Re: [PATCH v4 3/3] lib/test_crc: Add test cases for crc calculation

[Eric Biggers]

On Wed, Jul 25, 2018 at 12:28:15AM +0800, Coly Li wrote:
> On 2018/7/24 12:44 PM, Eric Biggers wrote:
> > On Thu, Jul 19, 2018 at 12:55:45AM +0800, Coly Li wrote:
> >> This patch adds a kernel module to test the consistency of multiple crc
> >> calculation in Linux kernel. It is enabled with CONFIG_TEST_CRC enabled.
> >>
> >> The test results are printed into kernel message, which look like,
> >>
> >> test_crc: crc64_be: FAILED (0x03d4d0d85685d9a1, expected 0x3d4d0d85685d9a1f)
> >>
> >> kernel 0day system has framework to check kernel message, then the above
> >> result can be handled by 0day system. If crc calculation inconsistency
> >> happens, it can be detected quite soon.
> >>
> >> lib/test_crc.c is a testing frame work for many crc consistency
> >> testings. For now, there is only one test caes for crc64_be().
> > 
> > Are you aware there's already a CRC-32 test module: CONFIG_CRC32_SELFTEST and
> > lib/crc32test.c?  Confusingly, your patch uses a different naming convention for
> > the new CRC-64 one, and puts the Kconfig option in a different place, and makes
> > it sound like it's a generic test for all CRC implementations rather than just
> > the CRC-64 one.  Please use the existing convention (i.e. add
> > CONFIG_CRC64_SELFTEST and lib/crc64test.c) unless you have a strong argument for
> > why it should be done differently.
> > 
> > (And I don't think it makes sense to combine all CRC tests into one module,
> > since you should be able to e.g. enable just CRC32 and CRC32_SELFTEST without
> > also pulling in a dependency on all the other CRC variants.)
> > 
> 
> Hi Eric,
> 
> The purpose of test_crc is to provide a unified crc calculation
> consistency testing for 0day. So far it is only crc64, and I will add
> more test cases later. I see there is crc-32 test module, which does
> more testing then consistency check, and no unified format for 0day
> system to detect. This is why people suggested me to add this test
> framework.
> 

Actually the code in crc32test is nearly the same as what you're adding for
CRC-64.  The CRC-32 test is longer because it's testing two different
polynomials "crc32" and "crc32c" as well as combining CRC's; neither of those is
relevant for CRC-64 yet, as you've implemented just one polynomial and there is
no function provided to combine CRC64's yet.  The CRC-32 test also tests
performance, but if you don't believe CRC performance should be tested, then you
should remove the performance test from the existing module rather than
implementing a brand new test module just to remove the performance test...

I still don't understand why you decided to do things differently for CRC-64,
when there were already CRC-32 tests that used a certain convention for the
Kconfig option, filename, etc.  It's inconsistent and confusing.  Again, please
use the existing convention unless you have a strong argument for why it should
be done differently.  (And if you do want to do things differently, the existing
test should be converted first.)

- Eric