[PATCH] isofs: fix KMSAN uninit-value bug in do_isofs_readdir()

[PATCH] isofs: fix KMSAN uninit-value bug in do_isofs_readdir()

[Qasim Ijaz]

In do_isofs_readdir() when assigning the variable 
"struct iso_directory_record *de" the b_data field of the buffer_head 
is accessed and an offset is added to it, the size of b_data is 2048 
and the offset size is 2047, meaning 
"de = (struct iso_directory_record *) (bh->b_data + offset);" 
yields the final byte of the 2048 sized b_data block.

The first byte of the directory record (de_len) is then read and 
found to be 31, meaning the directory record size is 31 bytes long. 
The directory record is defined by the structure:

	struct iso_directory_record {
		__u8 length;                     // 1 byte 
		__u8 ext_attr_length;            // 1 byte 
		__u8 extent[8];                  // 8 bytes 
		__u8 size[8];                    // 8 bytes 
		__u8 date[7];                    // 7 bytes 
		__u8 flags;                      // 1 byte 
		__u8 file_unit_size;             // 1 byte 
		__u8 interleave;                 // 1 byte 
		__u8 volume_sequence_number[4];  // 4 bytes
		__u8 name_len;                   // 1 byte
		char name[];                     // variable size
	} __attribute__((packed));

The fixed portion of this structure occupies 33 bytes. Therefore, a 
valid directory record must be at least 33 bytes long 
(even without considering the variable-length name field). 
Since de_len is only 31, it is insufficient to contain
the complete fixed header. 

The code later hits the following sanity check that 
compares de_len against the sum of de->name_len and 
sizeof(struct iso_directory_record):

	if (de_len < de->name_len[0] + sizeof(struct iso_directory_record)) {
		...
	}

Since the fixed portion of the structure is 
33 bytes (up to and including name_len member), 
a valid record should have de_len of at least 33 bytes; 
here, however, de_len is too short, and the field de->name_len 
(located at offset 32) is accessed even though it lies beyond 
the available 31 bytes. 

This access on the corrupted isofs data triggers a KASAN uninitialized 
memory warning. The fix would be to first verify that de_len is at least 
sizeof(struct iso_directory_record) before accessing any 
fields like de->name_len.

Reported-by: syzbot <syzbot+812641c6c3d7586a1613@syzkaller.appspotmail.com>
Tested-by: syzbot <syzbot+812641c6c3d7586a1613@syzkaller.appspotmail.com>
Closes: https://syzkaller.appspot.com/bug?extid=812641c6c3d7586a1613
Fixes: 2deb1acc653c ("isofs: fix access to unallocated memory when reading corrupted filesystem")
Signed-off-by: Qasim Ijaz <qasdev00@gmail.com>
---
 fs/isofs/dir.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/isofs/dir.c b/fs/isofs/dir.c
index eb2f8273e6f1..366ac8b95330 100644
--- a/fs/isofs/dir.c
+++ b/fs/isofs/dir.c
@@ -147,7 +147,8 @@ static int do_isofs_readdir(struct inode *inode, struct file *file,
 			de = tmpde;
 		}
 		/* Basic sanity check, whether name doesn't exceed dir entry */
-		if (de_len < de->name_len[0] +
+		if (de_len < sizeof(struct iso_directory_record) ||
+		    de_len < de->name_len[0] +
 					sizeof(struct iso_directory_record)) {
 			printk(KERN_NOTICE "iso9660: Corrupted directory entry"
 			       " in block %lu of inode %lu\n", block,
-- 
2.39.5

Re: [PATCH] isofs: fix KMSAN uninit-value bug in do_isofs_readdir()

[Jan Kara]

On Tue 11-02-25 19:59:00, Qasim Ijaz wrote:
> In do_isofs_readdir() when assigning the variable 
> "struct iso_directory_record *de" the b_data field of the buffer_head 
> is accessed and an offset is added to it, the size of b_data is 2048 
> and the offset size is 2047, meaning 
> "de = (struct iso_directory_record *) (bh->b_data + offset);" 
> yields the final byte of the 2048 sized b_data block.
> 
> The first byte of the directory record (de_len) is then read and 
> found to be 31, meaning the directory record size is 31 bytes long. 
> The directory record is defined by the structure:
> 
> 	struct iso_directory_record {
> 		__u8 length;                     // 1 byte 
> 		__u8 ext_attr_length;            // 1 byte 
> 		__u8 extent[8];                  // 8 bytes 
> 		__u8 size[8];                    // 8 bytes 
> 		__u8 date[7];                    // 7 bytes 
> 		__u8 flags;                      // 1 byte 
> 		__u8 file_unit_size;             // 1 byte 
> 		__u8 interleave;                 // 1 byte 
> 		__u8 volume_sequence_number[4];  // 4 bytes
> 		__u8 name_len;                   // 1 byte
> 		char name[];                     // variable size
> 	} __attribute__((packed));
> 
> The fixed portion of this structure occupies 33 bytes. Therefore, a 
> valid directory record must be at least 33 bytes long 
> (even without considering the variable-length name field). 
> Since de_len is only 31, it is insufficient to contain
> the complete fixed header. 
> 
> The code later hits the following sanity check that 
> compares de_len against the sum of de->name_len and 
> sizeof(struct iso_directory_record):
> 
> 	if (de_len < de->name_len[0] + sizeof(struct iso_directory_record)) {
> 		...
> 	}
> 
> Since the fixed portion of the structure is 
> 33 bytes (up to and including name_len member), 
> a valid record should have de_len of at least 33 bytes; 
> here, however, de_len is too short, and the field de->name_len 
> (located at offset 32) is accessed even though it lies beyond 
> the available 31 bytes. 
> 
> This access on the corrupted isofs data triggers a KASAN uninitialized 
> memory warning. The fix would be to first verify that de_len is at least 
> sizeof(struct iso_directory_record) before accessing any 
> fields like de->name_len.
> 
> Reported-by: syzbot <syzbot+812641c6c3d7586a1613@syzkaller.appspotmail.com>
> Tested-by: syzbot <syzbot+812641c6c3d7586a1613@syzkaller.appspotmail.com>
> Closes: https://syzkaller.appspot.com/bug?extid=812641c6c3d7586a1613
> Fixes: 2deb1acc653c ("isofs: fix access to unallocated memory when reading corrupted filesystem")
> Signed-off-by: Qasim Ijaz <qasdev00@gmail.com>

Thanks. Added to my tree.

								Honza

> ---
>  fs/isofs/dir.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/isofs/dir.c b/fs/isofs/dir.c
> index eb2f8273e6f1..366ac8b95330 100644
> --- a/fs/isofs/dir.c
> +++ b/fs/isofs/dir.c
> @@ -147,7 +147,8 @@ static int do_isofs_readdir(struct inode *inode, struct file *file,
>  			de = tmpde;
>  		}
>  		/* Basic sanity check, whether name doesn't exceed dir entry */
> -		if (de_len < de->name_len[0] +
> +		if (de_len < sizeof(struct iso_directory_record) ||
> +		    de_len < de->name_len[0] +
>  					sizeof(struct iso_directory_record)) {
>  			printk(KERN_NOTICE "iso9660: Corrupted directory entry"
>  			       " in block %lu of inode %lu\n", block,
> -- 
> 2.39.5
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR