On Wed, Feb 24, 2021 at 06:26:51PM -0300, Wainer dos Santos Moschetta wrote:
Currently the acceptance tests tagged with "machine" have the "-M TYPE"
automatically added to the list of arguments of the QEMUMachine object.
In other words, that option is passed to the launched QEMU. On this
series it is implemented the same feature but instead for tests marked
with "cpu".
Good!
There is a caveat, however, in case the test needs additional arguments to
the CPU type they cannot be passed via tag, because the tags parser split
values by comma. For example, in tests/acceptance/x86_cpu_model_versions.py,
there are cases where:
* -cpu is set to
"Cascadelake-Server,x-force-features=on,check=off,enforce=off"
* if it was tagged like
"cpu:Cascadelake-Server,x-force-features=on,check=off,enforce=off"
then the parser would break it into 4 tags ("cpu:Cascadelake-Server",
"x-force-features=on", "check=off", "enforce=off")
* resulting on "-cpu Cascadelake-Server" and the remaining arguments are
ignored.
For the example above, one should tag it (or not at all) as
"cpu:Cascadelake-Server"
AND self.vm.add_args('-cpu',
"Cascadelake-Server,x-force-features=on,check=off,enforce=off"),
and that results on something like:
"qemu-system-x86_64 (...) -cpu Cascadelake-Server -cpu
Cascadelake-Server,x-force-features=on,check=off,enforce=off".
There are clearly two problems here:
1) the tag is meant to be succinct, so that it can be used by users
selecting which tests to run. At the same time, it's a waste
to throw away the other information or keep it duplicate or
incosistent.
2) QEMUMachine doesn't keep track of command line arguments
(add_args() makes it pretty clear what's doing). But, on this type
of use case, a "set_args()" is desirable, in which case it would
overwrite the existing arguments for a given command line option.
QEMU is going to ignore the first -cpu argument. See the patch 0003 for a
reference.
But this would still be creating a QEMU command line with multiple
'-cpu' arguments, right? I understand this could be useful for
testing the behavior of the parameter parsing (if that's intended) but
it's bad practice to be generating incorrect command line in tests.
Maybe just by tackling issue #2 this could be avoided.
Cheers,
- Cleber.