[PATCH 0/2] ext4: fix unable to handle kernel paging request in do_split()

[PATCH 0/2] ext4: fix unable to handle kernel paging request in do_split()

[libaokun]

From: Baokun Li <libaokun1@huawei.com>

"kvm-xfstests -c ext4/all -g auto" has been executed with no new failures.

Baokun Li (2):
  ext4: check dot and dotdot of dx_root before making dir indexed
  ext4: make sure the first directory block is not a hole

 fs/ext4/namei.c | 73 ++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 57 insertions(+), 16 deletions(-)

-- 
2.39.2

[PATCH 1/2] ext4: check dot and dotdot of dx_root before making dir indexed

[libaokun]

From: Baokun Li <libaokun1@huawei.com>

Syzbot reports a issue as follows:
============================================
BUG: unable to handle page fault for address: ffffed11022e24fe
PGD 23ffee067 P4D 23ffee067 PUD 0
Oops: Oops: 0000 [#1] PREEMPT SMP KASAN PTI
CPU: 0 PID: 5079 Comm: syz-executor306 Not tainted 6.10.0-rc5-g55027e689933 #0
Call Trace:
 <TASK>
 make_indexed_dir+0xdaf/0x13c0 fs/ext4/namei.c:2341
 ext4_add_entry+0x222a/0x25d0 fs/ext4/namei.c:2451
 ext4_rename fs/ext4/namei.c:3936 [inline]
 ext4_rename2+0x26e5/0x4370 fs/ext4/namei.c:4214
[...]
============================================

The immediate cause of this problem is that there is only one valid dentry
for the block to be split during do_split, so split==0 results in out of
bounds accesses to the map triggering the issue.

    do_split
      unsigned split
      dx_make_map
       count = 1
      split = count/2 = 0;
      continued = hash2 == map[split - 1].hash;
       ---> map[4294967295]

The maximum length of a filename is 255 and the minimum block size is 1024,
so it is always guaranteed that the number of entries is greater than or
equal to 2 when do_split() is called.

But syzbot's crafted image has no dot and dotdot in dir, and the dentry
distribution in dirblock is as follows:

  bus     dentry1          hole           dentry2           free
|xx--|xx-------------|...............|xx-------------|...............|
0   12 (8+248)=256  268     256     524 (8+256)=264 788     236     1024

So when renaming dentry1 increases its name_len length by 1, neither hole
nor free is sufficient to hold the new dentry, and make_indexed_dir() is
called.

In make_indexed_dir() it is assumed that the first two entries of the
dirblock must be dot and dotdot, so bus and dentry1 are left in dx_root
because they are treated as dot and dotdot, and only dentry2 is moved
to the new leaf block. That's why count is equal to 1.

Therefore add the ext4_check_dx_root() helper function to add more sanity
checks to dot and dotdot before starting the conversion to avoid the above
issue.

Reported-by: syzbot+ae688d469e36fb5138d0@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=ae688d469e36fb5138d0
Fixes: ac27a0ec112a ("[PATCH] ext4: initial copy of files from ext3")
Cc: stable@kernel.org
Signed-off-by: Baokun Li <libaokun1@huawei.com>
---
 fs/ext4/namei.c | 56 ++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 51 insertions(+), 5 deletions(-)

diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
index e6769b97a970..35881e3dd880 100644
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c
@@ -2172,6 +2172,52 @@ static int add_dirent_to_buf(handle_t *handle, struct ext4_filename *fname,
 	return err ? err : err2;
 }
 
+static bool ext4_check_dx_root(struct inode *dir, struct dx_root *root)
+{
+	struct fake_dirent *fde;
+	const char *error_msg;
+	unsigned int rlen;
+	unsigned int blocksize = dir->i_sb->s_blocksize;
+	char *blockend = (char *)root + dir->i_sb->s_blocksize;
+
+	fde = &root->dot;
+	if (unlikely(fde->name_len != 1)) {
+		error_msg = "invalid name_len for '.'";
+		goto corrupted;
+	}
+	if (unlikely(strncmp(root->dot_name, ".", fde->name_len))) {
+		error_msg = "invalid name for '.'";
+		goto corrupted;
+	}
+	rlen = ext4_rec_len_from_disk(fde->rec_len, blocksize);
+	if (unlikely((char *)fde + rlen >= blockend)) {
+		error_msg = "invalid rec_len for '.'";
+		goto corrupted;
+	}
+
+	fde = &root->dotdot;
+	if (unlikely(fde->name_len != 2)) {
+		error_msg = "invalid name_len for '..'";
+		goto corrupted;
+	}
+	if (unlikely(strncmp(root->dotdot_name, "..", fde->name_len))) {
+		error_msg = "invalid name for '..'";
+		goto corrupted;
+	}
+	rlen = ext4_rec_len_from_disk(fde->rec_len, blocksize);
+	if (unlikely((char *)fde + rlen >= blockend)) {
+		error_msg = "invalid rec_len for '..'";
+		goto corrupted;
+	}
+
+	return true;
+
+corrupted:
+	EXT4_ERROR_INODE(dir, "Corrupt dir, %s, running e2fsck is recommended",
+			 error_msg);
+	return false;
+}
+
 /*
  * This converts a one block unindexed directory to a 3 block indexed
  * directory, and adds the dentry to the indexed directory.
@@ -2206,17 +2252,17 @@ static int make_indexed_dir(handle_t *handle, struct ext4_filename *fname,
 		brelse(bh);
 		return retval;
 	}
+
 	root = (struct dx_root *) bh->b_data;
+	if (!ext4_check_dx_root(dir, root)) {
+		brelse(bh);
+		return -EFSCORRUPTED;
+	}
 
 	/* The 0th block becomes the root, move the dirents out */
 	fde = &root->dotdot;
 	de = (struct ext4_dir_entry_2 *)((char *)fde +
 		ext4_rec_len_from_disk(fde->rec_len, blocksize));
-	if ((char *) de >= (((char *) root) + blocksize)) {
-		EXT4_ERROR_INODE(dir, "invalid rec_len for '..'");
-		brelse(bh);
-		return -EFSCORRUPTED;
-	}
 	len = ((char *) root) + (blocksize - csum_size) - (char *) de;
 
 	/* Allocate new block for the 0th block's dirents */
-- 
2.39.2

[PATCH 2/2] ext4: make sure the first directory block is not a hole

[libaokun]

From: Baokun Li <libaokun1@huawei.com>

The syzbot constructs a directory that has no dirblock but is non-inline,
i.e. the first directory block is a hole. And no errors are reported when
creating files in this directory in the following flow.

    ext4_mknod
     ...
      ext4_add_entry
        // Read block 0
        ext4_read_dirblock(dir, block, DIRENT)
          bh = ext4_bread(NULL, inode, block, 0)
          if (!bh && (type == INDEX || type == DIRENT_HTREE))
          // The first directory block is a hole
          // But type == DIRENT, so no error is reported.

After that, we get a directory block without '.' and '..' but with a valid
dentry. This may cause some code that relies on dot or dotdot (such as
make_indexed_dir()) to crash.

Therefore when ext4_read_dirblock() finds that the first directory block
is a hole report that the filesystem is corrupted and return an error to
avoid loading corrupted data from disk causing something bad.

Reported-by: syzbot+ae688d469e36fb5138d0@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=ae688d469e36fb5138d0
Fixes: 4e19d6b65fb4 ("ext4: allow directory holes")
Cc: stable@kernel.org
Signed-off-by: Baokun Li <libaokun1@huawei.com>
---
 fs/ext4/namei.c | 17 ++++++-----------
 1 file changed, 6 insertions(+), 11 deletions(-)

diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
index 35881e3dd880..6a95713f9193 100644
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c
@@ -151,10 +151,11 @@ static struct buffer_head *__ext4_read_dirblock(struct inode *inode,
 
 		return bh;
 	}
-	if (!bh && (type == INDEX || type == DIRENT_HTREE)) {
+	/* The first directory block must not be a hole. */
+	if (!bh && (type == INDEX || type == DIRENT_HTREE || block == 0)) {
 		ext4_error_inode(inode, func, line, block,
-				 "Directory hole found for htree %s block",
-				 (type == INDEX) ? "index" : "leaf");
+				 "Directory hole found for htree %s block %u",
+				 (type == INDEX) ? "index" : "leaf", block);
 		return ERR_PTR(-EFSCORRUPTED);
 	}
 	if (!bh)
@@ -3084,10 +3085,7 @@ bool ext4_empty_dir(struct inode *inode)
 		EXT4_ERROR_INODE(inode, "invalid size");
 		return false;
 	}
-	/* The first directory block must not be a hole,
-	 * so treat it as DIRENT_HTREE
-	 */
-	bh = ext4_read_dirblock(inode, 0, DIRENT_HTREE);
+	bh = ext4_read_dirblock(inode, 0, EITHER);
 	if (IS_ERR(bh))
 		return false;
 
@@ -3529,10 +3527,7 @@ static struct buffer_head *ext4_get_first_dir_block(handle_t *handle,
 		struct ext4_dir_entry_2 *de;
 		unsigned int offset;
 
-		/* The first directory block must not be a hole, so
-		 * treat it as DIRENT_HTREE
-		 */
-		bh = ext4_read_dirblock(inode, 0, DIRENT_HTREE);
+		bh = ext4_read_dirblock(inode, 0, EITHER);
 		if (IS_ERR(bh)) {
 			*retval = PTR_ERR(bh);
 			return NULL;
-- 
2.39.2

Re: [PATCH 1/2] ext4: check dot and dotdot of dx_root before making dir indexed

[Jan Kara]

On Tue 02-07-24 21:23:48, libaokun@huaweicloud.com wrote:
> From: Baokun Li <libaokun1@huawei.com>
> 
> Syzbot reports a issue as follows:
> ============================================
> BUG: unable to handle page fault for address: ffffed11022e24fe
> PGD 23ffee067 P4D 23ffee067 PUD 0
> Oops: Oops: 0000 [#1] PREEMPT SMP KASAN PTI
> CPU: 0 PID: 5079 Comm: syz-executor306 Not tainted 6.10.0-rc5-g55027e689933 #0
> Call Trace:
>  <TASK>
>  make_indexed_dir+0xdaf/0x13c0 fs/ext4/namei.c:2341
>  ext4_add_entry+0x222a/0x25d0 fs/ext4/namei.c:2451
>  ext4_rename fs/ext4/namei.c:3936 [inline]
>  ext4_rename2+0x26e5/0x4370 fs/ext4/namei.c:4214
> [...]
> ============================================
> 
> The immediate cause of this problem is that there is only one valid dentry
> for the block to be split during do_split, so split==0 results in out of
> bounds accesses to the map triggering the issue.
> 
>     do_split
>       unsigned split
>       dx_make_map
>        count = 1
>       split = count/2 = 0;
>       continued = hash2 == map[split - 1].hash;
>        ---> map[4294967295]
> 
> The maximum length of a filename is 255 and the minimum block size is 1024,
> so it is always guaranteed that the number of entries is greater than or
> equal to 2 when do_split() is called.
> 
> But syzbot's crafted image has no dot and dotdot in dir, and the dentry
> distribution in dirblock is as follows:
> 
>   bus     dentry1          hole           dentry2           free
> |xx--|xx-------------|...............|xx-------------|...............|
> 0   12 (8+248)=256  268     256     524 (8+256)=264 788     236     1024
> 
> So when renaming dentry1 increases its name_len length by 1, neither hole
> nor free is sufficient to hold the new dentry, and make_indexed_dir() is
> called.
> 
> In make_indexed_dir() it is assumed that the first two entries of the
> dirblock must be dot and dotdot, so bus and dentry1 are left in dx_root
> because they are treated as dot and dotdot, and only dentry2 is moved
> to the new leaf block. That's why count is equal to 1.
> 
> Therefore add the ext4_check_dx_root() helper function to add more sanity
> checks to dot and dotdot before starting the conversion to avoid the above
> issue.
> 
> Reported-by: syzbot+ae688d469e36fb5138d0@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=ae688d469e36fb5138d0
> Fixes: ac27a0ec112a ("[PATCH] ext4: initial copy of files from ext3")
> Cc: stable@kernel.org
> Signed-off-by: Baokun Li <libaokun1@huawei.com>

Thanks! The patch looks good to me. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
>  fs/ext4/namei.c | 56 ++++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 51 insertions(+), 5 deletions(-)
> 
> diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
> index e6769b97a970..35881e3dd880 100644
> --- a/fs/ext4/namei.c
> +++ b/fs/ext4/namei.c
> @@ -2172,6 +2172,52 @@ static int add_dirent_to_buf(handle_t *handle, struct ext4_filename *fname,
>  	return err ? err : err2;
>  }
>  
> +static bool ext4_check_dx_root(struct inode *dir, struct dx_root *root)
> +{
> +	struct fake_dirent *fde;
> +	const char *error_msg;
> +	unsigned int rlen;
> +	unsigned int blocksize = dir->i_sb->s_blocksize;
> +	char *blockend = (char *)root + dir->i_sb->s_blocksize;
> +
> +	fde = &root->dot;
> +	if (unlikely(fde->name_len != 1)) {
> +		error_msg = "invalid name_len for '.'";
> +		goto corrupted;
> +	}
> +	if (unlikely(strncmp(root->dot_name, ".", fde->name_len))) {
> +		error_msg = "invalid name for '.'";
> +		goto corrupted;
> +	}
> +	rlen = ext4_rec_len_from_disk(fde->rec_len, blocksize);
> +	if (unlikely((char *)fde + rlen >= blockend)) {
> +		error_msg = "invalid rec_len for '.'";
> +		goto corrupted;
> +	}
> +
> +	fde = &root->dotdot;
> +	if (unlikely(fde->name_len != 2)) {
> +		error_msg = "invalid name_len for '..'";
> +		goto corrupted;
> +	}
> +	if (unlikely(strncmp(root->dotdot_name, "..", fde->name_len))) {
> +		error_msg = "invalid name for '..'";
> +		goto corrupted;
> +	}
> +	rlen = ext4_rec_len_from_disk(fde->rec_len, blocksize);
> +	if (unlikely((char *)fde + rlen >= blockend)) {
> +		error_msg = "invalid rec_len for '..'";
> +		goto corrupted;
> +	}
> +
> +	return true;
> +
> +corrupted:
> +	EXT4_ERROR_INODE(dir, "Corrupt dir, %s, running e2fsck is recommended",
> +			 error_msg);
> +	return false;
> +}
> +
>  /*
>   * This converts a one block unindexed directory to a 3 block indexed
>   * directory, and adds the dentry to the indexed directory.
> @@ -2206,17 +2252,17 @@ static int make_indexed_dir(handle_t *handle, struct ext4_filename *fname,
>  		brelse(bh);
>  		return retval;
>  	}
> +
>  	root = (struct dx_root *) bh->b_data;
> +	if (!ext4_check_dx_root(dir, root)) {
> +		brelse(bh);
> +		return -EFSCORRUPTED;
> +	}
>  
>  	/* The 0th block becomes the root, move the dirents out */
>  	fde = &root->dotdot;
>  	de = (struct ext4_dir_entry_2 *)((char *)fde +
>  		ext4_rec_len_from_disk(fde->rec_len, blocksize));
> -	if ((char *) de >= (((char *) root) + blocksize)) {
> -		EXT4_ERROR_INODE(dir, "invalid rec_len for '..'");
> -		brelse(bh);
> -		return -EFSCORRUPTED;
> -	}
>  	len = ((char *) root) + (blocksize - csum_size) - (char *) de;
>  
>  	/* Allocate new block for the 0th block's dirents */
> -- 
> 2.39.2
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [PATCH 2/2] ext4: make sure the first directory block is not a hole

[Jan Kara]

On Tue 02-07-24 21:23:49, libaokun@huaweicloud.com wrote:
> From: Baokun Li <libaokun1@huawei.com>
> 
> The syzbot constructs a directory that has no dirblock but is non-inline,
> i.e. the first directory block is a hole. And no errors are reported when
> creating files in this directory in the following flow.
> 
>     ext4_mknod
>      ...
>       ext4_add_entry
>         // Read block 0
>         ext4_read_dirblock(dir, block, DIRENT)
>           bh = ext4_bread(NULL, inode, block, 0)
>           if (!bh && (type == INDEX || type == DIRENT_HTREE))
>           // The first directory block is a hole
>           // But type == DIRENT, so no error is reported.
> 
> After that, we get a directory block without '.' and '..' but with a valid
> dentry. This may cause some code that relies on dot or dotdot (such as
> make_indexed_dir()) to crash.
> 
> Therefore when ext4_read_dirblock() finds that the first directory block
> is a hole report that the filesystem is corrupted and return an error to
> avoid loading corrupted data from disk causing something bad.
> 
> Reported-by: syzbot+ae688d469e36fb5138d0@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=ae688d469e36fb5138d0
> Fixes: 4e19d6b65fb4 ("ext4: allow directory holes")
> Cc: stable@kernel.org
> Signed-off-by: Baokun Li <libaokun1@huawei.com>

Looks good to me. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
>  fs/ext4/namei.c | 17 ++++++-----------
>  1 file changed, 6 insertions(+), 11 deletions(-)
> 
> diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
> index 35881e3dd880..6a95713f9193 100644
> --- a/fs/ext4/namei.c
> +++ b/fs/ext4/namei.c
> @@ -151,10 +151,11 @@ static struct buffer_head *__ext4_read_dirblock(struct inode *inode,
>  
>  		return bh;
>  	}
> -	if (!bh && (type == INDEX || type == DIRENT_HTREE)) {
> +	/* The first directory block must not be a hole. */
> +	if (!bh && (type == INDEX || type == DIRENT_HTREE || block == 0)) {
>  		ext4_error_inode(inode, func, line, block,
> -				 "Directory hole found for htree %s block",
> -				 (type == INDEX) ? "index" : "leaf");
> +				 "Directory hole found for htree %s block %u",
> +				 (type == INDEX) ? "index" : "leaf", block);
>  		return ERR_PTR(-EFSCORRUPTED);
>  	}
>  	if (!bh)
> @@ -3084,10 +3085,7 @@ bool ext4_empty_dir(struct inode *inode)
>  		EXT4_ERROR_INODE(inode, "invalid size");
>  		return false;
>  	}
> -	/* The first directory block must not be a hole,
> -	 * so treat it as DIRENT_HTREE
> -	 */
> -	bh = ext4_read_dirblock(inode, 0, DIRENT_HTREE);
> +	bh = ext4_read_dirblock(inode, 0, EITHER);
>  	if (IS_ERR(bh))
>  		return false;
>  
> @@ -3529,10 +3527,7 @@ static struct buffer_head *ext4_get_first_dir_block(handle_t *handle,
>  		struct ext4_dir_entry_2 *de;
>  		unsigned int offset;
>  
> -		/* The first directory block must not be a hole, so
> -		 * treat it as DIRENT_HTREE
> -		 */
> -		bh = ext4_read_dirblock(inode, 0, DIRENT_HTREE);
> +		bh = ext4_read_dirblock(inode, 0, EITHER);
>  		if (IS_ERR(bh)) {
>  			*retval = PTR_ERR(bh);
>  			return NULL;
> -- 
> 2.39.2
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [PATCH 0/2] ext4: fix unable to handle kernel paging request in do_split()

[Theodore Ts'o]

On Tue, 02 Jul 2024 21:23:47 +0800, libaokun@huaweicloud.com wrote:
> "kvm-xfstests -c ext4/all -g auto" has been executed with no new failures.
> 
> Baokun Li (2):
>   ext4: check dot and dotdot of dx_root before making dir indexed
>   ext4: make sure the first directory block is not a hole
> 
> fs/ext4/namei.c | 73 ++++++++++++++++++++++++++++++++++++++-----------
>  1 file changed, 57 insertions(+), 16 deletions(-)
> 
> [...]

Applied, thanks!

[1/2] ext4: check dot and dotdot of dx_root before making dir indexed
      commit: 50ea741def587a64e08879ce6c6a30131f7111e7
[2/2] ext4: make sure the first directory block is not a hole
      commit: f9ca51596bbfd0f9c386dd1c613c394c78d9e5e6

Best regards,
-- 
Theodore Ts'o <tytso@mit.edu>