Re: [PATCH] memory hotadd fixes [4/5] avoid check in acpi

From: KAMEZAWA Hiroyuki
Date: Thu Aug 03 2006 - 23:09:05 EST

On Thu, 03 Aug 2006 20:00:08 -0700
keith mannthey <kmannth@xxxxxxxxxx> wrote:

> > > What protecting is there for calling add_memory on an already present
> > > memory range?
> > >
> > For example, considering ia64, which has 1Gbytes section...
> Maybe 1gb sections is too large?
ia64 machines sometimes to have crazy big 1gb section is requested.
Configurable section_size for small machines was rejected in old days.

> > hot add following region.
> > ==
> > (A) 0xc0000000 - 0xd7ffffff (section 3)
> > (B) 0xe0000000 - 0xffffffff (section 3)
> > ==
> > (A) and (B) will go to the same section, but there is a memory hole between
> > (A) and (B). Considering memory (B) appears after (A) in DSDT.
> >
> > After add_memory() against (A) is called, section 3 is ready.
> > Then, pfn_valid(0xe0000000) and pfn_valid(0xffffffff) returns true because
> > they are in section 3.
> > So, checking pfn_valid() for (B) will returns true and memory (B) cannot be
> > added. ioresouce collision check will help this situation.
> With iommus out there throwing aliment all off way the flexability is
> good.
> My question is this.
> Assuming 0-0xbfffffff is present.
> What keeps 0xa0000000 to 0xa1000000 from being re-onlined by a bad call
> to add_memory?

Usual sparsemem's add_memory() checks whether there are sections in
sparse_add_one_section(). then add_pages() returns -EEXIST (nothing to do).
And ioresouce collision check will finally find collision because 0-0xbffffff
resource will conflict with 0xa0000000 to 0xa10000000 area.
But, x86_64 's (not sparsemem) add_pages() doen't do collision check, so it panics.
I posted patch to catch collision before calling arch_add_memory(), I think it will
help x86_64 users. and no-re-online.


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at