Am 09.04.19 um 13:29 schrieb Magnus Karlsson:
> On Tue, Apr 9, 2019 at 11:11 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
>>
>> On 04/09/2019 08:44 AM, Magnus Karlsson wrote:
>>> The use of smp_rmb() and smp_wmb() creates a Linux header dependency
>>> on barrier.h that is uneccessary in most parts. This patch implements
>>> the two small defines that are needed from barrier.h. As a bonus, the
>>> new implementations are faster than the default ones as they default
>>> to sfence and lfence for x86, while we only need a compiler barrier in
>>> our case. Just as it is when the same ring access code is compiled in
>>> the kernel.
>>>
>>> Fixes: 1cad07884239 ("libbpf: add support for using AF_XDP sockets")
>>> Signed-off-by: Magnus Karlsson <magnus.karlsson@xxxxxxxxx>
>>> ---
>>> tools/lib/bpf/xsk.h | 19 +++++++++++++++++--
>>> 1 file changed, 17 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/tools/lib/bpf/xsk.h b/tools/lib/bpf/xsk.h
>>> index 3638147..317b44f 100644
>>> --- a/tools/lib/bpf/xsk.h
>>> +++ b/tools/lib/bpf/xsk.h
>>> @@ -39,6 +39,21 @@ DEFINE_XSK_RING(xsk_ring_cons);
>>> struct xsk_umem;
>>> struct xsk_socket;
>>>
>>> +#if !defined bpf_smp_rmb && !defined bpf_smp_wmb
>>> +# if defined(__i386__) || defined(__x86_64__)
>>> +# define bpf_smp_rmb() asm volatile("" : : : "memory")
>>> +# define bpf_smp_wmb() asm volatile("" : : : "memory")
>>> +# elif defined(__aarch64__)
>>> +# define bpf_smp_rmb() asm volatile("dmb ishld" : : : "memory")
>>> +# define bpf_smp_wmb() asm volatile("dmb ishst" : : : "memory")
>>> +# elif defined(__arm__)
>>> +# define bpf_smp_rmb() asm volatile("dmb ish" : : : "memory")
>>> +# define bpf_smp_wmb() asm volatile("dmb ishst" : : : "memory")
>>> +# else
>>> +# error Architecture not supported by the XDP socket code in libbpf.
>>> +# endif
>>> +#endif
>>
>> Hmm, bit unfortunate, but I can see why it's needed inside the header here
>> as it's installed into the system and needs to be included by applications.
>> Fine with me. In any case, I still think we should also fix it for tooling
>> infra to avoid others running into the same issue, lemme resurrect [0] from
>> back in the days.
>>
>> One question though regarding above arm32 implementation. You match these
>> barriers against the ones from af_xdp in kernel. So smp_* variants from the
>> arch/arm/include/asm/barrier.h header need to be matched. This selects the
>> dmb(ish) and dmb(ishst) variant. The implementation of dmb() however is
>> quite CPU dependent, the above assumes to match with __LINUX_ARM_ARCH__ >= 7.
>> Perhaps this assumption needs at minimum a comment to make it clear for
>> developers.
>
> Good idea. I will put a comment in there to clarify this.
>
> /Magnus
Since this is a userspace library, what about using the C11 atomic_thread_fence()
from stdatomic.h - wouldn't that be sufficient?
>> [0] https://lore.kernel.org/netdev/20181017144156.16639-2-daniel@xxxxxxxxxxxxx/
>>
>>> static inline __u64 *xsk_ring_prod__fill_addr(struct xsk_ring_prod *fill,
>>> __u32 idx)
>>> {
>>> @@ -119,7 +134,7 @@ static inline void xsk_ring_prod__submit(struct xsk_ring_prod *prod, size_t nb)
>>> /* Make sure everything has been written to the ring before signalling
>>> * this to the kernel.
>>> */
>>> - smp_wmb();
>>> + bpf_smp_wmb();
>>>
>>> *prod->producer += nb;
>>> }
>>> @@ -133,7 +148,7 @@ static inline size_t xsk_ring_cons__peek(struct xsk_ring_cons *cons,
>>> /* Make sure we do not speculatively read the data before
>>> * we have received the packet buffers from the ring.
>>> */
>>> - smp_rmb();
>>> + bpf_smp_rmb();
>>>
>>> *idx = cons->cached_cons;
>>> cons->cached_cons += entries;
>>>
>>
