On Tue, Mar 09, 2021 at 09:10:19AM -0800, Sean Christopherson wrote: > @@ -2024,9 +2021,6 @@ static int __init init_hw_perf_events(void) > if (!x86_pmu.read) > x86_pmu.read = _x86_pmu_read; > > - if (!x86_pmu.guest_get_msrs) > - x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;
I suspect I might've been over eager here and we're now in trouble when *_pmu_init() clears x86_pmu.guest_get_msrs (like for instance on AMD). When that happens we need to restore __static_call_return0, otherwise the following static_call_update() will patch in a NOP and RAX will be garbage again. So I've taken the liberty to update the patch as below. --- Subject: x86/perf: Use RET0 as default for guest_get_msrs to handle "no PMU" case From: Sean Christopherson <[email protected]> Date: Tue, 9 Mar 2021 09:10:19 -0800 From: Sean Christopherson <[email protected]> Initialize x86_pmu.guest_get_msrs to return 0/NULL to handle the "nop" case. Patching in perf_guest_get_msrs_nop() during setup does not work if there is no PMU, as setup bails before updating the static calls, leaving x86_pmu.guest_get_msrs NULL and thus a complete nop. Ultimately, this causes VMX abort on VM-Exit due to KVM putting random garbage from the stack into the MSR load list. Add a comment in KVM to note that nr_msrs is valid if and only if the return value is non-NULL. Fixes: abd562df94d1 ("x86/perf: Use static_call for x86_pmu.guest_get_msrs") Reported-by: Dmitry Vyukov <[email protected]> Reported-by: [email protected] Suggested-by: Peter Zijlstra <[email protected]> Signed-off-by: Sean Christopherson <[email protected]> Signed-off-by: Peter Zijlstra (Intel) <[email protected]> Link: https://lkml.kernel.org/r/[email protected] --- v2: - Use __static_call_return0 to return NULL instead of manually checking the hook at invocation. [Peter] - Rebase to tip/sched/core, commit 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups"). arch/x86/events/core.c | 15 ++++++--------- arch/x86/kvm/vmx/vmx.c | 2 +- 2 files changed, 7 insertions(+), 10 deletions(-) --- a/arch/x86/events/core.c +++ b/arch/x86/events/core.c @@ -81,7 +81,11 @@ DEFINE_STATIC_CALL_NULL(x86_pmu_swap_tas DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs, *x86_pmu.drain_pebs); DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases); -DEFINE_STATIC_CALL_NULL(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs); +/* + * This one is magic, it will get called even when PMU init fails (because + * there is no PMU), in which case it should simply return NULL. + */ +DEFINE_STATIC_CALL_RET0(x86_pmu_guest_get_msrs, *x86_pmu.guest_get_msrs); u64 __read_mostly hw_cache_event_ids [PERF_COUNT_HW_CACHE_MAX] @@ -1944,13 +1948,6 @@ static void _x86_pmu_read(struct perf_ev x86_perf_event_update(event); } -static inline struct perf_guest_switch_msr * -perf_guest_get_msrs_nop(int *nr) -{ - *nr = 0; - return NULL; -} - static int __init init_hw_perf_events(void) { struct x86_pmu_quirk *quirk; @@ -2025,7 +2022,7 @@ static int __init init_hw_perf_events(vo x86_pmu.read = _x86_pmu_read; if (!x86_pmu.guest_get_msrs) - x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop; + x86_pmu.guest_get_msrs = (void *)&__static_call_return0; x86_pmu_static_call_update(); --- a/arch/x86/kvm/vmx/vmx.c +++ b/arch/x86/kvm/vmx/vmx.c @@ -6580,8 +6580,8 @@ static void atomic_switch_perf_msrs(stru int i, nr_msrs; struct perf_guest_switch_msr *msrs; + /* Note, nr_msrs may be garbage if perf_guest_get_msrs() returns NULL. */ msrs = perf_guest_get_msrs(&nr_msrs); - if (!msrs) return;
