ufs: don't use lock_ufs() for block pointers tree protection

* stores to block pointers are under per-inode seqlock (meta_lock) and
mutex (truncate_mutex)
* fetches of block pointers are either under truncate_mutex, or wrapped
into seqretry loop on meta_lock
* all changes of ->i_size are under truncate_mutex and i_mutex
* all changes of ->i_lastfrag are under truncate_mutex

It's similar to what ext2 is doing; the main difference is that unlike
ext2 we can't rely upon the atomicity of stores into block pointers -
on UFS2 they are 64bit.  So we can't cut the corner when switching
a pointer from NULL to non-NULL as we could in ext2_splice_branch()
and need to use meta_lock on all modifications.

We use seqlock where ext2 uses rwlock; ext2 could probably also benefit
from such change...

Another non-trivial difference is that with UFS we *cannot* have reader
grab truncate_mutex in case of race - it has to keep retrying.  That
might be possible to change, but not until we lift tail unpacking
several levels up in call chain.

After that commit we do *NOT* hold fs-wide serialization on accesses
to block pointers anymore.  Moreover, lock_ufs() can become a normal
mutex now - it's only used on statfs, remount and sync_fs and none
of those uses are recursive.  As the matter of fact, *now* it can be
collapsed with ->s_lock, and be eventually replaced with saner
per-cylinder-group spinlocks, but that's a separate story.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

fs/ufs/balloc.c

@@ -417,7 +417,9 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 	if (oldcount == 0) {
 		result = ufs_alloc_fragments (inode, cgno, goal, count, err);
 		if (result) {
+			write_seqlock(&UFS_I(inode)->meta_lock);
 			ufs_cpu_to_data_ptr(sb, p, result);
+			write_sequnlock(&UFS_I(inode)->meta_lock);
 			*err = 0;
 			UFS_I(inode)->i_lastfrag =
 				max(UFS_I(inode)->i_lastfrag, fragment + count);
@@ -473,7 +475,9 @@ u64 ufs_new_fragments(struct inode *inode, void *p, u64 fragment,
 		ufs_change_blocknr(inode, fragment - oldcount, oldcount,
 				   uspi->s_sbbase + tmp,
 				   uspi->s_sbbase + result, locked_page);
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_cpu_to_data_ptr(sb, p, result);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 		*err = 0;
 		UFS_I(inode)->i_lastfrag = max(UFS_I(inode)->i_lastfrag,
 						fragment + count);

fs/ufs/inode.c

@@ -41,8 +41,6 @@
 #include "swab.h"
 #include "util.h"
 
-static u64 ufs_frag_map(struct inode *inode, sector_t frag, bool needs_lock);
-
 static int ufs_block_to_path(struct inode *inode, sector_t i_block, sector_t offsets[4])
 {
 	struct ufs_sb_private_info *uspi = UFS_SB(inode->i_sb)->s_uspi;
@@ -75,12 +73,53 @@ static int ufs_block_to_path(struct inode *inode, sector_t i_block, sector_t off
 	return n;
 }
 
+typedef struct {
+	void	*p;
+	union {
+		__fs32	key32;
+		__fs64	key64;
+	};
+	struct buffer_head *bh;
+} Indirect;
+
+static inline int grow_chain32(struct ufs_inode_info *ufsi,
+			       struct buffer_head *bh, __fs32 *v,
+			       Indirect *from, Indirect *to)
+{
+	Indirect *p;
+	unsigned seq;
+	to->bh = bh;
+	do {
+		seq = read_seqbegin(&ufsi->meta_lock);
+		to->key32 = *(__fs32 *)(to->p = v);
+		for (p = from; p <= to && p->key32 == *(__fs32 *)p->p; p++)
+			;
+	} while (read_seqretry(&ufsi->meta_lock, seq));
+	return (p > to);
+}
+
+static inline int grow_chain64(struct ufs_inode_info *ufsi,
+			       struct buffer_head *bh, __fs64 *v,
+			       Indirect *from, Indirect *to)
+{
+	Indirect *p;
+	unsigned seq;
+	to->bh = bh;
+	do {
+		seq = read_seqbegin(&ufsi->meta_lock);
+		to->key64 = *(__fs64 *)(to->p = v);
+		for (p = from; p <= to && p->key64 == *(__fs64 *)p->p; p++)
+			;
+	} while (read_seqretry(&ufsi->meta_lock, seq));
+	return (p > to);
+}
+
 /*
  * Returns the location of the fragment from
  * the beginning of the filesystem.
  */
 
-static u64 ufs_frag_map(struct inode *inode, sector_t frag, bool needs_lock)
+static u64 ufs_frag_map(struct inode *inode, sector_t frag)
 {
 	struct ufs_inode_info *ufsi = UFS_I(inode);
 	struct super_block *sb = inode->i_sb;
@@ -88,12 +127,10 @@ static u64 ufs_frag_map(struct inode *inode, sector_t frag, bool needs_lock)
 	u64 mask = (u64) uspi->s_apbmask>>uspi->s_fpbshift;
 	int shift = uspi->s_apbshift-uspi->s_fpbshift;
 	sector_t offsets[4], *p;
+	Indirect chain[4], *q = chain;
 	int depth = ufs_block_to_path(inode, frag >> uspi->s_fpbshift, offsets);
-	u64  ret = 0L;
-	__fs32 block;
-	__fs64 u2_block = 0L;
 	unsigned flags = UFS_SB(sb)->s_flags;
-	u64 temp = 0L;
+	u64 res = 0;
 
 	UFSD(": frag = %llu  depth = %d\n", (unsigned long long)frag, depth);
 	UFSD(": uspi->s_fpbshift = %d ,uspi->s_apbmask = %x, mask=%llx\n",
@@ -101,59 +138,73 @@ static u64 ufs_frag_map(struct inode *inode, sector_t frag, bool needs_lock)
 		(unsigned long long)mask);
 
 	if (depth == 0)
-		return 0;
+		goto no_block;
 
+again:
 	p = offsets;
 
-	if (needs_lock)
-		lock_ufs(sb);
 	if ((flags & UFS_TYPE_MASK) == UFS_TYPE_UFS2)
 		goto ufs2;
 
-	block = ufsi->i_u1.i_data[*p++];
+	if (!grow_chain32(ufsi, NULL, &ufsi->i_u1.i_data[*p++], chain, q))
-	if (!block)
+		goto changed;
-		goto out;
+	if (!q->key32)
+		goto no_block;
 	while (--depth) {
+		__fs32 *ptr;
 		struct buffer_head *bh;
 		sector_t n = *p++;
 
-		bh = sb_bread(sb, uspi->s_sbbase + fs32_to_cpu(sb, block)+(n>>shift));
+		bh = sb_bread(sb, uspi->s_sbbase +
+				  fs32_to_cpu(sb, q->key32) + (n>>shift));
 		if (!bh)
-			goto out;
+			goto no_block;
-		block = ((__fs32 *) bh->b_data)[n & mask];
+		ptr = (__fs32 *)bh->b_data + (n & mask);
-		brelse (bh);
+		if (!grow_chain32(ufsi, bh, ptr, chain, ++q))
-		if (!block)
+			goto changed;
-			goto out;
+		if (!q->key32)
+			goto no_block;
 	}
-	ret = (u64) (uspi->s_sbbase + fs32_to_cpu(sb, block) + (frag & uspi->s_fpbmask));
+	res = fs32_to_cpu(sb, q->key32);
-	goto out;
+	goto found;
+
 ufs2:
-	u2_block = ufsi->i_u1.u2_i_data[*p++];
+	if (!grow_chain64(ufsi, NULL, &ufsi->i_u1.u2_i_data[*p++], chain, q))
-	if (!u2_block)
+		goto changed;
-		goto out;
+	if (!q->key64)
-
+		goto no_block;
 
 	while (--depth) {
+		__fs64 *ptr;
 		struct buffer_head *bh;
 		sector_t n = *p++;
 
-
+		bh = sb_bread(sb, uspi->s_sbbase +
-		temp = (u64)(uspi->s_sbbase) + fs64_to_cpu(sb, u2_block);
+				  fs64_to_cpu(sb, q->key64) + (n>>shift));
-		bh = sb_bread(sb, temp +(u64) (n>>shift));
 		if (!bh)
-			goto out;
+			goto no_block;
-		u2_block = ((__fs64 *)bh->b_data)[n & mask];
+		ptr = (__fs64 *)bh->b_data + (n & mask);
-		brelse(bh);
+		if (!grow_chain64(ufsi, bh, ptr, chain, ++q))
-		if (!u2_block)
+			goto changed;
-			goto out;
+		if (!q->key64)
+			goto no_block;
 	}
-	temp = (u64)uspi->s_sbbase + fs64_to_cpu(sb, u2_block);
+	res = fs64_to_cpu(sb, q->key64);
-	ret = temp + (u64) (frag & uspi->s_fpbmask);
+found:
+	res += uspi->s_sbbase + (frag & uspi->s_fpbmask);
+no_block:
+	while (q > chain) {
+		brelse(q->bh);
+		q--;
+	}
+	return res;
 
-out:
+changed:
-	if (needs_lock)
+	while (q > chain) {
-		unlock_ufs(sb);
+		brelse(q->bh);
-	return ret;
+		q--;
+	}
+	goto again;
 }
 
 /**
@@ -421,10 +472,9 @@ int ufs_getfrag_block(struct inode *inode, sector_t fragment, struct buffer_head
 	int ret, err, new;
 	unsigned long ptr,phys;
 	u64 phys64 = 0;
-	bool needs_lock = (sbi->mutex_owner != current);
 	
 	if (!create) {
-		phys64 = ufs_frag_map(inode, fragment, needs_lock);
+		phys64 = ufs_frag_map(inode, fragment);
 		UFSD("phys64 = %llu\n", (unsigned long long)phys64);
 		if (phys64)
 			map_bh(bh_result, sb, phys64);
@@ -438,8 +488,7 @@ int ufs_getfrag_block(struct inode *inode, sector_t fragment, struct buffer_head
 	ret = 0;
 	bh = NULL;
 
-	if (needs_lock)
+	mutex_lock(&UFS_I(inode)->truncate_mutex);
-		lock_ufs(sb);
 
 	UFSD("ENTER, ino %lu, fragment %llu\n", inode->i_ino, (unsigned long long)fragment);
 	if (fragment >
@@ -501,8 +550,7 @@ out:
 		set_buffer_new(bh_result);
 	map_bh(bh_result, sb, phys);
 abort:
-	if (needs_lock)
+	mutex_unlock(&UFS_I(inode)->truncate_mutex);
-		unlock_ufs(sb);
 
 	return err;
 

fs/ufs/super.c

@@ -1429,6 +1429,8 @@ static struct inode *ufs_alloc_inode(struct super_block *sb)
 		return NULL;
 
 	ei->vfs_inode.i_version = 1;
+	seqlock_init(&ei->meta_lock);
+	mutex_init(&ei->truncate_mutex);
 	return &ei->vfs_inode;
 }
 

fs/ufs/truncate.c

@@ -128,7 +128,9 @@ next1:
 		tmp = ufs_data_ptr_to_cpu(sb, p);
 		if (!tmp)
 			continue;
+		write_seqlock(&ufsi->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&ufsi->meta_lock);
 
 		if (free_count == 0) {
 			frag_to_free = tmp;
@@ -157,7 +159,9 @@ next1:
 	if (!tmp )
 		ufs_panic(sb, "ufs_truncate_direct", "internal error");
 	frag4 = ufs_fragnum (frag4);
+	write_seqlock(&ufsi->meta_lock);
 	ufs_data_ptr_clear(uspi, p);
+	write_sequnlock(&ufsi->meta_lock);
 
 	ufs_free_fragments (inode, tmp, frag4);
 	mark_inode_dirty(inode);
@@ -199,7 +203,9 @@ static int ufs_trunc_indirect(struct inode *inode, u64 offset, void *p)
 		return 1;
 	}
 	if (!ind_ubh) {
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 		return 0;
 	}
 
@@ -210,7 +216,9 @@ static int ufs_trunc_indirect(struct inode *inode, u64 offset, void *p)
 		if (!tmp)
 			continue;
 
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_data_ptr_clear(uspi, ind);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 		ubh_mark_buffer_dirty(ind_ubh);
 		if (free_count == 0) {
 			frag_to_free = tmp;
@@ -235,7 +243,9 @@ static int ufs_trunc_indirect(struct inode *inode, u64 offset, void *p)
 			break;
 	if (i >= uspi->s_apb) {
 		tmp = ufs_data_ptr_to_cpu(sb, p);
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 
 		ubh_bforget(ind_ubh);
 		ufs_free_blocks (inode, tmp, uspi->s_fpb);
@@ -278,7 +288,9 @@ static int ufs_trunc_dindirect(struct inode *inode, u64 offset, void *p)
 		return 1;
 	}
 	if (!dind_bh) {
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 		return 0;
 	}
 
@@ -297,7 +309,9 @@ static int ufs_trunc_dindirect(struct inode *inode, u64 offset, void *p)
 			break;
 	if (i >= uspi->s_apb) {
 		tmp = ufs_data_ptr_to_cpu(sb, p);
+		write_seqlock(&UFS_I(inode)->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&UFS_I(inode)->meta_lock);
 
 		ubh_bforget(dind_bh);
 		ufs_free_blocks(inode, tmp, uspi->s_fpb);
@@ -339,7 +353,9 @@ static int ufs_trunc_tindirect(struct inode *inode)
 		return 1;
 	}
 	if (!tind_bh) {
+		write_seqlock(&ufsi->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&ufsi->meta_lock);
 		return 0;
 	}
 
@@ -355,7 +371,9 @@ static int ufs_trunc_tindirect(struct inode *inode)
 			break;
 	if (i >= uspi->s_apb) {
 		tmp = ufs_data_ptr_to_cpu(sb, p);
+		write_seqlock(&ufsi->meta_lock);
 		ufs_data_ptr_clear(uspi, p);
+		write_sequnlock(&ufsi->meta_lock);
 
 		ubh_bforget(tind_bh);
 		ufs_free_blocks(inode, tmp, uspi->s_fpb);
@@ -447,7 +465,7 @@ static void __ufs_truncate_blocks(struct inode *inode)
 	struct ufs_sb_private_info *uspi = UFS_SB(sb)->s_uspi;
 	int retry;
 
-	lock_ufs(sb);
+	mutex_lock(&ufsi->truncate_mutex);
 	while (1) {
 		retry = ufs_trunc_direct(inode);
 		retry |= ufs_trunc_indirect(inode, UFS_IND_BLOCK,
@@ -465,7 +483,7 @@ static void __ufs_truncate_blocks(struct inode *inode)
 	}
 
 	ufsi->i_lastfrag = DIRECT_FRAGMENT;
-	unlock_ufs(sb);
+	mutex_unlock(&ufsi->truncate_mutex);
 }
 
 int ufs_truncate(struct inode *inode, loff_t size)

fs/ufs/ufs.h

@@ -46,6 +46,8 @@ struct ufs_inode_info {
 	__u32	i_oeftflag;
 	__u16	i_osync;
 	__u64	i_lastfrag;
+	seqlock_t meta_lock;
+	struct mutex	truncate_mutex;
 	__u32   i_dir_start_lookup;
 	struct inode vfs_inode;
 };