On Tue, Aug 2, 2016 at 3:16 PM, Jan Kiszka <jan.kis...@siemens.com> wrote:

> On 2016-08-02 13:58, David Kiarie wrote:
> >
> >
> > On Tue, Aug 2, 2016 at 1:28 PM, Peter Xu <pet...@redhat.com
> > <mailto:pet...@redhat.com>> wrote:
> >
> >     On Tue, Aug 02, 2016 at 10:46:13AM +0200, Jan Kiszka wrote:
> >     > On 2016-08-02 10:36, Peter Xu wrote:
> >     > > On Mon, Aug 01, 2016 at 06:39:05PM +0200, Jan Kiszka wrote:
> >     > >
> >     > > [...]
> >     > >
> >     > >>>  static MemTxResult vtd_mem_ir_read(void *opaque, hwaddr addr,
> >     > >>> @@ -2209,11 +2250,17 @@ static MemTxResult
> >     vtd_mem_ir_write(void *opaque, hwaddr addr,
> >     > >>>  {
> >     > >>>      int ret = 0;
> >     > >>>      MSIMessage from = {}, to = {};
> >     > >>> +    uint16_t sid = X86_IOMMU_SID_INVALID;
> >     > >>>
> >     > >>>      from.address = (uint64_t) addr + VTD_INTERRUPT_ADDR_FIRST;
> >     > >>>      from.data = (uint32_t) value;
> >     > >>>
> >     > >>> -    ret = vtd_interrupt_remap_msi(opaque, &from, &to);
> >     > >>> +    if (!attrs.unspecified) {
> >     > >>> +        /* We have explicit Source ID */
> >     > >>> +        sid = attrs.requester_id;
> >     > >>> +    }
> >     > >>
> >     > >> ...here you fall back to X86_IOMMU_SID_INVALID if writer to
> >     this region
> >     > >> has not provided some valid attrs. That is questionable, defeats
> >     > >> validation of the IOAPIC e.g. (and you can see lots of
> >     > >> X86_IOMMU_SID_INVALID in vtd_irte_get when booting a guest).
> >     > >>
> >     > >> The credits also go to David who noticed that he still doesn't
> >     get a
> >     > >> proper ID from the IOAPIC while implementing AMD IR. Looks like
> >     we need
> >     > >> to enlighten the IOAPIC MSI writes...
> >     > >
> >     > > Jan, David,
> >     > >
> >     > > At the time when drafting the patch, I skipped SID verification
> for
> >     > > IOAPIC interrupts since it differs from generic PCI devices (no
> >     > > natural requester ID, so need some hacky lines to enable it).
> >     >
> >     > It's not hacky at all if done properly. For Intel it is simply
> >     > (Q35_PSEUDO_BUS_PLATFORM << 8) | Q35_PSEUDO_DEVFN_IOAPIC, but it
> >     will be
> >     > 0x00a0 (as constant as well) for AMD. So we need some interface to
> >     tell
> >     > those parameters to the IOMMU. Keep in mind that we will need a
> >     similar
> >     > interface for other platform devices, e.g. the HPET.
> >
> >     Okay.
> >
> >     >
> >     > >
> >     > > I can try to cook another seperate patch to enable it (for 2.8
> >     > > possibly?). Thanks for pointing out this issue.
> >     >
> >     > David needs that IOAPIC ID as well in order to finish interrupt
> >     > remapping on AMD. Please synchronize with him who will implement
> what.
> >
> >     Sure. David, so do you like to do it or I cook this patch? :)
> >
> >
> > If there are no objections I will look at this employing Jan's approach:
> > associating a write with an address space.
> >
>
> That actually means making it an IOMMU-local thing (if write_requester
> == invalid -> derive ID from addressed memory region, probably via the
> opaque passed to the ops). In that case, everyone could do this on himself.
>

Yes, each person can actually do their part without affecting the other ;-)


>
> Jan
>
> --
> Siemens AG, Corporate Technology, CT RDA ITP SES-DE
> Corporate Competence Center Embedded Linux
>

Reply via email to