1. Patchew
  2. linux
  3. View series

[PATCH v4] ocfs2: Invalidate inode if i_mode is zero after block read

Ahmet Eray Karadag posted 1 patch 2 months, 1 week ago
There is a newer version of this series
fs/ocfs2/inode.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
[PATCH v4] ocfs2: Invalidate inode if i_mode is zero after block read
Posted by Ahmet Eray Karadag 2 months, 1 week ago 
A panic occurs in ocfs2_unlink due to WARN_ON(inode->i_nlink == 0) when
handling a corrupted inode with i_mode=0 and i_nlink=0 in memory.

This "zombie" inode is created because ocfs2_read_locked_inode proceeds
even after ocfs2_validate_inode_block successfully validates a block
that structurally looks okay (passes checksum, signature etc.) but
contains semantically invalid data (specifically i_mode=0). The current
validation function doesn't check for i_mode being zero.

This results in an in-memory inode with i_mode=0 being added to the VFS
cache, which later triggers the panic during unlink.

Prevent this by adding an explicit check for (i_mode == 0, i_nlink == 0, non-orphan)
within ocfs2_validate_inode_block. If the check is true, return -EFSCORRUPTED to signal
corruption. This causes the caller (ocfs2_read_locked_inode) to invoke
make_bad_inode(), correctly preventing the zombie inode from entering
the cache.

Reported-by: syzbot+55c40ae8a0e5f3659f2b@syzkaller.appspotmail.com
Fixes: https://syzkaller.appspot.com/bug?extid=55c40ae8a0e5f3659f2b
Co-developed-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
Signed-off-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
Signed-off-by: Ahmet Eray Karadag <eraykrdg1@gmail.com>
Previous link: https://lore.kernel.org/all/20251022222752.46758-2-eraykrdg1@gmail.com/T/
---
v2:
 - Only checking either i_links_count == 0 or i_mode == 0
 - Not performing le16_to_cpu() anymore
 - Tested with ocfs2-test
---
v3:
 - Add checking both high and low bits of i_links_count
---
v4:
 - Reading i_links_count hi and low bits without helper function
  to save few cpu cycles
---
 fs/ocfs2/inode.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
index 14bf440ea4df..34c2882273ae 100644
--- a/fs/ocfs2/inode.c
+++ b/fs/ocfs2/inode.c
@@ -1455,7 +1455,13 @@ int ocfs2_validate_inode_block(struct super_block *sb,
 		     (unsigned long long)bh->b_blocknr);
 		goto bail;
 	}
-
+	if (!(di->i_links_count | di->i_links_count_hi) || !di->i_mode) {
+		mlog(ML_ERROR, "Invalid dinode #%llu: "
+			"Corrupt state (nlink=0 or mode=0,) detected!\n",
+		        (unsigned long long)bh->b_blocknr);
+		rc = -EFSCORRUPTED;
+		goto bail;
+	}
 	/*
 	 * Errors after here are fatal.
 	 */
-- 
2.43.0
Re: [PATCH v4] ocfs2: Invalidate inode if i_mode is zero after block read
Posted by Joseph Qi 2 months, 1 week ago 

On 2025/12/2 08:32, Ahmet Eray Karadag wrote:
> A panic occurs in ocfs2_unlink due to WARN_ON(inode->i_nlink == 0) when
> handling a corrupted inode with i_mode=0 and i_nlink=0 in memory.
> 
> This "zombie" inode is created because ocfs2_read_locked_inode proceeds
> even after ocfs2_validate_inode_block successfully validates a block
> that structurally looks okay (passes checksum, signature etc.) but
> contains semantically invalid data (specifically i_mode=0). The current
> validation function doesn't check for i_mode being zero.
> 
> This results in an in-memory inode with i_mode=0 being added to the VFS
> cache, which later triggers the panic during unlink.
> 
> Prevent this by adding an explicit check for (i_mode == 0, i_nlink == 0, non-orphan)
> within ocfs2_validate_inode_block. If the check is true, return -EFSCORRUPTED to signal
> corruption. This causes the caller (ocfs2_read_locked_inode) to invoke
> make_bad_inode(), correctly preventing the zombie inode from entering
> the cache.
> 
> Reported-by: syzbot+55c40ae8a0e5f3659f2b@syzkaller.appspotmail.com
> Fixes: https://syzkaller.appspot.com/bug?extid=55c40ae8a0e5f3659f2b
> Co-developed-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
> Signed-off-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
> Signed-off-by: Ahmet Eray Karadag <eraykrdg1@gmail.com>
> Previous link: https://lore.kernel.org/all/20251022222752.46758-2-eraykrdg1@gmail.com/T/
> ---
> v2:
>  - Only checking either i_links_count == 0 or i_mode == 0
>  - Not performing le16_to_cpu() anymore
>  - Tested with ocfs2-test
> ---
> v3:
>  - Add checking both high and low bits of i_links_count
> ---
> v4:
>  - Reading i_links_count hi and low bits without helper function
>   to save few cpu cycles
> ---
>  fs/ocfs2/inode.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
> index 14bf440ea4df..34c2882273ae 100644
> --- a/fs/ocfs2/inode.c
> +++ b/fs/ocfs2/inode.c
> @@ -1455,7 +1455,13 @@ int ocfs2_validate_inode_block(struct super_block *sb,
>  		     (unsigned long long)bh->b_blocknr);
>  		goto bail;
>  	}
> -
> +	if (!(di->i_links_count | di->i_links_count_hi) || !di->i_mode) {

"(di->i_links_count | di->i_links_count_hi)" looks meaningless.
So if we don't want to introduce the endian coversion here, how about:

if ((!di->i_links_count && !di->i_links_count_hi) || !di->i_mode) {
	...
}

Heming, what's your opinion?

> +		mlog(ML_ERROR, "Invalid dinode #%llu: "
> +			"Corrupt state (nlink=0 or mode=0,) detected!\n",

Better to log the actual i_nlink/i_mode. e.g.

"corrupted i_nlink %u or i_mode %u\n"
...

Joseph

> +		        (unsigned long long)bh->b_blocknr);
> +		rc = -EFSCORRUPTED;
> +		goto bail;
> +	}
>  	/*
>  	 * Errors after here are fatal.
>  	 */
Re: [PATCH v4] ocfs2: Invalidate inode if i_mode is zero after block read
Posted by Heming Zhao 2 months, 1 week ago 
On Tue, Dec 02, 2025 at 10:44:21AM +0800, Joseph Qi wrote:
> 
> 
> On 2025/12/2 08:32, Ahmet Eray Karadag wrote:
> > A panic occurs in ocfs2_unlink due to WARN_ON(inode->i_nlink == 0) when
> > handling a corrupted inode with i_mode=0 and i_nlink=0 in memory.
> > 
> > This "zombie" inode is created because ocfs2_read_locked_inode proceeds
> > even after ocfs2_validate_inode_block successfully validates a block
> > that structurally looks okay (passes checksum, signature etc.) but
> > contains semantically invalid data (specifically i_mode=0). The current
> > validation function doesn't check for i_mode being zero.
> > 
> > This results in an in-memory inode with i_mode=0 being added to the VFS
> > cache, which later triggers the panic during unlink.
> > 
> > Prevent this by adding an explicit check for (i_mode == 0, i_nlink == 0, non-orphan)
> > within ocfs2_validate_inode_block. If the check is true, return -EFSCORRUPTED to signal
> > corruption. This causes the caller (ocfs2_read_locked_inode) to invoke
> > make_bad_inode(), correctly preventing the zombie inode from entering
> > the cache.
> > 
> > Reported-by: syzbot+55c40ae8a0e5f3659f2b@syzkaller.appspotmail.com
> > Fixes: https://syzkaller.appspot.com/bug?extid=55c40ae8a0e5f3659f2b
> > Co-developed-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
> > Signed-off-by: Albin Babu Varghese <albinbabuvarghese20@gmail.com>
> > Signed-off-by: Ahmet Eray Karadag <eraykrdg1@gmail.com>
> > Previous link: https://lore.kernel.org/all/20251022222752.46758-2-eraykrdg1@gmail.com/T/
> > ---
> > v2:
> >  - Only checking either i_links_count == 0 or i_mode == 0
> >  - Not performing le16_to_cpu() anymore
> >  - Tested with ocfs2-test
> > ---
> > v3:
> >  - Add checking both high and low bits of i_links_count
> > ---
> > v4:
> >  - Reading i_links_count hi and low bits without helper function
> >   to save few cpu cycles
> > ---
> >  fs/ocfs2/inode.c | 8 +++++++-
> >  1 file changed, 7 insertions(+), 1 deletion(-)
> > 
> > diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
> > index 14bf440ea4df..34c2882273ae 100644
> > --- a/fs/ocfs2/inode.c
> > +++ b/fs/ocfs2/inode.c
> > @@ -1455,7 +1455,13 @@ int ocfs2_validate_inode_block(struct super_block *sb,
> >  		     (unsigned long long)bh->b_blocknr);
> >  		goto bail;
> >  	}
> > -
> > +	if (!(di->i_links_count | di->i_links_count_hi) || !di->i_mode) {
> 
> "(di->i_links_count | di->i_links_count_hi)" looks meaningless.
> So if we don't want to introduce the endian coversion here, how about:
> 
> if ((!di->i_links_count && !di->i_links_count_hi) || !di->i_mode) {
> 	...
> }
> 
> Heming, what's your opinion?

clearer, looks good to me.

> 
> > +		mlog(ML_ERROR, "Invalid dinode #%llu: "
> > +			"Corrupt state (nlink=0 or mode=0,) detected!\n",
> 
> Better to log the actual i_nlink/i_mode. e.g.
> 
> "corrupted i_nlink %u or i_mode %u\n"
> ...
> 
> Joseph

agree.

- Heming
> 
> > +		        (unsigned long long)bh->b_blocknr);
> > +		rc = -EFSCORRUPTED;
> > +		goto bail;
> > +	}
> >  	/*
> >  	 * Errors after here are fatal.
> >  	 */
>

Patchew 2.1-devel-b67e4c2

© 2016 - 2026 Red Hat, Inc.