[PATCH 10/12] KVM: PPC: Book3S HV: Fix accounting of stolen time
From: Paul Mackerras pau...@samba.org
Currently the code that accounts stolen time tends to overestimate the
stolen time, and will sometimes report more stolen time in a DTL
(dispatch trace log) entry than has elapsed since the last DTL entry.
This can cause guests to underflow the user or system time measured
for some tasks, leading to ridiculous CPU percentages and total runtimes
being reported by top and other utilities.
In addition, the current code was designed for the previous policy where
a vcore would only run when all the vcpus in it were runnable, and so
only counted stolen time on a per-vcore basis. Now that a vcore can
run while some of the vcpus in it are doing other things in the kernel
(e.g. handling a page fault), we need to count the time when a vcpu task
is preempted while it is not running as part of a vcore as stolen also.
To do this, we bring back the BUSY_IN_HOST vcpu state and extend the
vcpu_load/put functions to count preemption time while the vcpu is
in that state. Handling the transitions between the RUNNING and
BUSY_IN_HOST states requires checking and updating two variables
(accumulated time stolen and time last preempted), so we add a new
spinlock, vcpu-arch.tbacct_lock. This protects both the per-vcpu
stolen/preempt-time variables, and the per-vcore variables while this
vcpu is running the vcore.
Finally, we now don't count time spent in userspace as stolen time.
The task could be executing in userspace on behalf of the vcpu, or
it could be preempted, or the vcpu could be genuinely stopped. Since
we have no way of dividing up the time between these cases, we don't
count any of it as stolen.
Signed-off-by: Paul Mackerras pau...@samba.org
Signed-off-by: Alexander Graf ag...@suse.de
---
arch/powerpc/include/asm/kvm_host.h |5 ++
arch/powerpc/kvm/book3s_hv.c| 127 ++
2 files changed, 117 insertions(+), 15 deletions(-)
diff --git a/arch/powerpc/include/asm/kvm_host.h
b/arch/powerpc/include/asm/kvm_host.h
index 1e8cbd1..3093896 100644
--- a/arch/powerpc/include/asm/kvm_host.h
+++ b/arch/powerpc/include/asm/kvm_host.h
@@ -559,12 +559,17 @@ struct kvm_vcpu_arch {
unsigned long dtl_index;
u64 stolen_logged;
struct kvmppc_vpa slb_shadow;
+
+ spinlock_t tbacct_lock;
+ u64 busy_stolen;
+ u64 busy_preempt;
#endif
};
/* Values for vcpu-arch.state */
#define KVMPPC_VCPU_NOTREADY 0
#define KVMPPC_VCPU_RUNNABLE 1
+#define KVMPPC_VCPU_BUSY_IN_HOST 2
/* Values for vcpu-arch.io_gpr */
#define KVM_MMIO_REG_MASK 0x001f
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index 61d2934..8b3c470 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -60,23 +60,74 @@
/* Used to indicate that a guest page fault needs to be handled */
#define RESUME_PAGE_FAULT (RESUME_GUEST | RESUME_FLAG_ARCH1)
+/* Used as a null value for timebase values */
+#define TB_NIL (~(u64)0)
+
static void kvmppc_end_cede(struct kvm_vcpu *vcpu);
static int kvmppc_hv_setup_htab_rma(struct kvm_vcpu *vcpu);
+/*
+ * We use the vcpu_load/put functions to measure stolen time.
+ * Stolen time is counted as time when either the vcpu is able to
+ * run as part of a virtual core, but the task running the vcore
+ * is preempted or sleeping, or when the vcpu needs something done
+ * in the kernel by the task running the vcpu, but that task is
+ * preempted or sleeping. Those two things have to be counted
+ * separately, since one of the vcpu tasks will take on the job
+ * of running the core, and the other vcpu tasks in the vcore will
+ * sleep waiting for it to do that, but that sleep shouldn't count
+ * as stolen time.
+ *
+ * Hence we accumulate stolen time when the vcpu can run as part of
+ * a vcore using vc-stolen_tb, and the stolen time when the vcpu
+ * needs its task to do other things in the kernel (for example,
+ * service a page fault) in busy_stolen. We don't accumulate
+ * stolen time for a vcore when it is inactive, or for a vcpu
+ * when it is in state RUNNING or NOTREADY. NOTREADY is a bit of
+ * a misnomer; it means that the vcpu task is not executing in
+ * the KVM_VCPU_RUN ioctl, i.e. it is in userspace or elsewhere in
+ * the kernel. We don't have any way of dividing up that time
+ * between time that the vcpu is genuinely stopped, time that
+ * the task is actively working on behalf of the vcpu, and time
+ * that the task is preempted, so we don't count any of it as
+ * stolen.
+ *
+ * Updates to busy_stolen are protected by arch.tbacct_lock;
+ * updates to vc-stolen_tb are protected by the arch.tbacct_lock
+ * of the vcpu that has taken responsibility for running the vcore
+ * (i.e. vc-runner). The stolen times are measured in units of
+ * timebase ticks. (Note that the != TB_NIL checks below are
+ * purely defensive; they should never fail.)
+ */
+
void kvmppc_core_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
{
[PATCH 10/12] KVM: PPC: Book3S HV: Fix accounting of stolen time
From: Paul Mackerras pau...@samba.org
Currently the code that accounts stolen time tends to overestimate the
stolen time, and will sometimes report more stolen time in a DTL
(dispatch trace log) entry than has elapsed since the last DTL entry.
This can cause guests to underflow the user or system time measured
for some tasks, leading to ridiculous CPU percentages and total runtimes
being reported by top and other utilities.
In addition, the current code was designed for the previous policy where
a vcore would only run when all the vcpus in it were runnable, and so
only counted stolen time on a per-vcore basis. Now that a vcore can
run while some of the vcpus in it are doing other things in the kernel
(e.g. handling a page fault), we need to count the time when a vcpu task
is preempted while it is not running as part of a vcore as stolen also.
To do this, we bring back the BUSY_IN_HOST vcpu state and extend the
vcpu_load/put functions to count preemption time while the vcpu is
in that state. Handling the transitions between the RUNNING and
BUSY_IN_HOST states requires checking and updating two variables
(accumulated time stolen and time last preempted), so we add a new
spinlock, vcpu-arch.tbacct_lock. This protects both the per-vcpu
stolen/preempt-time variables, and the per-vcore variables while this
vcpu is running the vcore.
Finally, we now don't count time spent in userspace as stolen time.
The task could be executing in userspace on behalf of the vcpu, or
it could be preempted, or the vcpu could be genuinely stopped. Since
we have no way of dividing up the time between these cases, we don't
count any of it as stolen.
Signed-off-by: Paul Mackerras pau...@samba.org
Signed-off-by: Alexander Graf ag...@suse.de
---
arch/powerpc/include/asm/kvm_host.h |5 ++
arch/powerpc/kvm/book3s_hv.c| 127 ++
2 files changed, 117 insertions(+), 15 deletions(-)
diff --git a/arch/powerpc/include/asm/kvm_host.h
b/arch/powerpc/include/asm/kvm_host.h
index 1e8cbd1..3093896 100644
--- a/arch/powerpc/include/asm/kvm_host.h
+++ b/arch/powerpc/include/asm/kvm_host.h
@@ -559,12 +559,17 @@ struct kvm_vcpu_arch {
unsigned long dtl_index;
u64 stolen_logged;
struct kvmppc_vpa slb_shadow;
+
+ spinlock_t tbacct_lock;
+ u64 busy_stolen;
+ u64 busy_preempt;
#endif
};
/* Values for vcpu-arch.state */
#define KVMPPC_VCPU_NOTREADY 0
#define KVMPPC_VCPU_RUNNABLE 1
+#define KVMPPC_VCPU_BUSY_IN_HOST 2
/* Values for vcpu-arch.io_gpr */
#define KVM_MMIO_REG_MASK 0x001f
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index 61d2934..8b3c470 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -60,23 +60,74 @@
/* Used to indicate that a guest page fault needs to be handled */
#define RESUME_PAGE_FAULT (RESUME_GUEST | RESUME_FLAG_ARCH1)
+/* Used as a null value for timebase values */
+#define TB_NIL (~(u64)0)
+
static void kvmppc_end_cede(struct kvm_vcpu *vcpu);
static int kvmppc_hv_setup_htab_rma(struct kvm_vcpu *vcpu);
+/*
+ * We use the vcpu_load/put functions to measure stolen time.
+ * Stolen time is counted as time when either the vcpu is able to
+ * run as part of a virtual core, but the task running the vcore
+ * is preempted or sleeping, or when the vcpu needs something done
+ * in the kernel by the task running the vcpu, but that task is
+ * preempted or sleeping. Those two things have to be counted
+ * separately, since one of the vcpu tasks will take on the job
+ * of running the core, and the other vcpu tasks in the vcore will
+ * sleep waiting for it to do that, but that sleep shouldn't count
+ * as stolen time.
+ *
+ * Hence we accumulate stolen time when the vcpu can run as part of
+ * a vcore using vc-stolen_tb, and the stolen time when the vcpu
+ * needs its task to do other things in the kernel (for example,
+ * service a page fault) in busy_stolen. We don't accumulate
+ * stolen time for a vcore when it is inactive, or for a vcpu
+ * when it is in state RUNNING or NOTREADY. NOTREADY is a bit of
+ * a misnomer; it means that the vcpu task is not executing in
+ * the KVM_VCPU_RUN ioctl, i.e. it is in userspace or elsewhere in
+ * the kernel. We don't have any way of dividing up that time
+ * between time that the vcpu is genuinely stopped, time that
+ * the task is actively working on behalf of the vcpu, and time
+ * that the task is preempted, so we don't count any of it as
+ * stolen.
+ *
+ * Updates to busy_stolen are protected by arch.tbacct_lock;
+ * updates to vc-stolen_tb are protected by the arch.tbacct_lock
+ * of the vcpu that has taken responsibility for running the vcore
+ * (i.e. vc-runner). The stolen times are measured in units of
+ * timebase ticks. (Note that the != TB_NIL checks below are
+ * purely defensive; they should never fail.)
+ */
+
void kvmppc_core_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
{
