On Sun, Mar 12, 2023 at 10:29:11AM +0000, Marc Zyngier wrote:
Mark Brown broonie@kernel.org wrote:
static struct vcpu_config *vcpu_configs[] = { &vregs_config, @@ -1131,5 +1163,6 @@ static struct vcpu_config *vcpu_configs[] = { &sve_pmu_config, &pauth_config, &pauth_pmu_config,
- &mte_config,
}; static int vcpu_configs_n = ARRAY_SIZE(vcpu_configs);
Is there any reason why we sidestep the combinations of MTE with PAuth and PMU? I know this leads to an exponential set growth, but this is the very purpose of this test, and we found bugs related to this in the past.
The test is already not bothering with the combinations of SVE and pointer auth, it appeared that the intent of the test was only to test specific combinations. From what's there it looks more like there's something with PMU interacting specially with things (it's all X and X+PMU) that needs coverage. I couldn't see anything between it and MTE, though I nearly added a MTE+PMU combination just for the sake of it. It's one of those areas where it's hard to determine if there's an intent behind the implementation choices made or if they're just whatever someone happened to write and not particularly important or desired.
A good first step would be to be able to build these combinations dynamically, and only then add new sublists to the mix.
That would certainly be a good idea, if we were heading in that direction I'd also expect negative tests checking that for example pointer authentication registers don't appear when that's not enabled. I'm not sure that it's worth blocking all new coverage for that though, there is still value in having a bit of basic coverage even if not all the combinations are covered yet.
The test is also going to want extension for more gracefully handling registers that appear based on architecture extensions with no control over their exposure to the guest, potentially with tie in to handling based on configurable ID registers when that goes in. The way this test is written was also part of why I was wondering if PIE should be configurable.