Fix stash apply in sparse checkouts (and a submodule test)

Heavier usage of sparse-checkouts at $DAYJOB is commencing. And an issue with git stash apply was found.

git stash's implementation as a pipeline of forked commands presents some problems, especially when implemented atop of three commands that all behave differently in the presence of sparse checkouts. Add a testcase demonstrating some issues with git stash apply in a repository with a different set of sparse-checkout patterns at apply vs create time, clean up the relevant section of git stash code, and incidentally fix a submodule testcase unrelated to sparse checkouts. Provide some detailed commit messages explaining the issues along the way.

NOTE: I found a couple minor issues with other commands in sparse checkouts while debugging this issue, but I don't yet have fixes for them and I can submit them separately.

Elijah Newren (3):
t7012: add a testcase demonstrating stash apply bugs in sparse
checkouts
stash: remove unnecessary process forking
stash: fix stash application in sparse-checkouts

 builtin/stash.c                  | 130 +++++++++++++++++++++----------
 t/lib-submodule-update.sh        |  16 ++--
 t/t7012-skip-worktree-writing.sh |  88 +++++++++++++++++++++
 3 files changed, 184 insertions(+), 50 deletions(-)

base-commit: faefdd61ec7c7f6f3c8c9907891465ac9a2a1475
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-919%2Fnewren%2Fsparse-checkout-fixups-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-919/newren/sparse-checkout-fixups-v1
Pull-Request: https://github.com/git/git/pull/919

gitgitgadget

From: Elijah Newren <[email protected]>

Applying stashes in sparse-checkouts, particularly when the patterns used to define the present paths have changed between when the stash was created and when it is applied, has a number of bugs. The primary problem is perhaps that stashes can be only partially applied (sometimes without any warning or error being displayed and with 0 exit status). In addition, there are cases where partial stash application comes with non-translated error messages and an early abort. The first is when there are files present despite the SKIP_WORKTREE bit being set, in which case the error message shown is:

error: Entry 'PATHNAME' not uptodate. Cannot merge.

The other situation is when a stash contains new files to add to the working tree; in this case, the code aborts early (but after at least partial stash application) with the following error message being shown:

error: NEWFILE: does not exist and --remove not passed
fatal: Unable to process path NEWFILE

Add a test that can trigger all three of these problems that carefully checks that the working copy and SKIP_WORKTREE bits are as expected after the stash application...but currently marked as expected to fail.

---
 t/t7012-skip-worktree-writing.sh | 88 ++++++++++++++++++++++++++++++++
 1 file changed, 88 insertions(+)

index 7476781979..a184ee97fb 100755
@@ -149,6 +149,94 @@ test_expect_success '--ignore-skip-worktree-entries leaves worktree alone' '
 		--diff-filter=D -- keep-me.t
 '
 
+test_expect_failure 'stash restore in sparse checkout' '
+	test_create_repo stash-restore &&
+	(
+		cd stash-restore &&
+
+		mkdir subdir &&
+		echo A >subdir/A &&
+		echo untouched >untouched &&
+		echo removeme >removeme &&
+		echo modified >modified &&
+		git add . &&
+		git commit -m Initial &&
+
+		echo AA >>subdir/A &&
+		echo addme >addme &&
+		echo tweaked >>modified &&
+		rm removeme &&
+		git add addme &&
+
+		git stash push &&
+
+		git sparse-checkout set subdir &&
+
+		# Ensure after sparse-checkout we only have expected files
+		cat >expect <<-EOF &&
+		S modified
+		S removeme
+		H subdir/A
+		S untouched
+		EOF
+		git ls-files -t >actual &&
+		test_cmp expect actual &&
+
+		test_path_is_missing addme &&
+		test_path_is_missing modified &&
+		test_path_is_missing removeme &&
+		test_path_is_file    subdir/A &&
+		test_path_is_missing untouched &&
+
+		# Put a file in the working directory in the way
+		echo in the way >modified &&
+		git stash apply &&
+
+		# Ensure stash vivifies modifies paths...
+		cat >expect <<-EOF &&
+		H addme
+		H modified
+		H removeme
+		H subdir/A
+		S untouched
+		EOF
+		git ls-files -t >actual &&
+		test_cmp expect actual &&
+
+		# ...and that the paths show up in status as changed...
+		cat >expect <<-EOF &&
+		A  addme
+		 M modified
+		 D removeme
+		 M subdir/A
+		?? actual
+		?? expect
+		?? modified.stash.XXXXXX
+		EOF
+		git status --porcelain | \
+			sed -e s/stash......./stash.XXXXXX/ >actual &&
+		test_cmp expect actual &&
+
+		# ...and that working directory reflects the files correctly
+		test_path_is_file    addme &&
+		test_path_is_file    modified &&
+		test_path_is_missing removeme &&
+		test_path_is_file    subdir/A &&
+		test_path_is_missing untouched &&
+
+		# ...including that we have the expected "modified" file...
+		cat >expect <<-EOF &&
+		modified
+		tweaked
+		EOF
+		test_cmp expect modified &&
+
+		# ...and that the other "modified" file is still present...
+		echo in the way >expect &&
+		test_cmp expect modified.stash.*
+	)
+'
+
 #TODO test_expect_failure 'git-apply adds file' false
 #TODO test_expect_failure 'git-apply updates file' false
 #TODO test_expect_failure 'git-apply removes file' false

gitgitgadget

From: Elijah Newren <[email protected]>

When stash was converted from shell to a builtin, it merely transliterated the forking of various git commands from shell to a C program that would fork the same commands. Some of those were converted over to actual library calls, but much of the pipeline-of-commands design still remains. Fix some of this by replacing the portion corresponding to

    git diff-index --cached --name-only --diff-filter=A $CTREE >"$a"
    git read-tree --reset $CTREE
    git update-index --add --stdin <"$a"
    rm -f "$a"

into a library function that does the same thing. (The read-tree --reset was already partially converted over to a library call, but as an independent piece.) Note here that this came after a merge operation was performed. The merge machinery always staged anything that cleanly merges, and the above code only runs if there were no conflicts. Its purpose is to make it so that when there are no conflicts, all the changes from the stash are unstaged. However, that causes brand new files from the stash to become untracked, so the code above first saves those files off and then re-adds them afterwards.

We replace the whole series of commands with a simple function that will unstage files that are not newly added. This doesn't fix any bugs in the usage of these commands, it simply matches the existing behavior but making it an actual builtin that we can then operate on as a whole. A subsequent commit will take advantage of this to fix issues with these commands in sparse-checkouts.

This conversion incidentally fixes t3906.1, because the separate update-index process would die with the following error messages:
error: uninitialized_sub: is a directory - add files inside instead
fatal: Unable to process path uninitialized_sub
The unstaging of the directory as a submodule meant it was no longer tracked, and thus as an uninitialized directory it could not be added back using git update-index --add, thus resulting in this error and early abort. Most of the submodule tests in 3906 continue to fail after this change, this change was just enough to push the first of those tests to success.

---
 builtin/stash.c           | 96 ++++++++++++++++++++++-----------------
 t/lib-submodule-update.sh | 16 +++----
 2 files changed, 62 insertions(+), 50 deletions(-)

index 24ddb1bffa..8117d7647d 100644
@@ -325,33 +325,64 @@ static void add_diff_to_buf(struct diff_queue_struct *q,
 	}
 }
 
-static int get_newly_staged(struct strbuf *out, struct object_id *c_tree)
+static void unstage_changes_unless_new(struct object_id *cache_tree)
 {
-	struct child_process cp = CHILD_PROCESS_INIT;
-	const char *c_tree_hex = oid_to_hex(c_tree);
-
 	/*
-	 * diff-index is very similar to diff-tree above, and should be
-	 * converted together with update_index.
+	 * When we enter this function, there has been a clean merge of
+	 * relevant trees, and the merge logic always stages whatever merges
+	 * cleanly.  We want to unstage those changes, unless it corresponds
+	 * to a file that didn't exist as of cache_tree.
 	 */
-	cp.git_cmd = 1;
-	strvec_pushl(&cp.args, "diff-index", "--cached", "--name-only",
-		     "--diff-filter=A", NULL);
-	strvec_push(&cp.args, c_tree_hex);
-	return pipe_command(&cp, NULL, 0, out, 0, NULL, 0);
-}
 
-static int update_index(struct strbuf *out)
-{
-	struct child_process cp = CHILD_PROCESS_INIT;
+	struct diff_options diff_opts;
+	struct lock_file lock = LOCK_INIT;
+	int i;
 
-	/*
-	 * Update-index is very complicated and may need to have a public
-	 * function exposed in order to remove this forking.
-	 */
-	cp.git_cmd = 1;
-	strvec_pushl(&cp.args, "update-index", "--add", "--stdin", NULL);
-	return pipe_command(&cp, out->buf, out->len, NULL, 0, NULL, 0);
+	diff_setup(&diff_opts);
+	diff_opts.flags.recursive = 1;
+	diff_opts.detect_rename = 0;
+	diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
+	diff_setup_done(&diff_opts);
+
+	do_diff_cache(cache_tree, &diff_opts);
+	diffcore_std(&diff_opts);
+
+	for (i = 0; i < diff_queued_diff.nr; i++) {
+		struct diff_filepair *p;
+		struct cache_entry *ce;
+		int pos;
+
+		p = diff_queued_diff.queue[i];
+		pos = index_name_pos(&the_index, p->two->path,
+				     strlen(p->two->path));
+		if (pos < 0) {
+			assert(p->one->oid_valid);
+			assert(!p->two->oid_valid);
+			ce = make_cache_entry(&the_index,
+					      p->one->mode,
+					      &p->one->oid,
+					      p->one->path,
+					      0, 0);
+			add_index_entry(&the_index, ce, ADD_CACHE_OK_TO_ADD);
+			continue;
+		}
+		ce = active_cache[pos];
+		if (p->one->oid_valid) {
+			ce = make_cache_entry(&the_index,
+					      p->one->mode,
+					      &p->one->oid,
+					      p->one->path,
+					      0, 0);
+			add_index_entry(&the_index, ce,
+					ADD_CACHE_OK_TO_REPLACE);
+		}
+	}
+	diff_flush(&diff_opts);
+
+	repo_hold_locked_index(the_repository, &lock, LOCK_DIE_ON_ERROR);
+	if (write_locked_index(&the_index, &lock,
+			       COMMIT_LOCK | SKIP_IF_UNCHANGED))
+		die(_("Unable to write index."));
 }
 
 static int restore_untracked(struct object_id *u_tree)

@@ -467,26 +498,7 @@ static int do_apply_stash(const char *prefix, struct stash_info *info,
 		if (reset_tree(&index_tree, 0, 0))
 			return -1;
 	} else {
-		struct strbuf out = STRBUF_INIT;
-
-		if (get_newly_staged(&out, &c_tree)) {
-			strbuf_release(&out);
-			return -1;
-		}
-
-		if (reset_tree(&c_tree, 0, 1)) {
-			strbuf_release(&out);
-			return -1;
-		}
-
-		ret = update_index(&out);
-		strbuf_release(&out);
-		if (ret)
-			return -1;
-
-		/* read back the result of update_index() back from the disk */
-		discard_cache();
-		read_cache();
+		unstage_changes_unless_new(&c_tree);
 	}
 
 	if (!quiet) {

index bd3fa3c6da..4b714e9308 100644
@@ -316,14 +316,7 @@ test_submodule_switch_common () {
 	command="$1"
 	######################### Appearing submodule #########################
 	# Switching to a commit letting a submodule appear creates empty dir ...
-	if test "$KNOWN_FAILURE_STASH_DOES_IGNORE_SUBMODULE_CHANGES" = 1
-	then
-		# Restoring stash fails to restore submodule index entry
-		RESULT="failure"
-	else
-		RESULT="success"
-	fi
-	test_expect_$RESULT "$command: added submodule creates empty directory" '
+	test_expect_success "$command: added submodule creates empty directory" '
 		prolog &&
 		reset_work_tree_to no_submodule &&
 		(

@@ -337,6 +330,13 @@ test_submodule_switch_common () {
 		)
 	'
 	# ... and doesn't care if it already exists.
+	if test "$KNOWN_FAILURE_STASH_DOES_IGNORE_SUBMODULE_CHANGES" = 1
+	then
+		# Restoring stash fails to restore submodule index entry
+		RESULT="failure"
+	else
+		RESULT="success"
+	fi
 	test_expect_$RESULT "$command: added submodule leaves existing empty directory alone" '
 		prolog &&
 		reset_work_tree_to no_submodule &&

gitgitgadget

From: Elijah Newren <[email protected]>

sparse-checkouts are built on the patterns in the $GIT_DIR/info/sparse-checkout file, where commands have modified behavior for paths that do not match those patterns. The differences in behavior, as far as the bugs concerned here, fall into three different categories (with git subcommands that fall into each category listed):

* commands that only look at files matching the patterns:
* status
* diff
* clean
* update-index
* commands that remove files from the working tree that do not match the patterns, and restore files that do match them:
* read-tree
* switch
* checkout
* reset (--hard)
* commands that omit writing files to the working tree that do not match the patterns, unless those files are not clean:
* merge
* rebase
* cherry-pick
* revert

There are some caveats above, e.g. a plain git diff ignores files outside the sparsity patterns but will show diffs for paths outside the sparsity patterns when revision arguments are passed. (Technically, diff is treating the sparse paths as matching HEAD.) So, there is some internal inconsistency among these commands. There are also additional commands that should behave differently in the face of sparse-checkouts, as the sparse-checkout documentation alludes to, but the above is sufficient for me to explain how git stash is affected.

What is relevant here is that logically 'stash' should behave like a merge; it three-way merges the changes the user had in progress at stash creation time, the HEAD at the time the stash was created, and the current HEAD, in order to get the stashed changes applied to the current branch. However, this simplistic view doesn't quite work in practice, because stash tweaks it a bit due to two factors: (1) flags like --keep-index and --include-untracked (why we used two different verbs, 'keep' and 'include', is a rant for another day) modify what should be staged at the end and include more things that should be quasi-merged,
(2) stash generally wants changes to NOT be staged. It only provides exceptions when (a) some of the changes had conflicts and thus we want to use staged to denote the clean merges and higher order stages to mark the conflicts, or (b) if there is a brand new file we don't want it to become untracked.

stash has traditionally gotten this special behavior by first doing a merge, and then when it's clean, applying a pipeline of commands to modify the result. This series of commands for unstaging-non-newly-added-files came from the following commands:

    git diff-index --cached --name-only --diff-filter=A $CTREE >"$a"
    git read-tree --reset $CTREE
    git update-index --add --stdin <"$a"
    rm -f "$a"

You might that between the merge that proceeded these commands and these different commands here, that we have commands from each of the different types of special sparsity handling listed at the beginning of this message, and in fact this precisely led to the following buggy behaviors:

(1) If a path merged cleanly and it didn't match the sparsity patterns, the merge backend would know to avoid writing it to the working tree and keep the SKIP_WORKTREE bit, simply only updating it in the index. Unfortunately, the subsequent commands would essentially undo the changes in the index and thus simply toss the changes altogether since there was nothing left in the working tree. This means the stash is only partially applied.

(2) If a path existed in the worktree before git stash apply despite having the SKIP_WORKTREE bit set, then the git read-tree --reset would print an error message of the form
error: Entry 'modified' not uptodate. Cannot merge.
and cause stash to abort early.

(3) If there was a brand new file added by the stash, then the diff-index command would save that pathname to the temporary file, the read-tree --reset would remove it from the index, and the update-index command would barf due to no such file being present in the working copy; it would print a message of the form:
error: NEWFILE: does not exist and --remove not passed
fatal: Unable to process path NEWFILE
and then cause stash to abort early.

Basically, the whole idea of unstage-unless-brand-new requires special care when you are dealing with a sparse-checkout. Fix these problems by applying the following simple rule:

When we unstage files, if they have the SKIP_WORKTREE bit set, clear that bit and write the file out to the working directory.

(*) If there's already a file present in the way, rename it first.

This fixes all three problems in t7012.13 and allows us to mark it as passing.

---
 builtin/stash.c                  | 36 +++++++++++++++++++++++++++++++-
 t/t7012-skip-worktree-writing.sh |  2 +-
 2 files changed, 36 insertions(+), 2 deletions(-)

index 8117d7647d..0f7e78d315 100644
@@ -331,13 +331,23 @@ static void unstage_changes_unless_new(struct object_id *cache_tree)
 	 * When we enter this function, there has been a clean merge of
 	 * relevant trees, and the merge logic always stages whatever merges
 	 * cleanly.  We want to unstage those changes, unless it corresponds
-	 * to a file that didn't exist as of cache_tree.
+	 * to a file that didn't exist as of cache_tree.  However, if any
+	 * SKIP_WORKTREE path is modified relative to cache_tree, then we
+	 * want to clear the SKIP_WORKTREE bit and write it to the worktree
+	 * before unstaging.
 	 */
 
+	struct checkout state = CHECKOUT_INIT;
 	struct diff_options diff_opts;
 	struct lock_file lock = LOCK_INIT;
 	int i;
 
+	/* If any entries have skip_worktree set, we'll have to check 'em out */
+	state.force = 1;
+	state.quiet = 1;
+	state.refresh_cache = 1;
+	state.istate = &the_index;
+
 	diff_setup(&diff_opts);
 	diff_opts.flags.recursive = 1;
 	diff_opts.detect_rename = 0;

@@ -367,6 +377,30 @@ static void unstage_changes_unless_new(struct object_id *cache_tree)
 			continue;
 		}
 		ce = active_cache[pos];
+		if (ce_skip_worktree(ce)) {
+			struct stat st;
+			if (!lstat(ce->name, &st)) {
+				struct strbuf new_path = STRBUF_INIT;
+				int fd;
+
+				strbuf_addf(&new_path,
+					    "%s.stash.XXXXXX", ce->name);
+				fd = xmkstemp(new_path.buf);
+				close(fd);
+				printf(_("WARNING: Untracked file in way of "
+					 "tracked file!  Renaming\n "
+					 "           %s -> %s\n"
+					 "         to make room.\n"),
+				       ce->name, new_path.buf);
+				if (rename(ce->name, new_path.buf))
+					die("Failed to move %s to %s\n",
+					    ce->name, new_path.buf);
+				strbuf_release(&new_path);
+			}
+			checkout_entry(ce, &state, NULL, NULL);
+		}
+
+		ce->ce_flags &= ~CE_SKIP_WORKTREE;
 		if (p->one->oid_valid) {
 			ce = make_cache_entry(&the_index,
 					      p->one->mode,

index a184ee97fb..e5c6a038fb 100755
@@ -149,7 +149,7 @@ test_expect_success '--ignore-skip-worktree-entries leaves worktree alone' '
 		--diff-filter=D -- keep-me.t
 '
 
-test_expect_failure 'stash restore in sparse checkout' '
+test_expect_success 'stash restore in sparse checkout' '
 	test_create_repo stash-restore &&
 	(
 		cd stash-restore &&

gitgitgadget

:-)

Not that this should affect any of these changes, but I'd also note that the fact that using -u or -a makes a third commit that cannot be ignored later is a problem as well. (git stash list should probably annotate the listed stashes as to whether they are two- or three-commit stashes.)

s/might/might think/, perhaps? And, s/proceeded/preceded/, probably.

(I didn't look closely at the rest of this but the idea seems sound.)

Chris

Yeah, it seems a bit odd when you compare to saving and applying cached changes; for that, stash requires you to specify BOTH --keep-index to to push/save, AND --index to pop/apply.

However, for saving and applying untracked or ignored changes, you merely need to record them at push/save time and they automatically always apply.

Not very consistent. (But, again, unrelated to these patches, so this is mostly just recreational complaining at this point. I'm not planning to tackle these bits.)

Indeed; thanks for catching those. I will fix them up.

Any comments on this from reviewers? The second patch is a but too busy looking and I am having a bit of trouble convincing myself that it is doing the right thing.

Thanks.

Hmm, that diff is a little hard to read. It's a removal of two functions, and an addition of a new one, but the way the diff reads it looks like I'm modifying the existing functions because it catches some comment line markers and thinks they're similar. Maybe it'd be easier to read if I inserted that function elsewhere in the file?

I'll send a re-roll with that and add a comment or two to help explain it (as well as fix up the small issues Chris highlighted).

This series fixes some issues with git stash apply in sparse checkouts, when the sparsity patterns have changed between when the stash was created and when the stash is being applied. It also incidentally fixes a submodule testcase unrelated to sparse checkouts.

Changes since v1:

* Commit message cleanup, including a couple wording issues highlighted by Chris
* Cleaned up the code a bit and commented it better, to try to make it easier for Junio (and others) to follow.

Elijah Newren (3):
t7012: add a testcase demonstrating stash apply bugs in sparse
checkouts
stash: remove unnecessary process forking
stash: fix stash application in sparse-checkouts

 builtin/stash.c                  | 165 ++++++++++++++++++++++---------
 t/lib-submodule-update.sh        |  16 +--
 t/t7012-skip-worktree-writing.sh |  88 +++++++++++++++++
 3 files changed, 212 insertions(+), 57 deletions(-)

base-commit: 898f80736c75878acc02dc55672317fcc0e0a5a6
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-919%2Fnewren%2Fsparse-checkout-fixups-v2
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-919/newren/sparse-checkout-fixups-v2
Pull-Request: https://github.com/git/git/pull/919

Range-diff vs v1:

1: de4b4d207b ! 1: 2155bbfe20 t7012: add a testcase demonstrating stash apply bugs in sparse checkouts
@@ Commit message
t7012: add a testcase demonstrating stash apply bugs in sparse checkouts

Applying stashes in sparse-checkouts, particularly when the patterns
- used to define the present paths have changed between when the stash was
+ used to define the sparseness have changed between when the stash was
created and when it is applied, has a number of bugs. The primary
- problem is perhaps that stashes can be only partially applied (sometimes
- without any warning or error being displayed and with 0 exit status).
- In addition, there are cases where partial stash application comes with
- non-translated error messages and an early abort. The first is when
- there are files present despite the SKIP_WORKTREE bit being set, in
- which case the error message shown is:
+ problem is that stashes are sometimes only partially applied. In most
+ such cases, it does so silently without any warning or error being
+ displayed and with 0 exit status.
+
+ There are, however, a few cases when non-translated error messages are
+ shown and the stash application aborts early. The first is when there
+ are files present despite the SKIP_WORKTREE bit being set, in which case
+ the error message shown is:

error: Entry 'PATHNAME' not uptodate. Cannot merge.

The other situation is when a stash contains new files to add to the
- working tree; in this case, the code aborts early (but after at least
- partial stash application) with the following error message being shown:
+ working tree; in this case, the code aborts early but still has the
+ stash partially applied, and shows the following error message:

error: NEWFILE: does not exist and --remove not passed
fatal: Unable to process path NEWFILE

- Add a test that can trigger all three of these problems that carefully
- checks that the working copy and SKIP_WORKTREE bits are as expected
- after the stash application...but currently marked as expected to fail.
+ Add a test that can trigger all three of these problems. Have it
+ carefully check that the working copy and SKIP_WORKTREE bits are as
+ expected after the stash application. The test is currently marked as
+ expected to fail, but subsequent commits will implement the fixes and
+ toggle the expectation.

2: eb9ebcf0bd ! 2: 1fa263cf3c stash: remove unnecessary process forking

     @@ Commit message
          into a library function that does the same thing.  (The read-tree
          --reset was already partially converted over to a library call, but as
          an independent piece.)  Note here that this came after a merge operation
     -    was performed.  The merge machinery always staged anything that cleanly
     -    merges, and the above code only runs if there were no conflicts.  Its
     +    was performed.  The merge machinery always stages anything that cleanly
     +    merges, and the above code only runs if there are no conflicts.  Its
          purpose is to make it so that when there are no conflicts, all the
          changes from the stash are unstaged.  However, that causes brand new
          files from the stash to become untracked, so the code above first saves
     @@ Commit message
          We replace the whole series of commands with a simple function that will
          unstage files that are not newly added.  This doesn't fix any bugs in
          the usage of these commands, it simply matches the existing behavior but
     -    making it an actual builtin that we can then operate on as a whole.  A
     -    subsequent commit will take advantage of this to fix issues with these
     -    commands in sparse-checkouts.
     +    makes it into a single atomic operation that we can then operate on as a
     +    whole.  A subsequent commit will take advantage of this to fix issues
     +    with these commands in sparse-checkouts.
      
          This conversion incidentally fixes t3906.1, because the separate
          update-index process would die with the following error messages:
     @@ builtin/stash.c: static void add_diff_to_buf(struct diff_queue_struct *q,
       }
       
      -static int get_newly_staged(struct strbuf *out, struct object_id *c_tree)
     -+static void unstage_changes_unless_new(struct object_id *cache_tree)
     - {
     +-{
      -	struct child_process cp = CHILD_PROCESS_INIT;
      -	const char *c_tree_hex = oid_to_hex(c_tree);
      -
     - 	/*
     +-	/*
      -	 * diff-index is very similar to diff-tree above, and should be
      -	 * converted together with update_index.
     -+	 * When we enter this function, there has been a clean merge of
     -+	 * relevant trees, and the merge logic always stages whatever merges
     -+	 * cleanly.  We want to unstage those changes, unless it corresponds
     -+	 * to a file that didn't exist as of cache_tree.
     - 	 */
     +-	 */
      -	cp.git_cmd = 1;
      -	strvec_pushl(&cp.args, "diff-index", "--cached", "--name-only",
      -		     "--diff-filter=A", NULL);
      -	strvec_push(&cp.args, c_tree_hex);
      -	return pipe_command(&cp, NULL, 0, out, 0, NULL, 0);
      -}
     - 
     +-
      -static int update_index(struct strbuf *out)
      -{
      -	struct child_process cp = CHILD_PROCESS_INIT;
     -+	struct diff_options diff_opts;
     -+	struct lock_file lock = LOCK_INIT;
     -+	int i;
     - 
     +-
      -	/*
      -	 * Update-index is very complicated and may need to have a public
      -	 * function exposed in order to remove this forking.
     @@ builtin/stash.c: static void add_diff_to_buf(struct diff_queue_struct *q,
      -	cp.git_cmd = 1;
      -	strvec_pushl(&cp.args, "update-index", "--add", "--stdin", NULL);
      -	return pipe_command(&cp, out->buf, out->len, NULL, 0, NULL, 0);
     +-}
     +-
     + static int restore_untracked(struct object_id *u_tree)
     + {
     + 	int res;
     +@@ builtin/stash.c: static int restore_untracked(struct object_id *u_tree)
     + 	return res;
     + }
     + 
     ++static void unstage_changes_unless_new(struct object_id *orig_tree)
     ++{
     ++	/*
     ++	 * When we enter this function, there has been a clean merge of
     ++	 * relevant trees, and the merge logic always stages whatever merges
     ++	 * cleanly.  We want to unstage those changes, unless it corresponds
     ++	 * to a file that didn't exist as of orig_tree.
     ++	 */
     ++
     ++	struct diff_options diff_opts;
     ++	struct lock_file lock = LOCK_INIT;
     ++	int i;
     ++
     ++	/*
     ++	 * Step 1: get a difference between orig_tree (which corresponding
     ++	 * to the index before a merge was run) and the current index
     ++	 * (reflecting the changes brought in by the merge).
     ++	 */
      +	diff_setup(&diff_opts);
      +	diff_opts.flags.recursive = 1;
      +	diff_opts.detect_rename = 0;
      +	diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
      +	diff_setup_done(&diff_opts);
      +
     -+	do_diff_cache(cache_tree, &diff_opts);
     ++	do_diff_cache(orig_tree, &diff_opts);
      +	diffcore_std(&diff_opts);
      +
     ++	/* Iterate over the paths that changed due to the merge... */
      +	for (i = 0; i < diff_queued_diff.nr; i++) {
      +		struct diff_filepair *p;
      +		struct cache_entry *ce;
      +		int pos;
      +
     ++		/* Look up the path's position in the current index. */
      +		p = diff_queued_diff.queue[i];
      +		pos = index_name_pos(&the_index, p->two->path,
      +				     strlen(p->two->path));
     -+		if (pos < 0) {
     -+			assert(p->one->oid_valid);
     -+			assert(!p->two->oid_valid);
     -+			ce = make_cache_entry(&the_index,
     -+					      p->one->mode,
     -+					      &p->one->oid,
     -+					      p->one->path,
     -+					      0, 0);
     -+			add_index_entry(&the_index, ce, ADD_CACHE_OK_TO_ADD);
     -+			continue;
     -+		}
     -+		ce = active_cache[pos];
     ++
     ++		/*
     ++		 * Step 2: "unstage" changes, as long as they are still tracked
     ++		 */
      +		if (p->one->oid_valid) {
     ++			/*
     ++			 * Path existed in orig_tree; restore index entry
     ++			 * from that tree in order to "unstage" the changes.
     ++			 */
     ++			int option = ADD_CACHE_OK_TO_REPLACE;
     ++			if (pos < 0)
     ++				option = ADD_CACHE_OK_TO_ADD;
     ++
      +			ce = make_cache_entry(&the_index,
      +					      p->one->mode,
      +					      &p->one->oid,
      +					      p->one->path,
      +					      0, 0);
     -+			add_index_entry(&the_index, ce,
     -+					ADD_CACHE_OK_TO_REPLACE);
     ++			add_index_entry(&the_index, ce, option);
      +		}
      +	}
      +	diff_flush(&diff_opts);
      +
     ++	/*
     ++	 * Step 3: write the new index to disk
     ++	 */
      +	repo_hold_locked_index(the_repository, &lock, LOCK_DIE_ON_ERROR);
      +	if (write_locked_index(&the_index, &lock,
      +			       COMMIT_LOCK | SKIP_IF_UNCHANGED))
      +		die(_("Unable to write index."));
     - }
     - 
     - static int restore_untracked(struct object_id *u_tree)
     ++}
     ++
     + static int do_apply_stash(const char *prefix, struct stash_info *info,
     + 			  int index, int quiet)
     + {
      @@ builtin/stash.c: static int do_apply_stash(const char *prefix, struct stash_info *info,
       		if (reset_tree(&index_tree, 0, 0))
       			return -1;

3: 5143cba704 ! 3: ccfedc7140 stash: fix stash application in sparse-checkouts
@@ Commit message
staged at the end and include more things that should be quasi-merged,
(2) stash generally wants changes to NOT be staged. It only provides exceptions when (a) some of the changes had conflicts and thus we want
- to use staged to denote the clean merges and higher order stages to
+ to use stages to denote the clean merges and higher order stages to
mark the conflicts, or (b) if there is a brand new file we don't want it to become untracked.

@@ Commit message
git update-index --add --stdin <"$a" rm -f "$a"

- You might that between the merge that proceeded these commands and these
- different commands here, that we have commands from each of the
- different types of special sparsity handling listed at the beginning of
- this message, and in fact this precisely led to the following buggy
- behaviors:
+ Looking back at the different types of special sparsity handling listed
+ at the beginning of this message, you may note that we have at least one
+ of each type covered here: merge, diff-index, and read-tree. The weird
+ mix-and-match led to 3 different bugs:

(1) If a path merged cleanly and it didn't match the sparsity patterns, the merge backend would know to avoid writing it to the working tree and
@@ Commit message

## builtin/stash.c ##
-@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *cache_tree)
- * When we enter this function, there has been a clean merge of
+@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *orig_tree)
* relevant trees, and the merge logic always stages whatever merges
* cleanly. We want to unstage those changes, unless it corresponds
-- * to a file that didn't exist as of cache_tree.
-+ * to a file that didn't exist as of cache_tree. However, if any
-+ * SKIP_WORKTREE path is modified relative to cache_tree, then we
-+ * want to clear the SKIP_WORKTREE bit and write it to the worktree
-+ * before unstaging.
+ * to a file that didn't exist as of orig_tree.
++ *
++ * However, if any SKIP_WORKTREE path is modified relative to
++ * orig_tree, then we want to clear the SKIP_WORKTREE bit and write
++ * it to the worktree before unstaging.
*/

+ struct checkout state = CHECKOUT_INIT;
@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *cache_
+ state.refresh_cache = 1;
+ state.istate = &the_index;
+
- diff_setup(&diff_opts);
- diff_opts.flags.recursive = 1;
- diff_opts.detect_rename = 0;
-@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *cache_tree)
- continue;
- }
- ce = active_cache[pos];
-+ if (ce_skip_worktree(ce)) {
+ /*
+ * Step 1: get a difference between orig_tree (which corresponding
+ * to the index before a merge was run) and the current index
+@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *orig_tree)
+ strlen(p->two->path));
+
+ /*
+- * Step 2: "unstage" changes, as long as they are still tracked
++ * Step 2: Place changes in the working tree
++ *
++ * Stash is about restoring changes *to the working tree*.
++ * So if the merge successfully got a new version of some
++ * path, but left it out of the working tree, then clear the
++ * SKIP_WORKTREE bit and write it to the working tree.
++ */
++ if (pos >= 0 && ce_skip_worktree(active_cache[pos])) {
+ struct stat st;
++
++ ce = active_cache[pos];
+ if (!lstat(ce->name, &st)) {
++ /* Conflicting path present; relocate it */
+ struct strbuf new_path = STRBUF_INIT;
+ int fd;
+
@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *cache_
+ strbuf_release(&new_path);
+ }
+ checkout_entry(ce, &state, NULL, NULL);
++ ce->ce_flags &= ~CE_SKIP_WORKTREE;
+ }
+
-+ ce->ce_flags &= ~CE_SKIP_WORKTREE;
++ /*
++ * Step 3: "unstage" changes, as long as they are still tracked
+ */
if (p->one->oid_valid) {
- ce = make_cache_entry(&the_index,
- p->one->mode,
+ /*
+@@ builtin/stash.c: static void unstage_changes_unless_new(struct object_id *orig_tree)
+ diff_flush(&diff_opts);
+
+ /*
+- * Step 3: write the new index to disk
++ * Step 4: write the new index to disk
+ */
+ repo_hold_locked_index(the_repository, &lock, LOCK_DIE_ON_ERROR);
+ if (write_locked_index(&the_index, &lock,

## t/t7012-skip-worktree-writing.sh ##
@@ t/t7012-skip-worktree-writing.sh: test_expect_success '--ignore-skip-worktree-entries leaves worktree alone' '

gitgitgadget

From: Elijah Newren <[email protected]>

Applying stashes in sparse-checkouts, particularly when the patterns used to define the sparseness have changed between when the stash was created and when it is applied, has a number of bugs. The primary problem is that stashes are sometimes only partially applied. In most such cases, it does so silently without any warning or error being displayed and with 0 exit status.

There are, however, a few cases when non-translated error messages are shown and the stash application aborts early. The first is when there are files present despite the SKIP_WORKTREE bit being set, in which case the error message shown is:

error: Entry 'PATHNAME' not uptodate. Cannot merge.

The other situation is when a stash contains new files to add to the working tree; in this case, the code aborts early but still has the stash partially applied, and shows the following error message:

error: NEWFILE: does not exist and --remove not passed
fatal: Unable to process path NEWFILE

Add a test that can trigger all three of these problems. Have it carefully check that the working copy and SKIP_WORKTREE bits are as expected after the stash application. The test is currently marked as expected to fail, but subsequent commits will implement the fixes and toggle the expectation.

---
 t/t7012-skip-worktree-writing.sh | 88 ++++++++++++++++++++++++++++++++
 1 file changed, 88 insertions(+)

index 7476781979..a184ee97fb 100755
@@ -149,6 +149,94 @@ test_expect_success '--ignore-skip-worktree-entries leaves worktree alone' '
 		--diff-filter=D -- keep-me.t
 '
 
+test_expect_failure 'stash restore in sparse checkout' '
+	test_create_repo stash-restore &&
+	(
+		cd stash-restore &&
+
+		mkdir subdir &&
+		echo A >subdir/A &&
+		echo untouched >untouched &&
+		echo removeme >removeme &&
+		echo modified >modified &&
+		git add . &&
+		git commit -m Initial &&
+
+		echo AA >>subdir/A &&
+		echo addme >addme &&
+		echo tweaked >>modified &&
+		rm removeme &&
+		git add addme &&
+
+		git stash push &&
+
+		git sparse-checkout set subdir &&
+
+		# Ensure after sparse-checkout we only have expected files
+		cat >expect <<-EOF &&
+		S modified
+		S removeme
+		H subdir/A
+		S untouched
+		EOF
+		git ls-files -t >actual &&
+		test_cmp expect actual &&
+
+		test_path_is_missing addme &&
+		test_path_is_missing modified &&
+		test_path_is_missing removeme &&
+		test_path_is_file    subdir/A &&
+		test_path_is_missing untouched &&
+
+		# Put a file in the working directory in the way
+		echo in the way >modified &&
+		git stash apply &&
+
+		# Ensure stash vivifies modifies paths...
+		cat >expect <<-EOF &&
+		H addme
+		H modified
+		H removeme
+		H subdir/A
+		S untouched
+		EOF
+		git ls-files -t >actual &&
+		test_cmp expect actual &&
+
+		# ...and that the paths show up in status as changed...
+		cat >expect <<-EOF &&
+		A  addme
+		 M modified
+		 D removeme
+		 M subdir/A
+		?? actual
+		?? expect
+		?? modified.stash.XXXXXX
+		EOF
+		git status --porcelain | \
+			sed -e s/stash......./stash.XXXXXX/ >actual &&
+		test_cmp expect actual &&
+
+		# ...and that working directory reflects the files correctly
+		test_path_is_file    addme &&
+		test_path_is_file    modified &&
+		test_path_is_missing removeme &&
+		test_path_is_file    subdir/A &&
+		test_path_is_missing untouched &&
+
+		# ...including that we have the expected "modified" file...
+		cat >expect <<-EOF &&
+		modified
+		tweaked
+		EOF
+		test_cmp expect modified &&
+
+		# ...and that the other "modified" file is still present...
+		echo in the way >expect &&
+		test_cmp expect modified.stash.*
+	)
+'
+
 #TODO test_expect_failure 'git-apply adds file' false
 #TODO test_expect_failure 'git-apply updates file' false
 #TODO test_expect_failure 'git-apply removes file' false

gitgitgadget

From: Elijah Newren <[email protected]>

When stash was converted from shell to a builtin, it merely transliterated the forking of various git commands from shell to a C program that would fork the same commands. Some of those were converted over to actual library calls, but much of the pipeline-of-commands design still remains. Fix some of this by replacing the portion corresponding to

    git diff-index --cached --name-only --diff-filter=A $CTREE >"$a"
    git read-tree --reset $CTREE
    git update-index --add --stdin <"$a"
    rm -f "$a"

into a library function that does the same thing. (The read-tree --reset was already partially converted over to a library call, but as an independent piece.) Note here that this came after a merge operation was performed. The merge machinery always stages anything that cleanly merges, and the above code only runs if there are no conflicts. Its purpose is to make it so that when there are no conflicts, all the changes from the stash are unstaged. However, that causes brand new files from the stash to become untracked, so the code above first saves those files off and then re-adds them afterwards.

We replace the whole series of commands with a simple function that will unstage files that are not newly added. This doesn't fix any bugs in the usage of these commands, it simply matches the existing behavior but makes it into a single atomic operation that we can then operate on as a whole. A subsequent commit will take advantage of this to fix issues with these commands in sparse-checkouts.

This conversion incidentally fixes t3906.1, because the separate update-index process would die with the following error messages:
error: uninitialized_sub: is a directory - add files inside instead
fatal: Unable to process path uninitialized_sub
The unstaging of the directory as a submodule meant it was no longer tracked, and thus as an uninitialized directory it could not be added back using git update-index --add, thus resulting in this error and early abort. Most of the submodule tests in 3906 continue to fail after this change, this change was just enough to push the first of those tests to success.

---
 builtin/stash.c           | 119 ++++++++++++++++++++++----------------
 t/lib-submodule-update.sh |  16 ++---
 2 files changed, 78 insertions(+), 57 deletions(-)

index 3f811f3050..a5465b565a 100644
@@ -325,35 +325,6 @@ static void add_diff_to_buf(struct diff_queue_struct *q,
 	}
 }
 
-static int get_newly_staged(struct strbuf *out, struct object_id *c_tree)
-{
-	struct child_process cp = CHILD_PROCESS_INIT;
-	const char *c_tree_hex = oid_to_hex(c_tree);
-
-	/*
-	 * diff-index is very similar to diff-tree above, and should be
-	 * converted together with update_index.
-	 */
-	cp.git_cmd = 1;
-	strvec_pushl(&cp.args, "diff-index", "--cached", "--name-only",
-		     "--diff-filter=A", NULL);
-	strvec_push(&cp.args, c_tree_hex);
-	return pipe_command(&cp, NULL, 0, out, 0, NULL, 0);
-}
-
-static int update_index(struct strbuf *out)
-{
-	struct child_process cp = CHILD_PROCESS_INIT;
-
-	/*
-	 * Update-index is very complicated and may need to have a public
-	 * function exposed in order to remove this forking.
-	 */
-	cp.git_cmd = 1;
-	strvec_pushl(&cp.args, "update-index", "--add", "--stdin", NULL);
-	return pipe_command(&cp, out->buf, out->len, NULL, 0, NULL, 0);
-}
-
 static int restore_untracked(struct object_id *u_tree)
 {
 	int res;

@@ -385,6 +356,75 @@ static int restore_untracked(struct object_id *u_tree)
 	return res;
 }
 
+static void unstage_changes_unless_new(struct object_id *orig_tree)
+{
+	/*
+	 * When we enter this function, there has been a clean merge of
+	 * relevant trees, and the merge logic always stages whatever merges
+	 * cleanly.  We want to unstage those changes, unless it corresponds
+	 * to a file that didn't exist as of orig_tree.
+	 */
+
+	struct diff_options diff_opts;
+	struct lock_file lock = LOCK_INIT;
+	int i;
+
+	/*
+	 * Step 1: get a difference between orig_tree (which corresponding
+	 * to the index before a merge was run) and the current index
+	 * (reflecting the changes brought in by the merge).
+	 */
+	diff_setup(&diff_opts);
+	diff_opts.flags.recursive = 1;
+	diff_opts.detect_rename = 0;
+	diff_opts.output_format = DIFF_FORMAT_NO_OUTPUT;
+	diff_setup_done(&diff_opts);
+
+	do_diff_cache(orig_tree, &diff_opts);
+	diffcore_std(&diff_opts);
+
+	/* Iterate over the paths that changed due to the merge... */
+	for (i = 0; i < diff_queued_diff.nr; i++) {
+		struct diff_filepair *p;
+		struct cache_entry *ce;
+		int pos;
+
+		/* Look up the path's position in the current index. */
+		p = diff_queued_diff.queue[i];
+		pos = index_name_pos(&the_index, p->two->path,
+				     strlen(p->two->path));
+
+		/*
+		 * Step 2: "unstage" changes, as long as they are still tracked
+		 */
+		if (p->one->oid_valid) {
+			/*
+			 * Path existed in orig_tree; restore index entry
+			 * from that tree in order to "unstage" the changes.
+			 */
+			int option = ADD_CACHE_OK_TO_REPLACE;
+			if (pos < 0)
+				option = ADD_CACHE_OK_TO_ADD;
+
+			ce = make_cache_entry(&the_index,
+					      p->one->mode,
+					      &p->one->oid,
+					      p->one->path,
+					      0, 0);
+			add_index_entry(&the_index, ce, option);
+		}
+	}
+	diff_flush(&diff_opts);
+
+	/*
+	 * Step 3: write the new index to disk
+	 */
+	repo_hold_locked_index(the_repository, &lock, LOCK_DIE_ON_ERROR);
+	if (write_locked_index(&the_index, &lock,
+			       COMMIT_LOCK | SKIP_IF_UNCHANGED))
+		die(_("Unable to write index."));
+}
+
 static int do_apply_stash(const char *prefix, struct stash_info *info,
 			  int index, int quiet)
 {

@@ -467,26 +507,7 @@ static int do_apply_stash(const char *prefix, struct stash_info *info,
 		if (reset_tree(&index_tree, 0, 0))
 			return -1;
 	} else {
-		struct strbuf out = STRBUF_INIT;
-
-		if (get_newly_staged(&out, &c_tree)) {
-			strbuf_release(&out);
-			return -1;
-		}
-
-		if (reset_tree(&c_tree, 0, 1)) {
-			strbuf_release(&out);
-			return -1;
-		}
-
-		ret = update_index(&out);
-		strbuf_release(&out);
-		if (ret)
-			return -1;
-
-		/* read back the result of update_index() back from the disk */
-		discard_cache();
-		read_cache();
+		unstage_changes_unless_new(&c_tree);
 	}
 
 	if (!quiet) {

index 87a759149f..c6d7aab6df 100644
@@ -316,14 +316,7 @@ test_submodule_switch_common () {
 	command="$1"
 	######################### Appearing submodule #########################
 	# Switching to a commit letting a submodule appear creates empty dir ...
-	if test "$KNOWN_FAILURE_STASH_DOES_IGNORE_SUBMODULE_CHANGES" = 1
-	then
-		# Restoring stash fails to restore submodule index entry
-		RESULT="failure"
-	else
-		RESULT="success"
-	fi
-	test_expect_$RESULT "$command: added submodule creates empty directory" '
+	test_expect_success "$command: added submodule creates empty directory" '
 		prolog &&
 		reset_work_tree_to no_submodule &&
 		(

@@ -337,6 +330,13 @@ test_submodule_switch_common () {
 		)
 	'
 	# ... and doesn't care if it already exists.
+	if test "$KNOWN_FAILURE_STASH_DOES_IGNORE_SUBMODULE_CHANGES" = 1
+	then
+		# Restoring stash fails to restore submodule index entry
+		RESULT="failure"
+	else
+		RESULT="success"
+	fi
 	test_expect_$RESULT "$command: added submodule leaves existing empty directory alone" '
 		prolog &&
 		reset_work_tree_to no_submodule &&

gitgitgadget

This is orthogonal to what this patch does, as this is supposed to be just bug-for-bug compatible rewrite.

But I wonder if the above sequence, whether it is done as a series of plumbing invocations or subroutine calls, is a relic dating back in the days before i-t-a existed. If we want to revert the changes to the index for working tree files for removed or modified ones, I do not offhand see a good reason why we would want to keep the contents to new paths---if i-t-a were available when the sequence was designed, I suspect we would just have added the path as i-t-a in order to keep track of the presence of the path but not necessarily the contents in it.

Yeah, I thought a little bit about the same thing, but wasn't sure if there was some other reason for the current behavior or if there was some workflow that might be relying upon it. Rather than investigate and try to switch it over to i-t-a (which could still be done later), I more narrowly focused this series at just doing the "make the changes be available in the working tree, but remove them from the index unless it'd make it untracked" more carefully.

From: Elijah Newren <[email protected]>

sparse-checkouts are built on the patterns in the $GIT_DIR/info/sparse-checkout file, where commands have modified behavior for paths that do not match those patterns. The differences in behavior, as far as the bugs concerned here, fall into three different categories (with git subcommands that fall into each category listed):

* commands that only look at files matching the patterns:
* status
* diff
* clean
* update-index
* commands that remove files from the working tree that do not match the patterns, and restore files that do match them:
* read-tree
* switch
* checkout
* reset (--hard)
* commands that omit writing files to the working tree that do not match the patterns, unless those files are not clean:
* merge
* rebase
* cherry-pick
* revert

There are some caveats above, e.g. a plain git diff ignores files outside the sparsity patterns but will show diffs for paths outside the sparsity patterns when revision arguments are passed. (Technically, diff is treating the sparse paths as matching HEAD.) So, there is some internal inconsistency among these commands. There are also additional commands that should behave differently in the face of sparse-checkouts, as the sparse-checkout documentation alludes to, but the above is sufficient for me to explain how git stash is affected.

What is relevant here is that logically 'stash' should behave like a merge; it three-way merges the changes the user had in progress at stash creation time, the HEAD at the time the stash was created, and the current HEAD, in order to get the stashed changes applied to the current branch. However, this simplistic view doesn't quite work in practice, because stash tweaks it a bit due to two factors: (1) flags like --keep-index and --include-untracked (why we used two different verbs, 'keep' and 'include', is a rant for another day) modify what should be staged at the end and include more things that should be quasi-merged,
(2) stash generally wants changes to NOT be staged. It only provides exceptions when (a) some of the changes had conflicts and thus we want to use stages to denote the clean merges and higher order stages to mark the conflicts, or (b) if there is a brand new file we don't want it to become untracked.

stash has traditionally gotten this special behavior by first doing a merge, and then when it's clean, applying a pipeline of commands to modify the result. This series of commands for unstaging-non-newly-added-files came from the following commands:

    git diff-index --cached --name-only --diff-filter=A $CTREE >"$a"
    git read-tree --reset $CTREE
    git update-index --add --stdin <"$a"
    rm -f "$a"

Looking back at the different types of special sparsity handling listed at the beginning of this message, you may note that we have at least one of each type covered here: merge, diff-index, and read-tree. The weird mix-and-match led to 3 different bugs:

(1) If a path merged cleanly and it didn't match the sparsity patterns, the merge backend would know to avoid writing it to the working tree and keep the SKIP_WORKTREE bit, simply only updating it in the index. Unfortunately, the subsequent commands would essentially undo the changes in the index and thus simply toss the changes altogether since there was nothing left in the working tree. This means the stash is only partially applied.

(2) If a path existed in the worktree before git stash apply despite having the SKIP_WORKTREE bit set, then the git read-tree --reset would print an error message of the form
error: Entry 'modified' not uptodate. Cannot merge.
and cause stash to abort early.

(3) If there was a brand new file added by the stash, then the diff-index command would save that pathname to the temporary file, the read-tree --reset would remove it from the index, and the update-index command would barf due to no such file being present in the working copy; it would print a message of the form:
error: NEWFILE: does not exist and --remove not passed
fatal: Unable to process path NEWFILE
and then cause stash to abort early.

Basically, the whole idea of unstage-unless-brand-new requires special care when you are dealing with a sparse-checkout. Fix these problems by applying the following simple rule:

When we unstage files, if they have the SKIP_WORKTREE bit set, clear that bit and write the file out to the working directory.

(*) If there's already a file present in the way, rename it first.

This fixes all three problems in t7012.13 and allows us to mark it as passing.

---
 builtin/stash.c                  | 50 ++++++++++++++++++++++++++++++--
 t/t7012-skip-worktree-writing.sh |  2 +-
 2 files changed, 49 insertions(+), 3 deletions(-)

index a5465b565a..84886e84e0 100644
@@ -363,12 +363,23 @@ static void unstage_changes_unless_new(struct object_id *orig_tree)
 	 * relevant trees, and the merge logic always stages whatever merges
 	 * cleanly.  We want to unstage those changes, unless it corresponds
 	 * to a file that didn't exist as of orig_tree.
+	 *
+	 * However, if any SKIP_WORKTREE path is modified relative to
+	 * orig_tree, then we want to clear the SKIP_WORKTREE bit and write
+	 * it to the worktree before unstaging.
 	 */
 
+	struct checkout state = CHECKOUT_INIT;
 	struct diff_options diff_opts;
 	struct lock_file lock = LOCK_INIT;
 	int i;
 
+	/* If any entries have skip_worktree set, we'll have to check 'em out */
+	state.force = 1;
+	state.quiet = 1;
+	state.refresh_cache = 1;
+	state.istate = &the_index;
+
 	/*
 	 * Step 1: get a difference between orig_tree (which corresponding
 	 * to the index before a merge was run) and the current index

@@ -395,7 +406,42 @@ static void unstage_changes_unless_new(struct object_id *orig_tree)
 				     strlen(p->two->path));
 
 		/*
-		 * Step 2: "unstage" changes, as long as they are still tracked
+		 * Step 2: Place changes in the working tree
+		 *
+		 * Stash is about restoring changes *to the working tree*.
+		 * So if the merge successfully got a new version of some
+		 * path, but left it out of the working tree, then clear the
+		 * SKIP_WORKTREE bit and write it to the working tree.
+		 */
+		if (pos >= 0 && ce_skip_worktree(active_cache[pos])) {
+			struct stat st;
+
+			ce = active_cache[pos];
+			if (!lstat(ce->name, &st)) {
+				/* Conflicting path present; relocate it */
+				struct strbuf new_path = STRBUF_INIT;
+				int fd;
+
+				strbuf_addf(&new_path,
+					    "%s.stash.XXXXXX", ce->name);
+				fd = xmkstemp(new_path.buf);
+				close(fd);
+				printf(_("WARNING: Untracked file in way of "
+					 "tracked file!  Renaming\n "
+					 "           %s -> %s\n"
+					 "         to make room.\n"),
+				       ce->name, new_path.buf);
+				if (rename(ce->name, new_path.buf))
+					die("Failed to move %s to %s\n",
+					    ce->name, new_path.buf);
+				strbuf_release(&new_path);
+			}
+			checkout_entry(ce, &state, NULL, NULL);
+			ce->ce_flags &= ~CE_SKIP_WORKTREE;
+		}
+
+		/*
+		 * Step 3: "unstage" changes, as long as they are still tracked
 		 */
 		if (p->one->oid_valid) {
 			/*

@@ -417,7 +463,7 @@ static void unstage_changes_unless_new(struct object_id *orig_tree)
 	diff_flush(&diff_opts);
 
 	/*
-	 * Step 3: write the new index to disk
+	 * Step 4: write the new index to disk
 	 */
 	repo_hold_locked_index(the_repository, &lock, LOCK_DIE_ON_ERROR);
 	if (write_locked_index(&the_index, &lock,

index a184ee97fb..e5c6a038fb 100755
@@ -149,7 +149,7 @@ test_expect_success '--ignore-skip-worktree-entries leaves worktree alone' '
 		--diff-filter=D -- keep-me.t
 '
 
-test_expect_failure 'stash restore in sparse checkout' '
+test_expect_success 'stash restore in sparse checkout' '
 	test_create_repo stash-restore &&
 	(
 		cd stash-restore &&

gitgitgadget