Re: Xen at scale

[Christian Limpach <chris@pin.lu>]

> Here's a patch that adds back the CONFIG_HIGHMEM stuff.  It works
> fine with CONFIG_HIGHMEM off and with CONFIG_HIGHMEM on with
> dom0_memory <= 512MB, but fails with dom0_memory == 1GB (1048576)
> on my 2GB machine.

I think you'll need support from Xen for CONFIG_HIGHMEM:  the domain builder
maps all the domain's memory starting at 0xc0000000 so there's only 1GB
(less what's reserved for the Xen) available.  This is also how Linux maps
its memory and gets very simple kva->pa translation but is limited to ~1GB
of physical kernel memory (unless there's some trick hidden somewhere, I
don't know the Linux kernel that well (yet ;-)) which is also the reason for
CONFIG_HIGHMEM in the first place (again, as far as I can tell).

My guess is that one of the DOM0 operations fails in the builder when you
request a too large domain or that page table initialization fails in the
builder.

Now it would seem easy to just not map the memory which will end up in
highmem but we need to have all the memory mapped initially so that we can
build the phys_to_machine_mapping array.  I believe you can't build it only
from the machine_to_phys_mapping array provided by Xen since you don't know
which pages belong to the domain...

I think I would map the highmem pages starting at VA 0, the alternative
seems to be to pass an array with the page numbers to the domain.  I prefer
mapping the pages over using an array because we're already using the
pagetables to pass this kind of information and using an array introduces a
2nd interface between the domain's kernel and the domain builder.  Maybe
there's a simpler solution?

     christian



-------------------------------------------------------
This SF.Net email is sponsored by: IBM Linux Tutorials
Free Linux tutorial presented by Daniel Robbins, President and CEO of
GenToo technologies. Learn everything from fundamentals to system
administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click