block processor: recycle blocks to reduce allocation pressure

Instead of freeing/allocating blocks all the time in the locked,
serial path, use a free list to "recycle" blocks. Once a block is
no longer used, throw it onto the free list. If a new block is,
needed try to get one from the free list before calling malloc.

After a few iterations, the block processor should stop allocating
new blocks and only re-use the ones it already has.

Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>

4 files changed, 73 insertions, 39 deletions

diff --git a/lib/sqfs/block_processor/common.c b/lib/sqfs/block_processor/common.c
index 50c0573..b8f2760 100644
--- a/lib/sqfs/block_processor/common.c
+++ b/lib/sqfs/block_processor/common.c
@@ -44,6 +44,29 @@ static int set_block_size(sqfs_inode_generic_t **inode,
 	return 0;
 }
 
+static sqfs_block_t *get_new_block(sqfs_block_processor_t *proc)
+{
+	sqfs_block_t *blk;
+
+	if (proc->free_list != NULL) {
+		blk = proc->free_list;
+		proc->free_list = blk->next;
+	} else {
+		blk = malloc(sizeof(*blk) + proc->max_block_size);
+	}
+
+	if (blk != NULL)
+		memset(blk, 0, sizeof(*blk));
+
+	return blk;
+}
+
+static void release_old_block(sqfs_block_processor_t *proc, sqfs_block_t *blk)
+{
+	blk->next = proc->free_list;
+	proc->free_list = blk;
+}
+
 int process_completed_block(sqfs_block_processor_t *proc, sqfs_block_t *blk)
 {
 	sqfs_u64 location;
@@ -53,7 +76,7 @@ int process_completed_block(sqfs_block_processor_t *proc, sqfs_block_t *blk)
 	err = sqfs_block_writer_write(proc->wr, blk->size, blk->checksum,
 				      blk->flags, blk->data, &location);
 	if (err)
-		return err;
+		goto out;
 
 	if (blk->flags & SQFS_BLK_IS_SPARSE) {
 		sqfs_inode_make_extended(*(blk->inode));
@@ -61,7 +84,7 @@ int process_completed_block(sqfs_block_processor_t *proc, sqfs_block_t *blk)
 
 		err = set_block_size(blk->inode, blk->index, 0);
 		if (err)
-			return err;
+			goto out;
 
 		proc->stats.sparse_block_count += 1;
 	} else if (blk->size != 0) {
@@ -73,19 +96,20 @@ int process_completed_block(sqfs_block_processor_t *proc, sqfs_block_t *blk)
 			err = sqfs_frag_table_set(proc->frag_tbl, blk->index,
 						  location, size);
 			if (err)
-				return err;
+				goto out;
 		} else {
 			err = set_block_size(blk->inode, blk->index, size);
 			if (err)
-				return err;
+				goto out;
 			proc->stats.data_block_count += 1;
 		}
 	}
 
 	if (blk->flags & SQFS_BLK_LAST_BLOCK)
 		sqfs_inode_set_file_block_start(*(blk->inode), location);
-
-	return 0;
+out:
+	release_old_block(proc, blk);
+	return err;
 }
 
 static bool is_zero_block(unsigned char *ptr, size_t size)
@@ -137,6 +161,7 @@ int process_completed_fragment(sqfs_block_processor_t *proc, sqfs_block_t *frag,
 		(*(frag->inode))->data.file_ext.sparse += frag->size;
 
 		proc->stats.sparse_block_count += 1;
+		release_old_block(proc, frag);
 		return 0;
 	}
 
@@ -149,6 +174,7 @@ int process_completed_fragment(sqfs_block_processor_t *proc, sqfs_block_t *frag,
 		if (err == 0) {
 			sqfs_inode_set_frag_location(*(frag->inode),
 						     index, offset);
+			release_old_block(proc, frag);
 			return 0;
 		}
 	}
@@ -167,15 +193,12 @@ int process_completed_fragment(sqfs_block_processor_t *proc, sqfs_block_t *frag,
 		if (err)
 			goto fail;
 
-		size = sizeof(sqfs_block_t) + proc->max_block_size;
-		proc->frag_block = malloc(size);
-
+		proc->frag_block = get_new_block(proc);
 		if (proc->frag_block == NULL) {
 			err = SQFS_ERROR_ALLOC;
 			goto fail;
 		}
 
-		memset(proc->frag_block, 0, sizeof(sqfs_block_t));
 		proc->frag_block->index = index;
 		proc->frag_block->flags = SQFS_BLK_FRAGMENT_BLOCK;
 		proc->stats.frag_block_count += 1;
@@ -197,16 +220,21 @@ int process_completed_fragment(sqfs_block_processor_t *proc, sqfs_block_t *frag,
 	proc->frag_block->flags |= (frag->flags & SQFS_BLK_DONT_COMPRESS);
 	proc->frag_block->size += frag->size;
 	proc->stats.actual_frag_count += 1;
+
+	release_old_block(proc, frag);
 	return 0;
 fail:
-	free(*blk_out);
-	*blk_out = NULL;
+	release_old_block(proc, frag);
+	if (*blk_out != NULL) {
+		release_old_block(proc, *blk_out);
+		*blk_out = NULL;
+	}
 	return err;
 }
 
 static int add_sentinel_block(sqfs_block_processor_t *proc)
 {
-	sqfs_block_t *blk = calloc(1, sizeof(*blk));
+	sqfs_block_t *blk = get_new_block(proc);
 
 	if (blk == NULL)
 		return SQFS_ERROR_ALLOC;
@@ -269,11 +297,10 @@ int sqfs_block_processor_append(sqfs_block_processor_t *proc, const void *data,
 
 	while (size > 0) {
 		if (proc->blk_current == NULL) {
-			new = malloc(sizeof(*new) + proc->max_block_size);
+			new = get_new_block(proc);
 			if (new == NULL)
 				return SQFS_ERROR_ALLOC;
 
-			memset(new, 0, sizeof(*new));
 			proc->blk_current = new;
 			proc->blk_current->flags = proc->blk_flags;
 			proc->blk_current->inode = proc->inode;
diff --git a/lib/sqfs/block_processor/internal.h b/lib/sqfs/block_processor/internal.h
index fed0842..eff5374 100644
--- a/lib/sqfs/block_processor/internal.h
+++ b/lib/sqfs/block_processor/internal.h
@@ -54,6 +54,8 @@ struct sqfs_block_processor_t {
 	sqfs_u32 blk_flags;
 	sqfs_u32 blk_index;
 
+	sqfs_block_t *free_list;
+
 	size_t max_block_size;
 };
 
diff --git a/lib/sqfs/block_processor/serial.c b/lib/sqfs/block_processor/serial.c
index b797ca0..a9190b1 100644
--- a/lib/sqfs/block_processor/serial.c
+++ b/lib/sqfs/block_processor/serial.c
@@ -16,6 +16,13 @@ typedef struct {
 static void block_processor_destroy(sqfs_object_t *obj)
 {
 	sqfs_block_processor_t *proc = (sqfs_block_processor_t *)obj;
+	sqfs_block_t *blk;
+
+	while (proc->free_list != NULL) {
+		blk = proc->free_list;
+		proc->free_list = blk->next;
+		free(blk);
+	}
 
 	free(proc->blk_current);
 	free(proc->frag_block);
@@ -50,34 +57,31 @@ int append_to_work_queue(sqfs_block_processor_t *proc, sqfs_block_t *block)
 	serial_block_processor_t *sproc = (serial_block_processor_t *)proc;
 	sqfs_block_t *fragblk = NULL;
 
-	if (sproc->status != 0)
-		goto done;
+	if (sproc->status != 0) {
+		free(block);
+		return sproc->status;
+	}
 
 	sproc->status = block_processor_do_block(block, proc->cmp,
 						 sproc->scratch,
 						 proc->max_block_size);
 	if (sproc->status != 0)
-		goto done;
+		return sproc->status;
 
 	if (block->flags & SQFS_BLK_IS_FRAGMENT) {
 		sproc->status = process_completed_fragment(proc, block,
 							   &fragblk);
 		if (fragblk == NULL)
-			goto done;
-
-		free(block);
-		block = fragblk;
+			return sproc->status;
 
-		sproc->status = block_processor_do_block(block, proc->cmp,
+		sproc->status = block_processor_do_block(fragblk, proc->cmp,
 							 sproc->scratch,
 							 proc->max_block_size);
 		if (sproc->status != 0)
-			goto done;
+			return sproc->status;
 	}
 
 	sproc->status = process_completed_block(proc, block);
-done:
-	free(block);
 	return sproc->status;
 }
 
@@ -91,16 +95,17 @@ int sqfs_block_processor_finish(sqfs_block_processor_t *proc)
 	serial_block_processor_t *sproc = (serial_block_processor_t *)proc;
 
 	if (proc->frag_block == NULL || sproc->status != 0)
-		goto out;
+		goto fail;
 
 	sproc->status = block_processor_do_block(proc->frag_block, proc->cmp,
 						 sproc->scratch,
 						 proc->max_block_size);
 	if (sproc->status != 0)
-		goto out;
+		goto fail;
 
 	sproc->status = process_completed_block(proc, proc->frag_block);
-out:
+	return sproc->status;
+fail:
 	free(proc->frag_block);
 	proc->frag_block = NULL;
 	return sproc->status;
diff --git a/lib/sqfs/block_processor/winpthread.c b/lib/sqfs/block_processor/winpthread.c
index 80cbde9..281eb6b 100644
--- a/lib/sqfs/block_processor/winpthread.c
+++ b/lib/sqfs/block_processor/winpthread.c
@@ -192,6 +192,7 @@ static void block_processor_destroy(sqfs_object_t *obj)
 	free_blk_list(proc->proc_queue);
 	free_blk_list(proc->io_queue);
 	free_blk_list(proc->done);
+	free_blk_list(proc->base.free_list);
 	free(proc->base.blk_current);
 	free(proc->base.frag_block);
 	free(proc);
@@ -390,12 +391,13 @@ static void append_block(thread_pool_processor_t *proc, sqfs_block_t *block)
 
 static int handle_io_queue(thread_pool_processor_t *proc, sqfs_block_t *list)
 {
-	sqfs_block_t *it = list;
+	sqfs_block_t *it;
 	int status = 0;
 
-	while (status == 0 && it != NULL) {
+	while (status == 0 && list != NULL) {
+		it = list;
+		list = list->next;
 		status = process_completed_block(&proc->base, it);
-		it = it->next;
 
 		if (status != 0) {
 			LOCK(&proc->mtx);
@@ -413,7 +415,7 @@ int append_to_work_queue(sqfs_block_processor_t *proc, sqfs_block_t *block)
 {
 	thread_pool_processor_t *thproc = (thread_pool_processor_t *)proc;
 	sqfs_block_t *io_list = NULL, *io_list_last = NULL;
-	sqfs_block_t *blk, *fragblk, *free_list = NULL;
+	sqfs_block_t *blk, *fragblk;
 	int status;
 
 	LOCK(&thproc->mtx);
@@ -454,8 +456,6 @@ int append_to_work_queue(sqfs_block_processor_t *proc, sqfs_block_t *block)
 			fragblk = NULL;
 			thproc->status = process_completed_fragment(proc, blk,
 								    &fragblk);
-			blk->next = free_list;
-			free_list = blk;
 
 			if (fragblk != NULL) {
 				fragblk->io_seq_num = thproc->io_enq_id++;
@@ -472,11 +472,12 @@ int append_to_work_queue(sqfs_block_processor_t *proc, sqfs_block_t *block)
 	UNLOCK(&thproc->mtx);
 	free(block);
 
-	if (status == 0)
+	if (status == 0) {
 		status = handle_io_queue(thproc, io_list);
+	} else {
+		free_blk_list(io_list);
+	}
 
-	free_blk_list(io_list);
-	free_blk_list(free_list);
 	return status;
 }
 
@@ -504,7 +505,6 @@ int sqfs_block_processor_finish(sqfs_block_processor_t *proc)
 
 		if (status == 0)
 			status = handle_io_queue(thproc, blk);
-		free(blk);
 
 		if (status != 0) {
 			LOCK(&thproc->mtx);