Re: [PATCH 5/7] Use %gs for per-cpu sections in kernel

From: Rusty Russell
Date: Sun Sep 24 2006 - 21:16:38 EST

On Sun, 2006-09-24 at 18:03 -0700, Jeremy Fitzhardinge wrote:
> Rusty Russell wrote:
> >> So are symbols referencing the .data.percpu section 0-based? Wouldn't
> >> you need to subtract __per_cpu_start from the symbols to get a 0-based
> >> segment offset?
> >>
> >
> > I don't think I understand the question.
> >
> > The .data.percpu section is the "template" per-cpu section, freed along
> > with other initdata: after setup_percpu_areas() is called, it is not
> > supposed to be used. Around that time, the gs segment is set up based
> > at __per_cpu_offset[cpu], so "%gs:<varname>" accesses the local version.
> >
> If you do
> DEFINE_PER_CPU(int, foo);
> then this ends up defining per_cpu__foo in .data.percpu. Since
> .data.percpu is part of the init data section, it starts at some address
> X (not 0), so the real offset into the actual per-cpu memory is actually
> (per_cpu__foo - __per_cpu_start). setup_per_cpu_areas() builds this
> delta into the __per_cpu_offset[], and so it means that the base of your
> %gs segment is at -__per_cpu_start from the actual start of the CPU's
> per-cpu memory, and the limit has to be correspondingly larger. Which
> is a bit ugly.

Hi Jeremy!

You're thinking of it in a convoluted way, by converting to offsets
from the per-cpu section, then converting it back. How about this
explanation: the local cpu's versions are offset from where the compiler
thinks they are by __per_cpu_offset[cpu]. We set the segment base to
__per_cpu_offset[cpu], so "%gs:per_cpu__foo" gets us straight to the
local cpu version. __per_cpu_offset[cpu] is always positive (kernel
image sits at bottom of kernel address space).

> Especially since "__per_cpu_start" is actually very
> large, and so this scheme pretty much relies on being able to wrap
> around the segment limit, and will be very bad for Xen.

__per_cpu_start is large, yes. But there's no reason to use it in
address calculation. The second half of your statement is not correct.

> An alternative is to put the "-__per_cpu_start" into the addressing mode
> when constructing the address of the per-cpu variable.

I think you're thinking of TLS relocations? I don't use them...

Help! Save Australia from the worst of the DMCA:

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