GFS2: Reduce file fragmentation

This patch reduces GFS2 file fragmentation by pre-reserving blocks. The
resulting improved on disk layout greatly speeds up operations in cases
which would have resulted in interlaced allocation of blocks previously.
A typical example of this is 10 parallel dd processes, each writing to a
file in a common dirctory.

The implementation uses an rbtree of reservations attached to each
resource group (and each inode).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>

fs/gfs2/bmap.c

@@ -785,6 +785,9 @@ static int do_strip(struct gfs2_inode *ip, struct buffer_head *dibh,
 	if (error)
 		goto out_rlist;
 
+	if (gfs2_rs_active(ip->i_res)) /* needs to be done with the rgrp glock held */
+		gfs2_rs_deltree(ip->i_res);
+
 	error = gfs2_trans_begin(sdp, rg_blocks + RES_DINODE +
 				 RES_INDIRECT + RES_STATFS + RES_QUOTA,
 				 revokes);

fs/gfs2/file.c

@@ -383,6 +383,9 @@ static int gfs2_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 	if (ret)
 		return ret;
 
+	atomic_set(&ip->i_res->rs_sizehint,
+		   PAGE_CACHE_SIZE / sdp->sd_sb.sb_bsize);
+
 	gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &gh);
 	ret = gfs2_glock_nq(&gh);
 	if (ret)
@@ -571,22 +574,15 @@ static int gfs2_open(struct inode *inode, struct file *file)
 
 static int gfs2_release(struct inode *inode, struct file *file)
 {
-	struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
-	struct gfs2_file *fp;
 	struct gfs2_inode *ip = GFS2_I(inode);
 
-	fp = file->private_data;
+	kfree(file->private_data);
 	file->private_data = NULL;
 
-	if ((file->f_mode & FMODE_WRITE) && ip->i_res &&
+	if ((file->f_mode & FMODE_WRITE) &&
 	    (atomic_read(&inode->i_writecount) == 1))
 		gfs2_rs_delete(ip);
 
-	if (gfs2_assert_warn(sdp, fp))
-		return -EIO;
-
-	kfree(fp);
-
 	return 0;
 }
 
@@ -662,14 +658,18 @@ static ssize_t gfs2_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 				   unsigned long nr_segs, loff_t pos)
 {
 	struct file *file = iocb->ki_filp;
+	size_t writesize = iov_length(iov, nr_segs);
 	struct dentry *dentry = file->f_dentry;
 	struct gfs2_inode *ip = GFS2_I(dentry->d_inode);
+	struct gfs2_sbd *sdp;
 	int ret;
 
+	sdp = GFS2_SB(file->f_mapping->host);
 	ret = gfs2_rs_alloc(ip);
 	if (ret)
 		return ret;
 
+	atomic_set(&ip->i_res->rs_sizehint, writesize / sdp->sd_sb.sb_bsize);
 	if (file->f_flags & O_APPEND) {
 		struct gfs2_holder gh;
 
@@ -795,6 +795,8 @@ static long gfs2_fallocate(struct file *file, int mode, loff_t offset,
 	if (unlikely(error))
 		goto out_uninit;
 
+	atomic_set(&ip->i_res->rs_sizehint, len / sdp->sd_sb.sb_bsize);
+
 	while (len > 0) {
 		if (len < bytes)
 			bytes = len;
@@ -803,10 +805,6 @@ static long gfs2_fallocate(struct file *file, int mode, loff_t offset,
 			offset += bytes;
 			continue;
 		}
-		error = gfs2_rindex_update(sdp);
-		if (error)
-			goto out_unlock;
-
 		error = gfs2_quota_lock_check(ip);
 		if (error)
 			goto out_unlock;

fs/gfs2/incore.h

@@ -84,6 +84,7 @@ struct gfs2_rgrpd {
 	u32 rd_data;			/* num of data blocks in rgrp */
 	u32 rd_bitbytes;		/* number of bytes in data bitmaps */
 	u32 rd_free;
+	u32 rd_reserved;                /* number of blocks reserved */
 	u32 rd_free_clone;
 	u32 rd_dinodes;
 	u64 rd_igeneration;
@@ -96,6 +97,9 @@ struct gfs2_rgrpd {
 #define GFS2_RDF_UPTODATE	0x20000000 /* rg is up to date */
 #define GFS2_RDF_ERROR		0x40000000 /* error in rg */
 #define GFS2_RDF_MASK		0xf0000000 /* mask for internal flags */
+	spinlock_t rd_rsspin;           /* protects reservation related vars */
+	struct rb_root rd_rstree;       /* multi-block reservation tree */
+	u32 rd_rs_cnt;                  /* count of current reservations */
 };
 
 enum gfs2_state_bits {
@@ -233,6 +237,38 @@ struct gfs2_holder {
 	unsigned long gh_ip;
 };
 
+/* Resource group multi-block reservation, in order of appearance:
+
+   Step 1. Function prepares to write, allocates a mb, sets the size hint.
+   Step 2. User calls inplace_reserve to target an rgrp, sets the rgrp info
+   Step 3. Function get_local_rgrp locks the rgrp, determines which bits to use
+   Step 4. Bits are assigned from the rgrp based on either the reservation
+           or wherever it can.
+*/
+
+struct gfs2_blkreserv {
+	/* components used during write (step 1): */
+	atomic_t rs_sizehint;         /* hint of the write size */
+
+	/* components used during inplace_reserve (step 2): */
+	u32 rs_requested; /* Filled in by caller of gfs2_inplace_reserve() */
+
+	/* components used during get_local_rgrp (step 3): */
+	struct gfs2_rgrpd *rs_rgd;    /* pointer to the gfs2_rgrpd */
+	struct gfs2_holder rs_rgd_gh; /* Filled in by get_local_rgrp */
+	struct rb_node rs_node;       /* link to other block reservations */
+
+	/* components used during block searches and assignments (step 4): */
+	struct gfs2_bitmap *rs_bi;    /* bitmap for the current allocation */
+	u32 rs_biblk;                 /* start block relative to the bi */
+	u32 rs_free;                  /* how many blocks are still free */
+
+	/* ancillary quota stuff */
+	struct gfs2_quota_data *rs_qa_qd[2 * MAXQUOTAS];
+	struct gfs2_holder rs_qa_qd_ghs[2 * MAXQUOTAS];
+	unsigned int rs_qa_qd_num;
+};
+
 enum {
 	GLF_LOCK			= 1,
 	GLF_DEMOTE			= 3,
@@ -290,16 +326,6 @@ struct gfs2_glock {
 
 #define GFS2_MIN_LVB_SIZE 32	/* Min size of LVB that gfs2 supports */
 
-struct gfs2_blkreserv {
-	u32 rs_requested; /* Filled in by caller of gfs2_inplace_reserve() */
-	struct gfs2_holder rs_rgd_gh; /* Filled in by gfs2_inplace_reserve() */
-
-	/* ancillary quota stuff */
-	struct gfs2_quota_data *rs_qa_qd[2 * MAXQUOTAS];
-	struct gfs2_holder rs_qa_qd_ghs[2 * MAXQUOTAS];
-	unsigned int rs_qa_qd_num;
-};
-
 enum {
 	GIF_INVALID		= 0,
 	GIF_QD_LOCKED		= 1,
@@ -307,7 +333,6 @@ enum {
 	GIF_SW_PAGED		= 3,
 };
 
-
 struct gfs2_inode {
 	struct inode i_inode;
 	u64 i_no_addr;
@@ -318,7 +343,7 @@ struct gfs2_inode {
 	struct gfs2_glock *i_gl; /* Move into i_gh? */
 	struct gfs2_holder i_iopen_gh;
 	struct gfs2_holder i_gh; /* for prepare/commit_write only */
-	struct gfs2_blkreserv *i_res; /* resource group block reservation */
+	struct gfs2_blkreserv *i_res; /* rgrp multi-block reservation */
 	struct gfs2_rgrpd *i_rgd;
 	u64 i_goal;	/* goal block for allocations */
 	struct rw_semaphore i_rw_mutex;

fs/gfs2/inode.c

@@ -521,6 +521,9 @@ static int make_dinode(struct gfs2_inode *dip, struct gfs2_glock *gl,
 	int error;
 
 	munge_mode_uid_gid(dip, &mode, &uid, &gid);
+	error = gfs2_rindex_update(sdp);
+	if (error)
+		return error;
 
 	error = gfs2_quota_lock(dip, uid, gid);
 	if (error)
@@ -551,6 +554,10 @@ static int link_dinode(struct gfs2_inode *dip, const struct qstr *name,
 	struct buffer_head *dibh;
 	int error;
 
+	error = gfs2_rindex_update(sdp);
+	if (error)
+		return error;
+
 	error = gfs2_quota_lock(dip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 	if (error)
 		goto fail;
@@ -596,7 +603,8 @@ static int link_dinode(struct gfs2_inode *dip, const struct qstr *name,
 	gfs2_trans_end(sdp);
 
 fail_ipreserv:
-	gfs2_inplace_release(dip);
+	if (alloc_required)
+		gfs2_inplace_release(dip);
 
 fail_quota_locks:
 	gfs2_quota_unlock(dip);
@@ -647,7 +655,7 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,
 	const struct qstr *name = &dentry->d_name;
 	struct gfs2_holder ghs[2];
 	struct inode *inode = NULL;
-	struct gfs2_inode *dip = GFS2_I(dir);
+	struct gfs2_inode *dip = GFS2_I(dir), *ip;
 	struct gfs2_sbd *sdp = GFS2_SB(&dip->i_inode);
 	struct gfs2_inum_host inum = { .no_addr = 0, .no_formal_ino = 0 };
 	int error;
@@ -657,6 +665,11 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,
 	if (!name->len || name->len > GFS2_FNAMESIZE)
 		return -ENAMETOOLONG;
 
+	/* We need a reservation to allocate the new dinode block. The
+	   directory ip temporarily points to the reservation, but this is
+	   being done to get a set of contiguous blocks for the new dinode.
+	   Since this is a create, we don't have a sizehint yet, so it will
+	   have to use the minimum reservation size. */
 	error = gfs2_rs_alloc(dip);
 	if (error)
 		return error;
@@ -694,24 +707,29 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,
 	if (IS_ERR(inode))
 		goto fail_gunlock2;
 
-	error = gfs2_inode_refresh(GFS2_I(inode));
+	ip = GFS2_I(inode);
+	error = gfs2_inode_refresh(ip);
 	if (error)
 		goto fail_gunlock2;
 
-	/* the new inode needs a reservation so it can allocate xattrs. */
-	error = gfs2_rs_alloc(GFS2_I(inode));
-	if (error)
-		goto fail_gunlock2;
+	/* The newly created inode needs a reservation so it can allocate
+	   xattrs. At the same time, we want new blocks allocated to the new
+	   dinode to be as contiguous as possible. Since we allocated the
+	   dinode block under the directory's reservation, we transfer
+	   ownership of that reservation to the new inode. The directory
+	   doesn't need a reservation unless it needs a new allocation. */
+	ip->i_res = dip->i_res;
+	dip->i_res = NULL;
 
 	error = gfs2_acl_create(dip, inode);
 	if (error)
 		goto fail_gunlock2;
 
-	error = gfs2_security_init(dip, GFS2_I(inode), name);
+	error = gfs2_security_init(dip, ip, name);
 	if (error)
 		goto fail_gunlock2;
 
-	error = link_dinode(dip, name, GFS2_I(inode));
+	error = link_dinode(dip, name, ip);
 	if (error)
 		goto fail_gunlock2;
 
@@ -738,6 +756,7 @@ static int gfs2_create_inode(struct inode *dir, struct dentry *dentry,
 		iput(inode);
 	}
 fail:
+	gfs2_rs_delete(dip);
 	if (bh)
 		brelse(bh);
 	return error;

fs/gfs2/rgrp.h

@@ -13,6 +13,14 @@
 #include <linux/slab.h>
 #include <linux/uaccess.h>
 
+/* Since each block in the file system is represented by two bits in the
+ * bitmap, one 64-bit word in the bitmap will represent 32 blocks.
+ * By reserving 32 blocks at a time, we can optimize / shortcut how we search
+ * through the bitmaps by looking a word at a time.
+ */
+#define RGRP_RSRV_MINBYTES 8
+#define RGRP_RSRV_MINBLKS ((u32)(RGRP_RSRV_MINBYTES * GFS2_NBBY))
+
 struct gfs2_rgrpd;
 struct gfs2_sbd;
 struct gfs2_holder;
@@ -29,6 +37,8 @@ extern void gfs2_free_clones(struct gfs2_rgrpd *rgd);
 extern int gfs2_rgrp_go_lock(struct gfs2_holder *gh);
 extern void gfs2_rgrp_go_unlock(struct gfs2_holder *gh);
 
+extern struct gfs2_alloc *gfs2_alloc_get(struct gfs2_inode *ip);
+
 extern int gfs2_inplace_reserve(struct gfs2_inode *ip, u32 requested);
 extern void gfs2_inplace_release(struct gfs2_inode *ip);
 
@@ -36,6 +46,7 @@ extern int gfs2_alloc_blocks(struct gfs2_inode *ip, u64 *bn, unsigned int *n,
 			     bool dinode, u64 *generation);
 
 extern int gfs2_rs_alloc(struct gfs2_inode *ip);
+extern void gfs2_rs_deltree(struct gfs2_blkreserv *rs);
 extern void gfs2_rs_delete(struct gfs2_inode *ip);
 extern void __gfs2_free_blocks(struct gfs2_inode *ip, u64 bstart, u32 blen, int meta);
 extern void gfs2_free_meta(struct gfs2_inode *ip, u64 bstart, u32 blen);
@@ -62,7 +73,7 @@ extern int gfs2_rgrp_send_discards(struct gfs2_sbd *sdp, u64 offset,
 				   const struct gfs2_bitmap *bi, unsigned minlen, u64 *ptrimmed);
 extern int gfs2_fitrim(struct file *filp, void __user *argp);
 
-/* This is how to tell if a reservation is "inplace" reserved: */
+/* This is how to tell if a multi-block reservation is "inplace" reserved: */
 static inline int gfs2_mb_reserved(struct gfs2_inode *ip)
 {
 	if (ip->i_res && ip->i_res->rs_requested)
@@ -70,4 +81,22 @@ static inline int gfs2_mb_reserved(struct gfs2_inode *ip)
 	return 0;
 }
 
+/* This is how to tell if a multi-block reservation is in the rgrp tree: */
+static inline int gfs2_rs_active(struct gfs2_blkreserv *rs)
+{
+	if (rs && rs->rs_bi)
+		return 1;
+	return 0;
+}
+
+static inline u32 gfs2_bi2rgd_blk(const struct gfs2_bitmap *bi, u32 blk)
+{
+	return (bi->bi_start * GFS2_NBBY) + blk;
+}
+
+static inline u64 gfs2_rs_startblk(const struct gfs2_blkreserv *rs)
+{
+	return gfs2_bi2rgd_blk(rs->rs_bi, rs->rs_biblk) + rs->rs_rgd->rd_data0;
+}
+
 #endif /* __RGRP_DOT_H__ */

fs/gfs2/super.c

@@ -1420,6 +1420,10 @@ static int gfs2_dinode_dealloc(struct gfs2_inode *ip)
 		return -EIO;
 	}
 
+	error = gfs2_rindex_update(sdp);
+	if (error)
+		return error;
+
 	error = gfs2_quota_hold(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 	if (error)
 		return error;
@@ -1550,6 +1554,9 @@ static void gfs2_evict_inode(struct inode *inode)
 
 out_unlock:
 	/* Error path for case 1 */
+	if (gfs2_rs_active(ip->i_res))
+		gfs2_rs_deltree(ip->i_res);
+
 	if (test_bit(HIF_HOLDER, &ip->i_iopen_gh.gh_iflags))
 		gfs2_glock_dq(&ip->i_iopen_gh);
 	gfs2_holder_uninit(&ip->i_iopen_gh);

fs/gfs2/trace_gfs2.h

@@ -14,6 +14,7 @@
 #include <linux/ktime.h>
 #include "incore.h"
 #include "glock.h"
+#include "rgrp.h"
 
 #define dlm_state_name(nn) { DLM_LOCK_##nn, #nn }
 #define glock_trace_name(x) __print_symbolic(x,		\
@@ -31,6 +32,17 @@
 			    { GFS2_BLKST_DINODE, "dinode" },	\
 			    { GFS2_BLKST_UNLINKED, "unlinked" })
 
+#define TRACE_RS_DELETE  0
+#define TRACE_RS_TREEDEL 1
+#define TRACE_RS_INSERT  2
+#define TRACE_RS_CLAIM   3
+
+#define rs_func_name(x) __print_symbolic(x,	\
+					 { 0, "del " },	\
+					 { 1, "tdel" },	\
+					 { 2, "ins " },	\
+					 { 3, "clm " })
+
 #define show_glock_flags(flags) __print_flags(flags, "",	\
 	{(1UL << GLF_LOCK),			"l" },		\
 	{(1UL << GLF_DEMOTE),			"D" },		\
@@ -470,6 +482,7 @@ TRACE_EVENT(gfs2_block_alloc,
 		__field(	u8,	block_state		)
 		__field(        u64,	rd_addr			)
 		__field(        u32,	rd_free_clone		)
+		__field(	u32,	rd_reserved		)
 	),
 
 	TP_fast_assign(
@@ -480,16 +493,58 @@ TRACE_EVENT(gfs2_block_alloc,
 		__entry->block_state	= block_state;
 		__entry->rd_addr	= rgd->rd_addr;
 		__entry->rd_free_clone	= rgd->rd_free_clone;
+		__entry->rd_reserved	= rgd->rd_reserved;
 	),
 
-	TP_printk("%u,%u bmap %llu alloc %llu/%lu %s rg:%llu rf:%u",
+	TP_printk("%u,%u bmap %llu alloc %llu/%lu %s rg:%llu rf:%u rr:%lu",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
 		  (unsigned long long)__entry->inum,
 		  (unsigned long long)__entry->start,
 		  (unsigned long)__entry->len,
 		  block_state_name(__entry->block_state),
 		  (unsigned long long)__entry->rd_addr,
-		  __entry->rd_free_clone)
+		  __entry->rd_free_clone, (unsigned long)__entry->rd_reserved)
+);
+
+/* Keep track of multi-block reservations as they are allocated/freed */
+TRACE_EVENT(gfs2_rs,
+
+	TP_PROTO(const struct gfs2_inode *ip, const struct gfs2_blkreserv *rs,
+		 u8 func),
+
+	TP_ARGS(ip, rs, func),
+
+	TP_STRUCT__entry(
+		__field(        dev_t,  dev                     )
+		__field(	u64,	rd_addr			)
+		__field(	u32,	rd_free_clone		)
+		__field(	u32,	rd_reserved		)
+		__field(	u64,	inum			)
+		__field(	u64,	start			)
+		__field(	u32,	free			)
+		__field(	u8,	func			)
+	),
+
+	TP_fast_assign(
+		__entry->dev		= rs->rs_rgd ? rs->rs_rgd->rd_sbd->sd_vfs->s_dev : 0;
+		__entry->rd_addr	= rs->rs_rgd ? rs->rs_rgd->rd_addr : 0;
+		__entry->rd_free_clone	= rs->rs_rgd ? rs->rs_rgd->rd_free_clone : 0;
+		__entry->rd_reserved	= rs->rs_rgd ? rs->rs_rgd->rd_reserved : 0;
+		__entry->inum		= ip ? ip->i_no_addr : 0;
+		__entry->start		= gfs2_rs_startblk(rs);
+		__entry->free		= rs->rs_free;
+		__entry->func		= func;
+	),
+
+	TP_printk("%u,%u bmap %llu resrv %llu rg:%llu rf:%lu rr:%lu %s "
+		  "f:%lu",
+		  MAJOR(__entry->dev), MINOR(__entry->dev),
+		  (unsigned long long)__entry->inum,
+		  (unsigned long long)__entry->start,
+		  (unsigned long long)__entry->rd_addr,
+		  (unsigned long)__entry->rd_free_clone,
+		  (unsigned long)__entry->rd_reserved,
+		  rs_func_name(__entry->func), (unsigned long)__entry->free)
 );
 
 #endif /* _TRACE_GFS2_H */

fs/gfs2/xattr.c

@@ -327,6 +327,10 @@ static int ea_remove_unstuffed(struct gfs2_inode *ip, struct buffer_head *bh,
 {
 	int error;
 
+	error = gfs2_rindex_update(GFS2_SB(&ip->i_inode));
+	if (error)
+		return error;
+
 	error = gfs2_quota_hold(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 	if (error)
 		goto out_alloc;
@@ -710,6 +714,10 @@ static int ea_alloc_skeleton(struct gfs2_inode *ip, struct gfs2_ea_request *er,
 	struct buffer_head *dibh;
 	int error;
 
+	error = gfs2_rindex_update(GFS2_SB(&ip->i_inode));
+	if (error)
+		return error;
+
 	error = gfs2_quota_lock_check(ip);
 	if (error)
 		return error;
@@ -1483,6 +1491,10 @@ int gfs2_ea_dealloc(struct gfs2_inode *ip)
 {
 	int error;
 
+	error = gfs2_rindex_update(GFS2_SB(&ip->i_inode));
+	if (error)
+		return error;
+
 	error = gfs2_quota_hold(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 	if (error)
 		return error;