Fix AFFS race condition.

AFFS code preallocates several blocks as an optimisation. Unfortunately it's not protected by lock so the same blocks may end up allocated twice. Here is a fix.

Signed-off-by Vladimir Serbinenko <[email protected]>

fs/affs/affs.h +2 -0

index 45a0ce4..fc1d4ca 100644
@@ -66,6 +66,8 @@ struct affs_inode_info {
 	u32	 i_protect;			/* unused attribute bits */
 	u32	 i_lastalloc;			/* last allocated block */
 	int	 i_pa_cnt;			/* number of preallocated blocks */
+	spinlock_t i_alloc;		        /* Protects last 2 fields. */
+
 	struct inode vfs_inode;
 };
 

fs/affs/bitmap.c +23 -11

index 3e26271..3341bde 100644
@@ -151,12 +151,18 @@ affs_alloc_block(struct inode *inode, u32 goal)
 
 	pr_debug("AFFS: balloc(inode=%lu,goal=%u): ", inode->i_ino, goal);
 
+	spin_lock(&AFFS_I(inode)->i_alloc);
+
 	if (AFFS_I(inode)->i_pa_cnt) {
-		pr_debug("%d\n", AFFS_I(inode)->i_lastalloc+1);
+		u32 ret;
 		AFFS_I(inode)->i_pa_cnt--;
-		return ++AFFS_I(inode)->i_lastalloc;
+		ret = ++AFFS_I(inode)->i_lastalloc;
+		spin_unlock(&AFFS_I(inode)->i_alloc);
+		return ret;
 	}
 
+	spin_unlock(&AFFS_I(inode)->i_alloc);
+
 	if (!goal || goal > sbi->s_partition_size) {
 		if (goal)
 			affs_warning(sb, "affs_balloc", "invalid goal %d", goal);

@@ -230,16 +236,22 @@ find_bit:
 	bit = ffs(tmp & mask) - 1;
 	blk += bit + sbi->s_reserved;
 	mask2 = mask = 1 << (bit & 31);
-	AFFS_I(inode)->i_lastalloc = blk;
-
-	/* prealloc as much as possible within this word */
-	while ((mask2 <<= 1)) {
-		if (!(tmp & mask2))
-			break;
-		AFFS_I(inode)->i_pa_cnt++;
-		mask |= mask2;
+
+	spin_lock(&AFFS_I(inode)->i_alloc);
+	if (!AFFS_I(inode)->i_pa_cnt) {
+		AFFS_I(inode)->i_lastalloc = blk;
+
+		/* prealloc as much as possible within this word */
+		while ((mask2 <<= 1)) {
+			if (!(tmp & mask2))
+				break;
+			AFFS_I(inode)->i_pa_cnt++;
+			mask |= mask2;
+		}
+		bm->bm_free -= AFFS_I(inode)->i_pa_cnt + 1;
 	}
-	bm->bm_free -= AFFS_I(inode)->i_pa_cnt + 1;
+
+	spin_unlock(&AFFS_I(inode)->i_alloc);
 
 	*data = cpu_to_be32(tmp & ~mask);
 

fs/affs/file.c +12 -3

index 2f4c935..829e976 100644
@@ -795,12 +795,21 @@ void
 affs_free_prealloc(struct inode *inode)
 {
 	struct super_block *sb = inode->i_sb;
+	u32 first, cnt;
 
 	pr_debug("AFFS: free_prealloc(ino=%lu)\n", inode->i_ino);
 
-	while (AFFS_I(inode)->i_pa_cnt) {
-		AFFS_I(inode)->i_pa_cnt--;
-		affs_free_block(sb, ++AFFS_I(inode)->i_lastalloc);
+	spin_lock(&AFFS_I(inode)->i_alloc);
+	first = AFFS_I(inode)->i_lastalloc;
+	cnt = AFFS_I(inode)->i_pa_cnt;
+	AFFS_I(inode)->i_lastalloc += cnt;
+	AFFS_I(inode)->i_pa_cnt = 0;
+
+	spin_unlock(&AFFS_I(inode)->i_alloc);
+
+	while (cnt) {
+		cnt--;
+		affs_free_block(sb, ++first);
 	}
 }
 

fs/affs/super.c +1 -0

index 0782653..1df3c95 100644
@@ -91,6 +91,7 @@ static struct inode *affs_alloc_inode(struct super_block *sb)
 	i->i_lc = NULL;
 	i->i_ext_bh = NULL;
 	i->i_pa_cnt = 0;
+	spin_lock_init(&i->i_alloc);
 
 	return &i->vfs_inode;
 }

Quoted 4 lines

AFFS code preallocates several blocks as an optimisation. Unfortunately it's not protected by lock so the same blocks may end up allocated twice. Here is a fix.

Quoted 1 line

Signed-off-by Vladimir Serbinenko <[email protected]>

 The patch looks good to me now. Thanks! You can add:

Reviewed-by Jan Kara <[email protected]>

 Al, will you merge this patch through your tree? AFFS does not seem to have a maintainer so you are a default fallback...

                                                               Honza

2012/5/14 Jan Kara <[email protected]>:

Quoted 12 lines · Vladimir 'φ-coder/phcoder' Serbinenko

Quoted 5 lines

AFFS code preallocates several blocks as an optimisation. Unfortunately it's not protected by lock so the same blocks may end up allocated twice. Here is a fix.

Signed-off-by Vladimir Serbinenko <[email protected]>

 The patch looks good to me now. Thanks! You can add:

Reviewed-by Jan Kara <[email protected]>

 Al, will you merge this patch through your tree? AFFS does not seem to have a maintainer so you are a default fallback...

Quoted 2 lines

                                                               Honza

I don't know the AFFS code, so only a question. Instead to use a spin lock, I think we can use a simple mutex. Or is the spin lock mandatory?

2012/5/14 Jan Kara <[email protected]>:

Quoted 12 lines · Vladimir 'φ-coder/phcoder' Serbinenko

Quoted 5 lines

AFFS code preallocates several blocks as an optimisation. Unfortunately it's not protected by lock so the same blocks may end up allocated twice. Here is a fix.

Signed-off-by Vladimir Serbinenko <[email protected]>

The patch looks good to me now. Thanks! You can add:

Reviewed-by Jan Kara <[email protected]>

Al, will you merge this patch through your tree? AFFS does not seem to
have a maintainer so you are a default fallback...

Quoted 2 lines

Honza

I don't know the AFFS code, so only a question. Instead to use a spin lock, I think we can use a simple mutex. Or is the spin lock mandatory?

Quoted 16 lines

2012/5/14 Jan Kara <[email protected]>:

Quoted 14 lines · Vladimir 'φ-coder/phcoder' Serbinenko

Quoted 5 lines

AFFS code preallocates several blocks as an optimisation. Unfortunately it's not protected by lock so the same blocks may end up allocated twice. Here is a fix.

Signed-off-by Vladimir Serbinenko <[email protected]>

 The patch looks good to me now. Thanks! You can add:

Reviewed-by Jan Kara <[email protected]>

 Al, will you merge this patch through your tree? AFFS does not seem to have a maintainer so you are a default fallback...

                                                               Honza

Quoted 3 lines

I don't know the AFFS code, so only a question. Instead to use a spin lock, I think we can use a simple mutex. Or is the spin lock mandatory?

 So what would be an advantage of a mutex? Spinlock *is* the simple locking variant...

                                                               Honza --