Re: [PATCH 11/17] revert: Save data for continuing after conflict resolution

[Ramkumar Ramachandra <artagnon@gmail.com>]

Hi again,

Jonathan Nieder writes:
> Ramkumar Ramachandra wrote:
>
>> It is imperative to note that these two files alone are not enough to
>> implement "--continue"; we will also need to persist the options that
>> were specified on the command-line, and this is done later in the
>> series.
>
> I think you could remove the phrase "It is imperative to note that"
> and this would say the same thing. :)
>
> By the way, those two files are enough to implement --continue in some
> sense, no?  One implementation of --continue would forget options
> passed on the first run and require the user to specify them again; it
> would just be less useful.  The reminder that "head" and "todo" are
> not the only files in the .git/sequencer dir is useful, though; thanks
> for it.

New commit message:

revert: Save data for continuing after conflict resolution

Ever since v1.7.2-rc1~4^2~7 (revert: allow cherry-picking more than
one commit, 2010-06-02), a single invocation of "git cherry-pick" or
"git revert" can perform picks of several individual commits.  To
implement features like "--continue" to continue the whole operation,
we will need to store some information about the state and the plan at
the beginning.  Introduce a ".git/sequencer/head" file to store this
state, and ".git/sequencer/todo" file to store the plan.  The head
file contains the SHA-1 of the HEAD before the start of the operation,
and the todo file contains an instruction sheet whose format is
inspired by the format of the "rebase -i" instruction sheet.  As a
result, a typical todo file looks like:

pick 8537f0e submodule add: test failure when url is not configured
pick 4d68932 submodule add: allow relative repository path
pick f22a17e submodule add: clean up duplicated code
pick 59a5775 make copy_ref globally available

Since SHA-1 hex is abbreviated using an find_unique_abbrev(), it is
unambiguous.  This does not guarantee that there will be no ambiguity
when more objects are added to the repository.

That these two files alone are not enough to implement a "--continue"
that remembers the command-line options specified; later patches in
the series save them too.

These new files are unrelated to the existing .git/CHERRY_PICK_HEAD,
which will still be useful while committing after a conflict
resolution.

>> --- a/builtin/revert.c
>> +++ b/builtin/revert.c
> [...]
>> +static void format_todo(struct strbuf *buf, struct commit_list *todo_list,
>> +                     struct replay_opts *opts)
>> +{
>> +     struct commit_list *cur = NULL;
>> +     struct commit_message msg = { NULL, NULL, NULL, NULL, NULL };
>> +     const char *sha1_abbrev = NULL;
>> +     const char *action;
>> +
>> +     action = (opts->action == REVERT ? "revert" : "pick");
>> +     for (cur = todo_list; cur; cur = cur->next) {
>> +             sha1_abbrev = find_unique_abbrev(cur->item->object.sha1, DEFAULT_ABBREV);
>> +             if (get_message(cur->item, &msg))
>> +                     die(_("Cannot get commit message for %s"), sha1_abbrev);
>> +             strbuf_addf(buf, "%s %s %s\n", action, sha1_abbrev, msg.subject);
>
> This die() seems odd in the context of a series that starts by
> converting various die() calls to "return error()".

Fixed.  I made format_todo return an int now.

>> +static struct commit *parse_insn_line(char *start, struct replay_opts *opts)
>> +{
>> +     unsigned char commit_sha1[20];
>> +     char sha1_abbrev[40];
>> +     struct commit *commit;
>> +     enum replay_action action;
>> +     int insn_len = 0;
>> +     char *p;
>> +
>> +     p = start;
>> +     if (!(p = strchr(p, ' ')))
>> +             return NULL;
>
> Style: it is much, much clearer to write
>
>        p = strchr(start, ' ');
>        if (!p)
>                return NULL;
>
> In the git codebase, assignments in "if" conditionals are very rare, so
> code like the above sticks out.
>
>> +     insn_len = p - start;
>> +     if (!(p = strchr(p + 1, ' ')))
>> +             return NULL;
>> +     p += 1;
>
> Likewise.  This could say:
>
>        insn_len = p - start;
>        p = strchr(p + 1, ' ');
>        if (!p)
>                return NULL;
>        p++;

Right, got it.  Fixed.

>> +     if (!strncmp(start, "pick", insn_len))
>> +             action = CHERRY_PICK;
>> +     else if (!strncmp(start, "revert", insn_len))
>> +             action = REVERT;
>
> This code means that "p", "pi", and "pic" are accepted as
> abbreviations for "pick".   Maybe a comment would help to clarify
> that.  Are we okay with reserving these names, so "r" in a todo
> file means to "revert" and not to "reword" like it does in a
> rebase -i todo list?

That was the original intent, but I clearly hadn't thought about it
hard enough.  Changed this to accept only full words.

>> +     /*
>> +      * Verify that the action matches up with the one in
>> +      * opts; we don't support arbitrary instructions
>> +      */
>> +     if (action != opts->action)
>> +             return NULL;
>
> If I try "git cherry-pick foo..bar" and then "git revert --continue",
> what error message will I get?

fatal: Malformed instruction sheet: .git/sequencer/todo
Technically speaking, this is correct.  However, this may not be ideal
from an end-user's perspective.  Anyway, this is going to change soon
-- do you think this is worth correcting here and now?

>> +static void read_populate_todo(struct commit_list **todo_list,
>> +                     struct replay_opts *opts)
>> +{
> [...]
>> +     for (p = buf.buf; *p; p = strchr(p, '\n') + 1) {
>> +             if (!(commit = parse_insn_line(p, opts)))
>> +                     goto error;
>> +             new = xmalloc(sizeof(struct commit_list));
>> +             new->item = commit;
>> +             *next = new;
>> +             next = &new->next;
>
> commit_list_append()?

Ugh.  Fixed.

>> +static void create_seq_dir(void)
>> +{
>> +     if (file_exists(git_path(SEQ_DIR))) {
>> +             if (!is_directory(git_path(SEQ_DIR)) && remove_path(git_path(SEQ_DIR)) < 0)
>> +                     die(_("Could not remove %s"), git_path(SEQ_DIR));
>> +     } else if (mkdir(git_path(SEQ_DIR), 0777) < 0)
>> +             die_errno(_("Could not create sequencer directory '%s'."), git_path(SEQ_DIR));
>> +}
>
> As mentioned in [1], a local would make this more readable.  And it's
> not clear to me why one would want to delete a .git/sequencer file
> that the user carefully placed there (I don't think git itself ever
> writes such a thing).

Okay,  removed that hunk.

>> +static void save_head(const char *head)
>> +{
>> +     static struct lock_file head_lock;
>> +     struct strbuf buf = STRBUF_INIT;
>> +     int fd;
>> +
>> +     fd = hold_lock_file_for_update(&head_lock, git_path(SEQ_HEAD_FILE), LOCK_DIE_ON_ERROR);
>> +     strbuf_addf(&buf, "%s\n", head);
>> +     if (write_in_full(fd, buf.buf, buf.len) < 0)
>> +             die_errno(_("Could not write to %s."), git_path(SEQ_HEAD_FILE));
>> +     if (commit_lock_file(&head_lock) < 0)
>> +             die(_("Error wrapping up %s"), git_path(SEQ_HEAD_FILE));
>> +}
>
> A local would be helpful here, too.
>
>> +static void save_todo(struct commit_list *todo_list, struct replay_opts *opts)
>> +{
>> +     static struct lock_file todo_lock;
>> +     struct strbuf buf = STRBUF_INIT;
>> +     int fd;
>> +
>> +     fd = hold_lock_file_for_update(&todo_lock, git_path(SEQ_TODO_FILE), LOCK_DIE_ON_ERROR);
>> +     format_todo(&buf, todo_list, opts);
>> +     if (write_in_full(fd, buf.buf, buf.len) < 0) {
>> +             strbuf_release(&buf);
>> +             die_errno(_("Could not write to %s."), git_path(SEQ_TODO_FILE));
>> +     }
>> +     if (commit_lock_file(&todo_lock) < 0) {
>> +             strbuf_release(&buf);
>> +             die(_("Error wrapping up %s"), git_path(SEQ_TODO_FILE));
>> +     }
>> +     strbuf_release(&buf);
>> +}
>
> Likewise here.  As mentioned before, git_path() clobbers errno, so the
> above is likely to print a wrong error message on some systems (e.g.,
> "fatal: Cannot write to .git/sequencer/todo: Success").

Fixed these.  Thanks.

>> +static int pick_commits(struct replay_opts *opts)
>> +{
>> +     struct commit_list *todo_list = NULL;
>> +     unsigned char sha1[20];
>> +     struct commit_list *cur;
>> +     int res;
>>
>>       setenv(GIT_REFLOG_ACTION, me, 0);
>>       if (opts->allow_ff)
>> @@ -580,14 +764,24 @@ static int pick_commits(struct replay_opts *opts)
>>                               opts->record_origin || opts->edit));
>>       read_and_refresh_cache(me, opts);
>>
>> -     prepare_revs(&revs, opts);
>> +     walk_revs_populate_todo(&todo_list, opts);
>> +     create_seq_dir();
>> +     if (!get_sha1("HEAD", sha1))
>> +             save_head(sha1_to_hex(sha1));
>
> What happens if the .git/sequencer dir already exists (e.g., we were
> in the middle of a multiple-cherry-pick)?  What happens if I try to
> cherry-pick onto an unborn branch, or get_sha1() fails for some other
> reason?

Oops.  I lost this somewhere along the way -- fixed now.  Thanks :)

-- Ram