Re: [dpdk-dev] [PATCH v2] Fixes: ethdev: secondary process change shared memory

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 1/15/20 11:43 PM, Thomas Monjalon wrote:
> 15/01/2020 19:35, Ferruh Yigit:
>> On 1/15/2020 6:49 AM, 方统浩50450 wrote:
>>> Hi Ferruh, thanks for your message.
>>>
>>>
>>> We developed a ethtool-dpdk which is secondary process based dpdk 17.08 version. Our device
>>> support hotplug detach, but hotplug deatch is failed when we use ethtool-dpdk.We found the
>>> secondary process will change the shared memory when initializing.Secondary process calls
>>> "rte_eth_dev_pci_allocate" function and enters "rte_eth_copy_pci_info" function.
>>> (rte_eth_dev_pci_generic_probe -> rte_eth_dev_pci_allocate -> rte_eth_copy_pci_info)
>>> Then it sets the value of struct "rte_eth_dev_data.dev_flags" to zero.In our platform, this value
>>> is equal to 0x0003.(RTE_ETH_DEV_DETACHABLE | RTE_ETH_DEV_INTR_LSC),but after reset
>>> the "dev_flags", the value changed to 0x0002.(RTE_ETH_DEV_DETACHABLE).So, our device hotplug
>>> detach is failed.I found the similar problem in other dpdk version, include dpdk 19.11.Even though
>>> the deivce hotplug detach is discarded,but i think the shared memory changed is unexpected by primary
>>> process.

Hold on, just for my understanding. As far as I can see
RTE_ETH_DEV_DETACHABLE was removed in 17.11. Does it
change something in above description?

>> I agree this is the problem.
>> In the driver code, 'rte_eth_copy_pci_info' is called only by primary process,
>>
>> but the generic code is faulty.
>>
>> And in 19.11 additionally 'eth_dev_pci_specific_init' also seems has same problem.

Yes, as I understand RTE_ETH_DEV_CLOSE_REMOVE,
RTE_ETH_DEV_BONDED_SLAVE, RTE_ETH_DEV_REPRESENTOR and
RTE_ETH_DEV_NOLIVE_MAC_ADDR may be lost because of
reinit (if not restored in other branches). Bad anyway.

>>> Our driver is ixgbe, i think this problem has a little relationship with driver, Secondary process
>>> enters "rte_eth_copy_pci_info" by "rte_eth_dev_pci_allocate".And I agree your opinion, the helper
>>> function should simple on what it does.I have two ways to fix this problem, one is add an if-statement
>>>
>>> in "rte_eth_dev_pci_allocate" function to forbid secondary process enters "rte_eth_copy_pci_info" function,
>>> another way is add an if-statement in "rte_eth_copy_pci_info" function to forbid secondary process change
>>> shared memory.And First way need to ensure the "rte_eth_copy_pci_info" function won't be called anywhere else.
>>> I think the second way is simple and lower risk.
>>
>> Yes these are the two options.
>>
>> I agree adding check in the 'rte_eth_copy_pci_info' covers all cases and safer.
>> BUT my concern was adding decision making to simple/leaf function and make it
>> harder to debug/use, instead of giving what primary/secondary process should
>> call decision in higher level.
>>
>> But I just recognized that some PMDs are calling 'rte_eth_copy_pci_info' on
>> secondary process, like mlx4 or szedata2, and most probably this is not their
>> intention.
>> And 'eth_dev->intr_handle' set in 'rte_eth_copy_pci_info', not calling this
>> function may have side affect of 'eth_dev->intr_handle' not set in secondary.
>>
>> With above considerations I am OK to your proposal to cover all cases, Thomas,
>> Andrew, any concern?

I would put if condition in rte_eth_copy_pci_info().
It is the function which writes shared space from
secondary process when it should not be done and it
should be fixed there.

> Do you mean drivers need to be fixed?

I'm not sure that I fully understand it. Since copy function
cares about intr_handle copying I'm afraid that it is not
100% correct to skip it in secondary process completely as
many drivers do right now. Basically it makes eth_dev structure
in secondary process inconsistent. However, it looks like
most of these drivers simply obtain handle from pci_dev
directly and it explains why they are not affected.
There are exceptions which are potentially bugs, e.g.
drivers/net/ice/ice_ethdev.c: ice_interrupt_handler at the end.

I think that it would be better if intr_handle is always
correct in eth_dev (both primary and secondary cases) and
drivers use it instead of the same from pci_dev.