Re: [PATCH -next v3 2/2] nbd: Fix hungtask when nbd_config_put

[Josef Bacik]

On Fri, Oct 29, 2021 at 05:42:28PM +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:

"reproduce"

> 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

"release"

> 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, according to Josef's suggestion move 'recv_work'
> init from start device to nbd_dev_add, then destory 'recv_work'when

"destroy"

> nbd device teardown.
> 
> Signed-off-by: Ye Bin <yebin10@xxxxxxxxxx>
> ---
>  drivers/block/nbd.c | 30 ++++++++++++++----------------
>  1 file changed, 14 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
> index 096883ab9b76..c9a65a260668 100644
> --- a/drivers/block/nbd.c
> +++ b/drivers/block/nbd.c
> @@ -1314,10 +1314,6 @@ static void nbd_config_put(struct nbd_device *nbd)
>  		kfree(nbd->config);
>  		nbd->config = NULL;
>  
> -		if (nbd->recv_workq)
> -			destroy_workqueue(nbd->recv_workq);
> -		nbd->recv_workq = NULL;
> -
>  		nbd->tag_set.timeout = 0;
>  		nbd->disk->queue->limits.discard_granularity = 0;
>  		nbd->disk->queue->limits.discard_alignment = 0;
> @@ -1346,14 +1342,6 @@ static int nbd_start_device(struct nbd_device *nbd)
>  		return -EINVAL;
>  	}
>  
> -	nbd->recv_workq = alloc_workqueue("knbd%d-recv",
> -					  WQ_MEM_RECLAIM | WQ_HIGHPRI |
> -					  WQ_UNBOUND, 0, nbd->index);
> -	if (!nbd->recv_workq) {
> -		dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
> -		return -ENOMEM;
> -	}
> -
>  	blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
>  	nbd->pid = task_pid_nr(current);
>  
> @@ -1779,6 +1767,15 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
>  	}
>  	nbd->disk = disk;
>  
> +	nbd->recv_workq = alloc_workqueue("nbd%d-recv",
> +					  WQ_MEM_RECLAIM | WQ_HIGHPRI |
> +					  WQ_UNBOUND, 0, nbd->index);
> +	if (!nbd->recv_workq) {
> +		dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
> +		err = -ENOMEM;
> +		goto out_err_disk;
> +	}
> +

You never free this up, you need to add a destroy_workqueue(nbd->rsv_workq) to
nbd_dev_remove().

>  	/*
>  	 * Tell the block layer that we are not a rotational device
>  	 */
> @@ -1809,7 +1806,7 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
>  	disk->first_minor = index << part_shift;
>  	if (disk->first_minor > 0xff) {
>  		err = -EINVAL;
> -		goto out_err_disk;
> +		goto out_free_work;
>  	}
>  
>  	disk->minors = 1 << part_shift;
> @@ -1818,7 +1815,7 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
>  	sprintf(disk->disk_name, "nbd%d", index);
>  	err = add_disk(disk);
>  	if (err)
> -		goto out_err_disk;
> +		goto out_free_work;
>  
>  	/*
>  	 * Now publish the device.
> @@ -1827,6 +1824,8 @@ static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
>  	nbd_total_devices++;
>  	return nbd;
>  
> +out_free_work:
> +	destroy_workqueue(nbd->recv_workq);
>  out_err_disk:
>  	blk_cleanup_disk(disk);
>  out_free_idr:
> @@ -2087,8 +2086,7 @@ static void nbd_disconnect_and_put(struct nbd_device *nbd)
>  	 * queue. And this also ensure that we can safely call nbd_clear_que()
>  	 * to cancel the inflight I/Os.
>  	 */
> -	if (nbd->recv_workq)
> -		flush_workqueue(nbd->recv_workq);
> +	flush_workqueue(nbd->recv_workq);

The comment above this part needs to be updated, as we no longer have this
problem.  Thanks,

Josef