Re: [PATCH v3 02/11] merge-one-file: rewrite in C

[Junio C Hamano <gitster@pobox.com>]

Alban Gruin <alban.gruin@gmail.com> writes:

> This rewrites `git merge-one-file' from shell to C.  This port is not
> completely straightforward: to save precious cycles by avoiding reading
> and flushing the index repeatedly, write temporary files when an
> operation can be performed in-memory, or allow other function to use the
> rewrite without forking nor worrying about the index,...

So, the in-core index is still used, but when the contents of the in-core
index does not have to be written out disk, we just don't?  Makes sense.

> diff --git a/builtin/merge-one-file.c b/builtin/merge-one-file.c
> new file mode 100644
> index 0000000000..598338ba16
> --- /dev/null
> +++ b/builtin/merge-one-file.c
> @@ -0,0 +1,92 @@
> +/*
> + * Builtin "git merge-one-file"
> + *
> + * Copyright (c) 2020 Alban Gruin
> + *
> + * Based on git-merge-one-file.sh, written by Linus Torvalds.
> + *
> + * This is the git per-file merge utility, called with
> + *
> + *   argv[1] - original file SHA1 (or empty)
> + *   argv[2] - file in branch1 SHA1 (or empty)
> + *   argv[3] - file in branch2 SHA1 (or empty)

Let's modernize this comment while we are at it.

    SHA1 -> "object name" (or "blob object name")

> + *   argv[4] - pathname in repository
> + *   argv[5] - original file mode (or empty)
> + *   argv[6] - file in branch1 mode (or empty)
> + *   argv[7] - file in branch2 mode (or empty)
> + *
> + * Handle some trivial cases. The _really_ trivial cases have been
> + * handled already by git read-tree, but that one doesn't do any merges
> + * that might change the tree layout.
> + */
> +
> +#define USE_THE_INDEX_COMPATIBILITY_MACROS
> +#include "cache.h"
> +#include "builtin.h"
> +#include "lockfile.h"
> +#include "merge-strategies.h"
> +
> +static const char builtin_merge_one_file_usage[] =
> +	"git merge-one-file <orig blob> <our blob> <their blob> <path> "
> +	"<orig mode> <our mode> <their mode>\n\n"
> +	"Blob ids and modes should be empty for missing files.";
> +
> +static int read_mode(const char *name, const char *arg, unsigned int *mode)
> +{
> +	char *last;
> +	int ret = 0;
> +
> +	*mode = strtol(arg, &last, 8);
> +
> +	if (*last)
> +		ret = error(_("invalid '%s' mode: expected nothing, got '%c'"), name, *last);
> +	else if (!(S_ISREG(*mode) || S_ISDIR(*mode) || S_ISLNK(*mode)))
> +		ret = error(_("invalid '%s' mode: %o"), name, *mode);
> +
> +	return ret;
> +}
> +
> +int cmd_merge_one_file(int argc, const char **argv, const char *prefix)
> +{
> +	struct object_id orig_blob, our_blob, their_blob,
> +		*p_orig_blob = NULL, *p_our_blob = NULL, *p_their_blob = NULL;
> +	unsigned int orig_mode = 0, our_mode = 0, their_mode = 0, ret = 0;
> +	struct lock_file lock = LOCK_INIT;
> +
> +	if (argc != 8)
> +		usage(builtin_merge_one_file_usage);
> +
> +	if (read_cache() < 0)
> +		die("invalid index");
> +
> +	hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
> +
> +	if (!get_oid(argv[1], &orig_blob)) {
> +		p_orig_blob = &orig_blob;
> +		ret = read_mode("orig", argv[5], &orig_mode);
> +	}

argv[1] is defined as "either the object name of the blob in the
common ancestor, or an empty string".  So you need to distinguish
three cases here, but you are only catching two.

 - argv[1] is an empty string; p_orig_blob can legitimately be left
   NULL.

 - argv[1] is a valid blob object name.  orig_blob should be
   populated and p_orig_blob should point at it.

 - argv[1] is garbage, names a non-blob object, or there is no such
   object with that name.  Don't we want to catch it as a mistake?

Also, when argv[1] is an empty string, argv[5] must also be an empty
string, or we got a wrong input---don't we want to catch it as a
mistake?

The third case needs a bit of thought.  For example, if $1 and $2
are the same and points at a non-existent object, we know we won't
care because we only care about $3.  In a lazily-cloned repository,
that may matter---we would not want to fail even if we not have blob
$1 and $2, as long as they are reasonably spelled a full hexadecimal
object name.  But we would want to fail if blob object named by $3
is missing.

One way to achieve semantics closer to the above than the posted
patch may be to tighten the parsing.  Instead of using "anything
goes" get_oid(), use get_oid_hex(), perhaps.

> +	if (!get_oid(argv[2], &our_blob)) {
> +		p_our_blob = &our_blob;
> +		ret = read_mode("our", argv[6], &our_mode);
> +	}
> +
> +	if (!get_oid(argv[3], &their_blob)) {
> +		p_their_blob = &their_blob;
> +		ret = read_mode("their", argv[7], &their_mode);
> +	}
> +
> +	if (ret)
> +		return ret;
> +
> +	ret = merge_strategies_one_file(the_repository,
> +					p_orig_blob, p_our_blob, p_their_blob, argv[4],
> +					orig_mode, our_mode, their_mode);

That's a funny function name.  It's not like the function will be
taught different strategy to handle the three-way merge, no?  It
probably makes sense to name it after what it does, which is "three
way merge".

> +	if (ret) {
> +		rollback_lock_file(&lock);
> +		return !!ret;
> +	}
> +
> +	return write_locked_index(&the_index, &lock, COMMIT_LOCK);
> +}

> diff --git a/merge-strategies.c b/merge-strategies.c
> new file mode 100644
> index 0000000000..bbe6f48698
> --- /dev/null
> +++ b/merge-strategies.c
> @@ -0,0 +1,214 @@
> +#include "cache.h"
> +#include "dir.h"
> +#include "ll-merge.h"
> +#include "merge-strategies.h"
> +#include "xdiff-interface.h"
> +

> +static int add_to_index_cacheinfo(struct index_state *istate,
> +				  unsigned int mode,
> +				  const struct object_id *oid, const char *path)
> +{
> +	struct cache_entry *ce;
> +	int len, option;
> +
> +	if (!verify_path(path, mode))
> +		return error(_("Invalid path '%s'"), path);
> +
> +	len = strlen(path);
> +	ce = make_empty_cache_entry(istate, len);
> +
> +	oidcpy(&ce->oid, oid);
> +	memcpy(ce->name, path, len);
> +	ce->ce_flags = create_ce_flags(0);
> +	ce->ce_namelen = len;
> +	ce->ce_mode = create_ce_mode(mode);
> +	if (assume_unchanged)
> +		ce->ce_flags |= CE_VALID;
> +	option = ADD_CACHE_OK_TO_ADD | ADD_CACHE_OK_TO_REPLACE;
> +	if (add_index_entry(istate, ce, option))
> +		return error(_("%s: cannot add to the index"), path);
> +
> +	return 0;
> +}

The above correctly does 'git update-index --add --cacheinfo "$6"
"$2" "$4"' but don't copy-and-paste existing code to do so.  Add one
preliminary patch before everything else in the series to massage
and extract add_cacheinfo() function out of builtin/update-index.c,
move it to somewhere common like read-cache.c and so that we can
call it from here.

> +static int checkout_from_index(struct index_state *istate, const char *path)
> +{
> +	struct checkout state = CHECKOUT_INIT;
> +	struct cache_entry *ce;
> +
> +	state.istate = istate;
> +	state.force = 1;
> +	state.base_dir = "";
> +	state.base_dir_len = 0;
> +
> +	ce = index_file_exists(istate, path, strlen(path), 0);

This call is unfortunate for the reasons I mention later.

But if you must have this call, then you need to sanity check what
you get from index_file_exists().  ce must be a merged cache entry,
so

	if (!ce || ce_stage(ce))
		BUG(...);

> +	if (checkout_entry(ce, &state, NULL, NULL) < 0)
> +		return error(_("%s: cannot checkout file"), path);
> +	return 0;
> +}
> +
> +static int merge_one_file_deleted(struct index_state *istate,
> +				  const struct object_id *orig_blob,
> +				  const struct object_id *our_blob,
> +				  const struct object_id *their_blob, const char *path,
> +				  unsigned int orig_mode, unsigned int our_mode, unsigned int their_mode)
> +{
> +	if ((our_blob && orig_mode != our_mode) ||
> +	    (their_blob && orig_mode != their_mode))
> +		return error(_("File %s deleted on one branch but had its "
> +			       "permissions changed on the other."), path);
> +
> +	if (our_blob) {
> +		printf(_("Removing %s\n"), path);
> +
> +		if (file_exists(path))
> +			remove_path(path);
> +	}
> +
> +	if (remove_file_from_index(istate, path))
> +		return error("%s: cannot remove from the index", path);
> +	return 0;

If the side that did not remove changed the mode, we don't silently
remove but fail and give a chance to inspect the situation to the
end user.  If we had the blob and it is removed by them, we give a
message and only in that case we remove the file from the working
tree, together with any leading directory that has become empty.

And after that we make sure that the path is no longer in the
index.  The function removes entries for the path at all the stages,
which is exactly what we want.

OK.

> +}
> +
> +static int do_merge_one_file(struct index_state *istate,
> +			     const struct object_id *orig_blob,
> +			     const struct object_id *our_blob,
> +			     const struct object_id *their_blob, const char *path,
> +			     unsigned int orig_mode, unsigned int our_mode, unsigned int their_mode)
> +{
> +	int ret, i, dest;
> +	ssize_t written;
> +	mmbuffer_t result = {NULL, 0};
> +	mmfile_t mmfs[3];
> +	struct ll_merge_options merge_opts = {0};
> +	struct cache_entry *ce;
> +
> +	if (our_mode == S_IFLNK || their_mode == S_IFLNK)
> +		return error(_("%s: Not merging symbolic link changes."), path);
> +	else if (our_mode == S_IFGITLINK || their_mode == S_IFGITLINK)
> +		return error(_("%s: Not merging conflicting submodule changes."), path);
> +
> +	read_mmblob(mmfs + 1, our_blob);
> +	read_mmblob(mmfs + 2, their_blob);
> +
> +	if (orig_blob) {
> +		printf(_("Auto-merging %s\n"), path);
> +		read_mmblob(mmfs + 0, orig_blob);
> +	} else {
> +		printf(_("Added %s in both, but differently.\n"), path);
> +		read_mmblob(mmfs + 0, &null_oid);
> +	}
> +
> +	merge_opts.xdl_opts = XDL_MERGE_ZEALOUS_ALNUM;
> +	ret = ll_merge(&result, path,
> +		       mmfs + 0, "orig",
> +		       mmfs + 1, "our",
> +		       mmfs + 2, "their",
> +		       istate, &merge_opts);

Is it correct to call into ll_merge() here?  The original used to
call "git merge-file" which called into xdl_merge().  Calling into
ll_merge() means the path is used to look up the attributes and use
the custom merge driver, which I am not offhand sure is what we want
to see at this low level (and if it turns out to be a good idea, we
definitely should explain the change of semantics in the proposed
log message for this commit).

> +	for (i = 0; i < 3; i++)
> +		free(mmfs[i].ptr);
> +
> +	if (ret < 0) {
> +		free(result.ptr);
> +		return error(_("Failed to execute internal merge"));
> +	}
> +
> +	/*
> +	 * Create the working tree file, using "our tree" version from
> +	 * the index, and then store the result of the merge.
> +	 */

The above is copied from the original, to explain what it did after
the comment, but it does not seem to match what the new code does.

> +	ce = index_file_exists(istate, path, strlen(path), 0);
> +	if (!ce)
> +		BUG("file is not present in the cache?");
> +
> +	unlink(path);
> +	if ((dest = open(path, O_WRONLY | O_CREAT, ce->ce_mode)) < 0) {
> +		free(result.ptr);
> +		return error_errno(_("failed to open file '%s'"), path);
> +	}
> +
> +	written = write_in_full(dest, result.ptr, result.size);
> +	close(dest);
> +
> +	free(result.ptr);
> +
> +	if (written < 0)
> +		return error_errno(_("failed to write to '%s'"), path);
> +

This open(..., ce->ce_mode) call is way insufficient.

The comment we have above this part of the code talks about the
difficulty of doing this correctly in scripted version.  Creating a
file by 'git checkout-index -f --stage=2 -- "$4"' and reusing it to
store the merged contents was the cleanest and easiest way without
having direct access to adjust_shared_perm() to create a working
tree file with the correct permission bits.

We are writing in C, so we should be able to do much better than the
scripted version, as we can later call adjust_shared_perm().

> +	if (ret != 0 || !orig_blob)
> +		ret = error(_("content conflict in %s"), path);
> +	if (our_mode != their_mode)
> +		return error(_("permission conflict: %o->%o,%o in %s"),
> +			     orig_mode, our_mode, their_mode, path);
> +	if (ret)
> +		return -1;
> +
> +	return add_file_to_index(istate, path, 0);
> +}
> +
> +int merge_strategies_one_file(struct repository *r,
> +			      const struct object_id *orig_blob,
> +			      const struct object_id *our_blob,
> +			      const struct object_id *their_blob, const char *path,
> +			      unsigned int orig_mode, unsigned int our_mode,
> +			      unsigned int their_mode)
> +{

In a long if/else if/else if/.../else cascade, enclose all bodies in
braces, if any one of them has a multi-statement body, to avoid
being distracting.

> +	if (orig_blob &&
> +	    ((!their_blob && our_blob && oideq(orig_blob, our_blob)) ||
> +	     (!our_blob && their_blob && oideq(orig_blob, their_blob))))
> +		/* Deleted in both or deleted in one and unchanged in the other. */
> +		return merge_one_file_deleted(r->index,
> +					      orig_blob, our_blob, their_blob, path,
> +					      orig_mode, our_mode, their_mode);

OK, we've already reviewed that function.

> +	else if (!orig_blob && our_blob && !their_blob) {
> +		/*
> +		 * Added in one.  The other side did not add and we
> +		 * added so there is nothing to be done, except making
> +		 * the path merged.
> +		 */
> +		return add_to_index_cacheinfo(r->index, our_mode, our_blob, path);

OK, we've already reviewed that function.

> +	} else if (!orig_blob && !our_blob && their_blob) {
> +		printf(_("Adding %s\n"), path);
> +
> +		if (file_exists(path))
> +			return error(_("untracked %s is overwritten by the merge."), path);
> +
> +		if (add_to_index_cacheinfo(r->index, their_mode, their_blob, path))
> +			return -1;
> +		return checkout_from_index(r->index, path);

You did "add_to_index_cacheinfo()", so you MUST know which ce is to
be checked out.

Consider if it is worth to teach add_to_index_cacheinfo() to give
you ce back and pass it to checkout_from_index(); that way, you do
not have to call index_file_exists() based on path in the function.

> +	} else if (!orig_blob && our_blob && their_blob &&
> +		   oideq(our_blob, their_blob)) {
> +		/* Added in both, identically (check for same permissions). */
> +		if (our_mode != their_mode)
> +			return error(_("File %s added identically in both branches, "
> +				       "but permissions conflict %o->%o."),
> +				     path, our_mode, their_mode);
> +
> +		printf(_("Adding %s\n"), path);
> +
> +		if (add_to_index_cacheinfo(r->index, our_mode, our_blob, path))
> +			return -1;
> +		return checkout_from_index(r->index, path);

Ditto.

> +	} else if (our_blob && their_blob)
> +		/* Modified in both, but differently. */
> +		return do_merge_one_file(r->index,
> +					 orig_blob, our_blob, their_blob, path,
> +					 orig_mode, our_mode, their_mode);
> +	else {
> +		char orig_hex[GIT_MAX_HEXSZ] = {0}, our_hex[GIT_MAX_HEXSZ] = {0},
> +			their_hex[GIT_MAX_HEXSZ] = {0};
> +
> +		if (orig_blob)
> +			oid_to_hex_r(orig_hex, orig_blob);
> +		if (our_blob)
> +			oid_to_hex_r(our_hex, our_blob);
> +		if (their_blob)
> +			oid_to_hex_r(their_hex, their_blob);
> +
> +		return error(_("%s: Not handling case %s -> %s -> %s"),
> +			     path, orig_hex, our_hex, their_hex);
> +	}
> +
> +	return 0;
> +}

I can see that this does go in the right direction.  With a bit more
attention to details it would soon be production-ready quality.

Thanks.