Re: [PATCH -next] nbd: Fix hungtask when nbd_config_put

[Josef Bacik <josef@toxicpanda.com>]

On Wed, Oct 20, 2021 at 05:48:30PM +0800, Ye Bin wrote:
> I got follow issue:
> [  247.381177] INFO: task kworker/u10:0:47 blocked for more than 120 seconds.
> [  247.382644]       Not tainted 4.19.90-dirty #140
> [  247.383502] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> [  247.385027] Call Trace:
> [  247.388384]  schedule+0xb8/0x3c0
> [  247.388966]  schedule_timeout+0x2b4/0x380
> [  247.392815]  wait_for_completion+0x367/0x510
> [  247.397713]  flush_workqueue+0x32b/0x1340
> [  247.402700]  drain_workqueue+0xda/0x3c0
> [  247.403442]  destroy_workqueue+0x7b/0x690
> [  247.405014]  nbd_config_put.cold+0x2f9/0x5b6
> [  247.405823]  recv_work+0x1fd/0x2b0
> [  247.406485]  process_one_work+0x70b/0x1610
> [  247.407262]  worker_thread+0x5a9/0x1060
> [  247.408699]  kthread+0x35e/0x430
> [  247.410918]  ret_from_fork+0x1f/0x30
> 
> We can reprodeuce issue as follows:
> 1. Inject memory fault in nbd_start_device
> @@ -1244,10 +1248,18 @@ static int nbd_start_device(struct nbd_device *nbd)
>         nbd_dev_dbg_init(nbd);
>         for (i = 0; i < num_connections; i++) {
>                 struct recv_thread_args *args;
> -
> -               args = kzalloc(sizeof(*args), GFP_KERNEL);
> +
> +               if (i == 1) {
> +                       args = NULL;
> +                       printk("%s: inject malloc error\n", __func__);
> +               }
> +               else
> +                       args = kzalloc(sizeof(*args), GFP_KERNEL);
> 2. Inject delay in recv_work
> @@ -757,6 +760,8 @@ static void recv_work(struct work_struct *work)
> 
>                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
>         }
> +       printk("%s: comm=%s pid=%d\n", __func__, current->comm, current->pid);
> +       mdelay(5 * 1000);
>         nbd_config_put(nbd);
>         atomic_dec(&config->recv_threads);
>         wake_up(&config->recv_wq);
> 3. Create nbd server
> nbd-server 8000 /tmp/disk
> 4. Create nbd client
> nbd-client localhost 8000 /dev/nbd1
> Then will trigger above issue.
> 
> Reason is when add delay in recv_work, lead to relase the last reference
> of 'nbd->config_refs'. nbd_config_put will call flush_workqueue to make
> all work finish. Obviously, it will lead to deadloop.
> To solve this issue, we must ensure 'recv_work' all exit before release
> the last 'nbd->config_refs' reference count.
> 
> Signed-off-by: Ye Bin <yebin10@huawei.com>

This isn't the way to fix this.  The crux of the problem is the fact that we
destroy the recv_workq at config time anyway, since it is tied to the nbd
device.

Fix this instead by getting rid of that, and have the nbd device teardown the
recv_workq when we destory the nbd device.  Alloc it on device allocation and
stop alloc'ing it with the config.  That way we don't have this weird thing
where we need recv threads to drop their config before the last thing.

And while you're at it would you fix recv_work() so it doesn't modify config
after it drops it's reference?  Thanks,

Josef