[dpdk-dev] [PATCH 3/3] app/testpmd: fix callback issue for hot-unplug

Jeff Guo jia.guo at intel.com
Wed Nov 14 10:32:11 CET 2018


On 11/12/2018 9:35 AM, Thomas Monjalon wrote:
> 11/11/2018 08:31, Matan Azrad:
>> From: Jeff Guo
>>> On 11/9/2018 1:24 PM, Matan Azrad wrote:
>>>>    From: Jeff Guo
>>>>> On 11/8/2018 5:35 PM, Matan Azrad wrote:
>>>>>> From: Jeff Guo
>>>>>>> On 11/8/2018 3:28 PM, Matan Azrad wrote:
>>>>>>>> From: Ananyev, Konstantin
>>>>>>>>> From: Guo, Jia
>>>>>>>>>> On 11/6/2018 2:36 PM, Matan Azrad wrote:
>>>>>>>>>>> Hi Jeff
>>>>>>>>>>>
>>>>>>>>>>>       From: Jeff Guo <jia.guo at intel.com>
>>>>>>>>>>>> Before detach device when device be hot-unplugged, the failure
>>>>>>>>>>>> process in user space and kernel space both need to be
>>>>>>>>>>>> finished, such as eal interrupt callback need to be inactive
>>>>>>>>>>>> before the callback be unregistered when device is being
>>>>>>>>>>>> cleaned. This patch add rte alarm for device detaching, with
>>>>>>>>>>>> that it could finish interrupt callback soon and give time to
>>>>>>>>>>>> let the failure process done
>>>>>>>>> before detaching.
>>>>>>>>>>>> Fixes: 2049c5113fe8 ("app/testpmd: use hotplug failure
>>>>>>>>>>>> handler")
>>>>>>>>>>>> Signed-off-by: Jeff Guo <jia.guo at intel.com>
>>>>>>>>>>>> ---
>>>>>>>>>>>>       app/test-pmd/testpmd.c | 13 ++++++++++++-
>>>>>>>>>>>>       1 file changed, 12 insertions(+), 1 deletion(-)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
>>>>>>>>>>>> index 9c0edca..9c673cf 100644
>>>>>>>>>>>> --- a/app/test-pmd/testpmd.c
>>>>>>>>>>>> +++ b/app/test-pmd/testpmd.c
>>>>>>>>>>>> @@ -2620,7 +2620,18 @@ eth_dev_event_callback(const char
>>>>>>>>>>>> *device_name, enum rte_dev_event_type type,
>>>>>>>>>>>>       				device_name);
>>>>>>>>>>>>       			return;
>>>>>>>>>>>>       		}
>>>>>>>>>>>> -		rmv_event_callback((void *)(intptr_t)port_id);
>>>>>>>>>>>> +		/*
>>>>>>>>>>>> +		 * Before detach device, the hot-unplug failure
>>>>>>> process in
>>>>>>>>>>>> +		 * user space and kernel space both need to be
>>>>>>> finished,
>>>>>>>>>>>> +		 * such as eal interrupt callback need to be inactive
>>>>>>> before
>>>>>>>>>>>> +		 * the callback be unregistered when device is being
>>>>>>> cleaned.
>>>>>>>>>>>> +		 * So finished interrupt callback soon here and give
>>>>>>> time to
>>>>>>>>>>>> +		 * let the work done before detaching.
>>>>>>>>>>>> +		 */
>>>>>>>>>>>> +		if (rte_eal_alarm_set(100000,
>>>>>>>>>>>> +				rmv_event_callback, (void
>>>>>>>>>>>> *)(intptr_t)port_id))
>>>>>>>>>>>> +			RTE_LOG(ERR, EAL,
>>>>>>>>>>>> +				"Could not set up deferred device
>>>>>>>>>>> It looks me strange to use callback and alarm to remove a device:
>>>>>>>>>>> Why not to use callback and that is it?
>>>>>>>>>>>
>>>>>>>>>>> I think that it's better to let the EAL to detach the device
>>>>>>>>>>> after all the
>>>>>>>>> callbacks were done and not to do it by the user callback.
>>>>>>>>>>> So the application\callback owners just need to clean its
>>>>>>>>>>> resources with understanding that after the callback the
>>>>>>>>>>> device(and the callback
>>>>>>>>>> itself) will be detached by the EAL.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Firstly, at the currently framework and solution, such as
>>>>>>>>>> callback for RTE_ETH_EVENT_INTR_RMV, still need to use the
>>>>>>>>>> deferred device
>>>>>>>>> removal,
>>>>>>>>>> we tend to give the control of detaching device to the
>>>>>>>>>> application, and the whole process is located on the user's
>>>>>>>>>> callback. Notify app to detach device by callback but make it
>>>>>>>>>> deferred,
>>>>> i think it is fine.
>>>>>>>> But the device must be detached in remove event, why not to do it
>>>>>>>> in
>>>>> EAL?
>>>>>>> I think it because of before detached the device, application
>>>>>>> should be stop the forwarding at first, then stop the device, then
>>>>>>> close
>>>>>>>
>>>>>>> the device, finally call eal unplug API to detach device. If eal
>>>>>>> directly detach device at the first step, there will be mountain
>>>>>>> user space error need to handle, so that is one reason why need to
>>>>>>> provider the remove notification to app, and let app to process it.
>>>>>> This is why the EAL need to detach the device only after all the
>>>>>> user
>>>>> callbacks were done.
>>>>>
>>>>>
>>>>> If i correctly got your meaning, you suppose to let eal to mandatory
>>>>> detach device but not app, app just need to stop/close port, right?
>>>> Yes, the app should stop,close,clean its own resources of the removed
>>>> device, Then, EAL to detach the device.
>>>>
>>>>> If so, it will need to modify rmv_event_callback by not invoke the
>>>>> detaching func and add some detaching logic to hotplug framework in eal.
>>>>>
>>>> rmv_event_callback is using by other hotplug mechanism (ETHDEV RMV
>>> event), so you need to use another one\ add parameter to it.
>>>> And yes, you need to detach the device from EAL, should be simple.
>>>
>>> I think rmv_event_callback is original use for other hotplug event (ETHDEV
>>> RMV event), but it still use the common hotplug mechanism(app callback and
>>> app detach),
>>>
>>> so i think it will still need to face this callback issue and you could check that
>>> eth_event_callback also use the rte alarm to detach device.
>> The ETHDEV layer cannot know if more than one ethdev port is associated to the rte_device,
>> So, it is not correct to detach a rte_device by the ETHDEV layer. Hence, the application should decide in the ETHDEV event case.
> Yes, we must not detach a device (rte_dev_remove) if we are not sure all ports are closed.
>
>> But, in the EAL event case this is different as I explained before.
>>
>> Moreover, I think that the ethdev RMV event should be deprecated one day, after the EAL hot-plug mechanism will be stable.
> Yes, we may replace the ethdev RMV event by the EAL event RTE_DEV_EVENT_REMOVE.
>
>>> so my suggestion is that, you maybe propose a good idea but let we keep on
>>> current mechanism until we come to a final good solution agreement, before
>>> that, just let it functional.
>> I still suggest to fix the issue by an EAL detaching.
> I don't understand the issue, but yes we can call rte_dev_remove in EAL,
> after the callback return.
> Ideally, the callback should be able to return the decision of cleaning
> the device or not.
>
> I suggest several steps (a roadmap for HW unplug):
>
>

I mainly agree on that we need to modify currently hotplug framework for 
HW hot-unplug. But i think we might be think about that.

1) simple delete detach_port_device from rmv_event_callback will affect 
both currently ethdev event and eal event hot-unplug mechanism.

2) If we remove call to rmv_event_callback in eth_dev_event_callback() , 
we also need to remove call to rmv_event_callback in 
eth_event_callback(),  since it also use deferred device removal to 
waiting the interrupt callback be unregistered in 
mlx5_dev_interrupt_handler_uninstall.

so i think that basically the multiple port free and the hotplug 
framework integrate should be an RFC at the begin of 19.02, but should 
not affect currently mechanism. And i suggest the several steps might be 
better as.

	- in 18.11, rename eth_dev_event_callback to dev_event_callback
	- in 18.11, rename rmv_event_callback() to rmv_port_callback(), keep current function calling when detach one port both for ethdev event and eal event. (we could document it for the limitation.)

	- in 19.02, remove call to detach_port_device() from rmv_event_callback()
	- in 19.02, call rte_dev_remove() in EAL after RTE_DEV_EVENT_REMOVE callback
	- in 19.02, remove call to rmv_event_callback() in eth_dev_event_callback()
	- in 19.02, convert all PMDs to free port resources on rte_eth_dev_close()
	- in 19.05, automatically call rte_dev_remove() from PMD when closing last port
	- in 19.08, we may try to unbind the kernel driver in rte_dev_remove()




More information about the dev mailing list