Hi, Toshi,
On 12/25/15 at 03:09pm, Toshi Kani wrote:
> Change to call walk_iomem_res_desc() for searching resource entries
> with the following names:
> "ACPI Tables"
> "ACPI Non-volatile Storage"
> "Persistent Memory (legacy)"
> "Crash kernel"
>
> Note, the caller of walk_iomem_res() with "GART" is left unchanged
> because this entry may be initialized by out-of-tree drivers, which
> do not have 'desc' set to IORES_DESC_GART.
Found below commit which initialize the GART entry:
commit 56dd669a138c40ea6cdae487f233430d12372767
Author: Aaron Durbin <adurbin at google.com>
Date: Tue Sep 26 10:52:40 2006 +0200
[PATCH] Insert GART region into resource map
Patch inserts the GART region into the iomem resource map. The GART will then
be visible within /proc/iomem. It will also allow for other users
utilizing the GART to subreserve the region (agp or IOMMU).
Signed-off-by: Aaron Durbin <adurbin at google.com>
But later it was reverted:
commit 707d4eefbdb31f8e588277157056b0ce637d6c68
Author: Bjorn Helgaas <bhelgaas at google.com>
Date: Tue Mar 18 14:26:12 2014 -0600
Revert "[PATCH] Insert GART region into resource map"
This reverts commit 56dd669a138c, which makes the GART visible in
/proc/iomem. This fixes a regression: e501b3d87f00 ("agp: Support 64-bit
APBASE") exposed an existing problem with a conflict between the GART
region and a PCI BAR region.
The GART addresses are bus addresses, not CPU addresses, and therefore
should not be inserted in iomem_resource.
On many machines, the GART region is addressable by the CPU as well as by
an AGP master, but CPU addressability is not required by the spec. On some
of these machines, the GART is mapped by a PCI BAR, and in that case, the
PCI core automatically inserts it into iomem_resource, just as it does for
all BARs.
Inserting it here means we'll have a conflict if the PCI core later tries
to claim the GART region, so let's drop the insertion here.
The conflict indirectly causes X failures, as reported by Jouni in the
bugzilla below. We detected the conflict even before e501b3d87f00, but
after it the AGP code (fix_northbridge()) uses the PCI resource (which is
zeroed because of the conflict) instead of reading the BAR again.
Conflicts:
arch/x86_64/kernel/aperture.c
Fixes: e501b3d87f00 agp: Support 64-bit APBASE
Link: https://bugzilla.kernel.org/show_bug.cgi?id=72201
Reported-and-tested-by: Jouni Mett?l? <jtmettala at gmail.com>
Signed-off-by: Bjorn Helgaas <bhelgaas at google.com>
For amd64 agp, currently the region name is "aperture" instead:
drivers/char/agp/amd64-agp.c: agp_aperture_valid()
This may not be the only case, but I doubt that anyone is testing this since
long time ago kexec-tools excluding the 'GART' region. Kexec-tools and kexec_file
may need update to use "aperture" if someone can test it.
I think adding an enum value for compatibility is reasonable, we do not care
about third party drivers in mainline.
Thanks
Dave
