Re: [PATCH bpf-next] selftests/bpf: make verifier loop tests arch independent

[Ilya Leoshkevich <iii@linux.ibm.com>]

> Am 08.07.2019 um 23:20 schrieb Stanislav Fomichev <sdf@fomichev.me>:
> 
> On 07/08, Ilya Leoshkevich wrote:
>> 
>> 
>>> Am 08.07.2019 um 18:13 schrieb Stanislav Fomichev <sdf@fomichev.me>:
>>> 
>>> On 07/03, Y Song wrote:
>>>> On Wed, Jul 3, 2019 at 1:51 PM Stanislav Fomichev <sdf@google.com> wrote:
>>>>> 
>>>>> Take the first x bytes of pt_regs for scalability tests, there is
>>>>> no real reason we need x86 specific rax.
>>>>> 
>>>>> Signed-off-by: Stanislav Fomichev <sdf@google.com>
>>>>> ---
>>>>> tools/testing/selftests/bpf/progs/loop1.c | 3 ++-
>>>>> tools/testing/selftests/bpf/progs/loop2.c | 3 ++-
>>>>> tools/testing/selftests/bpf/progs/loop3.c | 3 ++-
>>>>> 3 files changed, 6 insertions(+), 3 deletions(-)
>>>>> 
>>>>> diff --git a/tools/testing/selftests/bpf/progs/loop1.c b/tools/testing/selftests/bpf/progs/loop1.c
>>>>> index dea395af9ea9..d530c61d2517 100644
>>>>> --- a/tools/testing/selftests/bpf/progs/loop1.c
>>>>> +++ b/tools/testing/selftests/bpf/progs/loop1.c
>>>>> @@ -14,11 +14,12 @@ SEC("raw_tracepoint/kfree_skb")
>>>>> int nested_loops(volatile struct pt_regs* ctx)
>>>>> {
>>>>>       int i, j, sum = 0, m;
>>>>> +       volatile int *any_reg = (volatile int *)ctx;
>>>>> 
>>>>>       for (j = 0; j < 300; j++)
>>>>>               for (i = 0; i < j; i++) {
>>>>>                       if (j & 1)
>>>>> -                               m = ctx->rax;
>>>>> +                               m = *any_reg;
>>>> 
>>>> I agree. ctx->rax here is only to generate some operations, which
>>>> cannot be optimized away by the compiler. dereferencing a volatile
>>>> pointee may just serve that purpose.
>>>> 
>>>> Comparing the byte code generated with ctx->rax and *any_reg, they are
>>>> slightly different. Using *any_reg is slighly worse, but this should
>>>> be still okay for the test.
>>>> 
>>>>>                       else
>>>>>                               m = j;
>>>>>                       sum += i * m;
>>>>> diff --git a/tools/testing/selftests/bpf/progs/loop2.c b/tools/testing/selftests/bpf/progs/loop2.c
>>>>> index 0637bd8e8bcf..91bb89d901e3 100644
>>>>> --- a/tools/testing/selftests/bpf/progs/loop2.c
>>>>> +++ b/tools/testing/selftests/bpf/progs/loop2.c
>>>>> @@ -14,9 +14,10 @@ SEC("raw_tracepoint/consume_skb")
>>>>> int while_true(volatile struct pt_regs* ctx)
>>>>> {
>>>>>       int i = 0;
>>>>> +       volatile int *any_reg = (volatile int *)ctx;
>>>>> 
>>>>>       while (true) {
>>>>> -               if (ctx->rax & 1)
>>>>> +               if (*any_reg & 1)
>>>>>                       i += 3;
>>>>>               else
>>>>>                       i += 7;
>>>>> diff --git a/tools/testing/selftests/bpf/progs/loop3.c b/tools/testing/selftests/bpf/progs/loop3.c
>>>>> index 30a0f6cba080..3a7f12d7186c 100644
>>>>> --- a/tools/testing/selftests/bpf/progs/loop3.c
>>>>> +++ b/tools/testing/selftests/bpf/progs/loop3.c
>>>>> @@ -14,9 +14,10 @@ SEC("raw_tracepoint/consume_skb")
>>>>> int while_true(volatile struct pt_regs* ctx)
>>>>> {
>>>>>       __u64 i = 0, sum = 0;
>>>>> +       volatile __u64 *any_reg = (volatile __u64 *)ctx;
>>>>>       do {
>>>>>               i++;
>>>>> -               sum += ctx->rax;
>>>>> +               sum += *any_reg;
>>>>>       } while (i < 0x100000000ULL);
>>>>>       return sum;
>>>>> }
>>>>> --
>>>>> 2.22.0.410.gd8fdbe21b5-goog
>>>> 
>>>> Ilya Leoshkevich (iii@linux.ibm.com, cc'ed) has another patch set
>>>> trying to solve this problem by introducing s360 arch register access
>>>> macros. I guess for now that patch set is not needed any more?
>>> Oh, I missed them. Do they fix the tests for other (non-s360) arches as
>>> well? I was trying to fix the issue by not depending on any arch
>>> specific stuff because the test really doesn't care :-)
>> 
>> They are supposed to work for everything that defines PT_REGS_RC in
>> bpf_helpers.h, but I have to admit I tested only x86_64 and s390.
>> 
>> The main source of problems with my approach were mismatching definitions
>> of struct pt_regs for userspace and kernel, and because of that there was
>> some tweaking required for both arches. I will double check how it looks
>> for others (arm, mips, ppc, sparc) tomorrow.
> Thanks, I've tested your patches and they fix my issue as well. So you
> can have my Tested-by if we'd go with your approach.
> 
> One thing I don't understand is: why do you add 'ifdef __KERNEL__' to
> the bpf_helpers.h for x86 case? Who is using bpf_helpers.h with
> __KERNEL__ defined? Is it perf?

That’s samples/bpf. Also, there is a modified copy of it in bcc
(src/cc/export/helpers.h), which also gets built with __KERNEL__.

Best regards,
Ilya