Message-ID: <CABVgOSnJPKRqAdYtu2bhSxBEehF7BVQH6Rvj1U4xOcGWei8Lig@mail.gmail.com>
Date:   Thu, 16 Mar 2023 12:15:46 +0800
From:   David Gow <davidgow@...gle.com>
To:     Frank Rowand <frowand.list@...il.com>
Cc:     Stephen Boyd <sboyd@...nel.org>,
        Michael Turquette <mturquette@...libre.com>,
        linux-kernel@...r.kernel.org, linux-clk@...r.kernel.org,
        patches@...ts.linux.dev,
        Brendan Higgins <brendan.higgins@...ux.dev>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        "Rafael J . Wysocki" <rafael@...nel.org>,
        Richard Weinberger <richard@....at>,
        Anton Ivanov <anton.ivanov@...bridgegreys.com>,
        Johannes Berg <johannes@...solutions.net>,
        Vincent Whitchurch <vincent.whitchurch@...s.com>,
        Rob Herring <robh+dt@...nel.org>,
        Christian Marangi <ansuelsmth@...il.com>,
        Krzysztof Kozlowski <krzysztof.kozlowski+dt@...aro.org>,
        devicetree@...r.kernel.org, linux-um@...ts.infradead.org,
        linux-kselftest@...r.kernel.org, kunit-dev@...glegroups.com
Subject: Re: [PATCH 2/8] of: Enable DTB loading on UML for KUnit tests

On Thu, 16 Mar 2023 at 08:45, Frank Rowand <frowand.list@...il.com> wrote:
>
> On 3/15/23 16:35, Frank Rowand wrote:
> > On 3/15/23 02:04, David Gow wrote:
> >> On Tue, 14 Mar 2023 at 12:28, Frank Rowand <frowand.list@...il.com> wrote:
> >>>
> >>> On 3/13/23 11:02, Frank Rowand wrote:
> >>>> On 3/11/23 00:42, David Gow wrote:
> >>>>> On Sat, 11 Mar 2023 at 07:34, Stephen Boyd <sboyd@...nel.org> wrote:
> >>>>>>
> >>>>>> Quoting David Gow (2023-03-10 00:09:48)
> >>>>>>> On Fri, 10 Mar 2023 at 07:19, Stephen Boyd <sboyd@...nel.org> wrote:
> >>>>>>>>
> >>>>>>>>
> >>>>>>>> Hmm. I think you're suggesting that the unit test data be loaded
> >>>>>>>> whenever CONFIG_OF=y and CONFIG_KUNIT=y. Then tests can check for
> >>>>>>>> CONFIG_OF and skip if it isn't enabled?
> >>>>>>>>
> >>>>>>>
> >>>>>>> More of the opposite: that we should have some way of supporting tests
> >>>>>>> which might want to use a DTB other than the built-in one. Mostly for
> >>>>>>> non-UML situations where an actual devicetree is needed to even boot
> >>>>>>> far enough to get test output (so we wouldn't be able to override it
> >>>>>>> with a compiled-in test one).
> >>>>>>
> >>>>>> Ok, got it.
> >>>>>>
> >>>>>>>
> >>>>>>> I think moving to overlays probably will render this idea obsolete:
> >>>>>>> but the thought was to give test code a way to check for the required
> >>>>>>> devicetree nodes at runtime, and skip the test if they weren't found.
> >>>>>>> That way, the failure mode for trying to boot this on something which
> >>>>>>> required another device tree for, e.g., serial, would be "these tests
> >>>>>>> are skipped because the wrong device tree is loaded", not "I get no
> >>>>>>> output because serial isn't working".
> >>>>>>>
> >>>>>>> Again, though, it's only really needed for non-UML, and just loading
> >>>>>>> overlays as needed should be much more sensible anyway.
> >>>>>>
> >>>>>> I still have one niggle here. Loading overlays requires
> >>>>>> CONFIG_OF_OVERLAY, and the overlay loading API returns -ENOTSUPP when
> >>>>>> CONFIG_OF_OVERLAY=n. For now I'm checking for the config being enabled
> >>>>>> in each test, but I'm thinking it may be better to simply call
> >>>>>> kunit_skip() from the overlay loading function if the config is
> >>>>>> disabled. This way tests can simply call the overlay loading function
> >>>>>> and we'll halt the test immediately if the config isn't enabled.
> >>>>>>
> >>>>>
> >>>>> That sounds sensible, though there is a potential pitfall. If
> >>>>> kunit_skip() is called directly from overlay code, might introduce a
> >>>>> dependency on kunit.ko from the DT overlay, which we might not want.
> >>>>> The solution there is either to have a kunit wrapper function (so the
> >>>>> call is already in kunit.ko), or to have a hook to skip the current
> >>>>> test (which probably makes sense to do anyway, but I think the wrapper
> >>>>> is the better option).
> >>>>>
> >>>>>
> >>>>>>>
> >>>>>>>>>
> >>>>>>>>> That being said, I do think that there's probably some sense in
> >>>>>>>>> supporting the compiled-in DTB as well (it's definitely simpler than
> >>>>>>>>> patching kunit.py to always pass the extra command-line option in, for
> >>>>>>>>> example).
> >>>>>>>>> But maybe it'd be nice to have the command-line option override the
> >>>>>>>>> built-in one if present.
> >>>>>>>>
> >>>>>>>> Got it. I need to test loading another DTB on the commandline still, but
> >>>>>>>> I think this won't be a problem. We'll load the unittest-data DTB even
> >>>>>>>> with KUnit on UML, so assuming that works on UML right now it should be
> >>>>>>>> unchanged by this series once I resend.
> >>>>>>>
> >>>>>>> Again, moving to overlays should render this mostly obsolete, no? Or
> >>>>>>> am I misunderstanding how the overlay stuff will work?
> >>>>>>
> >>>>>> Right, overlays make it largely a moot issue. The way the OF unit tests
> >>>>>> work today is by grafting a DTB onto the live tree. I'm reusing that
> >>>>>> logic to graft a container node target for kunit tests to add their
> >>>>>> overlays too. It will be clearer once I post v2.
> >>>>>>
> >>>>>>>
> >>>>>>> One possible future advantage of being able to test with custom DTs at
> >>>>>>> boot time would be for fuzzing (provide random DT properties, see what
> >>>>>>> happens in the test). We've got some vague plans to support a way of
> >>>>>>> passing custom data to tests to support this kind of case (though, if
> >>>>>>> we're using overlays, maybe the test could just patch those if we
> >>>>>>> wanted to do that).
> >>>>>>
> >>>>>> Ah ok. I can see someone making a fuzzer that modifies devicetree
> >>>>>> properties randomly, e.g. using different strings for clock-names.
> >>>>>>
> >>>>>> This reminds me of another issue I ran into. I wanted to test adding the
> >>>>>> same platform device to the platform bus twice to confirm that the
> >>>>>> second device can't be added. That prints a warning, which makes
> >>>>>> kunit.py think that the test has failed because it printed a warning. Is
> >>>>>> there some way to avoid that? I want something like
> >>>>>>
> >>>>>>         KUNIT_EXPECT_WARNING(test, <call some function>)
> >>>>>>
> >>>>>> so I can test error cases.
> >>>>
> >>>> DT unittests already have a similar concept.  A test can report that a
> >>>> kernel warning (or any other specific text) either (1) must occur for the
> >>>> test to pass or (2) must _not_ occur for the test to pass.  The check
> >>>> for the kernel warning is done by the test output parsing program
> >>>> scripts/dtc/of_unittest_expect.
> >>>>
> >>>> The reporting by a test of an expected error in drivers/of/unittest.c
> >>>> is done by EXPECT_BEGIN() and EXPECT_END().  These have been in
> >>>> unittest for a long time.
> >>>>
> >>>> The reporting by a test of a not expected to occur error is done
> >>>> by EXPECT_NOT_BEGIN() and EXPECT_NOT_END().  These are added to
> >>>> unittest in linux 6.3-rc1.
> >>>>
> >>>> I discussed this concept in one of the early TAP / KTAP discussion
> >>>
> >>> The link to the early KTAP discussion on this concept is:
> >>>
> >>>    https://lore.kernel.org/all/d38bf9f9-8a39-87a6-8ce7-d37e4a641675@gmail.com/T/#u
> >>>
> >>>
> >>
> >> Thanks -- I'd totally forgotten about that!
> >>
> >
> >> I still personally would prefer a way of checking this from within the
> >> kernel, as if we're just printing out "EXPECT: " lines, then it's not
> >> possible to know if a test passes just from the raw results (and
> >> things like statistics can't be updated without a separate tool like
> >> kunit.py parsing the KTAP.
> >
> > Yes, I totally agree with that.  If there is a reasonable way to
> > implement.  But in the DT unittest world, I have not found a
> > reasonable way.  Adding hooks is suggested below, but for DT
> > unittest _I_ (opinion) do not find that reasonable.  I voice no
> > vote for kunit - that decision is up to the kunit crowd.
> >
> >>
> >> Indeed, my personal preference is that this log-based way of doing
> >> expectations is probably best kept as a last resort. i.e.,
> >> 1. Try to add a hook to the code which prints the message, which can
> >> then fail the test (or set a flag for the test to check later). This
> >> probably needs some better KUnit-side helpers to be truly ergonomic,
> >> but at least avoids too strict a dependency on the exact formatting of
> >> the log messages.
> >
> > I'm not a fan of hooks.  I see them as a maintenance burden, dependent
> > upon the source version of the object being tested, yet another
> > thing that can go wrong, and adds complexity to creating a test
> > environment and running the test.  Again, this just a personal
> > opinion, and I'm not voting for or against this for kunit.
> >

I definitely agree that they've got their problems, and aren't the
right solution for every test.

That being said, I think a few of those problems also apply to
expecting individual error lines, which are basically doing the same
thing, just less explicitly (though with the advantage of serving
another real-world purpose).

> >> 2. If that doesn't work, use console tracepoints or similar to
> >> implement an EXPECT_BEGIN() / EXPECT_END() or similar API entirely
> >> within the kernel.
> >
> > Isn't this just another hook?  So same opinion.
> >

I see what you mean. I guess the distinction I was trying to draw was
between a specific hook, implemented as such explicitly on both sides,
and a generic mechanism for console-message-based expectations.

So the difference is that, once the EXPECT_BEGIN()/EXPECT_END() macros
have been implemented this way, the uses of them should look the same.

> >> 3. Only if we can't come up with a working way of doing the former
> >> options, resort to adding "EXPECT:" lines and having a parser pick up
> >> on this.
>
> Adding one more thought here so I don't forget it before the topic picks
> up again in the KTAP version 2 world...
>
> The test parser could generate an artificial subtest test case status line
> or normal test case status line to report the result of the EXPECT.  This
> also is ugly because it is creating a new requirement on the parser vs the
> expectations in the KTAP plan line (the plan line could include the EXPECT
> in the number of tests count, but then the raw KTAP test output would be
> missing the artificial EXPECT test result).  No need to hash out details
> here, just a thought...
>

Yeah, I think if we have this done at the parser level, our options
are either that or an override.

For KUnit, I think the override makes much more sense, as logically,
any expectation would be considered part of an existing test case.
Maybe we could treat an override as a "subtest" of the current test,
but injecting a test case anywhere else would go pretty seriously
against the KUnit model.

> -Frank
>
> >
> > Again, don't let my opinion affect the voting between 1, 2, 3, or other
> > for kunit.
> >
> >>
> >> One of the downsides of doing "EXPECT" lines in KTAP is that it'll
> >> suddenly be much more dependent on the exact layout of the tests, as
> >> we'd need to be able to override a test result if an expectation fails
> >> (at least, to maintain the KUnit structure). And overriding a result
> >> which is already in the output seems really, really ugly.
> >
> > I don't understand "dependent on the exact layout of the tests".
> > If you are saying that the test result parser has to figure out
> > which test result to override, that has not been an issue in
> > the cases that I use EXPECTs in DT unittest.  The EXPECT begin and
> > EXPECT end have always immediately surrounded a single test, so
> > when the parser processes the EXPECT end, only the most recent
> > test result could be overridden.  This has worked because the
> > kernel warning and error messages have been from kernel action
> > that happens synchronously with the test.  If the test prods the
> > kernel in a way that results in the kernel performing an
> > asynchronous activity (eg in another thread), then it becomes
> > more complex to structure the EXPECT end -- I would imagine that the
> > test would have to block on the asynchronous activity just before
> > reporting the normal KTAP status result for the test (and the
> > EXPECT end would normally be just after reporting the KTAP
> > status result for the test).

Okay: I agree 100% with you there.

I think the difference I was thinking of is more whether expectations
are considered part of an existing test (which may need overriding),
or exist as their own separate test result (as you suggest in the
follow-up email above).

> >
> > I agree with overriding being ugly.  For the DT unittest results
> > parser, the EXPECT summary results are reported separately from
> > the individual test summary results.  The parser also flags the
> > EXPECT failure in line with the normal individual test result
> > lines.
> >
> > I see both parsing results as valid, and as a policy choice for
> > each test parser.
> >

I agree that each parser should have some leeway here, but do think we
want to make sure the results have some sensible, standardised
interpretation, so that we can use parsers interchangeably without
getting totally inconsistent results. That's the big advantage of
standardisation, after all.

My philosophical objection to overriding is that it's really confusing
to have an "ok" line in the results, indicating that a test has
passed, when the test has in fact failed (because the expectation
doesn't match). This gets worse when subtests are considered, and we
have these misleading results bubbled up to reported overall "suite"
results.

I guess one solution is to have an extra layer of parsing, which takes
raw kernel output, verifies the expectations, and then outputs
"processed KTAP", with all of the final results resolved. But that's
ugly in a way, too.

> >>
> >> There's a patch to the KASAN tests to move from doing option 1 to
> >> option 2 above (in order to better support RCU, which didn't work with
> >> the hook):
> >> https://lore.kernel.org/all/ebf96ea600050f00ed567e80505ae8f242633640.1666113393.git.andreyknvl@google.com/
> >>
> >>
> >>>> threads and expect to start a discussion thread on this specific
> >>>> topic in the KTAP Specification V2 context.  I expect the discussion
> >>>> to result in a different implementation than what DT unittests are
> >>>> using (bike shedding likely to ensue) but whatever is agreed to
> >>>> should be easy for DT to switch to.
> >>>
> >>> The link to the KTAP Specification Version 2 process and progress is:
> >>>
> >>>    https://elinux.org/Test_Results_Format_Notes#KTAP_version_2
> >>>
> >>
> >> Thanks! We've got a few more KTAP ideas to air, so will hopefully send
> >> those out soon!
> >
> > Glad to hear, I'm hoping that process starts progressing a bit.
> >

Yeah. I'll shift further discussion of this to a KTAP proposal: I
don't want to derail this thread too much further.

We'll keep looking at fully in-kernel ways of achieving similar things
in the meantime.

Cheers,
-- David

> > -Frank
> >
> >>
> >> Cheers,
> >> -- David
> >>
> >>>>
> >>>>>
> >>>>> Hmm... I'd've thought that shouldn't be a problem: kunit.py should
> >>>>> ignore most messages during a test, unless it can't find a valid
> >>>>> result line. What does the raw KTAP output look like? (You can get it
> >>>>> from kunit.py by passing the --raw_output option).
> >>>>>
> >>>>> That being said, a KUNIT_EXPECT_LOG_MESSAGE() or similar is something
> >>>>> we've wanted for a while. I think that the KASAN folks have been
> >>>>> working on something similar using console tracepoints:
> >>>>> https://lore.kernel.org/all/ebf96ea600050f00ed567e80505ae8f242633640.1666113393.git.andreyknvl@google.com/
> >>>>>
> >>>>> Cheers,
> >>>>> -- David
> >>>>
> >>>
> >
>

Download attachment "smime.p7s" of type "application/pkcs7-signature" (4003 bytes)

Powered by blists - more mailing lists