pci overmapping

[Sethi Varun-B16395]

> -----Original Message-----
> From: Wood Scott-B07421
> Sent: Thursday, May 02, 2013 11:50 PM
> To: Sethi Varun-B16395
> Cc: Wood Scott-B07421; Yoder Stuart-B08248; galak at kernel.crashing.org;
> linuxppc-dev at lists.ozlabs.org
> Subject: Re: pci overmapping
> 
> On 05/02/2013 01:09:53 PM, Sethi Varun-B16395 wrote:
> >
> >
> > > -----Original Message-----
> > > From: Wood Scott-B07421
> > > Sent: Thursday, May 02, 2013 10:44 PM
> > > To: Yoder Stuart-B08248
> > > Cc: galak at kernel.crashing.org; Sethi Varun-B16395; linuxppc-
> > > dev at lists.ozlabs.org
> > > Subject: Re: pci overmapping
> > >
> > > On 05/02/2013 12:05:42 PM, Yoder Stuart-B08248 wrote:
> > > > Kumar,
> > > >
> > > > In fsl_pci.c there is a change you made a while back:
> > > >   "powerpc/fsl: Setup PCI inbound window based on actual amount of
> > > > memory"
> > > >
> > > > ...and there is this comment in the code:
> > > >
> > > >     /* PCIe can overmap inbound & outbound since RX & TX are
> > separated
> > > > */
> > > >     if (early_find_capability(hose, 0, 0, PCI_CAP_ID_EXP)) {
> > > >
> > > > You are implying that PCIe can overmap and PCI can't.  Why is
> > > > that?   (I'm assuming that 'overmap' means that inbound window
> > > > can extend beyond the end of ram.)
> > >
> > > Shouldn't the concern be whether we're overlapping outbound, not
> > merely
> > > whether we go beyond the end of RAM?
> > >
> > > And couldn't inbound/outbound overlap be an issue even on PCIe, if
> > > there's a PCI bridge underneath it?
> > I believe that the overlap problem would be avoided in case of 36 bit
> > physical support (outbound window address would be high order 36 bit
> > value), right?
> 
> Not necessarily -- it depends on what the bus addresses are, not the host
> addresses.
The fsl_pci code already checks for the overlap condition. The inbound window range is set to the minimum of memory end address or the start of pci outbound address. In our case, the guest memory end address (host physical address derived from the huge tlb page) is less than the pci outbound start address. But, it could be the other way round as well. In both cases if the inbound memory can't map the end of RAM address (for the guest), the fsl_pci driver would enable swiotlb.

For the first case where end of memory is < start of pci outbound address, we can avoid enabling swiotlb (when end address is not a power of two) by over committing the inbound address window.

-Varun