Skip to contentSkip to conversation

fs/minix: bugfix, number of indirect block ptrs per block depends on block size

PATCH In review Erik van der Kouwe 3 participants 6 messages May 26, 2010 – May 27, 2010 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs/minix: bugfix, number of indirect block ptrs per block depends on block size

From: Erik van der Kouwe <[email protected]>

The MINIX filesystem driver used a constant number of indirect block pointers in an indirect block. This worked only for filesystems with 1kb block, while the MINIX default block size is now 4kb. As a consequence, large files were read incorrectly on such filesystems and writing a large file would cause the filesystem to become corrupted. This patch computes the number of indirect block pointers based on the block size, making the driver work for each block size.

I would like to thank Feiran Zheng ('Fam') for pointing out the cause of the corruption.

---

minix/itree_v2.c renamed from minix/itree_v2.c.orig +0 -0
@@ -20,6 +20,9 @@ static inline block_t *i_data(struct ino
	return (block_t *)minix_i(inode)->u.i2_data;
}

+#define DIRCOUNT 7
+#define INDIRCOUNT(sb) ((sb)->s_blocksize / 4)
+
static int block_to_path(struct inode * inode, long block, int offsets[DEPTH])

{
	int n = 0;

@@ -34,21 +37,21 @@ static int block_to_path(struct inode *
printk("MINIX-fs: block_to_path: "
"block %ld too big on dev %s\n",
block, bdevname(sb->s_bdev, b));
- } else if (block < 7) {

+	} else if (block < DIRCOUNT) {
		offsets[n++] = block;
-	} else if ((block -= 7) < 256) {
-		offsets[n++] = 7;
+	} else if ((block -= DIRCOUNT) < INDIRCOUNT(sb)) {
+		offsets[n++] = DIRCOUNT;
		offsets[n++] = block;

- } else if ((block -= 256) < 256*256) {
- offsets[n++] = 8;
- offsets[n++] = block>>8;
- offsets[n++] = block & 255;
+ } else if ((block -= INDIRCOUNT(sb)) < INDIRCOUNT(sb) * INDIRCOUNT(sb)) {
+ offsets[n++] = DIRCOUNT + 1;
+ offsets[n++] = block / INDIRCOUNT(sb);

+		offsets[n++] = block % INDIRCOUNT(sb);
	} else {

- block -= 256*256;
- offsets[n++] = 9;
- offsets[n++] = block>>16;
- offsets[n++] = (block>>8) & 255;
- offsets[n++] = block & 255;
+ block -= INDIRCOUNT(sb) * INDIRCOUNT(sb);
+ offsets[n++] = DIRCOUNT + 2;
+ offsets[n++] = (block / INDIRCOUNT(sb)) / INDIRCOUNT(sb);
+ offsets[n++] = (block / INDIRCOUNT(sb)) % INDIRCOUNT(sb);

+		offsets[n++] = block % INDIRCOUNT(sb);
	}
	return n;
}
maintainer
#1 lore raw
Quoted 1 line · Erik van der Kouwe

+#define INDIRCOUNT(sb) ((sb)->s_blocksize / 4)

ITYM (1 << ((sb)->s_blocksize_bits - 2))

Give gcc at least a chance to turn the uses of that into shifts... Applied with such change.

reviewer
#2 lore raw
In reply to Erik van der Kouwe
Quoted 9 lines · Erik van der Kouwe

From: Erik van der Kouwe <[email protected]>

The MINIX filesystem driver used a constant number of indirect block pointers in an indirect block. This worked only for filesystems with 1kb block, while the MINIX default block size is now 4kb. As a consequence, large files were read incorrectly on such filesystems and writing a large file would cause the filesystem to become corrupted. This patch computes the number of indirect block pointers based on the block size, making the driver work for each block size.

Out of curiosity: Are you really using MINIX? The code is dead for a
long time...

Quoted 19 lines

I would like to thank Feiran Zheng ('Fam') for pointing out the cause of the corruption.

---

minix/itree_v2.c renamed from minix/itree_v2.c.orig +3 -0
@@ -20,6 +20,9 @@ static inline block_t *i_data(struct ino
 	return (block_t *)minix_i(inode)->u.i2_data;
 }
 
+#define DIRCOUNT 7
+#define INDIRCOUNT(sb) ((sb)->s_blocksize / 4)
+
 static int block_to_path(struct inode * inode, long block, int offsets[DEPTH])
 {
 	int n = 0;

minix/itree_v2.c static int block_to_path(struct inode * · line 42

Quoted 9 lines
minix/itree_v2.c renamed from minix/itree_v2.c.orig +2 -3
@@ -34,21 +37,21 @@ static int block_to_path(struct inode *
 			printk("MINIX-fs: block_to_path: "
 			       "block %ld too big on dev %s\n",
 				block, bdevname(sb->s_bdev, b));
-	} else if (block < 7) {
+	} else if (block < DIRCOUNT) {
 		offsets[n++] = block;
-	} else if ((block -= 7) < 256) {
-		offsets[n++] = 7;
+	} else if ((block -= DIRCOUNT) < INDIRCOUNT(sb)) {

This modification of 'block' in the if condition is a blatant violation
of the kernel coding style. If you feel like it, you could fix it when changing the conditions anyway. Something like:

  if (block < DIRCOUNT) {
	offsets[n++] = block;
	return n;

}
block -= DIRCOUNT;
if (block < INDIRCOUNT(sb)) {
...

minix/itree_v2.c static int block_to_path(struct inode * · line 48

Quoted 10 lines
minix/itree_v2.c renamed from minix/itree_v2.c.orig +5 -4
+		offsets[n++] = DIRCOUNT;
 		offsets[n++] = block;
-	} else if ((block -= 256) < 256*256) {
-		offsets[n++] = 8;
-		offsets[n++] = block>>8;
-		offsets[n++] = block & 255;
+	} else if ((block -= INDIRCOUNT(sb)) < INDIRCOUNT(sb) * INDIRCOUNT(sb)) {
+		offsets[n++] = DIRCOUNT + 1;
+		offsets[n++] = block / INDIRCOUNT(sb);
+		offsets[n++] = block % INDIRCOUNT(sb);

This division and modulo aren't really necessary. You can define
INDIRCOUNT_BITS(sb) as (sb->s_blocksize_bits - 2) and use it instead of the division and use & (INDIRCOUNT(sb)-1) instead of the modulo. Not that it would seriously matter but it's an improvement.

Quoted 4 lines
minix/itree_v2.c renamed from minix/itree_v2.c.orig +0 -3
 	} else {
-		block -= 256*256;
-		offsets[n++] = 9;
-		offsets[n++] = block>>16;

minix/itree_v2.c static int block_to_path(struct inode * · line 57

Quoted 10 lines
minix/itree_v2.c renamed from minix/itree_v2.c.orig +5 -2
-		offsets[n++] = (block>>8) & 255;
-		offsets[n++] = block & 255;
+		block -= INDIRCOUNT(sb) * INDIRCOUNT(sb);
+		offsets[n++] = DIRCOUNT + 2;
+		offsets[n++] = (block / INDIRCOUNT(sb)) / INDIRCOUNT(sb);
+		offsets[n++] = (block / INDIRCOUNT(sb)) % INDIRCOUNT(sb);
+		offsets[n++] = block % INDIRCOUNT(sb);
 	}
 	return n;
 }

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
maintainer
#3 lore raw
Quoted 5 lines · Jan Kara

Review

Quoted 3 lines
(quoted diff) +1 -2
-	} else if ((block -= 7) < 256) {
-		offsets[n++] = 7;
+	} else if ((block -= DIRCOUNT) < INDIRCOUNT(sb)) {

This modification of 'block' in the if condition is a blatant violation
of the kernel coding style.

The hell it is. [...said the guy who'd written it that way in the first place] It _is_ idiomatic C for that kind of thing ("we've got several adjacent ranges, find which one had we hit and how far from its start").

#4 lore raw
In reply to Jan Kara

Dear all,

Quoted 2 lines

Out of curiosity: Are you really using MINIX? The code is dead for a
long time...

MINIX is very much alive, see http://www.minix3.org/.

I work in the team developing MINIX. Being able to exchange data between Linux and MINIX is very useful for us and there is no other filesystem that the two share. We have been aware of (and annoyed by) this problem for a long time, but didn't know the cause (and didn't have much time to check). Feiran is a GSoC student working on our project and he found the issue when setting up his development environment so then we decided to bring out a patch.

With kind regards,
Erik

reviewer
#5 lore raw
Quoted 7 lines · Erik van der Kouwe

Dear all,

Quoted 2 lines

Out of curiosity: Are you really using MINIX? The code is dead for a
long time...

MINIX is very much alive, see http://www.minix3.org/.

Quoted 7 lines

I work in the team developing MINIX. Being able to exchange data between Linux and MINIX is very useful for us and there is no other filesystem that the two share. We have been aware of (and annoyed by) this problem for a long time, but didn't know the cause (and didn't have much time to check). Feiran is a GSoC student working on our project and he found the issue when setting up his development environment so then we decided to bring out a patch.

OK, I was more speaking about the Linux MINIX filesystem driver but it's
good to know someone is still using it :) Thanks for education.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR