Re: [dpdk-dev] [PATCH v2] app/testpmd: fix pmd_test_exit function for vdevs

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 5/21/2018 11:54 AM, Thomas Monjalon wrote:
> 19/05/2018 16:19, Thomas Monjalon:
>> 18/05/2018 18:29, Ferruh Yigit:
>>> On 5/18/2018 4:55 PM, Matan Azrad wrote:
>>>> Hi all
>>>>
>>>> While this patch also applied I don't understand it.
>>>> Is it mandatory for each PMD to free all its resources in dev_close()?
>>>> Or it should be done by the rte_device remove function?
>>>>
>>>> If the resource cleanup should be done by the remove function I think it
>>>> should be called for all the devices (pci, vdev, etc).
>>>>
>>>> Is there an exit function for EAL to clean rte_eal_init()? If no, looks like we need it...
>>>
>>> Hi Matan,
>>>
>>> I believe there is a gap in resource cleanup.
>>> dev_close() it not for resource cleanup, it should be in PMD remove() functions,
>>> and PMDs have it. The problem is remove path is not called in application exit.
>>>
>>> As far as I know there is no simple API to clean the resources, having it may
>>> help application to do the cleanup.
>>>
>>> I have seen the rte_eal_cleanup() API by Harry, that can be extended to cover
>>> PMD resource cleanup if there is enough motivation for it.
>>
>> Yes, EAL resources should be removed by the function rte_eal_cleanup().
>> And ethdev ports must be removed by rte_eth_dev_close().
> 
> This patch is relying on rte_eth_dev_detach() to remove the EAL device.
> It should be done in rte_eal_cleanup().
> 
> I am concerned that this patch is workarounding a miss in rte_eal_cleanup,
> and takes a different action only for vdev. It is a bad example.

Indeed it does workaround, but it is needed to fix a defect in virtio-user.

And currently rte_eal_cleanup() is not complete, it is not doing any device
related cleanup.

> 
> And the function rte_eth_dev_detach() is fundamentally wrong and should be deprecated:
> 	http://dpdk.org/commit/b05b444d22
> 	http://dpdk.org/commit/b0fb266855
> 	http://dpdk.org/commit/df3e8ad73f
> 
> One more concern: it seems this patch is breaking failsafe use case.
> Note: bonding is managed as an exception in rte_eth_dev_detach().
> 
>