On Sun, Feb 04, 2007 at 04:51:38PM +0100, Eric Piel wrote:
 > Hello,
 > 
 > I've got a regression in 2.6.20-rc7 (-rc6 was fine) due to commit 
 > 4b95320f ([AGPGART] intel_agp: restore 
 > graphics device's pci space early in resume).

I think the key to this failure is the last line here ..

 > agpgart-intel 0000:00:00.0: resuming
 > PM: Writing back config space on device 0000:00:02.0 at offset f (was 10b, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset d (was dc, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset b (was 10161025, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset 5 (was f4000000, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset 4 (was f8000008, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset 2 (was 3000011, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset 1 (was 2b00007, writing 0)
 > PM: Writing back config space on device 0000:00:02.0 at offset 0 (was 11328086, writing 0)
 > agpgart: Unable to remap memory.

This then blows up the next access to intel_i810_private.registers, which happens to
be intel_i810_insert_entries.

Either we need .suspend methods which unmap these regions, or we need
to skip trying to map them a second time on resume.

There's an ugly patch below which does the latter. Give it a try?

The intel-agp suspend/resume code has really grown into something
of a monster, and could use some refactoring in a big way.

		Dave


From: Dave Jones <davej@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>