10:25 p.m.
On Mon, Jun 07, 2021 at 01:40:02PM +0200, Vlastimil Babka wrote:
On 6/6/21 1:08 PM, Hyeonggon Yoo wrote:
> On Sat, Jun 05, 2021 at 02:10:46PM +0800, kernel test robot wrote:
>> tree: https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git
master
>> head: ccc252d2e818f6a479441119ad453c3ce7c7c461
>> commit: a7ba988ff9de37f0961b4bf96d17aca73d0d2e25 [7012/7430] mm, slub: change
run-time assertion in kmalloc_index() to compile-time
>> config: parisc-randconfig-r014-20210604 (attached as .config)
>> compiler: hppa-linux-gcc (GCC) 9.3.0
>> reproduce (this is a W=1 build):
>> wget
https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O
~/bin/make.cross
>> chmod +x ~/bin/make.cross
>> #
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commi...
>> git remote add linux-next
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git
>> git fetch --no-tags linux-next master
>> git checkout a7ba988ff9de37f0961b4bf96d17aca73d0d2e25
>> # save the attached .config to linux build tree
>> COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross
ARCH=parisc
>>
>> If you fix the issue, kindly add following tag as appropriate
>> Reported-by: kernel test robot <lkp(a)intel.com>
>>
>> All errors (new ones prefixed by >>):
>> In file included from <command-line>:
>> In function 'kmalloc_index',
>> inlined from 'kmalloc_node' at include/linux/slab.h:572:20,
>> inlined from 'bpf_map_kmalloc_node.isra.0.part.0' at
include/linux/bpf.h:1319:9:
>> >> include/linux/compiler_types.h:328:38: error: call to
'__compiletime_assert_183' declared with attribute error: unexpected size in
kmalloc_index()
>> 328 | _compiletime_assert(condition, msg, __compiletime_assert_,
__COUNTER__)
>> | ^
>> include/linux/compiler_types.h:309:4: note: in definition of macro
'__compiletime_assert'
>> 309 | prefix ## suffix(); \
>> | ^~~~~~
>> include/linux/compiler_types.h:328:2: note: in expansion of macro
'_compiletime_assert'
>> 328 | _compiletime_assert(condition, msg, __compiletime_assert_,
__COUNTER__)
>> | ^~~~~~~~~~~~~~~~~~~
>> include/linux/build_bug.h:39:37: note: in expansion of macro
'compiletime_assert'
>> 39 | #define BUILD_BUG_ON_MSG(cond, msg) compiletime_assert(!(cond), msg)
>> | ^~~~~~~~~~~~~~~~~~
>> include/linux/slab.h:389:2: note: in expansion of macro
'BUILD_BUG_ON_MSG'
>> 389 | BUILD_BUG_ON_MSG(1, "unexpected size in
kmalloc_index()");
>> | ^~~~~~~~~~~~~~~~
>
> Reproduce with attached config, and read the code.
> It has no problem in clang (clang-10/clang-11). it is problem with gcc.
But what exactly is the gcc problem here?
Did you have to reproduce it with specific gcc version and/or architecture?
Before replying, I should say that I'm not an expert on gcc.
I just tried some ways to fix the error, and it seemed to me that
gcc is doing something wrong.
I found the error was in kernel/bpf/local_storage.c
and I copied the bot's config in linux-next (20210607), and just entered all new
config.
running 'make kernel/bpf/local_storage.o CC=gcc' can reproduce the
error.
the bot says it is error with parisc, but I was able to reproduce
it in my x86 machine.
I tested on gcc-9.3.0 and gcc-10.2.0 both makes an error,
and clang 10.0.1, clang 11.0.0 didn't make an error.
> I found two ways to solve the error (maybe there would be
better
> solution) Any thoughts or opinions?
>
>
> First ways is to change condition of kmalloc_index macro.
>
> diff --git a/include/linux/slab.h b/include/linux/slab.h
> index 70e46db766ca..be2c900cba4b 100644
> --- a/include/linux/slab.h
> +++ b/include/linux/slab.h
> @@ -397,7 +397,7 @@ static __always_inline unsigned int __kmalloc_index(size_t
size,
> /* Will never be reached. Needed because the compiler may complain */
> return -1;
> }
> -#define kmalloc_index(s) __kmalloc_index(s, true)
> +#define kmalloc_index(s) __kmalloc_index(s, __builtin_constant_p(s) &&
true)
I wonder how this extra guard can possibly matter?
bpf_map_kmalloc_node()
kmalloc_node()
if (__builtin_constant_p(size) ...)
unsigned int i = kmalloc_index(size);
We shouldn't be even reaching kmalloc_index() unless __builtin_constant_p(size)
is already true.
Yes, I knew that when I wrote the code. That totally doesn't make sense.
why __builtin_constant_p(size) was true in kmalloc_node,
and false in the evalaution (__builtin_constant_p(s) && true)?
but it actually solves the error.
At this point, I thought gcc was doing something wrong....
Well, I know we need more evidence to conclude gcc is wrong.
(Or we made wrong code that makes compiler confusing.)
I want to hear you and some other people's opinion.
> #endif /* !CONFIG_SLOB */
>
> void *__kmalloc(size_t size, gfp_t flags) __assume_kmalloc_alignment __malloc;
>
>
> Second way is making bpf_map_kmalloc_node always inline.
If bpf_map_kmalloc_node() is not being compiled as inline, how can it possibly
evaluate __builtin_constant_p(size) as true when processing the inlined
kmalloc_node()?
As I said above - it doesn't make sense, but gcc is acting differently
on __inline and inline, in bpf_map_kmalloc_node function. just making
bpf_map_kmalloc_node always solves the error, but I have no clue WHY...
anyway, in summary:
- the diff I sent should not make sense, but it works for gcc.
- So I think gcc is doing something wrong (but more evidence needed)
I can be missing something. So If I said something wrong, or if you
can't reproduce the error, please tell me!
Thanks,
Hyeonggon Yoo
> > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > index 02b02cb29ce2..09379d705349 100644
> > --- a/include/linux/bpf.h
> > +++ b/include/linux/bpf.h
> > @@ -1312,7 +1312,7 @@ void *bpf_map_kzalloc(const struct bpf_map *map, size_t
size, gfp_t flags);
> > void __percpu *bpf_map_alloc_percpu(const struct bpf_map *map, size_t size,
> > size_t align, gfp_t flags);
> > #else
> > -static inline void *
> > +static __always_inline void *
> > bpf_map_kmalloc_node(const struct bpf_map *map, size_t size, gfp_t flags,
> > int node)
> > {
> >
>