On Tue, Dec 1, 2020 at 9:21 AM Arpitha Raghunandan <[email protected]> wrote:
> Convert test lib/test_hexdump.c to KUnit. More information about
> KUnit can be found at:
> https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html.
> KUnit provides a common framework for unit tests in the kernel.
> KUnit and kselftest are standardizing around KTAP, converting this
> test to KUnit makes this test output in KTAP which we are trying to
> make the standard test result format for the kernel.
Below doesn't suit commit message, perhaps adding it after '---' line
would be good. In the commit message you can choose one failed case
followed by all OK and show the difference.
> I ran both the original and converted tests as is to produce the
> output for success of the test in the two cases. I also ran these
> tests with a small modification to show the difference in the output
> for failure of the test in both cases. The modification I made is:
> static const char * const test_data_4_le[] __initconst = {
> - "7bdb32be", "b293180a", "24c4ba70", "9b34837d",
> + "7bdb32be", "b293180a", "24c4ba70", "9b3483d",
>
> The difference in outputs can be seen here:
> https://gist.github.com/arpi-r/38f53a3c65692bf684a6bf3453884b6e
Looks pretty much good, what I'm sad to see is the absence of the test
statistics. Any ideas if we can get it back?
--
With Best Regards,
Andy Shevchenko
On 01/12/20 4:36 pm, Andy Shevchenko wrote:
> On Tue, Dec 1, 2020 at 9:21 AM Arpitha Raghunandan <[email protected]> wrote:
>> Convert test lib/test_hexdump.c to KUnit. More information about
>> KUnit can be found at:
>> https://www.kernel.org/doc/html/latest/dev-tools/kunit/index.html.
>> KUnit provides a common framework for unit tests in the kernel.
>> KUnit and kselftest are standardizing around KTAP, converting this
>> test to KUnit makes this test output in KTAP which we are trying to
>> make the standard test result format for the kernel.
>
> Below doesn't suit commit message, perhaps adding it after '---' line
> would be good. In the commit message you can choose one failed case
> followed by all OK and show the difference.
>
Okay, I will make this change.
>> I ran both the original and converted tests as is to produce the
>> output for success of the test in the two cases. I also ran these
>> tests with a small modification to show the difference in the output
>> for failure of the test in both cases. The modification I made is:
>> static const char * const test_data_4_le[] __initconst = {
>> - "7bdb32be", "b293180a", "24c4ba70", "9b34837d",
>> + "7bdb32be", "b293180a", "24c4ba70", "9b3483d",
>>
>> The difference in outputs can be seen here:
>> https://gist.github.com/arpi-r/38f53a3c65692bf684a6bf3453884b6e
>
> Looks pretty much good, what I'm sad to see is the absence of the test
> statistics. Any ideas if we can get it back?
>
I could again include variable total_tests as was in the original test.
Would that be fine?
Thanks!
On Wed, Dec 02, 2020 at 09:51:19AM +0530, Arpitha Raghunandan wrote:
> On 01/12/20 4:36 pm, Andy Shevchenko wrote:
> > On Tue, Dec 1, 2020 at 9:21 AM Arpitha Raghunandan <[email protected]> wrote:
...
> >> I ran both the original and converted tests as is to produce the
> >> output for success of the test in the two cases. I also ran these
> >> tests with a small modification to show the difference in the output
> >> for failure of the test in both cases. The modification I made is:
> >> static const char * const test_data_4_le[] __initconst = {
> >> - "7bdb32be", "b293180a", "24c4ba70", "9b34837d",
> >> + "7bdb32be", "b293180a", "24c4ba70", "9b3483d",
> >>
> >> The difference in outputs can be seen here:
> >> https://gist.github.com/arpi-r/38f53a3c65692bf684a6bf3453884b6e
> >
> > Looks pretty much good, what I'm sad to see is the absence of the test
> > statistics. Any ideas if we can get it back?
> >
>
> I could again include variable total_tests as was in the original test.
> Would that be fine?
What I;m talking about is the output. How it will be implemented (using the
same variable or differently) is up to you. So the point is I want to see the
statistics of success/total at the end.
I think this should be done in KUNIT rather than in the individual test cases.
--
With Best Regards,
Andy Shevchenko
On Wed, Dec 2, 2020 at 6:06 PM Andy Shevchenko
<[email protected]> wrote:
>
> On Wed, Dec 02, 2020 at 09:51:19AM +0530, Arpitha Raghunandan wrote:
> > On 01/12/20 4:36 pm, Andy Shevchenko wrote:
> > > On Tue, Dec 1, 2020 at 9:21 AM Arpitha Raghunandan <[email protected]> wrote:
>
> ...
>
> > >> I ran both the original and converted tests as is to produce the
> > >> output for success of the test in the two cases. I also ran these
> > >> tests with a small modification to show the difference in the output
> > >> for failure of the test in both cases. The modification I made is:
> > >> static const char * const test_data_4_le[] __initconst = {
> > >> - "7bdb32be", "b293180a", "24c4ba70", "9b34837d",
> > >> + "7bdb32be", "b293180a", "24c4ba70", "9b3483d",
> > >>
> > >> The difference in outputs can be seen here:
> > >> https://gist.github.com/arpi-r/38f53a3c65692bf684a6bf3453884b6e
> > >
> > > Looks pretty much good, what I'm sad to see is the absence of the test
> > > statistics. Any ideas if we can get it back?
> > >
> >
> > I could again include variable total_tests as was in the original test.
> > Would that be fine?
>
> What I;m talking about is the output. How it will be implemented (using the
> same variable or differently) is up to you. So the point is I want to see the
> statistics of success/total at the end.
>
> I think this should be done in KUNIT rather than in the individual test cases.
I tend to agree here that this really is something for KUnit. At the
moment, the tools/testing/kunit/kunit.py script will parse the kernel
log and generate these sorts of statistics. I know that needing to run
it through a script might seem like a step backwards, but there's no
formal place for statistics in the KTAP specification[1] being worked
on to standardise kselftest/kunit output formats. Note that there are
other parsers for TAP-like formats which are being used with KUnit
results, so systems like LAVA could also sum up these statistics. It's
also possible, as Arpitha alluded to, to have the test dump them out
as a comment.
This won't actually work for this test as-is, though, as the KUnit
version is running as a single giant test case (so KUnit believes that
1/1 tests have passed, rather than having any more-detailed
statistics). It looks like there are a few ways to split it up a bit
which would make it neater (a test each for the for() loops in
test_hexdump_init() seems sensible to me), but at the moment, there's
not really a way of programmatically generating test cases which KUnit
then counts
The "Parameterised Tests"[2] work Arpitha has been working on ought to
go some way to helping here, though it won't solve this completely in
this initial version. The problem there is that parameterised tests
are not reported individually in a way the kunit.py parser can report
cleanly, yet, so it'll still only be counted as one test until that's
changed (though, at least, that shouldn't require any test-specific
work).
My suggestion for the ultimate state of the test would be:
- Split up the test into separate KUnit tests for the different
"categories" of tests: (e.g., test_hexdump_set,
test_hexdump_overflow_set_ascii, etc)
- Replace the for loops in test_hexdump_init() with parameters, so
that KUnit is aware of the original runs.
- Once KUnit and the tooling supports it, these will be reported as
subtests. (In the meantime, the results will be listed individually,
commented out)
Of course, it'll take a while before all of those KUnit pieces are in
place. I personally think that a good compromise would be to just do
the first of these for now, which would make kunit_tool give at least
a 4/4 rather than 1/1 result. Then, once the parameterised testing
work is merged (and perhaps the tooling fixes are finished), the tests
could be updated to take advantage of that.
Cheres,
-- David
[1]: https://lore.kernel.org/linux-kselftest/CY4PR13MB1175B804E31E502221BC8163FD830@CY4PR13MB1175.namprd13.prod.outlook.com/T/
[2]: https://lore.kernel.org/linux-kselftest/[email protected]/
On Wed, Dec 2, 2020 at 1:57 PM David Gow <[email protected]> wrote:
> On Wed, Dec 2, 2020 at 6:06 PM Andy Shevchenko
> <[email protected]> wrote:
> > On Wed, Dec 02, 2020 at 09:51:19AM +0530, Arpitha Raghunandan wrote:
...
> > What I;m talking about is the output. How it will be implemented (using the
> > same variable or differently) is up to you. So the point is I want to see the
> > statistics of success/total at the end.
> >
> > I think this should be done in KUNIT rather than in the individual test cases.
>
> I tend to agree here that this really is something for KUnit. At the
> moment, the tools/testing/kunit/kunit.py script will parse the kernel
> log and generate these sorts of statistics. I know that needing to run
> it through a script might seem like a step backwards, but there's no
> formal place for statistics in the KTAP specification[1] being worked
> on to standardise kselftest/kunit output formats.
Then it sucks. Fix specification (in a long term) and does it have a
comment style of messages that we can have this statistics printed
(but maybe not parsed)?
> Note that there are
> other parsers for TAP-like formats which are being used with KUnit
> results, so systems like LAVA could also sum up these statistics. It's
> also possible, as Arpitha alluded to, to have the test dump them out
> as a comment.
Fine to me.
> This won't actually work for this test as-is, though, as the KUnit
> version is running as a single giant test case (so KUnit believes that
> 1/1 tests have passed, rather than having any more-detailed
> statistics). It looks like there are a few ways to split it up a bit
> which would make it neater (a test each for the for() loops in
> test_hexdump_init() seems sensible to me), but at the moment, there's
> not really a way of programmatically generating test cases which KUnit
> then counts
Fix it, please. We rely on this statistics pretty much.
> The "Parameterised Tests"[2] work Arpitha has been working on ought to
> go some way to helping here, though it won't solve this completely in
> this initial version. The problem there is that parameterised tests
> are not reported individually in a way the kunit.py parser can report
> cleanly, yet, so it'll still only be counted as one test until that's
> changed (though, at least, that shouldn't require any test-specific
> work).
>
> My suggestion for the ultimate state of the test would be:
> - Split up the test into separate KUnit tests for the different
> "categories" of tests: (e.g., test_hexdump_set,
> test_hexdump_overflow_set_ascii, etc)
> - Replace the for loops in test_hexdump_init() with parameters, so
> that KUnit is aware of the original runs.
> - Once KUnit and the tooling supports it, these will be reported as
> subtests. (In the meantime, the results will be listed individually,
> commented out)
I'm fine as long as we have this information printed to the user.
> Of course, it'll take a while before all of those KUnit pieces are in
> place. I personally think that a good compromise would be to just do
> the first of these for now, which would make kunit_tool give at least
> a 4/4 rather than 1/1 result. Then, once the parameterised testing
> work is merged (and perhaps the tooling fixes are finished), the tests
> could be updated to take advantage of that.
How can we guarantee it will be not forgotten?
> [1]: https://lore.kernel.org/linux-kselftest/CY4PR13MB1175B804E31E502221BC8163FD830@CY4PR13MB1175.namprd13.prod.outlook.com/T/
> [2]: https://lore.kernel.org/linux-kselftest/[email protected]/
--
With Best Regards,
Andy Shevchenko
On Wed, Dec 2, 2020 at 8:21 PM Andy Shevchenko
<[email protected]> wrote:
>
> On Wed, Dec 2, 2020 at 1:57 PM David Gow <[email protected]> wrote:
> > On Wed, Dec 2, 2020 at 6:06 PM Andy Shevchenko
> > <[email protected]> wrote:
> > > On Wed, Dec 02, 2020 at 09:51:19AM +0530, Arpitha Raghunandan wrote:
>
> ...
>
> > > What I;m talking about is the output. How it will be implemented (using the
> > > same variable or differently) is up to you. So the point is I want to see the
> > > statistics of success/total at the end.
> > >
> > > I think this should be done in KUNIT rather than in the individual test cases.
> >
> > I tend to agree here that this really is something for KUnit. At the
> > moment, the tools/testing/kunit/kunit.py script will parse the kernel
> > log and generate these sorts of statistics. I know that needing to run
> > it through a script might seem like a step backwards, but there's no
> > formal place for statistics in the KTAP specification[1] being worked
> > on to standardise kselftest/kunit output formats.
>
> Then it sucks. Fix specification (in a long term) and does it have a
> comment style of messages that we can have this statistics printed
> (but maybe not parsed)?
I should clarify: there's nothing in the spec which explicitly defines
a place for such statistics (nor anything which requires them). There
are "diagnostic" lines which are not parsed, and so it'd be possible
to output statistics there. KUnit itself doesn't at present, but
allows individual tests to log diagnostic lines which could be such
statistics, particularly in cases like this where the full structure
of the tests aren't quite exposed to the framework.
> > Note that there are
> > other parsers for TAP-like formats which are being used with KUnit
> > results, so systems like LAVA could also sum up these statistics. It's
> > also possible, as Arpitha alluded to, to have the test dump them out
> > as a comment.
>
> Fine to me.
>
> > This won't actually work for this test as-is, though, as the KUnit
> > version is running as a single giant test case (so KUnit believes that
> > 1/1 tests have passed, rather than having any more-detailed
> > statistics). It looks like there are a few ways to split it up a bit
> > which would make it neater (a test each for the for() loops in
> > test_hexdump_init() seems sensible to me), but at the moment, there's
> > not really a way of programmatically generating test cases which KUnit
> > then counts
>
> Fix it, please. We rely on this statistics pretty much.
The hope is that the Parameterised Test feature will make this
possible (though, as mentioned, there are a few other issues around
then making those statistics available, but we should be able to work
through those).
It may be a silly question, but what is it that makes these statistics
useful in this test? Maybe I'm misunderstanding something, but I'd've
thought that the important things were whether or not _all_ tests had
passed, and -- if not --- _which_ ones had failed. Is the count of
failing cases within a test like this really that useful for
debugging, or is it more for comparing against different versions?
Either way, we'll try to make sure they're available.
> > The "Parameterised Tests"[2] work Arpitha has been working on ought to
> > go some way to helping here, though it won't solve this completely in
> > this initial version. The problem there is that parameterised tests
> > are not reported individually in a way the kunit.py parser can report
> > cleanly, yet, so it'll still only be counted as one test until that's
> > changed (though, at least, that shouldn't require any test-specific
> > work).
> >
> > My suggestion for the ultimate state of the test would be:
> > - Split up the test into separate KUnit tests for the different
> > "categories" of tests: (e.g., test_hexdump_set,
> > test_hexdump_overflow_set_ascii, etc)
> > - Replace the for loops in test_hexdump_init() with parameters, so
> > that KUnit is aware of the original runs.
> > - Once KUnit and the tooling supports it, these will be reported as
> > subtests. (In the meantime, the results will be listed individually,
> > commented out)
>
> I'm fine as long as we have this information printed to the user.
Okay -- Arpitha: does this seem like a sensible approach to you?
If printing it to the kernel log is really essential, I'll have a look
into how we can do that in KUnit.
> > Of course, it'll take a while before all of those KUnit pieces are in
> > place. I personally think that a good compromise would be to just do
> > the first of these for now, which would make kunit_tool give at least
> > a 4/4 rather than 1/1 result. Then, once the parameterised testing
> > work is merged (and perhaps the tooling fixes are finished), the tests
> > could be updated to take advantage of that.
>
> How can we guarantee it will be not forgotten?
Thinking about it further, if we can get the parameterised testing
stuff in 5.11 as planned, then any tooling/output fixes done later
would automatically apply. Maybe rather than doing an intermediate
version with just the first splitting up of tests, we try it with the
current parameterised test patches, and we can possibly prototype some
kernel-side statistics reporting which should work. To be honest,
though, the subtest support on the kunit_tool side is likely to take
quite a while, so it would be nice to have something (like statistics
in the kernel log) which ameliorate the problem in the meantime.
I'll have a bit of a play around with the test output and parsing code
this week and see if there's a simple change that can get us most of
the way there. I think something should be possible if the test uses
the Parameterised testing feature.
Cheers,
-- David