Hello, Jan.

On Fri, Nov 23, 2012 at 09:44:02PM +0100, Jan Engelhardt wrote:

when compiling a kernel with the gold linker (3.7.0-rc6 26d29d06ea0204, 
gcc-4.7 and binutils-2.23 in my case), certain pcpu symbols are 
seemingly errneously copied over from .o files to .ko files, leading to 
a hard warning during depmod:

	gold$ make -j8 LD=gold HOSTLD=gold
	gold$ make modules_install INSTALL_MOD_PATH=/tmp/foo
	[...]
	WARNING: /tmp/foo/lib/modules/3.7.0-rc6-jng6-default+
	/kernel/net/rds/rds_tcp.ko needs unknown symbol
	__pcpu_scope_rds_tcp_stats

Yeah, this is from the nasty tricks percpu plays to get percpu decls
working on s390 and alpha.  Take a look at
DECLARE/DEFINE_PER_CPU_SECTION() definitions in
include/linux/percpu-defs.h for details.

This happens with many modules using percpu; looking at things with nm 
reveals:

	gold/net/ipv6$ nm ipv6.o | grep __pcpu_
	0000000000000000 D __pcpu_unique_ipv6_cookie_scratch
	gold/net/ipv6$ nm ipv6.ko | grep __pcpu_
	                 U __pcpu_unique_ipv6_cookie_scratch

On the other hand, in a linux tree built with the original ld (ld.bfd), 
things look like:

	bfd$ make -j8
	[...]
	bfd/net/ipv6$ nm ipv6.o | grep pcpu
	0000000000000000 D __pcpu_unique_ipv6_cookie_scratch
	bfd/net/ipv6$ nm ipv6.ko | grep pcpu
	(no result)

So, those __pcpu_unique and __pcpu_scope variables are only used so
that multiple copy of them collide with each other and trigger compile
failure (as the actual percpu variable needs to be declared __weak).
They're all put into .discard section, so should go away once it
passes through the linker.  gold left reference to that symbol while
discarding the definition.  Hmmm... different interpretation of
.discard?  Still weird tho.  There's nothing dereferencing that
symbol.

Thanks.

-- 
tejun

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>