git/list[1] front-page[2] threads[3] people[4] search[5] about
 

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

From
JTJonathan Tan <jonathantanmy@google.com>
Date
Sep 8, 2020, 19:48 UTC
Message-ID
<01d6d882768127df1596e3a4c63d9b1a5a9ba093.1599594441.git.jonathantanmy@google.com>
In-Reply-To
<cover.1599594441.git.jonathantanmy@google.com>

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.

Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
---
 builtin/index-pack.c | 348 +++++++++++++++++++++++++------------------
 1 file changed, 200 insertions(+), 148 deletions(-)
diff --git a/builtin/index-pack.c b/builtin/index-pack.c
index f69a50d46b..8acd078aa0 100644
--- a/builtin/index-pack.c
+++ b/builtin/index-pack.c
@@ -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);
-- 
2.28.0.526.ge36021eeef-goog
Previous: Jonathan TanNext: Jonathan Tan
Message 23 of 28 in “Better threaded delta resolution in index-pack (another try)”
  1. 0/7 Better threaded delta resolution in index-pack (another try)Jonathan Tan, Aug 24, 2020
  2. 1/7 Documentation: deltaBaseCacheLimit is per-threadJonathan Tan, Aug 24, 2020
  3. fetch-pack: in partial clone, pass --promisorJonathan Tan, Aug 24, 2020
  4. Jonathan TanAug 24, 2020
  5. 2/7 index-pack: remove redundant parameterJonathan Tan, Aug 24, 2020
  6. Junio C HamanoAug 24, 2020
  7. 3/7 index-pack: unify threaded and unthreaded codeJonathan Tan, Aug 24, 2020
  8. Junio C HamanoAug 24, 2020
  9. 5/7 index-pack: calculate {ref,ofs}_{first,last} earlyJonathan Tan, Aug 24, 2020
  10. 7/7 index-pack: make quantum of work smallerJonathan Tan, Aug 24, 2020
  11. Junio C HamanoAug 24, 2020
  12. 6/7 index-pack: make resolve_delta() assume base dataJonathan Tan, Aug 24, 2020
  13. 4/7 index-pack: remove redundant child fieldJonathan Tan, Aug 24, 2020
  14. Junio C HamanoAug 24, 2020
  15. fixup! index-pack: make quantum of work smallerJonathan Tan, Aug 24, 2020
  16. Jeff KingAug 24, 2020
  17. Jonathan TanAug 25, 2020
  18. Jeff KingAug 25, 2020
  19. Jeff KingAug 25, 2020
  20. 0/7 Better threaded delta resolution in index-pack (another try)Jonathan Tan, Sep 8, 2020
  21. 2/7 index-pack: remove redundant parameterJonathan Tan, Sep 8, 2020
  22. 3/7 index-pack: unify threaded and unthreaded codeJonathan Tan, Sep 8, 2020
  23. 7/7 index-pack: make quantum of work smallerJonathan Tan, Sep 8, 2020
  24. 6/7 index-pack: make resolve_delta() assume base dataJonathan Tan, Sep 8, 2020
  25. 5/7 index-pack: calculate {ref,ofs}_{first,last} earlyJonathan Tan, Sep 8, 2020
  26. 4/7 index-pack: remove redundant child fieldJonathan Tan, Sep 8, 2020
  27. 1/7 Documentation: deltaBaseCacheLimit is per-threadJonathan Tan, Sep 8, 2020
  28. Junio C HamanoSep 8, 2020

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.