9:26 a.m.
On Fri, 2021-02-12 at 14:12 +0000, Daniel P. Berrangé wrote:
On Fri, Feb 12, 2021 at 02:25:24PM +0100, Tim Wiederhake wrote:
> "clang-tidy" is a static code analysis tool for c and c++ code. See
> https://listman.redhat.com/archives/libvir-list/2021-January/msg01152.html
> for some issues found by clang-tidy and more background
> information.
>
> Meson has support for clang-tidy and generates target named "clang-
> tidy" if
> it finds a ".clang-tidy" configuration file in the project's root
> directory.
>
> There are some problems with this approach though, with regards to
> inclusion
> in GitLab's CI:
>
> * Single-threaded runtime of a complete scan takes between 95 and
> 110 minutes,
> depending on the enabled checks, which is significantly longer than
> GitLab's
> maximum run time of 60 minutes, after which jobs are aborted.
IIUC, you can override the timeout setting per job...
IMO, extending the "hard" timeout for the job is not an option anyway,
for the reason presented in the next point:
> * Even without this limit on runtime, this new check would double
> to triple
> the run time of the libVirt pipeline in GitLab.
↑ This one.
Yeah that's a problem, but bear in mind there are two separate
scenarios
Running post merge we need to check all files, but the length of time
is a non-issue since no one is waiting for a post-merge check to
complete.
I am vary of the possibility of the pre-merge job running successfully
(e.g. because it ran into timeout before finding a certain issue) and
then failing for the post-merge job. One idea would be to set the post-
merge job to "allow-failure: true", i.e. making it a "notification"
only. But I am absolutely unsure on what the best approach is here,
hence the "RFC".
Running pre- merge we only need to check files which are modified by
the patches on the current branch. That ought to be orders of
magnitude
faster.
I see where you are coming from, but I believe this does not catch
cases where e.g. a change in a header file leads to a clang-tidy issue
in an otherwise untouched .c file, or vice versa.
The approach to cache files keyed on the result of the preprocessor
stage ("hash(gcc -E file.c)") on the other hand does catch this kind of
changes and reruns clang-tidy for all affected files.
> * clang-tidy finds a lot of false positives (see link above for
> explanation)
> and has checks that contradict the libVirt code style (e.g. braces
> around
> blocks). This makes a quite complex configuration in ".clang-tidy"
> necessary.
IMHO for code checkers to add real value to people you need the
existing
codebase in git to be 100% clean.
If you're not going to make the existing code compliant, then users
will
never be sure of what they should do for new code. We see this all
the
time in QEMU where its patch checker complains about problems that
were
pre-existing.
If we could get clang-tidy to do a subset of checks where we can make
libvirt 100% clean
I believe there is a misunderstanding here, please see explanation
below.
> * I was unable to make clang-tidy recognize the settings from
the
> configuration file for generated files, leading clang-tidy to
> always add some
> checks. These checks were among those that produced false
> positives.
>
> * The list of enabled / disabled checks in the yaml configuration
> file is a
> quite long string, making it hard to weave in some comments /
> documentation
> on which checks are enabled / disabled for what reason.
>
> This series introduces a new script, "run-clang-tidy.py". This is a
> replacement for the script of the same name from clang-tools-extra.
> It offers
> parallel execution, caching of results and a configurable soft-
> timeout.
>
> Please see the individual commits for more details. Comments
> welcome.
>
> https://gitlab.com/twiederh/libvirt/-/pipelines/255321968 → "clang-
> tidy".
What am I missing here - that job just seems to show a list of
filenames, but isn't reporting any issues for them ?
There are three "types" of checks:
* Irrelevant checks that we want to disable permanently, e.g. checks
for some c++ specific issue that does not apply to c code,
* Checks that the libvirt code base currently passes,
* Checks that the libvirt code base currently fails.
Path #09 disables the first and last category of checks. Note that the
list is not defined in terms of "enable these checks" but "disable
these checks". The idea is to have checks that are introduced in new
versions of clang-tidy enabled by default and disabled checks commented
on and justified.
This is why you see no findings in the output -- there are none,
currently only "passing" checks are enabled. I disabled all failing
checks indiscriminately, as a discussion of which particular checks are
actually useful and which are not would distract from the general
question of "Do we want clang-tidy integration in the CI and if so,
how".
See for example the check "readability-braces-around-statements" [1],
which mandates braces around single-line "if" or "while" blocks,
which
contradicts the libvirt code style. I believe that the verdict for this
check will be permanent disablement. Other checks, for example "cert-
err34-c" [2] are issues that in my opinion should be fixed and I
started doing so already, see [3].
For comparison, have a look at [4]: This is the output of run-clang-
tidy.py with only the "irrelevant" checks disabled.
Regards,
Daniel
Regards,
Tim
[1]
https://clang.llvm.org/extra/clang-tidy/checks/readability-braces-around-...
[2] https://clang.llvm.org/extra/clang-tidy/checks/cert-err34-c.html
[3]
https://listman.redhat.com/archives/libvir-list/2021-January/msg01152.html
[4] https://gitlab.com/twiederh/libvirt/-/jobs/1026797599