Add more debug logging to debug rt-app calibration issue.

FEATURE

Allow setting the --log option of rt-app.

Also set the level to "debug" while calibrating.

FIX

Fix CI by ignoring deprecation warnings triggered internally by
Jupyterlab/IPython themselves.

Dietmar Eggemann authored Nov 14, 2022 and Douglas Raillard committed Nov 14, 2022

Occasionally the load_tracking invariance tests
[load\|util]_[correctness\|behaviour] fail on rb5 little/medium CPUs
on low frequencies.
The root cause is wrong rt-app calibration for these CPUs/frequencies.
The reason for the rt-app calibration issue is not known at the moment.
Add an assert in _get_rtapp_profile() of class InvarianceItemBase to
catch the issue earlier.

Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>

FIX

Do not error if a class does not define __init__.

FEATURE

Recognize PEP 673 Self return annotation on classmethods as being a
factory class method.

FIX

Older kernels do not have this folder and also do not suffer from the
broken sha1 in the /sys/kheaders.tar.xz headers, so ignore them.

Patrick Bellasi authored Nov 04, 2022 and Douglas Raillard committed Nov 04, 2022

FIX

A Systrace generated by the Perfetto's traceconv tool has records with this
format:
```
    <idle>-0     (-----) [000] .... 608397.919018: sched_wakeup: comm=traced pid=403293 prio=120 target_cpu=000
 ```
which timestamp gets parsed as `7.919018`, i.e. by ignoring all the digits but
one decimal.

The regexp also picks up all the spaces before the task name, by ending up with
duplicated names when the same task name is parsed also (without space) form an
event field (e.g. a `sched_switch`'s `next_comm`)

Fix the timestamp parsing by ensuring that, after the `__cpu` filed has been
parsed, we skip all and only the non digit chars following it.
Fix the command name parsing by ensuring we discard all the space before the
name start.

FIX

Headers found at /sys/kheaders.tar.xz generated when CONFIG_IKHEADERS=y
contain a broken version of the include/linux/atomic/*. headers. They
embed a sha1sum of the file on their last line, but
kernel/gen_kheaders.sh strips some multiline comments from the file,
changing the SHA1.

Work around the issue by updating the SHA1 in the file.

FIX

Disable Tabulator pagination as it is now fast enough to load all the
data. This allows scrolling when clicking on the plot to work correctly.

FIX

Fix scrolling in the table when clicking on the plot. Instead of making
the plot disappear, the corresponding row is selected in the table.

Note that the table will not change page automatically.

FIX

Ensure hv.extension() is done after all imports, otherwise plots are not
displayed in some circumstances:

https://github.com/holoviz/holoviews/issues/5491

FEATURE

Pass path to module and kernel path instead of relying on Kbuild's
$(src) and $(srctree) variables, as they seem to be broken on some
android's kernel.

FEATURE

Expose devlib's max_async Target parameter. It is used to limit the
max number of in-flight async commands at any point in time.

FIX

Update HRTxtTraceParser skeleton regex to allow for a TGID field, which
is part of recent systrace files.

Pierre Gondois authored Oct 14, 2022 and Douglas Raillard committed Oct 14, 2022

commit 487e1ac4

 ("lisa.tests.scheduler.eas_behaviour:
Add big-task-duty-cycle platform parameter")
modified the prototype of _get_rtapp_profile(), but an implementation
was left with the initial prototype, triggering errors. Fix that.

Reported-by: Deepak Kumar Mishra <deepakkumar.mishra@arm.com>
Signed-off-by: Pierre Gondois <pierre.gondois@arm.com>

Pierre Gondois authored Oct 07, 2022 and Douglas Raillard committed Oct 13, 2022

On a rb5, the first OPPs of the little CPUs are less energy
efficient than the first OPPs of the big CPUs:
little CPUs:
freq=300000  power=28864  capa=59  power/capa=489.22
freq=403200  power=38774  capa=79  power/capa=490.81
freq=480000  power=46183  capa=95  power/capa=486.13
freq=576000  power=55419  capa=114 power/capa=486.13
freq=652800  power=62732  capa=129 power/capa=486.29
freq=748800  power=71968  capa=148 power/capa=486.27
freq=825600  power=82719  capa=163 power/capa=507.47 <- [3]
freq=902400  power=94168  capa=178 power/capa=529.03
freq=979200  power=113400 capa=194 power/capa=584.53
freq=1056000 power=114700 capa=209 power/capa=548.80
freq=1132800 power=126190 capa=224 power/capa=563.34
[...]
BIG CPUs:
freq=825600 power=133030  capa=301 power/capa=441.96
freq=902400 power=151283  capa=329 power/capa=459.82
freq=979200 power=168490  capa=357 power/capa=471.96 <- [1]
freq=1056000 power=188799 capa=385 power/capa=490.38 <- [2]
freq=1209600 power=224387 capa=441 power/capa=508.81
freq=1286400 power=244608 capa=469 power/capa=521.55
freq=1363200 power=275303 capa=497 power/capa=553.92
freq=1459200 power=305249 capa=533 power/capa=572.69
freq=1536000 power=340314 capa=561 power/capa=606.62
freq=1612800 power=381776 capa=589 power/capa=648.17
[...]

When running the TwoBigThreeSmall test, the big task has a
utilization such as the task cannot fit on the little CPUs.
util~=290 so that 290 * 1.2 = 348. The capacity of the little CPUs
is 350. The small task will run at ~50% of the maximum capacity of
the little CPUs, so util~=174.
The big CPUs will run at the OPP [1], which is more energy efficient
than all the little CPUs' OPPs. Now if the big CPU utilization jumps
to [2], the little CPUs' OPPs below [3] become more energy efficient.

Thus, it appears that for the rb5, the utilization chosen for the
big tasks makes the big CPUs either attractive or repulsive for the
small tasks, and this for a range of 20 util. This makes the
TwoBigThreeSmall test for rb5 unstable.

Add a platform specific 'big-task-duty-cycle' parameter to handle
this specific test case. This parameter allows to choose a higher
utilization for the big tasks, making the big CPUs less attractive
for small tasks.

Suggested-by: Douglas Raillard <douglas.raillard@arm.com>
Signed-off-by: Pierre Gondois <pierre.gondois@arm.com>

FIX

Using a trace window can remove all activations for a given task. That
leads to <NA> values in dataframes, that bokeh chokes on. Avoid creating
such <NA> values in the first place fixes that.

FIX

Change expected output for very recent versions of trace-cmd.

Use the new rtapp_profile_kwargs infrastructure to avoid having to
redefine the rtapp_profile property.

FEATURE

Allow forwarding custom keyword arguments to _get_rtapp_profile() when
it is called via the "rtapp_profile" property.

This allows _get_rtapp_profile() to depend not only on the platform info
but also on user-defined values.

FIX

Do not raise exceptions when accessing class attributes on partially
initialized objects.

FIX

iteritems() has been deprecated in favor of items()

FEATURE

Use the trace-cmd binary shipped by LISA to improve reproducibility and
for when lisa is installed from PyPI.

It is still possible to make LISA prefer the system binary by setting
LISA_USE_SYSTEM_BIN=1 environment variable.

FEATURE

If env var LISA_USE_SYSTEM_BIN=1, HOST_PATH will give priority to
existing PATH, otherwise it will give priority to LISA's PATH containing
the shipped binaries.

Also add get_bin() function that allows getting the path to a binary.

FIX

Ensure exceptions are re-raised unless explicitly swallowed by the user
by returning a truthy value. This is necessary as the expected pattern
is to handle ContextManagerExit and not re-raise anything explicitly,
which would otherwise be necessary with a normal use of
contextlib.contextmanager.

FIX

Make try_loaders() robust against inner context manager swallowing an
exception by forcefully re-raising it.

Qais Yousef authored Sep 29, 2022 and Douglas Raillard committed Oct 03, 2022

By design, migration margin shouldn't be considered when comparing
uclamp with capacities as explained in this series:

	https://lore.kernel.org/lkml/20220804143609.515789-1-qais.yousef@arm.com/



It is considered a bug - that the above series fixes it.

Integration branch was failing with this series applied because of this
margin.

Fix the test to ignore the margin to reflect the actual uclamp design.

Older kernels without the series might fail now, which is the correct
result.

The series should trickle to stable kernels once merged in mainline
Linux. The failure should signal to any user that they need to backport
this series.

Signed-off-by: Qais Yousef <qais.yousef@arm.com>

FEATURE

Add a new tool that displays the changes since a given git ref.
Also call this tool automatically after a git pull.

FIX

Avoid error if lisa package cannot be imported (e.g. if source init_env
has not been done)

FIX

Also watch for new git hooks to install.

FIX

Avoid showing warnings when git notes are not available.

Document the use of lisa-load-kmod

Add command line tool to compile and load LISA's kernel module using the
facilites of the Python API (such as Alpine container build).

FIX

Do not add a leading -- to the parameter name, so that it stays a
positional parameter.

FEATURE

If path=None, use the output_path specified to __init__.

FEATURE

Allow passing custom CLI tool description to argparse.

FIX

FIX

Equivalent to contextlib.contextmanager except that this pattern will
not run the cleanup code when context manager is destroyed without
calling __exit__() (explicitly or via "with" statement).

The usual pattern will run the cleanup code even if __exit__() is not
called, as destructing the generator will raise a GeneratorExit in it no
matter what:

    @contextlib.contextmanager
    def f():
	try:
	    yield
	finally:
	    cleanup()

This version will avoid that issue and will therefore behave like
context managers directly defined with of __enter__() and __exit__():

    @destroyablecontextmanager
    def f():
	try:
	    yield
	except ContextManagerExit:
	    cleanup()

This allows the final user of the context manager to skip cleanup if
they want to, which might be useful in some cases.

FEATURE

Very similar to contextlib.contextmanager but allows graceful handling
of destruction without calling __exit__().

Since most context managers
only use it with try/yield/finally pattern, this avoids running the
"finally" clause in case of destruction without calling __exit__(), just
like for context managers not created with contextlib.contextmanager().