* [BUG] ext4: NULL pointer dereference in jbd2_journal_file_inode() during inline data conversion
@ 2026-11-15 0:43 Ahmed Khan
2026-11-16 21:39 ` Theodore Tso
0 siblings, 1 reply; 2+ messages in thread
From: Ahmed Khan @ 2026-11-15 0:43 UTC (permalink / raw)
To: tytso, adilger.kernel; +Cc: linux-ext4, linux-fsdevel, linux-kernel
Hi Theodore and Andreas,
My name is Ahmed Khan, I am a security researcher from Canada. I
believe I found a bug within the Linux Kernel, specifically, version
7.3.4 affecting all major distros. I will be candid and state that I
used AI heavily in fuzzing the kernel using "syzkaller". Below are its
findings, the proof of concept worked on my QEMU VM repeatedly.
I found and verified a deterministic NULL pointer dereference in
jbd2_journal_file_inode() triggered when expanding an inline_data ext4
file via path-based truncate() on a filesystem mounted in data=ordered
mode.
This corresponds to syzbot crash ID:
3eb9083d961b1b97587d103692546ba06797de4e
1. Root Cause Analysis
----------------------
In ext4, EXT4_I(inode)->jinode is lazily allocated via
ext4_inode_attach_jinode(). Normally this occurs during
ext4_file_open(), ext4_write_begin(), or late inside ext4_setattr().
When path-based truncate() is called on a file without opening a file
descriptor, the inode enters ext4_setattr() with EXT4_I(inode)->jinode
still NULL:
/* fs/ext4/inode.c — ext4_setattr() */
if (ext4_has_inline_data(inode) &&
attr->ia_size > EXT4_I(inode)->i_inline_size) {
error = ext4_convert_inline_data(inode); /* jinode is still NULL */
if (error)
goto err_out;
}
...
/* ext4_inode_attach_jinode() is only called later in ext4_setattr() */
During inline conversion, ext4_convert_inline_data_nolock() calls
ext4_map_blocks(..., EXT4_GET_BLOCKS_CREATE). When the inode also has
EXT4_INODE_EXTENTS set and the filesystem is mounted with data=ordered,
ext4_map_blocks() calls ext4_jbd2_inode_add_write(), passing
EXT4_I(inode)->jinode (NULL) directly to jbd2_journal_file_inode().
jbd2_journal_file_inode() then acquires journal->j_list_lock (with IRQs
disabled) and dereferences the NULL jinode pointer at offset 0x28/0x38,
causing a fatal Oops/GPF while holding the journal list lock.
2. Reproduction Steps
---------------------
1. Mount a crafted ext4 image in default data=ordered mode containing a
regular file with both EXT4_INLINE_DATA_FL and EXT4_EXTENTS_FL set.
2. Without opening a file descriptor on the target file, invoke:
truncate("./file3", 0x10001);
* Proof of Concept: Linux Kernel ext4/jbd2 NULL Pointer Dereference
*
* Target Linux Kernel Versions:
* - Upstream Linux: 7.0 – 7.3.4+
* - Confirmed Distro Kernels (6/6 Tested Distros Crashed):
* * Ubuntu 26.04.1 LTS (Linux 7.0.0-38-generic #38-Ubuntu)
* * Ubuntu 26.10 (Linux 7.3.0-10-generic #10-Ubuntu)
* * Fedora Cloud 45 (Linux 7.2.11-300.fc45.x86_64)
* * Debian Sid / Forky (Linux 7.2.15+deb14-cloud-amd64)
* * Arch Linux (Linux 7.2.11-arch1-1)
* * openSUSE Tumbleweed (Linux 7.3.3-1-default)
* * Upstream Base / CachyOS (Linux 7.3.4 SMP KASAN)
*
* Vulnerability Summary:
* Triggered via truncate("./file3", 0x10001) on a crafted ext4 filesystem
* where the target inode has both EXT4_INLINE_DATA_FL and
EXT4_EXTENTS_FL set.
* ext4_setattr() calls ext4_convert_inline_data() before attaching a jinode,
* causing ext4_jbd2_inode_add_write() to pass jinode == NULL into
* jbd2_journal_file_inode(), causing an immediate NULL pointer
dereference (GPF)
* while holding journal->j_list_lock with interrupts disabled.
*
* Crash Signature:
* Oops: general protection fault, address: 0x0000000000000028 (or 0x38)
* RIP: 0010:jbd2_journal_file_inode+0x14b/0x5b0
* note: poc1 exited with irqs disabled / preempt_count 1
*
* Compilation:
* gcc -O2 -o poc1 poc1_ext4_jbd2_nullderef.c
* Prerequisite: ensure /dev/loop0 exists (modprobe loop && losetup -f)
*
* Originally autogenerated by syzkaller (https://github.com/google/syzkaller)
3. Crash Log (KASAN)
--------------------
[ 18.161021] Oops: general protection fault, probably for
non-canonical address
0xdffffc0000000005: 0000 [#1] SMP KASAN NOPTI
[ 18.162724] RIP: 0010:jbd2_journal_file_inode+0x14b/0x5b0
[ 18.163481] Code: 48 c1 e8 03 42 80 3c 30 00 74 08 4c 89 e7 e8 4a
41 8c ff 49 8b 44 24 28 ...
[ 18.170125] Call Trace:
[ 18.170410] <TASK>
[ 18.170685] ext4_jbd2_inode_add_write+0x138/0x1e0
[ 18.171220] ext4_map_blocks+0x1102/0x1b40
[ 18.171691] ext4_convert_inline_data_nolock+0x18e/0x450
[ 18.172281] ext4_convert_inline_data+0x17c/0x250
[ 18.172780] ext4_setattr+0xa2c/0x1ea0
[ 18.173200] notify_change+0x628/0xd60
[ 18.173620] do_truncate+0x144/0x230
[ 18.174020] do_sys_truncate+0x17e/0x200
[ 18.174450] ksys_truncate+0x2c/0x40
[ 18.174820] do_syscall_64+0x10e/0x5e0
[ 18.176450] entry_SYSCALL_64_after_hwframe+0x77/0x7f
[ 18.178920] note: poc1[318] exited with irqs disabled / preempt_count
4. Suggested Fix Direction
--------------------------
Calling ext4_inode_attach_jinode(inode) inside ext4_convert_inline_data()
(or prior to calling ext4_convert_inline_data() in ext4_setattr())
ensures jinode is allocated before ext4_map_blocks() attempts to add
the inode to the journal's ordered data list.
Please let me know if you would like the standalone C reproducer or
the minimal ext4 disk image attached inline. I have this in a private
repo and would be happy to add you as a project contributor.
Best regards,
Ahmed Khan
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [BUG] ext4: NULL pointer dereference in jbd2_journal_file_inode() during inline data conversion
2026-11-15 0:43 [BUG] ext4: NULL pointer dereference in jbd2_journal_file_inode() during inline data conversion Ahmed Khan
@ 2026-11-16 21:39 ` Theodore Tso
0 siblings, 0 replies; 2+ messages in thread
From: Theodore Tso @ 2026-11-16 21:39 UTC (permalink / raw)
To: Ahmed Khan; +Cc: adilger.kernel, linux-ext4, linux-fsdevel, linux-kernel
On Sat, Nov 14, 2026 at 07:43:08PM -0500, Ahmed Khan wrote:
>
> My name is Ahmed Khan, I am a security researcher from Canada. I
> believe I found a bug within the Linux Kernel, specifically, version
> 7.3.4 affecting all major distros. I will be candid and state that I
> used AI heavily in fuzzing the kernel using "syzkaller"....
When you say "syzkaller", are you referring to your own instance of
syzkaller, which you may have modified to use AI-directed fuzzing?
> This corresponds to syzbot crash ID:
> 3eb9083d961b1b97587d103692546ba06797de4e
This crash ID doesn't appear to be one which is recognized by
https://syzkaller.appspot.com --- which would make sense if it's your
own Syzkaller instance.
> Please let me know if you would like the standalone C reproducer or
> the minimal ext4 disk image attached inline. I have this in a private
> repo and would be happy to add you as a project contributor.
Yes, please send the standalone C reproducer and minimal crafted ext4 image.
That would be most helpful.
Note that we don't really consider crashes which require a crafted
file system image to trigger to be a security vulnerability, unless
the file system corruption is not caught by e2fsck (or triggers a
memory pointer bug or other issue in e2fsck). Not running fsck on an
untrusted file system before introducing it to the Trusted Computing
Base of the linux kernel is roughly equivalent to running "dd
if=/dev/zero of=/dev/mem". Sure, it might crash the system, but
that's really a PEBCAK system administrator issue. :-)
Regards,
- Ted
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-16 21:40 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-15 0:43 [BUG] ext4: NULL pointer dereference in jbd2_journal_file_inode() during inline data conversion Ahmed Khan
2026-11-16 21:39 ` Theodore Tso
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox