Re: [dpdk-dev] [PATCH] bus/vdev: reduce scope of device list lock

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 21-May-18 5:11 PM, Thomas Monjalon wrote:
> The lock vdev_device_list_lock was taken before calling
> "remove" function for the device.
> So it prevents to remove sub-devices (as in failsafe) inside
> its own "remove" function, because of a deadlock.
> 
> The lock is now only protecting the device list inside
> the bus driver.
> 
> Fixes: 35f462839b69 ("bus/vdev: add lock on device list")
> 
> Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> ---
>   drivers/bus/vdev/vdev.c | 10 ++++------
>   1 file changed, 4 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/bus/vdev/vdev.c b/drivers/bus/vdev/vdev.c
> index 099b9ff85..2fbc86806 100644
> --- a/drivers/bus/vdev/vdev.c
> +++ b/drivers/bus/vdev/vdev.c
> @@ -293,25 +293,23 @@ rte_vdev_uninit(const char *name)
>   	if (name == NULL)
>   		return -EINVAL;
>   
> -	rte_spinlock_lock(&vdev_device_list_lock);
> -
>   	dev = find_vdev(name);
>   	if (!dev) {
>   		ret = -ENOENT;
> -		goto unlock;
> +		return ret;
>   	}

Without that lock, all of this would be racy - find_dev would iterate a 
tailq that might change under its feet, and tailq_remove may be called 
with a pointer that has already been removed.

How about changing the lock to a recursive lock? Failsafe would be 
removing devices from within the same thread, correct?

>   
>   	ret = vdev_remove_driver(dev);
>   	if (ret)
> -		goto unlock;
> +		return ret;
>   
> +	rte_spinlock_lock(&vdev_device_list_lock);
>   	TAILQ_REMOVE(&vdev_device_list, dev, next);
>   	devargs = dev->device.devargs;
>   	rte_devargs_remove(devargs->bus->name, devargs->name);
>   	free(dev);
> -
> -unlock:
>   	rte_spinlock_unlock(&vdev_device_list_lock);
> +
>   	return ret;
>   }
>   
> 


-- 
Thanks,
Anatoly