Re: linux-next: Tree for Nov 20 (kcsan + objtool)

[Marco Elver <elver@google.com>]

On Wed, 20 Nov 2019 at 17:18, Randy Dunlap <rdunlap@infradead.org> wrote:
>
> On 11/20/19 1:34 AM, Stephen Rothwell wrote:
> > Hi all,
> >
> > Changes since 20191119:
> >
>
> on x86_64:
>
> kernel/kcsan/core.o: warning: objtool: kcsan_found_watchpoint()+0xa: call to kcsan_is_enabled() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read1()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write1()+0x10: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read2()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write2()+0x10: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read4()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write4()+0x10: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read8()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write8()+0x10: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read16()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write16()+0x10: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_read_range()+0x13: call to find_watchpoint() with UACCESS enabled
> kernel/kcsan/core.o: warning: objtool: __tsan_write_range()+0x10: call to find_watchpoint() with UACCESS enabled
>
> kernel/trace/trace_branch.o: warning: objtool: ftrace_likely_update()+0x361: call to __stack_chk_fail() with UACCESS enabled
>
>
> Full randconfig file is attached.

Thanks.

This is due to CONFIG_CC_OPTIMIZE_FOR_SIZE=y. It seems the compiler
decides to not even inline small static inline functions. I tried to
make this go away by adding __always_inline, but then we're also left
with atomic64_try_cmpxchg which never gets inlined.

The optimized build simply inlines the small static inline functions.
We certainly do not want to add more functions to the objtool
whitelist, especially those that are private to KCSAN.

We could fix it by either:

1. Adding __always_inline to every function used by the KCSAN runtime
outside user_access_save + also fix atomic64_try_cmpxchg
(atomic-instrumented.h).

2. Just not compile KCSAN with -Os, i.e. have the Makefile strip -Os
and replace it with -O2 for kcsan/core.c. #2 is the simpler option,
and would probably make KCSAN more effective even with -Os. Although
it might violate the assumption of whoever decided they want both
CC_OPTIMIZE_FOR_SIZE and KCSAN. It might also mean that future
compilers that have a new inlining algorithm will have the same
problem.

What do people think is better?

Thanks,
-- Marco

> --
> ~Randy
> Reported-by: Randy Dunlap <rdunlap@infradead.org>