Re: 2.5.68-mm4

[Eric W. Biederman]

Anton Blanchard <[e-mail]> writes:

> Hi,
> 
> > . Included the `kexec' patch - load Linux from Linux.  Various people want
> >   this for various reasons.  I like the idea of going from a login prompt to
> >   "Calibrating delay loop" in 0.5 seconds.
> 
> One thing that bothers me about kexec is how we grab low pages in
> kimage_alloc_page(). On a partitioned ppc64 box I will need to grab
> memory in the low 256MB and the machine might have 500GB of memory
> free. Thats going to take some time :)

Could you explain to me the need to allocate memory in the low 256MB.
Generally the design is that you can allocate the memory anywhere
and then relocate_kernel.S will move where it needs to be kept.

I have had people wanting to use 300MB initial ramdisks and the like.  
If you have 500GB of memory what is the point of keeping anything on a disk?

When you have 4TB on a cluster or a NUMA machine I can understand
wanting to keep things local to a node.  But in those cases you want
to have local node zones so the problem does not come up.

In general I hate restricting the memory you can use, because kexec is
not just about booting linux.  But it is about booting anything that
we reasonably can.  The only case I have seen so far that makes sense
is when your physical memory is larger than your virtual memory.

> Id hate to introduce a separate zone just for this sort of stuff (we
> currently throw all memory in the DMA zone). Could we add a hint to
> the page allocator where it makes a best effort to grab memory below
> a threshold?

I suspect so.  And I can't imagine it would be that hard to implement.

But I think I would like to see why you need that.

Eric