Re: Problems with csum_partial with misaligned buffers on sh4 platform

["D. Jeff Dionne" <djeffdionne@gmail.com>]

I remember there being problems with alignment on SH targets in the network stack.  IIRC, wireguard triggered it in actual use, seems to me it had to do with skb alignment.

Rich Felker may remember more, but I don’t think we implemented a (complete) solution.

Cheers,
J.

> On 11 Feb 2024, at 07:03, Guenter Roeck <linux@roeck-us.net> wrote:
> 
> Hi Adrian,
> 
>> On 2/10/24 12:12, John Paul Adrian Glaubitz wrote:
>> Hi Guenter,
>>> On Sat, 2024-02-10 at 07:12 -0800, Guenter Roeck wrote:
>>> when running checksum unit tests on sh4 qemu emulations, I get the following
>>> errors.
>>> 
>>>     KTAP version 1
>>>     # Subtest: checksum
>>>     # module: checksum_kunit
>>>     1..5
>>>     # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:500
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 53378 (0xd082)
>>>         ( u64)expec == 33488 (0x82d0)
>>>     not ok 1 test_csum_fixed_random_inputs
>>>     # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:525
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 65281 (0xff01)
>>>         ( u64)expec == 65280 (0xff00)
>>>     not ok 2 test_csum_all_carry_inputs
>>>     # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:573
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 65535 (0xffff)
>>>         ( u64)expec == 65534 (0xfffe)
>>>     not ok 3 test_csum_no_carry_inputs
>>>     ok 4 test_ip_fast_csum
>>>     ok 5 test_csum_ipv6_magic
>>> # checksum: pass:2 fail:3 skip:0 total:5
>>> 
>>> The above is with from a little endian system. On a big endian system,
>>> the test result is as follows.
>>> 
>>>     KTAP version 1
>>>     # Subtest: checksum
>>>     # module: checksum_kunit
>>>     1..5
>>>     # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:500
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 33488 (0x82d0)
>>>         ( u64)expec == 53378 (0xd082)
>>>     not ok 1 test_csum_fixed_random_inputs
>>>     # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:525
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 65281 (0xff01)
>>>         ( u64)expec == 255 (0xff)
>>>     not ok 2 test_csum_all_carry_inputs
>>>     # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:565
>>>     Expected ( u64)result == ( u64)expec, but
>>>         ( u64)result == 1020 (0x3fc)
>>>         ( u64)expec == 0 (0x0)
>>>     not ok 3 test_csum_no_carry_inputs
>>>     # test_ip_fast_csum: ASSERTION FAILED at lib/checksum_kunit.c:589
>>>     Expected ( u64)expected == ( u64)csum_result, but
>>>         ( u64)expected == 55939 (0xda83)
>>>         ( u64)csum_result == 33754 (0x83da)
>>>     not ok 4 test_ip_fast_csum
>>>     # test_csum_ipv6_magic: ASSERTION FAILED at lib/checksum_kunit.c:617
>>>     Expected ( u64)expected_csum_ipv6_magic[i] == ( u64)csum_ipv6_magic(saddr, daddr, len, proto, csum), but
>>>         ( u64)expected_csum_ipv6_magic[i] == 6356 (0x18d4)
>>>         ( u64)csum_ipv6_magic(saddr, daddr, len, proto, csum) == 43586 (0xaa42)
>>>     not ok 5 test_csum_ipv6_magic
>>> # checksum: pass:0 fail:5 skip:0 total:5
>>> 
>>> Note that test_ip_fast_csum and test_csum_ipv6_magic fail on all big endian
>>> systems due to a bug in the test code, unrelated to this problem.
>>> 
>>> Analysis shows that the errors are seen only if the buffer is misaligned.
>>> Looking into arch/sh/lib/checksum.S, I found commit cadc4e1a2b4d2 ("sh:
>>> Handle calling csum_partial with misaligned data") which seemed to be
>>> related. Reverting that commit fixes the problem.
>>> This suggests that something may be wrong with that commit. Alternatively,
>>> of course, it may be possible that something is wrong with the qemu
>>> emulation, but that seems unlikely.
>> I have not run these tests before. Can you tell me how these are run,
>> so I can verify these reproduce on real hardware?
> 
> Enabling CONFIG_KUNIT and CONFIG_CHECKSUM_KUNIT on top of a working
> configuration should do the trick. Both can be built as module,
> so presumably one can build and load them separately. I have not tried
> that, though - I always build them into the kernel and boot the resulting
> image.
> 
> Hope this helps,
> Guenter
> 
>