qemu: add per-vcpu delay stats

This patch adds delay time (steal time inside guest) to libvirt
domain per-vcpu stats. Delay time is an important performance metric.
It is a consequence of the overloaded CPU. Knowledge of the delay
time of a virtual machine helps to understand if it is affected and
estimate the impact.

As a result, it is possible to react exactly when needed and
rebalance the load between hosts. This is used by cloud providers
to provide quality of service, especially when the CPU is
oversubscribed.

It's more convenient to work with this metric in a context of a
libvirt domain. Any monitoring software may use this information.

Signed-off-by: Aleksei Zakharov <zaharov@selectel.ru>
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>

diff --git a/docs/manpages/virsh.rst b/docs/manpages/virsh.rst
index d44c36e2e9806c1e8d7a579aad0b7bf7e7b26bc3..9ff31a01605965477945451193e536b9b8b473aa 100644 (file)
--- a/docs/manpages/virsh.rst
+++ b/docs/manpages/virsh.rst
@@ -2295,6 +2295,10 @@ When selecting the *--state* group the following fields are returned:
 * ``vcpu.<num>.halted`` - virtual CPU <num> is halted: yes or
   no (may indicate the processor is idle or even disabled,
   depending on the architecture)
+* ``vcpu.<num>.delay`` - time the vCPU <num> thread was enqueued by the
+  host scheduler, but was waiting in the queue instead of running.
+  Exposed to the VM as a steal time.
+
 
 
 *--interface* returns:

diff --git a/src/libvirt-domain.c b/src/libvirt-domain.c
index 4af016687280e2fff90466c4192d930501ceea4c..e54d11e513414c398704b87ba1ceae3c580563e9 100644 (file)
--- a/src/libvirt-domain.c
+++ b/src/libvirt-domain.c
@@ -11693,6 +11693,10 @@ virConnectGetDomainCapabilities(virConnectPtr conn,
  *     "vcpu.<num>.halted" - virtual CPU <num> is halted, may indicate the
  *                           processor is idle or even disabled, depending
  *                           on the architecture)
+ *     "vcpu.<num>.delay" - time the vCPU <num> thread was enqueued by the
+ *                          host scheduler, but was waiting in the queue
+ *                          instead of running. Exposed to the VM as a steal
+ *                          time.
  *
  * VIR_DOMAIN_STATS_INTERFACE:
  *     Return network interface statistics (from domain point of view).

diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c
index 185fe903b2eca6b93d2da49c10da27ba56cfec4b..c7323e1fa532cab3bec4d333cb07c8c5fa12f436 100644 (file)
--- a/src/qemu/qemu_driver.c
+++ b/src/qemu/qemu_driver.c
@@ -1333,6 +1333,37 @@ static char *qemuConnectGetCapabilities(virConnectPtr conn) {
 }
 
 
+static int
+qemuGetSchedstatDelay(unsigned long long *cpudelay,
+                      pid_t pid,
+                      pid_t tid)
+{
+    g_autofree char *path = NULL;
+    g_autofree char *buf = NULL;
+
+    if (tid)
+        path = g_strdup_printf("/proc/%d/task/%d/schedstat", (int)pid, (int)tid);
+    else
+        path = g_strdup_printf("/proc/%d/schedstat", (int)pid);
+
+    /* This file might not exist (needs CONFIG_SCHED_INFO) */
+    if (!virFileExists(path))
+        return 0;
+
+    if (virFileReadAll(path, 1024, &buf) < 0)
+        return -1;
+
+    if (sscanf(buf, "%*u %llu", cpudelay) != 1) {
+        virReportError(VIR_ERR_INTERNAL_ERROR,
+                       _("Unable to parse schedstat info at '%s'"),
+                       path);
+        return -1;
+    }
+
+    return 0;
+}
+
+
 static int
 qemuGetSchedInfo(unsigned long long *cpuWait,
                  pid_t pid, pid_t tid)
@@ -1470,6 +1501,7 @@ static int
 qemuDomainHelperGetVcpus(virDomainObjPtr vm,
                          virVcpuInfoPtr info,
                          unsigned long long *cpuwait,
+                         unsigned long long *cpudelay,
                          int maxinfo,
                          unsigned char *cpumaps,
                          int maplen)
@@ -1529,6 +1561,11 @@ qemuDomainHelperGetVcpus(virDomainObjPtr vm,
                 return -1;
         }
 
+        if (cpudelay) {
+            if (qemuGetSchedstatDelay(&(cpudelay[ncpuinfo]), vm->pid, vcpupid) < 0)
+                return -1;
+        }
+
         ncpuinfo++;
     }
 
@@ -4873,7 +4910,7 @@ qemuDomainGetVcpus(virDomainPtr dom,
         goto cleanup;
     }
 
-    ret = qemuDomainHelperGetVcpus(vm, info, NULL, maxinfo, cpumaps, maplen);
+    ret = qemuDomainHelperGetVcpus(vm, info, NULL, NULL, maxinfo, cpumaps, maplen);
 
  cleanup:
     virDomainObjEndAPI(&vm);
@@ -17928,6 +17965,7 @@ qemuDomainGetStatsVcpu(virQEMUDriverPtr driver,
     int ret = -1;
     virVcpuInfoPtr cpuinfo = NULL;
     g_autofree unsigned long long *cpuwait = NULL;
+    g_autofree unsigned long long *cpudelay = NULL;
 
     if (virTypedParamListAddUInt(params, virDomainDefGetVcpus(dom->def),
                                  "vcpu.current") < 0)
@@ -17939,6 +17977,7 @@ qemuDomainGetStatsVcpu(virQEMUDriverPtr driver,
 
     cpuinfo = g_new0(virVcpuInfo, virDomainDefGetVcpus(dom->def));
     cpuwait = g_new0(unsigned long long, virDomainDefGetVcpus(dom->def));
+    cpudelay = g_new0(unsigned long long, virDomainDefGetVcpus(dom->def));
 
     if (HAVE_JOB(privflags) && virDomainObjIsActive(dom) &&
         qemuDomainRefreshVcpuHalted(driver, dom, QEMU_ASYNC_JOB_NONE) < 0) {
@@ -17947,7 +17986,7 @@ qemuDomainGetStatsVcpu(virQEMUDriverPtr driver,
             virResetLastError();
     }
 
-    if (qemuDomainHelperGetVcpus(dom, cpuinfo, cpuwait,
+    if (qemuDomainHelperGetVcpus(dom, cpuinfo, cpuwait, cpudelay,
                                  virDomainDefGetVcpus(dom->def),
                                  NULL, 0) < 0) {
         virResetLastError();
@@ -17972,6 +18011,10 @@ qemuDomainGetStatsVcpu(virQEMUDriverPtr driver,
                                        "vcpu.%u.wait", cpuinfo[i].number) < 0)
             goto cleanup;
 
+        if (virTypedParamListAddULLong(params, cpudelay[i],
+                                        "vcpu.%u.delay", cpuinfo[i].number) < 0)
+            goto cleanup;
+
         /* state below is extracted from the individual vcpu structs */
         if (!(vcpu = virDomainDefGetVcpu(dom->def, cpuinfo[i].number)))
             continue;