Re: [PATCH v18 0/9] Compile-time stack metadata validation

From: Josh Poimboeuf
Date: Sun Feb 28 2016 - 12:00:20 EST


On Sun, Feb 28, 2016 at 07:57:29AM +0100, Ingo Molnar wrote:
>
> Another thing I've noticed are _thousands_ of these warnings:
>
> objtool: kernel/sched/core.o: preempt_schedule_irq()+0x20: function has unreachable instruction
> objtool: kernel/sched/core.o: preempt_schedule_irq()+0x27: function has unreachable instruction
> objtool: kernel/sched/core.o: preempt_schedule_irq()+0x2c: function has unreachable instruction
> objtool: kernel/sched/core.o: preempt_schedule_irq()+0x33: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_put_page_vector()+0x171: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_put_page_vector()+0x178: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_put_page_vector()+0x17d: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_put_page_vector()+0x184: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_get_direct_page_vector()+0xe2: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_get_direct_page_vector()+0xe9: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_get_direct_page_vector()+0xee: function has unreachable instruction
> objtool: net/ceph/pagevec.o: ceph_get_direct_page_vector()+0xf5: function has unreachable instruction
> objtool: net/ceph/pagevec.o: zero_user_segment()+0x120: function has unreachable instruction
> objtool: net/ceph/pagevec.o: zero_user_segment()+0x127: function has unreachable instruction
> objtool: net/ceph/pagevec.o: zero_user_segment()+0x12c: function has unreachable instruction
> objtool: net/ceph/pagevec.o: zero_user_segment()+0x133: function has unreachable instruction
>
> that's on an x86-64 allyesconfig kernel, with objtool merged to tip:master, using
> GCC 4.9.2:
>
> gcc version 4.9.2 20150212 (Red Hat 4.9.2-6) (GCC)

Those are caused by CONFIG_KASAN, CONFIG_UBSAN, and CONFIG_GCOV_KERNEL,
which can all add unreachable instructions. Technically, an unreachable
instruction isn't really a problem, but objtool warns about it because
it often means there's something going on in the control flow which it
doesn't understand.

I guess we could make CONFIG_STACK_VALIDATION conflict with those
options, but I think that would disable it on allyesconfig.

I'll see if I can make objtool smarter so that it detects these special
cases of unreachable instructions and ignores them.

> Also, please prefix such warnings with the standard compiler prefix, something
> like this:
>
> kernel/sched/core.c: warning: objtool: preempt_schedule_irq()+0x20: function has unreachable instruction
>
> so that scripts/tools monitoring new build warnings can pick them up
> automatically?

Yeah, good idea.

--
Josh