[PATCH v2.1 00/10] btrfs-progs: image: Enhancement with new data dump feature

[PATCH v2.1 00/10] btrfs-progs: image: Enhancement with new data dump feature

[Qu Wenruo]

This updated patchset is rebased on devel branch, whose HEAD is:
commit 73289664917ad7c73f3f3393593e93a118228ad8 (david/devel)
Author: David Sterba <dsterba@suse.com>
Date:   Thu Jul 4 01:42:15 2019 +0200

    btrfs-progs: kerncompat: define __always_inline conditionally

This updated patchset includes the following features:

- Small fixes to improve btrfs-image error report
  * Output error message for chunk tree build error
  * Fix error output to show correct return value
  Patch 1 and 2.

- Reduce memory usage when decompress super block
  Independent change, for current btrfs-image, it will reduce buffer
  size from 256K to fixed 4K.
  Patch 3.

- Rework how we search chunk tree blocks
  Instead of iterating clusters again and again for each chunk tree
  block, record system chunk array and iterate clusters once for all
  chunk tree blocks.
  This should reduce restore time for large image dump.
  Patch 4, 5 and 6.

- Introduce data dump feature to dump the whole fs.
  This will introduce a new magic number to prevent old btrfs-image to
  hit failure as the item size limit is enlarged.
  Patch 7 and 8.

- Reduce memory usage for data dump restore
  This is mostly due to the fact that we have much larger
  max_pending_size introduced by data dump(256K -> 256M).
  Using 4 * max_pending_size for each decompress thread as buffer is way
  too expensive now.
  Use proper inflate() to replace uncompress() calls.
  Patch 9 and 10.

Changelog:
v2:
- New small fixes:
  * Fix a confusing error message due to unpopulated errno
  * Output error message for chunk tree build error
  
- Fix a regression of previous version
  Patch "btrfs-progs: image: Rework how we search chunk tree blocks"
  deleted a "ret = 0" line which could cause false early exit.

- Reduce memory usage for data dump

v2.1:
- Rebased to devel branch
  Removing 4 already merged patches from the patchset.

- Re-order the patchset
  Put small and independent patches at the top of queue, and put the
  data dump related feature at the end.

- Fix -Wmaybe-uninitialized warnings
  Strangely, D=1 won't trigger these warnings thus they sneak into v2
  without being detected.

- Fix FROM: line
  Reverted to old smtp setup. The new setup will override FROM: line,
  messing up the name of author.

Qu Wenruo (10):
  btrfs-progs: image: Output error message for chunk tree build error
  btrfs-progs: image: Fix error output to show correct return value
  btrfs-progs: image: Don't waste memory when we're just extracting
    super block
  btrfs-progs: image: Allow restore to record system chunk ranges for
    later usage
  btrfs-progs: image: Introduce helper to determine if a tree block is
    in the range of system chunks
  btrfs-progs: image: Rework how we search chunk tree blocks
  btrfs-progs: image: Introduce framework for more dump versions
  btrfs-progs: image: Introduce -d option to dump data
  btrfs-progs: image: Reduce memory requirement for decompression
  btrfs-progs: image: Reduce memory usage for chunk tree search

 image/main.c     | 866 +++++++++++++++++++++++++++++++++++------------
 image/metadump.h |  13 +-
 2 files changed, 654 insertions(+), 225 deletions(-)

-- 
2.22.0

[PATCH v2.1 01/10] btrfs-progs: image: Output error message for chunk tree build error

[Qu Wenruo]

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/image/main.c b/image/main.c
index 28ef1609b5ff..fb407f33f858 100644
--- a/image/main.c
+++ b/image/main.c
@@ -2406,8 +2406,10 @@ static int restore_metadump(const char *input, FILE *out, int old_restore,
 
 	if (!multi_devices && !old_restore) {
 		ret = build_chunk_tree(&mdrestore, cluster);
-		if (ret)
+		if (ret) {
+			error("failed to build chunk tree");
 			goto out;
+		}
 		if (!list_empty(&mdrestore.overlapping_chunks))
 			remap_overlapping_chunks(&mdrestore);
 	}
-- 
2.22.0

[PATCH v2.1 02/10] btrfs-progs: image: Fix error output to show correct return value

[Qu Wenruo]

We can easily get confusing error message like:
  ERROR: restore failed: Success

This is caused by wrong "%m" usage, as we normally use ret to indicate
error, without populating errno.

This patch will fix it by output the return value directly as normally
we have extra error message to show more meaning message than the return
value.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/image/main.c b/image/main.c
index fb407f33f858..7c499c0853d0 100644
--- a/image/main.c
+++ b/image/main.c
@@ -2734,7 +2734,7 @@ int BOX_MAIN(image)(int argc, char *argv[])
 				       0, target, multi_devices);
 	}
 	if (ret) {
-		error("%s failed: %m", (create) ? "create" : "restore");
+		error("%s failed: %d", (create) ? "create" : "restore", ret);
 		goto out;
 	}
 
-- 
2.22.0

[PATCH v2.1 03/10] btrfs-progs: image: Don't waste memory when we're just extracting super block

[Qu Wenruo]

There is no need to allocate 2 * max_pending_size (which can be 256M) if
we're just extracting super block.

We only need to prepare BTRFS_SUPER_INFO_SIZE as buffer size.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/image/main.c b/image/main.c
index 7c499c0853d0..162a578a3ff8 100644
--- a/image/main.c
+++ b/image/main.c
@@ -1522,9 +1522,14 @@ static int fill_mdres_info(struct mdrestore_struct *mdres,
 		return 0;
 
 	if (mdres->compress_method == COMPRESS_ZLIB) {
-		size_t size = MAX_PENDING_SIZE * 2;
+		/*
+		 * We know this item is superblock, its should only be 4K.
+		 * Don't need to waste memory following max_pending_size as it
+		 * can be as large as 256M.
+		 */
+		size_t size = BTRFS_SUPER_INFO_SIZE;
 
-		buffer = malloc(MAX_PENDING_SIZE * 2);
+		buffer = malloc(size);
 		if (!buffer)
 			return -ENOMEM;
 		ret = uncompress(buffer, (unsigned long *)&size,
@@ -1963,10 +1968,10 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 	}
 
 	if (mdres->compress_method == COMPRESS_ZLIB) {
-		size_t size = MAX_PENDING_SIZE * 2;
+		size_t size = BTRFS_SUPER_INFO_SIZE;
 		u8 *tmp;
 
-		tmp = malloc(MAX_PENDING_SIZE * 2);
+		tmp = malloc(size);
 		if (!tmp) {
 			free(buffer);
 			return -ENOMEM;
-- 
2.22.0

[PATCH v2.1 04/10] btrfs-progs: image: Allow restore to record system chunk ranges for later usage

[Qu Wenruo]

Currently we are doing a pretty slow search for system chunks before
restoring real data.
The current behavior is to search all clusters for chunk tree root
first, then search all clusters again and again for every chunk tree
block.

This causes recursive calls and pretty slow start up, the only good news
is since chunk tree are normally small, we don't need to iterate too
many times, thus overall it's acceptable.

To address such bad behavior, we could take usage of system chunk array
in the super block.
By recording all system chunks ranges, we could easily determine if an
extent belongs to chunk tree, thus do one loop simple linear search for
chunk tree leaves.

This patch only introduces the code base for later patches.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 103 +++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 103 insertions(+)

diff --git a/image/main.c b/image/main.c
index 162a578a3ff8..0a68ea8e70d9 100644
--- a/image/main.c
+++ b/image/main.c
@@ -35,6 +35,7 @@
 #include "common/utils.h"
 #include "volumes.h"
 #include "extent_io.h"
+#include "extent-cache.h"
 #include "common/help.h"
 #include "common/device-utils.h"
 #include "image/metadump.h"
@@ -93,6 +94,11 @@ struct mdrestore_struct {
 	pthread_mutex_t mutex;
 	pthread_cond_t cond;
 
+	/*
+	 * Records system chunk ranges, so restore can use this to determine
+	 * if an item is in chunk tree range.
+	 */
+	struct cache_tree sys_chunks;
 	struct rb_root chunk_tree;
 	struct rb_root physical_tree;
 	struct list_head list;
@@ -102,6 +108,8 @@ struct mdrestore_struct {
 	u64 devid;
 	u64 alloced_chunks;
 	u64 last_physical_offset;
+	/* An quicker checker for if a item is in sys chunk range */
+	u64 sys_chunk_end;
 	u8 uuid[BTRFS_UUID_SIZE];
 	u8 fsid[BTRFS_FSID_SIZE];
 
@@ -1456,6 +1464,7 @@ static void mdrestore_destroy(struct mdrestore_struct *mdres, int num_threads)
 		rb_erase(&entry->p, &mdres->physical_tree);
 		free(entry);
 	}
+	free_extent_cache_tree(&mdres->sys_chunks);
 	pthread_mutex_lock(&mdres->mutex);
 	mdres->done = 1;
 	pthread_cond_broadcast(&mdres->cond);
@@ -1480,6 +1489,7 @@ static int mdrestore_init(struct mdrestore_struct *mdres,
 	pthread_mutex_init(&mdres->mutex, NULL);
 	INIT_LIST_HEAD(&mdres->list);
 	INIT_LIST_HEAD(&mdres->overlapping_chunks);
+	cache_tree_init(&mdres->sys_chunks);
 	mdres->in = in;
 	mdres->out = out;
 	mdres->old_restore = old_restore;
@@ -1903,6 +1913,92 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 	return ret;
 }
 
+/*
+ * Add system chunks in super blocks into mdres->sys_chunks, so later
+ * we can determine if an item is a chunk tree block.
+ */
+static int add_sys_array(struct mdrestore_struct *mdres,
+			 struct btrfs_super_block *sb)
+{
+	struct btrfs_disk_key *disk_key;
+	struct btrfs_key key;
+	struct btrfs_chunk *chunk;
+	struct cache_extent *cache;
+	u32 cur_offset;
+	u32 len = 0;
+	u32 array_size;
+	u8 *array_ptr;
+	int ret = 0;
+
+	array_size = btrfs_super_sys_array_size(sb);
+	array_ptr = sb->sys_chunk_array;
+	cur_offset = 0;
+
+	while (cur_offset < array_size) {
+		u32 num_stripes;
+
+		disk_key = (struct btrfs_disk_key *)array_ptr;
+		len = sizeof(*disk_key);
+		if (cur_offset + len > array_size)
+			goto out_short_read;
+		btrfs_disk_key_to_cpu(&key, disk_key);
+
+		array_ptr += len;
+		cur_offset += len;
+
+		if (key.type == BTRFS_CHUNK_ITEM_KEY) {
+			chunk = (struct btrfs_chunk *)array_ptr;
+
+			/*
+			 * At least one btrfs_chunk with one stripe must be
+			 * present, exact stripe count check comes afterwards
+			 */
+			len = btrfs_chunk_item_size(1);
+			if (cur_offset + len > array_size)
+				goto out_short_read;
+			num_stripes = btrfs_stack_chunk_num_stripes(chunk);
+			if (!num_stripes) {
+				printk(
+	    "ERROR: invalid number of stripes %u in sys_array at offset %u\n",
+					num_stripes, cur_offset);
+				ret = -EIO;
+				break;
+			}
+			len = btrfs_chunk_item_size(num_stripes);
+			if (cur_offset + len > array_size)
+				goto out_short_read;
+			if (btrfs_stack_chunk_type(chunk) &
+			    BTRFS_BLOCK_GROUP_SYSTEM) {
+				ret = add_merge_cache_extent(&mdres->sys_chunks,
+					key.offset,
+					btrfs_stack_chunk_length(chunk));
+				if (ret < 0)
+					break;
+			}
+		} else {
+			error("unexpected item type %u in sys_array offset %u",
+			      key.type, cur_offset);
+			ret = -EUCLEAN;
+			break;
+		}
+		array_ptr += len;
+		cur_offset += len;
+	}
+
+	/* Get the last system chunk end as a quicker check */
+	cache = last_cache_extent(&mdres->sys_chunks);
+	if (!cache) {
+		error("no system chunk found in super block");
+		return -EUCLEAN;
+	}
+	mdres->sys_chunk_end = cache->start + cache->size - 1;
+	return ret;
+out_short_read:
+	error("sys_array too short to read %u bytes at offset %u\n",
+		len, cur_offset);
+	return -EUCLEAN;
+}
+
 static int build_chunk_tree(struct mdrestore_struct *mdres,
 			    struct meta_cluster *cluster)
 {
@@ -1995,6 +2091,13 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 		error("invalid superblock");
 		return ret;
 	}
+	ret = add_sys_array(mdres, super);
+	if (ret < 0) {
+		error("failed to read system chunk array");
+		free(buffer);
+		pthread_mutex_unlock(&mdres->mutex);
+		return ret;
+	}
 	chunk_root_bytenr = btrfs_super_chunk_root(super);
 	mdres->nodesize = btrfs_super_nodesize(super);
 	if (btrfs_super_incompat_flags(super) &
-- 
2.22.0

[PATCH v2.1 05/10] btrfs-progs: image: Introduce helper to determine if a tree block is in the range of system chunks

[Qu Wenruo]

Introduce a new helper function, is_in_sys_chunks(), to determine if an
item is in the range of system chunks.

Since btrfs-image will merge adjacent same type extents into one item,
this function is designed to return true for any bytes in system chunk
range.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 48 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 48 insertions(+)

diff --git a/image/main.c b/image/main.c
index 0a68ea8e70d9..2ceaada7784b 100644
--- a/image/main.c
+++ b/image/main.c
@@ -1658,6 +1658,54 @@ static int wait_for_worker(struct mdrestore_struct *mdres)
 	return ret;
 }
 
+/*
+ * Check if a range [start ,start + len] has ANY bytes covered by
+ * system chunks ranges.
+ */
+static bool is_in_sys_chunks(struct mdrestore_struct *mdres, u64 start,
+			     u64 len)
+{
+	struct rb_node *node = mdres->sys_chunks.root.rb_node;
+	struct cache_extent *entry;
+	struct cache_extent *next;
+	struct cache_extent *prev;
+
+	if (start > mdres->sys_chunk_end)
+		return false;
+
+	while (node) {
+		entry = rb_entry(node, struct cache_extent, rb_node);
+		if (start > entry->start) {
+			if (!node->rb_right)
+				break;
+			node = node->rb_right;
+		} else if (start < entry->start) {
+			if (!node->rb_left)
+				break;
+			node = node->rb_left;
+		} else {
+			/* already in a system chunk */
+			return true;
+		}
+	}
+	if (!node)
+		return false;
+	entry = rb_entry(node, struct cache_extent, rb_node);
+	/* Now we have entry which is the nearst chunk around @start */
+	if (start > entry->start) {
+		prev = entry;
+		next = next_cache_extent(entry);
+	} else {
+		prev = prev_cache_extent(entry);
+		next = entry;
+	}
+	if (prev && prev->start + prev->size > start)
+		return true;
+	if (next && start + len > next->start)
+		return true;
+	return false;
+}
+
 static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
 			    u64 bytenr, u64 item_bytenr, u32 bufsize,
 			    u64 cluster_bytenr)
-- 
2.22.0

[PATCH v2.1 06/10] btrfs-progs: image: Rework how we search chunk tree blocks

[Qu Wenruo]

Before this patch, we were using a very inefficient way to search
chunks:

We iterate through all clusters to find the chunk root tree block first,
then re-iterate all clusters again to find every child tree blocks.

Every time we need to iterate all clusters just to find a chunk tree
block.
This is obviously inefficient, specially when chunk tree get larger.
So the original author leaves a comment on it:
  /* If you have to ask you aren't worthy */
  static int search_for_chunk_blocks()

This patch will change the behavior so that we will only iterate all
clusters once.

The idea behind the optimization is, since we have the superblock
restored first, we could use the CHUNK_ITEMs in
super_block::sys_chunk_array to build a SYSTEM chunk mapping.

Then when we start to iterate through all items, we can easily skip
unrelated items at different level:
- At cluster level
  If a cluster starts beyond last system chunk map, it must not contain
  any chunk tree blocks (as chunk tree blocks only lives inside system
  chunks)

- At item level
  If one item has no intersection with any system chunk map, then it
  must not contain any tree blocks.

By this, we can iterate through all clusters just once, and find out all
CHUNK_ITEMs.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 214 +++++++++++++++++++++++++++------------------------
 1 file changed, 115 insertions(+), 99 deletions(-)

diff --git a/image/main.c b/image/main.c
index 2ceaada7784b..b59ae137076a 100644
--- a/image/main.c
+++ b/image/main.c
@@ -123,8 +123,6 @@ struct mdrestore_struct {
 	struct btrfs_fs_info *info;
 };
 
-static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
-				   u64 search, u64 cluster_bytenr);
 static struct extent_buffer *alloc_dummy_eb(u64 bytenr, u32 size);
 
 static void csum_block(u8 *buf, size_t len)
@@ -1706,67 +1704,17 @@ static bool is_in_sys_chunks(struct mdrestore_struct *mdres, u64 start,
 	return false;
 }
 
-static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
-			    u64 bytenr, u64 item_bytenr, u32 bufsize,
-			    u64 cluster_bytenr)
+static int read_chunk_tree_block(struct mdrestore_struct *mdres,
+				 struct extent_buffer *eb)
 {
-	struct extent_buffer *eb;
-	int ret = 0;
 	int i;
 
-	eb = alloc_dummy_eb(bytenr, mdres->nodesize);
-	if (!eb) {
-		ret = -ENOMEM;
-		goto out;
-	}
-
-	while (item_bytenr != bytenr) {
-		buffer += mdres->nodesize;
-		item_bytenr += mdres->nodesize;
-	}
-
-	memcpy(eb->data, buffer, mdres->nodesize);
-	if (btrfs_header_bytenr(eb) != bytenr) {
-		error("eb bytenr does not match found bytenr: %llu != %llu",
-				(unsigned long long)btrfs_header_bytenr(eb),
-				(unsigned long long)bytenr);
-		ret = -EIO;
-		goto out;
-	}
-
-	if (memcmp(mdres->fsid, eb->data + offsetof(struct btrfs_header, fsid),
-		   BTRFS_FSID_SIZE)) {
-		error("filesystem metadata UUID of eb %llu does not match",
-				(unsigned long long)bytenr);
-		ret = -EIO;
-		goto out;
-	}
-
-	if (btrfs_header_owner(eb) != BTRFS_CHUNK_TREE_OBJECTID) {
-		error("wrong eb %llu owner %llu",
-				(unsigned long long)bytenr,
-				(unsigned long long)btrfs_header_owner(eb));
-		ret = -EIO;
-		goto out;
-	}
-
 	for (i = 0; i < btrfs_header_nritems(eb); i++) {
 		struct btrfs_chunk *chunk;
 		struct fs_chunk *fs_chunk;
 		struct btrfs_key key;
 		u64 type;
 
-		if (btrfs_header_level(eb)) {
-			u64 blockptr = btrfs_node_blockptr(eb, i);
-
-			ret = search_for_chunk_blocks(mdres, blockptr,
-						      cluster_bytenr);
-			if (ret)
-				break;
-			continue;
-		}
-
-		/* Yay a leaf!  We loves leafs! */
 		btrfs_item_key_to_cpu(eb, &key, i);
 		if (key.type != BTRFS_CHUNK_ITEM_KEY)
 			continue;
@@ -1774,8 +1722,7 @@ static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
 		fs_chunk = malloc(sizeof(struct fs_chunk));
 		if (!fs_chunk) {
 			error("not enough memory to allocate chunk");
-			ret = -ENOMEM;
-			break;
+			return -ENOMEM;
 		}
 		memset(fs_chunk, 0, sizeof(*fs_chunk));
 		chunk = btrfs_item_ptr(eb, i, struct btrfs_chunk);
@@ -1784,19 +1731,18 @@ static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
 		fs_chunk->physical = btrfs_stripe_offset_nr(eb, chunk, 0);
 		fs_chunk->bytes = btrfs_chunk_length(eb, chunk);
 		INIT_LIST_HEAD(&fs_chunk->list);
+
 		if (tree_search(&mdres->physical_tree, &fs_chunk->p,
 				physical_cmp, 1) != NULL)
 			list_add(&fs_chunk->list, &mdres->overlapping_chunks);
 		else
 			tree_insert(&mdres->physical_tree, &fs_chunk->p,
 				    physical_cmp);
-
 		type = btrfs_chunk_type(eb, chunk);
 		if (type & BTRFS_BLOCK_GROUP_DUP) {
 			fs_chunk->physical_dup =
 					btrfs_stripe_offset_nr(eb, chunk, 1);
 		}
-
 		if (fs_chunk->physical_dup + fs_chunk->bytes >
 		    mdres->last_physical_offset)
 			mdres->last_physical_offset = fs_chunk->physical_dup +
@@ -1811,19 +1757,80 @@ static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
 			mdres->alloced_chunks += fs_chunk->bytes;
 		tree_insert(&mdres->chunk_tree, &fs_chunk->l, chunk_cmp);
 	}
-out:
+	return 0;
+}
+
+static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
+			    u64 item_bytenr, u32 bufsize,
+			    u64 cluster_bytenr)
+{
+	struct extent_buffer *eb;
+	u32 nodesize = mdres->nodesize;
+	u64 bytenr;
+	size_t cur_offset;
+	int ret = 0;
+
+	eb = alloc_dummy_eb(0, mdres->nodesize);
+	if (!eb)
+		return -ENOMEM;
+
+	for (cur_offset = 0; cur_offset < bufsize; cur_offset += nodesize) {
+		bytenr = item_bytenr + cur_offset;
+		if (!is_in_sys_chunks(mdres, bytenr, nodesize))
+			continue;
+		memcpy(eb->data, buffer + cur_offset, nodesize);
+		if (btrfs_header_bytenr(eb) != bytenr) {
+			error(
+			"eb bytenr does not match found bytenr: %llu != %llu",
+				(unsigned long long)btrfs_header_bytenr(eb),
+				(unsigned long long)bytenr);
+			ret = -EUCLEAN;
+			break;
+		}
+		if (memcmp(mdres->fsid, eb->data +
+			   offsetof(struct btrfs_header, fsid),
+		    BTRFS_FSID_SIZE)) {
+			error(
+			"filesystem metadata UUID of eb %llu does not match",
+				bytenr);
+			ret = -EUCLEAN;
+			break;
+		}
+		if (btrfs_header_owner(eb) != BTRFS_CHUNK_TREE_OBJECTID) {
+			error("wrong eb %llu owner %llu",
+				(unsigned long long)bytenr,
+				(unsigned long long)btrfs_header_owner(eb));
+			ret = -EUCLEAN;
+			break;
+		}
+		/*
+		 * No need to search node, as we will iterate all tree blocks
+		 * in chunk tree, only need to bother leaves.
+		 */
+		if (btrfs_header_level(eb))
+			continue;
+		ret = read_chunk_tree_block(mdres, eb);
+		if (ret < 0)
+			break;
+	}
 	free(eb);
 	return ret;
 }
 
-/* If you have to ask you aren't worthy */
-static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
-				   u64 search, u64 cluster_bytenr)
+/*
+ * This function will try to find all chunk items in the dump image.
+ *
+ * This function will iterate all clusters, and find any item inside
+ * system chunk ranges.
+ * For such item, it will try to read them as tree blocks, and find
+ * CHUNK_ITEMs, add them to @mdres.
+ */
+static int search_for_chunk_blocks(struct mdrestore_struct *mdres)
 {
 	struct meta_cluster *cluster;
 	struct meta_cluster_header *header;
 	struct meta_cluster_item *item;
-	u64 current_cluster = cluster_bytenr, bytenr;
+	u64 current_cluster = 0, bytenr;
 	u64 item_bytenr;
 	u32 bufsize, nritems, i;
 	u32 max_size = MAX_PENDING_SIZE * 2;
@@ -1854,30 +1861,27 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 	}
 
 	bytenr = current_cluster;
+	/* Main loop, iterating all clusters */
 	while (1) {
 		if (fseek(mdres->in, current_cluster, SEEK_SET)) {
 			error("seek failed: %m");
 			ret = -EIO;
-			break;
+			goto out;
 		}
 
 		ret = fread(cluster, BLOCK_SIZE, 1, mdres->in);
 		if (ret == 0) {
-			if (cluster_bytenr != 0) {
-				cluster_bytenr = 0;
-				current_cluster = 0;
-				bytenr = 0;
-				continue;
-			}
+			if (feof(mdres->in))
+				goto out;
 			error(
 	"unknown state after reading cluster at %llu, probably corrupted data",
-					cluster_bytenr);
+					current_cluster);
 			ret = -EIO;
-			break;
+			goto out;
 		} else if (ret < 0) {
 			error("unable to read image at %llu: %m",
-					(unsigned long long)cluster_bytenr);
-			break;
+					current_cluster);
+			goto out;
 		}
 		ret = 0;
 
@@ -1886,11 +1890,17 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 		    le64_to_cpu(header->bytenr) != current_cluster) {
 			error("bad header in metadump image");
 			ret = -EIO;
-			break;
+			goto out;
 		}
 
+		/* We're already over the system chunk end, no need to search*/
+		if (current_cluster > mdres->sys_chunk_end)
+			goto out;
+
 		bytenr += BLOCK_SIZE;
 		nritems = le32_to_cpu(header->nritems);
+
+		/* Search items for tree blocks in sys chunks */
 		for (i = 0; i < nritems; i++) {
 			size_t size;
 
@@ -1898,11 +1908,23 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 			bufsize = le32_to_cpu(item->size);
 			item_bytenr = le64_to_cpu(item->bytenr);
 
-			if (bufsize > max_size) {
-				error("item %u too big: %u > %u", i, bufsize,
-						max_size);
-				ret = -EIO;
-				break;
+			/*
+			 * Only data extent/free space cache can be that big,
+			 * adjacent tree blocks won't be able to be merged
+			 * beyond max_size.
+			 * Also, we can skip super block.
+			 */
+			if (bufsize > max_size ||
+			    !is_in_sys_chunks(mdres, item_bytenr, bufsize) ||
+			    item_bytenr == BTRFS_SUPER_INFO_OFFSET) {
+				ret = fseek(mdres->in, bufsize, SEEK_CUR);
+				if (ret < 0) {
+					error("failed to seek: %m");
+					ret = -errno;
+					goto out;
+				}
+				bytenr += bufsize;
+				continue;
 			}
 
 			if (mdres->compress_method == COMPRESS_ZLIB) {
@@ -1910,7 +1932,7 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 				if (ret != 1) {
 					error("read error: %m");
 					ret = -EIO;
-					break;
+					goto out;
 				}
 
 				size = max_size;
@@ -1921,40 +1943,36 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres,
 					error("decompression failed with %d",
 							ret);
 					ret = -EIO;
-					break;
+					goto out;
 				}
 			} else {
 				ret = fread(buffer, bufsize, 1, mdres->in);
 				if (ret != 1) {
 					error("read error: %m");
 					ret = -EIO;
-					break;
+					goto out;
 				}
 				size = bufsize;
 			}
 			ret = 0;
 
-			if (item_bytenr <= search &&
-			    item_bytenr + size > search) {
-				ret = read_chunk_block(mdres, buffer, search,
-						       item_bytenr, size,
-						       current_cluster);
-				if (!ret)
-					ret = 1;
-				break;
+			ret = read_chunk_block(mdres, buffer,
+					       item_bytenr, size,
+					       current_cluster);
+			if (ret < 0) {
+				error(
+	"failed to search tree blocks in item bytenr %llu size %lu",
+					item_bytenr, size);
+				goto out;
 			}
 			bytenr += bufsize;
 		}
-		if (ret) {
-			if (ret > 0)
-				ret = 0;
-			break;
-		}
 		if (bytenr & BLOCK_MASK)
 			bytenr += BLOCK_SIZE - (bytenr & BLOCK_MASK);
 		current_cluster = bytenr;
 	}
 
+out:
 	free(tmp);
 	free(buffer);
 	free(cluster);
@@ -2053,7 +2071,6 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 	struct btrfs_super_block *super;
 	struct meta_cluster_header *header;
 	struct meta_cluster_item *item = NULL;
-	u64 chunk_root_bytenr = 0;
 	u32 i, nritems;
 	u64 bytenr = 0;
 	u8 *buffer;
@@ -2146,7 +2163,6 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 		pthread_mutex_unlock(&mdres->mutex);
 		return ret;
 	}
-	chunk_root_bytenr = btrfs_super_chunk_root(super);
 	mdres->nodesize = btrfs_super_nodesize(super);
 	if (btrfs_super_incompat_flags(super) &
 	    BTRFS_FEATURE_INCOMPAT_METADATA_UUID)
@@ -2159,7 +2175,7 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 	free(buffer);
 	pthread_mutex_unlock(&mdres->mutex);
 
-	return search_for_chunk_blocks(mdres, chunk_root_bytenr, 0);
+	return search_for_chunk_blocks(mdres);
 }
 
 static int range_contains_super(u64 physical, u64 bytes)
-- 
2.22.0

[PATCH v2.1 07/10] btrfs-progs: image: Introduce framework for more dump versions

[Qu Wenruo]

The original dump format only contains a @magic member to verify the
format, this means if we want to introduce new on-disk format or change
certain size limit, we can only introduce new magic as kind of version.

This patch will introduce the framework to allow multiple magic to
co-exist for further functions.

This patch will introduce the following members for each dump version.

- max_pending_size
  The threshold size for an cluster. It's not a hard limit but a soft
  one. One cluster can go larger than max_pending_size for one item, but
  next item would go to next cluster.

- magic_cpu
  The magic number in CPU endian.

- extra_sb_flags
  If the super block of this restore needs extra super block flags like
  BTRFS_SUPER_FLAG_METADUMP_V2.
  For incoming data dump feature, we don't need any extra super block
  flags.

This change also implies that all image dumps will use the same magic
for all clusters. No mixing is allowed, as we will use the first cluster
to determine the dump version.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c     | 72 ++++++++++++++++++++++++++++++++++++++++++------
 image/metadump.h | 13 +++++++--
 2 files changed, 74 insertions(+), 11 deletions(-)

diff --git a/image/main.c b/image/main.c
index b59ae137076a..a82d21b01b46 100644
--- a/image/main.c
+++ b/image/main.c
@@ -44,6 +44,19 @@
 
 #define MAX_WORKER_THREADS	(32)
 
+const struct dump_version dump_versions[NR_DUMP_VERSIONS] = {
+	/*
+	 * The original format, which only supports tree blocks and
+	 * free space cache dump.
+	 */
+	{ .version = 0,
+	  .max_pending_size = SZ_256K,
+	  .magic_cpu = 0xbd5c25e27295668bULL,
+	  .extra_sb_flags = 1 }
+};
+
+const struct dump_version *current_version = &dump_versions[0];
+
 struct async_work {
 	struct list_head list;
 	struct list_head ordered;
@@ -403,7 +416,7 @@ static void meta_cluster_init(struct metadump_struct *md, u64 start)
 	md->num_items = 0;
 	md->num_ready = 0;
 	header = &md->cluster.header;
-	header->magic = cpu_to_le64(HEADER_MAGIC);
+	header->magic = cpu_to_le64(current_version->magic_cpu);
 	header->bytenr = cpu_to_le64(start);
 	header->nritems = cpu_to_le32(0);
 	header->compress = md->compress_level > 0 ?
@@ -715,7 +728,7 @@ static int add_extent(u64 start, u64 size, struct metadump_struct *md,
 {
 	int ret;
 	if (md->data != data ||
-	    md->pending_size + size > MAX_PENDING_SIZE ||
+	    md->pending_size + size > current_version->max_pending_size ||
 	    md->pending_start + md->pending_size != start) {
 		ret = flush_pending(md, 0);
 		if (ret)
@@ -1044,7 +1057,8 @@ static void update_super_old(u8 *buffer)
 	u32 sectorsize = btrfs_super_sectorsize(super);
 	u64 flags = btrfs_super_flags(super);
 
-	flags |= BTRFS_SUPER_FLAG_METADUMP;
+	if (current_version->extra_sb_flags)
+		flags |= BTRFS_SUPER_FLAG_METADUMP;
 	btrfs_set_super_flags(super, flags);
 
 	key = (struct btrfs_disk_key *)(super->sys_chunk_array);
@@ -1137,7 +1151,8 @@ static int update_super(struct mdrestore_struct *mdres, u8 *buffer)
 	if (mdres->clear_space_cache)
 		btrfs_set_super_cache_generation(super, 0);
 
-	flags |= BTRFS_SUPER_FLAG_METADUMP_V2;
+	if (current_version->extra_sb_flags)
+		flags |= BTRFS_SUPER_FLAG_METADUMP_V2;
 	btrfs_set_super_flags(super, flags);
 	btrfs_set_super_sys_array_size(super, new_array_size);
 	btrfs_set_super_num_devices(super, 1);
@@ -1325,7 +1340,7 @@ static void *restore_worker(void *data)
 	u8 *outbuf;
 	int outfd;
 	int ret;
-	int compress_size = MAX_PENDING_SIZE * 4;
+	int compress_size = current_version->max_pending_size * 4;
 
 	outfd = fileno(mdres->out);
 	buffer = malloc(compress_size);
@@ -1475,6 +1490,42 @@ static void mdrestore_destroy(struct mdrestore_struct *mdres, int num_threads)
 	pthread_mutex_destroy(&mdres->mutex);
 }
 
+static int detect_version(FILE *in)
+{
+	struct meta_cluster *cluster;
+	u8 buf[BLOCK_SIZE];
+	bool found = false;
+	int i;
+	int ret;
+
+	if (fseek(in, 0, SEEK_SET) < 0) {
+		error("seek failed: %m");
+		return -errno;
+	}
+	ret = fread(buf, BLOCK_SIZE, 1, in);
+	if (!ret) {
+		error("failed to read header");
+		return -EIO;
+	}
+
+	fseek(in, 0, SEEK_SET);
+	cluster = (struct meta_cluster *)buf;
+	for (i = 0; i < NR_DUMP_VERSIONS; i++) {
+		if (le64_to_cpu(cluster->header.magic) ==
+		    dump_versions[i].magic_cpu) {
+			found = true;
+			current_version = &dump_versions[i];
+			break;
+		}
+	}
+
+	if (!found) {
+		error("unrecognized header format");
+		return -EINVAL;
+	}
+	return 0;
+}
+
 static int mdrestore_init(struct mdrestore_struct *mdres,
 			  FILE *in, FILE *out, int old_restore,
 			  int num_threads, int fixup_offset,
@@ -1482,6 +1533,9 @@ static int mdrestore_init(struct mdrestore_struct *mdres,
 {
 	int i, ret = 0;
 
+	ret = detect_version(in);
+	if (ret < 0)
+		return ret;
 	memset(mdres, 0, sizeof(*mdres));
 	pthread_cond_init(&mdres->cond, NULL);
 	pthread_mutex_init(&mdres->mutex, NULL);
@@ -1833,7 +1887,7 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres)
 	u64 current_cluster = 0, bytenr;
 	u64 item_bytenr;
 	u32 bufsize, nritems, i;
-	u32 max_size = MAX_PENDING_SIZE * 2;
+	u32 max_size = current_version->max_pending_size * 2;
 	u8 *buffer, *tmp = NULL;
 	int ret = 0;
 
@@ -1886,7 +1940,7 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres)
 		ret = 0;
 
 		header = &cluster->header;
-		if (le64_to_cpu(header->magic) != HEADER_MAGIC ||
+		if (le64_to_cpu(header->magic) != current_version->magic_cpu ||
 		    le64_to_cpu(header->bytenr) != current_cluster) {
 			error("bad header in metadump image");
 			ret = -EIO;
@@ -2088,7 +2142,7 @@ static int build_chunk_tree(struct mdrestore_struct *mdres,
 	ret = 0;
 
 	header = &cluster->header;
-	if (le64_to_cpu(header->magic) != HEADER_MAGIC ||
+	if (le64_to_cpu(header->magic) != current_version->magic_cpu ||
 	    le64_to_cpu(header->bytenr) != 0) {
 		error("bad header in metadump image");
 		return -EIO;
@@ -2597,7 +2651,7 @@ static int restore_metadump(const char *input, FILE *out, int old_restore,
 			break;
 
 		header = &cluster->header;
-		if (le64_to_cpu(header->magic) != HEADER_MAGIC ||
+		if (le64_to_cpu(header->magic) != current_version->magic_cpu ||
 		    le64_to_cpu(header->bytenr) != bytenr) {
 			error("bad header in metadump image");
 			ret = -EIO;
diff --git a/image/metadump.h b/image/metadump.h
index f85c9bcfb813..941d4b827a24 100644
--- a/image/metadump.h
+++ b/image/metadump.h
@@ -22,8 +22,6 @@
 #include "kernel-lib/list.h"
 #include "ctree.h"
 
-#define HEADER_MAGIC		0xbd5c25e27295668bULL
-#define MAX_PENDING_SIZE	SZ_256K
 #define BLOCK_SIZE		SZ_1K
 #define BLOCK_MASK		(BLOCK_SIZE - 1)
 
@@ -33,6 +31,17 @@
 #define COMPRESS_NONE		0
 #define COMPRESS_ZLIB		1
 
+struct dump_version {
+	u64 magic_cpu;
+	int version;
+	int max_pending_size;
+	unsigned int extra_sb_flags:1;
+};
+
+#define NR_DUMP_VERSIONS	1
+extern const struct dump_version dump_versions[NR_DUMP_VERSIONS];
+const extern struct dump_version *current_version;
+
 struct meta_cluster_item {
 	__le64 bytenr;
 	__le32 size;
-- 
2.22.0

[PATCH v2.1 08/10] btrfs-progs: image: Introduce -d option to dump data

[Qu Wenruo]

This new data dump feature will dump the whole image, not long the
existing tree blocks but also all its data extents(*).

This feature will rely on the new dump format (_DUmP_v1), as it needs
extra large extent size limit, and older btrfs-image dump can't handle
such large item/cluster size.

Since we're dumping all extents including data extents, for the restored
image there is no need to use any extra super block flags to inform
kernel.
Kernel should just treat the restored image as any ordinary btrfs.

*: The data extents will be dumped as is, that's to say, even for
preallocated extent, its (meaningless) data will be read out and
dumpped.
This behavior will cause extra space usage for the image, but we can
skip all the complex partially shared preallocated extent check.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c     | 53 +++++++++++++++++++++++++++++++++++++-----------
 image/metadump.h |  2 +-
 2 files changed, 42 insertions(+), 13 deletions(-)

diff --git a/image/main.c b/image/main.c
index a82d21b01b46..922a046272a8 100644
--- a/image/main.c
+++ b/image/main.c
@@ -52,7 +52,15 @@ const struct dump_version dump_versions[NR_DUMP_VERSIONS] = {
 	{ .version = 0,
 	  .max_pending_size = SZ_256K,
 	  .magic_cpu = 0xbd5c25e27295668bULL,
-	  .extra_sb_flags = 1 }
+	  .extra_sb_flags = 1 },
+	/*
+	 * The newer format, with much larger item size to contain
+	 * any data extent.
+	 */
+	{ .version = 1,
+	  .max_pending_size = SZ_256M,
+	  .magic_cpu = 0x31765f506d55445fULL, /* ascii _DUmP_v1, no null */
+	  .extra_sb_flags = 0 },
 };
 
 const struct dump_version *current_version = &dump_versions[0];
@@ -452,10 +460,14 @@ static void metadump_destroy(struct metadump_struct *md, int num_threads)
 
 static int metadump_init(struct metadump_struct *md, struct btrfs_root *root,
 			 FILE *out, int num_threads, int compress_level,
-			 enum sanitize_mode sanitize_names)
+			 bool dump_data, enum sanitize_mode sanitize_names)
 {
 	int i, ret = 0;
 
+	/* We need larger item/cluster limit for data extents */
+	if (dump_data)
+		current_version = &dump_versions[1];
+
 	memset(md, 0, sizeof(*md));
 	INIT_LIST_HEAD(&md->list);
 	INIT_LIST_HEAD(&md->ordered);
@@ -883,7 +895,7 @@ static int copy_space_cache(struct btrfs_root *root,
 }
 
 static int copy_from_extent_tree(struct metadump_struct *metadump,
-				 struct btrfs_path *path)
+				 struct btrfs_path *path, bool dump_data)
 {
 	struct btrfs_root *extent_root;
 	struct extent_buffer *leaf;
@@ -948,9 +960,15 @@ static int copy_from_extent_tree(struct metadump_struct *metadump,
 			ei = btrfs_item_ptr(leaf, path->slots[0],
 					    struct btrfs_extent_item);
 			if (btrfs_extent_flags(leaf, ei) &
-			    BTRFS_EXTENT_FLAG_TREE_BLOCK) {
+			    BTRFS_EXTENT_FLAG_TREE_BLOCK ||
+			    btrfs_extent_flags(leaf, ei) &
+			    BTRFS_EXTENT_FLAG_DATA) {
+				bool is_data;
+
+				is_data = btrfs_extent_flags(leaf, ei) &
+					  BTRFS_EXTENT_FLAG_DATA;
 				ret = add_extent(bytenr, num_bytes, metadump,
-						 0);
+						 is_data);
 				if (ret) {
 					error("unable to add block %llu: %d",
 						(unsigned long long)bytenr, ret);
@@ -973,7 +991,7 @@ static int copy_from_extent_tree(struct metadump_struct *metadump,
 
 static int create_metadump(const char *input, FILE *out, int num_threads,
 			   int compress_level, enum sanitize_mode sanitize,
-			   int walk_trees)
+			   int walk_trees, bool dump_data)
 {
 	struct btrfs_root *root;
 	struct btrfs_path path;
@@ -988,7 +1006,7 @@ static int create_metadump(const char *input, FILE *out, int num_threads,
 	}
 
 	ret = metadump_init(&metadump, root, out, num_threads,
-			    compress_level, sanitize);
+			    compress_level, dump_data, sanitize);
 	if (ret) {
 		error("failed to initialize metadump: %d", ret);
 		close_ctree(root);
@@ -1020,7 +1038,7 @@ static int create_metadump(const char *input, FILE *out, int num_threads,
 			goto out;
 		}
 	} else {
-		ret = copy_from_extent_tree(&metadump, &path);
+		ret = copy_from_extent_tree(&metadump, &path, dump_data);
 		if (ret) {
 			err = ret;
 			goto out;
@@ -2813,6 +2831,7 @@ static void print_usage(int ret)
 	printf("\t-s      \tsanitize file names, use once to just use garbage, use twice if you want crc collisions\n");
 	printf("\t-w      \twalk all trees instead of using extent tree, do this if your extent tree is broken\n");
 	printf("\t-m	   \trestore for multiple devices\n");
+	printf("\t-d	   \talso dump data, conflicts with -w\n");
 	printf("\n");
 	printf("\tIn the dump mode, source is the btrfs device and target is the output file (use '-' for stdout).\n");
 	printf("\tIn the restore mode, source is the dumped image and target is the btrfs device/file.\n");
@@ -2832,6 +2851,7 @@ int BOX_MAIN(image)(int argc, char *argv[])
 	int ret;
 	enum sanitize_mode sanitize = SANITIZE_NONE;
 	int dev_cnt = 0;
+	bool dump_data = false;
 	int usage_error = 0;
 	FILE *out;
 
@@ -2840,7 +2860,7 @@ int BOX_MAIN(image)(int argc, char *argv[])
 			{ "help", no_argument, NULL, GETOPT_VAL_HELP},
 			{ NULL, 0, NULL, 0 }
 		};
-		int c = getopt_long(argc, argv, "rc:t:oswm", long_options, NULL);
+		int c = getopt_long(argc, argv, "rc:t:oswmd", long_options, NULL);
 		if (c < 0)
 			break;
 		switch (c) {
@@ -2880,6 +2900,9 @@ int BOX_MAIN(image)(int argc, char *argv[])
 			create = 0;
 			multi_devices = 1;
 			break;
+		case 'd':
+			dump_data = true;
+			break;
 		case GETOPT_VAL_HELP:
 		default:
 			print_usage(c != GETOPT_VAL_HELP);
@@ -2898,10 +2921,15 @@ int BOX_MAIN(image)(int argc, char *argv[])
 			"create and restore cannot be used at the same time");
 			usage_error++;
 		}
+		if (dump_data && walk_trees) {
+			error("-d conflicts with -f option");
+			usage_error++;
+		}
 	} else {
-		if (walk_trees || sanitize != SANITIZE_NONE || compress_level) {
+		if (walk_trees || sanitize != SANITIZE_NONE || compress_level ||
+		    dump_data) {
 			error(
-			"using -w, -s, -c options for restore makes no sense");
+		"using -w, -s, -c, -d options for restore makes no sense");
 			usage_error++;
 		}
 		if (multi_devices && dev_cnt < 2) {
@@ -2954,7 +2982,8 @@ int BOX_MAIN(image)(int argc, char *argv[])
 		}
 
 		ret = create_metadump(source, out, num_threads,
-				      compress_level, sanitize, walk_trees);
+				      compress_level, sanitize, walk_trees,
+				      dump_data);
 	} else {
 		ret = restore_metadump(source, out, old_restore, num_threads,
 				       0, target, multi_devices);
diff --git a/image/metadump.h b/image/metadump.h
index 941d4b827a24..a04f63a910d6 100644
--- a/image/metadump.h
+++ b/image/metadump.h
@@ -38,7 +38,7 @@ struct dump_version {
 	unsigned int extra_sb_flags:1;
 };
 
-#define NR_DUMP_VERSIONS	1
+#define NR_DUMP_VERSIONS	2
 extern const struct dump_version dump_versions[NR_DUMP_VERSIONS];
 const extern struct dump_version *current_version;
 
-- 
2.22.0

Re: [PATCH v2.1 08/10] btrfs-progs: image: Introduce -d option to dump data

[Anand Jain]

> +		if (dump_data && walk_trees) {
> +			error("-d conflicts with -f option");

  should be ..with -w option

Thanks, Anand

Re: [PATCH v2.1 08/10] btrfs-progs: image: Introduce -d option to dump data

[Anand Jain]

On 4/7/19 2:11 PM, Qu Wenruo wrote:
>   static int copy_from_extent_tree(struct metadump_struct *metadump,
> -				 struct btrfs_path *path)
> +				 struct btrfs_path *path, bool dump_data)
>   {
>   	struct btrfs_root *extent_root;
>   	struct extent_buffer *leaf;
> @@ -948,9 +960,15 @@ static int copy_from_extent_tree(struct metadump_struct *metadump,
>   			ei = btrfs_item_ptr(leaf, path->slots[0],
>   					    struct btrfs_extent_item);
>   			if (btrfs_extent_flags(leaf, ei) &
> -			    BTRFS_EXTENT_FLAG_TREE_BLOCK) {
> +			    BTRFS_EXTENT_FLAG_TREE_BLOCK ||
> +			    btrfs_extent_flags(leaf, ei) &
> +			    BTRFS_EXTENT_FLAG_DATA) {
> +				bool is_data;
> +
> +				is_data = btrfs_extent_flags(leaf, ei) &
> +					  BTRFS_EXTENT_FLAG_DATA;
>   				ret = add_extent(bytenr, num_bytes, metadump,
> -						 0);
> +						 is_data);


   Both with and without -d option copies the data.
   Should check dump_data.

Thanks, Anand

[PATCH v2.1 09/10] btrfs-progs: image: Reduce memory requirement for decompression

[Qu Wenruo]

With recent change to enlarge max_pending_size to 256M for data dump,
the decompress code requires quite a lot of memory space. (256M * 4).

The main reason behind it is, we're using wrapped uncompress() function
call, which needs the buffer to be large enough to contain the
decompressed data.

This patch will re-work the decompress work to use inflate() which can
resume it decompression so that we can use a much smaller buffer size.

This patch choose to use 512K buffer size.

Now the memory consumption for restore is reduced to
 Cluster data size + 512K * nr_running_threads

Instead of the original one:
 Cluster data size + 1G * nr_running_threads

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 220 +++++++++++++++++++++++++++++++++------------------
 1 file changed, 145 insertions(+), 75 deletions(-)

diff --git a/image/main.c b/image/main.c
index 922a046272a8..9ac921659c8b 100644
--- a/image/main.c
+++ b/image/main.c
@@ -1349,128 +1349,198 @@ static void write_backup_supers(int fd, u8 *buf)
 	}
 }
 
-static void *restore_worker(void *data)
+/*
+ * Restore one item.
+ *
+ * For uncompressed data, it's just reading from work->buf then write to output.
+ * For compressed data, since we can have very large decompressed data
+ * (up to 256M), we need to consider memory usage. So here we will fill buffer
+ * then write the decompressed buffer to output.
+ */
+static int restore_one_work(struct mdrestore_struct *mdres,
+			    struct async_work *async, u8 *buffer, int bufsize)
 {
-	struct mdrestore_struct *mdres = (struct mdrestore_struct *)data;
-	struct async_work *async;
-	size_t size;
-	u8 *buffer;
-	u8 *outbuf;
-	int outfd;
+	z_stream strm;
+	int buf_offset = 0;	/* offset inside work->buffer */
+	int out_offset = 0;	/* offset for output */
+	int out_len;
+	int outfd = fileno(mdres->out);
+	int compress_method = mdres->compress_method;
 	int ret;
-	int compress_size = current_version->max_pending_size * 4;
 
-	outfd = fileno(mdres->out);
-	buffer = malloc(compress_size);
-	if (!buffer) {
-		error("not enough memory for restore worker buffer");
-		pthread_mutex_lock(&mdres->mutex);
-		if (!mdres->error)
-			mdres->error = -ENOMEM;
-		pthread_mutex_unlock(&mdres->mutex);
-		pthread_exit(NULL);
+	ASSERT(is_power_of_2(bufsize));
+
+	if (compress_method == COMPRESS_ZLIB) {
+		strm.zalloc = Z_NULL;
+		strm.zfree = Z_NULL;
+		strm.opaque = Z_NULL;
+		strm.avail_in = async->bufsize;
+		strm.next_in = async->buffer;
+		strm.avail_out = 0;
+		strm.next_out = Z_NULL;
+		ret = inflateInit(&strm);
+		if (ret != Z_OK) {
+			error("failed to initialize decompress parameters: %d",
+				ret);
+			return ret;
+		}
 	}
+	while (buf_offset < async->bufsize) {
+		bool compress_end = false;
+		int read_size = min_t(u64, async->bufsize - buf_offset,
+				      bufsize);
 
-	while (1) {
-		u64 bytenr, physical_dup;
-		off_t offset = 0;
-		int err = 0;
-
-		pthread_mutex_lock(&mdres->mutex);
-		while (!mdres->nodesize || list_empty(&mdres->list)) {
-			if (mdres->done) {
-				pthread_mutex_unlock(&mdres->mutex);
-				goto out;
+		/* Read part */
+		if (compress_method == COMPRESS_ZLIB) {
+			if (strm.avail_out == 0) {
+				strm.avail_out = bufsize;
+				strm.next_out = buffer;
 			}
-			pthread_cond_wait(&mdres->cond, &mdres->mutex);
-		}
-		async = list_entry(mdres->list.next, struct async_work, list);
-		list_del_init(&async->list);
-
-		if (mdres->compress_method == COMPRESS_ZLIB) {
-			size = compress_size;
 			pthread_mutex_unlock(&mdres->mutex);
-			ret = uncompress(buffer, (unsigned long *)&size,
-					 async->buffer, async->bufsize);
+			ret = inflate(&strm, Z_NO_FLUSH);
 			pthread_mutex_lock(&mdres->mutex);
-			if (ret != Z_OK) {
-				error("decompression failed with %d", ret);
-				err = -EIO;
+			switch (ret) {
+			case Z_NEED_DICT:
+				ret = Z_DATA_ERROR; /* fallthrough */
+				__attribute__ ((fallthrough));
+			case Z_DATA_ERROR:
+			case Z_MEM_ERROR:
+				goto out;
+			}
+			if (ret == Z_STREAM_END) {
+				ret = 0;
+				compress_end = true;
 			}
-			outbuf = buffer;
+			out_len = bufsize - strm.avail_out;
 		} else {
-			outbuf = async->buffer;
-			size = async->bufsize;
+			/* No compress, read as many data as possible */
+			memcpy(buffer, async->buffer + buf_offset, read_size);
+
+			buf_offset += read_size;
+			out_len = read_size;
 		}
 
+		/* Fixup part */
 		if (!mdres->multi_devices) {
 			if (async->start == BTRFS_SUPER_INFO_OFFSET) {
 				if (mdres->old_restore) {
-					update_super_old(outbuf);
+					update_super_old(buffer);
 				} else {
-					ret = update_super(mdres, outbuf);
-					if (ret)
-						err = ret;
+					ret = update_super(mdres, buffer);
+					if (ret < 0) 
+						goto out;
 				}
 			} else if (!mdres->old_restore) {
-				ret = fixup_chunk_tree_block(mdres, async, outbuf, size);
+				ret = fixup_chunk_tree_block(mdres, async,
+							     buffer, out_len);
 				if (ret)
-					err = ret;
+					goto out;
 			}
 		}
 
+		/* Write part */
 		if (!mdres->fixup_offset) {
+			int size = out_len;
+			off_t offset = 0;
+
 			while (size) {
+				u64 logical = async->start + out_offset + offset;
 				u64 chunk_size = size;
-				physical_dup = 0;
+				u64 physical_dup = 0;
+				u64 bytenr;
+
 				if (!mdres->multi_devices && !mdres->old_restore)
 					bytenr = logical_to_physical(mdres,
-						     async->start + offset,
-						     &chunk_size,
-						     &physical_dup);
+							logical, &chunk_size,
+							&physical_dup);
 				else
-					bytenr = async->start + offset;
+					bytenr = logical;
 
-				ret = pwrite64(outfd, outbuf+offset, chunk_size,
-					       bytenr);
+				ret = pwrite64(outfd, buffer + offset, chunk_size, bytenr);
 				if (ret != chunk_size)
-					goto error;
+					goto write_error;
 
 				if (physical_dup)
-					ret = pwrite64(outfd, outbuf+offset,
-						       chunk_size,
-						       physical_dup);
+					ret = pwrite64(outfd, buffer + offset,
+						       chunk_size, physical_dup);
 				if (ret != chunk_size)
-					goto error;
+					goto write_error;
 
 				size -= chunk_size;
 				offset += chunk_size;
 				continue;
-
-error:
-				if (ret < 0) {
-					error("unable to write to device: %m");
-					err = errno;
-				} else {
-					error("short write");
-					err = -EIO;
-				}
 			}
 		} else if (async->start != BTRFS_SUPER_INFO_OFFSET) {
-			ret = write_data_to_disk(mdres->info, outbuf, async->start, size, 0);
+			ret = write_data_to_disk(mdres->info, buffer,
+						 async->start, out_len, 0);
 			if (ret) {
 				error("failed to write data");
 				exit(1);
 			}
 		}
 
-
 		/* backup super blocks are already there at fixup_offset stage */
-		if (!mdres->multi_devices && async->start == BTRFS_SUPER_INFO_OFFSET)
-			write_backup_supers(outfd, outbuf);
+		if (async->start == BTRFS_SUPER_INFO_OFFSET &&
+		    !mdres->multi_devices)
+			write_backup_supers(outfd, buffer);
+		out_offset += out_len;
+		if (compress_end) {
+			inflateEnd(&strm);
+			break;
+		}
+	}
+	return ret;
+
+write_error:
+	if (ret < 0) {
+		error("unable to write to device: %m");
+		ret = -errno;
+	} else {
+		error("short write");
+		ret = -EIO;
+	}
+out:
+	if (compress_method == COMPRESS_ZLIB)
+		inflateEnd(&strm);
+	return ret;
+}
+
+static void *restore_worker(void *data)
+{
+	struct mdrestore_struct *mdres = (struct mdrestore_struct *)data;
+	struct async_work *async;
+	u8 *buffer;
+	int ret;
+	int buffer_size = SZ_512K;
+
+	buffer = malloc(buffer_size);
+	if (!buffer) {
+		error("not enough memory for restore worker buffer");
+		pthread_mutex_lock(&mdres->mutex);
+		if (!mdres->error)
+			mdres->error = -ENOMEM;
+		pthread_mutex_unlock(&mdres->mutex);
+		pthread_exit(NULL);
+	}
+
+	while (1) {
+		pthread_mutex_lock(&mdres->mutex);
+		while (!mdres->nodesize || list_empty(&mdres->list)) {
+			if (mdres->done) {
+				pthread_mutex_unlock(&mdres->mutex);
+				goto out;
+			}
+			pthread_cond_wait(&mdres->cond, &mdres->mutex);
+		}
+		async = list_entry(mdres->list.next, struct async_work, list);
+		list_del_init(&async->list);
 
-		if (err && !mdres->error)
-			mdres->error = err;
+		ret = restore_one_work(mdres, async, buffer, buffer_size);
+		if (ret < 0) {
+			mdres->error = ret;
+			pthread_mutex_unlock(&mdres->mutex);
+			goto out;
+		}
 		mdres->num_items--;
 		pthread_mutex_unlock(&mdres->mutex);
 
-- 
2.22.0

[PATCH v2.1 10/10] btrfs-progs: image: Reduce memory usage for chunk tree search

[Qu Wenruo]

Just like original restore_worker, search_for_chunk_blocks() will also
use a lot of memory for restoring large uncompressed extent or
compressed extent with data dump.

Reduce the memory usage by:
- Use fixed buffer size for uncompressed extent
  Now we will use fixed 512K as buffer size for reading uncompressed
  extent.

  Now chunk tree search will read out 512K data at most, then search
  chunk trees in the buffer.

  Reduce the memory usage from as large as item size, to fixed 512K.

- Use inflate() for compressed extent
  For compressed extent, we need two buffers, one for compressed data,
  and one for uncompressed data.
  For compressed data, we will use the item size as buffer size, since
  compressed extent should be small enough.
  For uncompressed data, we use 512K as buffer size.

  Now chunk tree search will fill the first 512K, then search chunk
  trees blocks in the uncompressed 512K buffer, then loop until the
  compressed data is exhausted.

  Reduce the memory usage from as large as 256M * 2 to 512K + compressed
  extent size.

Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 image/main.c | 159 ++++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 126 insertions(+), 33 deletions(-)

diff --git a/image/main.c b/image/main.c
index 9ac921659c8b..32767eed100b 100644
--- a/image/main.c
+++ b/image/main.c
@@ -1959,6 +1959,126 @@ static int read_chunk_block(struct mdrestore_struct *mdres, u8 *buffer,
 	return ret;
 }
 
+static int search_chunk_uncompressed(struct mdrestore_struct *mdres,
+				     struct meta_cluster_item *item,
+				     u64 current_cluster)
+{
+	u32 item_size = le32_to_cpu(item->size);
+	u64 item_bytenr = le64_to_cpu(item->bytenr);
+	int bufsize = SZ_512K;
+	int read_size;
+	u32 offset = 0;
+	u8 *buffer;
+	int ret = 0;
+
+	ASSERT(mdres->compress_method == COMPRESS_NONE);
+	buffer = malloc(bufsize);
+	if (!buffer)
+		return -ENOMEM;
+
+	while (offset < item_size) {
+		read_size = min_t(u32, bufsize, item_size - offset);
+		ret = fread(buffer, read_size, 1, mdres->in);
+		if (ret != 1) {
+			error("read error: %m");
+			ret = -EIO;
+			goto out;
+		}
+		ret = read_chunk_block(mdres, buffer, item_bytenr, read_size,
+					current_cluster);
+		if (ret < 0) {
+			error(
+	"failed to search tree blocks in item bytenr %llu size %u",
+				item_bytenr, item_size);
+			goto out;
+		}
+		offset += read_size;
+	}
+out:
+	free(buffer);
+	return ret;
+}
+
+static int search_chunk_compressed(struct mdrestore_struct *mdres,
+				   struct meta_cluster_item *item,
+				   u64 current_cluster)
+{
+	z_stream strm;
+	u32 item_size = le32_to_cpu(item->size);
+	u64 item_bytenr = le64_to_cpu(item->bytenr);
+	int bufsize = SZ_512K;
+	int read_size;
+	u8 *out_buf = NULL;	/* uncompressed data */
+	u8 *in_buf = NULL;	/* compressed data */
+	bool end = false;
+	int ret;
+
+	ASSERT(mdres->compress_method != COMPRESS_NONE);
+	strm.zalloc = Z_NULL;
+	strm.zfree = Z_NULL;
+	strm.opaque = Z_NULL;
+	strm.avail_in = 0;
+	strm.next_in = Z_NULL;
+	strm.avail_out = 0;
+	strm.next_out = Z_NULL;
+	ret = inflateInit(&strm);
+	if (ret != Z_OK) {
+		error("failed to initialize decompress parameters: %d", ret);
+		return ret;
+	}
+
+	out_buf = malloc(bufsize);
+	in_buf = malloc(item_size);
+	if (!in_buf || !out_buf) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	ret = fread(in_buf, item_size, 1, mdres->in);
+	if (ret != 1) {
+		error("read error: %m");
+		ret = -EIO;
+		goto out;
+	}
+	strm.avail_in = item_size;
+	strm.next_in = in_buf;
+	while (!end) {
+		if (strm.avail_out == 0) {
+			strm.avail_out = bufsize;
+			strm.next_out = out_buf;
+		}
+		ret = inflate(&strm, Z_NO_FLUSH);
+		switch (ret) {
+		case Z_NEED_DICT:
+			ret = Z_DATA_ERROR; /* fallthrough */
+			__attribute__ ((fallthrough));
+		case Z_DATA_ERROR:
+		case Z_MEM_ERROR:
+			goto out;
+		}
+		if (ret == Z_STREAM_END) {
+			ret = 0;
+			end = true;
+		}
+		read_size = bufsize - strm.avail_out;
+
+		ret = read_chunk_block(mdres, out_buf, item_bytenr, read_size,
+					current_cluster);
+		if (ret < 0) {
+			error(
+	"failed to search tree blocks in item bytenr %llu size %u",
+				item_bytenr, item_size);
+			goto out;
+		}
+	}
+
+out:
+	free(in_buf);
+	free(out_buf);
+	inflateEnd(&strm);
+	return ret;
+}
+
 /*
  * This function will try to find all chunk items in the dump image.
  *
@@ -2044,8 +2164,6 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres)
 
 		/* Search items for tree blocks in sys chunks */
 		for (i = 0; i < nritems; i++) {
-			size_t size;
-
 			item = &cluster->items[i];
 			bufsize = le32_to_cpu(item->size);
 			item_bytenr = le64_to_cpu(item->bytenr);
@@ -2070,41 +2188,16 @@ static int search_for_chunk_blocks(struct mdrestore_struct *mdres)
 			}
 
 			if (mdres->compress_method == COMPRESS_ZLIB) {
-				ret = fread(tmp, bufsize, 1, mdres->in);
-				if (ret != 1) {
-					error("read error: %m");
-					ret = -EIO;
-					goto out;
-				}
-
-				size = max_size;
-				ret = uncompress(buffer,
-						 (unsigned long *)&size, tmp,
-						 bufsize);
-				if (ret != Z_OK) {
-					error("decompression failed with %d",
-							ret);
-					ret = -EIO;
-					goto out;
-				}
+				ret = search_chunk_compressed(mdres, item,
+						current_cluster);
 			} else {
-				ret = fread(buffer, bufsize, 1, mdres->in);
-				if (ret != 1) {
-					error("read error: %m");
-					ret = -EIO;
-					goto out;
-				}
-				size = bufsize;
+				ret = search_chunk_uncompressed(mdres, item,
+						current_cluster);
 			}
-			ret = 0;
-
-			ret = read_chunk_block(mdres, buffer,
-					       item_bytenr, size,
-					       current_cluster);
 			if (ret < 0) {
 				error(
-	"failed to search tree blocks in item bytenr %llu size %lu",
-					item_bytenr, size);
+	"failed to search tree blocks in item bytenr %llu size %u",
+					item_bytenr, bufsize);
 				goto out;
 			}
 			bytenr += bufsize;
-- 
2.22.0

Re: [PATCH v2.1 00/10] btrfs-progs: image: Enhancement with new data dump feature

[WenRuo Qu]

[-- Attachment #1.1: Type: text/plain, Size: 3688 bytes --]

Gentle ping?

This feature should be pretty useful for both full fs backup and debug
purpose.

Any feedback?
It can still be applied to latest stable branch without any conflict.

Thanks,
Qu

On 2019/7/4 下午2:10, Qu Wenruo wrote:
> This updated patchset is rebased on devel branch, whose HEAD is:
> commit 73289664917ad7c73f3f3393593e93a118228ad8 (david/devel)
> Author: David Sterba <dsterba@suse.com>
> Date:   Thu Jul 4 01:42:15 2019 +0200
> 
>     btrfs-progs: kerncompat: define __always_inline conditionally
> 
> This updated patchset includes the following features:
> 
> - Small fixes to improve btrfs-image error report
>   * Output error message for chunk tree build error
>   * Fix error output to show correct return value
>   Patch 1 and 2.
> 
> - Reduce memory usage when decompress super block
>   Independent change, for current btrfs-image, it will reduce buffer
>   size from 256K to fixed 4K.
>   Patch 3.
> 
> - Rework how we search chunk tree blocks
>   Instead of iterating clusters again and again for each chunk tree
>   block, record system chunk array and iterate clusters once for all
>   chunk tree blocks.
>   This should reduce restore time for large image dump.
>   Patch 4, 5 and 6.
> 
> - Introduce data dump feature to dump the whole fs.
>   This will introduce a new magic number to prevent old btrfs-image to
>   hit failure as the item size limit is enlarged.
>   Patch 7 and 8.
> 
> - Reduce memory usage for data dump restore
>   This is mostly due to the fact that we have much larger
>   max_pending_size introduced by data dump(256K -> 256M).
>   Using 4 * max_pending_size for each decompress thread as buffer is way
>   too expensive now.
>   Use proper inflate() to replace uncompress() calls.
>   Patch 9 and 10.
> 
> Changelog:
> v2:
> - New small fixes:
>   * Fix a confusing error message due to unpopulated errno
>   * Output error message for chunk tree build error
>   
> - Fix a regression of previous version
>   Patch "btrfs-progs: image: Rework how we search chunk tree blocks"
>   deleted a "ret = 0" line which could cause false early exit.
> 
> - Reduce memory usage for data dump
> 
> v2.1:
> - Rebased to devel branch
>   Removing 4 already merged patches from the patchset.
> 
> - Re-order the patchset
>   Put small and independent patches at the top of queue, and put the
>   data dump related feature at the end.
> 
> - Fix -Wmaybe-uninitialized warnings
>   Strangely, D=1 won't trigger these warnings thus they sneak into v2
>   without being detected.
> 
> - Fix FROM: line
>   Reverted to old smtp setup. The new setup will override FROM: line,
>   messing up the name of author.
> 
> Qu Wenruo (10):
>   btrfs-progs: image: Output error message for chunk tree build error
>   btrfs-progs: image: Fix error output to show correct return value
>   btrfs-progs: image: Don't waste memory when we're just extracting
>     super block
>   btrfs-progs: image: Allow restore to record system chunk ranges for
>     later usage
>   btrfs-progs: image: Introduce helper to determine if a tree block is
>     in the range of system chunks
>   btrfs-progs: image: Rework how we search chunk tree blocks
>   btrfs-progs: image: Introduce framework for more dump versions
>   btrfs-progs: image: Introduce -d option to dump data
>   btrfs-progs: image: Reduce memory requirement for decompression
>   btrfs-progs: image: Reduce memory usage for chunk tree search
> 
>  image/main.c     | 866 +++++++++++++++++++++++++++++++++++------------
>  image/metadump.h |  13 +-
>  2 files changed, 654 insertions(+), 225 deletions(-)
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH v2.1 00/10] btrfs-progs: image: Enhancement with new data dump feature

[Anand Jain]

On 9/19/19 3:19 PM, WenRuo Qu wrote:
> Gentle ping?
> 
> This feature should be pretty useful for both full fs backup and debug
> purpose.
> 
> Any feedback?

  Did you miss my review? There are bugs in the patch.

Thanks, Anand

> It can still be applied to latest stable branch without any conflict.
> 
> Thanks,
> Qu

Re: [PATCH v2.1 00/10] btrfs-progs: image: Enhancement with new data dump feature

[Qu Wenruo]

[-- Attachment #1.1: Type: text/plain, Size: 537 bytes --]



On 2019/9/24 下午7:31, Anand Jain wrote:
> On 9/19/19 3:19 PM, WenRuo Qu wrote:
>> Gentle ping?
>>
>> This feature should be pretty useful for both full fs backup and debug
>> purpose.
>>
>> Any feedback?
> 
>  Did you miss my review? There are bugs in the patch.

Oh, my bad.

Indeed there is a bug.

I'll update the patchset soon.

Thanks for reminding me about that.
Qu

> 
> Thanks, Anand
> 
>> It can still be applied to latest stable branch without any conflict.
>>
>> Thanks,
>> Qu
> 
> 
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]