[PATCH] x86/kexec: Exclude GART aperture from vmcore

Jiri Bohac jbohac at suse.cz
Tue Nov 7 05:42:12 PST 2017

On Tue, Nov 07, 2017 at 07:39:56PM +0800, Baoquan He wrote:
> I saw you defined the variable as xx_stolen_xx, does it mean that the
> memory region where aperture located will be stolen from memory domain?
> I am wondering if it will cause error when access this region from direct
> mapping since allocate_aperture() is called in pci_iommu_alloc(), while
> in setup_arch() we have built the direct mapping for all system
> ram. Did I miss anything?

yes, I believe accessing it through the direct mapping would
cause the error; but the range is allocated by the memblock
allocator and never given back, which effectively steals it from
any other use; The memory will never be given to any user by any
subsequent allocation, so nothing will access it.

> Anyway, if it should be excluded from crash memory region, can we dig it
> away from /proc/iomem so that it's a hole in /proc/vmcore? Like this, we

Not sure this would work. In bko#72201, marking the range as used
caused pci_claim_resource() to error out with -EBUSY after
request_resource_conflict() (the error message has changed in
29003be but the logic remains).

My (possibly wrong?) reading of pci_claim_resource() tells me
that leaving the range out of the map would cause it to error out
a little earlier with -EINVAL just after
pci_find_parent_resource().  I'd rather avoid such experiments in
fear of causing regressions similar to bko#72201. This particular
one remained unnoticed for 8 years. I can't possibly test all AGP
drivers :/ 
(So much for my justification of passing this to crash
independently of the iomem_resource/e820 infrastructure.)

> don't worry about the user space kexec utility either. 

What's the problem with the userspace kexec? The bug is in
reading /proc/vmcore by makedumpfile. kexec would only operate
within the preallocated crashkernel area, right?


Jiri Bohac <jbohac at suse.cz>
SUSE Labs, Prague, Czechia

More information about the kexec mailing list