Re: [PATCH v6 1/2] kunit: Support for Parameterized Testing

[Marco Elver]

On Thu, Nov 12, 2020 at 04:18PM +0800, David Gow wrote:
> On Thu, Nov 12, 2020 at 12:55 AM Bird, Tim <Tim.Bird@xxxxxxxx> wrote:
[...]
> > > kunit_tool has a bug when parsing the comments / diagnostic lines,
> > > which requires a ": " to be present. This is a bug, which is being
> > > fixed[1], so while it's _nice_ to not trigger it, it's not really an
> > > important long-term goal of the format. In any case, this kunit_tool
> > > issue only affects the comment lines: if the per-parameter result line
> > > is an actual result, rather than just a diagnostic, this shouldn't be
> > > a problem.
> > >
> > > In any case, I still prefer my proposed option for now -- noting that
> > > these per-parameter results are not actually supposed to be parsed --
> > > with then the possibility of expanding them to actual nested results
> > > later should we wish. But if the idea of having TAP-like lines in
> > > diagnostics seems too dangerous (e.g. because people will try to parse
> > > them anyway), then I think the options we have are to stick to the
> > > output format given in the current version of this patch (which
> > > doesn't resemble a TAP result), make each parameterised version its
> > > own test (without a "container test", which would require a bit of
> > > extra work while computing test plans), or to hold this whole feature
> > > back until we can support arbitrary test hierarchies in KUnit.
> > It seems like you're missing a 4th option, which is just tack the parameter
> > name on, without using a colon, and have these testcases treated
> > as unique within the context of the super-test.  Is there some reason
> > these can't be expressed as individual testcases in the parent test?
> >
> 
> No: there's no fundamental reason why we couldn't do that, though
> there are some things which make it suboptiomal, IMHO.
>
> Firstly, there could be a lot of parameters, and that by not grouping
> them together it could make dealing with the results a little
> unwieldy. The other side of that is that it'll result in tests being
> split up and renamed as they're ported to use this, whereas if the
> whole test shows up once (with subtests or without), the old test name
> can still be there, with a single PASS/FAIL for the whole test.

Agree, it's suboptimal and having the parameterized not be absorbed by
the whole suite would be strongly preferred.

> (It also might be a little tricky with the current implementation to
> produce the test plan, as the parameters come from a generator, and I
> don't think there's a way of getting the number of parameters ahead of
> time. That's a problem with the sub-subtest model, too, though at
> least there it's a little more isolated from other tests.)

The whole point of generators, as I envisage it, is to also provide the
ability for varying parameters dependent on e.g. environment,
configuration, number of CPUs, etc. The current array-based generator is
the simplest possible use-case.

However, we *can* require generators generate a deterministic number of
parameters when called multiple times on the same system.

To that end, I propose a v7 (below) that takes care of getting number of
parameters (and also displays descriptions for each parameter where
available).

Now it is up to you how you want to turn the output from diagnostic
lines into something TAP compliant, because now we have the number of
parameters and can turn it into a subsubtest. But I think kunit-tool
doesn't understand subsubtests yet, so I suggest we take these patches,
and then somebody can prepare kunit-tool.

Or did I miss something else?

> > > Personally, I'd rather not hold this feature back, and prefer to have
> > > a single combined result available, so would just stick with v6 if
> > > so...
> > >
> > > Does that make sense?
> >
> > I understand what you are saying, but this seems like a step backwards.
> > We already know that having just numbers to represent a test case is not
> > very human friendly. The same will go for these parameter case numbers.
> > I admit to not following this kunit test parameterization effort, so I don't
> > know the background of how the numbers relate to the parameters.
> > But there must be some actual semantic meaning to each of the parameter
> > cases.  Not conveying that meaning as part of the test case name seems like
> > a missed opportunity.
> 
> Yeah: I'm not a big fan of just numbering the parameters either: the
> plan is to eventually support naming these. Basically, the goal is to
> be able to run the same test code repeatedly with different inputs
> (which may be programmatically generated): depending on the testcase /
> parameter generator in question, the numbers may mean something
> specific, but it's not necessarily the case. Certainly in most cases,
> the order of these parameters is unlikely to matter, so having the
> number be part of the test name isn't ideal there: it's unlikely to
> have semantic meaning, and worst-case could be unstable due to code
> changes.

We can name them. Probably a good idea to do it while we can, as I think
the best design is changing the generator function signature.

> > I'm at a little of a loss as to why, if you have valid testcase results, you would
> > shy away from putting them into a format that is machine-readable.  If it's because
> > the tooling is not there, then IMHO you should fix the tooling.
> 
> I think the real disconnect here is the ambiguity between treating
> each run-through with a different parameter as its own test case,
> versus an implementation detail of the single "meta testcase". Since
> parameters are not really named/ordered properly, (and the example is
> replacing a single test) it feels more like an implementation detail
> to me.
> 
> > I realize that perfect is the enemy of good enough, and that there's value for humans
> > to see these testcase results and manually interpret them, even if they are put into
> > a syntax that automated parsers will ignore.  However, I do think there's a danger that
> > this syntax will get canonicalized. Personally, I'd rather see the testcases
> > with parameters show up in the parsable results.  I'd rather sacrifice the hierarchy
> > than the results.
> 
> With the state of things at the moment, I don't think the individual
> results for given parameters are as useful as the overall result for
> the test (run over all parameters), so for me the hierarchy is more
> important than the actual results. There are certainly a lot of things
> we can do to make the results more useful (supporting named
> parameters, for one), and actually supporting the extra level of
> nesting in the tooling would make it possible to have both.

Named parameters are supported in my proposed v7 (see below). I think
it's now up to you what the format should be, as now it's just a matter
of changing 2 prints' formats to some non-diagnostic TAP compliant
output (but I have no idea what ;-)).

> There is of course another possibility -- to just not print the
> individual parameter results at all (the parameters will likely show
> up in the assertion messages of failures anyway -- especially if, as
> in the example, the _MSG() variants are used). That's probably
> slightly easier to read than a huge list of parameters which are all
> "ok" anyway...

To me this is not acceptable, as I'd like to see some progress feedback
for long tests.
 
> In any case, I'm happy to leave the final decision here to Arpitha and
> Marco, so long as we don't actually violate the TAP/KTAP spec and
> kunit_tool is able to read at least the top-level result. My
> preference is still to go either with the "# [test_case->name]:
> [ok|not ok] [index] - param-[index]", or to get rid of the
> per-parameter results entirely for now (or just print out a diagnostic
> message on failure). In any case, it's a decision we can revisit once
> we have support for named parameters, better tooling, or a better idea
> of how people are actually using this.

Right, so I think we'll be in a better place if we implement: 1)
parameter to description conversion support, 2) counting parameters. So
I decided to see what it looks like, and it wasn't too bad. I just don't
know how you want to fix kunit-tool to make these non-diagnostic lines
and not complain, but as I said, it'd be good to not block these
patches.

It currently looks like this:

| TAP version 14
| 1..6
|     # Subtest: ext4_inode_test
|     1..1
|     # inode_test_xtimestamp_decoding: Test with 1..16 params
|     # inode_test_xtimestamp_decoding: ok 1 - [1901-12-13 Lower bound of 32bit < 0 timestamp, no extra bits]
|     # inode_test_xtimestamp_decoding: ok 2 - [1969-12-31 Upper bound of 32bit < 0 timestamp, no extra bits]
|     # inode_test_xtimestamp_decoding: ok 3 - [1970-01-01 Lower bound of 32bit >=0 timestamp, no extra bits]
|     # inode_test_xtimestamp_decoding: ok 4 - [2038-01-19 Upper bound of 32bit >=0 timestamp, no extra bits]
|     # inode_test_xtimestamp_decoding: ok 5 - [2038-01-19 Lower bound of 32bit <0 timestamp, lo extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 6 - [2106-02-07 Upper bound of 32bit <0 timestamp, lo extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 7 - [2106-02-07 Lower bound of 32bit >=0 timestamp, lo extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 8 - [2174-02-25 Upper bound of 32bit >=0 timestamp, lo extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 9 - [2174-02-25 Lower bound of 32bit <0 timestamp, hi extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 10 - [2242-03-16 Upper bound of 32bit <0 timestamp, hi extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 11 - [2242-03-16 Lower bound of 32bit >=0 timestamp, hi extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 12 - [2310-04-04 Upper bound of 32bit >=0 timestamp, hi extra sec bit on]
|     # inode_test_xtimestamp_decoding: ok 13 - [2310-04-04 Upper bound of 32bit>=0 timestamp, hi extra sec bit 1. 1 ns]
|     # inode_test_xtimestamp_decoding: ok 14 - [2378-04-22 Lower bound of 32bit>= timestamp. Extra sec bits 1. Max ns]
|     # inode_test_xtimestamp_decoding: ok 15 - [2378-04-22 Lower bound of 32bit >=0 timestamp. All extra sec bits on]
|     # inode_test_xtimestamp_decoding: ok 16 - [2446-05-10 Upper bound of 32bit >=0 timestamp. All extra sec bits on]
|     ok 1 - inode_test_xtimestamp_decoding
| ok 1 - ext4_inode_test

Changing the format of the 2 prints to something else TAP-compliant
should be easy enough once kunit-tool supports subsubtests. :-)

I hope this is a reasonable compromise for now.

[Arpitha: I suggest waiting a day or two and see what further comments
 we get, and then take the 2 patches and send the v7 if we decide this is
 what we want.]

Thanks,
-- Marco

------ >8 ------