Re: [PATCH 1/2] iommu/omap: Reverse dependency between omap-iommu and omap-iommu2

[Laurent Pinchart <laurent.pinchart-ryLnwIuWjnjg/C1BVhZhaw@public.gmane.org>]

Hi Suman,

On Tuesday 09 September 2014 17:31:44 Suman Anna wrote:
> > On Tuesday 09 September 2014 16:33:11 Suman Anna wrote:
> >> On 09/09/2014 10:45 AM, Laurent Pinchart wrote:
> >>> The OMAP IOMMU driver supports both the OMAP1 and OMAP2+ IOMMU variants
> >>> by splitting the driver into a core module and a thin arch-specific
> >>> operations module.
> >>> 
> >>> (In practice only the OMAP2+ module omap-iommu2 is implemented, but
> >>> let's not denigrate the effort.)
> >> 
> >> Thank you for the patch. I had something similar in my list of cleanup
> >> TODO items on the OMAP IOMMU driver, but I was intending to cut down
> >> more and consolidate the omap-iommu.c and omap-iommu2.c files into a
> >> single one.
> >> 
> >> This had been the case from the introduction of the driver going back to
> >> v2.6.30, and OMAP1 was never added and I doubt it would be added anytime
> >> in the foreseeable future.
> >> 
> >>> The arch-specific operations module registers itself with the OMAP IOMMU
> >>> core module at initialization time. This initializes a module global
> >>> arch-specific operations pointer, used at runtime by the IOMMU
> >>> instances.
> >>> 
> >>> This scheme causes several issues. In addition to making it impossible
> >>> to support different OMAP IOMMU types in a single system (which in all
> >>> fairness is quite unlikely to happen),
> >> 
> >> Yep, except for a few enhancements (like reporting Fault PC address on
> >> OMAP4 DSPs, and dropping both endianness support), the core IOMMU
> >> functionality hasn't changed much between OMAP2 and the latest OMAP4+
> >> SoCs. So, my plan was to completely get rid of the iommu_functions (it
> >> also eliminates the need for exporting most of the OMAP IOMMU API). So
> >> while I am ok with the current patch, I prefer consolidation than
> >> keeping the scalability alive, it can always be added if a need for that
> >> arises. What do you think?
> > 
> > I agree with your approach, but in the meantime we have a problem to
> > solve.
> > How about applying this patch now (it goes in the right direction anyway),
> > and then removing the iommu functions when you will have time ?
> 
> Can you give the subsys_initcall solution a try to see if that resolves
> the problem at hand? That would be a much smaller change, if that
> doesn't work we can go with this patch.

It seems to work.

> I will work on my cleanup list for 3.19.

Does that schedule still hold ? If so I'll submit a simple subsys_initcall() 
patch for v3.18.

> >>  it also causes initialization
> >>  
> >>> ordering issues by requiring the arch-specific operations module to be
> >>> loaded before any IOMMU user. This results in a probe breakage with the
> >>> OMAP3 ISP driver when not compiled as a module.
> >> 
> >> This can be fixed if we make the current omap-iommu2.c as a
> >> subsys_initcall as well, right?
> >> 
> >> regards
> >> Suman
> >> 
> >>> Fix the problem by inverting the dependency. Instead of having the
> >>> omap-iommu2 module register itself to iommu-omap, make the iommu-omap
> >>> retrieve the omap-iommu2 operations structure directly when probing the
> >>> IOMMU device. This ensures that a probed IOMMU will always have valid
> >>> arch-specific operations.
> >>> 
> >>> As the arch-specific operations pointer is now initialized at probe
> >>> time, this change requires turning it from a global variable into a
> >>> per-device variable.
> >>> 
> >>> Signed-off-by: Laurent Pinchart <laurent.pinchart-ryLnwIuWjnjg/C1BVhZhaw@public.gmane.org>
> >>> ---
> >>> 
> >>>  drivers/iommu/omap-iommu-debug.c |  6 ++-
> >>>  drivers/iommu/omap-iommu.c       | 94 +++++++++++++--------------------
> >>>  drivers/iommu/omap-iommu.h       | 10 ++++-
> >>>  drivers/iommu/omap-iommu2.c      | 18 +-------
> >>>  4 files changed, 45 insertions(+), 83 deletions(-)

-- 
Regards,

Laurent Pinchart