Re: [dpdk-users] [dpdk-dev] IGB_UIO: PCI Resources Management

["Tan, Jianfeng" <jianfeng.tan@intel.com>]

On 1/19/2017 11:59 PM, Ferruh Yigit wrote:
> On 1/13/2017 5:33 AM, Tan, Jianfeng wrote:
>>
>>> -----Original Message-----
>>> From: Yigit, Ferruh
>>> Sent: Friday, January 13, 2017 10:05 AM
>>> To: Tan, Jianfeng; Alejandro Lucero
>>> Cc: Gregory Etelson; dev; users@dpdk.org
>>> Subject: Re: [dpdk-users] [dpdk-dev] IGB_UIO: PCI Resources Management
>>>
>>> On 1/13/2017 1:51 AM, Tan, Jianfeng wrote:
>>>>
>>>>> -----Original Message-----
>>>>> From: users [mailto:users-bounces@dpdk.org] On Behalf Of Ferruh Yigit
>>>>> Sent: Thursday, January 12, 2017 8:22 PM
>>>>> To: Alejandro Lucero
>>>>> Cc: Gregory Etelson; dev; users@dpdk.org
>>>>> Subject: Re: [dpdk-users] [dpdk-dev] IGB_UIO: PCI Resources
>>> Management
>>>>> On 1/12/2017 12:12 PM, Alejandro Lucero wrote:
>>>>>>
>>>>>> On Thu, Jan 12, 2017 at 11:55 AM, Ferruh Yigit <ferruh.yigit@intel.com
>>>>>> <mailto:ferruh.yigit@intel.com>> wrote:
>>>>>>
>>>>>>      On 12/9/2016 8:54 AM, Gregory Etelson wrote:
>>>>>>      > Hello,
>>>>>>      >
>>>>>>      > IGB_UIO driver does not close port PCI activities after DPDK process
>>>>> exits.
>>>>>>      > DPDK API provides rte_eth_dev_close() to manage port PCI,
>>>>>>      > but it can be skipped if process receives SIGKILL signal
>>>>>>
>>>>>>      I guess I understand the problem.
>>>>>>
>>>>>>
>>>>>> This is a known problem, but it is not just a UIO problem, and this
>>>>>> patch does not solve it, maybe it just solves part of it.
>>>>>>
>>>>>> In fact, a DPDK program crashing could imply the NIC DMAing after that
>>>>>> and after that memory was assigned to another program.
>>>>> Yes.
>>>>> Can there be a way to stop NIC DMA, (or prevent it access to mem
>>>>> anymore) when app crashes?
>>>>> I think that is what this patch is looking for.
>>>> If I understand it correctly, you are looking for this patch?
>>>> http://dpdk.org/dev/patchwork/patch/17495/
>>>>
>>> That is good, thanks Jianfeng, I will check it.
>>>
>>> btw, patch's current state is rejected, which is by mistake, it seems I
>>> confused it with "iomem and ioport mapping" patch, sorry about it, I
>>> will update its status immediately.
>> No problem at all. This patch is rejected as it's based on "iomem and ioport mapping" patch. As "iomem and ioport mapping" patch has backward compatibility issue, we need to figure out a way to resubmit this patch without changing the original "iomem and ioport mapping" in igb_uio.
> I thinks implementing uio_info->release and uio_info.open is good idea,
> but I have a few questions:
>
> 1- What is the the dependency to "iomem and ioport mapping" patch?

igb_uio is based on UIO framework to do iomem and ioport mapping, and 
it's done once in probe. If we do pci_disable_device() in release(), 
iomem/ioport mapping will be missing. And I did not figure out a way to 
do the iomem/ioport remapping in open() (may be we can check how 
vfio-pci succeeds to do this).


>
> 2- If we keep pci_enable_device() in probe() can this prevent moving
> registering/freeing interrupts in open()/release()

Yes. But then how can we stop a device in release()?

>
> 3- And is pci_disable_device() done in release is enough to stop NIC DMA
> to access memory?

To find out the answer, we can also use vfio-pci as a reference. Please 
refer to vfio_pci_release().

>
>
> I did a simple test, implemented simple uio_info->release and
> uio_info.open, which only does pci_disable_device() and
> pci_enable_device(),
> but this prevent app receiving packets in its second run, independent
> from app terminated gracefully or not. Any idea why this is not working?

After calling pci_disable_device() in first uio_info->release, 
iomem/ioport remap is missing. So my original idea is to let DPDK 
initialization not depends on this igb_uio's sysfs files, instead use 
the info what pci bus driver provides.

Thanks,
Jianfeng

>
>
> btw, I can produce the problematic case, as George Prekas described in:
> http://dpdk.org/ml/archives/users/2016-September/001026.html
>
> CC'ed George, since he also seems interested in issue.
>
>> Thanks,
>> Jianfeng
>>