KVM: Inject asynchronous page fault into a PV guest if page is swapped out.
authorGleb Natapov <gleb@redhat.com>
Thu, 14 Oct 2010 09:22:53 +0000 (11:22 +0200)
committerAvi Kivity <avi@redhat.com>
Wed, 12 Jan 2011 09:23:17 +0000 (11:23 +0200)
Send async page fault to a PV guest if it accesses swapped out memory.
Guest will choose another task to run upon receiving the fault.

Allow async page fault injection only when guest is in user mode since
otherwise guest may be in non-sleepable context and will not be able
to reschedule.

Vcpu will be halted if guest will fault on the same page again or if
vcpu executes kernel code.

Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
arch/x86/include/asm/kvm_host.h
arch/x86/kvm/mmu.c
arch/x86/kvm/x86.c
include/trace/events/kvm.h
virt/kvm/async_pf.c

index 0d7039804b4cadd9f4df97845caf89450aa1fe86..167375cc49ff580ab1cbc9adbf8fc9da6e8140d9 100644 (file)
@@ -421,6 +421,7 @@ struct kvm_vcpu_arch {
                gfn_t gfns[roundup_pow_of_two(ASYNC_PF_PER_VCPU)];
                struct gfn_to_hva_cache data;
                u64 msr_val;
+               u32 id;
        } apf;
 };
 
@@ -596,6 +597,7 @@ struct kvm_x86_ops {
 };
 
 struct kvm_arch_async_pf {
+       u32 token;
        gfn_t gfn;
 };
 
@@ -819,6 +821,7 @@ void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
                                 struct kvm_async_pf *work);
 void kvm_arch_async_page_ready(struct kvm_vcpu *vcpu,
                               struct kvm_async_pf *work);
+bool kvm_arch_can_inject_async_page_present(struct kvm_vcpu *vcpu);
 extern bool kvm_find_async_pf_gfn(struct kvm_vcpu *vcpu, gfn_t gfn);
 
 #endif /* _ASM_X86_KVM_HOST_H */
index b2c60986a7ce361f33bedb52235367b8ffb3dfc4..64f90db369fba94d9c44bbdb3023bd9b8c20e0bb 100644 (file)
@@ -2592,6 +2592,7 @@ static int nonpaging_page_fault(struct kvm_vcpu *vcpu, gva_t gva,
 int kvm_arch_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn)
 {
        struct kvm_arch_async_pf arch;
+       arch.token = (vcpu->arch.apf.id++ << 12) | vcpu->vcpu_id;
        arch.gfn = gfn;
 
        return kvm_setup_async_pf(vcpu, gva, gfn, &arch);
index 063c0729676494e8155c3ece90df8af51ec428c3..ac4c368afd40747587f65ed2f0d6671901400a8d 100644 (file)
@@ -6248,20 +6248,53 @@ static void kvm_del_async_pf_gfn(struct kvm_vcpu *vcpu, gfn_t gfn)
        }
 }
 
+static int apf_put_user(struct kvm_vcpu *vcpu, u32 val)
+{
+
+       return kvm_write_guest_cached(vcpu->kvm, &vcpu->arch.apf.data, &val,
+                                     sizeof(val));
+}
+
 void kvm_arch_async_page_not_present(struct kvm_vcpu *vcpu,
                                     struct kvm_async_pf *work)
 {
-       trace_kvm_async_pf_not_present(work->gva);
-
-       kvm_make_request(KVM_REQ_APF_HALT, vcpu);
+       trace_kvm_async_pf_not_present(work->arch.token, work->gva);
        kvm_add_async_pf_gfn(vcpu, work->arch.gfn);
+
+       if (!(vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED) ||
+           kvm_x86_ops->get_cpl(vcpu) == 0)
+               kvm_make_request(KVM_REQ_APF_HALT, vcpu);
+       else if (!apf_put_user(vcpu, KVM_PV_REASON_PAGE_NOT_PRESENT)) {
+               vcpu->arch.fault.error_code = 0;
+               vcpu->arch.fault.address = work->arch.token;
+               kvm_inject_page_fault(vcpu);
+       }
 }
 
 void kvm_arch_async_page_present(struct kvm_vcpu *vcpu,
                                 struct kvm_async_pf *work)
 {
-       trace_kvm_async_pf_ready(work->gva);
-       kvm_del_async_pf_gfn(vcpu, work->arch.gfn);
+       trace_kvm_async_pf_ready(work->arch.token, work->gva);
+       if (is_error_page(work->page))
+               work->arch.token = ~0; /* broadcast wakeup */
+       else
+               kvm_del_async_pf_gfn(vcpu, work->arch.gfn);
+
+       if ((vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED) &&
+           !apf_put_user(vcpu, KVM_PV_REASON_PAGE_READY)) {
+               vcpu->arch.fault.error_code = 0;
+               vcpu->arch.fault.address = work->arch.token;
+               kvm_inject_page_fault(vcpu);
+       }
+}
+
+bool kvm_arch_can_inject_async_page_present(struct kvm_vcpu *vcpu)
+{
+       if (!(vcpu->arch.apf.msr_val & KVM_ASYNC_PF_ENABLED))
+               return true;
+       else
+               return !kvm_event_needs_reinjection(vcpu) &&
+                       kvm_x86_ops->interrupt_allowed(vcpu);
 }
 
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_exit);
index a78a5e5746320d290721cbcde47b6edf2b300688..9c2cc6a96e825b6853016b090734f31a539b4f52 100644 (file)
@@ -204,34 +204,39 @@ TRACE_EVENT(
 
 TRACE_EVENT(
        kvm_async_pf_not_present,
-       TP_PROTO(u64 gva),
-       TP_ARGS(gva),
+       TP_PROTO(u64 token, u64 gva),
+       TP_ARGS(token, gva),
 
        TP_STRUCT__entry(
+               __field(__u64, token)
                __field(__u64, gva)
                ),
 
        TP_fast_assign(
+               __entry->token = token;
                __entry->gva = gva;
                ),
 
-       TP_printk("gva %#llx not present", __entry->gva)
+       TP_printk("token %#llx gva %#llx not present", __entry->token,
+                 __entry->gva)
 );
 
 TRACE_EVENT(
        kvm_async_pf_ready,
-       TP_PROTO(u64 gva),
-       TP_ARGS(gva),
+       TP_PROTO(u64 token, u64 gva),
+       TP_ARGS(token, gva),
 
        TP_STRUCT__entry(
+               __field(__u64, token)
                __field(__u64, gva)
                ),
 
        TP_fast_assign(
+               __entry->token = token;
                __entry->gva = gva;
                ),
 
-       TP_printk("gva %#llx ready", __entry->gva)
+       TP_printk("token %#llx gva %#llx ready", __entry->token, __entry->gva)
 );
 
 TRACE_EVENT(
index 1f59498561b29a8d6fc50d00e174ed0fadca580e..60df9e059e6929fde5a71d2d70ed00fd02ef9365 100644 (file)
@@ -124,7 +124,8 @@ void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
 {
        struct kvm_async_pf *work;
 
-       if (list_empty_careful(&vcpu->async_pf.done))
+       if (list_empty_careful(&vcpu->async_pf.done) ||
+           !kvm_arch_can_inject_async_page_present(vcpu))
                return;
 
        spin_lock(&vcpu->async_pf.lock);