On Mon, Sep 19, 2016 at 01:41:47PM +0100, Robin Murphy wrote:
> On 19/09/16 13:24, Will Deacon wrote:
> > On Mon, Sep 19, 2016 at 02:13:45PM +0200, Auger Eric wrote:
> >> On 16/09/2016 18:18, Robin Murphy wrote:
> >>> What I probably will do, though, since we have the functionality in
> >>> place for the sake of the old binding, and I think there are other folks
> >>> who want PCI iommu-map support but would prefer not to bother with DMA
> >>> ops on the host, is add a command-line option to disable DMA domains
> >>> even for the generic bindings.
> >>
> >> Yes this would be a good thing I think. This series has an important
> >> impact on platforms which do not have smmu v3, where contexts are scarce
> >> HW resources.
> > 
> > Rather than disabling DMA domains entirely, we could specify a number
> > of contexts to reserve for other use (e.g. VFIO). It's a pity that these
> > options are global for the system, as opposed to per SMMU instance,
> > but I can't see a good way around that.
> The problem with that approach is that due to bus traversal order you'd
> typically end up with the even-more-limited number of non-reserved
> contexts getting consumed by bridges that are unlikely to ever actually
> use their DMA ops, so you end up barely any better off than just not
> doing DMA ops at all, which we already have the code for.

For Seattle, perhaps, but that's quite a big generalisation. If we're
going to add an option, I'd much rather add something with some flexibility,
rather than end up supporting both "disable_dma_domains" and
"reserved_s2_contexts" in the long run.

> And yeah, if you had, say, "arm-smmu.reserved_s2_contexts=4", for the
> benefit of your PCI SMMU, what do the SMMUs in front of your other
> platform devices with only 2 contexts (and which only want DMA ops) do?

There's two things in this example:

  (1) If other SMMUs only have 2 contexts, then they reserve both and warn
  (2) If some SMMUs want DMA ops and others don't, then you have an issue
      with *any* sort of cmdline option to control this

iommu mailing list

Reply via email to