On Wed, Jun 09, 2021 at 03:45:01AM +0900, Hyeonggon Yoo wrote:
> static __always_inline void *kmalloc_node(size_t size, gfp_t flags, int node)
> {
>
> if ( (
> __builtin_constant_p(
> !!(__builtin_constant_p(size) && size <= (1UL << ((11 + 12 - 1) <= 25 ? (11 + 12 - 1) : 25)))
> )
> ? (!!(__builtin_constant_p(size) && size <= (1UL << ((11 + 12 - 1) <= 25 ? (11 + 12 - 1) : 25))))
> : ({ static struct ftrace_branch_data __attribute__((__aligned__(4))) __attribute__((__section__("_ftrace_branch"))) __if_trace = { .func = __func__, .file = "include/linux/slab.h", .line = 601, }; (!!(__builtin_constant_p(size) && size <= (1UL << ((11 + 12 - 1) <= 25 ? (11 + 12 - 1) : 25)))) ? (__if_trace.miss_hit[1]++,1) : (__if_trace.miss_hit[0]++,0); })) )
> {
> unsigned int i = __kmalloc_index(size, true);
>
>
> they are so ugly but the point is:
>
> > > It seems that gcc evaluates
> > >
> > > __builtin_constant_p(
> > > !!(builtin_constant_p(size)
> > > && size <= KMALLOC_MAX_CACHE_SIZE)
> > > )
> > > as true.
> > >
> > > mm.. when the size passed to bpf_map_kmalloc_node is not constant,
> > > (__builtin_constant_p(size) && size <= KMALLOC_MAX_CACHE_SIZE) is false.
> > > then __builtin_constant_p(!!false) is true. So it calls kmalloc_index.
> > >
There were some logical errors on my words. I still think the compiler
evaluate it as true, but that is not reason why kmalloc_node calls
__kmalloc_index for non-constant size.
I wonder why kmalloc_node called __kmalloc_index(size, size_is_constant=true)
for non-constant size.
Thanks,
Hyeonggon
> > > what change should be made?
> > > just checking CONFIG_PROFILE_ALL_BRANCHES to kmalloc_index's if condition
> > > doesn't make sense, because kmalloc_node is not working as expected.
> some evidence to add:
>
> there are 4 calls of bpf_map_kmalloc_node.
> if you comment out two calls of bpf_map_kmalloc_node with non-constant
> (in line 168, 508), it doesn't make an error. So I thought
> it was problem when non-constant size was passed.
>
> And if "kmalloc_index makes problem with non-constant size" is
> true, then it is caller's fault because it is not allowed (except
> in __kmalloc_index)
>
> kmalloc_node shouldn't call kmalloc_index if the size was not
> constant.
