Date:   Fri, 20 Jan 2023 06:33:56 -0800
From:   Ricardo Koller <ricarkol@...gle.com>
To:     Ben Gardon <bgardon@...gle.com>
Cc:     David Matlack <dmatlack@...gle.com>, linux-kernel@...r.kernel.org,
        kvm@...r.kernel.org, Paolo Bonzini <pbonzini@...hat.com>,
        Peter Xu <peterx@...hat.com>,
        Sean Christopherson <seanjc@...gle.com>,
        Vipin Sharma <vipinsh@...gle.com>
Subject: Re: [PATCH 2/2] selftests: KVM: Add page splitting test

On Fri, Jan 20, 2023 at 6:23 AM Ricardo Koller <ricarkol@...gle.com> wrote:
>
> On Thu, Jan 19, 2023 at 03:48:14PM -0800, Ben Gardon wrote:
> > On Thu, Jan 19, 2023 at 2:56 PM David Matlack <dmatlack@...gle.com> wrote:
> > ...
> > > > +static int NR_VCPUS = 2;
> > > > +static int NR_SLOTS = 2;
> > > > +static int NR_ITERATIONS = 2;
> > >
> > > These should be macros or at least const?
> >
> > Yikes, woops, that was a basic mistake.
> >
> > >
> > > > +static uint64_t guest_percpu_mem_size = DEFAULT_PER_VCPU_MEM_SIZE;
> > > > +
> > > > +/* Host variables */
> > >
> > > What does "Host variables" mean? (And why is guest_percpu_mem_size not a
> > > "Host variable"?)
> > >
> > > I imagine this is copy-pasta from a test that has some global variables
> > > that are used by guest code? If that's correct, it's probably best to
> > > just drop this comment.
> >
> > Yeah, shameful copypasta. I'll drop it.
> >
> > >
> > > > +static u64 dirty_log_manual_caps;
> > ...
> >
> > > > +             /*
> > > > +              * Incrementing the iteration number will start the vCPUs
> > > > +              * dirtying memory again.
> > > > +              */
> > > > +             iteration++;
> > > > +
> > > > +             for (i = 0; i < NR_VCPUS; i++) {
> > > > +                     while (READ_ONCE(vcpu_last_completed_iteration[i])
> > > > +                            != iteration)
> > > > +                             ;
> > > > +             }
> > > > +
> > > > +             pr_debug("\nGetting stats after dirtying memory on pass %d:\n", iteration);
> > > > +             get_page_stats(vm, &stats_dirty_pass[iteration - 1]);
> > >
> > > Incrementing iteration, waiting for vCPUs, and grabbing stats is
> > > repeated below. Throw it in a helper function?
> >
> > Good call.
> >
> > >
> > > > +
> > > > +             memstress_get_dirty_log(vm, bitmaps, NR_SLOTS);
> > > > +
> > > > +             if (dirty_log_manual_caps) {
> > > > +                     memstress_clear_dirty_log(vm, bitmaps, NR_SLOTS, pages_per_slot);
> > > > +
> > > > +                     pr_debug("\nGetting stats after clearing dirty log pass %d:\n", iteration);
> > > > +                     get_page_stats(vm, &stats_clear_pass[iteration - 1]);
> > > > +             }
> > > > +     }
> > > > +
> > > > +     /* Disable dirty logging */
> > > > +     memstress_disable_dirty_logging(vm, NR_SLOTS);
> > > > +
> > > > +     pr_debug("\nGetting stats after disabling dirty logging:\n");
> > > > +     get_page_stats(vm, &stats_dirty_logging_disabled);
> > > > +
> > > > +     /* Run vCPUs again to fault pages back in. */
> > > > +     iteration++;
> > > > +     for (i = 0; i < NR_VCPUS; i++) {
> > > > +             while (READ_ONCE(vcpu_last_completed_iteration[i]) != iteration)
> > > > +                     ;
> > > > +     }
> > > > +
> > > > +     pr_debug("\nGetting stats after repopulating memory:\n");
> > > > +     get_page_stats(vm, &stats_repopulated);
> > > > +
> > > > +     /*
> > > > +      * Tell the vCPU threads to quit.  No need to manually check that vCPUs
> > > > +      * have stopped running after disabling dirty logging, the join will
> > > > +      * wait for them to exit.
> > > > +      */
> > > > +     host_quit = true;
> > > > +     memstress_join_vcpu_threads(NR_VCPUS);
> > > > +
> > > > +     memstress_free_bitmaps(bitmaps, NR_SLOTS);
> > > > +     memstress_destroy_vm(vm);
> > > > +
> > > > +     /* Make assertions about the page counts. */
> > > > +     total_4k_pages = stats_populated.pages_4k;
> > > > +     total_4k_pages += stats_populated.pages_2m * 512;
> > > > +     total_4k_pages += stats_populated.pages_1g * 512 * 512;
> > > > +
> > > > +     /*
> > > > +      * Check that all huge pages were split. Since large pages can only
> > > > +      * exist in the data slot, and the vCPUs should have dirtied all pages
> > > > +      * in the data slot, there should be no huge pages left after splitting.
> > > > +      * Splitting happens at dirty log enable time without
> > > > +      * KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 and after the first clear pass
> > > > +      * with that capability.
> > > > +      */
> > > > +     if (dirty_log_manual_caps) {
> > > > +             TEST_ASSERT(stats_clear_pass[0].hugepages == 0,
> > >
> > > Consider using ASSERT_EQ() to simplify these checks. It will
> > > automatically print out the values for you, but you'll lose the
> > > contextual error message ("Unexpected huge page count after
> > > splitting..."). But maybe we could add support for a custom extra error
> > > string?
> > >
> > > __ASSERT_EQ(stats_clear_pass[0].hugepages, 0,
> > >             "Expected 0 hugepages after splitting");
> > >
> > > Or use a comment to document the context for the assertion. Whoever is
> > > debugging a failure is going to come look at the selftest code no matter
> > > what.
> > >
> > > I think I prefer ASSERT_EQ() + comment, especially since the comment
> > > pretty much already exists above.
> >
> > That's fair. I prefer the way it is because the resulting error
> > message is a lot easier to read and I don't need to look at the test
> > code to decrypt it. If I'm developing a feature and just running all
> > tests, it's nice to not have to track down the test source code.
> >
> > >
> > > > +                         "Unexpected huge page count after splitting. Expected 0, got %ld",
> > > > +                         stats_clear_pass[0].hugepages);
> > > > +             TEST_ASSERT(stats_clear_pass[0].pages_4k == total_4k_pages,
> > > > +                         "All memory should be mapped at 4k. Expected %ld 4k pages, got %ld",
> > > > +                         total_4k_pages, stats_clear_pass[0].pages_4k);
> > >
> > > Also assert that huge pages are *not* split when dirty logging is first
> > > enabled.
> >
> > Ah great idea. I felt like I was a little light on the assertions.
> > That'll be a good addition.
> >
> > >
> > > > +     } else {
> > ...
> > > > +
> > > > +     dirty_log_manual_caps =
> > > > +             kvm_check_cap(KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2);
> > > > +     dirty_log_manual_caps &= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE |
> > > > +                               KVM_DIRTY_LOG_INITIALLY_SET);
> > >
> > > Since this is a correctness test I think the test should, by default,
> > > test both KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE and 0, to ensure we get
> > > test coverage of both.
> > >
> > > And with that in place, there's probably no need for the -g flag.
> >
> > Good idea.
> >
> > >
> > > > +
> > > > +     guest_modes_append_default();
> > ...
> > > > +
> > > > +     if (!is_backing_src_hugetlb(p.backing_src)) {
> > > > +             pr_info("This test will only work reliably with HugeTLB memory. "
> > > > +                     "It can work with THP, but that is best effort.");
> > > > +             return KSFT_SKIP;
> > > > +     }
> > >
> > > backing_src only controls the memstress data slots. The rest of guest
> > > memory could be a source of noise for this test.
> >
> > That's true, but we compensate for that noise by taking a measurement
> > after the population pass. At that point the guest has executed all
> > it's code (or at least from all it's code pages) and touched every
> > page in the data slot. Since the other slots aren't backed with huge
> > pages, NX hugepages shouldn't be an issue. As a result, I would be
> > surprised if noise from the other memory became a problem.
> >
> > >
> > > > +
> > > > +     run_test(&p);
> > >
> > > Use for_each_guest_mode() to run against all supported guest modes.
> >
> > I'm not sure that would actually improve coverage. None of the page
> > splitting behavior depends on the mode AFAICT.
>
> You need to use for_each_guest_mode() for the ARM case. The issue is
> that whatever mode (guest page size and VA size) you pick might not be
> supported by the host. So, you first to explore what's available (via
> for_each_guest_mode()).

Actually, that's fixed by using the default mode, which picks the
first available
mode. I would prefer to use for_each_guest_mode() though, who knows and
something fails with some specific guest page size for some reason.

>
> Thanks,
> Ricardo
>
> >
> > >
> > > > +
> > > > +     return 0;
> > > > +}
> > > > --
> > > > 2.39.1.405.gd4c25cc71f-goog
> > > >

Powered by blists - more mailing lists