Re: [PATCH] usertools: fix bind failure from dpdk to kernel

[Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>]

On 09/08/2022 14:44, lihuisong (C) wrote:
> 
> 在 2022/8/5 23:35, Stephen Hemminger 写道:
>> On Fri, 5 Aug 2022 11:10:22 +0800
>> Huisong Li <lihuisong@huawei.com> wrote:
>>
>>> Currently, the steps for binding device from dpdk driver to kernel
>>> driver is as follows:
>>> echo $BDF > /sys/bus/pci/drivers/vfio-pci/unbind
>>> echo $BDF > /sys/bus/pci/drivers/$kernel_driver/bind
>>>
>>> This steps cannot bind device from dpdk driver to kernel driver on
>>> platform with kernel 5.19. The 'driver_override' must be specify
>>> kernel driver before binding device to kernel driver.
>>>
>>> Fixes: 720b7a058260 ("usertools: fix device binding with kernel tools")
>>> Cc: stable@dpdk.org
>>>
>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>> Not sure exactly what you did and why.
>> The patch seems to just remove the check that the driver
>> is in the set of dpdk_drivers.
>> .
> Currently, the end of the operation binding device from kernel driver to
> dpdk driver write '\00' to driver_override file so as to this device can
> be bound to any other driver. 

This could have work but this was not the way to use the
driver_override. The kernel ABI document clearly states:
"and  may be cleared with an empty string (echo > driver_override)."
Documentation/ABI/testing/sysfs-bus-pci

Please use the kernel ABI how it is described. Using it in wrong way
might sometimes work, sometimes not.


> And perform following steps to
> bind device dpdk driver to kernel driver:
> echo $BDF > /sys/bus/pci/drivers/vfio-pci/unbind
> echo $BDF > /sys/bus/pci/drivers/$kernel_driver/bind
> 
> However, due to the patch[1] merged into 5.19 kernel, 'driver_override'
> in the pci_dev is no longer NULL by writing '\00' to driver_override file.
> This causes PCI match device failure and the device will never be bound to
> their kernel driver.
> 
> In 5.19 kernel, I found that dpdk-devbind.py need to write '\n' to
> driver_override file if we want to bind divce to any other driver.
> But I think it is not necessary to write empty to driver_override
> file. 

It is necessary because in 2014 it was described that PCI
driver_override works like that. What you are implying here is that "it
is not necessary to follow the API and we can do it differently"...

> After all, the device has only one kernel driver, and binding
> to dpdk driver(like, vfio-pci) must specify driver_override.
> 
> [1] 23d99baf9d72 ("PCI: Use driver_set_override() instead of open-coding")


Best regards,
Krzysztof