Skip to contentSkip to conversation

udf: Don't corrupt unalloc spacetable when writing it

PATCH In review Steven J. Magnani 2 participants 4 messages Jul 7, 2015 – Jul 10, 2015 UDF filesystem

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 udf: Don't corrupt unalloc spacetable when writing it

For a UDF filesystem configured with an Unallocated Space Table, a filesystem operation that triggers an update to the table results in on-disk corruption that prevents remounting:

  udf_read_tagged: tag version 0x0000 != 0x0002 || 0x0003, block 274

For example:
1. Create a filesystem

      $ mkudffs --media-type=hd --blocksize=512 --lvid=BUGTEST \
              --vid=BUGTEST --fsid=BUGTEST --space=unalloctable \
              /dev/mmcblk0

2. Mount it

      # mount /dev/mmcblk0 /mnt

3. Create a file

      $ echo "No corruption, please" > /mnt/new.file

4. Umount

      # umount /mnt

5. Attempt remount

      # mount /dev/mmcblk0 /mnt

This appears to be a longstanding bug caused by zero-initialization of the Unallocated Space Entry block buffer and only partial repopulation of required fields before writing to disk.

Commit 0adfb339fd64 ("udf: Fix unalloc space handling in udf_update_inode") addressed one such field, but several others are required.

---

fs/udf/inode.c +7 -12
diff -uprN a/fs/udf/inode.c b/fs/udf/inode.c
@@ -1652,17 +1652,9 @@ static int udf_update_inode(struct inode
 		       iinfo->i_ext.i_data, inode->i_sb->s_blocksize -
 					sizeof(struct unallocSpaceEntry));
 		use->descTag.tagIdent = cpu_to_le16(TAG_IDENT_USE);
-		use->descTag.tagLocation =
-				cpu_to_le32(iinfo->i_location.logicalBlockNum);
-		crclen = sizeof(struct unallocSpaceEntry) +
-				iinfo->i_lenAlloc - sizeof(struct tag);
-		use->descTag.descCRCLength = cpu_to_le16(crclen);
-		use->descTag.descCRC = cpu_to_le16(crc_itu_t(0, (char *)use +
-							   sizeof(struct tag),
-							   crclen));
-		use->descTag.tagChecksum = udf_tag_checksum(&use->descTag);
+		crclen = sizeof(struct unallocSpaceEntry);
 
-		goto out;
+		goto finish;
 	}
 
 	if (UDF_QUERY_FLAG(inode->i_sb, UDF_FLAG_UID_FORGET))
@@ -1782,6 +1774,8 @@ static int udf_update_inode(struct inode
 		efe->descTag.tagIdent = cpu_to_le16(TAG_IDENT_EFE);
 		crclen = sizeof(struct extendedFileEntry);
 	}
+
+finish:
 	if (iinfo->i_strat4096) {
 		fe->icbTag.strategyType = cpu_to_le16(4096);
 		fe->icbTag.strategyParameter = cpu_to_le16(1);
@@ -1791,7 +1785,9 @@ static int udf_update_inode(struct inode
 		fe->icbTag.numEntries = cpu_to_le16(1);
 	}
 
-	if (S_ISDIR(inode->i_mode))
+	if (iinfo->i_use)
+		fe->icbTag.fileType = ICBTAG_FILE_TYPE_USE;
+	else if (S_ISDIR(inode->i_mode))
 		fe->icbTag.fileType = ICBTAG_FILE_TYPE_DIRECTORY;
 	else if (S_ISREG(inode->i_mode))
 		fe->icbTag.fileType = ICBTAG_FILE_TYPE_REGULAR;
@@ -1828,7 +1824,6 @@ static int udf_update_inode(struct inode
 						  crclen));
 	fe->descTag.tagChecksum = udf_tag_checksum(&fe->descTag);
 
-out:
 	set_buffer_uptodate(bh);
 	unlock_buffer(bh);
 
#1 lore raw
Quoted 24 lines · Steven J. Magnani

For a UDF filesystem configured with an Unallocated Space Table, a filesystem operation that triggers an update to the table results in on-disk corruption that prevents remounting:

  udf_read_tagged: tag version 0x0000 != 0x0002 || 0x0003, block 274

For example:
1. Create a filesystem

      $ mkudffs --media-type=hd --blocksize=512 --lvid=BUGTEST \
              --vid=BUGTEST --fsid=BUGTEST --space=unalloctable \
              /dev/mmcblk0

2. Mount it

      # mount /dev/mmcblk0 /mnt

3. Create a file

      $ echo "No corruption, please" > /mnt/new.file

4. Umount

      # umount /mnt

5. Attempt remount

      # mount /dev/mmcblk0 /mnt
Quoted 8 lines

This appears to be a longstanding bug caused by zero-initialization of the Unallocated Space Entry block buffer and only partial repopulation of required fields before writing to disk.

Commit 0adfb339fd64 ("udf: Fix unalloc space handling in udf_update_inode") addressed one such field, but several others are required.

So the patch looks good to me. But what kind of mkudffs are you using? Because when I use the command you wrote into the changelog, I cannot mount the filesystem. The kernel complains about:

UDF-fs: error (device ubdb): udf_read_inode: (ino 274) failed ident=256
mount: /dev/ubdb: can't read superblock

Now ident 256 is TAG_IDENT_FSD (file set descriptor). So apparently my mkudffs doesn't create proper partition table with your parameters...

Honza

Quoted 52 lines

---
diff -uprN a/fs/udf/inode.c b/fs/udf/inode.c

fs/udf/inode.c +7 -12
@@ -1652,17 +1652,9 @@ static int udf_update_inode(struct inode
 		       iinfo->i_ext.i_data, inode->i_sb->s_blocksize -
 					sizeof(struct unallocSpaceEntry));
 		use->descTag.tagIdent = cpu_to_le16(TAG_IDENT_USE);
-		use->descTag.tagLocation =
-				cpu_to_le32(iinfo->i_location.logicalBlockNum);
-		crclen = sizeof(struct unallocSpaceEntry) +
-				iinfo->i_lenAlloc - sizeof(struct tag);
-		use->descTag.descCRCLength = cpu_to_le16(crclen);
-		use->descTag.descCRC = cpu_to_le16(crc_itu_t(0, (char *)use +
-							   sizeof(struct tag),
-							   crclen));
-		use->descTag.tagChecksum = udf_tag_checksum(&use->descTag);
+		crclen = sizeof(struct unallocSpaceEntry);
 
-		goto out;
+		goto finish;
 	}
 
 	if (UDF_QUERY_FLAG(inode->i_sb, UDF_FLAG_UID_FORGET))
@@ -1782,6 +1774,8 @@ static int udf_update_inode(struct inode
 		efe->descTag.tagIdent = cpu_to_le16(TAG_IDENT_EFE);
 		crclen = sizeof(struct extendedFileEntry);
 	}
+
+finish:
 	if (iinfo->i_strat4096) {
 		fe->icbTag.strategyType = cpu_to_le16(4096);
 		fe->icbTag.strategyParameter = cpu_to_le16(1);
@@ -1791,7 +1785,9 @@ static int udf_update_inode(struct inode
 		fe->icbTag.numEntries = cpu_to_le16(1);
 	}
 
-	if (S_ISDIR(inode->i_mode))
+	if (iinfo->i_use)
+		fe->icbTag.fileType = ICBTAG_FILE_TYPE_USE;
+	else if (S_ISDIR(inode->i_mode))
 		fe->icbTag.fileType = ICBTAG_FILE_TYPE_DIRECTORY;
 	else if (S_ISREG(inode->i_mode))
 		fe->icbTag.fileType = ICBTAG_FILE_TYPE_REGULAR;
@@ -1828,7 +1824,6 @@ static int udf_update_inode(struct inode
 						  crclen));
 	fe->descTag.tagChecksum = udf_tag_checksum(&fe->descTag);
 
-out:
 	set_buffer_uptodate(bh);
 	unlock_buffer(bh);
 
Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#2 lore raw
Quoted 21 lines · Jan Kara
Quoted 8 lines · Steven J. Magnani

For a UDF filesystem configured with an Unallocated Space Table, a filesystem operation that triggers an update to the table results in on-disk corruption that prevents remounting:

   udf_read_tagged: tag version 0x0000 != 0x0002 || 0x0003, block 274

For example:
1. Create a filesystem

Quoted 8 lines
       $ mkudffs --media-type=hd --blocksize=512 --lvid=BUGTEST \
               --vid=BUGTEST --fsid=BUGTEST --space=unalloctable \
               /dev/mmcblk0

2. Mount it

       # mount /dev/mmcblk0 /mnt

...

So the patch looks good to me. But what kind of mkudffs are you using? Because when I use the command you wrote into the changelog, I cannot mount the filesystem. The kernel complains about:

Quoted 7 lines

UDF-fs: error (device ubdb): udf_read_inode: (ino 274) failed ident=256
mount: /dev/ubdb: can't read superblock

Now ident 256 is TAG_IDENT_FSD (file set descriptor). So apparently my mkudffs doesn't create proper partition table with your parameters...

It identifies itself as:

    mkudffs 1.0.0b2 for UDF FS 1.0.0-cvs, 2002/02/09

...from Ubuntu udftools 1.0.0b3-14.2.

------------------------------------------------------------------------
Steven J. Magnani "I claim this network for MARS!
www.digidescorp.com Earthling, return my space modulator!"

#include <standard.disclaimer>

#3 lore raw
Quoted 27 lines · Steve Magnani
Quoted 24 lines · Jan Kara
Quoted 8 lines · Steven J. Magnani

For a UDF filesystem configured with an Unallocated Space Table, a filesystem operation that triggers an update to the table results in on-disk corruption that prevents remounting:

  udf_read_tagged: tag version 0x0000 != 0x0002 || 0x0003, block 274

For example:
1. Create a filesystem

Quoted 8 lines
      $ mkudffs --media-type=hd --blocksize=512 --lvid=BUGTEST \
              --vid=BUGTEST --fsid=BUGTEST --space=unalloctable \
              /dev/mmcblk0

2. Mount it

      # mount /dev/mmcblk0 /mnt

...

So the patch looks good to me. But what kind of mkudffs are you using? Because when I use the command you wrote into the changelog, I cannot mount the filesystem. The kernel complains about:

UDF-fs: error (device ubdb): udf_read_inode: (ino 274) failed ident=256
mount: /dev/ubdb: can't read superblock

Quoted 8 lines
Quoted 4 lines

Now ident 256 is TAG_IDENT_FSD (file set descriptor). So apparently my mkudffs doesn't create proper partition table with your parameters...

It identifies itself as:

    mkudffs 1.0.0b2 for UDF FS 1.0.0-cvs, 2002/02/09

...from Ubuntu udftools 1.0.0b3-14.2.

OK, I've fixed my mkudffs and verified that your patch indeed fixes the problem. Thanks. I have merged the patch to my tree and will push it to Linus.

Honza

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