mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Bernd Schubert <bernd@bsbernd.com>
To: n3rd <18670333770@163.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Cc: Joanne Koong <joannelkoong@gmail.com>,
	Miklos Szeredi <miklos@szeredi.hu>,
	fuse-devel <fuse-devel@lists.linux.dev>
Subject: Re: forged reply bit triggers WARN_ON in fuse-over-io_uring
Date: Fri, 20 Nov 2026 17:53:47 +0100	[thread overview]
Message-ID: <23404e53-49d6-4775-ae96-f8abaac7a145@bsbernd.com> (raw)
In-Reply-To: <6b97fb98.dc1ff.1a1f6b99657.Coremail.18670333770@163.com>



On 11/20/26 03:12, n3rd wrote:
> 
> Hello,
> 
> I am reporting a bug in the Linux kernel.
> 
> 1. Affected kernel versions
> Linux v7.3-rc7 (mainline, tested as 7.3.0-rc7) is affected and was
> reproduced here. The same code is present in mainline 65ba8f34ec88
> (checked on 2026-10-28, fs/fuse/dev_uring.c unchanged in that area). The
> v7.2.x and 6.x stable branches were not tested or bisected, so no claim
> is made about them. The code has been there since the fuse-over-io_uring
> channel was introduced.


Feel free to send a patch that changes to pr_info_ratelimited().

Thanks,
Bernd


> 
> 2. Problem description
> fuse-over-io_uring keeps the FUSE request and reply headers in daemon
> memory. When a daemon commits a reply, fuse_uring_commit_fetch() uses
> copy_header_from_ring() to copy the 128-byte in_out area of the ring
> entry into a kernel buffer; the first 16 bytes are struct
> fuse_out_header (len, error, unique), entirely supplied by the daemon.
> fuse_uring_out_header_has_err() then validates that header: it only
> rejects unique == 0, only checks the error range, and at the unique
> comparison it uses (oh->unique & ~FUSE_INT_REQ_BIT) != req->in.h.unique,
> which masks off exactly the bit the later WARN_ON_ONCE tests. Kernel-
> generated unique values are always even (fuse_get_unique_locked() adds
> FUSE_REQ_ID_STEP, defined as 1ULL << 1), so bit 0 is normally always
> zero; a daemon that replies with unique = commit_id | 1 passes the mask
> comparison and then necessarily hits WARN_ON_ONCE(oh->unique &
> FUSE_INT_REQ_BIT) at fs/fuse/dev_uring.c:653. On a default kernel this
> prints one WARNING per boot and the reply is still treated as a normal
> success (err = 0), so there is no memory corruption, no out-of-bounds
> access, no use-after-free and no privilege escalation. On a kernel built
> or booted with panic_on_warn=1, which is common for Android, ChromeOS
> and hardened/cloud kernels, the same single forged bit panics the whole
> machine, and the user can do it again after every reboot. The legacy /
> dev/fuse path (fs/fuse/dev.c, interrupt reply branch) handles the very
> same input as an ordinary error branch and does not warn, which is why
> this looks like a bug in the io_uring path rather than intended behaviour.
> 
> 3. Reproducer
> A self-contained C program (about 450 lines, static, no libfuse) mounts
> its own fuse filesystem, answers FUSE_INIT with FUSE_OVER_IO_URING,
> registers ring entries with IORING_OP_URING_CMD through a raw SQE128
> io_uring, then replies to the FUSE_LOOKUP generated by a child open()
> with an out header whose unique is commit_id | FUSE_INT_REQ_BIT and
> commits it. A negative control replies with unique = commit_id (bit 0
> clear) through the same mount and the same path. The complete reproducer
> and the serial logs are available on request; they are not posted here.
> 
> 4. Conditions
> Four conditions: (1) fuse.enable_uring=1. That is module_param(bool,
> 0644), off by default and writable by root only; I found no distribution
> that enables it by default, so an operator has to turn it on for
> performance, usually via /etc/modprobe.d or the kernel command line. (2)
> The attacker can open /dev/fuse, which udev's 50-udev-default.rules
> makes mode 0666 on a normal distribution. (3) kernel.io_uring_disabled=0
> and support for IORING_SETUP_SQE128 and IORING_OP_URING_CMD. (4) The
> attacker's daemon completes FUSE_INIT and registers ring entries. The
> attacker does not need root or CAP_SYS_ADMIN; with CONFIG_USER_NS the
> whole sequence runs inside an unprivileged user namespace, which is what
> was tested here (uid 1000, setuid then unshare(CLONE_NEWUSER|
> CLONE_NEWNS); the kernel printed CPU: 0 UID: 1000 before panicking).
> Triggering is deterministic and needs no race. Tested on x86_64, QEMU
> 6.2 TCG, 2 vCPUs, 2048 MB, initramfs, 7.3.0-rc7, with KASAN enabled as a
> control that stayed silent.
> 
> 5. Suspected location
> fs/fuse/dev_uring.c: fuse_uring_out_header_has_err() (v7.3-rc7 line 653,
> the WARN_ON_ONCE), reached from fuse_uring_commit_fetch() and
> copy_header_from_ring(); contrast fs/fuse/dev.c line 1947 (legacy
> interrupt reply handling).
> 
> 6. Suggested fix
> Do not assert on a field the daemon controls. Handle it like the legacy
> path: if (oh->unique & FUSE_INT_REQ_BIT) { pr_warn_ratelimited(...); err
> = -ENOENT; goto err; }. A regression test that replies with unique =
> commit_id | 1 and expects no WARN and an -ENOENT rejection would cover it.
> 
> 7. Mitigation
> Keep fuse.enable_uring=0 (the default) and do not enable it for
> performance; deployments that already enabled it should turn it off and
> reload fuse. On hardened kernels, setting panic_on_warn=0 only
> downgrades the panic to a WARNING and does not fix the bug.
> 
> 8. Disclosure
> Public. Per the kernel security process this report is being sent to the
> public mailing list and the fuse maintainers with the security team
> copied; no embargo is requested and no CVE was assigned here. The
> reproducer is not posted publicly; it is available on request.
> 
> 9. AI-assistance note
> This report was prepared with an AI coding assistant, which was used for
> source review, for building and driving the QEMU guest used to reproduce
> the issue, and for drafting this text. Per the kernel security process I
> am treating the issue as public and am not posting the reproducer.
> 


           reply	other threads:[~2026-11-20 16:53 UTC|newest]

Thread overview: expand[flat|nested]  mbox.gz  Atom feed
 [parent not found: <6b97fb98.dc1ff.1a1f6b99657.Coremail.18670333770@163.com>]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=23404e53-49d6-4775-ae96-f8abaac7a145@bsbernd.com \
    --to=bernd@bsbernd.com \
    --cc=18670333770@163.com \
    --cc=fuse-devel@lists.linux.dev \
    --cc=joannelkoong@gmail.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=miklos@szeredi.hu \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®