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

[PATCH 11/12] reftable/record: decode keys in place

From
Patrick Steinhardt <ps@pks.im>
Date
Feb 14, 2024, 07:46 UTC
Message-ID
<f0663c1d62d13d01afb1fa6c3760a38741bdfc8e.1707895758.git.ps@pks.im>
In-Reply-To
<cover.1707895758.git.ps@pks.im>

When reading a record from a block, we need to decode the record's key. As reftable keys are prefix-compressed, meaning they reuse a prefix from the preceding record's key, this is a bit more involved than just having to copy the relevant bytes: we need to figure out the prefix and suffix lengths, copy the prefix from the preceding record and finally copy the suffix from the current record.

This is done by passing three buffers to `reftable_decode_key()`: one buffer that holds the result, one buffer that holds the last key, and one buffer that points to the current record. The final key is then assembled by calling `strbuf_add()` twice to copy over the prefix and suffix.

Performing two memory copies is inefficient though. And we can indeed do better by decoding keys in place. Instead of providing two buffers, the caller may only call a single buffer that is already pre-populated with the last key. Like this, we only have to call `strbuf_setlen()` to trim the record to its prefix and then `strbuf_add()` to add the suffix.

This refactoring leads to a noticeable performance bump when iterating over 1 million refs:

  Benchmark 1: show-ref: single matching ref (revision = HEAD~)
    Time (mean ± σ):     112.2 ms ±   3.9 ms    [User: 109.3 ms, System: 2.8 ms]
    Range (min … max):   109.2 ms … 149.6 ms    1000 runs
  Benchmark 2: show-ref: single matching ref (revision = HEAD)
    Time (mean ± σ):     106.0 ms ±   3.5 ms    [User: 103.2 ms, System: 2.7 ms]
    Range (min … max):   103.2 ms … 133.7 ms    1000 runs
  Summary
    show-ref: single matching ref (revision = HEAD) ran
      1.06 ± 0.05 times faster than show-ref: single matching ref (revision = HEAD~)
Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 reftable/block.c       | 25 +++++++++++--------------
 reftable/block.h       |  2 --
 reftable/record.c      | 19 +++++++++----------
 reftable/record.h      |  9 ++++++---
 reftable/record_test.c |  3 ++-
 5 files changed, 28 insertions(+), 30 deletions(-)
diff --git a/reftable/block.c b/reftable/block.c
index 72eb73b380..ad9074dba6 100644
--- a/reftable/block.c
+++ b/reftable/block.c
@@ -291,9 +291,8 @@ static int restart_key_less(size_t idx, void *args)
 	/* the restart key is verbatim in the block, so this could avoid the
 	   alloc for decoding the key */
 	struct strbuf rkey = STRBUF_INIT;
-	struct strbuf last_key = STRBUF_INIT;
 	uint8_t unused_extra;
-	int n = reftable_decode_key(&rkey, &unused_extra, last_key, in);
+	int n = reftable_decode_key(&rkey, &unused_extra, in);
 	int result;
 	if (n < 0) {
 		a->error = 1;
@@ -326,35 +325,34 @@ int block_iter_next(struct block_iter *it, struct reftable_record *rec)
 	if (it->next_off >= it->br->block_len)
 		return 1;
 
-	n = reftable_decode_key(&it->key, &extra, it->last_key, in);
+	n = reftable_decode_key(&it->last_key, &extra, in);
 	if (n < 0)
 		return -1;
-
-	if (!it->key.len)
+	if (!it->last_key.len)
 		return REFTABLE_FORMAT_ERROR;
 
 	string_view_consume(&in, n);
-	n = reftable_record_decode(rec, it->key, extra, in, it->br->hash_size);
+	n = reftable_record_decode(rec, it->last_key, extra, in, it->br->hash_size);
 	if (n < 0)
 		return -1;
 	string_view_consume(&in, n);
 
-	strbuf_swap(&it->last_key, &it->key);
 	it->next_off += start.len - in.len;
 	return 0;
 }
 
 int block_reader_first_key(struct block_reader *br, struct strbuf *key)
 {
-	struct strbuf empty = STRBUF_INIT;
-	int off = br->header_off + 4;
+	int off = br->header_off + 4, n;
 	struct string_view in = {
 		.buf = br->block.data + off,
 		.len = br->block_len - off,
 	};
-
 	uint8_t extra = 0;
-	int n = reftable_decode_key(key, &extra, empty, in);
+
+	strbuf_reset(key);
+
+	n = reftable_decode_key(key, &extra, in);
 	if (n < 0)
 		return n;
 	if (!key->len)
@@ -371,7 +369,6 @@ int block_iter_seek(struct block_iter *it, struct strbuf *want)
 void block_iter_close(struct block_iter *it)
 {
 	strbuf_release(&it->last_key);
-	strbuf_release(&it->key);
 }
 
 int block_reader_seek(struct block_reader *br, struct block_iter *it,
@@ -408,8 +405,8 @@ int block_reader_seek(struct block_reader *br, struct block_iter *it,
 		if (err < 0)
 			goto done;
 
-		reftable_record_key(&rec, &it->key);
-		if (err > 0 || strbuf_cmp(&it->key, want) >= 0) {
+		reftable_record_key(&rec, &it->last_key);
+		if (err > 0 || strbuf_cmp(&it->last_key, want) >= 0) {
 			err = 0;
 			goto done;
 		}
diff --git a/reftable/block.h b/reftable/block.h
index 17481e6331..51699af233 100644
--- a/reftable/block.h
+++ b/reftable/block.h
@@ -84,12 +84,10 @@ struct block_iter {
 
 	/* key for last entry we read. */
 	struct strbuf last_key;
-	struct strbuf key;
 };
 
 #define BLOCK_ITER_INIT { \
 	.last_key = STRBUF_INIT, \
-	.key = STRBUF_INIT, \
 }
 
 /* initializes a block reader. */
diff --git a/reftable/record.c b/reftable/record.c
index 3f2a639036..37682cc7d0 100644
--- a/reftable/record.c
+++ b/reftable/record.c
@@ -159,20 +159,19 @@ int reftable_encode_key(int *restart, struct string_view dest,
 	return start.len - dest.len;
 }
 
-int reftable_decode_key(struct strbuf *key, uint8_t *extra,
-			struct strbuf last_key, struct string_view in)
+int reftable_decode_key(struct strbuf *last_key, uint8_t *extra,
+			struct string_view in)
 {
 	int start_len = in.len;
 	uint64_t prefix_len = 0;
 	uint64_t suffix_len = 0;
-	int n = get_var_int(&prefix_len, &in);
+	int n;
+
+	n = get_var_int(&prefix_len, &in);
 	if (n < 0)
 		return -1;
 	string_view_consume(&in, n);
 
-	if (prefix_len > last_key.len)
-		return -1;
-
 	n = get_var_int(&suffix_len, &in);
 	if (n <= 0)
 		return -1;
@@ -181,12 +180,12 @@ int reftable_decode_key(struct strbuf *key, uint8_t *extra,
 	*extra = (uint8_t)(suffix_len & 0x7);
 	suffix_len >>= 3;
 
-	if (in.len < suffix_len)
+	if (in.len < suffix_len ||
+	    prefix_len > last_key->len)
 		return -1;
 
-	strbuf_reset(key);
-	strbuf_add(key, last_key.buf, prefix_len);
-	strbuf_add(key, in.buf, suffix_len);
+	strbuf_setlen(last_key, prefix_len);
+	strbuf_add(last_key, in.buf, suffix_len);
 	string_view_consume(&in, suffix_len);
 
 	return start_len - in.len;
diff --git a/reftable/record.h b/reftable/record.h
index a05e2be179..91c9c6ebfd 100644
--- a/reftable/record.h
+++ b/reftable/record.h
@@ -81,9 +81,12 @@ int reftable_encode_key(int *is_restart, struct string_view dest,
 			struct strbuf prev_key, struct strbuf key,
 			uint8_t extra);
 
-/* Decode into `key` and `extra` from `in` */
-int reftable_decode_key(struct strbuf *key, uint8_t *extra,
-			struct strbuf last_key, struct string_view in);
+/*
+ * Decode into `last_key` and `extra` from `in`. `last_key` is expected to
+ * contain the decoded key of the preceding record, if any.
+ */
+int reftable_decode_key(struct strbuf *last_key, uint8_t *extra,
+			struct string_view in);
 
 /* reftable_index_record are used internally to speed up lookups. */
 struct reftable_index_record {
diff --git a/reftable/record_test.c b/reftable/record_test.c
index a86cff5526..89209894d8 100644
--- a/reftable/record_test.c
+++ b/reftable/record_test.c
@@ -295,7 +295,8 @@ static void test_key_roundtrip(void)
 	EXPECT(!restart);
 	EXPECT(n > 0);
 
-	m = reftable_decode_key(&roundtrip, &rt_extra, last_key, dest);
+	strbuf_addstr(&roundtrip, "refs/heads/master");
+	m = reftable_decode_key(&roundtrip, &rt_extra, dest);
 	EXPECT(n == m);
 	EXPECT(0 == strbuf_cmp(&key, &roundtrip));
 	EXPECT(rt_extra == extra);
-- 
2.43.GIT
Previous: James LiuNext: James Liu
Message 20 of 51 in “reftable: improve ref iteration performance (pt.2)”
  1. 00/12 reftable: improve ref iteration performance (pt.2)Patrick Steinhardt, Feb 14, 2024
  2. 01/12 reftable/pq: use `size_t` to track iterator indexPatrick Steinhardt, Feb 14, 2024
  3. 02/12 reftable/merged: make `merged_iter` structure privatePatrick Steinhardt, Feb 14, 2024
  4. Justin ToblerFeb 20, 2024
  5. Patrick SteinhardtFeb 27, 2024
  6. 03/12 reftable/merged: advance subiter on subsequent iterationPatrick Steinhardt, Feb 14, 2024
  7. Justin ToblerFeb 20, 2024
  8. Patrick SteinhardtFeb 27, 2024
  9. 04/12 reftable/merged: make subiters own their recordsPatrick Steinhardt, Feb 14, 2024
  10. 05/12 reftable/merged: remove unnecessary null check for subitersPatrick Steinhardt, Feb 14, 2024
  11. 06/12 reftable/merged: handle subiter cleanup on close onlyPatrick Steinhardt, Feb 14, 2024
  12. 07/12 reftable/merged: circumvent pqueue with single subiterPatrick Steinhardt, Feb 14, 2024
  13. 08/12 reftable/merged: avoid duplicate pqueue emptiness checkPatrick Steinhardt, Feb 14, 2024
  14. James LiuFeb 27, 2024
  15. 09/12 reftable/record: reuse refname when decodingPatrick Steinhardt, Feb 14, 2024
  16. James LiuFeb 28, 2024
  17. Patrick SteinhardtMar 4, 2024
  18. 10/12 reftable/record: reuse refname when copyingPatrick Steinhardt, Feb 14, 2024
  19. James LiuFeb 28, 2024
  20. 11/12 reftable/record: decode keys in placePatrick Steinhardt, Feb 14, 2024
  21. James LiuFeb 28, 2024
  22. Patrick SteinhardtMar 4, 2024
  23. 12/12 reftable: allow inlining of a few functionsPatrick Steinhardt, Feb 14, 2024
  24. 00/13 reftable: improve ref iteration performance (pt.2)Patrick Steinhardt, Feb 27, 2024
  25. 01/13 reftable/pq: use `size_t` to track iterator indexPatrick Steinhardt, Feb 27, 2024
  26. 02/13 reftable/merged: make `merged_iter` structure privatePatrick Steinhardt, Feb 27, 2024
  27. 03/13 reftable/merged: advance subiter on subsequent iterationPatrick Steinhardt, Feb 27, 2024
  28. 04/13 reftable/merged: make subiters own their recordsPatrick Steinhardt, Feb 27, 2024
  29. 05/13 reftable/merged: remove unnecessary null check for subitersPatrick Steinhardt, Feb 27, 2024
  30. 06/13 reftable/merged: handle subiter cleanup on close onlyPatrick Steinhardt, Feb 27, 2024
  31. 07/13 reftable/merged: circumvent pqueue with single subiterPatrick Steinhardt, Feb 27, 2024
  32. 08/13 reftable/merged: avoid duplicate pqueue emptiness checkPatrick Steinhardt, Feb 27, 2024
  33. 09/13 reftable/record: reuse refname when decodingPatrick Steinhardt, Feb 27, 2024
  34. 10/13 reftable/record: reuse refname when copyingPatrick Steinhardt, Feb 27, 2024
  35. 11/13 reftable/record: decode keys in placePatrick Steinhardt, Feb 27, 2024
  36. 12/13 reftable: allow inlining of a few functionsPatrick Steinhardt, Feb 27, 2024
  37. 13/13 refs/reftable: precompute prefix lengthPatrick Steinhardt, Feb 27, 2024
  38. 00/13 reftable: improve ref iteration performance (pt.2)Patrick Steinhardt, Mar 4, 2024
  39. 01/13 reftable/pq: use `size_t` to track iterator indexPatrick Steinhardt, Mar 4, 2024
  40. 02/13 reftable/merged: make `merged_iter` structure privatePatrick Steinhardt, Mar 4, 2024
  41. 03/13 reftable/merged: advance subiter on subsequent iterationPatrick Steinhardt, Mar 4, 2024
  42. 04/13 reftable/merged: make subiters own their recordsPatrick Steinhardt, Mar 4, 2024
  43. 05/13 reftable/merged: remove unnecessary null check for subitersPatrick Steinhardt, Mar 4, 2024
  44. 06/13 reftable/merged: handle subiter cleanup on close onlyPatrick Steinhardt, Mar 4, 2024
  45. 07/13 reftable/merged: circumvent pqueue with single subiterPatrick Steinhardt, Mar 4, 2024
  46. 08/13 reftable/merged: avoid duplicate pqueue emptiness checkPatrick Steinhardt, Mar 4, 2024
  47. 09/13 reftable/record: reuse refname when decodingPatrick Steinhardt, Mar 4, 2024
  48. 10/13 reftable/record: reuse refname when copyingPatrick Steinhardt, Mar 4, 2024
  49. 11/13 reftable/record: decode keys in placePatrick Steinhardt, Mar 4, 2024
  50. 12/13 reftable: allow inlining of a few functionsPatrick Steinhardt, Mar 4, 2024
  51. 13/13 refs/reftable: precompute prefix lengthPatrick Steinhardt, Mar 4, 2024

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.