Re: [PATCH V8 7/9] acpi: Add generic MCFG table handling

[Bjorn Helgaas <helgaas@kernel.org>]

On Wed, Jun 08, 2016 at 02:21:30PM +0200, Tomasz Nowicki wrote:
> On 08.06.2016 03:56, Bjorn Helgaas wrote:
> >On Mon, May 30, 2016 at 05:14:20PM +0200, Tomasz Nowicki wrote:
> >>In order to handle PCI config space regions properly in ACPI, new MCFG
> >>interface is defined which does sanity checks on MCFG table and keeps its
> >>root pointer. The user is able to lookup MCFG regions based on
> >>host bridge root structure and domain:bus_start:bus_end touple.
> >>Use pci_mmcfg_late_init old prototype to avoid another function name.

> >>+	/* found matching entry, bus range check */
> >>+	if (entry->end_bus_number != bus_res->end) {
> >>+		resource_size_t bus_end = min_t(resource_size_t,
> >>+					entry->end_bus_number, bus_res->end);
> >>+		pr_warn("%04x:%pR bus end mismatch, using %02lx\n",
> >>+			root->segment, bus_res, (unsigned long)bus_end);
> >>+		bus_res->end = bus_end;
> >>+	}
> 
> What about bus end mismatch case? Should we trim the host bridge bus
> range or expect MCFG entry covers that range? Sometimes we get
> _BBN-0xFF bus range, not from _CRS.

Lack of a bus range in _CRS is a firmware defect.  There's a comment
about this in acpi_pci_root_add().  On x86, we probably had to live
with firmware in the field that had this defect.  I think we should
expect all ARM64 systems to provide a bus number range in _CRS, and
fail the attach if it's not there.

I don't think we should warn about an MCFG entry that covers more than
the _CRS bus range.  On x86, it's common to have something like:

  ACPI: PCI Root Bridge [PCI0] (domain 0000 [bus 00-7f])
  ACPI: PCI Root Bridge [PCI1] (domain 0000 [bus 80-ff])

with a single MCFG entry that covers [bus 00-ff].  That seems
reasonable and I don't think it's worth warning about it.

If the MCFG entry doesn't cover all of a _CRS bus range, we should
just fail so we can find and fix broken firmware.

> >>+/* Interface called by ACPI - parse and save MCFG table */
> >
> >I think we save a *pointer* to the MCFG table, not the table itself.
> 
> Right, the comment is broken.
> 
> >And acpi_table_parse() calls early_acpi_os_unmap_memory() immediately
> >after it calls pci_mcfg_parse(), so I'm doubtful that the pointer
> >remains valid.
> 
> At this stage early_acpi_os_unmap_memory() is doing nothing since
> acpi_early_init() set acpi_gbl_permanent_mmap to 1 way before. The
> pointer is fine then.

Hmmm...  I see your argument, but this is a problem waiting to happen.
We should not depend on the internal implementation of
early_acpi_os_unmap_memory().  The pattern of:

  y = x;
  unmap(x);
  z = *y;

is just broken and we shouldn't expect readers to recognize that "oh,
unmap() isn't really unmapping anything in this special case, so this
looks wrong but is really fine."

Bjorn