Skip to contentSkip to conversation

Better threaded delta resolution in index-pack (another try)

PATCH Superseded Jonathan Tan 3 participants 28 messages Aug 24, 2020 – Sep 8, 2020 Documentation

lore
Share & tools
lore
#0 lore raw

cover Better threaded delta resolution in index-pack (another try)

I'm trying to resurrect [1] and have rebased it to latest master (675a4aaf3b ("Ninth batch", 2020-08-19)).

Peff said [2] (of v1) that the overall direction seems reasonable and Josh Steadmon said [3] (of v2) that it looks mostly good except for possible improvements to commit messages and comments. Josh did not list out specific improvements to commit messages but I have taken his suggestions for comments.

[1] https://lore.kernel.org/git/[email protected]/
[2] https://lore.kernel.org/git/[email protected]/
[3] https://lore.kernel.org/git/[email protected]/

Jonathan Tan (7):
Documentation: deltaBaseCacheLimit is per-thread
index-pack: remove redundant parameter
index-pack: unify threaded and unthreaded code
index-pack: remove redundant child field
index-pack: calculate {ref,ofs}_{first,last} early
index-pack: make resolve_delta() assume base data
index-pack: make quantum of work smaller

 Documentation/config/core.txt |   2 +-
 builtin/index-pack.c          | 449 ++++++++++++++++++----------------
 2 files changed, 244 insertions(+), 207 deletions(-)
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#1 lore raw

PATCH 1/7 Documentation: deltaBaseCacheLimit is per-thread

Clarify that core.deltaBaseCacheLimit is per-thread, as can be seen from the fact that cache usage (base_cache_used in struct thread_local in builtin/index-pack.c) is tracked individually for each thread and compared against delta_base_cache_limit.

---
 Documentation/config/core.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Documentation/config/core.txt +1 -1
index 74619a9c03..02002cf109 100644
@@ -399,7 +399,7 @@ the largest projects.  You probably do not need to adjust this value.
 Common unit suffixes of 'k', 'm', or 'g' are supported.
 
 core.deltaBaseCacheLimit::
-	Maximum number of bytes to reserve for caching base objects
+	Maximum number of bytes per thread to reserve for caching base objects
 	that may be referenced by multiple deltified objects.  By storing the
 	entire decompressed base objects in a cache Git is able
 	to avoid unpacking and decompressing frequently used base
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#2 lore raw

PATCH 1/1 fetch-pack: in partial clone, pass --promisor

In reply to Jonathan Tan

When fetching a pack from a promisor remote, the corresponding .promisor file needs to be created. "fetch-pack" originally did this by passing "--promisor" to "index-pack", but in 5374a290aa ("fetch-pack: write fetched refs to .promisor", 2019-10-16), "fetch-pack" was taught to do this itself instead, because it needed to store ref information in the .promisor file.

This causes a problem with superprojects when transfer.fsckobjects is set, because in the current implementation, it is "index-pack" that calls fsck_finish() to check the objects; before 5374a290aa, fsck_finish() would see that .gitmodules is a promisor object and tolerate it being missing, but after, there is no .promisor file (at the time of the invocation of fsck_finish() by "index-pack") to tell it that .gitmodules is a promisor object, so it returns an error.

Therefore, teach "fetch-pack" to pass "--promisor" to index pack once again. "fetch-pack" will subsequently overwrite this file with the ref information.

An alternative is to instead move object checking to "fetch-pack", and let "index-pack" only index the files. However, since "index-pack" has to inflate objects in order to index them, it seems reasonable to also let it check the objects (which also require inflated files).

---
 fetch-pack.c             | 17 ++++++++++-------
 t/t5616-partial-clone.sh | 16 ++++++++++++++++
 2 files changed, 26 insertions(+), 7 deletions(-)
fetch-pack.c +10 -7
index 7f20eca4f8..d467edc24e 100644
@@ -866,13 +866,16 @@ static int get_pack(struct fetch_pack_args *args,
 			 * have this responsibility.
 			 */
 			args->check_self_contained_and_connected = 0;
-		/*
-		 * If we're obtaining the filename of a lockfile, we'll use
-		 * that filename to write a .promisor file with more
-		 * information below. If not, we need index-pack to do it for
-		 * us.
-		 */
-		if (!(do_keep && pack_lockfiles) && args->from_promisor)
+
+		if (args->from_promisor)
+			/*
+			 * write_promisor_file() may be called afterwards but
+			 * we still need index-pack to know that this is a
+			 * promisor pack. For example, if transfer.fsckobjects
+			 * is true, index-pack needs to know that .gitmodules
+			 * is a promisor object (so that it won't complain if
+			 * it is missing).
+			 */
 			strvec_push(&cmd.args, "--promisor");
 	}
 	else {
t/t5616-partial-clone.sh +16 -0
index 8827c2ed18..5a01466db4 100755
@@ -163,6 +163,22 @@ test_expect_success 'manual prefetch of missing objects' '
 	test_line_count = 0 observed.oids
 '
 
+test_expect_success 'partial clone with transfer.fsckobjects=1 works with submodules' '
+	test_create_repo submodule &&
+	test_commit -C submodule mycommit &&
+
+	test_create_repo src_with_sub &&
+	test_config -C src_with_sub uploadpack.allowfilter 1 &&
+	test_config -C src_with_sub uploadpack.allowanysha1inwant 1 &&
+
+	git -C src_with_sub submodule add "file://$(pwd)/submodule" mysub &&
+	git -C src_with_sub commit -m "commit with submodule" &&
+
+	git -c transfer.fsckobjects=1 \
+		clone --filter="blob:none" "file://$(pwd)/src_with_sub" dst &&
+	test_when_finished rm -rf dst
+'
+
 test_expect_success 'partial clone with transfer.fsckobjects=1 uses index-pack --fsck-objects' '
 	git init src &&
 	test_commit -C src x &&
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#3 lore raw

Re: [PATCH] fetch-pack: in partial clone, pass --promisor

Quoted 6 lines

When fetching a pack from a promisor remote, the corresponding .promisor file needs to be created. "fetch-pack" originally did this by passing "--promisor" to "index-pack", but in 5374a290aa ("fetch-pack: write fetched refs to .promisor", 2019-10-16), "fetch-pack" was taught to do this itself instead, because it needed to store ref information in the .promisor file.

Oops...please ignore this. This has already been sent out as
https://lore.kernel.org/git/[email protected]/

#4 lore raw

PATCH 2/7 index-pack: remove redundant parameter

In reply to Jonathan Tan

find_{ref,ofs}_delta_{,children} take an enum object_type parameter, but the object type is already present in the name of the function. Remove that parameter from these functions.

---
 builtin/index-pack.c | 26 ++++++++++++--------------
 1 file changed, 12 insertions(+), 14 deletions(-)
builtin/index-pack.c +12 -14
index f865666db9..a8444d550b 100644
@@ -614,7 +614,7 @@ static int compare_ofs_delta_bases(off_t offset1, off_t offset2,
 	       0;
 }
 
-static int find_ofs_delta(const off_t offset, enum object_type type)
+static int find_ofs_delta(const off_t offset)
 {
 	int first = 0, last = nr_ofs_deltas;
 
@@ -624,7 +624,8 @@ static int find_ofs_delta(const off_t offset, enum object_type type)
 		int cmp;
 
 		cmp = compare_ofs_delta_bases(offset, delta->offset,
-					      type, objects[delta->obj_no].type);
+					      OBJ_OFS_DELTA,
+					      objects[delta->obj_no].type);
 		if (!cmp)
 			return next;
 		if (cmp < 0) {
@@ -637,10 +638,9 @@ static int find_ofs_delta(const off_t offset, enum object_type type)
 }
 
 static void find_ofs_delta_children(off_t offset,
-				    int *first_index, int *last_index,
-				    enum object_type type)
+				    int *first_index, int *last_index)
 {
-	int first = find_ofs_delta(offset, type);
+	int first = find_ofs_delta(offset);
 	int last = first;
 	int end = nr_ofs_deltas - 1;
 
@@ -668,7 +668,7 @@ static int compare_ref_delta_bases(const struct object_id *oid1,
 	return oidcmp(oid1, oid2);
 }
 
-static int find_ref_delta(const struct object_id *oid, enum object_type type)
+static int find_ref_delta(const struct object_id *oid)
 {
 	int first = 0, last = nr_ref_deltas;
 
@@ -678,7 +678,8 @@ static int find_ref_delta(const struct object_id *oid, enum object_type type)
 		int cmp;
 
 		cmp = compare_ref_delta_bases(oid, &delta->oid,
-					      type, objects[delta->obj_no].type);
+					      OBJ_REF_DELTA,
+					      objects[delta->obj_no].type);
 		if (!cmp)
 			return next;
 		if (cmp < 0) {
@@ -691,10 +692,9 @@ static int find_ref_delta(const struct object_id *oid, enum object_type type)
 }
 
 static void find_ref_delta_children(const struct object_id *oid,
-				    int *first_index, int *last_index,
-				    enum object_type type)
+				    int *first_index, int *last_index)
 {
-	int first = find_ref_delta(oid, type);
+	int first = find_ref_delta(oid);
 	int last = first;
 	int end = nr_ref_deltas - 1;
 
@@ -983,12 +983,10 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 {
 	if (base->ref_last == -1 && base->ofs_last == -1) {
 		find_ref_delta_children(&base->obj->idx.oid,
-					&base->ref_first, &base->ref_last,
-					OBJ_REF_DELTA);
+					&base->ref_first, &base->ref_last);
 
 		find_ofs_delta_children(base->obj->idx.offset,
-					&base->ofs_first, &base->ofs_last,
-					OBJ_OFS_DELTA);
+					&base->ofs_first, &base->ofs_last);
 
 		if (base->ref_last == -1 && base->ofs_last == -1) {
 			free(base->data);
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#5 lore raw

Re: [PATCH 2/7] index-pack: remove redundant parameter

Quoted 3 lines · Jonathan Tan

find_{ref,ofs}_delta_{,children} take an enum object_type parameter, but the object type is already present in the name of the function. Remove that parameter from these functions.

Interesting.

These "find children" are the only caller for each delta type and they always pass OFS/REF constant that corresponds to them, so find_*_delta() can lose the type. But compare_*_delta_bases() cannot, as the other one in the objects[] array may be of different type.

Makes sense. Thanks.

#6 lore raw

PATCH 3/7 index-pack: unify threaded and unthreaded code

In reply to Jonathan Tan
---
 builtin/index-pack.c | 10 +---------
 1 file changed, 1 insertion(+), 9 deletions(-)
builtin/index-pack.c +1 -9
index a8444d550b..357e03b5aa 100644
@@ -1211,15 +1211,7 @@ static void resolve_deltas(void)
 		cleanup_thread();
 		return;
 	}
-
-	for (i = 0; i < nr_objects; i++) {
-		struct object_entry *obj = &objects[i];
-
-		if (is_delta_type(obj->type))
-			continue;
-		resolve_base(obj);
-		display_progress(progress, nr_resolved_deltas);
-	}
+	threaded_second_pass(&nothread_data);
 }
 
 /*
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#7 lore raw

Re: [PATCH 3/7] index-pack: unify threaded and unthreaded code

Quoted 14 lines · Jonathan Tan
---
 builtin/index-pack.c | 10 +---------
 1 file changed, 1 insertion(+), 9 deletions(-)
builtin/index-pack.c +0 -1
index a8444d550b..357e03b5aa 100644
@@ -1211,15 +1211,7 @@ static void resolve_deltas(void)
 		cleanup_thread();
 		return;
 	}
-

builtin/index-pack.c static void resolve_deltas(void) · line 1215

Quoted 10 lines
builtin/index-pack.c +1 -8
-	for (i = 0; i < nr_objects; i++) {
-		struct object_entry *obj = &objects[i];
-
-		if (is_delta_type(obj->type))
-			continue;
-		resolve_base(obj);
-		display_progress(progress, nr_resolved_deltas);
-	}
+	threaded_second_pass(&nothread_data);
 }

The threaded_second_pass() function has a lot more noise than the above simple and easy-to-grok loop, managing the locks, but those all turn into no-op unless init_thread() is called, which happens only when this code is not reached. So OK, this looks essentially a no-op clean-up.

Thanks.

#8 lore raw

PATCH 4/7 index-pack: remove redundant child field

In reply to Jonathan Tan

This is refactoring 1 of 2 to simplify struct base_data.

In index-pack, each thread maintains a doubly-linked list of the delta chain that it is currently processing (the "base" and "child" pointers in struct base_data). When a thread exceeds the delta base cache limit and needs to reclaim memory, it uses the "child" pointers to traverse the lineage, reclaiming the memory of the eldest delta bases first.

A subsequent patch will perform memory reclaiming in a different way and will thus no longer need the "child" pointer. Because the "child" pointer is redundant even now, remove it so that the aforementioned subsequent patch will be clearer. In the meantime, reclaim memory in the reverse order of the "base" pointers.

---
 builtin/index-pack.c | 41 ++++++++++++++++++++++-------------------
 1 file changed, 22 insertions(+), 19 deletions(-)
builtin/index-pack.c +22 -19
index 357e03b5aa..032716553c 100644
@@ -34,7 +34,6 @@ struct object_stat {
 
 struct base_data {
 	struct base_data *base;
-	struct base_data *child;
 	struct object_entry *obj;
 	void *data;
 	unsigned long size;
@@ -44,7 +43,6 @@ struct base_data {
 
 struct thread_local {
 	pthread_t thread;
-	struct base_data *base_cache;
 	size_t base_cache_used;
 	int pack_fd;
 };
@@ -380,27 +378,37 @@ static void free_base_data(struct base_data *c)
 	}
 }
 
-static void prune_base_data(struct base_data *retain)
+static void prune_base_data(struct base_data *youngest_child)
 {
 	struct base_data *b;
 	struct thread_local *data = get_thread_data();
-	for (b = data->base_cache;
-	     data->base_cache_used > delta_base_cache_limit && b;
-	     b = b->child) {
-		if (b->data && b != retain)
-			free_base_data(b);
+	struct base_data **ancestry = NULL;
+	size_t nr = 0, alloc = 0;
+	ssize_t i;
+
+	if (data->base_cache_used <= delta_base_cache_limit)
+		return;
+
+	/*
+	 * Free all ancestors of youngest_child until we have enough space,
+	 * starting with the oldest. (We cannot free youngest_child itself.)
+	 */
+	for (b = youngest_child->base; b != NULL; b = b->base) {
+		ALLOC_GROW(ancestry, nr + 1, alloc);
+		ancestry[nr++] = b;
 	}
+	for (i = nr - 1;
+	     i >= 0 && data->base_cache_used > delta_base_cache_limit;
+	     i--) {
+		if (ancestry[i]->data)
+			free_base_data(ancestry[i]);
+	}
+	free(ancestry);
 }
 
 static void link_base_data(struct base_data *base, struct base_data *c)
 {
-	if (base)
-		base->child = c;
-	else
-		get_thread_data()->base_cache = c;
-
 	c->base = base;
-	c->child = NULL;
 	if (c->data)
 		get_thread_data()->base_cache_used += c->size;
 	prune_base_data(c);
@@ -408,11 +416,6 @@ static void link_base_data(struct base_data *base, struct base_data *c)
 
 static void unlink_base_data(struct base_data *c)
 {
-	struct base_data *base = c->base;
-	if (base)
-		base->child = NULL;
-	else
-		get_thread_data()->base_cache = NULL;
 	free_base_data(c);
 }
 
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#9 lore raw

PATCH 5/7 index-pack: calculate {ref,ofs}_{first,last} early

In reply to Jonathan Tan

This is refactoring 2 of 2 to simplify struct base_data.

Whenever we make a struct base_data, immediately calculate its delta children. This eliminates confusion as to when the {ref,ofs}_{first,last} fields are initialized.

Before this patch, the delta children were calculated at the last possible moment. This allowed the members of struct base_data to be populated in any order, superficially useful when we have the object contents before the struct object_entry. But this makes reasoning about the state of struct base_data more complicated, hence this patch.

---
 builtin/index-pack.c | 123 +++++++++++++++++++++----------------------
 1 file changed, 60 insertions(+), 63 deletions(-)
builtin/index-pack.c +60 -63
index 032716553c..e98b11ab37 100644
@@ -33,12 +33,15 @@ struct object_stat {
 };
 
 struct base_data {
+	/* Initialized by make_base(). */
 	struct base_data *base;
 	struct object_entry *obj;
-	void *data;
-	unsigned long size;
 	int ref_first, ref_last;
 	int ofs_first, ofs_last;
+
+	/* Not initialized by make_base(). */
+	void *data;
+	unsigned long size;
 };
 
 struct thread_local {
@@ -362,14 +365,6 @@ static void set_thread_data(struct thread_local *data)
 		pthread_setspecific(key, data);
 }
 
-static struct base_data *alloc_base_data(void)
-{
-	struct base_data *base = xcalloc(1, sizeof(struct base_data));
-	base->ref_last = -1;
-	base->ofs_last = -1;
-	return base;
-}
-
 static void free_base_data(struct base_data *c)
 {
 	if (c->data) {
@@ -406,19 +401,6 @@ static void prune_base_data(struct base_data *youngest_child)
 	free(ancestry);
 }
 
-static void link_base_data(struct base_data *base, struct base_data *c)
-{
-	c->base = base;
-	if (c->data)
-		get_thread_data()->base_cache_used += c->size;
-	prune_base_data(c);
-}
-
-static void unlink_base_data(struct base_data *c)
-{
-	free_base_data(c);
-}
-
 static int is_delta_type(enum object_type type)
 {
 	return (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA);
@@ -929,10 +911,25 @@ static void *get_base_data(struct base_data *c)
 	return c->data;
 }
 
-static void resolve_delta(struct object_entry *delta_obj,
-			  struct base_data *base, struct base_data *result)
+static struct base_data *make_base(struct object_entry *obj,
+				   struct base_data *parent)
 {
-	void *base_data, *delta_data;
+	struct base_data *base = xcalloc(1, sizeof(struct base_data));
+	base->base = parent;
+	base->obj = obj;
+	find_ref_delta_children(&obj->idx.oid,
+				&base->ref_first, &base->ref_last);
+	find_ofs_delta_children(obj->idx.offset,
+				&base->ofs_first, &base->ofs_last);
+	return base;
+}
+
+static struct base_data *resolve_delta(struct object_entry *delta_obj,
+				       struct base_data *base)
+{
+	void *base_data, *delta_data, *result_data;
+	struct base_data *result;
+	unsigned long result_size;
 
 	if (show_stat) {
 		int i = delta_obj - objects;
@@ -946,19 +943,31 @@ static void resolve_delta(struct object_entry *delta_obj,
 	}
 	delta_data = get_data_from_pack(delta_obj);
 	base_data = get_base_data(base);
-	result->obj = delta_obj;
-	result->data = patch_delta(base_data, base->size,
-				   delta_data, delta_obj->size, &result->size);
+	result_data = patch_delta(base_data, base->size,
+				  delta_data, delta_obj->size, &result_size);
 	free(delta_data);
-	if (!result->data)
+	if (!result_data)
 		bad_object(delta_obj->idx.offset, _("failed to apply delta"));
-	hash_object_file(the_hash_algo, result->data, result->size,
+	hash_object_file(the_hash_algo, result_data, result_size,
 			 type_name(delta_obj->real_type), &delta_obj->idx.oid);
-	sha1_object(result->data, NULL, result->size, delta_obj->real_type,
+	sha1_object(result_data, NULL, result_size, delta_obj->real_type,
 		    &delta_obj->idx.oid);
+
+	result = make_base(delta_obj, base);
+	if (result->ref_last == -1 && result->ofs_last == -1) {
+		free(result_data);
+	} else {
+		result->data = result_data;
+		result->size = result_size;
+		get_thread_data()->base_cache_used += result->size;
+		prune_base_data(result);
+	}
+
 	counter_lock();
 	nr_resolved_deltas++;
 	counter_unlock();
+
+	return result;
 }
 
 /*
@@ -984,24 +993,9 @@ static int compare_and_swap_type(signed char *type,
 static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 						  struct base_data *prev_base)
 {
-	if (base->ref_last == -1 && base->ofs_last == -1) {
-		find_ref_delta_children(&base->obj->idx.oid,
-					&base->ref_first, &base->ref_last);
-
-		find_ofs_delta_children(base->obj->idx.offset,
-					&base->ofs_first, &base->ofs_last);
-
-		if (base->ref_last == -1 && base->ofs_last == -1) {
-			free(base->data);
-			return NULL;
-		}
-
-		link_base_data(prev_base, base);
-	}
-
 	if (base->ref_first <= base->ref_last) {
 		struct object_entry *child = objects + ref_deltas[base->ref_first].obj_no;
-		struct base_data *result = alloc_base_data();
+		struct base_data *result;
 
 		if (!compare_and_swap_type(&child->real_type, OBJ_REF_DELTA,
 					   base->obj->real_type))
@@ -1009,7 +1003,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 			    (uintmax_t)child->idx.offset,
 			    oid_to_hex(&base->obj->idx.oid));
 
-		resolve_delta(child, base, result);
+		result = resolve_delta(child, base);
 		if (base->ref_first == base->ref_last && base->ofs_last == -1)
 			free_base_data(base);
 
@@ -1019,11 +1013,11 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 
 	if (base->ofs_first <= base->ofs_last) {
 		struct object_entry *child = objects + ofs_deltas[base->ofs_first].obj_no;
-		struct base_data *result = alloc_base_data();
+		struct base_data *result;
 
 		assert(child->real_type == OBJ_OFS_DELTA);
 		child->real_type = base->obj->real_type;
-		resolve_delta(child, base, result);
+		result = resolve_delta(child, base);
 		if (base->ofs_first == base->ofs_last)
 			free_base_data(base);
 
@@ -1031,7 +1025,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 		return result;
 	}
 
-	unlink_base_data(base);
+	free_base_data(base);
 	return NULL;
 }
 
@@ -1074,9 +1068,8 @@ static int compare_ref_delta_entry(const void *a, const void *b)
 
 static void resolve_base(struct object_entry *obj)
 {
-	struct base_data *base_obj = alloc_base_data();
-	base_obj->obj = obj;
-	base_obj->data = NULL;
+	struct base_data *base_obj = make_base(obj, NULL);
+
 	find_unresolved_deltas(base_obj);
 }
 
@@ -1369,22 +1362,26 @@ static void fix_unresolved_deltas(struct hashfile *f)
 	for (i = 0; i < nr_ref_deltas; i++) {
 		struct ref_delta_entry *d = sorted_by_pos[i];
 		enum object_type type;
-		struct base_data *base_obj = alloc_base_data();
+		struct base_data *base;
+		void *data;
+		unsigned long size;
+		struct object_entry *obj;
 
 		if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
 			continue;
-		base_obj->data = read_object_file(&d->oid, &type,
-						  &base_obj->size);
-		if (!base_obj->data)
+		data = read_object_file(&d->oid, &type, &size);
+		if (!data)
 			continue;
 
 		if (check_object_signature(the_repository, &d->oid,
-					   base_obj->data, base_obj->size,
+					   data, size,
 					   type_name(type)))
 			die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
-		base_obj->obj = append_obj_to_pack(f, d->oid.hash,
-					base_obj->data, base_obj->size, type);
-		find_unresolved_deltas(base_obj);
+		obj = append_obj_to_pack(f, d->oid.hash, data, size, type);
+		base = make_base(obj, NULL);
+		base->data = data;
+		base->size = size;
+		find_unresolved_deltas(base);
 		display_progress(progress, nr_resolved_deltas);
 	}
 	free(sorted_by_pos);
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#10 lore raw

PATCH 6/7 index-pack: make resolve_delta() assume base data

In reply to Jonathan Tan

A subsequent commit will make the quantum of work smaller, necessitating more locking. This commit allows resolve_delta() to be called outside the lock.

---
 builtin/index-pack.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
builtin/index-pack.c +5 -3
index e98b11ab37..c6d2acc13a 100644
@@ -927,7 +927,7 @@ static struct base_data *make_base(struct object_entry *obj,
 static struct base_data *resolve_delta(struct object_entry *delta_obj,
 				       struct base_data *base)
 {
-	void *base_data, *delta_data, *result_data;
+	void *delta_data, *result_data;
 	struct base_data *result;
 	unsigned long result_size;
 
@@ -942,8 +942,8 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 		obj_stat[i].base_object_no = j;
 	}
 	delta_data = get_data_from_pack(delta_obj);
-	base_data = get_base_data(base);
-	result_data = patch_delta(base_data, base->size,
+	assert(base->data);
+	result_data = patch_delta(base->data, base->size,
 				  delta_data, delta_obj->size, &result_size);
 	free(delta_data);
 	if (!result_data)
@@ -1003,6 +1003,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 			    (uintmax_t)child->idx.offset,
 			    oid_to_hex(&base->obj->idx.oid));
 
+		get_base_data(base);
 		result = resolve_delta(child, base);
 		if (base->ref_first == base->ref_last && base->ofs_last == -1)
 			free_base_data(base);
@@ -1017,6 +1018,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 
 		assert(child->real_type == OBJ_OFS_DELTA);
 		child->real_type = base->obj->real_type;
+		get_base_data(base);
 		result = resolve_delta(child, base);
 		if (base->ofs_first == base->ofs_last)
 			free_base_data(base);
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#11 lore raw

PATCH 7/7 index-pack: make quantum of work smaller

In reply to Jonathan Tan

Currently, when index-pack resolves deltas, it does not split up delta trees into threads: each delta base root (an object that is not a REF_DELTA or OFS_DELTA) can go into its own thread, but all deltas on that root (direct or indirect) are processed in the same thread.

This is a problem when a repository contains a large text file (thus, delta-able) that is modified many times - delta resolution time during fetching is dominated by processing the deltas corresponding to that text file.

This patch contains a solution to that. When cloning using

  git -c core.deltabasecachelimit=1g clone \
    https://fuchsia.googlesource.com/third_party/vulkan-cts

on my laptop, clone time improved from 3m2s to 2m5s (using 3 threads, which is the default).

The solution is to have a global work stack. This stack contains delta bases (objects, whether appearing directly in the packfile or generated by delta resolution, that themselves have delta children) that need to be processed; whenever a thread needs work, it peeks at the top of the stack and processes its next unprocessed child. If a thread finds the stack empty, it will look for more delta base roots to push on the stack instead.

The main weakness of having a global work stack is that more time is spent in the mutex, but profiling has shown that most time is spent in the resolution of the deltas themselves, so this shouldn't be an issue in practice. In any case, experimentation (as described in the clone command above) shows that this patch is a net improvement.

---
 builtin/index-pack.c | 341 ++++++++++++++++++++++++-------------------
 1 file changed, 193 insertions(+), 148 deletions(-)
builtin/index-pack.c +193 -148
index c6d2acc13a..0a5b938e1e 100644
@@ -38,15 +38,49 @@ struct base_data {
 	struct object_entry *obj;
 	int ref_first, ref_last;
 	int ofs_first, ofs_last;
+	/*
+	 * Threads should increment retain_data if they are about to call
+	 * patch_delta() using this struct's data as a base, and decrement this
+	 * when they are done. While retain_data is nonzero, this struct's data
+	 * will not be freed even if the delta base cache limit is exceeded.
+	 */
+	int retain_data;
+	/*
+	 * The number of direct children that have not been fully processed
+	 * (entered work_head, entered done_head, left done_head). When this
+	 * number reaches zero, this struct base_data can be freed.
+	 */
+	int children_remaining;
 
 	/* Not initialized by make_base(). */
+	struct list_head list;
 	void *data;
 	unsigned long size;
 };
 
+/*
+ * Stack of struct base_data that have unprocessed children.
+ * threaded_second_pass() uses this as a source of work (the other being the
+ * objects array).
+ */
+LIST_HEAD(work_head);
+
+/*
+ * Stack of struct base_data that have children, all of whom have been
+ * processed or are being processed, and at least one child is being processed.
+ * These struct base_data must be kept around until the last child is
+ * processed.
+ */
+LIST_HEAD(done_head);
+
+/*
+ * All threads share one delta base cache.
+ */
+size_t base_cache_used;
+size_t base_cache_limit;
+
 struct thread_local {
 	pthread_t thread;
-	size_t base_cache_used;
 	int pack_fd;
 };
 
@@ -369,36 +403,38 @@ static void free_base_data(struct base_data *c)
 {
 	if (c->data) {
 		FREE_AND_NULL(c->data);
-		get_thread_data()->base_cache_used -= c->size;
+		base_cache_used -= c->size;
 	}
 }
 
-static void prune_base_data(struct base_data *youngest_child)
+static void prune_base_data(struct base_data *retain)
 {
-	struct base_data *b;
-	struct thread_local *data = get_thread_data();
-	struct base_data **ancestry = NULL;
-	size_t nr = 0, alloc = 0;
-	ssize_t i;
+	struct list_head *pos;
 
-	if (data->base_cache_used <= delta_base_cache_limit)
+	if (base_cache_used <= base_cache_limit)
 		return;
 
-	/*
-	 * Free all ancestors of youngest_child until we have enough space,
-	 * starting with the oldest. (We cannot free youngest_child itself.)
-	 */
-	for (b = youngest_child->base; b != NULL; b = b->base) {
-		ALLOC_GROW(ancestry, nr + 1, alloc);
-		ancestry[nr++] = b;
+	list_for_each_prev(pos, &done_head) {
+		struct base_data *b = list_entry(pos, struct base_data, list);
+		if (b->retain_data || b == retain)
+			continue;
+		if (b->data) {
+			free_base_data(b);
+			if (base_cache_used <= base_cache_limit)
+				return;
+		}
 	}
-	for (i = nr - 1;
-	     i >= 0 && data->base_cache_used > delta_base_cache_limit;
-	     i--) {
-		if (ancestry[i]->data)
-			free_base_data(ancestry[i]);
+
+	list_for_each_prev(pos, &work_head) {
+		struct base_data *b = list_entry(pos, struct base_data, list);
+		if (b->retain_data || b == retain)
+			continue;
+		if (b->data) {
+			free_base_data(b);
+			if (base_cache_used <= base_cache_limit)
+				return;
+		}
 	}
-	free(ancestry);
 }
 
 static int is_delta_type(enum object_type type)
@@ -851,15 +887,7 @@ static void sha1_object(const void *data, struct object_entry *obj_entry,
 }
 
 /*
- * This function is part of find_unresolved_deltas(). There are two
- * walkers going in the opposite ways.
- *
- * The first one in find_unresolved_deltas() traverses down from
- * parent node to children, deflating nodes along the way. However,
- * memory for deflated nodes is limited by delta_base_cache_limit, so
- * at some point parent node's deflated content may be freed.
- *
- * The second walker is this function, which goes from current node up
+ * Walk from current node up
  * to top parent if necessary to deflate the node. In normal
  * situation, its parent node would be already deflated, so it just
  * needs to apply delta.
@@ -887,7 +915,7 @@ static void *get_base_data(struct base_data *c)
 		if (!delta_nr) {
 			c->data = get_data_from_pack(obj);
 			c->size = obj->size;
-			get_thread_data()->base_cache_used += c->size;
+			base_cache_used += c->size;
 			prune_base_data(c);
 		}
 		for (; delta_nr > 0; delta_nr--) {
@@ -903,7 +931,7 @@ static void *get_base_data(struct base_data *c)
 			free(raw);
 			if (!c->data)
 				bad_object(obj->idx.offset, _("failed to apply delta"));
-			get_thread_data()->base_cache_used += c->size;
+			base_cache_used += c->size;
 			prune_base_data(c);
 		}
 		free(delta);
@@ -921,6 +949,8 @@ static struct base_data *make_base(struct object_entry *obj,
 				&base->ref_first, &base->ref_last);
 	find_ofs_delta_children(obj->idx.offset,
 				&base->ofs_first, &base->ofs_last);
+	base->children_remaining = base->ref_last - base->ref_first +
+		base->ofs_last - base->ofs_first + 2;
 	return base;
 }
 
@@ -954,14 +984,8 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 		    &delta_obj->idx.oid);
 
 	result = make_base(delta_obj, base);
-	if (result->ref_last == -1 && result->ofs_last == -1) {
-		free(result_data);
-	} else {
-		result->data = result_data;
-		result->size = result_size;
-		get_thread_data()->base_cache_used += result->size;
-		prune_base_data(result);
-	}
+	result->data = result_data;
+	result->size = result_size;
 
 	counter_lock();
 	nr_resolved_deltas++;
@@ -970,86 +994,6 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 	return result;
 }
 
-/*
- * Standard boolean compare-and-swap: atomically check whether "*type" is
- * "want"; if so, swap in "set" and return true. Otherwise, leave it untouched
- * and return false.
- */
-static int compare_and_swap_type(signed char *type,
-				 enum object_type want,
-				 enum object_type set)
-{
-	enum object_type old;
-
-	type_cas_lock();
-	old = *type;
-	if (old == want)
-		*type = set;
-	type_cas_unlock();
-
-	return old == want;
-}
-
-static struct base_data *find_unresolved_deltas_1(struct base_data *base,
-						  struct base_data *prev_base)
-{
-	if (base->ref_first <= base->ref_last) {
-		struct object_entry *child = objects + ref_deltas[base->ref_first].obj_no;
-		struct base_data *result;
-
-		if (!compare_and_swap_type(&child->real_type, OBJ_REF_DELTA,
-					   base->obj->real_type))
-			die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
-			    (uintmax_t)child->idx.offset,
-			    oid_to_hex(&base->obj->idx.oid));
-
-		get_base_data(base);
-		result = resolve_delta(child, base);
-		if (base->ref_first == base->ref_last && base->ofs_last == -1)
-			free_base_data(base);
-
-		base->ref_first++;
-		return result;
-	}
-
-	if (base->ofs_first <= base->ofs_last) {
-		struct object_entry *child = objects + ofs_deltas[base->ofs_first].obj_no;
-		struct base_data *result;
-
-		assert(child->real_type == OBJ_OFS_DELTA);
-		child->real_type = base->obj->real_type;
-		get_base_data(base);
-		result = resolve_delta(child, base);
-		if (base->ofs_first == base->ofs_last)
-			free_base_data(base);
-
-		base->ofs_first++;
-		return result;
-	}
-
-	free_base_data(base);
-	return NULL;
-}
-
-static void find_unresolved_deltas(struct base_data *base)
-{
-	struct base_data *new_base, *prev_base = NULL;
-	for (;;) {
-		new_base = find_unresolved_deltas_1(base, prev_base);
-
-		if (new_base) {
-			prev_base = base;
-			base = new_base;
-		} else {
-			free(base);
-			base = prev_base;
-			if (!base)
-				return;
-			prev_base = base->base;
-		}
-	}
-}
-
 static int compare_ofs_delta_entry(const void *a, const void *b)
 {
 	const struct ofs_delta_entry *delta_a = a;
@@ -1068,33 +1012,131 @@ static int compare_ref_delta_entry(const void *a, const void *b)
 	return oidcmp(&delta_a->oid, &delta_b->oid);
 }
 
-static void resolve_base(struct object_entry *obj)
-{
-	struct base_data *base_obj = make_base(obj, NULL);
-
-	find_unresolved_deltas(base_obj);
-}
-
 static void *threaded_second_pass(void *data)
 {
-	set_thread_data(data);
+	if (data)
+		set_thread_data(data);
 	for (;;) {
-		int i;
-		counter_lock();
-		display_progress(progress, nr_resolved_deltas);
-		counter_unlock();
+		struct base_data *parent = NULL;
+		struct object_entry *child_obj;
+		struct base_data *child;
+
 		work_lock();
-		while (nr_dispatched < nr_objects &&
-		       is_delta_type(objects[nr_dispatched].type))
-			nr_dispatched++;
-		if (nr_dispatched >= nr_objects) {
-			work_unlock();
-			break;
+		if (list_empty(&work_head)) {
+			/*
+			 * Take an object from the object array.
+			 */
+			while (nr_dispatched < nr_objects &&
+			       is_delta_type(objects[nr_dispatched].type))
+				nr_dispatched++;
+			if (nr_dispatched >= nr_objects) {
+				work_unlock();
+				break;
+			}
+			child_obj = &objects[nr_dispatched++];
+		} else {
+			/*
+			 * Peek at the top of the stack, and take a child from
+			 * it.
+			 */
+			parent = list_first_entry(&work_head, struct base_data,
+						  list);
+
+			if (parent->ref_first <= parent->ref_last) {
+				int offset = ref_deltas[parent->ref_first++].obj_no;
+				child_obj = objects + offset;
+				if (child_obj->real_type != OBJ_REF_DELTA)
+					die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
+					    (uintmax_t) child_obj->idx.offset,
+					    oid_to_hex(&parent->obj->idx.oid));
+				child_obj->real_type = parent->obj->real_type;
+			} else {
+				child_obj = objects +
+					ofs_deltas[parent->ofs_first++].obj_no;
+				assert(child_obj->real_type == OBJ_OFS_DELTA);
+				child_obj->real_type = parent->obj->real_type;
+			}
+
+			if (parent->ref_first > parent->ref_last &&
+			    parent->ofs_first > parent->ofs_last) {
+				/*
+				 * This parent has run out of children, so move
+				 * it to done_head.
+				 */
+				list_del(&parent->list);
+				list_add(&parent->list, &done_head);
+			}
+
+			/*
+			 * Ensure that the parent has data, since we will need
+			 * it later.
+			 *
+			 * NEEDSWORK: If parent data needs to be reloaded, this
+			 * prolongs the time that the current thread spends in
+			 * the mutex. A mitigating factor is that parent data
+			 * needs to be reloaded only if the delta base cache
+			 * limit is exceeded, so in the typical case, this does
+			 * not happen.
+			 */
+			get_base_data(parent);
+			parent->retain_data++;
 		}
-		i = nr_dispatched++;
 		work_unlock();
 
-		resolve_base(&objects[i]);
+		if (parent) {
+			child = resolve_delta(child_obj, parent);
+			if (!child->children_remaining)
+				FREE_AND_NULL(child->data);
+		} else {
+			child = make_base(child_obj, NULL);
+			if (child->children_remaining) {
+				/*
+				 * Since this child has its own delta children,
+				 * we will need this data in the future.
+				 * Inflate now so that future iterations will
+				 * have access to this object's data while
+				 * outside the work mutex.
+				 */
+				child->data = get_data_from_pack(child_obj);
+				child->size = child_obj->size;
+			}
+		}
+
+		work_lock();
+		if (parent)
+			parent->retain_data--;
+		if (child->data) {
+			/*
+			 * This child has its own children, so add it to
+			 * work_head.
+			 */
+			list_add(&child->list, &work_head);
+			base_cache_used += child->size;
+			prune_base_data(NULL);
+		} else {
+			/*
+			 * This child does not have its own children. It may be
+			 * the last descendant of its ancestors; free those
+			 * that we can.
+			 */
+			struct base_data *p = parent;
+
+			while (p) {
+				struct base_data *next_p;
+
+				p->children_remaining--;
+				if (p->children_remaining)
+					break;
+
+				next_p = p->base;
+				free_base_data(p);
+				list_del(&p->list);
+				free(p);
+
+				p = next_p;
+			}
+		}
+		work_unlock();
 	}
 	return NULL;
 }
@@ -1195,6 +1237,7 @@ static void resolve_deltas(void)
 					  nr_ref_deltas + nr_ofs_deltas);
 
 	nr_dispatched = 0;
+	base_cache_limit = delta_base_cache_limit * nr_threads;
 	if (nr_threads > 1 || getenv("GIT_FORCE_THREADS")) {
 		init_thread();
 		for (i = 0; i < nr_threads; i++) {
@@ -1364,10 +1407,8 @@ static void fix_unresolved_deltas(struct hashfile *f)
 	for (i = 0; i < nr_ref_deltas; i++) {
 		struct ref_delta_entry *d = sorted_by_pos[i];
 		enum object_type type;
-		struct base_data *base;
 		void *data;
 		unsigned long size;
-		struct object_entry *obj;
 
 		if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
 			continue;
@@ -1379,11 +1420,15 @@ static void fix_unresolved_deltas(struct hashfile *f)
 					   data, size,
 					   type_name(type)))
 			die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
-		obj = append_obj_to_pack(f, d->oid.hash, data, size, type);
-		base = make_base(obj, NULL);
-		base->data = data;
-		base->size = size;
-		find_unresolved_deltas(base);
+
+		/*
+		 * Add this as an object to the objects array and call
+		 * threaded_second_pass() (which will pick up the added
+		 * object).
+		 */
+		append_obj_to_pack(f, d->oid.hash, data, size, type);
+		threaded_second_pass(NULL);
+
 		display_progress(progress, nr_resolved_deltas);
 	}
 	free(sorted_by_pos);
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#12 lore raw

Re: [PATCH 7/7] index-pack: make quantum of work smaller

Quoted 9 lines · Jonathan Tan

Currently, when index-pack resolves deltas, it does not split up delta trees into threads: each delta base root (an object that is not a REF_DELTA or OFS_DELTA) can go into its own thread, but all deltas on that root (direct or indirect) are processed in the same thread.

This is a problem when a repository contains a large text file (thus, delta-able) that is modified many times - delta resolution time during fetching is dominated by processing the deltas corresponding to that text file.

We favor wide and shallow delta trees for both reduced storage footprint and lower runtime overhead, so optimizing the index-pack for that use case makes a lot of sense.

#14 lore raw

PATCH 1/1 fixup! index-pack: make quantum of work smaller

In reply to Jonathan Tan

---

Quoted 3 lines

Josh did not list
out specific improvements to commit messages but I have taken his suggestions for comments.

...and somehow I forgot to commit them before sending out this patch set, so here they are.

 builtin/index-pack.c | 7 +++++++
 1 file changed, 7 insertions(+)
builtin/index-pack.c +7 -0
index 0a5b938e1e..4fa9127c35 100644
@@ -62,6 +62,8 @@ struct base_data {
  * Stack of struct base_data that have unprocessed children.
  * threaded_second_pass() uses this as a source of work (the other being the
  * objects array).
+ *
+ * Guarded by work_mutex.
  */
 LIST_HEAD(work_head);
 
@@ -70,11 +72,16 @@ LIST_HEAD(work_head);
  * processed or are being processed, and at least one child is being processed.
  * These struct base_data must be kept around until the last child is
  * processed.
+ *
+ * Guarded by work_mutex.
  */
 LIST_HEAD(done_head);
 
 /*
  * All threads share one delta base cache.
+ *
+ * base_cache_used is guarded by work_mutex, and base_cache_limit is read-only
+ * in a thread.
  */
 size_t base_cache_used;
 size_t base_cache_limit;
Signature · 1 line
2.28.0.297.g1956fa8f8d-goog
#15 lore raw
In reply to Jonathan Tan
Quoted 8 lines · Jonathan Tan

I'm trying to resurrect [1] and have rebased it to latest master (675a4aaf3b ("Ninth batch", 2020-08-19)).

Peff said [2] (of v1) that the overall direction seems reasonable and Josh Steadmon said [3] (of v2) that it looks mostly good except for possible improvements to commit messages and comments. Josh did not list out specific improvements to commit messages but I have taken his suggestions for comments.

I haven't looked closely yet, but since I was doing timings of index-pack recently[1], I wondered if this might change anything (spoiler: it doesn't really seem to).

Here's the result of p5302 with my extra tests on my 8-core (16 with hyperthreading) laptop against linux.git:

  5302.3: index-pack 0 threads                   266.66(263.85+2.71)
  5302.4: index-pack 1 threads                   275.06(272.11+2.85)
  5302.5: index-pack 2 threads                   159.49(285.44+3.51)
  5302.6: index-pack 4 threads                   102.54(318.86+4.30)
  5302.7: index-pack 8 threads                   75.60(391.39+6.56) 
  5302.8: index-pack 16 threads                  75.56(748.45+13.37)
  5302.9: index-pack default number of threads   75.01(389.33+6.59) 

So the conclusions from that other series remain pretty similar: nothing gets faster as we move past the number of actual cores. The penalty for doing so seems less than what I got before, though (though it might just be a fluke; it was something like 2s worse before your patches, and there's a bit of noise; the increased CPU time can be disregarded as the processors are throttled down when more are running).

The overall time seems to get slightly worse, though (HEAD~7 is before your patch, HEAD is with it):

Test HEAD~7 HEAD
--------------------------------------------------------------------------------------------
5302.9: index-pack default number of threads 71.96(376.11+3.66) 74.18(390.62+6.08) +3.1%

There may be other cases that get better, though. A 3% increase here is probably OK if we get something for it. But if our primary goal here is increasing multithread efficiency, then we should be able to show some benchmark that improves. :)

-Peff

[1] https://lore.kernel.org/git/[email protected]/

#16 lore raw
Quoted 11 lines

The overall time seems to get slightly worse, though (HEAD~7 is before your patch, HEAD is with it):

Test HEAD~7 HEAD
--------------------------------------------------------------------------------------------
5302.9: index-pack default number of threads 71.96(376.11+3.66) 74.18(390.62+6.08) +3.1%

There may be other cases that get better, though. A 3% increase here is probably OK if we get something for it. But if our primary goal here is increasing multithread efficiency, then we should be able to show some benchmark that improves. :)

Ah...good question. Cloning from
https://fuchsia.googlesource.com/third_party/vulkan-cts (mentioned in patch 7), cd-ing to the pack dir, and running:

  git index-pack --stdin -o foo <*.pack

I got 8m2.878s with my patches and 12m6.365s without. But I ran this on a cloud virtual machine (what I have access to right now) so the numbers might look different on a dedicated machine.

#17 lore raw
Quoted 9 lines · Jonathan Tan
Quoted 4 lines

There may be other cases that get better, though. A 3% increase here is probably OK if we get something for it. But if our primary goal here is increasing multithread efficiency, then we should be able to show some benchmark that improves. :)

Ah...good question. Cloning from
https://fuchsia.googlesource.com/third_party/vulkan-cts (mentioned in patch 7), cd-ing to the pack dir, and running:

Quoted 5 lines
  git index-pack --stdin -o foo <*.pack

I got 8m2.878s with my patches and 12m6.365s without. But I ran this on a cloud virtual machine (what I have access to right now) so the numbers might look different on a dedicated machine.

Thanks, that's a much more interesting example. Here's what I get on my 8-core machine:

5302.9: index-pack default number of threads 167.70(546.19+12.00) 83.69(585.61+6.95) -50.1%

So that's a considerable improvement. And hardly surprising given the repository structure. I used the script below to show the size of the delta families, and the vk-master ones really dominate in size and object number (the biggest is 50GB in one delta family).

I also ran my PERF_EXTRA tests on them to see if it behaved differently as the threads increased. Nope:

  5302.3: index-pack 0 threads                   434.13(425.90+8.16)
  5302.4: index-pack 1 threads                   428.65(421.82+6.77)
  5302.5: index-pack 2 threads                   224.05(424.13+6.21)
  5302.6: index-pack 4 threads                   125.43(457.68+5.77)
  5302.7: index-pack 8 threads                   82.60(579.10+7.78) 
  5302.8: index-pack 16 threads                  82.89(1147.82+9.66)
  5302.9: index-pack default number of threads   83.91(576.92+8.52) 

Still maxes out at the number of physical cores (not unexpected, but that was the thing I was most curious about ;) ). I may run it on the 40-core machine, too. It's possible that with the new threading we're able to do better going past 20-threads. I doubt it, because I think it's mostly a function of Git's locking granularity, but worth checking.

-Peff

-- >8 --
#!/bin/sh

# script to output size, count, and filenames for each delta family

git rev-list --objects --all |
git cat-file --buffer \
--batch-check='%(objectname) %(deltabase) %(objectsize) %(rest)' |
perl -alne '

  if ($F[1] =~ /[^0]/) {
    push @{$children{$F[1]}}, $F[0];
  } else {
    push @bases, $F[0];
  }
  $size{$F[0]} = $F[2];
  $name{$F[0]} = $F[3];
  END {
    sub add_to_component {
      my ($oid, $data) = @_;
      $data->{names}->{$name{$oid}}++;
      $data->{size} += $size{$oid};
      $data->{nr}++;
      add_to_component($_, $data) for @{$children{$oid}};
    }
    for my $b (@bases) {
      my $data = { size => 0, nr => 0, names => {} };
      add_to_component($b, $data);
      print join(" ",
                 $data->{size}, $data->{nr},
		 sort keys(%{$data->{names}})
            ), "\n";
    }

}
' |
sort -rn

#18 lore raw
Quoted 5 lines · Jeff King

Still maxes out at the number of physical cores (not unexpected, but that was the thing I was most curious about ;) ). I may run it on the 40-core machine, too. It's possible that with the new threading we're able to do better going past 20-threads. I doubt it, because I think it's mostly a function of Git's locking granularity, but worth checking.

For the curious (I know you were all on the edge of your seat):

  5302.3: index-pack 0 threads                    519.23(479.10+40.10) 
  5302.4: index-pack 1 threads                    525.94(476.27+49.64) 
  5302.5: index-pack 2 threads                    271.82(458.93+55.52) 
  5302.6: index-pack 5 threads                    115.88(463.84+50.69) 
  5302.7: index-pack 10 threads                   67.26(478.37+57.38)  
  5302.8: index-pack 20 threads                   43.02(524.01+77.33)  
  5302.9: index-pack 40 threads                   33.42(709.86+100.24) 
  5302.10: index-pack 80 threads                  32.02(1030.75+228.28)
  5302.11: index-pack default number of threads   43.58(551.13+68.92)  

So it actually does do a slight improvement to go from 20 to 40 threads on this repository/machine combo. Not enough to make me revise the code I sent the other day, though. And we still get nothing from going past the number of physical cores.

-Peff

#19 lore raw

v2 cover Better threaded delta resolution in index-pack (another try)

In reply to Jonathan Tan

From What's Cooking [1]:

Quoted 11 lines

* jt/threaded-index-pack (2020-08-27) 9 commits
- builtin/index-pack.c: fix some sparse warnings
- fixup! index-pack: make quantum of work smaller
- index-pack: make quantum of work smaller
- index-pack: make resolve_delta() assume base data
- index-pack: calculate {ref,ofs}_{first,last} early
- index-pack: remove redundant child field
- index-pack: unify threaded and unthreaded code
- index-pack: remove redundant parameter
- Documentation: deltaBaseCacheLimit is per-thread

Quoted 5 lines

"git index-pack" learned to resolve deltified objects with greater parallelism.

Expecting the final reroll.
cf. https://colabti.org/irclogger/irclogger_log/git-devel?date=2020-08-31#l82

Here's the reroll.

[1] https://lore.kernel.org/git/[email protected]/

Jonathan Tan (7):
Documentation: deltaBaseCacheLimit is per-thread
index-pack: remove redundant parameter
index-pack: unify threaded and unthreaded code
index-pack: remove redundant child field
index-pack: calculate {ref,ofs}_{first,last} early
index-pack: make resolve_delta() assume base data
index-pack: make quantum of work smaller

 Documentation/config/core.txt |   2 +-
 builtin/index-pack.c          | 456 +++++++++++++++++++---------------
 2 files changed, 251 insertions(+), 207 deletions(-)
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#20 lore raw

PATCH v2 1/7 Documentation: deltaBaseCacheLimit is per-thread

Clarify that core.deltaBaseCacheLimit is per-thread, as can be seen from the fact that cache usage (base_cache_used in struct thread_local in builtin/index-pack.c) is tracked individually for each thread and compared against delta_base_cache_limit.

---
 Documentation/config/core.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Documentation/config/core.txt +1 -1
index 74619a9c03..02002cf109 100644
@@ -399,7 +399,7 @@ the largest projects.  You probably do not need to adjust this value.
 Common unit suffixes of 'k', 'm', or 'g' are supported.
 
 core.deltaBaseCacheLimit::
-	Maximum number of bytes to reserve for caching base objects
+	Maximum number of bytes per thread to reserve for caching base objects
 	that may be referenced by multiple deltified objects.  By storing the
 	entire decompressed base objects in a cache Git is able
 	to avoid unpacking and decompressing frequently used base
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#21 lore raw

PATCH v2 2/7 index-pack: remove redundant parameter

In reply to Jonathan Tan

find_{ref,ofs}_delta_{,children} take an enum object_type parameter, but the object type is already present in the name of the function. Remove that parameter from these functions.

---
 builtin/index-pack.c | 26 ++++++++++++--------------
 1 file changed, 12 insertions(+), 14 deletions(-)
builtin/index-pack.c +12 -14
index 9721bf1ffe..0e0889afc4 100644
@@ -614,7 +614,7 @@ static int compare_ofs_delta_bases(off_t offset1, off_t offset2,
 	       0;
 }
 
-static int find_ofs_delta(const off_t offset, enum object_type type)
+static int find_ofs_delta(const off_t offset)
 {
 	int first = 0, last = nr_ofs_deltas;
 
@@ -624,7 +624,8 @@ static int find_ofs_delta(const off_t offset, enum object_type type)
 		int cmp;
 
 		cmp = compare_ofs_delta_bases(offset, delta->offset,
-					      type, objects[delta->obj_no].type);
+					      OBJ_OFS_DELTA,
+					      objects[delta->obj_no].type);
 		if (!cmp)
 			return next;
 		if (cmp < 0) {
@@ -637,10 +638,9 @@ static int find_ofs_delta(const off_t offset, enum object_type type)
 }
 
 static void find_ofs_delta_children(off_t offset,
-				    int *first_index, int *last_index,
-				    enum object_type type)
+				    int *first_index, int *last_index)
 {
-	int first = find_ofs_delta(offset, type);
+	int first = find_ofs_delta(offset);
 	int last = first;
 	int end = nr_ofs_deltas - 1;
 
@@ -668,7 +668,7 @@ static int compare_ref_delta_bases(const struct object_id *oid1,
 	return oidcmp(oid1, oid2);
 }
 
-static int find_ref_delta(const struct object_id *oid, enum object_type type)
+static int find_ref_delta(const struct object_id *oid)
 {
 	int first = 0, last = nr_ref_deltas;
 
@@ -678,7 +678,8 @@ static int find_ref_delta(const struct object_id *oid, enum object_type type)
 		int cmp;
 
 		cmp = compare_ref_delta_bases(oid, &delta->oid,
-					      type, objects[delta->obj_no].type);
+					      OBJ_REF_DELTA,
+					      objects[delta->obj_no].type);
 		if (!cmp)
 			return next;
 		if (cmp < 0) {
@@ -691,10 +692,9 @@ static int find_ref_delta(const struct object_id *oid, enum object_type type)
 }
 
 static void find_ref_delta_children(const struct object_id *oid,
-				    int *first_index, int *last_index,
-				    enum object_type type)
+				    int *first_index, int *last_index)
 {
-	int first = find_ref_delta(oid, type);
+	int first = find_ref_delta(oid);
 	int last = first;
 	int end = nr_ref_deltas - 1;
 
@@ -983,12 +983,10 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 {
 	if (base->ref_last == -1 && base->ofs_last == -1) {
 		find_ref_delta_children(&base->obj->idx.oid,
-					&base->ref_first, &base->ref_last,
-					OBJ_REF_DELTA);
+					&base->ref_first, &base->ref_last);
 
 		find_ofs_delta_children(base->obj->idx.offset,
-					&base->ofs_first, &base->ofs_last,
-					OBJ_OFS_DELTA);
+					&base->ofs_first, &base->ofs_last);
 
 		if (base->ref_last == -1 && base->ofs_last == -1) {
 			free(base->data);
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#22 lore raw

PATCH v2 3/7 index-pack: unify threaded and unthreaded code

In reply to Jonathan Tan
---
 builtin/index-pack.c | 10 +---------
 1 file changed, 1 insertion(+), 9 deletions(-)
builtin/index-pack.c +1 -9
index 0e0889afc4..c7b4aef4e4 100644
@@ -1211,15 +1211,7 @@ static void resolve_deltas(void)
 		cleanup_thread();
 		return;
 	}
-
-	for (i = 0; i < nr_objects; i++) {
-		struct object_entry *obj = &objects[i];
-
-		if (is_delta_type(obj->type))
-			continue;
-		resolve_base(obj);
-		display_progress(progress, nr_resolved_deltas);
-	}
+	threaded_second_pass(&nothread_data);
 }
 
 /*
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#23 lore raw

PATCH v2 4/7 index-pack: remove redundant child field

In reply to Jonathan Tan

This is refactoring 1 of 2 to simplify struct base_data.

In index-pack, each thread maintains a doubly-linked list of the delta chain that it is currently processing (the "base" and "child" pointers in struct base_data). When a thread exceeds the delta base cache limit and needs to reclaim memory, it uses the "child" pointers to traverse the lineage, reclaiming the memory of the eldest delta bases first.

A subsequent patch will perform memory reclaiming in a different way and will thus no longer need the "child" pointer. Because the "child" pointer is redundant even now, remove it so that the aforementioned subsequent patch will be clearer. In the meantime, reclaim memory in the reverse order of the "base" pointers.

---
 builtin/index-pack.c | 41 ++++++++++++++++++++++-------------------
 1 file changed, 22 insertions(+), 19 deletions(-)
builtin/index-pack.c +22 -19
index c7b4aef4e4..c8db464557 100644
@@ -34,7 +34,6 @@ struct object_stat {
 
 struct base_data {
 	struct base_data *base;
-	struct base_data *child;
 	struct object_entry *obj;
 	void *data;
 	unsigned long size;
@@ -44,7 +43,6 @@ struct base_data {
 
 struct thread_local {
 	pthread_t thread;
-	struct base_data *base_cache;
 	size_t base_cache_used;
 	int pack_fd;
 };
@@ -380,27 +378,37 @@ static void free_base_data(struct base_data *c)
 	}
 }
 
-static void prune_base_data(struct base_data *retain)
+static void prune_base_data(struct base_data *youngest_child)
 {
 	struct base_data *b;
 	struct thread_local *data = get_thread_data();
-	for (b = data->base_cache;
-	     data->base_cache_used > delta_base_cache_limit && b;
-	     b = b->child) {
-		if (b->data && b != retain)
-			free_base_data(b);
+	struct base_data **ancestry = NULL;
+	size_t nr = 0, alloc = 0;
+	ssize_t i;
+
+	if (data->base_cache_used <= delta_base_cache_limit)
+		return;
+
+	/*
+	 * Free all ancestors of youngest_child until we have enough space,
+	 * starting with the oldest. (We cannot free youngest_child itself.)
+	 */
+	for (b = youngest_child->base; b != NULL; b = b->base) {
+		ALLOC_GROW(ancestry, nr + 1, alloc);
+		ancestry[nr++] = b;
 	}
+	for (i = nr - 1;
+	     i >= 0 && data->base_cache_used > delta_base_cache_limit;
+	     i--) {
+		if (ancestry[i]->data)
+			free_base_data(ancestry[i]);
+	}
+	free(ancestry);
 }
 
 static void link_base_data(struct base_data *base, struct base_data *c)
 {
-	if (base)
-		base->child = c;
-	else
-		get_thread_data()->base_cache = c;
-
 	c->base = base;
-	c->child = NULL;
 	if (c->data)
 		get_thread_data()->base_cache_used += c->size;
 	prune_base_data(c);
@@ -408,11 +416,6 @@ static void link_base_data(struct base_data *base, struct base_data *c)
 
 static void unlink_base_data(struct base_data *c)
 {
-	struct base_data *base = c->base;
-	if (base)
-		base->child = NULL;
-	else
-		get_thread_data()->base_cache = NULL;
 	free_base_data(c);
 }
 
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#24 lore raw

PATCH v2 5/7 index-pack: calculate {ref,ofs}_{first,last} early

In reply to Jonathan Tan

This is refactoring 2 of 2 to simplify struct base_data.

Whenever we make a struct base_data, immediately calculate its delta children. This eliminates confusion as to when the {ref,ofs}_{first,last} fields are initialized.

Before this patch, the delta children were calculated at the last possible moment. This allowed the members of struct base_data to be populated in any order, superficially useful when we have the object contents before the struct object_entry. But this makes reasoning about the state of struct base_data more complicated, hence this patch.

---
 builtin/index-pack.c | 123 +++++++++++++++++++++----------------------
 1 file changed, 60 insertions(+), 63 deletions(-)
builtin/index-pack.c +60 -63
index c8db464557..94d0f53b03 100644
@@ -33,12 +33,15 @@ struct object_stat {
 };
 
 struct base_data {
+	/* Initialized by make_base(). */
 	struct base_data *base;
 	struct object_entry *obj;
-	void *data;
-	unsigned long size;
 	int ref_first, ref_last;
 	int ofs_first, ofs_last;
+
+	/* Not initialized by make_base(). */
+	void *data;
+	unsigned long size;
 };
 
 struct thread_local {
@@ -362,14 +365,6 @@ static void set_thread_data(struct thread_local *data)
 		pthread_setspecific(key, data);
 }
 
-static struct base_data *alloc_base_data(void)
-{
-	struct base_data *base = xcalloc(1, sizeof(struct base_data));
-	base->ref_last = -1;
-	base->ofs_last = -1;
-	return base;
-}
-
 static void free_base_data(struct base_data *c)
 {
 	if (c->data) {
@@ -406,19 +401,6 @@ static void prune_base_data(struct base_data *youngest_child)
 	free(ancestry);
 }
 
-static void link_base_data(struct base_data *base, struct base_data *c)
-{
-	c->base = base;
-	if (c->data)
-		get_thread_data()->base_cache_used += c->size;
-	prune_base_data(c);
-}
-
-static void unlink_base_data(struct base_data *c)
-{
-	free_base_data(c);
-}
-
 static int is_delta_type(enum object_type type)
 {
 	return (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA);
@@ -929,10 +911,25 @@ static void *get_base_data(struct base_data *c)
 	return c->data;
 }
 
-static void resolve_delta(struct object_entry *delta_obj,
-			  struct base_data *base, struct base_data *result)
+static struct base_data *make_base(struct object_entry *obj,
+				   struct base_data *parent)
 {
-	void *base_data, *delta_data;
+	struct base_data *base = xcalloc(1, sizeof(struct base_data));
+	base->base = parent;
+	base->obj = obj;
+	find_ref_delta_children(&obj->idx.oid,
+				&base->ref_first, &base->ref_last);
+	find_ofs_delta_children(obj->idx.offset,
+				&base->ofs_first, &base->ofs_last);
+	return base;
+}
+
+static struct base_data *resolve_delta(struct object_entry *delta_obj,
+				       struct base_data *base)
+{
+	void *base_data, *delta_data, *result_data;
+	struct base_data *result;
+	unsigned long result_size;
 
 	if (show_stat) {
 		int i = delta_obj - objects;
@@ -946,19 +943,31 @@ static void resolve_delta(struct object_entry *delta_obj,
 	}
 	delta_data = get_data_from_pack(delta_obj);
 	base_data = get_base_data(base);
-	result->obj = delta_obj;
-	result->data = patch_delta(base_data, base->size,
-				   delta_data, delta_obj->size, &result->size);
+	result_data = patch_delta(base_data, base->size,
+				  delta_data, delta_obj->size, &result_size);
 	free(delta_data);
-	if (!result->data)
+	if (!result_data)
 		bad_object(delta_obj->idx.offset, _("failed to apply delta"));
-	hash_object_file(the_hash_algo, result->data, result->size,
+	hash_object_file(the_hash_algo, result_data, result_size,
 			 type_name(delta_obj->real_type), &delta_obj->idx.oid);
-	sha1_object(result->data, NULL, result->size, delta_obj->real_type,
+	sha1_object(result_data, NULL, result_size, delta_obj->real_type,
 		    &delta_obj->idx.oid);
+
+	result = make_base(delta_obj, base);
+	if (result->ref_last == -1 && result->ofs_last == -1) {
+		free(result_data);
+	} else {
+		result->data = result_data;
+		result->size = result_size;
+		get_thread_data()->base_cache_used += result->size;
+		prune_base_data(result);
+	}
+
 	counter_lock();
 	nr_resolved_deltas++;
 	counter_unlock();
+
+	return result;
 }
 
 /*
@@ -984,24 +993,9 @@ static int compare_and_swap_type(signed char *type,
 static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 						  struct base_data *prev_base)
 {
-	if (base->ref_last == -1 && base->ofs_last == -1) {
-		find_ref_delta_children(&base->obj->idx.oid,
-					&base->ref_first, &base->ref_last);
-
-		find_ofs_delta_children(base->obj->idx.offset,
-					&base->ofs_first, &base->ofs_last);
-
-		if (base->ref_last == -1 && base->ofs_last == -1) {
-			free(base->data);
-			return NULL;
-		}
-
-		link_base_data(prev_base, base);
-	}
-
 	if (base->ref_first <= base->ref_last) {
 		struct object_entry *child = objects + ref_deltas[base->ref_first].obj_no;
-		struct base_data *result = alloc_base_data();
+		struct base_data *result;
 
 		if (!compare_and_swap_type(&child->real_type, OBJ_REF_DELTA,
 					   base->obj->real_type))
@@ -1009,7 +1003,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 			    (uintmax_t)child->idx.offset,
 			    oid_to_hex(&base->obj->idx.oid));
 
-		resolve_delta(child, base, result);
+		result = resolve_delta(child, base);
 		if (base->ref_first == base->ref_last && base->ofs_last == -1)
 			free_base_data(base);
 
@@ -1019,11 +1013,11 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 
 	if (base->ofs_first <= base->ofs_last) {
 		struct object_entry *child = objects + ofs_deltas[base->ofs_first].obj_no;
-		struct base_data *result = alloc_base_data();
+		struct base_data *result;
 
 		assert(child->real_type == OBJ_OFS_DELTA);
 		child->real_type = base->obj->real_type;
-		resolve_delta(child, base, result);
+		result = resolve_delta(child, base);
 		if (base->ofs_first == base->ofs_last)
 			free_base_data(base);
 
@@ -1031,7 +1025,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 		return result;
 	}
 
-	unlink_base_data(base);
+	free_base_data(base);
 	return NULL;
 }
 
@@ -1074,9 +1068,8 @@ static int compare_ref_delta_entry(const void *a, const void *b)
 
 static void resolve_base(struct object_entry *obj)
 {
-	struct base_data *base_obj = alloc_base_data();
-	base_obj->obj = obj;
-	base_obj->data = NULL;
+	struct base_data *base_obj = make_base(obj, NULL);
+
 	find_unresolved_deltas(base_obj);
 }
 
@@ -1369,22 +1362,26 @@ static void fix_unresolved_deltas(struct hashfile *f)
 	for (i = 0; i < nr_ref_deltas; i++) {
 		struct ref_delta_entry *d = sorted_by_pos[i];
 		enum object_type type;
-		struct base_data *base_obj = alloc_base_data();
+		struct base_data *base;
+		void *data;
+		unsigned long size;
+		struct object_entry *obj;
 
 		if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
 			continue;
-		base_obj->data = read_object_file(&d->oid, &type,
-						  &base_obj->size);
-		if (!base_obj->data)
+		data = read_object_file(&d->oid, &type, &size);
+		if (!data)
 			continue;
 
 		if (check_object_signature(the_repository, &d->oid,
-					   base_obj->data, base_obj->size,
+					   data, size,
 					   type_name(type)))
 			die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
-		base_obj->obj = append_obj_to_pack(f, d->oid.hash,
-					base_obj->data, base_obj->size, type);
-		find_unresolved_deltas(base_obj);
+		obj = append_obj_to_pack(f, d->oid.hash, data, size, type);
+		base = make_base(obj, NULL);
+		base->data = data;
+		base->size = size;
+		find_unresolved_deltas(base);
 		display_progress(progress, nr_resolved_deltas);
 	}
 	free(sorted_by_pos);
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#25 lore raw

PATCH v2 6/7 index-pack: make resolve_delta() assume base data

In reply to Jonathan Tan

A subsequent commit will make the quantum of work smaller, necessitating more locking. This commit allows resolve_delta() to be called outside the lock.

---
 builtin/index-pack.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
builtin/index-pack.c +5 -3
index 94d0f53b03..f69a50d46b 100644
@@ -927,7 +927,7 @@ static struct base_data *make_base(struct object_entry *obj,
 static struct base_data *resolve_delta(struct object_entry *delta_obj,
 				       struct base_data *base)
 {
-	void *base_data, *delta_data, *result_data;
+	void *delta_data, *result_data;
 	struct base_data *result;
 	unsigned long result_size;
 
@@ -942,8 +942,8 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 		obj_stat[i].base_object_no = j;
 	}
 	delta_data = get_data_from_pack(delta_obj);
-	base_data = get_base_data(base);
-	result_data = patch_delta(base_data, base->size,
+	assert(base->data);
+	result_data = patch_delta(base->data, base->size,
 				  delta_data, delta_obj->size, &result_size);
 	free(delta_data);
 	if (!result_data)
@@ -1003,6 +1003,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 			    (uintmax_t)child->idx.offset,
 			    oid_to_hex(&base->obj->idx.oid));
 
+		get_base_data(base);
 		result = resolve_delta(child, base);
 		if (base->ref_first == base->ref_last && base->ofs_last == -1)
 			free_base_data(base);
@@ -1017,6 +1018,7 @@ static struct base_data *find_unresolved_deltas_1(struct base_data *base,
 
 		assert(child->real_type == OBJ_OFS_DELTA);
 		child->real_type = base->obj->real_type;
+		get_base_data(base);
 		result = resolve_delta(child, base);
 		if (base->ofs_first == base->ofs_last)
 			free_base_data(base);
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#26 lore raw

PATCH v2 7/7 index-pack: make quantum of work smaller

In reply to Jonathan Tan

Currently, when index-pack resolves deltas, it does not split up delta trees into threads: each delta base root (an object that is not a REF_DELTA or OFS_DELTA) can go into its own thread, but all deltas on that root (direct or indirect) are processed in the same thread.

This is a problem when a repository contains a large text file (thus, delta-able) that is modified many times - delta resolution time during fetching is dominated by processing the deltas corresponding to that text file.

This patch contains a solution to that. When cloning using

  git -c core.deltabasecachelimit=1g clone \
    https://fuchsia.googlesource.com/third_party/vulkan-cts

on my laptop, clone time improved from 3m2s to 2m5s (using 3 threads, which is the default).

The solution is to have a global work stack. This stack contains delta bases (objects, whether appearing directly in the packfile or generated by delta resolution, that themselves have delta children) that need to be processed; whenever a thread needs work, it peeks at the top of the stack and processes its next unprocessed child. If a thread finds the stack empty, it will look for more delta base roots to push on the stack instead.

The main weakness of having a global work stack is that more time is spent in the mutex, but profiling has shown that most time is spent in the resolution of the deltas themselves, so this shouldn't be an issue in practice. In any case, experimentation (as described in the clone command above) shows that this patch is a net improvement.

---
 builtin/index-pack.c | 348 +++++++++++++++++++++++++------------------
 1 file changed, 200 insertions(+), 148 deletions(-)
builtin/index-pack.c +200 -148
index f69a50d46b..8acd078aa0 100644
@@ -38,15 +38,56 @@ struct base_data {
 	struct object_entry *obj;
 	int ref_first, ref_last;
 	int ofs_first, ofs_last;
+	/*
+	 * Threads should increment retain_data if they are about to call
+	 * patch_delta() using this struct's data as a base, and decrement this
+	 * when they are done. While retain_data is nonzero, this struct's data
+	 * will not be freed even if the delta base cache limit is exceeded.
+	 */
+	int retain_data;
+	/*
+	 * The number of direct children that have not been fully processed
+	 * (entered work_head, entered done_head, left done_head). When this
+	 * number reaches zero, this struct base_data can be freed.
+	 */
+	int children_remaining;
 
 	/* Not initialized by make_base(). */
+	struct list_head list;
 	void *data;
 	unsigned long size;
 };
 
+/*
+ * Stack of struct base_data that have unprocessed children.
+ * threaded_second_pass() uses this as a source of work (the other being the
+ * objects array).
+ *
+ * Guarded by work_mutex.
+ */
+static LIST_HEAD(work_head);
+
+/*
+ * Stack of struct base_data that have children, all of whom have been
+ * processed or are being processed, and at least one child is being processed.
+ * These struct base_data must be kept around until the last child is
+ * processed.
+ *
+ * Guarded by work_mutex.
+ */
+static LIST_HEAD(done_head);
+
+/*
+ * All threads share one delta base cache.
+ *
+ * base_cache_used is guarded by work_mutex, and base_cache_limit is read-only
+ * in a thread.
+ */
+static size_t base_cache_used;
+static size_t base_cache_limit;
+
 struct thread_local {
 	pthread_t thread;
-	size_t base_cache_used;
 	int pack_fd;
 };
 
@@ -369,36 +410,38 @@ static void free_base_data(struct base_data *c)
 {
 	if (c->data) {
 		FREE_AND_NULL(c->data);
-		get_thread_data()->base_cache_used -= c->size;
+		base_cache_used -= c->size;
 	}
 }
 
-static void prune_base_data(struct base_data *youngest_child)
+static void prune_base_data(struct base_data *retain)
 {
-	struct base_data *b;
-	struct thread_local *data = get_thread_data();
-	struct base_data **ancestry = NULL;
-	size_t nr = 0, alloc = 0;
-	ssize_t i;
+	struct list_head *pos;
 
-	if (data->base_cache_used <= delta_base_cache_limit)
+	if (base_cache_used <= base_cache_limit)
 		return;
 
-	/*
-	 * Free all ancestors of youngest_child until we have enough space,
-	 * starting with the oldest. (We cannot free youngest_child itself.)
-	 */
-	for (b = youngest_child->base; b != NULL; b = b->base) {
-		ALLOC_GROW(ancestry, nr + 1, alloc);
-		ancestry[nr++] = b;
+	list_for_each_prev(pos, &done_head) {
+		struct base_data *b = list_entry(pos, struct base_data, list);
+		if (b->retain_data || b == retain)
+			continue;
+		if (b->data) {
+			free_base_data(b);
+			if (base_cache_used <= base_cache_limit)
+				return;
+		}
 	}
-	for (i = nr - 1;
-	     i >= 0 && data->base_cache_used > delta_base_cache_limit;
-	     i--) {
-		if (ancestry[i]->data)
-			free_base_data(ancestry[i]);
+
+	list_for_each_prev(pos, &work_head) {
+		struct base_data *b = list_entry(pos, struct base_data, list);
+		if (b->retain_data || b == retain)
+			continue;
+		if (b->data) {
+			free_base_data(b);
+			if (base_cache_used <= base_cache_limit)
+				return;
+		}
 	}
-	free(ancestry);
 }
 
 static int is_delta_type(enum object_type type)
@@ -851,15 +894,7 @@ static void sha1_object(const void *data, struct object_entry *obj_entry,
 }
 
 /*
- * This function is part of find_unresolved_deltas(). There are two
- * walkers going in the opposite ways.
- *
- * The first one in find_unresolved_deltas() traverses down from
- * parent node to children, deflating nodes along the way. However,
- * memory for deflated nodes is limited by delta_base_cache_limit, so
- * at some point parent node's deflated content may be freed.
- *
- * The second walker is this function, which goes from current node up
+ * Walk from current node up
  * to top parent if necessary to deflate the node. In normal
  * situation, its parent node would be already deflated, so it just
  * needs to apply delta.
@@ -887,7 +922,7 @@ static void *get_base_data(struct base_data *c)
 		if (!delta_nr) {
 			c->data = get_data_from_pack(obj);
 			c->size = obj->size;
-			get_thread_data()->base_cache_used += c->size;
+			base_cache_used += c->size;
 			prune_base_data(c);
 		}
 		for (; delta_nr > 0; delta_nr--) {
@@ -903,7 +938,7 @@ static void *get_base_data(struct base_data *c)
 			free(raw);
 			if (!c->data)
 				bad_object(obj->idx.offset, _("failed to apply delta"));
-			get_thread_data()->base_cache_used += c->size;
+			base_cache_used += c->size;
 			prune_base_data(c);
 		}
 		free(delta);
@@ -921,6 +956,8 @@ static struct base_data *make_base(struct object_entry *obj,
 				&base->ref_first, &base->ref_last);
 	find_ofs_delta_children(obj->idx.offset,
 				&base->ofs_first, &base->ofs_last);
+	base->children_remaining = base->ref_last - base->ref_first +
+		base->ofs_last - base->ofs_first + 2;
 	return base;
 }
 
@@ -954,14 +991,8 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 		    &delta_obj->idx.oid);
 
 	result = make_base(delta_obj, base);
-	if (result->ref_last == -1 && result->ofs_last == -1) {
-		free(result_data);
-	} else {
-		result->data = result_data;
-		result->size = result_size;
-		get_thread_data()->base_cache_used += result->size;
-		prune_base_data(result);
-	}
+	result->data = result_data;
+	result->size = result_size;
 
 	counter_lock();
 	nr_resolved_deltas++;
@@ -970,86 +1001,6 @@ static struct base_data *resolve_delta(struct object_entry *delta_obj,
 	return result;
 }
 
-/*
- * Standard boolean compare-and-swap: atomically check whether "*type" is
- * "want"; if so, swap in "set" and return true. Otherwise, leave it untouched
- * and return false.
- */
-static int compare_and_swap_type(signed char *type,
-				 enum object_type want,
-				 enum object_type set)
-{
-	enum object_type old;
-
-	type_cas_lock();
-	old = *type;
-	if (old == want)
-		*type = set;
-	type_cas_unlock();
-
-	return old == want;
-}
-
-static struct base_data *find_unresolved_deltas_1(struct base_data *base,
-						  struct base_data *prev_base)
-{
-	if (base->ref_first <= base->ref_last) {
-		struct object_entry *child = objects + ref_deltas[base->ref_first].obj_no;
-		struct base_data *result;
-
-		if (!compare_and_swap_type(&child->real_type, OBJ_REF_DELTA,
-					   base->obj->real_type))
-			die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
-			    (uintmax_t)child->idx.offset,
-			    oid_to_hex(&base->obj->idx.oid));
-
-		get_base_data(base);
-		result = resolve_delta(child, base);
-		if (base->ref_first == base->ref_last && base->ofs_last == -1)
-			free_base_data(base);
-
-		base->ref_first++;
-		return result;
-	}
-
-	if (base->ofs_first <= base->ofs_last) {
-		struct object_entry *child = objects + ofs_deltas[base->ofs_first].obj_no;
-		struct base_data *result;
-
-		assert(child->real_type == OBJ_OFS_DELTA);
-		child->real_type = base->obj->real_type;
-		get_base_data(base);
-		result = resolve_delta(child, base);
-		if (base->ofs_first == base->ofs_last)
-			free_base_data(base);
-
-		base->ofs_first++;
-		return result;
-	}
-
-	free_base_data(base);
-	return NULL;
-}
-
-static void find_unresolved_deltas(struct base_data *base)
-{
-	struct base_data *new_base, *prev_base = NULL;
-	for (;;) {
-		new_base = find_unresolved_deltas_1(base, prev_base);
-
-		if (new_base) {
-			prev_base = base;
-			base = new_base;
-		} else {
-			free(base);
-			base = prev_base;
-			if (!base)
-				return;
-			prev_base = base->base;
-		}
-	}
-}
-
 static int compare_ofs_delta_entry(const void *a, const void *b)
 {
 	const struct ofs_delta_entry *delta_a = a;
@@ -1068,33 +1019,131 @@ static int compare_ref_delta_entry(const void *a, const void *b)
 	return oidcmp(&delta_a->oid, &delta_b->oid);
 }
 
-static void resolve_base(struct object_entry *obj)
-{
-	struct base_data *base_obj = make_base(obj, NULL);
-
-	find_unresolved_deltas(base_obj);
-}
-
 static void *threaded_second_pass(void *data)
 {
-	set_thread_data(data);
+	if (data)
+		set_thread_data(data);
 	for (;;) {
-		int i;
-		counter_lock();
-		display_progress(progress, nr_resolved_deltas);
-		counter_unlock();
+		struct base_data *parent = NULL;
+		struct object_entry *child_obj;
+		struct base_data *child;
+
 		work_lock();
-		while (nr_dispatched < nr_objects &&
-		       is_delta_type(objects[nr_dispatched].type))
-			nr_dispatched++;
-		if (nr_dispatched >= nr_objects) {
-			work_unlock();
-			break;
+		if (list_empty(&work_head)) {
+			/*
+			 * Take an object from the object array.
+			 */
+			while (nr_dispatched < nr_objects &&
+			       is_delta_type(objects[nr_dispatched].type))
+				nr_dispatched++;
+			if (nr_dispatched >= nr_objects) {
+				work_unlock();
+				break;
+			}
+			child_obj = &objects[nr_dispatched++];
+		} else {
+			/*
+			 * Peek at the top of the stack, and take a child from
+			 * it.
+			 */
+			parent = list_first_entry(&work_head, struct base_data,
+						  list);
+
+			if (parent->ref_first <= parent->ref_last) {
+				int offset = ref_deltas[parent->ref_first++].obj_no;
+				child_obj = objects + offset;
+				if (child_obj->real_type != OBJ_REF_DELTA)
+					die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
+					    (uintmax_t) child_obj->idx.offset,
+					    oid_to_hex(&parent->obj->idx.oid));
+				child_obj->real_type = parent->obj->real_type;
+			} else {
+				child_obj = objects +
+					ofs_deltas[parent->ofs_first++].obj_no;
+				assert(child_obj->real_type == OBJ_OFS_DELTA);
+				child_obj->real_type = parent->obj->real_type;
+			}
+
+			if (parent->ref_first > parent->ref_last &&
+			    parent->ofs_first > parent->ofs_last) {
+				/*
+				 * This parent has run out of children, so move
+				 * it to done_head.
+				 */
+				list_del(&parent->list);
+				list_add(&parent->list, &done_head);
+			}
+
+			/*
+			 * Ensure that the parent has data, since we will need
+			 * it later.
+			 *
+			 * NEEDSWORK: If parent data needs to be reloaded, this
+			 * prolongs the time that the current thread spends in
+			 * the mutex. A mitigating factor is that parent data
+			 * needs to be reloaded only if the delta base cache
+			 * limit is exceeded, so in the typical case, this does
+			 * not happen.
+			 */
+			get_base_data(parent);
+			parent->retain_data++;
 		}
-		i = nr_dispatched++;
 		work_unlock();
 
-		resolve_base(&objects[i]);
+		if (parent) {
+			child = resolve_delta(child_obj, parent);
+			if (!child->children_remaining)
+				FREE_AND_NULL(child->data);
+		} else {
+			child = make_base(child_obj, NULL);
+			if (child->children_remaining) {
+				/*
+				 * Since this child has its own delta children,
+				 * we will need this data in the future.
+				 * Inflate now so that future iterations will
+				 * have access to this object's data while
+				 * outside the work mutex.
+				 */
+				child->data = get_data_from_pack(child_obj);
+				child->size = child_obj->size;
+			}
+		}
+
+		work_lock();
+		if (parent)
+			parent->retain_data--;
+		if (child->data) {
+			/*
+			 * This child has its own children, so add it to
+			 * work_head.
+			 */
+			list_add(&child->list, &work_head);
+			base_cache_used += child->size;
+			prune_base_data(NULL);
+		} else {
+			/*
+			 * This child does not have its own children. It may be
+			 * the last descendant of its ancestors; free those
+			 * that we can.
+			 */
+			struct base_data *p = parent;
+
+			while (p) {
+				struct base_data *next_p;
+
+				p->children_remaining--;
+				if (p->children_remaining)
+					break;
+
+				next_p = p->base;
+				free_base_data(p);
+				list_del(&p->list);
+				free(p);
+
+				p = next_p;
+			}
+		}
+		work_unlock();
 	}
 	return NULL;
 }
@@ -1195,6 +1244,7 @@ static void resolve_deltas(void)
 					  nr_ref_deltas + nr_ofs_deltas);
 
 	nr_dispatched = 0;
+	base_cache_limit = delta_base_cache_limit * nr_threads;
 	if (nr_threads > 1 || getenv("GIT_FORCE_THREADS")) {
 		init_thread();
 		for (i = 0; i < nr_threads; i++) {
@@ -1364,10 +1414,8 @@ static void fix_unresolved_deltas(struct hashfile *f)
 	for (i = 0; i < nr_ref_deltas; i++) {
 		struct ref_delta_entry *d = sorted_by_pos[i];
 		enum object_type type;
-		struct base_data *base;
 		void *data;
 		unsigned long size;
-		struct object_entry *obj;
 
 		if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
 			continue;
@@ -1379,11 +1427,15 @@ static void fix_unresolved_deltas(struct hashfile *f)
 					   data, size,
 					   type_name(type)))
 			die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
-		obj = append_obj_to_pack(f, d->oid.hash, data, size, type);
-		base = make_base(obj, NULL);
-		base->data = data;
-		base->size = size;
-		find_unresolved_deltas(base);
+
+		/*
+		 * Add this as an object to the objects array and call
+		 * threaded_second_pass() (which will pick up the added
+		 * object).
+		 */
+		append_obj_to_pack(f, d->oid.hash, data, size, type);
+		threaded_second_pass(NULL);
+
 		display_progress(progress, nr_resolved_deltas);
 	}
 	free(sorted_by_pos);
Signature · 1 line
2.28.0.526.ge36021eeef-goog
#27 lore raw
In reply to Jonathan Tan
Quoted 3 lines · Jonathan Tan

Here's the reroll.

[1] https://lore.kernel.org/git/[email protected]/

It seems that the only changes (as expected) are squashing in the follow-up "oops" fixes into the right place in the series?

Will replace; thanks.

Quoted 13 lines

Jonathan Tan (7):
Documentation: deltaBaseCacheLimit is per-thread
index-pack: remove redundant parameter
index-pack: unify threaded and unthreaded code
index-pack: remove redundant child field
index-pack: calculate {ref,ofs}_{first,last} early
index-pack: make resolve_delta() assume base data
index-pack: make quantum of work smaller

 Documentation/config/core.txt |   2 +-
 builtin/index-pack.c          | 456 +++++++++++++++++++---------------
 2 files changed, 251 insertions(+), 207 deletions(-)