KVM: async_pf: Fix #DF due to inject "Page not Present" and "Page Ready" exceptions...
authorWanpeng Li <wanpeng.li@hotmail.com>
Thu, 14 Sep 2017 10:54:16 +0000 (03:54 -0700)
committerRadim Krčmář <rkrcmar@redhat.com>
Thu, 14 Sep 2017 16:43:43 +0000 (18:43 +0200)
qemu-system-x86-8600  [004] d..1  7205.687530: kvm_entry: vcpu 2
qemu-system-x86-8600  [004] ....  7205.687532: kvm_exit: reason EXCEPTION_NMI rip 0xffffffffa921297d info ffffeb2c0e44e018 80000b0e
qemu-system-x86-8600  [004] ....  7205.687532: kvm_page_fault: address ffffeb2c0e44e018 error_code 0
qemu-system-x86-8600  [004] ....  7205.687620: kvm_try_async_get_page: gva = 0xffffeb2c0e44e018, gfn = 0x427e4e
qemu-system-x86-8600  [004] .N..  7205.687628: kvm_async_pf_not_present: token 0x8b002 gva 0xffffeb2c0e44e018
    kworker/4:2-7814  [004] ....  7205.687655: kvm_async_pf_completed: gva 0xffffeb2c0e44e018 address 0x7fcc30c4e000
qemu-system-x86-8600  [004] ....  7205.687703: kvm_async_pf_ready: token 0x8b002 gva 0xffffeb2c0e44e018
qemu-system-x86-8600  [004] d..1  7205.687711: kvm_entry: vcpu 2

After running some memory intensive workload in guest, I catch the kworker
which completes the GUP too quickly, and queues an "Page Ready" #PF exception
after the "Page not Present" exception before the next vmentry as the above
trace which will result in #DF injected to guest.

This patch fixes it by clearing the queue for "Page not Present" if "Page Ready"
occurs before the next vmentry since the GUP has already got the required page
and shadow page table has already been fixed by "Page Ready" handler.

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
Fixes: 7c90705bf2a3 ("KVM: Inject asynchronous page fault into a PV guest if page is swapped out.")
[Changed indentation and added clearing of injected. - Radim]
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/x86.c

index bfda79fb102c459c05393311fd81633cd8ba45d6..cd17b7d9a1076c1d28904bc4cd1ea06af5e0f2c0 100644 (file)
@@ -8631,6 +8631,13 @@ static int apf_put_user(struct kvm_vcpu *vcpu, u32 val)
                                      sizeof(val));
 }
 
+static int apf_get_user(struct kvm_vcpu *vcpu, u32 *val)
+{
+
+       return kvm_read_guest_cached(vcpu->kvm, &vcpu->arch.apf.data, val,
+                                     sizeof(u32));
+}
+
 void kvm_arch_async_page_not_present(struct kvm_vcpu *vcpu,
                                     struct kvm_async_pf *work)
 {
@@ -8658,6 +8665,7 @@ void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
                                 struct kvm_async_pf *work)
 {
        struct x86_exception fault;
+       u32 val;
 
        if (work->wakeup_all)
                work->arch.token = ~0; /* broadcast wakeup */
@@ -8665,15 +8673,26 @@ void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
                kvm_del_async_pf_gfn(vcpu, work->arch.gfn);
        trace_kvm_async_pf_ready(work->arch.token, work->gva);
 
-       if ((vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED) &&
-           !apf_put_user(vcpu, KVM_PV_REASON_PAGE_READY)) {
-               fault.vector = PF_VECTOR;
-               fault.error_code_valid = true;
-               fault.error_code = 0;
-               fault.nested_page_fault = false;
-               fault.address = work->arch.token;
-               fault.async_page_fault = true;
-               kvm_inject_page_fault(vcpu, &fault);
+       if (vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED &&
+           !apf_get_user(vcpu, &val)) {
+               if (val == KVM_PV_REASON_PAGE_NOT_PRESENT &&
+                   vcpu->arch.exception.pending &&
+                   vcpu->arch.exception.nr == PF_VECTOR &&
+                   !apf_put_user(vcpu, 0)) {
+                       vcpu->arch.exception.injected = false;
+                       vcpu->arch.exception.pending = false;
+                       vcpu->arch.exception.nr = 0;
+                       vcpu->arch.exception.has_error_code = false;
+                       vcpu->arch.exception.error_code = 0;
+               } else if (!apf_put_user(vcpu, KVM_PV_REASON_PAGE_READY)) {
+                       fault.vector = PF_VECTOR;
+                       fault.error_code_valid = true;
+                       fault.error_code = 0;
+                       fault.nested_page_fault = false;
+                       fault.address = work->arch.token;
+                       fault.async_page_fault = true;
+                       kvm_inject_page_fault(vcpu, &fault);
+               }
        }
        vcpu->arch.apf.halted = false;
        vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;