Re: [PATCH v2 1/2] igb_uio: compatible with upstream longterm kernel and RHEL6

[Jincheng Miao <jmiao-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On 11/29/2014 12:42 AM, Thomas Monjalon wrote:
> 2014-11-28 16:13, Jincheng Miao:
>> On 11/28/2014 01:01 AM, Thomas Monjalon wrote:
>>> 2014-10-31 15:37, Jincheng Miao:
>>>> Function pci_num_vf() is introduced from upstream linux-2.6.34. So
>>>> this patch make compatible with longterm kernel linux-2.6.32.63.
>>>>
>>>> For RHEL6's kernel, although it is based on linux-2.6.32, it has
>>>> pci_num_vf() implementation. As the same with commit 11ba0426,
>>>> pci_num_vf() is defined from RHEL6. So we should check the macro
>>>> RHEL_RELEASE_CODE to consider this situation.
>>> Please, could you explain in which case CONFIG_PCI_IOV is defined?
>>> The logic is a bit difficult to understand.
>> Yep, there is a little confusion for pci_num_vf():
>> 1. it is available when CONFIG_PCI_IOV is defined.
>> 2. it is introduced from upstream kernel v2.6.34 (fb8a0d9)
>> 3. it is implemented from RHEL6.0, although the kernel version is 2.6.32.
> Sorry, you didn't described when CONFIG_PCI_IOV is defined.
> Is it defined since 2.6.34 upstream? In lower stable versions?
> Is it defined since RHEL 6.0?
> Why checking CONFIG_PCI_IOV is not sufficient?
>
> When pci_num_vf will be backported in other distributions, we will have to
> tune this check and clearly understand what was the situation.

I am not the expert on this, the only thing I know is:
CONFIG_PCI_IOV is config switch for the I/O device virtualization, it will
enable some SRIOV feature for PCI devices supports IO virtualization.
It is defined since upstream kernel v2.6.30 (d1b054d).

But as you said, function pci_num_vf() is stub-defined in kernel when
CONFIG_PCI_IOV disabled. So checking CONFIG_PCI_IOV is meaningless
for pci_num_vf().

For the function pci_num_vf(), I wrote it is defined since RHEL6.0.
But after some searching, I found the earliest is RHEL5 Update9
kernel-devel-2.6.18-348.el5.x86_64.rpm.

I think it should be modified to:
```
  #if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 34) && \
-    !defined(CONFIG_PCI_IOV)
+       (!(defined(RHEL_RELEASE_CODE) && \
+          RHEL_RELEASE_CODE >= RHEL_RELEASE_VERSION(5, 9)))
```

After that, if there is other distro need to be compatible,
we could simply add condition bellow:
(!(defined(OTHER_RELEASE_CODE) && OTHER_RELEASE_CODE >= 
OTHER_RELEASE_VERSION(X, Y)))


Thanks for your patient for pointing out the issue of my patch.
Jincheng Miao

>> The logic of this patch is:
>> #if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 34) && \
>> (!(defined(RHEL_RELEASE_CODE) && RHEL_RELEASE_CODE >=
>> RHEL_RELEASE_VERSION(6, 0) && defined(CONFIG_PCI_IOV)))
>>
>> Firstly it detects kernel version, if it is less than 2.6.34, and it is
>> not RHEL-specified, then define pci_num_vf().
>>
>> Secondly, it deals with RHEL-specified. If it is RHEL6.0 or later, and
>> CONFIG_PCI_IOV is defined. we should not define pci_num_vf(). If any of
>> these conditions is not reached, pci_num_vf() should be defined.
> I can read the check but I don't know why CONFIG_PCI_IOV is checked in the
> RHEL case.
>
>> Some days ago, I setup dpdk for longterm kernel 2.6.32.63, and got error:
>> ```
>> CC [M]
>> /root/dpdk-source/build/build/lib/librte_eal/linuxapp/igb_uio/igb_uio.o
>> /root/dpdk-source/build/build/lib/librte_eal/linuxapp/igb_uio/igb_uio.c:
>> In function ‘show_max_vfs’:
>> /root/dpdk-source/build/build/lib/librte_eal/linuxapp/igb_uio/igb_uio.c:75:
>> error: implicit declaration of function ‘pci_num_vf’
>> ```
> Thank you. Describing the problem is helpful for the commit log.
>   
>> This problem is introduced by commit 11ba04265
>>
>> commit 11ba04265cfd2a53c12c030fcaa5dfe7eed39a42
>> Author: Guillaume Gaudonville<guillaume.gaudonville-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>
>> Date: Wed Sep 3 10:18:23 2014 +0200
>>
>> igb_uio: fix build on RHEL 6.3
>>
>> - pci_num_vf() is already defined in RHEL 6
>> - pci_intx_mask_supported is already defined in RHEL 6.3
>> - pci_check_and_mask_intx is already defined in RHEL 6.3
>>
>> Signed-off-by: Guillaume Gaudonville<guillaume.gaudonville-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>
>> Signed-off-by: David Marchand<david.marchand-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>
>> Signed-off-by: Thomas Monjalon<thomas.monjalon-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>
>>
>> +#if LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 34) && \
>> + !defined(CONFIG_PCI_IOV)
>>
>> That is because longterm kernel 2.6.32.63 defined CONFIG_PCI_IOV, but it
>> lacks pci_num_vf(),
>> after above processing, pci_num_vf() is still not existed, then build fail.
>>
>> My patch could work around it, and can deal with RHEL-specified kernel.
> Thanks, we just need to understand the matrix of combinations to be sure
> it will be well maintained.
>