Re: [PATCH 1/6] RDMA/rxe: handle ICRC correctly on big endian systems

[Zhu Yanjun <yanjun.zhu@linux.dev>]

在 2025/1/29 19:30, Jason Gunthorpe 写道:
> On Wed, Jan 29, 2025 at 10:44:39AM +0100, Zhu Yanjun wrote:
>> 在 2025/1/27 23:38, Eric Biggers 写道:
>>> From: Eric Biggers <ebiggers@google.com>
>>>
>>> The usual big endian convention of InfiniBand does not apply to the
>>> ICRC field, whose transmission is specified in terms of the CRC32
>>> polynomial coefficients.
> 
> This patch is on to something but this is not a good explanation.
> 
> The CRC32 in IB is stored as big endian and computed in big endian,
> the spec says so explicitly:
> 
> 2) The CRC calculation is done in big endian byte order with the least
>     31 significant bit of the most significant byte being the first
>     bits in the 32 CRC calculation.
> 
> In this context saying it is not "big endian" doesn't seem to be quite
> right..
> 
> The spec gives a sample data packet (in offset/value pairs):
> 
> 0  0xF0 15 0xB3 30 0x7A 45 0x8B
> 1  0x12 16 0x00 31 0x05 46 0xC0
> 2  0x37 17 0x0D 32 0x00 47 0x69
> 3  0x5C 18 0xEC 33 0x00 48 0x0E
> 4  0x00 19 0x2A 34 0x00 49 0xD4
> 5  0x0E 20 0x01 35 0x0E 50 0x00
> 6  0x17 21 0x71 36 0xBB 51 0x00
> 7  0xD2 22 0x0A 37 0x88
> 8  0x0A 23 0x1C 38 0x4D
> 9  0x20 24 0x01 39 0x85
> 10 0x24 25 0x5D 40 0xFD
> 11 0x87 26 0x40 41 0x5C
> 12 0xFF 27 0x02 42 0xFB
> 13 0x87 28 0x38 43 0xA4
> 14 0xB1 29 0xF2 44 0x72
> 
> If you feed that to the CRC32 you should get 0x9625B75A in a CPU
> register u32.
> 
> cpu_to_be32() will put it in the right order for on the wire.
> 
> Since rxe doesn't have any cpu_to_be32() on this path, I'm guessing
> the Linux CRC32 implementations gives a u32 with the
> value = 0x5AB72596 ie swapped.
> 
> Probably the issue here is that the Linux CRC32 and the IBTA CRC32 are
> using a different mapping of LFSR bits. I love CRCs. So many different
> ways to implement the same thing.
> 
> Thus, I guess, the code should really read:
>   linux_crc32 = swab32(be32_to_cpu(val));
> 
> Which is a NOP on x86 and requires a swap on BE.
> 
> Zhu, can you check it for Eric? (this is page 229 in the spec).

Got it. In my IBTA spec, I can find what you mentioned in the IBTA spec.
Your explanation is in details and very nice.
Thanks a lot.

Zhu Yanjun

> 
> I assume the Linux CRC32 always gives the same CPU value regardless of
> LE or BE?
> 
> Jason