On 2/14/19 12:50 AM, Alexey Brodkin wrote:
>>>>> I suspect the slab allocator should be returning 8 byte aligned addresses
>>>>> on all systems....
>>>>
>>>> why ? As I understand it is still not fool proof against the expected alignment of
>>>> inner members. There ought to be a better way to enforce all this.
>>>
>>> I agree that for ARC ARCH_SLAB_MINALIGN should be at least 8.
>>
>> This issue aside, are there other reasons ? Because making it 8 on ARC is just
>> pending the eventuality for later.
>
> But that's pretty much the same for other 32-bit arches that have 64-bit atomics
> like ARM etc. From what I may see from ARM's documentation for LDREXD/SRREXD they
> require double-word alignment of data as well.
Right LLOCKD/SCONDD (64-bit exclusive load/store) needs 64-bit aligned effective
addresses for micro-arch reasons (1 vs 2 cache lines) etc.
So lets try to unpack this for me. Say we had.
struct foo {
int a;
atomic64_t b;
};
The atomic64_t (which for ARC and most others is u64 __attribute__((aligned(8))
*already ensures* that there a 4 b padding is generated by gcc (I just confirmed
with a simple test case).
#ifdef DOALIGN__
#define my_u64 __u64 __attribute__((aligned(8)))
#else
#define my_u64 __u64
#endif
struct foo on_heap;
printf(%d", &on_heap.b)
$ arc-linux-gcc -O2 test.c -DDOALIGN__ -c --save-temps
main:
mov_s r1,@on_heap+8 <----
mov_s r0,@.LC0
b @printf
W/o the alignment attribute (say normal LDD/STD)
$ arc-linux-gcc -O2 test.c -c --save-temps
main:
mov_s r1,@on_heap+4
mov_s r0,@.LC0
b @printf
So indeed your patch aligns dynamic structs to 64-bit, ensuring any embedded
aligned_u64 to be 64-bit aligned as well. Phew !
> That said if for some reason atomic64_t variable is unaligned execution on
> any (or at least most) 32-bit architectures will lead to run-time failure,
> i.e. we'll know about it and this will be fixed.
>
> And what I'm doing by that change (ARCH_SLAB_MINALIGN=8 for ARC) I'm just
> working-around peculiarity of ARC ABI.
Right.
>
> Out of curiosity I checked if there're any other occurrences of "alingof(long long)"
> and there seems to be a couple of more:
> ----------------------------------->8-----------------------------
> # git grep alignof | grep "long long"
>
> ...
>
> kernel/workqueue.c:5693: WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
> mm/slab.c:155:#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
For ARC, it will be max(4,4) so 4
for others 32-bit,it will be max(4,8)
So indeed it makes sense to change it.
> mm/slab.c:2034: if (ralign > __alignof__(unsigned long long))
> ----------------------------------->8-----------------------------
>
> Not really sure how important is "kernel/workqueue.c" part but in case of "mm/slab.c"
> shouldn't we use ARCH_SLAB_MINALIGN there instead of that "not very meaningful" __alignof__(long long)?
