[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Xen-ia64-devel] [patch] calculate dom0 metaphysical load address correctly
Isaku Yamahata wrote:
> On Thu, May 24, 2007 at 11:42:56AM +0900, Isaku Yamahata wrote:
>> elilo tries the physical address given by its elf header at first,
>> then tries to relocate the image only when administrater allows elilo
>> to relocate explicitly with warning message.
> I only took look at elilo code so that I don't know how distros
> handle relocations.
> I don't think relocation harms current dom0 builder as long as
> the loading address is properly chosen.
I don't really like relying on elilo for this, but maybe thats ok
to do. The problem is that the Xen ELF code looks in the ELF
headers and determines the address from there without taking into
account if the address is valid.
>> + /*
>> + * We cannot rely on the load address in the ELF headers to
>> + * determine the meta physical address the image is loaded
>> + * at since meta physical memory needs to match physical
>> + * memory of the system for NUMA awareness. Patch the address
>> + * to match the real one, based on xen_pstart
>> + */
> I don't understand this comment.
> I think it's quite feasible to support NUMA with current memory model.
> The pseudo physical address model is independ of NUMA support.
> Presumably you have many patches locally behind this patch.
For dom0 the metaphysical addresses must match real physical addresses,
ie. on SN2 we don't have anything below 0x3000000000 and we also encode
the node ID in the address so on multi node systems the addresses go up.
dom0 needs the metaphysical addresses or nothing can work as we use the
node info for talking to the PCI controllers, memory controllers (IPIs
and TLB flushes) etc.
Xen-ia64-devel mailing list