Rusty Russell wrote:

	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).
  

We're talking kernel virtual addresses, so the physical load address 
doesn't matter, of course.

So, take this kernel I have here as an explicit example:

$ nm -n vmlinux
[...]
c0431100 A __per_cpu_start
[...]
c0433800 D per_cpu__cpu_gdt_descr
c0433880 D per_cpu__cpu_tlbstate


And say that this CPU has its percpu data allocated at 0xc100000.

So, in this case the %gs base will be loaded with 0xc100000-0xc0431100 = 
0x4bccef00
The offset of per_cpu__cpu_gdt_descr is 0xc0433800, so 
%gs:per_cpu__cpu_gdt_descr will compute 0x4bccef00+0xc0433800 to get the 
final linear address.  Since 0xc0433800 is negative, this is actually a 
subtraction, and it therefore requires the segment to have a 4G limit.  
Which makes Xen sad.

quoted

  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.
  

__per_cpu_start is added to all per_cpu__* addresses.

quoted

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...
  

No, but this is just as bad.

    J