[dpdk-dev] [PATCH] bus/pci: fix vfio mode

[Gaëtan Rivet]

Hi Ferruh,

On Mon, Oct 30, 2017 at 02:00:43AM -0700, Ferruh Yigit wrote:
> On 10/30/2017 1:06 AM, Gaëtan Rivet wrote:
> > Hi Jerin,
> > 
> > On Sat, Oct 28, 2017 at 11:50:52AM +0530, Jerin Jacob wrote:
> >> The definition of VFIO_PRESENT is "eal_vfio.h", Fail to
> >> include eal_vfio.h will result in disabling vfio.
> >>
> >> Fixes: 279b581c897d ("vfio: expose functions")
> >>
> > 
> > Thanks for the fix, sorry for VFIO.
> > I tried to let go of VFIO_PRESENT in the PCI patchset, unfortunately I did
> > not do a good-enough job.
> > 
> > Instead of reinstating the dependency on the private eal_vfio.h header,
> > I'd suggest replacing all VFIO_PRESENT references within the PCI bus by
> > RTE_EAL_VFIO, and make the pci_vfio.c compilation depend on it within
> > the linux Makefile. Something like:
> > 
> > ---8<---
> > 
> > grep -rl VFIO_PRESENT drivers/bus/pci/linux/ |while read -r file
> > do sed -i 's;VFIO_PRESENT;RTE_EAL_VFIO;' $file
> > done
> 
> VFIO_PRESENT is the combination of the if user enabled VFIO and if Linux kernel
> supports it.
> 
> Why not add same check and VFIO_PRESENT definition to rte_vfio.h:
> 
>   --- a/lib/librte_eal/common/include/rte_vfio.h
>   +++ b/lib/librte_eal/common/include/rte_vfio.h
>   @@ -34,7 +34,13 @@
>    #ifndef _RTE_VFIO_H_
>    #define _RTE_VFIO_H_
> 
>   +#if !defined(VFIO_PRESENT) && defined(RTE_EAL_VFIO)
>   +#include <linux/version.h>
>   +#if LINUX_VERSION_CODE >= KERNEL_VERSION(3, 6, 0)
>    #include <linux/vfio.h>
>   +#define VFIO_PRESENT
>   +#endif /* >= 3.6.0 */
>   +#endif
>   +
>   +#ifdef VFIO_PRESENT
> 
>   ... Need to wrap here in case VFIO disabled ...
> 
>   #endif
> 

It would work indeed.

I mostly dislike having whole compilation units disabled silently on a
linux version check. I think that if someone wanted to support kernels <
3.6, they ought to do the work of disabling RTE_EAL_VFIO.

A build error could be thrown to help those users toward the right
solution, but I think that the meaning of having this option enabled
should be enforced: if it is enabled, it is compiled. If dependencies
are not met, then the option should be disabled.

> > 
> > patch -p1 <<___HERE
> > diff --git a/drivers/bus/pci/linux/Makefile b/drivers/bus/pci/linux/Makefile
> > index 77c5f97..b5b9c54 100644
> > --- a/drivers/bus/pci/linux/Makefile
> > +++ b/drivers/bus/pci/linux/Makefile
> > @@ -31,6 +31,8 @@
> > 
> >  SRCS += pci.c
> >  SRCS += pci_uio.c
> > +ifeq (\$(CONFIG_RTE_EAL_VFIO),y)
> >  SRCS += pci_vfio.c
> > +endif
> 
> +1
> 
> Also ?
>   +ifeq ($(CONFIG_RTE_EAL_VFIO),y)
>    SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += eal_vfio.c
>    SRCS-$(CONFIG_RTE_EXEC_ENV_LINUXAPP) += eal_vfio_mp_sync.c
>   +endif
> 
> Build shouldn't broken when VFIO disabled form config
> 
> > 
> >  CFLAGS += -D_GNU_SOURCE
> > ___HERE
> > 
> > --->8---
> > 
> > Do you think it could work? If so, I think it would be preferable
> > to re-including eal_vfio.h. Private EAL headers are bound to be removed
> > from other subsystems at some point. I tried to start with VFIO, others
> > should follow.
> > 
> >> Cc: Gaetan Rivet <gaetan.rivet at 6wind.com>
> >>
> >> Signed-off-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> >> ---
> >>  drivers/bus/pci/linux/pci.c | 1 +
> >>  1 file changed, 1 insertion(+)
> >>
> >> diff --git a/drivers/bus/pci/linux/pci.c b/drivers/bus/pci/linux/pci.c
> >> index cdf810693..d0ce0207a 100644
> >> --- a/drivers/bus/pci/linux/pci.c
> >> +++ b/drivers/bus/pci/linux/pci.c
> >> @@ -46,6 +46,7 @@
> >>  
> >>  #include "eal_private.h"
> >>  #include "eal_filesystem.h"
> >> +#include "eal_vfio.h"
> >>  
> >>  #include "private.h"
> >>  #include "pci_init.h"
> >> -- 
> >> 2.14.3
> >>
> > 
> 

-- 
Gaëtan Rivet
6WIND