Re: [dpdk-dev] VFIO no-iommu

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 17 Dec 2015 20:38:16 +0100
Jan Viktorin <viktorin@rehivetech.com> wrote:

> On Thu, 17 Dec 2015 11:09:23 +0100
> Thomas Monjalon <thomas.monjalon@6wind.com> wrote:
> 
> > Hi,
> > 
> > 2015-12-17 09:52, Burakov, Anatoly:
> > > > >  > > On Tue, Dec 15, 2015 at 09:53:18AM -0700, Alex Williamson wrote:
> > > > > > > So it works.  Is it acceptable?  Useful?  Sufficiently complete?
> > > > > > > Does it imply deprecating the uio interface?  I believe the
> > > > > > > feature that started this discussion was support for MSI/X
> > > > > > > interrupts so that VFs can support some kind of interrupt (uio
> > > > > > > only supports INTx since it doesn't allow DMA).  Implementing that
> > > > > > > would be the ultimate test of whether this provides dpdk with not
> > > > > > > only a more consistent interface, but the feature dpdk wants
> > > > > > > that's missing in uio. Thanks,  
> > > > >
> > > > > Ferruh has done a great job so far testing Alex's patch, very few changes  
> > > > from DPDK side seem to be required as far as existing functionality goes (not
> > > > sure about VF interrupts mentioned by Alex). However, one thing that
> > > > concerns me is usability. While it is true that no-IOMMU mode in VFIO would
> > > > mean uio interfaces could be deprecated in time, the no-iommu mode is way
> > > > more hassle than using igb_uio/uio_pci_generic because it will require a
> > > > kernel recompile as opposed to simply compiling and insmod'ding an out-of-
> > > > tree driver. So, in essence, if you don't want an IOMMU, it's becoming that
> > > > much harder to use DPDK. Would that be something DPDK is willing to live
> > > > with in the absence of uio interfaces?
> > > > 
> > > > Excuse me if I missed something obvious.
> > > > Why a kernel compilation is needed?  
> > > 
> > > Well, not really full kernel compilation, but in the default configuration, VFIO driver would not support NOIOMMU mode. I.e. it's not compiled by default. Support for no-iommu should be enabled in kernel config and compiled in. So, whoever is going to use DPDK with VFIO-no-iommu will have to download kernel tree and recompile the VFIO module and install it. That's obviously way more hassle than simply compiling an out-of-tree driver that's already included and works with an out-of-the-box kernel.  
> > 
> > The "out-of-the-box kernel" is configured by your distribution.
> > So we don't know yet what will be their choice.
> > If the distribution supports DPDK, it should be enabled.
> 
> I have a question as I am not involved in all possible DPDK
> configurations, platforms, etc. and not yet very involved in vfio. What
> are the devices which do not have IOMMU? If I have, say, DPDK 2.3 with
> vfio-noiommu, which platforms (or computer systems) I am targeting?
> 
> Would it be an Intel-based system? Would it be PPC8, ARM?
> 
> If it is ARMv7... I would say that the fact I have to explicitly enable
> the no-IOMMU feature and rebuild the kernel (or whatever) is just OK. As
> for such systems, it is common to have a quite customized OS. Well,
> the big distributions are able to run on those devices, that's true...
> However, in such case, the users are usually skilled enough to take
> care of having their own special Linux kernel.
> 
> So, is the fact the distributions would not support the no-IOMMU setup
> in their default configuration really an issue? Will some very common
> Intel/DPDK-based box need this?
> 
> Regards
> Jan

So far:
  * broken hardware (many systems including those from Dell) do not provide
    working IOMMU because of Bios bugs etc.
  * Linux guest in VMware/KVM/Hyper-V. There is no IOMMU emulation in most
    of these systems.
  * Older smaller systems (ie Atom) may not have IOMMU