RE: [PATCH 8/9] dspbridge: add map support for big buffers

["Guzman Lugo, Fernando" <fernando.lugo@ti.com>]

Hi Hari,

> -----Original Message-----
> From: Kanigeri, Hari
> Sent: Friday, July 02, 2010 12:03 PM
> To: Guzman Lugo, Fernando; linux-omap@vger.kernel.org; linux-
> kernel@vger.kernel.org
> Cc: ohad@wizery.com; hiroshi.doyu@nokia.com; ameya.palande@nokia.com;
> felipe.contreras@nokia.com
> Subject: RE: [PATCH 8/9] dspbridge: add map support for big buffers
> 
> Fernando,
> 
> > -----Original Message-----
> > From: Guzman Lugo, Fernando
> > Sent: Friday, July 02, 2010 11:27 AM
> > To: Kanigeri, Hari; linux-omap@vger.kernel.org; linux-
> > kernel@vger.kernel.org
> > Cc: ohad@wizery.com; hiroshi.doyu@nokia.com; ameya.palande@nokia.com;
> > felipe.contreras@nokia.com
> > Subject: RE: [PATCH 8/9] dspbridge: add map support for big buffers
> >
> >
> >
> > Hi Hari,
> >
> > > -----Original Message-----
> > > From: Kanigeri, Hari
> > > Sent: Thursday, July 01, 2010 6:36 PM
> > > To: Guzman Lugo, Fernando; linux-omap@vger.kernel.org; linux-
> > > kernel@vger.kernel.org
> > > Cc: ohad@wizery.com; hiroshi.doyu@nokia.com; ameya.palande@nokia.com;
> > > felipe.contreras@nokia.com; Guzman Lugo, Fernando
> > > Subject: RE: [PATCH 8/9] dspbridge: add map support for big buffers
> > >
> > > Fernando,
> > >
> > > > -	for_each_sg(sgt->sgl, sg, sgt->nents, i)
> > > > -		sg_set_page(sg, usr_pgs[i], PAGE_SIZE, 0);
> > > > +		da = iommu_vmap(mmu, da, sgt, IOVMF_ENDIAN_LITTLE |
> > > > +							IOVMF_ELSZ_32);
> > >
> > > -- iommu_vmap does the Kernel mapping to the buffers you are mapping
> to
> > > DSP MMU. Why do you need Kernel mappings ?
> > >
> > > If there is no benefit in maintaining Kernel mapping I would rather
> call
> > > iopgtable_store_entry directly to map the entries.
> >
> > Where inside iommu_vmap is the mapping done?
> 
> -- The mapping is done to track down the Device mappings. But since you
> already have it in dmm.c this is kind of redundant right now, and we might
> see performance impact due to this.
> 
> I think it might be good to transition to iovmm when we phase out dmm.c.

it could be a good time now, I think.

> Few things to take into account transitioning to iovmm approach:
> 	- DSPBridge used to have linked list approach to track down the
> mapped entries and profiling showed it took considerable amount of
> traversing through the list. Jeff Taylor's algorithm in dmm.c helped to
> reduce this impact.

That can be implemented in the iovmmu without too many changes, right?

> 	- How would you manage the Device virtual pool moving to iovmm ? And
> how about the reservation ?

I think, it would be good if we get rid of DMMPOOL size, if the liked list grow up as it is needed, there is no memory penalty of have all the possible iommu addresses valid (11000000 - FFFFFFFF). The reservation will only fail when there is no memory. If a software restriction is needed we could define a start and end addresses for iommu module (maybe as a parameter when the iommu handle for iva2 is got) and that boundaries can be taking in account at the moment of reserve the memory.


I think the reserve/unreserved dspbridge api can disappear and just return the da address in the map function.


Please let me know what you think.


Thanks for the comments,
Fernando.


> 
> Thank you,
> Best regards,
> Hari