Re: [RFC 00/15] x86_64: Optimize percpu accesses

[ebiederm@xmission.com (Eric W. Biederman)]

Jeremy Fitzhardinge <jeremy@goop.org> writes:

> H. Peter Anvin wrote:
>> Thinking about this some more, I don't know if it would make sense to put the
>> x86-64 stack canary at the *end* of the percpu area, and otherwise use
>> negative offsets.  That would make sure they were readily reachable from
>> %rip-based references from within the kernel text area.
>
> If we can move the canary then a whole pile of options open up.  But the problem
> is that we can't.

But we can pick an arbitrary point where %gs points at.

Hmm.  This whole thing is even sillier then I thought.
Why can't we access per cpu vars as:
%gs:(per_cpu__var - __per_cpu_start) ?

If we can subtract constants and allow the linker to perform that resolution
at link.  A zero based per cpu segment becomes a moot issue.

We may need to change the definition of PERCPU in vmlinux.lds.h to
#define PERCPU(align)							\
	. = ALIGN(align);						\
-	__per_cpu_start = .;					\
	.data.percpu  : AT(ADDR(.data.percpu) - LOAD_OFFSET) {		\
+		__per_cpu_start = .;					\
		*(.data.percpu)						\
		*(.data.percpu.shared_aligned)				\
+		__per_cpu_end = .;					\
+	}
-	}								\
-	__per_cpu_end = .;


So that the linker knows  __per_cpu_start and __per_cpu_end are in the same section
but otherwise it sounds entirely reasonable.  Just slightly trickier math at link
time.

Eric