On Mon, Aug 22, 2022, Nadav Amit wrote: > On Aug 22, 2022, at 8:42 AM, Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > On Mon, Aug 22, 2022, Michal Luczaj wrote: > >> On 8/22/22 00:06, Michal Luczaj wrote: > >>> Note that doing this the ASM_TRY() way would require extending > >>> setup_idt() (to handle #DB) and introducing another ASM_TRY() variant > >>> (one without the initial `movl $0, %%gs:4`). > >> > >> Replying to self as I was wrong regarding the need for another ASM_TRY() variant. > >> Once setup_idt() is told to handle #DB, > > > > Hmm, it might be a moot point for this patch (see below), but my vote is to have > > setup_idt() wire up all known handlers to check_exception_table(). I don't see > > any reason to skip some vectors. Code with __ASM_TRY() will explode no matter what, > > so it's not like it risks suppressing completely unexpected faults. > > I would just like to point out that this whole FEP approach is not friendly > for other hypervisors or bare-metal testings. While people might have mixed > feelings about other hypervisors, the benefit of running the tests on > bare-metal should be clear. If people want to run KUT on other hypervisors that support a different FEP (KVM stole it from Xen after all), then I have no objection to genericizing KVM_FEP and making it a config option. > Specifically in this test, it is rather simple to avoid using FEP. I'm saying KUT should do both, i.e. run the same sequences with and without forced emulation on the "critical" instructions. That provides full coverage for bare metal, and also provides coverage for both "fast" and "slow" emulation in KVM (and other future hypervisors if other prefixes are ever supported).
