While examining holes in percpu section I found this :

c05f5000 D per_cpu__current_task
c05f5000 D __per_cpu_start
c05f5004 D per_cpu__cpu_number
c05f5008 D per_cpu__irq_regs
c05f500c d per_cpu__cpu_devices
c05f5040 D per_cpu__cyc2ns

<Big Hole of about 4000 bytes>

c05f6000 d per_cpu__cpuid4_info
c05f6004 d per_cpu__cache_kobject
c05f6008 d per_cpu__index_kobject

<Big Hole of about 4000 bytes>

c05f7000 D per_cpu__gdt_page

This is because gdt_page is a percpu variable, defined with
a page alignement, and linker is doing its job, two times because of .o 
nesting in the build process.

I introduced a new macro DEFINE_PER_CPU_PAGE_ALIGNED() to avoid
wasting this space. All page aligned variables (only one at this time) 
are put in a separate
subsection .data.percpu.page_aligned, at the very begining of percpu zone.

Before patch , on a x86_32 machine :

# size -A vmlinux.old vmlinux | grep percpu
.data.percpu                30232   3227471872
.data.percpu                22168   3227471872

Thats 8064 bytes saved for each CPU.

Patch against x86.git/latest

Note : This will collide with Christoph efforts to put pda at offset 0 
of percpu zone...

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>

 arch/x86/kernel/cpu/common.c      |    2 +-
 arch/x86/kernel/vmlinux_32.lds.S  |    1 +
 include/asm-generic/vmlinux.lds.h |    1 +
 include/linux/percpu.h            |    7 +++++++
 4 files changed, 10 insertions(+), 1 deletion(-)