System freeze after suspending two instances of Neovim with io_uring enabled

[pajlada]

Hi!
When running Arch Linux or Fedora Rawhide and suspending two instances of Neovim, which uses libuv, which uses io_uring, I experience a system freeze. It stops me from typing anything in any shell, or spawn any new shell, but I'm able to run some simple commands over ssh (e.g. ssh myserver ls -la).
dmesg doesn't report anything interesting as far as I could tell, other than some of the apps that were running not being responsive.

Mar 31 11:53:31 billy kernel: INFO: task st:29757 blocked for more than 122 seconds.
Mar 31 11:53:31 billy kernel:       Tainted: P           OE      6.8.2-arch2-1 #1
Mar 31 11:53:31 billy kernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
Mar 31 11:53:31 billy kernel: task:st              state:D stack:0     pid:29757 tgid:29757 ppid:29751  flags:0x00004006
Mar 31 11:53:31 billy kernel: Call Trace:
Mar 31 11:53:31 billy kernel:  <TASK>
Mar 31 11:53:31 billy kernel:  __schedule+0x3e6/0x1520
Mar 31 11:53:31 billy kernel:  schedule+0x32/0xd0
Mar 31 11:53:31 billy kernel:  schedule_timeout+0x151/0x160
Mar 31 11:53:31 billy kernel:  wait_for_completion+0x86/0x170
Mar 31 11:53:31 billy kernel:  __flush_work.isra.0+0x173/0x280
Mar 31 11:53:31 billy kernel:  ? __pfx_wq_barrier_func+0x10/0x10
Mar 31 11:53:31 billy kernel:  n_tty_poll+0x134/0x1e0
Mar 31 11:53:31 billy kernel:  tty_poll+0x57/0xc0
Mar 31 11:53:31 billy kernel:  do_select+0x362/0x880
Mar 31 11:53:31 billy kernel:  ? pollwake+0x50/0xa0
Mar 31 11:53:31 billy kernel:  ? __pfx_pollwake+0x10/0x10
Mar 31 11:53:31 billy kernel:  ? __pfx_pollwake+0x10/0x10
Mar 31 11:53:31 billy kernel:  ? __pfx_pollwake+0x10/0x10
Mar 31 11:53:31 billy kernel:  core_sys_select+0x36b/0x530
Mar 31 11:53:31 billy kernel:  do_pselect.constprop.0+0xe9/0x180
Mar 31 11:53:31 billy kernel:  __x64_sys_pselect6+0x3d/0x70
Mar 31 11:53:31 billy kernel:  do_syscall_64+0x86/0x170
Mar 31 11:53:31 billy kernel:  ? do_syscall_64+0x96/0x170
Mar 31 11:53:31 billy kernel:  ? do_syscall_64+0x96/0x170
Mar 31 11:53:31 billy kernel:  ? exc_page_fault+0x7f/0x180
Mar 31 11:53:31 billy kernel:  entry_SYSCALL_64_after_hwframe+0x6e/0x76
Mar 31 11:53:31 billy kernel: RIP: 0033:0x7541e536b640
Mar 31 11:53:31 billy kernel: RSP: 002b:00007ffec990be70 EFLAGS: 00000202 ORIG_RAX: 000000000000010e
Mar 31 11:53:31 billy kernel: RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007541e536b640
Mar 31 11:53:31 billy kernel: RDX: 0000000000000000 RSI: 00007ffec990bf50 RDI: 0000000000000005
Mar 31 11:53:31 billy kernel: RBP: 0000000000000000 R08: 0000000000000000 R09: 00007ffec990beb0
Mar 31 11:53:31 billy kernel: R10: 0000000000000000 R11: 0000000000000202 R12: bff0000000000000
Mar 31 11:53:31 billy kernel: R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000010
Mar 31 11:53:31 billy kernel:  </TASK>

The freeze doesn't occur after disabling io_uring in libuv using UV_USE_IO_URING=0 or in the kernel with sysctl kernel.io_uring_disabled=1

uname -a from the tested systems

• Linux billy 6.6.23-1-lts #1 SMP PREEMPT_DYNAMIC Wed, 27 Mar 2024 07:47:20 +0000 x86_64 GNU/Linux running Arch Linux
• Linux yolen 6.8.2-arch2-1 #1 SMP PREEMPT_DYNAMIC Thu, 28 Mar 2024 17:06:35 +0000 x86_64 GNU/Linux running Arch Linux
• Linux localhost 6.9.0-0.rc1.20240329git317c7bc0ef03.20.fc41.x86_64 #1 SMP PREEMPT_DYNAMIC Fri Mar 29 14:04:53 UTC 2024 x86_64 GNU/Linux running Fedora Rawhide

Reproduction steps

• Be on Arch Linux (bug reproducible on netcup & hetzner servers) or Fedora Rawhide
• Install Neovim (pacman -S neovim)
• Run Neovim
• Suspend Neovim (by pressing CTRL+Z)
• Run Neovim again
• Suspend Neovim (by pressing CTRL+Z)
• Your system is now most likely unresponsive

The suspension can be done in separate shells, or as different users with the same results.

Video showing off the freeze

nvim-suspend-freeze.mp4

I'm still able to run certain apps on the system, but not open a shell

If the io-uring@vger.kernel.org email is a better place for this report let me know and I'll report it there instead.
Originally reported in libuv/libuv#4377

[axboe]

Funky! Thanks for the report, I'll take a look.

[axboe]

Was able to reproduce the stall. Wow, libuv does some funky stuff.

[axboe]

Just to be clear, this is obviously an io_uring bug, regardless of what libuv does!

[pajlada]

Awesome! Glad to hear you were able to reproduce it. Thank you!

[axboe]

Unsure if you're able to test kernel patches, but I believe the below should do it. Looks like we get into an inversion between the events workqueue being flushed for console output, and io_uring ring exits for some weird cases. If not, then I'll get it into 6.9-rc3 end of this week and it can bubble back to stable from there.

diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 5d4b448fdc50..f6277e029d5f 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -147,6 +147,7 @@ static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 static void io_queue_sqe(struct io_kiocb *req);
 
 struct kmem_cache *req_cachep;
+static struct workqueue_struct *iou_wq __ro_after_init;
 
 static int __read_mostly sysctl_io_uring_disabled;
 static int __read_mostly sysctl_io_uring_group = -1;
@@ -3161,7 +3162,7 @@ static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
 	 * noise and overhead, there's no discernable change in runtime
 	 * over using system_wq.
 	 */
-	queue_work(system_unbound_wq, &ctx->exit_work);
+	queue_work(iou_wq, &ctx->exit_work);
 }
 
 static int io_uring_release(struct inode *inode, struct file *file)
@@ -4185,6 +4186,8 @@ static int __init io_uring_init(void)
 	io_buf_cachep = KMEM_CACHE(io_buffer,
 					  SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
 
+	iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
+
 #ifdef CONFIG_SYSCTL
 	register_sysctl_init("kernel", kernel_io_uring_disabled_table);
 #endif

[axboe]

Oh, and if you want a Reported-by: tag in the commit, please do let me know and I'll update it with that. Just need an identity + email for that. Queued up:

https://git.kernel.dk/cgit/linux/commit/?h=io_uring-6.9&id=d1a9cef84784f873b457f2622ca2415c4b4db748

[pajlada]

A Reported-by tag would be appreciated yeah! My identity + email below:
Rasmus Karlsson <rasmus.karlsson@pajlada.com>

I'll see if I can test the patch, or if not test the rc3 kernel when that's released

[axboe]

Perfect, commit updated:

https://git.kernel.dk/cgit/linux/commit/?h=io_uring-6.9&id=e5444baa42e545bb929ba56c497e7f3c73634099

[pajlada]

Just applied the patch to my 6.8.2 kernel on a system that previously experienced the issue and I can confirm that it fixes it. Thanks for the quick turnaround!

[redbaron]

out of curiousity, why did all shells freeze, but ssh didn't?

[ichernev]

Perfect, commit updated:

https://git.kernel.dk/cgit/linux/commit/?h=io_uring-6.9&id=e5444baa42e545bb929ba56c497e7f3c73634099

I tested this on top of arch linux kernel 6.8.2-arch2-2 and it seems to work. You have my permission to add Tested-by: Iskren Chernev <me@iskren.info>

[ichernev]

out of curiousity, why did all shells freeze, but ssh didn't?

For me ssh was freezing too. But I also failed to kill the nvim after it froze... maybe there are a few variations of this.

[axboe]

Thanks everyone, patch will go upstream later this week, and it'll bubble back to -stable post that. Marking this one as closed as fix exists.