Merge branch 'xfs-sparse-inode' into for-next

fs/xfs/libxfs/xfs_alloc.c

@@ -149,13 +149,27 @@ xfs_alloc_compute_aligned(
 {
 	xfs_agblock_t	bno;
 	xfs_extlen_t	len;
+	xfs_extlen_t	diff;
 
 	/* Trim busy sections out of found extent */
 	xfs_extent_busy_trim(args, foundbno, foundlen, &bno, &len);
 
+	/*
+	 * If we have a largish extent that happens to start before min_agbno,
+	 * see if we can shift it into range...
+	 */
+	if (bno < args->min_agbno && bno + len > args->min_agbno) {
+		diff = args->min_agbno - bno;
+		if (len > diff) {
+			bno += diff;
+			len -= diff;
+		}
+	}
+
 	if (args->alignment > 1 && len >= args->minlen) {
 		xfs_agblock_t	aligned_bno = roundup(bno, args->alignment);
-		xfs_extlen_t	diff = aligned_bno - bno;
+
+		diff = aligned_bno - bno;
 
 		*resbno = aligned_bno;
 		*reslen = diff >= len ? 0 : len - diff;
@@ -795,9 +809,13 @@ xfs_alloc_find_best_extent(
 		 * The good extent is closer than this one.
 		 */
 		if (!dir) {
+			if (*sbnoa > args->max_agbno)
+				goto out_use_good;
 			if (*sbnoa >= args->agbno + gdiff)
 				goto out_use_good;
 		} else {
+			if (*sbnoa < args->min_agbno)
+				goto out_use_good;
 			if (*sbnoa <= args->agbno - gdiff)
 				goto out_use_good;
 		}
@@ -884,6 +902,17 @@ xfs_alloc_ag_vextent_near(
 	dofirst = prandom_u32() & 1;
 #endif
 
+	/* handle unitialized agbno range so caller doesn't have to */
+	if (!args->min_agbno && !args->max_agbno)
+		args->max_agbno = args->mp->m_sb.sb_agblocks - 1;
+	ASSERT(args->min_agbno <= args->max_agbno);
+
+	/* clamp agbno to the range if it's outside */
+	if (args->agbno < args->min_agbno)
+		args->agbno = args->min_agbno;
+	if (args->agbno > args->max_agbno)
+		args->agbno = args->max_agbno;
+
 restart:
 	bno_cur_lt = NULL;
 	bno_cur_gt = NULL;
@@ -976,6 +1005,8 @@ restart:
 						  &ltbnoa, &ltlena);
 			if (ltlena < args->minlen)
 				continue;
+			if (ltbnoa < args->min_agbno || ltbnoa > args->max_agbno)
+				continue;
 			args->len = XFS_EXTLEN_MIN(ltlena, args->maxlen);
 			xfs_alloc_fix_len(args);
 			ASSERT(args->len >= args->minlen);
@@ -1096,11 +1127,11 @@ restart:
 			XFS_WANT_CORRUPTED_GOTO(args->mp, i == 1, error0);
 			xfs_alloc_compute_aligned(args, ltbno, ltlen,
 						  &ltbnoa, &ltlena);
-			if (ltlena >= args->minlen)
+			if (ltlena >= args->minlen && ltbnoa >= args->min_agbno)
 				break;
 			if ((error = xfs_btree_decrement(bno_cur_lt, 0, &i)))
 				goto error0;
-			if (!i) {
+			if (!i || ltbnoa < args->min_agbno) {
 				xfs_btree_del_cursor(bno_cur_lt,
 						     XFS_BTREE_NOERROR);
 				bno_cur_lt = NULL;
@@ -1112,11 +1143,11 @@ restart:
 			XFS_WANT_CORRUPTED_GOTO(args->mp, i == 1, error0);
 			xfs_alloc_compute_aligned(args, gtbno, gtlen,
 						  &gtbnoa, &gtlena);
-			if (gtlena >= args->minlen)
+			if (gtlena >= args->minlen && gtbnoa <= args->max_agbno)
 				break;
 			if ((error = xfs_btree_increment(bno_cur_gt, 0, &i)))
 				goto error0;
-			if (!i) {
+			if (!i || gtbnoa > args->max_agbno) {
 				xfs_btree_del_cursor(bno_cur_gt,
 						     XFS_BTREE_NOERROR);
 				bno_cur_gt = NULL;
@@ -1216,6 +1247,7 @@ restart:
 	ASSERT(ltnew >= ltbno);
 	ASSERT(ltnew + rlen <= ltbnoa + ltlena);
 	ASSERT(ltnew + rlen <= be32_to_cpu(XFS_BUF_TO_AGF(args->agbp)->agf_length));
+	ASSERT(ltnew >= args->min_agbno && ltnew <= args->max_agbno);
 	args->agbno = ltnew;
 
 	if ((error = xfs_alloc_fixup_trees(cnt_cur, bno_cur_lt, ltbno, ltlen,

fs/xfs/libxfs/xfs_alloc.h

@@ -112,6 +112,8 @@ typedef struct xfs_alloc_arg {
 	xfs_extlen_t	total;		/* total blocks needed in xaction */
 	xfs_extlen_t	alignment;	/* align answer to multiple of this */
 	xfs_extlen_t	minalignslop;	/* slop for minlen+alignment calcs */
+	xfs_agblock_t	min_agbno;	/* set an agbno range for NEAR allocs */
+	xfs_agblock_t	max_agbno;	/* ... */
 	xfs_extlen_t	len;		/* output: actual size of extent */
 	xfs_alloctype_t	type;		/* allocation type XFS_ALLOCTYPE_... */
 	xfs_alloctype_t	otype;		/* original allocation type */

fs/xfs/libxfs/xfs_format.h

@@ -170,7 +170,7 @@ typedef struct xfs_sb {
 	__uint32_t	sb_features_log_incompat;
 
 	__uint32_t	sb_crc;		/* superblock crc */
-	__uint32_t	sb_pad;
+	xfs_extlen_t	sb_spino_align;	/* sparse inode chunk alignment */
 
 	xfs_ino_t	sb_pquotino;	/* project quota inode */
 	xfs_lsn_t	sb_lsn;		/* last write sequence */
@@ -256,7 +256,7 @@ typedef struct xfs_dsb {
 	__be32		sb_features_log_incompat;
 
 	__le32		sb_crc;		/* superblock crc */
-	__be32		sb_pad;
+	__be32		sb_spino_align;	/* sparse inode chunk alignment */
 
 	__be64		sb_pquotino;	/* project quota inode */
 	__be64		sb_lsn;		/* last write sequence */
@@ -457,8 +457,10 @@ xfs_sb_has_ro_compat_feature(
 }
 
 #define XFS_SB_FEAT_INCOMPAT_FTYPE	(1 << 0)	/* filetype in dirent */
+#define XFS_SB_FEAT_INCOMPAT_SPINODES	(1 << 1)	/* sparse inode chunks */
 #define XFS_SB_FEAT_INCOMPAT_ALL \
-		(XFS_SB_FEAT_INCOMPAT_FTYPE)
+		(XFS_SB_FEAT_INCOMPAT_FTYPE|	\
+		 XFS_SB_FEAT_INCOMPAT_SPINODES)
 
 #define XFS_SB_FEAT_INCOMPAT_UNKNOWN	~XFS_SB_FEAT_INCOMPAT_ALL
 static inline bool
@@ -506,6 +508,12 @@ static inline int xfs_sb_version_hasfinobt(xfs_sb_t *sbp)
 		(sbp->sb_features_ro_compat & XFS_SB_FEAT_RO_COMPAT_FINOBT);
 }
 
+static inline bool xfs_sb_version_hassparseinodes(struct xfs_sb *sbp)
+{
+	return XFS_SB_VERSION_NUM(sbp) == XFS_SB_VERSION_5 &&
+		xfs_sb_has_incompat_feature(sbp, XFS_SB_FEAT_INCOMPAT_SPINODES);
+}
+
 /*
  * end of superblock version macros
  */
@@ -1216,26 +1224,54 @@ typedef	__uint64_t	xfs_inofree_t;
 #define	XFS_INOBT_ALL_FREE		((xfs_inofree_t)-1)
 #define	XFS_INOBT_MASK(i)		((xfs_inofree_t)1 << (i))
 
+#define XFS_INOBT_HOLEMASK_FULL		0	/* holemask for full chunk */
+#define XFS_INOBT_HOLEMASK_BITS		(NBBY * sizeof(__uint16_t))
+#define XFS_INODES_PER_HOLEMASK_BIT	\
+	(XFS_INODES_PER_CHUNK / (NBBY * sizeof(__uint16_t)))
+
 static inline xfs_inofree_t xfs_inobt_maskn(int i, int n)
 {
 	return ((n >= XFS_INODES_PER_CHUNK ? 0 : XFS_INOBT_MASK(n)) - 1) << i;
 }
 
 /*
- * Data record structure
+ * The on-disk inode record structure has two formats. The original "full"
+ * format uses a 4-byte freecount. The "sparse" format uses a 1-byte freecount
+ * and replaces the 3 high-order freecount bytes wth the holemask and inode
+ * count.
+ *
+ * The holemask of the sparse record format allows an inode chunk to have holes
+ * that refer to blocks not owned by the inode record. This facilitates inode
+ * allocation in the event of severe free space fragmentation.
  */
 typedef struct xfs_inobt_rec {
 	__be32		ir_startino;	/* starting inode number */
-	__be32		ir_freecount;	/* count of free inodes (set bits) */
+	union {
+		struct {
+			__be32	ir_freecount;	/* count of free inodes */
+		} f;
+		struct {
+			__be16	ir_holemask;/* hole mask for sparse chunks */
+			__u8	ir_count;	/* total inode count */
+			__u8	ir_freecount;	/* count of free inodes */
+		} sp;
+	} ir_u;
 	__be64		ir_free;	/* free inode mask */
 } xfs_inobt_rec_t;
 
 typedef struct xfs_inobt_rec_incore {
 	xfs_agino_t	ir_startino;	/* starting inode number */
-	__int32_t	ir_freecount;	/* count of free inodes (set bits) */
+	__uint16_t	ir_holemask;	/* hole mask for sparse chunks */
+	__uint8_t	ir_count;	/* total inode count */
+	__uint8_t	ir_freecount;	/* count of free inodes (set bits) */
 	xfs_inofree_t	ir_free;	/* free inode mask */
 } xfs_inobt_rec_incore_t;
 
+static inline bool xfs_inobt_issparse(uint16_t holemask)
+{
+	/* non-zero holemask represents a sparse rec. */
+	return holemask;
+}
 
 /*
  * Key structure

fs/xfs/libxfs/xfs_fs.h

@@ -239,6 +239,7 @@ typedef struct xfs_fsop_resblks {
 #define XFS_FSOP_GEOM_FLAGS_V5SB	0x8000	/* version 5 superblock */
 #define XFS_FSOP_GEOM_FLAGS_FTYPE	0x10000	/* inode directory types */
 #define XFS_FSOP_GEOM_FLAGS_FINOBT	0x20000	/* free inode btree */
+#define XFS_FSOP_GEOM_FLAGS_SPINODES	0x40000	/* sparse inode chunks	*/
 
 /*
  * Minimum and maximum sizes need for growth checks.

fs/xfs/libxfs/xfs_ialloc.c

@@ -65,6 +65,8 @@ xfs_inobt_lookup(
 	int			*stat)	/* success/failure */
 {
 	cur->bc_rec.i.ir_startino = ino;
+	cur->bc_rec.i.ir_holemask = 0;
+	cur->bc_rec.i.ir_count = 0;
 	cur->bc_rec.i.ir_freecount = 0;
 	cur->bc_rec.i.ir_free = 0;
 	return xfs_btree_lookup(cur, dir, stat);
@@ -82,7 +84,14 @@ xfs_inobt_update(
 	union xfs_btree_rec	rec;
 
 	rec.inobt.ir_startino = cpu_to_be32(irec->ir_startino);
-	rec.inobt.ir_freecount = cpu_to_be32(irec->ir_freecount);
+	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
+		rec.inobt.ir_u.sp.ir_holemask = cpu_to_be16(irec->ir_holemask);
+		rec.inobt.ir_u.sp.ir_count = irec->ir_count;
+		rec.inobt.ir_u.sp.ir_freecount = irec->ir_freecount;
+	} else {
+		/* ir_holemask/ir_count not supported on-disk */
+		rec.inobt.ir_u.f.ir_freecount = cpu_to_be32(irec->ir_freecount);
+	}
 	rec.inobt.ir_free = cpu_to_be64(irec->ir_free);
 	return xfs_btree_update(cur, &rec);
 }
@@ -100,12 +109,27 @@ xfs_inobt_get_rec(
 	int			error;
 
 	error = xfs_btree_get_rec(cur, &rec, stat);
-	if (!error && *stat == 1) {
+	if (error || *stat == 0)
-		irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
+		return error;
-		irec->ir_freecount = be32_to_cpu(rec->inobt.ir_freecount);
+
-		irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
+	irec->ir_startino = be32_to_cpu(rec->inobt.ir_startino);
+	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
+		irec->ir_holemask = be16_to_cpu(rec->inobt.ir_u.sp.ir_holemask);
+		irec->ir_count = rec->inobt.ir_u.sp.ir_count;
+		irec->ir_freecount = rec->inobt.ir_u.sp.ir_freecount;
+	} else {
+		/*
+		 * ir_holemask/ir_count not supported on-disk. Fill in hardcoded
+		 * values for full inode chunks.
+		 */
+		irec->ir_holemask = XFS_INOBT_HOLEMASK_FULL;
+		irec->ir_count = XFS_INODES_PER_CHUNK;
+		irec->ir_freecount =
+				be32_to_cpu(rec->inobt.ir_u.f.ir_freecount);
 	}
-	return error;
+	irec->ir_free = be64_to_cpu(rec->inobt.ir_free);
+
+	return 0;
 }
 
 /*
@@ -114,10 +138,14 @@ xfs_inobt_get_rec(
 STATIC int
 xfs_inobt_insert_rec(
 	struct xfs_btree_cur	*cur,
+	__uint16_t		holemask,
+	__uint8_t		count,
 	__int32_t		freecount,
 	xfs_inofree_t		free,
 	int			*stat)
 {
+	cur->bc_rec.i.ir_holemask = holemask;
+	cur->bc_rec.i.ir_count = count;
 	cur->bc_rec.i.ir_freecount = freecount;
 	cur->bc_rec.i.ir_free = free;
 	return xfs_btree_insert(cur, stat);
@@ -154,7 +182,9 @@ xfs_inobt_insert(
 		}
 		ASSERT(i == 0);
 
-		error = xfs_inobt_insert_rec(cur, XFS_INODES_PER_CHUNK,
+		error = xfs_inobt_insert_rec(cur, XFS_INOBT_HOLEMASK_FULL,
+					     XFS_INODES_PER_CHUNK,
+					     XFS_INODES_PER_CHUNK,
 					     XFS_INOBT_ALL_FREE, &i);
 		if (error) {
 			xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
@@ -220,6 +250,7 @@ xfs_ialloc_inode_init(
 	struct xfs_mount	*mp,
 	struct xfs_trans	*tp,
 	struct list_head	*buffer_list,
+	int			icount,
 	xfs_agnumber_t		agno,
 	xfs_agblock_t		agbno,
 	xfs_agblock_t		length,
@@ -275,7 +306,7 @@ xfs_ialloc_inode_init(
 		 * they track in the AIL as if they were physically logged.
 		 */
 		if (tp)
-			xfs_icreate_log(tp, agno, agbno, mp->m_ialloc_inos,
+			xfs_icreate_log(tp, agno, agbno, icount,
 					mp->m_sb.sb_inodesize, length, gen);
 	} else
 		version = 2;
@@ -346,6 +377,214 @@ xfs_ialloc_inode_init(
 	return 0;
 }
 
+/*
+ * Align startino and allocmask for a recently allocated sparse chunk such that
+ * they are fit for insertion (or merge) into the on-disk inode btrees.
+ *
+ * Background:
+ *
+ * When enabled, sparse inode support increases the inode alignment from cluster
+ * size to inode chunk size. This means that the minimum range between two
+ * non-adjacent inode records in the inobt is large enough for a full inode
+ * record. This allows for cluster sized, cluster aligned block allocation
+ * without need to worry about whether the resulting inode record overlaps with
+ * another record in the tree. Without this basic rule, we would have to deal
+ * with the consequences of overlap by potentially undoing recent allocations in
+ * the inode allocation codepath.
+ *
+ * Because of this alignment rule (which is enforced on mount), there are two
+ * inobt possibilities for newly allocated sparse chunks. One is that the
+ * aligned inode record for the chunk covers a range of inodes not already
+ * covered in the inobt (i.e., it is safe to insert a new sparse record). The
+ * other is that a record already exists at the aligned startino that considers
+ * the newly allocated range as sparse. In the latter case, record content is
+ * merged in hope that sparse inode chunks fill to full chunks over time.
+ */
+STATIC void
+xfs_align_sparse_ino(
+	struct xfs_mount		*mp,
+	xfs_agino_t			*startino,
+	uint16_t			*allocmask)
+{
+	xfs_agblock_t			agbno;
+	xfs_agblock_t			mod;
+	int				offset;
+
+	agbno = XFS_AGINO_TO_AGBNO(mp, *startino);
+	mod = agbno % mp->m_sb.sb_inoalignmt;
+	if (!mod)
+		return;
+
+	/* calculate the inode offset and align startino */
+	offset = mod << mp->m_sb.sb_inopblog;
+	*startino -= offset;
+
+	/*
+	 * Since startino has been aligned down, left shift allocmask such that
+	 * it continues to represent the same physical inodes relative to the
+	 * new startino.
+	 */
+	*allocmask <<= offset / XFS_INODES_PER_HOLEMASK_BIT;
+}
+
+/*
+ * Determine whether the source inode record can merge into the target. Both
+ * records must be sparse, the inode ranges must match and there must be no
+ * allocation overlap between the records.
+ */
+STATIC bool
+__xfs_inobt_can_merge(
+	struct xfs_inobt_rec_incore	*trec,	/* tgt record */
+	struct xfs_inobt_rec_incore	*srec)	/* src record */
+{
+	uint64_t			talloc;
+	uint64_t			salloc;
+
+	/* records must cover the same inode range */
+	if (trec->ir_startino != srec->ir_startino)
+		return false;
+
+	/* both records must be sparse */
+	if (!xfs_inobt_issparse(trec->ir_holemask) ||
+	    !xfs_inobt_issparse(srec->ir_holemask))
+		return false;
+
+	/* both records must track some inodes */
+	if (!trec->ir_count || !srec->ir_count)
+		return false;
+
+	/* can't exceed capacity of a full record */
+	if (trec->ir_count + srec->ir_count > XFS_INODES_PER_CHUNK)
+		return false;
+
+	/* verify there is no allocation overlap */
+	talloc = xfs_inobt_irec_to_allocmask(trec);
+	salloc = xfs_inobt_irec_to_allocmask(srec);
+	if (talloc & salloc)
+		return false;
+
+	return true;
+}
+
+/*
+ * Merge the source inode record into the target. The caller must call
+ * __xfs_inobt_can_merge() to ensure the merge is valid.
+ */
+STATIC void
+__xfs_inobt_rec_merge(
+	struct xfs_inobt_rec_incore	*trec,	/* target */
+	struct xfs_inobt_rec_incore	*srec)	/* src */
+{
+	ASSERT(trec->ir_startino == srec->ir_startino);
+
+	/* combine the counts */
+	trec->ir_count += srec->ir_count;
+	trec->ir_freecount += srec->ir_freecount;
+
+	/*
+	 * Merge the holemask and free mask. For both fields, 0 bits refer to
+	 * allocated inodes. We combine the allocated ranges with bitwise AND.
+	 */
+	trec->ir_holemask &= srec->ir_holemask;
+	trec->ir_free &= srec->ir_free;
+}
+
+/*
+ * Insert a new sparse inode chunk into the associated inode btree. The inode
+ * record for the sparse chunk is pre-aligned to a startino that should match
+ * any pre-existing sparse inode record in the tree. This allows sparse chunks
+ * to fill over time.
+ *
+ * This function supports two modes of handling preexisting records depending on
+ * the merge flag. If merge is true, the provided record is merged with the
+ * existing record and updated in place. The merged record is returned in nrec.
+ * If merge is false, an existing record is replaced with the provided record.
+ * If no preexisting record exists, the provided record is always inserted.
+ *
+ * It is considered corruption if a merge is requested and not possible. Given
+ * the sparse inode alignment constraints, this should never happen.
+ */
+STATIC int
+xfs_inobt_insert_sprec(
+	struct xfs_mount		*mp,
+	struct xfs_trans		*tp,
+	struct xfs_buf			*agbp,
+	int				btnum,
+	struct xfs_inobt_rec_incore	*nrec,	/* in/out: new/merged rec. */
+	bool				merge)	/* merge or replace */
+{
+	struct xfs_btree_cur		*cur;
+	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
+	xfs_agnumber_t			agno = be32_to_cpu(agi->agi_seqno);
+	int				error;
+	int				i;
+	struct xfs_inobt_rec_incore	rec;
+
+	cur = xfs_inobt_init_cursor(mp, tp, agbp, agno, btnum);
+
+	/* the new record is pre-aligned so we know where to look */
+	error = xfs_inobt_lookup(cur, nrec->ir_startino, XFS_LOOKUP_EQ, &i);
+	if (error)
+		goto error;
+	/* if nothing there, insert a new record and return */
+	if (i == 0) {
+		error = xfs_inobt_insert_rec(cur, nrec->ir_holemask,
+					     nrec->ir_count, nrec->ir_freecount,
+					     nrec->ir_free, &i);
+		if (error)
+			goto error;
+		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
+
+		goto out;
+	}
+
+	/*
+	 * A record exists at this startino. Merge or replace the record
+	 * depending on what we've been asked to do.
+	 */
+	if (merge) {
+		error = xfs_inobt_get_rec(cur, &rec, &i);
+		if (error)
+			goto error;
+		XFS_WANT_CORRUPTED_GOTO(mp, i == 1, error);
+		XFS_WANT_CORRUPTED_GOTO(mp,
+					rec.ir_startino == nrec->ir_startino,
+					error);
+
+		/*
+		 * This should never fail. If we have coexisting records that
+		 * cannot merge, something is seriously wrong.
+		 */
+		XFS_WANT_CORRUPTED_GOTO(mp, __xfs_inobt_can_merge(nrec, &rec),
+					error);
+
+		trace_xfs_irec_merge_pre(mp, agno, rec.ir_startino,
+					 rec.ir_holemask, nrec->ir_startino,
+					 nrec->ir_holemask);
+
+		/* merge to nrec to output the updated record */
+		__xfs_inobt_rec_merge(nrec, &rec);
+
+		trace_xfs_irec_merge_post(mp, agno, nrec->ir_startino,
+					  nrec->ir_holemask);
+
+		error = xfs_inobt_rec_check_count(mp, nrec);
+		if (error)
+			goto error;
+	}
+
+	error = xfs_inobt_update(cur, nrec);
+	if (error)
+		goto error;
+
+out:
+	xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
+	return 0;
+error:
+	xfs_btree_del_cursor(cur, XFS_BTREE_ERROR);
+	return error;
+}
+
 /*
  * Allocate new inodes in the allocation group specified by agbp.
  * Return 0 for success, else error code.
@@ -364,11 +603,22 @@ xfs_ialloc_ag_alloc(
 	xfs_agino_t	newlen;		/* new number of inodes */
 	int		isaligned = 0;	/* inode allocation at stripe unit */
 					/* boundary */
+	uint16_t	allocmask = (uint16_t) -1; /* init. to full chunk */
+	struct xfs_inobt_rec_incore rec;
 	struct xfs_perag *pag;
 
+	int		do_sparse = 0;
+
+#ifdef DEBUG
+	/* randomly do sparse inode allocations */
+	if (xfs_sb_version_hassparseinodes(&tp->t_mountp->m_sb))
+		do_sparse = prandom_u32() & 1;
+#endif
+
 	memset(&args, 0, sizeof(args));
 	args.tp = tp;
 	args.mp = tp->t_mountp;
+	args.fsbno = NULLFSBLOCK;
 
 	/*
 	 * Locking will ensure that we don't have two callers in here
@@ -390,6 +640,8 @@ xfs_ialloc_ag_alloc(
 	agno = be32_to_cpu(agi->agi_seqno);
 	args.agbno = XFS_AGINO_TO_AGBNO(args.mp, newino) +
 		     args.mp->m_ialloc_blks;
+	if (do_sparse)
+		goto sparse_alloc;
 	if (likely(newino != NULLAGINO &&
 		  (args.agbno < be32_to_cpu(agi->agi_length)))) {
 		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
@@ -428,8 +680,7 @@ xfs_ialloc_ag_alloc(
 		 * subsequent requests.
 		 */
 		args.minalignslop = 0;
-	} else
+	}
-		args.fsbno = NULLFSBLOCK;
 
 	if (unlikely(args.fsbno == NULLFSBLOCK)) {
 		/*
@@ -480,6 +731,46 @@ xfs_ialloc_ag_alloc(
 			return error;
 	}
 
+	/*
+	 * Finally, try a sparse allocation if the filesystem supports it and
+	 * the sparse allocation length is smaller than a full chunk.
+	 */
+	if (xfs_sb_version_hassparseinodes(&args.mp->m_sb) &&
+	    args.mp->m_ialloc_min_blks < args.mp->m_ialloc_blks &&
+	    args.fsbno == NULLFSBLOCK) {
+sparse_alloc:
+		args.type = XFS_ALLOCTYPE_NEAR_BNO;
+		args.agbno = be32_to_cpu(agi->agi_root);
+		args.fsbno = XFS_AGB_TO_FSB(args.mp, agno, args.agbno);
+		args.alignment = args.mp->m_sb.sb_spino_align;
+		args.prod = 1;
+
+		args.minlen = args.mp->m_ialloc_min_blks;
+		args.maxlen = args.minlen;
+
+		/*
+		 * The inode record will be aligned to full chunk size. We must
+		 * prevent sparse allocation from AG boundaries that result in
+		 * invalid inode records, such as records that start at agbno 0
+		 * or extend beyond the AG.
+		 *
+		 * Set min agbno to the first aligned, non-zero agbno and max to
+		 * the last aligned agbno that is at least one full chunk from
+		 * the end of the AG.
+		 */
+		args.min_agbno = args.mp->m_sb.sb_inoalignmt;
+		args.max_agbno = round_down(args.mp->m_sb.sb_agblocks,
+					    args.mp->m_sb.sb_inoalignmt) -
+				 args.mp->m_ialloc_blks;
+
+		error = xfs_alloc_vextent(&args);
+		if (error)
+			return error;
+
+		newlen = args.len << args.mp->m_sb.sb_inopblog;
+		allocmask = (1 << (newlen / XFS_INODES_PER_HOLEMASK_BIT)) - 1;
+	}
+
 	if (args.fsbno == NULLFSBLOCK) {
 		*alloc = 0;
 		return 0;
@@ -495,8 +786,8 @@ xfs_ialloc_ag_alloc(
 	 * rather than a linear progression to prevent the next generation
 	 * number from being easily guessable.
 	 */
-	error = xfs_ialloc_inode_init(args.mp, tp, NULL, agno, args.agbno,
+	error = xfs_ialloc_inode_init(args.mp, tp, NULL, newlen, agno,
-			args.len, prandom_u32());
+			args.agbno, args.len, prandom_u32());
 
 	if (error)
 		return error;
@@ -504,6 +795,73 @@ xfs_ialloc_ag_alloc(
 	 * Convert the results.
 	 */
 	newino = XFS_OFFBNO_TO_AGINO(args.mp, args.agbno, 0);
+
+	if (xfs_inobt_issparse(~allocmask)) {
+		/*
+		 * We've allocated a sparse chunk. Align the startino and mask.
+		 */
+		xfs_align_sparse_ino(args.mp, &newino, &allocmask);
+
+		rec.ir_startino = newino;
+		rec.ir_holemask = ~allocmask;
+		rec.ir_count = newlen;
+		rec.ir_freecount = newlen;
+		rec.ir_free = XFS_INOBT_ALL_FREE;
+
+		/*
+		 * Insert the sparse record into the inobt and allow for a merge
+		 * if necessary. If a merge does occur, rec is updated to the
+		 * merged record.
+		 */
+		error = xfs_inobt_insert_sprec(args.mp, tp, agbp, XFS_BTNUM_INO,
+					       &rec, true);
+		if (error == -EFSCORRUPTED) {
+			xfs_alert(args.mp,
+	"invalid sparse inode record: ino 0x%llx holemask 0x%x count %u",
+				  XFS_AGINO_TO_INO(args.mp, agno,
+						   rec.ir_startino),
+				  rec.ir_holemask, rec.ir_count);
+			xfs_force_shutdown(args.mp, SHUTDOWN_CORRUPT_INCORE);
+		}
+		if (error)
+			return error;
+
+		/*
+		 * We can't merge the part we've just allocated as for the inobt
+		 * due to finobt semantics. The original record may or may not
+		 * exist independent of whether physical inodes exist in this
+		 * sparse chunk.
+		 *
+		 * We must update the finobt record based on the inobt record.
+		 * rec contains the fully merged and up to date inobt record
+		 * from the previous call. Set merge false to replace any
+		 * existing record with this one.
+		 */
+		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
+			error = xfs_inobt_insert_sprec(args.mp, tp, agbp,
+						       XFS_BTNUM_FINO, &rec,
+						       false);
+			if (error)
+				return error;
+		}
+	} else {
+		/* full chunk - insert new records to both btrees */
+		error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
+					 XFS_BTNUM_INO);
+		if (error)
+			return error;
+
+		if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
+			error = xfs_inobt_insert(args.mp, tp, agbp, newino,
+						 newlen, XFS_BTNUM_FINO);
+			if (error)
+				return error;
+		}
+	}
+
+	/*
+	 * Update AGI counts and newino.
+	 */
 	be32_add_cpu(&agi->agi_count, newlen);
 	be32_add_cpu(&agi->agi_freecount, newlen);
 	pag = xfs_perag_get(args.mp, agno);
@@ -511,20 +869,6 @@ xfs_ialloc_ag_alloc(
 	xfs_perag_put(pag);
 	agi->agi_newino = cpu_to_be32(newino);
 
-	/*
-	 * Insert records describing the new inode chunk into the btrees.
-	 */
-	error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
-				 XFS_BTNUM_INO);
-	if (error)
-		return error;
-
-	if (xfs_sb_version_hasfinobt(&args.mp->m_sb)) {
-		error = xfs_inobt_insert(args.mp, tp, agbp, newino, newlen,
-					 XFS_BTNUM_FINO);
-		if (error)
-			return error;
-	}
 	/*
 	 * Log allocation group header fields
 	 */
@@ -645,7 +989,7 @@ xfs_ialloc_ag_select(
 		 * if we fail allocation due to alignment issues then it is most
 		 * likely a real ENOSPC condition.
 		 */
-		ineed = mp->m_ialloc_blks;
+		ineed = mp->m_ialloc_min_blks;
 		if (flags && ineed > 1)
 			ineed += xfs_ialloc_cluster_alignment(mp);
 		longest = pag->pagf_longest;
@@ -731,6 +1075,27 @@ xfs_ialloc_get_rec(
 	return 0;
 }
 
+/*
+ * Return the offset of the first free inode in the record. If the inode chunk
+ * is sparsely allocated, we convert the record holemask to inode granularity
+ * and mask off the unallocated regions from the inode free mask.
+ */
+STATIC int
+xfs_inobt_first_free_inode(
+	struct xfs_inobt_rec_incore	*rec)
+{
+	xfs_inofree_t			realfree;
+
+	/* if there are no holes, return the first available offset */
+	if (!xfs_inobt_issparse(rec->ir_holemask))
+		return xfs_lowbit64(rec->ir_free);
+
+	realfree = xfs_inobt_irec_to_allocmask(rec);
+	realfree &= rec->ir_free;
+
+	return xfs_lowbit64(realfree);
+}
+
 /*
  * Allocate an inode using the inobt-only algorithm.
  */
@@ -961,7 +1326,7 @@ newino:
 	}
 
 alloc_inode:
-	offset = xfs_lowbit64(rec.ir_free);
+	offset = xfs_inobt_first_free_inode(&rec);
 	ASSERT(offset >= 0);
 	ASSERT(offset < XFS_INODES_PER_CHUNK);
 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
@@ -1210,7 +1575,7 @@ xfs_dialloc_ag(
 	if (error)
 		goto error_cur;
 
-	offset = xfs_lowbit64(rec.ir_free);
+	offset = xfs_inobt_first_free_inode(&rec);
 	ASSERT(offset >= 0);
 	ASSERT(offset < XFS_INODES_PER_CHUNK);
 	ASSERT((XFS_AGINO_TO_OFFSET(mp, rec.ir_startino) %
@@ -1439,6 +1804,83 @@ out_error:
 	return error;
 }
 
+/*
+ * Free the blocks of an inode chunk. We must consider that the inode chunk
+ * might be sparse and only free the regions that are allocated as part of the
+ * chunk.
+ */
+STATIC void
+xfs_difree_inode_chunk(
+	struct xfs_mount		*mp,
+	xfs_agnumber_t			agno,
+	struct xfs_inobt_rec_incore	*rec,
+	struct xfs_bmap_free		*flist)
+{
+	xfs_agblock_t	sagbno = XFS_AGINO_TO_AGBNO(mp, rec->ir_startino);
+	int		startidx, endidx;
+	int		nextbit;
+	xfs_agblock_t	agbno;
+	int		contigblk;
+	DECLARE_BITMAP(holemask, XFS_INOBT_HOLEMASK_BITS);
+
+	if (!xfs_inobt_issparse(rec->ir_holemask)) {
+		/* not sparse, calculate extent info directly */
+		xfs_bmap_add_free(XFS_AGB_TO_FSB(mp, agno,
+				  XFS_AGINO_TO_AGBNO(mp, rec->ir_startino)),
+				  mp->m_ialloc_blks, flist, mp);
+		return;
+	}
+
+	/* holemask is only 16-bits (fits in an unsigned long) */
+	ASSERT(sizeof(rec->ir_holemask) <= sizeof(holemask[0]));
+	holemask[0] = rec->ir_holemask;
+
+	/*
+	 * Find contiguous ranges of zeroes (i.e., allocated regions) in the
+	 * holemask and convert the start/end index of each range to an extent.
+	 * We start with the start and end index both pointing at the first 0 in
+	 * the mask.
+	 */
+	startidx = endidx = find_first_zero_bit(holemask,
+						XFS_INOBT_HOLEMASK_BITS);
+	nextbit = startidx + 1;
+	while (startidx < XFS_INOBT_HOLEMASK_BITS) {
+		nextbit = find_next_zero_bit(holemask, XFS_INOBT_HOLEMASK_BITS,
+					     nextbit);
+		/*
+		 * If the next zero bit is contiguous, update the end index of
+		 * the current range and continue.
+		 */
+		if (nextbit != XFS_INOBT_HOLEMASK_BITS &&
+		    nextbit == endidx + 1) {
+			endidx = nextbit;
+			goto next;
+		}
+
+		/*
+		 * nextbit is not contiguous with the current end index. Convert
+		 * the current start/end to an extent and add it to the free
+		 * list.
+		 */
+		agbno = sagbno + (startidx * XFS_INODES_PER_HOLEMASK_BIT) /
+				  mp->m_sb.sb_inopblock;
+		contigblk = ((endidx - startidx + 1) *
+			     XFS_INODES_PER_HOLEMASK_BIT) /
+			    mp->m_sb.sb_inopblock;
+
+		ASSERT(agbno % mp->m_sb.sb_spino_align == 0);
+		ASSERT(contigblk % mp->m_sb.sb_spino_align == 0);
+		xfs_bmap_add_free(XFS_AGB_TO_FSB(mp, agno, agbno), contigblk,
+				  flist, mp);
+
+		/* reset range to current bit and carry on... */
+		startidx = endidx = nextbit;
+
+next:
+		nextbit++;
+	}
+}
+
 STATIC int
 xfs_difree_inobt(
 	struct xfs_mount		*mp,
@@ -1446,8 +1888,7 @@ xfs_difree_inobt(
 	struct xfs_buf			*agbp,
 	xfs_agino_t			agino,
 	struct xfs_bmap_free		*flist,
-	int				*deleted,
+	struct xfs_icluster		*xic,
-	xfs_ino_t			*first_ino,
 	struct xfs_inobt_rec_incore	*orec)
 {
 	struct xfs_agi			*agi = XFS_BUF_TO_AGI(agbp);
@@ -1501,20 +1942,23 @@ xfs_difree_inobt(
 	rec.ir_freecount++;
 
 	/*
-	 * When an inode cluster is free, it becomes eligible for removal
+	 * When an inode chunk is free, it becomes eligible for removal. Don't
+	 * remove the chunk if the block size is large enough for multiple inode
+	 * chunks (that might not be free).
 	 */
 	if (!(mp->m_flags & XFS_MOUNT_IKEEP) &&
-	    (rec.ir_freecount == mp->m_ialloc_inos)) {
+	    rec.ir_free == XFS_INOBT_ALL_FREE &&
-
+	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK) {
-		*deleted = 1;
+		xic->deleted = 1;
-		*first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
+		xic->first_ino = XFS_AGINO_TO_INO(mp, agno, rec.ir_startino);
+		xic->alloc = xfs_inobt_irec_to_allocmask(&rec);
 
 		/*
 		 * Remove the inode cluster from the AGI B+Tree, adjust the
 		 * AGI and Superblock inode counts, and mark the disk space
 		 * to be freed when the transaction is committed.
 		 */
-		ilen = mp->m_ialloc_inos;
+		ilen = rec.ir_freecount;
 		be32_add_cpu(&agi->agi_count, -ilen);
 		be32_add_cpu(&agi->agi_freecount, -(ilen - 1));
 		xfs_ialloc_log_agi(tp, agbp, XFS_AGI_COUNT | XFS_AGI_FREECOUNT);
@@ -1530,11 +1974,9 @@ xfs_difree_inobt(
 			goto error0;
 		}
 
-		xfs_bmap_add_free(XFS_AGB_TO_FSB(mp, agno,
+		xfs_difree_inode_chunk(mp, agno, &rec, flist);
-				  XFS_AGINO_TO_AGBNO(mp, rec.ir_startino)),
-				  mp->m_ialloc_blks, flist, mp);
 	} else {
-		*deleted = 0;
+		xic->deleted = 0;
 
 		error = xfs_inobt_update(cur, &rec);
 		if (error) {
@@ -1599,7 +2041,9 @@ xfs_difree_finobt(
 		 */
 		XFS_WANT_CORRUPTED_GOTO(mp, ibtrec->ir_freecount == 1, error);
 
-		error = xfs_inobt_insert_rec(cur, ibtrec->ir_freecount,
+		error = xfs_inobt_insert_rec(cur, ibtrec->ir_holemask,
+					     ibtrec->ir_count,
+					     ibtrec->ir_freecount,
 					     ibtrec->ir_free, &i);
 		if (error)
 			goto error;
@@ -1634,8 +2078,13 @@ xfs_difree_finobt(
 	 * free inode. Hence, if all of the inodes are free and we aren't
 	 * keeping inode chunks permanently on disk, remove the record.
 	 * Otherwise, update the record with the new information.
+	 *
+	 * Note that we currently can't free chunks when the block size is large
+	 * enough for multiple chunks. Leave the finobt record to remain in sync
+	 * with the inobt.
 	 */
-	if (rec.ir_freecount == mp->m_ialloc_inos &&
+	if (rec.ir_free == XFS_INOBT_ALL_FREE &&
+	    mp->m_sb.sb_inopblock <= XFS_INODES_PER_CHUNK &&
 	    !(mp->m_flags & XFS_MOUNT_IKEEP)) {
 		error = xfs_btree_delete(cur, &i);
 		if (error)
@@ -1671,8 +2120,7 @@ xfs_difree(
 	struct xfs_trans	*tp,		/* transaction pointer */
 	xfs_ino_t		inode,		/* inode to be freed */
 	struct xfs_bmap_free	*flist,		/* extents to free */
-	int			*deleted,/* set if inode cluster was deleted */
+	struct xfs_icluster	*xic)	/* cluster info if deleted */
-	xfs_ino_t		*first_ino)/* first inode in deleted cluster */
 {
 	/* REFERENCED */
 	xfs_agblock_t		agbno;	/* block number containing inode */
@@ -1723,8 +2171,7 @@ xfs_difree(
 	/*
 	 * Fix up the inode allocation btree.
 	 */
-	error = xfs_difree_inobt(mp, tp, agbp, agino, flist, deleted, first_ino,
+	error = xfs_difree_inobt(mp, tp, agbp, agino, flist, xic, &rec);
-				 &rec);
 	if (error)
 		goto error0;
 

fs/xfs/libxfs/xfs_ialloc.h

@@ -28,6 +28,13 @@ struct xfs_btree_cur;
 /* Move inodes in clusters of this size */
 #define	XFS_INODE_BIG_CLUSTER_SIZE	8192
 
+struct xfs_icluster {
+	bool		deleted;	/* record is deleted */
+	xfs_ino_t	first_ino;	/* first inode number */
+	uint64_t	alloc;		/* inode phys. allocation bitmap for
+					 * sparse chunks */
+};
+
 /* Calculate and return the number of filesystem blocks per inode cluster */
 static inline int
 xfs_icluster_size_fsb(
@@ -90,8 +97,7 @@ xfs_difree(
 	struct xfs_trans *tp,		/* transaction pointer */
 	xfs_ino_t	inode,		/* inode to be freed */
 	struct xfs_bmap_free *flist,	/* extents to free */
-	int		*deleted,	/* set if inode cluster was deleted */
+	struct xfs_icluster *ifree);	/* cluster info if deleted */
-	xfs_ino_t	*first_ino);	/* first inode in deleted cluster */
 
 /*
  * Return the location of the inode in imap, for mapping it into a buffer.
@@ -156,7 +162,7 @@ int xfs_inobt_get_rec(struct xfs_btree_cur *cur,
  * Inode chunk initialisation routine
  */
 int xfs_ialloc_inode_init(struct xfs_mount *mp, struct xfs_trans *tp,
-			  struct list_head *buffer_list,
+			  struct list_head *buffer_list, int icount,
 			  xfs_agnumber_t agno, xfs_agblock_t agbno,
 			  xfs_agblock_t length, unsigned int gen);
 

fs/xfs/libxfs/xfs_ialloc_btree.c

@@ -167,7 +167,16 @@ xfs_inobt_init_rec_from_cur(
 	union xfs_btree_rec	*rec)
 {
 	rec->inobt.ir_startino = cpu_to_be32(cur->bc_rec.i.ir_startino);
-	rec->inobt.ir_freecount = cpu_to_be32(cur->bc_rec.i.ir_freecount);
+	if (xfs_sb_version_hassparseinodes(&cur->bc_mp->m_sb)) {
+		rec->inobt.ir_u.sp.ir_holemask =
+					cpu_to_be16(cur->bc_rec.i.ir_holemask);
+		rec->inobt.ir_u.sp.ir_count = cur->bc_rec.i.ir_count;
+		rec->inobt.ir_u.sp.ir_freecount = cur->bc_rec.i.ir_freecount;
+	} else {
+		/* ir_holemask/ir_count not supported on-disk */
+		rec->inobt.ir_u.f.ir_freecount =
+					cpu_to_be32(cur->bc_rec.i.ir_freecount);
+	}
 	rec->inobt.ir_free = cpu_to_be64(cur->bc_rec.i.ir_free);
 }
 
@@ -418,3 +427,85 @@ xfs_inobt_maxrecs(
 		return blocklen / sizeof(xfs_inobt_rec_t);
 	return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t));
 }
+
+/*
+ * Convert the inode record holemask to an inode allocation bitmap. The inode
+ * allocation bitmap is inode granularity and specifies whether an inode is
+ * physically allocated on disk (not whether the inode is considered allocated
+ * or free by the fs).
+ *
+ * A bit value of 1 means the inode is allocated, a value of 0 means it is free.
+ */
+uint64_t
+xfs_inobt_irec_to_allocmask(
+	struct xfs_inobt_rec_incore	*rec)
+{
+	uint64_t			bitmap = 0;
+	uint64_t			inodespbit;
+	int				nextbit;
+	uint				allocbitmap;
+
+	/*
+	 * The holemask has 16-bits for a 64 inode record. Therefore each
+	 * holemask bit represents multiple inodes. Create a mask of bits to set
+	 * in the allocmask for each holemask bit.
+	 */
+	inodespbit = (1 << XFS_INODES_PER_HOLEMASK_BIT) - 1;
+
+	/*
+	 * Allocated inodes are represented by 0 bits in holemask. Invert the 0
+	 * bits to 1 and convert to a uint so we can use xfs_next_bit(). Mask
+	 * anything beyond the 16 holemask bits since this casts to a larger
+	 * type.
+	 */
+	allocbitmap = ~rec->ir_holemask & ((1 << XFS_INOBT_HOLEMASK_BITS) - 1);
+
+	/*
+	 * allocbitmap is the inverted holemask so every set bit represents
+	 * allocated inodes. To expand from 16-bit holemask granularity to
+	 * 64-bit (e.g., bit-per-inode), set inodespbit bits in the target
+	 * bitmap for every holemask bit.
+	 */
+	nextbit = xfs_next_bit(&allocbitmap, 1, 0);
+	while (nextbit != -1) {
+		ASSERT(nextbit < (sizeof(rec->ir_holemask) * NBBY));
+
+		bitmap |= (inodespbit <<
+			   (nextbit * XFS_INODES_PER_HOLEMASK_BIT));
+
+		nextbit = xfs_next_bit(&allocbitmap, 1, nextbit + 1);
+	}
+
+	return bitmap;
+}
+
+#if defined(DEBUG) || defined(XFS_WARN)
+/*
+ * Verify that an in-core inode record has a valid inode count.
+ */
+int
+xfs_inobt_rec_check_count(
+	struct xfs_mount		*mp,
+	struct xfs_inobt_rec_incore	*rec)
+{
+	int				inocount = 0;
+	int				nextbit = 0;
+	uint64_t			allocbmap;
+	int				wordsz;
+
+	wordsz = sizeof(allocbmap) / sizeof(unsigned int);
+	allocbmap = xfs_inobt_irec_to_allocmask(rec);
+
+	nextbit = xfs_next_bit((uint *) &allocbmap, wordsz, nextbit);
+	while (nextbit != -1) {
+		inocount++;
+		nextbit = xfs_next_bit((uint *) &allocbmap, wordsz,
+				       nextbit + 1);
+	}
+
+	if (inocount != rec->ir_count)
+		return -EFSCORRUPTED;
+
+	return 0;
+}
+#endif	/* DEBUG */

fs/xfs/libxfs/xfs_ialloc_btree.h

@@ -62,4 +62,14 @@ extern struct xfs_btree_cur *xfs_inobt_init_cursor(struct xfs_mount *,
 		xfs_btnum_t);
 extern int xfs_inobt_maxrecs(struct xfs_mount *, int, int);
 
+/* ir_holemask to inode allocation bitmap conversion */
+uint64_t xfs_inobt_irec_to_allocmask(struct xfs_inobt_rec_incore *);
+
+#if defined(DEBUG) || defined(XFS_WARN)
+int xfs_inobt_rec_check_count(struct xfs_mount *,
+			      struct xfs_inobt_rec_incore *);
+#else
+#define xfs_inobt_rec_check_count(mp, rec)	0
+#endif	/* DEBUG */
+
 #endif	/* __XFS_IALLOC_BTREE_H__ */

fs/xfs/libxfs/xfs_sb.c

@@ -174,6 +174,27 @@ xfs_mount_validate_sb(
 			return -EFSCORRUPTED;
 	}
 
+	/*
+	 * Full inode chunks must be aligned to inode chunk size when
+	 * sparse inodes are enabled to support the sparse chunk
+	 * allocation algorithm and prevent overlapping inode records.
+	 */
+	if (xfs_sb_version_hassparseinodes(sbp)) {
+		uint32_t	align;
+
+		xfs_alert(mp,
+	"EXPERIMENTAL sparse inode feature enabled. Use at your own risk!");
+
+		align = XFS_INODES_PER_CHUNK * sbp->sb_inodesize
+				>> sbp->sb_blocklog;
+		if (sbp->sb_inoalignmt != align) {
+			xfs_warn(mp,
+"Inode block alignment (%u) must match chunk size (%u) for sparse inodes.",
+				 sbp->sb_inoalignmt, align);
+			return -EINVAL;
+		}
+	}
+
 	if (unlikely(
 	    sbp->sb_logstart == 0 && mp->m_logdev_targp == mp->m_ddev_targp)) {
 		xfs_warn(mp,
@@ -374,7 +395,7 @@ __xfs_sb_from_disk(
 				be32_to_cpu(from->sb_features_log_incompat);
 	/* crc is only used on disk, not in memory; just init to 0 here. */
 	to->sb_crc = 0;
-	to->sb_pad = 0;
+	to->sb_spino_align = be32_to_cpu(from->sb_spino_align);
 	to->sb_pquotino = be64_to_cpu(from->sb_pquotino);
 	to->sb_lsn = be64_to_cpu(from->sb_lsn);
 	/* Convert on-disk flags to in-memory flags? */
@@ -516,7 +537,7 @@ xfs_sb_to_disk(
 				cpu_to_be32(from->sb_features_incompat);
 		to->sb_features_log_incompat =
 				cpu_to_be32(from->sb_features_log_incompat);
-		to->sb_pad = 0;
+		to->sb_spino_align = cpu_to_be32(from->sb_spino_align);
 		to->sb_lsn = cpu_to_be64(from->sb_lsn);
 	}
 }
@@ -689,6 +710,11 @@ xfs_sb_mount_common(
 	mp->m_ialloc_inos = (int)MAX((__uint16_t)XFS_INODES_PER_CHUNK,
 					sbp->sb_inopblock);
 	mp->m_ialloc_blks = mp->m_ialloc_inos >> sbp->sb_inopblog;
+
+	if (sbp->sb_spino_align)
+		mp->m_ialloc_min_blks = sbp->sb_spino_align;
+	else
+		mp->m_ialloc_min_blks = mp->m_ialloc_blks;
 }
 
 /*

fs/xfs/xfs_fsops.c

@@ -101,7 +101,9 @@ xfs_fs_geometry(
 			(xfs_sb_version_hasftype(&mp->m_sb) ?
 				XFS_FSOP_GEOM_FLAGS_FTYPE : 0) |
 			(xfs_sb_version_hasfinobt(&mp->m_sb) ?
-				XFS_FSOP_GEOM_FLAGS_FINOBT : 0);
+				XFS_FSOP_GEOM_FLAGS_FINOBT : 0) |
+			(xfs_sb_version_hassparseinodes(&mp->m_sb) ?
+				XFS_FSOP_GEOM_FLAGS_SPINODES : 0);
 		geo->logsectsize = xfs_sb_version_hassector(&mp->m_sb) ?
 				mp->m_sb.sb_logsectsize : BBSIZE;
 		geo->rtsectsize = mp->m_sb.sb_blocksize;

fs/xfs/xfs_inode.c

@@ -2235,9 +2235,9 @@ xfs_iunlink_remove(
  */
 STATIC int
 xfs_ifree_cluster(
-	xfs_inode_t	*free_ip,
+	xfs_inode_t		*free_ip,
-	xfs_trans_t	*tp,
+	xfs_trans_t		*tp,
-	xfs_ino_t	inum)
+	struct xfs_icluster	*xic)
 {
 	xfs_mount_t		*mp = free_ip->i_mount;
 	int			blks_per_cluster;
@@ -2250,13 +2250,26 @@ xfs_ifree_cluster(
 	xfs_inode_log_item_t	*iip;
 	xfs_log_item_t		*lip;
 	struct xfs_perag	*pag;
+	xfs_ino_t		inum;
 
+	inum = xic->first_ino;
 	pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
 	blks_per_cluster = xfs_icluster_size_fsb(mp);
 	inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
 	nbufs = mp->m_ialloc_blks / blks_per_cluster;
 
 	for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
+		/*
+		 * The allocation bitmap tells us which inodes of the chunk were
+		 * physically allocated. Skip the cluster if an inode falls into
+		 * a sparse region.
+		 */
+		if ((xic->alloc & XFS_INOBT_MASK(inum - xic->first_ino)) == 0) {
+			ASSERT(((inum - xic->first_ino) %
+				inodes_per_cluster) == 0);
+			continue;
+		}
+
 		blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
 					 XFS_INO_TO_AGBNO(mp, inum));
 
@@ -2414,8 +2427,7 @@ xfs_ifree(
 	xfs_bmap_free_t	*flist)
 {
 	int			error;
-	int			delete;
+	struct xfs_icluster	xic = { 0 };
-	xfs_ino_t		first_ino;
 
 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 	ASSERT(ip->i_d.di_nlink == 0);
@@ -2431,7 +2443,7 @@ xfs_ifree(
 	if (error)
 		return error;
 
-	error = xfs_difree(tp, ip->i_ino, flist, &delete, &first_ino);
+	error = xfs_difree(tp, ip->i_ino, flist, &xic);
 	if (error)
 		return error;
 
@@ -2448,8 +2460,8 @@ xfs_ifree(
 	ip->i_d.di_gen++;
 	xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 
-	if (delete)
+	if (xic.deleted)
-		error = xfs_ifree_cluster(ip, tp, first_ino);
+		error = xfs_ifree_cluster(ip, tp, &xic);
 
 	return error;
 }

fs/xfs/xfs_itable.c

@@ -252,7 +252,7 @@ xfs_bulkstat_grab_ichunk(
 		}
 
 		irec->ir_free |= xfs_inobt_maskn(0, idx);
-		*icount = XFS_INODES_PER_CHUNK - irec->ir_freecount;
+		*icount = irec->ir_count - irec->ir_freecount;
 	}
 
 	return 0;
@@ -415,6 +415,8 @@ xfs_bulkstat(
 				goto del_cursor;
 			if (icount) {
 				irbp->ir_startino = r.ir_startino;
+				irbp->ir_holemask = r.ir_holemask;
+				irbp->ir_count = r.ir_count;
 				irbp->ir_freecount = r.ir_freecount;
 				irbp->ir_free = r.ir_free;
 				irbp++;
@@ -447,13 +449,15 @@ xfs_bulkstat(
 			 * If this chunk has any allocated inodes, save it.
 			 * Also start read-ahead now for this chunk.
 			 */
-			if (r.ir_freecount < XFS_INODES_PER_CHUNK) {
+			if (r.ir_freecount < r.ir_count) {
 				xfs_bulkstat_ichunk_ra(mp, agno, &r);
 				irbp->ir_startino = r.ir_startino;
+				irbp->ir_holemask = r.ir_holemask;
+				irbp->ir_count = r.ir_count;
 				irbp->ir_freecount = r.ir_freecount;
 				irbp->ir_free = r.ir_free;
 				irbp++;
-				icount += XFS_INODES_PER_CHUNK - r.ir_freecount;
+				icount += r.ir_count - r.ir_freecount;
 			}
 			error = xfs_btree_increment(cur, 0, &stat);
 			if (error || stat == 0) {
@@ -599,8 +603,7 @@ xfs_inumbers(
 		agino = r.ir_startino + XFS_INODES_PER_CHUNK - 1;
 		buffer[bufidx].xi_startino =
 			XFS_AGINO_TO_INO(mp, agno, r.ir_startino);
-		buffer[bufidx].xi_alloccount =
+		buffer[bufidx].xi_alloccount = r.ir_count - r.ir_freecount;
-			XFS_INODES_PER_CHUNK - r.ir_freecount;
 		buffer[bufidx].xi_allocmask = ~r.ir_free;
 		if (++bufidx == bcount) {
 			long	written;

fs/xfs/xfs_log_recover.c

@@ -3068,12 +3068,22 @@ xlog_recover_do_icreate_pass2(
 		return -EINVAL;
 	}
 
-	/* existing allocation is fixed value */
+	/*
-	ASSERT(count == mp->m_ialloc_inos);
+	 * The inode chunk is either full or sparse and we only support
-	ASSERT(length == mp->m_ialloc_blks);
+	 * m_ialloc_min_blks sized sparse allocations at this time.
-	if (count != mp->m_ialloc_inos ||
+	 */
-	     length != mp->m_ialloc_blks) {
+	if (length != mp->m_ialloc_blks &&
-		xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count 2");
+	    length != mp->m_ialloc_min_blks) {
+		xfs_warn(log->l_mp,
+			 "%s: unsupported chunk length", __FUNCTION__);
+		return -EINVAL;
+	}
+
+	/* verify inode count is consistent with extent length */
+	if ((count >> mp->m_sb.sb_inopblog) != length) {
+		xfs_warn(log->l_mp,
+			 "%s: inconsistent inode count and chunk length",
+			 __FUNCTION__);
 		return -EINVAL;
 	}
 
@@ -3091,8 +3101,8 @@ xlog_recover_do_icreate_pass2(
 			XFS_AGB_TO_DADDR(mp, agno, agbno), length, 0))
 		return 0;
 
-	xfs_ialloc_inode_init(mp, NULL, buffer_list, agno, agbno, length,
+	xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno, length,
-					be32_to_cpu(icl->icl_gen));
+			      be32_to_cpu(icl->icl_gen));
 	return 0;
 }
 

fs/xfs/xfs_mount.c

@@ -724,6 +724,22 @@ xfs_mountfs(
 			mp->m_inode_cluster_size = new_size;
 	}
 
+	/*
+	 * If enabled, sparse inode chunk alignment is expected to match the
+	 * cluster size. Full inode chunk alignment must match the chunk size,
+	 * but that is checked on sb read verification...
+	 */
+	if (xfs_sb_version_hassparseinodes(&mp->m_sb) &&
+	    mp->m_sb.sb_spino_align !=
+			XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size)) {
+		xfs_warn(mp,
+	"Sparse inode block alignment (%u) must match cluster size (%llu).",
+			 mp->m_sb.sb_spino_align,
+			 XFS_B_TO_FSBT(mp, mp->m_inode_cluster_size));
+		error = -EINVAL;
+		goto out_remove_uuid;
+	}
+
 	/*
 	 * Set inode alignment fields
 	 */

fs/xfs/xfs_mount.h

@@ -101,6 +101,8 @@ typedef struct xfs_mount {
 	__uint64_t		m_flags;	/* global mount flags */
 	int			m_ialloc_inos;	/* inodes in inode allocation */
 	int			m_ialloc_blks;	/* blocks in inode allocation */
+	int			m_ialloc_min_blks;/* min blocks in sparse inode
+						   * allocation */
 	int			m_inoalign_mask;/* mask sb_inoalignmt if used */
 	uint			m_qflags;	/* quota status flags */
 	struct xfs_trans_resv	m_resv;		/* precomputed res values */

fs/xfs/xfs_trace.h

@@ -738,6 +738,53 @@ TRACE_EVENT(xfs_iomap_prealloc_size,
 		  __entry->blocks, __entry->shift, __entry->writeio_blocks)
 )
 
+TRACE_EVENT(xfs_irec_merge_pre,
+	TP_PROTO(struct xfs_mount *mp, xfs_agnumber_t agno, xfs_agino_t agino,
+		 uint16_t holemask, xfs_agino_t nagino, uint16_t nholemask),
+	TP_ARGS(mp, agno, agino, holemask, nagino, nholemask),
+	TP_STRUCT__entry(
+		__field(dev_t, dev)
+		__field(xfs_agnumber_t, agno)
+		__field(xfs_agino_t, agino)
+		__field(uint16_t, holemask)
+		__field(xfs_agino_t, nagino)
+		__field(uint16_t, nholemask)
+	),
+	TP_fast_assign(
+		__entry->dev = mp->m_super->s_dev;
+		__entry->agno = agno;
+		__entry->agino = agino;
+		__entry->holemask = holemask;
+		__entry->nagino = nagino;
+		__entry->nholemask = holemask;
+	),
+	TP_printk("dev %d:%d agno %d inobt (%u:0x%x) new (%u:0x%x)",
+		  MAJOR(__entry->dev), MINOR(__entry->dev), __entry->agno,
+		  __entry->agino, __entry->holemask, __entry->nagino,
+		  __entry->nholemask)
+)
+
+TRACE_EVENT(xfs_irec_merge_post,
+	TP_PROTO(struct xfs_mount *mp, xfs_agnumber_t agno, xfs_agino_t agino,
+		 uint16_t holemask),
+	TP_ARGS(mp, agno, agino, holemask),
+	TP_STRUCT__entry(
+		__field(dev_t, dev)
+		__field(xfs_agnumber_t, agno)
+		__field(xfs_agino_t, agino)
+		__field(uint16_t, holemask)
+	),
+	TP_fast_assign(
+		__entry->dev = mp->m_super->s_dev;
+		__entry->agno = agno;
+		__entry->agino = agino;
+		__entry->holemask = holemask;
+	),
+	TP_printk("dev %d:%d agno %d inobt (%u:0x%x)", MAJOR(__entry->dev),
+		  MINOR(__entry->dev), __entry->agno, __entry->agino,
+		  __entry->holemask)
+)
+
 #define DEFINE_IREF_EVENT(name) \
 DEFINE_EVENT(xfs_iref_class, name, \
 	TP_PROTO(struct xfs_inode *ip, unsigned long caller_ip), \