FIX

If the base path where we look for the lisa kmod exist but no module
exist, find will return an empty string without error. Handle that case
and treat it as having not found the kmod.

FIX

Report the actual list of available events when reporting the missing
optional events.

FIX

Ensure the order of failed event checker is preserved, as this matters
when displaying an "or" group.

FEATURE

Log the events that LISA can detect in the kernel module (note that this
is a superset of what feature can actually be enabled).

FIX

Remove useless duplicate in the list of cross compile prefix tried.

FIX

Allow passing a filesystem path as CC value instead of just a command
name.

FIX

Caching of the compilation artifact was keyed on the compiler name so
far. This breaks if the user manipulates PATH and swaps different
binaries under the same name, e.g. multiple versions of clang all called
"clang".

Fix that by calling "cc --version" and caching based on the result. This
ensures we will not reuse artifacts created with a different toolchain.

FIX

Update trace-cmd to include zlib support

FIX

Enable zlib support in trace-cmd report.

FEATURE

Reorder event fields to reduce as much padding as possible.

FEATURE

Allow attaching the same probes to multiple trace points.

Make most functions in sched_helpers.h take and return pointers to const.

FIX

CFS_UCLAMP makes no sense since there is no uclamp info at the CFS rq
level, only at the rq level.

FEATURE

Same as DEFINE_TP_EVENT_FEATURE() but with an extra deprecation warning
upon init.

FIX

__pycache__ is sometimes created around Python sources which later can
affect the checksum of the module.

Ensure that the __pycache__ folders and all their contents are ignored
when gathering module sources.

FEATURE

Allow debugging checksum issues by showing the content of the module
sources.

FIX

Fix alignment inference on packed structs to return 1 instead of what
would be the alignment without __attribute__((packed)).

FIX

Ensure that anonymous struct/union are dumped correctly as anonymous
struct/union instead of being put behind a typdef, which breaks the
semantic.

FIX

Avoid accidentally stopping the recursive type printing when
encountering a forward declerence, since user code might be following a
pointer chain (a->b->c) where "b" is of an internal type, and suddenly
overnight becomes a forward decl, making "b->c" break.

Instead, resolve forward decl to the type they reference. This is
inaccurate since "struct foo;" is not necessarily a forward declaration
of "struct foo {...};" as they could both appear in totally unrelated
files.

However, the risk of that happening _in the set of types_ we care about
is quite low, and is lower than the risk of having forward decl popping
up randomly.

FIX

FIX

Ensure we don't include __pycache__ in the checksum.

FEATURE

Since this kernel module is meant to host more than just ftrace events,
rename it to a more general name.

FIX

Modules can manipulate sys.modules[__name__], and the Python import
statement will use that result, instead of necessarily returning the
module that was actually executed.

Make a shim with a deprecation warning to avoid breaking existing
imports.

Add README stating how the namespace package is to be organized.

FEATURE

Comment-out a return annotation so that exekall will not pickup the
example test.

BREAKING CHANGE

This is a breaking change from a setup point of view (separate package)
and when using "exekall run lisa" that now needs to be
"exekall run lisa lisa_tests" since the most class definitions are moved
to the new package.

However, from the API perspective this is not a breaking change, as a
later commit adds import shims (with a deprecation warning) in the old
locations.

FIX

Install location for git hooks is relative to $LISA_HOME, so prepend
"$LISA_HOME" to the path.

FIX

Fix type hints in _ATTRIBUTES so that they are accurate, and then use
them to validate only as they cannot generally be used to convert a
value.

FIX

Ensure uclamp values are integers and not float.

FEATURE

Use the type of each attribute to process user input. For example, this
will ensure uclamp values are integers and not user-provided floats.

FIX

Avoid choosing uclamp in CPU capacity ranges that are overlapping
between CPUs. Instead, ignore all the capacities on faster CPUs that are
lower than the max capa of any slower CPU.

Also ensure that the capacity bands considered exist inside a given CPU
and are not a band between the highest capacity of a CPU and the lowest
capacity of another one.

FIX

Factor in the schedutil capacity margin to ask for a util clamp that
should map right in the middle of a frequency band, instead of
middle + 25%.

FIX

Use the reliable square bracket syntax for column access instead of the
dodgy attribute syntax.

FIX

Base the expected CPU capacity on the CPU the tasks is currently
scheduled on rather than picking a single CPU for the entire phase and
hope for the task to stay on that CPU.