[PATCH v2 09/10] reftable/block: reuse `zstream` state on inflation

From: Patrick Steinhardt <ps@pks.im>
Date: Apr 8, 2024, 12:17 UTC
Message-ID: <587b5601c0c51a8f041258a27e693a2590468eed.1712578376.git.ps@pks.im>
In-Reply-To: <cover.1712578376.git.ps@pks.im>

When calling `inflateInit()` and `inflate()`, the zlib library will allocate several data structures for the underlying `zstream` to keep track of various information. Thus, when inflating repeatedly, it is possible to optimize memory allocation patterns by reusing the `zstream` and then calling `inflateReset()` on it to prepare it for the next chunk of data to inflate.

This is exactly what the reftable code is doing: when iterating through reflogs we need to potentially inflate many log blocks, but we discard the `zstream` every single time. Instead, as we reuse the `block_reader` for each of the blocks anyway, we can initialize the `zstream` once and then reuse it for subsequent inflations.

Refactor the code to do so, which leads to a significant reduction in the number of allocations. The following measurements were done when iterating through 1 million reflog entries. Before:

  HEAP SUMMARY:
      in use at exit: 13,473 bytes in 122 blocks
    total heap usage: 23,028 allocs, 22,906 frees, 162,813,552 bytes allocated

After:

  HEAP SUMMARY:
      in use at exit: 13,473 bytes in 122 blocks
    total heap usage: 302 allocs, 180 frees, 88,352 bytes allocated

Signed-off-by: Patrick Steinhardt <ps@pks.im>
---
 reftable/block.c  | 25 +++++++++++++++----------
 reftable/block.h  |  3 +++
 reftable/reader.c |  1 +
 3 files changed, 19 insertions(+), 10 deletions(-)

diff --git a/reftable/block.c b/reftable/block.c
index 435922b569..c6c4a68ea1 100644
--- a/reftable/block.c
+++ b/reftable/block.c
@@ -198,7 +198,6 @@ int block_reader_init(struct block_reader *br, struct reftable_block *block,
 		uint32_t block_header_skip = 4 + header_off;
 		uLong dst_len = sz - block_header_skip;
 		uLong src_len = block->len - block_header_skip;
-		z_stream stream = {0};
 
 		/* Log blocks specify the *uncompressed* size in their header. */
 		REFTABLE_ALLOC_GROW(br->uncompressed_data, sz,
@@ -207,16 +206,21 @@ int block_reader_init(struct block_reader *br, struct reftable_block *block,
 		/* Copy over the block header verbatim. It's not compressed. */
 		memcpy(br->uncompressed_data, block->data, block_header_skip);
 
-		err = inflateInit(&stream);
+		if (!br->zstream) {
+			REFTABLE_CALLOC_ARRAY(br->zstream, 1);
+			err = inflateInit(br->zstream);
+		} else {
+			err = inflateReset(br->zstream);
+		}
 		if (err != Z_OK) {
 			err = REFTABLE_ZLIB_ERROR;
 			goto done;
 		}
 
-		stream.next_in = block->data + block_header_skip;
-		stream.avail_in = src_len;
-		stream.next_out = br->uncompressed_data + block_header_skip;
-		stream.avail_out = dst_len;
+		br->zstream->next_in = block->data + block_header_skip;
+		br->zstream->avail_in = src_len;
+		br->zstream->next_out = br->uncompressed_data + block_header_skip;
+		br->zstream->avail_out = dst_len;
 
 		/*
 		 * We know both input as well as output size, and we know that
@@ -225,15 +229,14 @@ int block_reader_init(struct block_reader *br, struct reftable_block *block,
 		 * here to instruct zlib to inflate the data in one go, which
 		 * is more efficient than using `Z_NO_FLUSH`.
 		 */
-		err = inflate(&stream, Z_FINISH);
-		inflateEnd(&stream);
+		err = inflate(br->zstream, Z_FINISH);
 		if (err != Z_STREAM_END) {
 			err = REFTABLE_ZLIB_ERROR;
 			goto done;
 		}
 		err = 0;
 
-		if (stream.total_out + block_header_skip != sz) {
+		if (br->zstream->total_out + block_header_skip != sz) {
 			err = REFTABLE_FORMAT_ERROR;
 			goto done;
 		}
@@ -242,7 +245,7 @@ int block_reader_init(struct block_reader *br, struct reftable_block *block,
 		reftable_block_done(block);
 		block->data = br->uncompressed_data;
 		block->len = sz;
-		full_block_size = src_len + block_header_skip - stream.avail_in;
+		full_block_size = src_len + block_header_skip - br->zstream->avail_in;
 	} else if (full_block_size == 0) {
 		full_block_size = sz;
 	} else if (sz < full_block_size && sz < block->len &&
@@ -275,6 +278,8 @@ int block_reader_init(struct block_reader *br, struct reftable_block *block,
 
 void block_reader_release(struct block_reader *br)
 {
+	inflateEnd(br->zstream);
+	reftable_free(br->zstream);
 	reftable_free(br->uncompressed_data);
 	reftable_block_done(&br->block);
 }
diff --git a/reftable/block.h b/reftable/block.h
index 12414eb642..c1bd1892cb 100644
--- a/reftable/block.h
+++ b/reftable/block.h
@@ -56,6 +56,8 @@ int block_writer_finish(struct block_writer *w);
 /* clears out internally allocated block_writer members. */
 void block_writer_release(struct block_writer *bw);
 
+struct z_stream;
+
 /* Read a block. */
 struct block_reader {
 	/* offset of the block header; nonzero for the first block in a
@@ -67,6 +69,7 @@ struct block_reader {
 	int hash_size;
 
 	/* Uncompressed data for log entries. */
+	z_stream *zstream;
 	unsigned char *uncompressed_data;
 	size_t uncompressed_cap;
 
diff --git a/reftable/reader.c b/reftable/reader.c
index aacd5f1337..481dff10d4 100644
--- a/reftable/reader.c
+++ b/reftable/reader.c
@@ -459,6 +459,7 @@ static int reader_seek_linear(struct table_iter *ti,
 		 * we would not do a linear search there anymore.
 		 */
 		memset(&next.br.block, 0, sizeof(next.br.block));
+		next.br.zstream = NULL;
 		next.br.uncompressed_data = NULL;
 		next.br.uncompressed_cap = 0;

-- 
2.44.GIT