strategy

The current code is dealing with step 5. Patch [3/6] was to prevent failures in step 3 from users creating stat-dirty files in step 1. Once step 3 runs, the only way to become stat-dirty again is if the user simultaneously messes with their checkout while the "git merge" command is running. Attempting to preventatively handle users modifying the working tree simultaneously with concurrent git commands like `git merge` seems like a losing proposition to me; it'd be a huge can of worms and have a million holes. I don't think that's worth it.

Show 9 quoted lines

> >  }
> >
> >  /* This is called when no merge was necessary. */
>
> > diff --git a/t/t7607-merge-state.sh b/t/t7607-merge-state.sh
> > new file mode 100755
>
> As long we are adding a brand-new script for new tests, probably we
> should add tests for other steps (like [4/6]) here, perhaps?

Yes.

Show 18 quoted lines

>
> > index 00000000000..655478cd0b3
> > --- /dev/null
> > +++ b/t/t7607-merge-state.sh
> > @@ -0,0 +1,25 @@
> > +#!/bin/sh
> > +
> > +test_description="Test that merge state is as expected after failed merge"
> > +
> > +GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME=main
> > +export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
> > +. ./test-lib.sh
> > +
> > +test_expect_success 'set up custom strategy' '
> > +     test_commit --no-tag "Initial" base base &&
> > +git show-ref &&
>
> Is this part of the test, or a leftover debugging aid?

Looks like part of a leftover debugging aid; sorry about that.  Will clean up.

Show 16 quoted lines

> > +
> > +     for b in branch1 branch2 branch3
> > +     do
> > +             git checkout -b $b main &&
> > +             test_commit --no-tag "Change on $b" base $b
> > +     done &&
> > +
> > +     git checkout branch1 &&
> > +     test_must_fail git merge branch2 branch3 &&
> > +     git diff --exit-code --name-status &&
> > +     test_path_is_missing .git/MERGE_HEAD
> > +'
>
> Hmph, I am not sure if the new behaviour is not too pessimistic.
> When octopus fails after successfully merging branch2 and then
> failing the merge of branch3 (i.e. the last one) due to conflict,

That's not what's happening here.  It is not failing due to conflict,
octopus is reporting that it cannot even leave things in a conflicted
state for the user, and is actually incapable of handling this
particular type of merge.  Part of the output seen when attempting
this merge includes:
    fatal: merge program failed
    Should not be doing an octopus.

See previous discussion at https://lore.kernel.org/git/xmqq35hdd205.fsf@gitster.g/. To make this clearer, perhaps I should use "test_expect_code 2" instead of test_must_fail, and also grep the output/error for the above messages.

> I think octpus users are used to be able to resolve it manually
> and make a commit.  Are we making it impossible by doing the
> reset-restore dance here?

No, we are not changing what octopus can handle here. The code above is not triggered when octopus returns that it ran into conflicts for the user to resolve. It is only triggered when octopus says it cannot handle the merge in question. See your commit 98efc8f3d8 ("octopus: allow manual resolve on the last round.", 2006-01-13), and note that the the "exit 2" code path is the one that we are hitting. I'll add some comments to the testcase and rewrite the commit message to try to make this clearer.