Generating elf kernel ?

tiejun.chen tiejun.chen at windriver.com
Sun Sep 19 11:40:15 EST 2010


Scott Wood wrote:
> On Fri, 17 Sep 2010 09:58:41 +0800
> "tiejun.chen" <tiejun.chen at windriver.com> wrote:
> 
>> Scott Wood wrote:
>>> The guest OS *is* the same as native Linux, as far as TLB handling is
>>> concerned.
>> Looks you means the TLB exception handler should be same between the native and
>> the guest OS. Right?
> 
> Yes.

I don't think so. The HY should assist the guest OS on MMU since I already point
the guest OS have no authority to create a real TLB directly as I previously said.

> 
>> Here I assume we're talking about e500mc since as far as I know for Freescale
>> only e500mc is designed to support virtual machine based on ISA 2.0.6.
> 
> Yes, though there's nothing preventing virtualization on cores without
> category E.HV (KVM supports this) -- it's just slower.

Absolutely.

> 
>> I also know all TLB exceptions can direct to the guest OS when we enable
>> EPCR[DTLBGS|ITLBGS|DSIGS|ISIGS]. But some TLB instructions (i.e. tlbwe )are the
>> privileged instructions. So the guest OS always trap into the hypervisor and
>> then the hypervisor should complete the real action with appropriate physical
>> address. 
> 
> Yes, of course.  But that's not the point.  I was just using it as a
> convenient example because that's what I've recently done ELF loading
> with...  There's no reason U-Boot couldn't do the same if its ELF
> loader were updated to support device trees.  Currently U-Boot loads
> bootwrapperless uImages to physical address zero.

I never doubt the U-boot can do this for uImage. But I think we're always
talking about vmlinux, a bare Image.

Here you already assume so many conditions for vmlinux before we were
discussing. Such as bootwrapperlee uImage, its ELF loader can update/support
dtb, the HY... I think this is just why I say we cannot boot vmlinux based on
common boot loader if only change entry point of vmlinux.

> 
> And FWIW, we have run setups where our hv loads Linux to true
> physical zero (with the hv living elsewhere), not just guest physical.

That's true. The HY should be allowed to access any address.

Best Regards
Tiejun

> 
> -Scott
> 
> 



More information about the Linuxppc-dev mailing list