[dpdk-dev] [PATCHv3 1/2] config/armv8a: disable igb_uio

[Jianbo Liu]

On 12 May 2016 at 16:57, Santosh Shukla
<santosh.shukla at caviumnetworks.com> wrote:
> On Thu, May 12, 2016 at 01:54:13PM +0800, Jianbo Liu wrote:
>> On 12 May 2016 at 13:06, Santosh Shukla
>> <santosh.shukla at caviumnetworks.com> wrote:
>> > On Thu, May 12, 2016 at 11:42:26AM +0800, Jianbo Liu wrote:
>> >> On 12 May 2016 at 11:17, Santosh Shukla
>> >> <santosh.shukla at caviumnetworks.com> wrote:
>> >> > On Thu, May 12, 2016 at 10:01:05AM +0800, Jianbo Liu wrote:
>> >> >> On 12 May 2016 at 02:25, Stephen Hemminger <stephen at networkplumber.org> wrote:
>> >> >> > On Wed, 11 May 2016 22:32:16 +0530
>> >> >> > Jerin Jacob <jerin.jacob at caviumnetworks.com> wrote:
>> >> >> >
>> >> >> >> On Wed, May 11, 2016 at 08:22:59AM -0700, Stephen Hemminger wrote:
>> >> >> >> > On Wed, 11 May 2016 19:17:58 +0530
>> >> >> >> > Hemant Agrawal <hemant.agrawal at nxp.com> wrote:
>> >> >> >> >
>> >> >> >> > > IGB_UIO not supported for arm64 arch in kernel so disable.
>> >> >> >> > >
>> >> >> >> > > Signed-off-by: Hemant Agrawal <hemant.agrawal at nxp.com>
>> >> >> >> > > Reviewed-by: Santosh Shukla <santosh.shukla at caviumnetworks.com>
>> >> >> >> >
>> >> >> >> > Really, I have use IGB_UIO on ARM64
>> >> >> >>
>> >> >> >> May I know what is the technical use case for igb_uio on arm64
>> >> >> >> which cannot be addressed through vfio or vfioionommu.
>> >> >> >
>> >> >> > I was running on older kernel which did not support vfioionommu mode.
>> >> >>
>> >> >> As I said, most of DPDK developers are not kernel developers. They may
>> >> >> have their own kernel tree, and couldn't like to upgrade to latest
>> >> >> kernel.
>> >> >> They can choose to use or not use igb_uio when binding the driver. But
>> >> >> blindly disabling it in the base config seems unreasonable.
>> >> >
>> >> > if user keeping his own kernel so they could also keep IGB_UIO=y in their local
>> >> Most likely they don't have local dpdk tree. They write their own
>> >> applications, complie and link to dpdk lib, then done.
>> >>
>> >> > dpdk tree. Why are you imposing user-x custome depedancy on upstream dpdk base
>> >> Customer requiremnts is important. I want they can choose the way they like.
>> >>
>> >
>> > so you choose to keep igb_uio option, provided arch doesn't support?
>> > new user did reported issues with igb_uio for arm64, refer this thread [1], as
>> > well hemanth too faced issues. we want to avoid that.
>> >
>> > If customer maintaing out-of-tree kernel then he can also switch to vfio-way.
>> > isn;t it?
>> >
>> >> > config. Is it not enough for explanation that - Base config ie.. armv8 doesn;t
>> >> > support pci mmap, so igb_uio is n/a. New user wont able to build/run dpdk/arm64
>> >> > in igb_uio-way, He'll prefer to use upstream stuff. I think, you are not making
>> >> You are wrong, he can build dpdk. If he like to use upstream without
>> >> patching, he can use vfio.
>> >
>> > I disagree, we want to avoid [1] for new user.
>> >
>> >> But you can't ignore the need from old user which is more comfortable
>> >> with older kernel.
>> >>
>> > arm/arm64 dpdk support recently added and I am guessing, most likely customer
>> > using near latest kernel, switching to vfio won't be so difficult.
>> >
>> > Or can you take up responsibility of upstreaming pci mmap patch, then we don't
>> > need this patch.
>> >
>> > [1] http://dpdk.org/ml/archives/dev/2016-January/031313.html
>>
>> Can you read carefully about the guide at
>> http://dpdk.org/doc/guides/linux_gsg/build_dpdk.html? It says to use
>> uio_pci_generic, igb_uio or vfio-pci.
>
> *** applicable and works for x86 only, not for arm64: because pci mmap support
> not present for arm64, in that case we should update the doc.
>
>> Could it be possible that the user in that thread has already read and
>> tried them all and found that he can't enable vifo with his kernel,
>> and igb_uio is the easy way for him and asked for help from community?
>> If so, we have no choice but keeping igb_uio enabled.
>
> By then vfionoiommu support was wip progress in dpdk/linux. but now it merged
> and it works. So no need to retain igb_uio in base config for which to work -
> user need to use mmap patch at linux side.

We can't decide which kernel user will use.

>
> Or can you maintain out-of-tree pci mmap patch/ kerne source and make it
> explicit somewhere in dpdk build doc that - if user want igb_uio way then
> use kernel/mmap patch from x location.

The patch is in the kernel maillist, and user google it.
And isn't funny to ask someone to do something again and again (3
times) in this thread?

>
>> He use lsmod to show us the modules, most likely he know vifo-pci.
>>
>> Below are the details on modules, hugepages and device binding.
>> root at arm64:~# lsmod
>> Module                  Size  Used by
>> rte_kni               292795  0
>> igb_uio                 4338  0
>> ixgbe                 184456  0