-----Original Message-----
From: Kees Cook <redacted>

[+kernelci, +Guillaume]

Hi!

Please keep me in CC for these kinds of proposals. And thanks for looking
at this again! Please understand that while I may be coming across as
rather negative here, I would like to have a rational and documented
specification for the kernel's test output, too. My main objection here
is that we already _have_ a specification, and it's already being parsed
by machines, so making changes without strong justification is going to
be met with resistance. :) So, with that in mind, on to my reply...

On Tue, Aug 10, 2021 at 04:25:03PM -0700, Rae Moar wrote:

quoted

We are looking to further standardise the output format used by kernel
test frameworks like kselftest and KUnit. Thus far we have used the
TAP (Test Anything Protocol) specification, but it has been extended
in many different ways, so we would like to agree on a common "Kernel
TAP" (KTAP) format to resolve these differences. Thus, below is a
draft of a specification of KTAP. Note that this specification is
largely based on the current format of test results for KUnit tests.

The kernel's largest producer of TAP is kselftest, and the largest
consumer is LAVA[1]. I would want buy-in from at least those responsible
for either side of those two things. (And I'm one of the people working
on both sides of it.)

The fundamental purpose of the kernel's TAP is to have many independent
tests runnable and parseable, specifically without any embedded framework
knowledge (or, even, any knowledge of TAP).

The tests run by kselftest come from 2 different TAP-producing harnesses
(kselftest.h for C, kselftest/runner.sh for TAP-agnostic tests) as well
as several hand-rolled instances in Shell, Python, and C. There used to
be more, but I have been steadily fixing their syntax[2] and merging[3]
separate implementations for a while now.

[1] https://github.com/Linaro/test-definitions/commit/8bd338bbcfa5a03efcf1d12e25b5d341d5a29cbc
[2] https://git.kernel.org/linus/b0df366bbd701c45e93af0dcb87ce22398589d1d
[3] https://git.kernel.org/linus/e80068be21824e4d2726eeea41cac6178d212415

quoted

Additionally, this specification was heavily inspired by the KTAP
specification draft by Tim Bird
(https://lore.kernel.org/linux-

kselftest/CY4PR13MB1175B804E31E502221BC8163FD830@CY4PR13MB1175.namprd13.prod.outlook.com/T/).

quoted

However, there are some notable differences to his specification. One
such difference is the format of nested tests is more fully specified
in the following specification. However, they are specified in a way
which may not be compatible with many kselftest nested tests.

I commented extensively on that thread. :)

quoted

=====================
Specification of KTAP
=====================

TAP, or the Test Anything Protocol is a format for specifying test
results used by a number of projects. It's website and specification
are found at: https://testanything.org/. The Linux Kernel uses TAP
output for test results. However, KUnit (and other Kernel testing
frameworks such as kselftest) have some special needs for test results
which don't gel perfectly with the original TAP specification. Thus, a
"Kernel TAP" (KTAP) format is specified to extend and alter TAP to
support these use-cases.

KTAP Output consists of 5 major elements (all line-based):
- The version line
- Plan lines
- Test case result lines

These are required.

quoted

- Diagnostic lines

This is optional.

quoted

- A bail out line

Bail out should be optional, and possibly not included at all. (See
below.)

quoted

An important component in this specification of KTAP is the
specification of the format of nested tests. This can be found in the
section on nested tests below.

The version line
----------------

The first line of KTAP output must be the version line. As this
specification documents the first version of KTAP,  the recommended
version line is "KTAP version 1". However, since all kernel testing
frameworks use TAP version lines, "TAP version 14" and "TAP version
13" are all acceptable version lines. Version lines with other
versions of TAP or KTAP will not cause the parsing of the test results
to fail but it will produce an error.

Maybe "TAP version Linux.1" ? I don't want to needlessly break existing
parsers.

quoted

Plan lines
----------

Plan lines must follow the format of "1..N" where N is the number of
subtests. The second line of KTAP output must be a plan line, which
indicates the number of tests at the highest level, such that the
tests do not have a parent. Also, in the instance of a test having
subtests, the second line of the test after the subtest header must be
a plan line which indicates the number of subtests within that test.

I do not want "subtests" as a specification concept, only "nested
tests".

quoted

Test case result lines
----------------------

Test case result lines must have the format:

    <result> <number> [-] [<description>] [<directive>] [<diagnostic data>]

"[<diagnostic data>]" is not defined below. I think what you mean is
"directive details".

IMHO it's useful to keep this called 'diagnostic data'.

The directive is optional, and having this be diagnostic data means you can
put explanatory text, or result details, that are independent of the description,
on the same line as the result.  This is convenient in some cases.

I think diagnostic data should always be prefixed by a hash sign.
Or, in the case where diagnostic data appears on the result line,
the hash serves as the delimiter.  This means that the description
text must not include a hash character.

For example:
ok 5 - check return code # rcode=0
or
not ok 5 - check return code # rcode=17

I suggest:

<result> <number>[ [- ]<description>][# <directive> [<directive details>]]

quoted

The result can be either "ok", which indicates the test case passed,
or "not ok", which indicates that the test case failed.

Yes.

quoted

The number represents the number of the test case or suite being
performed. The first test case or suite must have the number 1 and the
number must increase by 1 for each additional test case or result at
the same level and within the same testing suite.

Yes, though parsers should handle ordering failures and missed test
results (this is the purpose of the "plan" line).

quoted

The "-" character is optional.

The description is a description of the test, generally the name of
the test, and can be any string of words (can't include #). The
description is optional.

Yes, though the optional "- " is strictly part of the optional
description.

It's mildly annoying that "-" is optional.  It's trivial to deal with in the parser
to just ignore it if it's present.  But it has no semantic meaning whatsoever.
IMHO it would be nice to either mandate it or remove it, for consistency's sake.
This could be based solely on the consensus for whether it added or detracted
from human readability, since parsers don't care.

quoted

The directive is used to indicate if a test was skipped. The format

The directive is a single word -- currently only "SKIP" is recognized
by TAP 14(?), but kselftest uses also XFAIL, XPASS, TIMEOUT, and
error. (Though I would love to hear what "error" is intended to be used
for, if different from regular "not ok".)

ERROR usually means that there was some kind of failure
in the testing process that makes the testcase result indeterminate.
(like loss of connection to a server, or crash of the program executing
the testcase).  This can be due to a bug in the test or the test system.
I'm, not sure if ERROR is used this way in kselftest.  It is, IMHO, a useful
distinction to make.  TIMEOUT is a special case of ERROR, IMHO.

As an aside, I've never liked XFAIL or XPASS.  Semantically, I believe they
are quite confusing, and depend on the level of testcase execution.  XPASS is also
redundant.  Isn't the default case always XPASS?

In the case of XFAIL, does it mean that an operation that was expected to fail
all the time, failed, and so the test passed?  Or does it mean that this is a known
bug that a particular developer or site has decided to ignore (as in "we know this
is a bug, but we expected it to fail, so just ignore it for now).  The latter, I really
don't think should be encoded in the tests themselves, but should be put
into a layer above the test and parser, that is applied according to developer
or testlab policy.

(This could also be called "context" rather than "directive".)

quoted

for the directive is: "# SKIP [<skip_description>]". The
skip_description is optional and can be any string of words to
describe why the test was skipped.

I would call this "directive details".

I would not, since it is human readable text, and may be present
whether the directive is there or not, I would still call it 'diagnostic data'.

quoted

The result of the test case result
line can be either "ok" or "not ok" if the skip directive is used.
Finally, note that TAP 14 specification includes TODO directives but
these are not supported for KTAP.

Examples of test case result lines:

Test passed:

    ok 1 - test_case_name

Test was skipped:

    not ok 1 - test_case_name # SKIP test_case_name should be skipped

Test failed:

    not_ok 1 - test_case_name

This isn't valid. No "_" is recognized for "ok" vs "not ok".

quoted

Diagnostic lines
----------------

Diagnostic lines are used for description of testing operations.
Diagnostic lines are generally formatted as "#
<diagnostic_description>", where the description can be any string.
However, in practice, diagnostic lines are all lines that don't follow
the format of any other KTAP line format.

I still think there should be a distinction between "diagnostic lines"
and "unknown lines". For example, if kselftest is running on a console,
the dmesg output may be intermixed, and possibly temporally offset. Such
lines may not be useful, and may not be strictly correlated, where-as
the output from kselftest is at least self-consistent.

I think this would be a very good distinction to make.  You might have
unknown lines that (inadvertently?) start with a hash, but trying to 
make sure that the test itself always outputs diagnostic data with a preceding
hash gives a reader potentially more information about where the message
came from, and could be useful.

quoted

Diagnostic lines can be
anywhere in the test output after the first two lines.

I don't see a reason for this strictness. They can be anywhere.

quoted

There are a few
special diagnostic lines.

No; diagnostic lines must have no meaning: they are for humans and nothing
else. If other details are needed for machines, we should explicitly
create new format lines. I made a mistake when I used a diagnostic line
for embedding the test names. :( There is a need for parsers to discover
the name of a given test, though, so we do likely need something for this.

I agree it is handy to have the test name for sub-tests.  However, I don't
see any semantic difference whether we define sub-test declarations via
diagnostic lines with special syntax or by creating a new syntax separate
from diagnostic lines.  Either way, we have a new syntax for parsers to
understand.

quoted

Diagnostic lines of the format "# Subtest:
<test_name>" indicate the start of a test with subtests. Also,
diagnostic lines of the format "# <test_name>: <description>" refer to
a specific test and tend to occur before the test result line of that
test but are optional.

I don't think the above should be included in the spec -- diag lines
should have no parseable meaning.

quoted

Bail out line
-------------

A bail out line can occur anywhere in the KTAP output and will
indicate that a test has crashed. The format of a bail out line is
"Bail out! [<description>]",  where the description can give
information on why the bail out occurred and can be any string.

I'm not a fan of the Bail out line. It's not a problem, exactly, but I
find it redundant. If we want an "end of test" line, let's make one.
"Bail out" is a special case of exit. If we want to handle test exit,
let's define it. E.g. make kselftest's test summary lines not
diagnostic lines:

# FAILED: 85 / 87 tests passed.
# Totals: pass:83 fail:2 xfail:0 xpass:0 skip:2 error:0

How consistent is it for selftests to produce test summary lines?

Also, parsers should treat a new "TAP version..." line at the same
nesting level as indication that the prior test has ended (and any tests
without result lines should be considered failed).

Agreed.

quoted

Nested tests
------------

The new specification for KTAP will support an arbitrary number of
nested subtests. Thus, tests can now have subtests and those subtests
can have subtests. This can be useful to further categorize tests and
organize test results.

The new required format for a test with subtests consists of: a
subtest header line, a plan line, all subtests, and a final test
result line.

The first line of the test must be the subtest header line with the
format: "# Subtest: <test_name>".

Please no. There is no reason to force a nested test to suddenly have
to know about its test execution depth/environment. A subtest is just a
nested TAP result. That it is nested is only meaningful to the parser, not
the executing test, and it must have the same output, nested or not. (e.g.
running the test alone or running the test under a full kselftest run,
the only difference is the indentation level.)

I agree that the sub-test should not know about this line.  I thought
the "parent" test wrote the line.

quoted

The second line of the test must be the plan line, which is formatted
as "1..N", where N is the number of subtests.

Following the plan line, all lines pertaining to the subtests will follow.

Finally, the last line of the test is a final test result line with
the format: "(ok|not ok) <number> [-] [<description>] [<directive>]
[<diagnostic data>]", which follows the same format as the general
test result lines described in this section. The result line should
indicate the result of the subtests. Thus, if one of the subtests
fail, the test should fail. The description in the final test result
line should match the name of the test in the subtest header.

An example format:

KTAP version 1
1..1
    # Subtest: test_suite
    1..2
    ok 1 - test_1
    ok 2 - test_2
ok 1 - test_suite

Again, I see only downsides to this. Nesting for the spec is simple
indentation-based recursion. Let's just keep what we already have:

TAP version 13
1..1
# TAP version 13
# 1..2
# ok 1 - test_1
# ok 2 - test_2
ok 1 - test_suite

I prefer spaces, but whatever.  Is the prefix on the 'TAP version 13' line guaranteed
to be the prefix on all lines for the (sub-)test?
 

quoted

An example format with multiple levels of nested testing:

KTAP version 1
1..1
    # Subtest: test_suite
    1..2
        # Subtest: sub_test_suite
        1..2
        ok 1 - test_1
        ok 2 test_2
    ok 1 - sub_test_suite
    ok 2 - test
ok 1 - test_suite

In the instance that the plan line is missing, the end of the test
will be denoted by the final result line containing a description that
matches the name of the test given in the subtest header. Note that
thus, if the plan line is missing and one of the subtests have a
matching name to the test suite this will cause errors.

A plan line is required. No special cases are needed. :)

If nesting level is lost, a parser will understand the nested test to
have ended, but probably so did the test runner:

TAP version 13
1..1
    TAP version 13
    1..2
        TAP version 13
        1..3
        ok 1 - test_1
        ok 2 test_2
    not ok 1 - sub test unfinished plan
    ok 2 - test
not ok 1 - test_suite

quoted

Lastly, indentation is also recommended for improved readability.

Indentation is _required_. :)

I strongly agree.  You can't disambiguate some outputs, unless the
indentation is present.

Whether this is "# " or "  " I don't really care, as long as the change
is coordinated. "  " is easier for humans to read, but may make parsing of
"unknown" lines more difficult for machines.

I prefer spaces for human readability.

Just as a side note, in some Fuego tests, it was very useful to include an identifier
in thethe prefix nested tests.  The output looked like this:

TAP version 13
1..2
[batch_id 4] TAP version 13
[batch_id 4] 1..2
[batch_id 4] ok 1 - cyclictest with 1000 cycles
[batch_id 4] # problem setting CLOCK_REALTIME
[batch_id 4] not ok 2 - cyclictest with CLOCK_REALTIME
not ok 1 - check realtime
[batch_id 4] TAP version 13
[batch_id 4] 1..1
[batch_id 4] ok 1 - IOZone read/write 4k blocks
ok 2 - check I/O performance

Can I propose that the prefix not be fixed by the spec, but that the spec indicates that
whatever the prefix is on the TAP version line, that prefix must be used with the output for
all lines from the test (with the exception of unknown lines)?

quoted

Major differences between TAP 14 and KTAP specification
-------------------------------------------------------

Note that the major differences between TAP 14 and KTAP specification:
- yaml and json are not allowed in diagnostic messages

Agreed -- these are overkill (and very difficult to implement as they
_follow_ the test result line: anything generating them has already
finished running).

I thought that some tests put the diagnostic data before the result line.

I don't care whether it is before or after (and indeed Fuego's parser
can handle either case).  But it can't handle it when some testcases
put diagnostic data before the result line, and some testcases put
it after, in the same test.  There's no delimiter for the separation
between the two different testcases diagnostic data.

Could we possibly have a rule about diagnostic lines positioning relative
to the testcase result line?  Maybe a rule like this:  For a single test,
diagnostic lines should either always preceded the testcase they
are related to, or follow the testcase they are related to.

quoted

- TODO directive not allowed

I would just say "unrecognized".

quoted

- KTAP allows for an arbitrary number of tests to be nested with
specified nested test format

Yup.

quoted

Example of KTAP
---------------

KTAP version 1
1..1
    # Subtest: test_suite
    1..1
        # Subtest: sub_test_suite
        1..2
        ok 1 - test_1
        ok 2 test_2
    ok 1 - sub_test_suite
ok 1 - test_suite

For a good example, please include all the possible combinations (SKIP,
not ok, diagnostics, etc etc)

quoted

=========================================
Note on incompatibilities with kselftests
=========================================

To my knowledge, the above specification seems to generally accept the
TAP format of many non-nested test results of kselftests.

An example of a common kselftests TAP format for non-nested test
results that are accepted by the above specification:

TAP version 13
1..2
# selftests: vDSO: vdso_test_gettimeofday
# The time is 1628024856.096879
ok 1 selftests: vDSO: vdso_test_gettimeofday
# selftests: vDSO: vdso_test_getcpu
# Could not find __vdso_getcpu
ok 2 selftests: vDSO: vdso_test_getcpu # SKIP

However, one major difference noted with kselftests is the use of more
directives than the "# SKIP" directive. kselftest also supports XPASS
and XFAIL directives. Some additional examples found in kselftests:

    not ok 5 selftests: netfilter: nft_concat_range.sh # TIMEOUT 45 seconds

    not ok 45 selftests: kvm: kvm_binary_stats_test # exit=127

Should the specification be expanded to include these directives?

Yes. (Though "exit=" is a mistake in runner.sh -- this should probably
be reported without the '#')

Please No!  (Unless I'm misinterpreting that line).

If you remove the '#', that makes the 'exit=127' part of the test description.
The test description should be invariant between success and failure cases.
The exit=127 (if I'm reading that line right) is part of data explaining or
describing the failure. It would be 'exit=0' if the test succeeded.

In order to correlate test cases (ie show results over time), the description
should not change from one invocation of the test to the next.  At least Fuego
uses the test description this way.  That is, the testcase description is used as
a testcase identifier, and if it changes from run to run things get confusing.

quoted

However, the general format for kselftests with nested test results
seems to differ from the above specification. It seems that a general
format for nested tests is as follows:

TAP version 13
1..2
# selftests: membarrier: membarrier_test_single_thread
# TAP version 13
# 1..2
# ok 1 sys_membarrier available
# ok 2 sys membarrier invalid command test: command = -1, flags = 0,
errno = 22. Failed as expected
ok 1 selftests: membarrier: membarrier_test_single_thread
# selftests: membarrier: membarrier_test_multi_thread
# TAP version 13
# 1..2
# ok 1 sys_membarrier available
# ok 2 sys membarrier invalid command test: command = -1, flags = 0,
errno = 22. Failed as expected
ok 2 selftests: membarrier: membarrier_test_multi_thread

The major differences here, that do not match the above specification,
are use of "# " as an indentation and using a TAP version line to
denote a new test with subtests rather than the subtest header line
described above. If these are widely utilized formats in kselftests,
should we include both versions in the specification or should we
attempt to agree on a single format for nested tests? I personally
believe we should try to agree on a single format for nested tests.
This would allow for a cleaner specification of KTAP and would reduce
possible confusion.

We already use "# " and the nested TAP lines to denote subtests. Without
a good reason to change it, we should avoid the churn with the existing
parsers.

quoted

====

So what do people think about the above specification?
How should we handle the differences with kselftests?

I'm probably a broken record by now, but kselftests _is_ the
specification. ;) What about it needs changing, and why?

quoted

If this specification is accepted, where should the specification be documented?

I imagine some .rst file under Documentation/dev-tools/, linked to from
kselftest.rst and kunit/...rst

--
Kees Cook