[PATCH V7 0/4] erofs-utils: compressed fragments feature

[PATCH V7 0/4] erofs-utils: compressed fragments feature

[Yue Hu]

From: Yue Hu <huyue2@coolpad.com>

In order to achieve greater compression ratio, let's introduce
compressed fragments feature which can merge tail of per-file or the
whole files into one special inode to reach the target.

And we can also set pcluster size to fragments inode for different
compression requirments.

In this patchset, we also improve the uncompressed data layout of
compressed files. Just write it from 'clusterofs' instead of 0 since it
can benefit from in-place I/O. For now, it only goes with fragments.

The main idea above is from Xiang.

Here is some test data of Linux 5.10.87 source code under Ubuntu 18.04:

linux-5.10.87 (erofs, uncompressed)                1.1G

linux-5.10.87 (erofs, lz4hc,12 4k fragments,4k)    301M
linux-5.10.87 (erofs, lz4hc,12 8k fragments,8k)    268M
linux-5.10.87 (erofs, lz4hc,12 16k fragments,16k)  242M
linux-5.10.87 (erofs, lz4hc,12 32k fragments,32k)  225M
linux-5.10.87 (erofs, lz4hc,12 64k fragments,64k)  217M

linux-5.10.87 (erofs, lz4hc,12 4k vanilla)         396M
linux-5.10.87 (erofs, lz4hc,12 8k vanilla)         376M
linux-5.10.87 (erofs, lz4hc,12 16k vanilla)        364M
linux-5.10.87 (erofs, lz4hc,12 32k vanilla)        359M
linux-5.10.87 (erofs, lz4hc,12 64k vanilla)        358M

Usage:
mkfs.erofs -zlz4hc,12 -C65536 -Efragments,65536 foo.erofs.img foo/

Changes since v6/v5:
 - mainly optimize the packed inode related pointed out in v6 by Xiang (v6)
 - mainly optimize interlaced uncompress data code (v5)

Changes since v4:
 - naming on-disk offset as interlaced offset and add a advise to hint the data
   layout (Gao Xiang);
 - improve the packed inode pcluster size checking condition (Gao Xiang);
 - add field is_packed_inode instead of using is_src parameter passed (Gao Xiang);
 - per-inode eof bug_on (Gao Xiang);
 - extract the interlaced data layout part as a seprate patch;
 - change the advise bit of fragment feature.
 - re-implement the function to read interlaced data;
 - update compresstion ratio check;

Changes since v3:
 - fuse minor change and modify fragments naming related suggested by Xiang;
 - refator fragments build code;
 - rebase to latest commit 547bea3cb71a.

Changes since v2:
 - mainly reimplment the decompression logic for fragment inode due to
   kernel side;
 - fix compatibility issue to old image with ztailpacking feature;
 - move code of super.c in patch 3/3 to patch 1/3;
 - minor naming change.

Changes since v1:
 - mainly optimize index space for fragment inode;
 - add merging tail with len <= pclustersize into fragments directly;
 - use a inode instead of nid to avoid multiple load fragments;
 - fix memory leak of building fragments;
 - minor change to diff special fragments with normal inode.
 - rebase to commit cb058526 with patch [1];
 - code cleanup.

Note that inode will be extended version (64 bytes) due to mtime, may
use 'force-inode-compact' option to reduce the size if mtime careless.

[1] https://lore.kernel.org/linux-erofs/20220722053610.23912-1-huyue2@coolpad.com/

Yue Hu (4):
  erofs-utils: fuse: support interlaced uncompressed pcluster
  erofs-utils: lib: support fragments data decompression
  erofs-utils: mkfs: support interlaced uncompressed data layout
  erofs-utils: mkfs: introduce compressed fragments support

 include/erofs/compress.h   |  8 +++-
 include/erofs/config.h     |  3 +-
 include/erofs/decompress.h |  3 ++
 include/erofs/fragments.h  | 28 ++++++++++++
 include/erofs/inode.h      |  1 +
 include/erofs/internal.h   |  9 ++++
 include/erofs_fs.h         | 28 +++++++++---
 lib/Makefile.am            |  4 +-
 lib/compress.c             | 90 ++++++++++++++++++++++++++++----------
 lib/data.c                 | 37 +++++++++++++++-
 lib/decompress.c           | 17 ++++++-
 lib/fragments.c            | 59 +++++++++++++++++++++++++
 lib/inode.c                | 47 ++++++++++++++++++--
 lib/super.c                |  1 +
 lib/zmap.c                 | 26 +++++++++++
 mkfs/main.c                | 61 +++++++++++++++++++++++---
 16 files changed, 378 insertions(+), 44 deletions(-)
 create mode 100644 include/erofs/fragments.h
 create mode 100644 lib/fragments.c

-- 
2.17.1

[PATCH V7 1/4] erofs-utils: fuse: support interlaced uncompressed pcluster

[Yue Hu]

From: Yue Hu <huyue2@coolpad.com>

Support uncompressed data layout with on-disk interlaced offset in
compression mode for erofsfuse.

Signed-off-by: Yue Hu <huyue2@coolpad.com>
---
 include/erofs/decompress.h |  3 +++
 include/erofs_fs.h         |  2 ++
 lib/data.c                 |  8 +++++++-
 lib/decompress.c           | 17 +++++++++++++++--
 4 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/include/erofs/decompress.h b/include/erofs/decompress.h
index 82bf7b8..a9067cb 100644
--- a/include/erofs/decompress.h
+++ b/include/erofs/decompress.h
@@ -23,6 +23,9 @@ struct z_erofs_decompress_req {
 	unsigned int decodedskip;
 	unsigned int inputsize, decodedlength;
 
+	/* cut point of interlaced uncompressed data */
+	unsigned int interlaced_offset;
+
 	/* indicate the algorithm will be used for decompression */
 	unsigned int alg;
 	bool partial_decoding;
diff --git a/include/erofs_fs.h b/include/erofs_fs.h
index 08f9761..b8a7421 100644
--- a/include/erofs_fs.h
+++ b/include/erofs_fs.h
@@ -294,11 +294,13 @@ struct z_erofs_lzma_cfgs {
  * bit 1 : HEAD1 big pcluster (0 - off; 1 - on)
  * bit 2 : HEAD2 big pcluster (0 - off; 1 - on)
  * bit 3 : tailpacking inline pcluster (0 - off; 1 - on)
+ * bit 4 : interlaced plain pcluster (0 - off; 1 - on)
  */
 #define Z_EROFS_ADVISE_COMPACTED_2B		0x0001
 #define Z_EROFS_ADVISE_BIG_PCLUSTER_1		0x0002
 #define Z_EROFS_ADVISE_BIG_PCLUSTER_2		0x0004
 #define Z_EROFS_ADVISE_INLINE_PCLUSTER		0x0008
+#define Z_EROFS_ADVISE_INTERLACED_PCLUSTER	0x0010
 
 struct z_erofs_map_header {
 	__le16	h_reserved1;
diff --git a/lib/data.c b/lib/data.c
index ad7b2cb..2d76816 100644
--- a/lib/data.c
+++ b/lib/data.c
@@ -226,7 +226,7 @@ static int z_erofs_read_data(struct erofs_inode *inode, char *buffer,
 	};
 	struct erofs_map_dev mdev;
 	bool partial;
-	unsigned int bufsize = 0;
+	unsigned int bufsize = 0, interlaced_offset;
 	char *raw = NULL;
 	int ret = 0;
 
@@ -287,10 +287,16 @@ static int z_erofs_read_data(struct erofs_inode *inode, char *buffer,
 		if (ret < 0)
 			break;
 
+		interlaced_offset = 0;
+		if ((inode->z_advise & Z_EROFS_ADVISE_INTERLACED_PCLUSTER) &&
+		    map.m_algorithmformat == Z_EROFS_COMPRESSION_SHIFTED)
+			interlaced_offset = erofs_blkoff(map.m_la);
+
 		ret = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 					.in = raw,
 					.out = buffer + end - offset,
 					.decodedskip = skip,
+					.interlaced_offset = interlaced_offset,
 					.inputsize = map.m_plen,
 					.decodedlength = length,
 					.alg = map.m_algorithmformat,
diff --git a/lib/decompress.c b/lib/decompress.c
index 1661f91..39da555 100644
--- a/lib/decompress.c
+++ b/lib/decompress.c
@@ -132,14 +132,27 @@ out:
 int z_erofs_decompress(struct z_erofs_decompress_req *rq)
 {
 	if (rq->alg == Z_EROFS_COMPRESSION_SHIFTED) {
+		unsigned int count, rightpart, skip;
+
 		if (rq->inputsize > EROFS_BLKSIZ)
 			return -EFSCORRUPTED;
 
 		DBG_BUGON(rq->decodedlength > EROFS_BLKSIZ);
 		DBG_BUGON(rq->decodedlength < rq->decodedskip);
 
-		memcpy(rq->out, rq->in + rq->decodedskip,
-		       rq->decodedlength - rq->decodedskip);
+		count = rq->decodedlength - rq->decodedskip;
+
+		if (rq->interlaced_offset) {
+			skip = erofs_blkoff(rq->interlaced_offset +
+					    rq->decodedskip);
+			rightpart = min(EROFS_BLKSIZ - skip, count);
+		} else {
+			skip = rq->decodedskip;
+			rightpart = count;
+		}
+		memcpy(rq->out, rq->in + skip, rightpart);
+		memcpy(rq->out + rightpart, rq->in, count - rightpart);
+
 		return 0;
 	}
 
-- 
2.17.1

[PATCH V7 2/4] erofs-utils: lib: support fragments data decompression

[Yue Hu]

From: Yue Hu <huyue2@coolpad.com>

Add compressed fragments support for erofsfuse.

Signed-off-by: Yue Hu <huyue2@coolpad.com>
---
 include/erofs/internal.h |  6 ++++++
 include/erofs_fs.h       | 26 ++++++++++++++++++++------
 lib/data.c               | 29 +++++++++++++++++++++++++++++
 lib/super.c              |  1 +
 lib/zmap.c               | 26 ++++++++++++++++++++++++++
 5 files changed, 82 insertions(+), 6 deletions(-)

diff --git a/include/erofs/internal.h b/include/erofs/internal.h
index 2e0aae8..742bfed 100644
--- a/include/erofs/internal.h
+++ b/include/erofs/internal.h
@@ -102,6 +102,7 @@ struct erofs_sb_info {
 		u16 devt_slotoff;		/* used for mkfs */
 		u16 device_id_mask;		/* used for others */
 	};
+	erofs_nid_t packed_nid;
 };
 
 /* global sbi */
@@ -132,6 +133,7 @@ EROFS_FEATURE_FUNCS(big_pcluster, incompat, INCOMPAT_BIG_PCLUSTER)
 EROFS_FEATURE_FUNCS(chunked_file, incompat, INCOMPAT_CHUNKED_FILE)
 EROFS_FEATURE_FUNCS(device_table, incompat, INCOMPAT_DEVICE_TABLE)
 EROFS_FEATURE_FUNCS(ztailpacking, incompat, INCOMPAT_ZTAILPACKING)
+EROFS_FEATURE_FUNCS(fragments, incompat, INCOMPAT_FRAGMENTS)
 EROFS_FEATURE_FUNCS(sb_chksum, compat, COMPAT_SB_CHKSUM)
 
 #define EROFS_I_EA_INITED	(1 << 0)
@@ -209,6 +211,7 @@ struct erofs_inode {
 #ifdef WITH_ANDROID
 	uint64_t capabilities;
 #endif
+	erofs_off_t fragmentoff;
 };
 
 static inline bool is_inode_layout_compression(struct erofs_inode *inode)
@@ -279,6 +282,7 @@ enum {
 	BH_Mapped,
 	BH_Encoded,
 	BH_FullMapped,
+	BH_Fragment,
 };
 
 /* Has a disk mapping */
@@ -289,6 +293,8 @@ enum {
 #define EROFS_MAP_ENCODED	(1 << BH_Encoded)
 /* The length of extent is full */
 #define EROFS_MAP_FULL_MAPPED	(1 << BH_FullMapped)
+/* Located in the special packed inode */
+#define EROFS_MAP_FRAGMENT	(1 << BH_Fragment)
 
 struct erofs_map_blocks {
 	char mpage[EROFS_BLKSIZ];
diff --git a/include/erofs_fs.h b/include/erofs_fs.h
index b8a7421..e492ad9 100644
--- a/include/erofs_fs.h
+++ b/include/erofs_fs.h
@@ -25,13 +25,15 @@
 #define EROFS_FEATURE_INCOMPAT_CHUNKED_FILE	0x00000004
 #define EROFS_FEATURE_INCOMPAT_DEVICE_TABLE	0x00000008
 #define EROFS_FEATURE_INCOMPAT_ZTAILPACKING	0x00000010
+#define EROFS_FEATURE_INCOMPAT_FRAGMENTS	0x00000020
 #define EROFS_ALL_FEATURE_INCOMPAT		\
 	(EROFS_FEATURE_INCOMPAT_LZ4_0PADDING | \
 	 EROFS_FEATURE_INCOMPAT_COMPR_CFGS | \
 	 EROFS_FEATURE_INCOMPAT_BIG_PCLUSTER | \
 	 EROFS_FEATURE_INCOMPAT_CHUNKED_FILE | \
 	 EROFS_FEATURE_INCOMPAT_DEVICE_TABLE | \
-	 EROFS_FEATURE_INCOMPAT_ZTAILPACKING)
+	 EROFS_FEATURE_INCOMPAT_ZTAILPACKING | \
+	 EROFS_FEATURE_INCOMPAT_FRAGMENTS)
 
 #define EROFS_SB_EXTSLOT_SIZE	16
 
@@ -73,7 +75,9 @@ struct erofs_super_block {
 	} __packed u1;
 	__le16 extra_devices;	/* # of devices besides the primary device */
 	__le16 devt_slotoff;	/* startoff = devt_slotoff * devt_slotsize */
-	__u8 reserved2[38];
+	__u8 reserved[6];
+	__le64 packed_nid;	/* nid of the special packed inode */
+	__u8 reserved2[24];
 };
 
 /*
@@ -295,17 +299,26 @@ struct z_erofs_lzma_cfgs {
  * bit 2 : HEAD2 big pcluster (0 - off; 1 - on)
  * bit 3 : tailpacking inline pcluster (0 - off; 1 - on)
  * bit 4 : interlaced plain pcluster (0 - off; 1 - on)
+ * bit 5 : fragment pcluster (0 - off; 1 - on)
  */
 #define Z_EROFS_ADVISE_COMPACTED_2B		0x0001
 #define Z_EROFS_ADVISE_BIG_PCLUSTER_1		0x0002
 #define Z_EROFS_ADVISE_BIG_PCLUSTER_2		0x0004
 #define Z_EROFS_ADVISE_INLINE_PCLUSTER		0x0008
 #define Z_EROFS_ADVISE_INTERLACED_PCLUSTER	0x0010
+#define Z_EROFS_ADVISE_FRAGMENT_PCLUSTER	0x0020
 
+#define Z_EROFS_FRAGMENT_INODE_BIT		7
 struct z_erofs_map_header {
-	__le16	h_reserved1;
-	/* record the size of tailpacking data */
-	__le16  h_idata_size;
+	union {
+		/* direct addressing for fragment offset */
+		__le32	h_fragmentoff;
+		struct {
+			__le16  h_reserved1;
+			/* record the size of tailpacking data */
+			__le16	h_idata_size;
+		};
+	};
 	__le16	h_advise;
 	/*
 	 * bit 0-3 : algorithm type of head 1 (logical cluster type 01);
@@ -314,7 +327,8 @@ struct z_erofs_map_header {
 	__u8	h_algorithmtype;
 	/*
 	 * bit 0-2 : logical cluster bits - 12, e.g. 0 for 4096;
-	 * bit 3-7 : reserved.
+	 * bit 3-6 : reserved;
+	 * bit 7   : move the whole file into packed inode or not.
 	 */
 	__u8	h_clusterbits;
 };
diff --git a/lib/data.c b/lib/data.c
index 2d76816..8e85ae7 100644
--- a/lib/data.c
+++ b/lib/data.c
@@ -275,6 +275,35 @@ static int z_erofs_read_data(struct erofs_inode *inode, char *buffer,
 			continue;
 		}
 
+		if (map.m_flags & EROFS_MAP_FRAGMENT) {
+			struct erofs_inode packed_inode = {
+				.nid = sbi.packed_nid,
+			};
+			char *out;
+
+			ret = erofs_read_inode_from_disk(&packed_inode);
+			if (ret) {
+				erofs_err("failed to read packed inode from disk");
+				return ret;
+			}
+
+			out = malloc(length - skip);
+			if (!out) {
+				ret = -ENOMEM;
+				break;
+			}
+			ret = z_erofs_read_data(&packed_inode, out,
+						length - skip,
+						inode->fragmentoff + skip);
+			if (ret < 0) {
+				free(out);
+				break;
+			}
+			memcpy(buffer + end - offset, out, length - skip);
+			free(out);
+			continue;
+		}
+
 		if (map.m_plen > bufsize) {
 			bufsize = map.m_plen;
 			raw = realloc(raw, bufsize);
diff --git a/lib/super.c b/lib/super.c
index b267412..30aeb36 100644
--- a/lib/super.c
+++ b/lib/super.c
@@ -104,6 +104,7 @@ int erofs_read_superblock(void)
 	sbi.xattr_blkaddr = le32_to_cpu(dsb->xattr_blkaddr);
 	sbi.islotbits = EROFS_ISLOTBITS;
 	sbi.root_nid = le16_to_cpu(dsb->root_nid);
+	sbi.packed_nid = le64_to_cpu(dsb->packed_nid);
 	sbi.inos = le64_to_cpu(dsb->inos);
 	sbi.checksum = le32_to_cpu(dsb->checksum);
 
diff --git a/lib/zmap.c b/lib/zmap.c
index abe0d31..f9d8d5f 100644
--- a/lib/zmap.c
+++ b/lib/zmap.c
@@ -83,6 +83,20 @@ static int z_erofs_fill_inode_lazy(struct erofs_inode *vi)
 		if (ret < 0)
 			return ret;
 	}
+	if (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER) {
+		vi->fragmentoff = le32_to_cpu(h->h_fragmentoff);
+
+		if (h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT) {
+			vi->z_tailextent_headlcn = 0;
+		} else {
+			struct erofs_map_blocks map = { .index = UINT_MAX };
+
+			ret = z_erofs_do_map_blocks(vi, &map,
+						    EROFS_GET_BLOCKS_FINDTAIL);
+			if (ret < 0)
+				return ret;
+		}
+	}
 	vi->flags |= EROFS_I_Z_INITED;
 	return 0;
 }
@@ -546,6 +560,7 @@ static int z_erofs_do_map_blocks(struct erofs_inode *vi,
 				 int flags)
 {
 	bool ztailpacking = vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER;
+	bool inpacked = vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
 	struct z_erofs_maprecorder m = {
 		.inode = vi,
 		.map = map,
@@ -609,6 +624,9 @@ static int z_erofs_do_map_blocks(struct erofs_inode *vi,
 		map->m_flags |= EROFS_MAP_META;
 		map->m_pa = vi->z_idataoff;
 		map->m_plen = vi->z_idata_size;
+	} else if (inpacked && m.lcn == vi->z_tailextent_headlcn) {
+		map->m_flags |= EROFS_MAP_FRAGMENT;
+		DBG_BUGON(!map->m_la);
 	} else {
 		map->m_pa = blknr_to_addr(m.pblk);
 		err = z_erofs_get_extent_compressedlen(&m, initial_lcn);
@@ -652,6 +670,14 @@ int z_erofs_map_blocks_iter(struct erofs_inode *vi,
 	if (err)
 		goto out;
 
+	if ((vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER) &&
+	    !vi->z_tailextent_headlcn) {
+		map->m_llen = map->m_la + 1;
+		map->m_la = 0;
+		map->m_flags = EROFS_MAP_MAPPED | EROFS_MAP_FRAGMENT;
+		goto out;
+	}
+
 	err = z_erofs_do_map_blocks(vi, map, flags);
 out:
 	DBG_BUGON(err < 0 && err != -ENOMEM);
-- 
2.17.1

[PATCH V7 3/4] erofs-utils: mkfs: support interlaced uncompressed data layout

[Yue Hu]

From: Yue Hu <huyue2@coolpad.com>

This can benefit from in-place I/O we are using.

Signed-off-by: Yue Hu <huyue2@coolpad.com>
---
 lib/compress.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

diff --git a/lib/compress.c b/lib/compress.c
index fd02053..4bd4e6b 100644
--- a/lib/compress.c
+++ b/lib/compress.c
@@ -139,11 +139,12 @@ static void vle_write_indexes(struct z_erofs_vle_compress_ctx *ctx,
 	ctx->clusterofs = clusterofs + count;
 }
 
-static int write_uncompressed_extent(struct z_erofs_vle_compress_ctx *ctx,
+static int write_uncompressed_extent(struct erofs_inode *inode,
+				     struct z_erofs_vle_compress_ctx *ctx,
 				     unsigned int *len, char *dst)
 {
 	int ret;
-	unsigned int count;
+	unsigned int count, interlaced_offset, rightpart;
 
 	/* reset clusterofs to 0 if permitted */
 	if (!erofs_sb_has_lz4_0padding() && ctx->clusterofs &&
@@ -153,11 +154,19 @@ static int write_uncompressed_extent(struct z_erofs_vle_compress_ctx *ctx,
 		ctx->clusterofs = 0;
 	}
 
-	/* write uncompressed data */
 	count = min(EROFS_BLKSIZ, *len);
 
-	memcpy(dst, ctx->queue + ctx->head, count);
-	memset(dst + count, 0, EROFS_BLKSIZ - count);
+	/*
+	 * write uncompressed data from clusterofs which can benefit from
+	 * in-place I/O, loop shift right when to exceed EROFS_BLKSIZ.
+	 */
+	interlaced_offset = 0; /* will set it to clusterofs */
+	rightpart = min(EROFS_BLKSIZ - interlaced_offset, count);
+
+	memset(dst, 0, EROFS_BLKSIZ);
+
+	memcpy(dst + interlaced_offset, ctx->queue + ctx->head, rightpart);
+	memcpy(dst, ctx->queue + ctx->head + rightpart, count - rightpart);
 
 	erofs_dbg("Writing %u uncompressed data to block %u",
 		  count, ctx->blkaddr);
@@ -263,7 +272,8 @@ static int vle_compress_one(struct erofs_inode *inode,
 						len, true);
 			else
 nocompression:
-				ret = write_uncompressed_extent(ctx, &len, dst);
+				ret = write_uncompressed_extent(inode, ctx,
+								&len, dst);
 
 			if (ret < 0)
 				return ret;
-- 
2.17.1

[PATCH V7 4/4] erofs-utils: mkfs: introduce compressed fragments support

[Yue Hu]

From: Yue Hu <huyue2@coolpad.com>

This approach can merge tail pclusters or the whole files into a special
inode in order to achieve greater compression ratio. And an option of
pcluster size is provided for different compression requirments.

Also enable interlaced uncompressed data layout for compressed files at
the same time.

Signed-off-by: Yue Hu <huyue2@coolpad.com>
---
 include/erofs/compress.h  |  8 ++++-
 include/erofs/config.h    |  3 +-
 include/erofs/fragments.h | 28 ++++++++++++++++
 include/erofs/inode.h     |  1 +
 include/erofs/internal.h  |  3 ++
 lib/Makefile.am           |  4 ++-
 lib/compress.c            | 70 +++++++++++++++++++++++++++++----------
 lib/fragments.c           | 59 +++++++++++++++++++++++++++++++++
 lib/inode.c               | 47 ++++++++++++++++++++++++--
 mkfs/main.c               | 61 ++++++++++++++++++++++++++++++----
 10 files changed, 254 insertions(+), 30 deletions(-)
 create mode 100644 include/erofs/fragments.h
 create mode 100644 lib/fragments.c

diff --git a/include/erofs/compress.h b/include/erofs/compress.h
index 24f6204..2e0f3f5 100644
--- a/include/erofs/compress.h
+++ b/include/erofs/compress.h
@@ -18,13 +18,19 @@ extern "C"
 #define EROFS_CONFIG_COMPR_MIN_SZ           (32   * 1024)
 
 void z_erofs_drop_inline_pcluster(struct erofs_inode *inode);
-int erofs_write_compressed_file(struct erofs_inode *inode);
+int erofs_write_compressed_file_from_fd(struct erofs_inode *inode, int fd);
 
 int z_erofs_compress_init(struct erofs_buffer_head *bh);
 int z_erofs_compress_exit(void);
 
 const char *z_erofs_list_available_compressors(unsigned int i);
 
+static inline bool is_inode_packed(struct erofs_inode *inode)
+{
+	return (inode->nid == sbi.packed_nid ||
+		inode->nid == EROFS_PACKED_NID_UNALLOCATED);
+}
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/include/erofs/config.h b/include/erofs/config.h
index 539d813..764b0f7 100644
--- a/include/erofs/config.h
+++ b/include/erofs/config.h
@@ -44,6 +44,7 @@ struct erofs_configure {
 	char c_chunkbits;
 	bool c_noinline_data;
 	bool c_ztailpacking;
+	bool c_fragments;
 	bool c_ignore_mtime;
 	bool c_showprogress;
 
@@ -62,7 +63,7 @@ struct erofs_configure {
 	/* < 0, xattr disabled and INT_MAX, always use inline xattrs */
 	int c_inline_xattr_tolerance;
 
-	u32 c_pclusterblks_max, c_pclusterblks_def;
+	u32 c_pclusterblks_max, c_pclusterblks_def, c_pclusterblks_packed;
 	u32 c_max_decompressed_extent_bytes;
 	u32 c_dict_size;
 	u64 c_unix_timestamp;
diff --git a/include/erofs/fragments.h b/include/erofs/fragments.h
new file mode 100644
index 0000000..05a88da
--- /dev/null
+++ b/include/erofs/fragments.h
@@ -0,0 +1,28 @@
+/* SPDX-License-Identifier: GPL-2.0+ OR Apache-2.0 */
+/*
+ * Copyright (C), 2022, Coolpad Group Limited.
+ */
+#ifndef __EROFS_FRAGMENTS_H
+#define __EROFS_FRAGMENTS_H
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+#include "erofs/internal.h"
+
+const char *frags_packedname;
+#define EROFS_PACKED_INODE	frags_packedname
+
+int z_erofs_pack_fragments(struct erofs_inode *inode, void *data,
+			   unsigned int len);
+struct erofs_inode *erofs_mkfs_build_fragments(void);
+int erofs_fragments_init(void);
+void erofs_fragments_exit(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/include/erofs/inode.h b/include/erofs/inode.h
index 79b8d89..bf20cd3 100644
--- a/include/erofs/inode.h
+++ b/include/erofs/inode.h
@@ -22,6 +22,7 @@ unsigned int erofs_iput(struct erofs_inode *inode);
 erofs_nid_t erofs_lookupnid(struct erofs_inode *inode);
 struct erofs_inode *erofs_mkfs_build_tree_from_path(struct erofs_inode *parent,
 						    const char *path);
+struct erofs_inode *erofs_mkfs_build_special_from_fd(int fd, const char *name);
 
 #ifdef __cplusplus
 }
diff --git a/include/erofs/internal.h b/include/erofs/internal.h
index 742bfed..8abcfc6 100644
--- a/include/erofs/internal.h
+++ b/include/erofs/internal.h
@@ -70,6 +70,8 @@ struct erofs_device_info {
 	u32 mapped_blkaddr;
 };
 
+#define EROFS_PACKED_NID_UNALLOCATED	-1
+
 struct erofs_sb_info {
 	struct erofs_device_info *devs;
 
@@ -212,6 +214,7 @@ struct erofs_inode {
 	uint64_t capabilities;
 #endif
 	erofs_off_t fragmentoff;
+	unsigned int fragment_size;
 };
 
 static inline bool is_inode_layout_compression(struct erofs_inode *inode)
diff --git a/lib/Makefile.am b/lib/Makefile.am
index 3fad357..95f1d55 100644
--- a/lib/Makefile.am
+++ b/lib/Makefile.am
@@ -22,12 +22,14 @@ noinst_HEADERS = $(top_srcdir)/include/erofs_fs.h \
       $(top_srcdir)/include/erofs/trace.h \
       $(top_srcdir)/include/erofs/xattr.h \
       $(top_srcdir)/include/erofs/compress_hints.h \
+      $(top_srcdir)/include/erofs/fragments.h \
       $(top_srcdir)/lib/liberofs_private.h
 
 noinst_HEADERS += compressor.h
 liberofs_la_SOURCES = config.c io.c cache.c super.c inode.c xattr.c exclude.c \
 		      namei.c data.c compress.c compressor.c zmap.c decompress.c \
-		      compress_hints.c hashmap.c sha256.c blobchunk.c dir.c
+		      compress_hints.c hashmap.c sha256.c blobchunk.c dir.c \
+		      fragments.c
 liberofs_la_CFLAGS = -Wall -I$(top_srcdir)/include
 if ENABLE_LZ4
 liberofs_la_CFLAGS += ${LZ4_CFLAGS}
diff --git a/lib/compress.c b/lib/compress.c
index 4bd4e6b..368db1f 100644
--- a/lib/compress.c
+++ b/lib/compress.c
@@ -18,6 +18,7 @@
 #include "compressor.h"
 #include "erofs/block_list.h"
 #include "erofs/compress_hints.h"
+#include "erofs/fragments.h"
 
 static struct erofs_compress compresshandle;
 static unsigned int algorithmtype[2];
@@ -160,7 +161,12 @@ static int write_uncompressed_extent(struct erofs_inode *inode,
 	 * write uncompressed data from clusterofs which can benefit from
 	 * in-place I/O, loop shift right when to exceed EROFS_BLKSIZ.
 	 */
-	interlaced_offset = 0; /* will set it to clusterofs */
+	if (cfg.c_fragments && ctx->clusterofs) {
+		interlaced_offset = ctx->clusterofs;
+		inode->z_advise |= Z_EROFS_ADVISE_INTERLACED_PCLUSTER;
+	} else {
+		interlaced_offset = 0;
+	}
 	rightpart = min(EROFS_BLKSIZ - interlaced_offset, count);
 
 	memset(dst, 0, EROFS_BLKSIZ);
@@ -178,6 +184,8 @@ static int write_uncompressed_extent(struct erofs_inode *inode,
 
 static unsigned int z_erofs_get_max_pclusterblks(struct erofs_inode *inode)
 {
+	if (cfg.c_fragments && is_inode_packed(inode))
+		return cfg.c_pclusterblks_packed;
 #ifndef NDEBUG
 	if (cfg.c_random_pclusterblks)
 		return 1 + rand() % cfg.c_pclusterblks_max;
@@ -246,11 +254,17 @@ static int vle_compress_one(struct erofs_inode *inode,
 		unsigned int pclustersize =
 			z_erofs_get_max_pclusterblks(inode) * EROFS_BLKSIZ;
 		bool may_inline = (cfg.c_ztailpacking && final);
+		bool may_packing = (cfg.c_fragments && final &&
+				   !is_inode_packed(inode));
 		bool raw;
 
 		if (len <= pclustersize) {
 			if (!final)
 				break;
+			if (may_packing) {
+				count = len;
+				goto frag_packing;
+			}
 			if (!may_inline && len <= EROFS_BLKSIZ)
 				goto nocompression;
 		}
@@ -265,7 +279,6 @@ static int vle_compress_one(struct erofs_inode *inode,
 					  inode->i_srcpath,
 					  erofs_strerror(ret));
 			}
-
 			if (may_inline && len < EROFS_BLKSIZ)
 				ret = z_erofs_fill_inline_data(inode,
 						ctx->queue + ctx->head,
@@ -304,6 +317,19 @@ nocompression:
 				return ret;
 			ctx->compressedblks = 1;
 			raw = false;
+		} else if (may_packing && len == count && ret < pclustersize) {
+frag_packing:
+			ret = z_erofs_pack_fragments(inode,
+						     ctx->queue + ctx->head,
+						     len);
+			if (ret < 0)
+				return ret;
+			if (inode->i_size == inode->fragment_size) {
+				ctx->head += len;
+				return 0;
+			}
+			ctx->compressedblks = 0;
+			raw = true;
 		} else {
 			unsigned int tailused, padding;
 
@@ -556,13 +582,20 @@ static void z_erofs_write_mapheader(struct erofs_inode *inode,
 {
 	struct z_erofs_map_header h = {
 		.h_advise = cpu_to_le16(inode->z_advise),
-		.h_idata_size = cpu_to_le16(inode->idata_size),
 		.h_algorithmtype = inode->z_algorithmtype[1] << 4 |
 				   inode->z_algorithmtype[0],
 		/* lclustersize */
 		.h_clusterbits = inode->z_logical_clusterbits - 12,
 	};
 
+	if (cfg.c_fragments)
+		h.h_fragmentoff = cpu_to_le32(inode->fragmentoff);
+	else
+		h.h_idata_size = cpu_to_le16(inode->idata_size);
+
+	if (inode->fragment_size && inode->i_size == inode->fragment_size)
+		h.h_clusterbits |=  1 << Z_EROFS_FRAGMENT_INODE_BIT;
+
 	memset(compressmeta, 0, Z_EROFS_LEGACY_MAP_HEADER_SIZE);
 	/* write out map header */
 	memcpy(compressmeta, &h, sizeof(struct z_erofs_map_header));
@@ -615,30 +648,24 @@ void z_erofs_drop_inline_pcluster(struct erofs_inode *inode)
 	inode->eof_tailraw = NULL;
 }
 
-int erofs_write_compressed_file(struct erofs_inode *inode)
+int erofs_write_compressed_file_from_fd(struct erofs_inode *inode, int fd)
 {
 	struct erofs_buffer_head *bh;
 	static struct z_erofs_vle_compress_ctx ctx;
 	erofs_off_t remaining;
 	erofs_blk_t blkaddr, compressed_blocks;
 	unsigned int legacymetasize;
-	int ret, fd;
+	int ret;
 	u8 *compressmeta = malloc(vle_compressmeta_capacity(inode->i_size));
 
 	if (!compressmeta)
 		return -ENOMEM;
 
-	fd = open(inode->i_srcpath, O_RDONLY | O_BINARY);
-	if (fd < 0) {
-		ret = -errno;
-		goto err_free_meta;
-	}
-
 	/* allocate main data buffer */
 	bh = erofs_balloc(DATA, 0, 0, 0);
 	if (IS_ERR(bh)) {
 		ret = PTR_ERR(bh);
-		goto err_close;
+		goto err_free_meta;
 	}
 
 	/* initialize per-file compression setting */
@@ -659,6 +686,9 @@ int erofs_write_compressed_file(struct erofs_inode *inode)
 	inode->z_algorithmtype[1] = algorithmtype[1];
 	inode->z_logical_clusterbits = LOG_BLOCK_SIZE;
 
+	inode->idata_size = 0;
+	inode->fragment_size = 0;
+
 	blkaddr = erofs_mapbh(bh->block);	/* start_blkaddr */
 	ctx.blkaddr = blkaddr;
 	ctx.metacur = compressmeta + Z_EROFS_LEGACY_MAP_HEADER_SIZE;
@@ -692,19 +722,20 @@ int erofs_write_compressed_file(struct erofs_inode *inode)
 	vle_write_indexes_final(&ctx);
 	legacymetasize = ctx.metacur - compressmeta;
 	/* estimate if data compression saves space or not */
-	if (compressed_blocks * EROFS_BLKSIZ + inode->idata_size +
+	if (!inode->fragment_size && !is_inode_packed(inode) &&
+	    compressed_blocks * EROFS_BLKSIZ + inode->idata_size +
 	    legacymetasize >= inode->i_size) {
 		ret = -ENOSPC;
 		goto err_free_idata;
 	}
 	z_erofs_write_mapheader(inode, compressmeta);
 
-	close(fd);
 	if (compressed_blocks) {
 		ret = erofs_bh_balloon(bh, blknr_to_addr(compressed_blocks));
 		DBG_BUGON(ret != EROFS_BLKSIZ);
 	} else {
-		DBG_BUGON(!inode->idata_size);
+		if (!cfg.c_fragments)
+			DBG_BUGON(!inode->idata_size);
 	}
 
 	erofs_info("compressed %s (%llu bytes) into %u blocks",
@@ -727,7 +758,8 @@ int erofs_write_compressed_file(struct erofs_inode *inode)
 		DBG_BUGON(ret);
 	}
 	inode->compressmeta = compressmeta;
-	erofs_droid_blocklist_write(inode, blkaddr, compressed_blocks);
+	if (!is_inode_packed(inode))
+		erofs_droid_blocklist_write(inode, blkaddr, compressed_blocks);
 	return 0;
 
 err_free_idata:
@@ -737,8 +769,6 @@ err_free_idata:
 	}
 err_bdrop:
 	erofs_bdrop(bh, true);	/* revoke buffer */
-err_close:
-	close(fd);
 err_free_meta:
 	free(compressmeta);
 	return ret;
@@ -852,6 +882,10 @@ int z_erofs_compress_init(struct erofs_buffer_head *sb_bh)
 		}
 		erofs_sb_set_big_pcluster();
 	}
+	if (cfg.c_pclusterblks_packed > cfg.c_pclusterblks_max) {
+		erofs_err("invalid physical cluster size for the packed file");
+		return -EINVAL;
+	}
 
 	if (ret != Z_EROFS_COMPRESSION_LZ4)
 		erofs_sb_set_compr_cfgs();
diff --git a/lib/fragments.c b/lib/fragments.c
new file mode 100644
index 0000000..61e711e
--- /dev/null
+++ b/lib/fragments.c
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: GPL-2.0+ OR Apache-2.0
+/*
+ * Copyright (C), 2022, Coolpad Group Limited.
+ * Created by Yue Hu <huyue2@coolpad.com>
+ */
+#define _GNU_SOURCE
+#include <stdlib.h>
+#include <unistd.h>
+#include "erofs/err.h"
+#include "erofs/inode.h"
+#include "erofs/compress.h"
+#include "erofs/print.h"
+#include "erofs/fragments.h"
+
+static FILE *packedfile;
+const char *frags_packedname = "frags_packed_file";
+
+int z_erofs_pack_fragments(struct erofs_inode *inode, void *data,
+			   unsigned int len)
+{
+	inode->z_advise |= Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
+	inode->fragmentoff = ftell(packedfile);
+	inode->fragment_size = len;
+
+	if (write(fileno(packedfile), data, len) < 0)
+		return -EIO;
+
+	erofs_sb_set_fragments();
+
+	erofs_dbg("Recording %u fragment data at %lu", inode->fragment_size,
+		  inode->fragmentoff);
+	return len;
+}
+
+struct erofs_inode *erofs_mkfs_build_fragments(void)
+{
+	fseek(packedfile, 0, SEEK_SET);
+
+	return erofs_mkfs_build_special_from_fd(fileno(packedfile),
+						frags_packedname);
+}
+
+void erofs_fragments_exit(void)
+{
+	if (packedfile)
+		fclose(packedfile);
+}
+
+int erofs_fragments_init(void)
+{
+#ifdef HAVE_TMPFILE64
+	packedfile = tmpfile64();
+#else
+	packedfile = tmpfile();
+#endif
+	if (!packedfile)
+		return -ENOMEM;
+	return 0;
+}
diff --git a/lib/inode.c b/lib/inode.c
index 4da28b3..43de1d0 100644
--- a/lib/inode.c
+++ b/lib/inode.c
@@ -25,6 +25,7 @@
 #include "erofs/block_list.h"
 #include "erofs/compress_hints.h"
 #include "erofs/blobchunk.h"
+#include "erofs/fragments.h"
 #include "liberofs_private.h"
 
 #define S_SHIFT                 12
@@ -424,7 +425,11 @@ int erofs_write_file(struct erofs_inode *inode)
 	}
 
 	if (cfg.c_compr_alg_master && erofs_file_is_compressible(inode)) {
-		ret = erofs_write_compressed_file(inode);
+		fd = open(inode->i_srcpath, O_RDONLY | O_BINARY);
+		if (fd < 0)
+			return -errno;
+		ret = erofs_write_compressed_file_from_fd(inode, fd);
+		close(fd);
 
 		if (!ret || ret != -ENOSPC)
 			return ret;
@@ -844,8 +849,7 @@ static int erofs_droid_inode_fsconfig(struct erofs_inode *inode,
 }
 #endif
 
-static int erofs_fill_inode(struct erofs_inode *inode,
-			    struct stat64 *st,
+static int erofs_fill_inode(struct erofs_inode *inode, struct stat64 *st,
 			    const char *path)
 {
 	int err = erofs_droid_inode_fsconfig(inode, st, path);
@@ -1180,3 +1184,40 @@ struct erofs_inode *erofs_mkfs_build_tree_from_path(struct erofs_inode *parent,
 
 	return erofs_mkfs_build_tree(inode);
 }
+
+struct erofs_inode *erofs_mkfs_build_special_from_fd(int fd, const char *name)
+{
+	struct stat64 st;
+	struct erofs_inode *inode;
+	int ret;
+
+	ret = fstat64(fd, &st);
+	if (ret)
+		return ERR_PTR(-errno);
+
+	inode = erofs_new_inode();
+	if (IS_ERR(inode))
+		return inode;
+
+	ret = erofs_fill_inode(inode, &st, name);
+	if (ret) {
+		free(inode);
+		return ERR_PTR(ret);
+	}
+
+	if (name == EROFS_PACKED_INODE) {
+		sbi.packed_nid = EROFS_PACKED_NID_UNALLOCATED;
+		inode->nid = sbi.packed_nid;
+	}
+
+	/* only for compressed file now */
+	ret = erofs_write_compressed_file_from_fd(inode, fd);
+	if (ret) {
+		DBG_BUGON(ret == -ENOSPC);
+		return ERR_PTR(ret);
+	}
+
+	erofs_prepare_inode_buffer(inode);
+
+	return inode;
+}
diff --git a/mkfs/main.c b/mkfs/main.c
index b969b35..adad0d7 100644
--- a/mkfs/main.c
+++ b/mkfs/main.c
@@ -23,6 +23,7 @@
 #include "erofs/block_list.h"
 #include "erofs/compress_hints.h"
 #include "erofs/blobchunk.h"
+#include "erofs/fragments.h"
 #include "../lib/liberofs_private.h"
 
 #ifdef HAVE_LIBUUID
@@ -133,9 +134,9 @@ static int parse_extended_opts(const char *opts)
 		const char *p = strchr(token, ',');
 
 		next = NULL;
-		if (p)
+		if (p) {
 			next = p + 1;
-		else {
+		} else {
 			p = token + strlen(token);
 			next = p;
 		}
@@ -202,7 +203,31 @@ static int parse_extended_opts(const char *opts)
 				return -EINVAL;
 			cfg.c_ztailpacking = true;
 		}
+
+		if (MATCH_EXTENTED_OPT("fragments", token, keylen)) {
+			char *endptr;
+			u64 i;
+
+			if (vallen || cfg.c_ztailpacking)
+				return -EINVAL;
+			cfg.c_fragments = true;
+
+			i = strtoull(next, &endptr, 0);
+			if ((*endptr != '\0' && *endptr != ',') ||
+			    i < EROFS_BLKSIZ || i % EROFS_BLKSIZ) {
+				erofs_err("invalid physical cluster size for the packed file %s",
+					  next);
+				return -EINVAL;
+			}
+			cfg.c_pclusterblks_packed = i / EROFS_BLKSIZ;
+
+			if (*endptr == ',')
+				next = strchr(next, ',')  + 1;
+			else
+				goto out;
+		}
 	}
+out:
 	return 0;
 }
 
@@ -458,7 +483,8 @@ static int mkfs_parse_options_cfg(int argc, char *argv[])
 
 int erofs_mkfs_update_super_block(struct erofs_buffer_head *bh,
 				  erofs_nid_t root_nid,
-				  erofs_blk_t *blocks)
+				  erofs_blk_t *blocks,
+				  erofs_nid_t packed_nid)
 {
 	struct erofs_super_block sb = {
 		.magic     = cpu_to_le32(EROFS_SUPER_MAGIC_V1),
@@ -482,6 +508,7 @@ int erofs_mkfs_update_super_block(struct erofs_buffer_head *bh,
 	*blocks         = erofs_mapbh(NULL);
 	sb.blocks       = cpu_to_le32(*blocks);
 	sb.root_nid     = cpu_to_le16(root_nid);
+	sb.packed_nid    = cpu_to_le64(packed_nid);
 	memcpy(sb.uuid, sbi.uuid, sizeof(sb.uuid));
 
 	if (erofs_sb_has_compr_cfgs())
@@ -599,8 +626,8 @@ int main(int argc, char **argv)
 {
 	int err = 0;
 	struct erofs_buffer_head *sb_bh;
-	struct erofs_inode *root_inode;
-	erofs_nid_t root_nid;
+	struct erofs_inode *root_inode, *packed_inode;
+	erofs_nid_t root_nid, packed_nid;
 	struct stat64 st;
 	erofs_blk_t nblocks;
 	struct timeval t;
@@ -670,6 +697,14 @@ int main(int argc, char **argv)
 		erofs_warn("EXPERIMENTAL chunked file feature in use. Use at your own risk!");
 	if (cfg.c_ztailpacking)
 		erofs_warn("EXPERIMENTAL compressed inline data feature in use. Use at your own risk!");
+	if (cfg.c_fragments) {
+		err = erofs_fragments_init();
+		if (err) {
+			erofs_err("failed to initialize fragments");
+			return 1;
+		}
+		erofs_warn("EXPERIMENTAL compressed fragments feature in use. Use at your own risk!");
+	}
 	erofs_set_fs_root(cfg.c_src_path);
 #ifndef NDEBUG
 	if (cfg.c_random_pclusterblks)
@@ -739,7 +774,19 @@ int main(int argc, char **argv)
 			goto exit;
 	}
 
-	err = erofs_mkfs_update_super_block(sb_bh, root_nid, &nblocks);
+	packed_nid = 0;
+	if (cfg.c_fragments) {
+		packed_inode = erofs_mkfs_build_fragments();
+		if (IS_ERR(packed_inode)) {
+			err = PTR_ERR(packed_inode);
+			goto exit;
+		}
+		packed_nid = erofs_lookupnid(packed_inode);
+		erofs_iput(packed_inode);
+	}
+
+	err = erofs_mkfs_update_super_block(sb_bh, root_nid, &nblocks,
+					    packed_nid);
 	if (err)
 		goto exit;
 
@@ -761,6 +808,8 @@ exit:
 	erofs_cleanup_exclude_rules();
 	if (cfg.c_chunkbits)
 		erofs_blob_exit();
+	if (cfg.c_fragments)
+		erofs_fragments_exit();
 	erofs_exit_configure();
 
 	if (err) {
-- 
2.17.1