* 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
*
* Authors:
- * Aurelien Reynaud <collectd at wattapower.net>
+ * Aurélien Reynaud <collectd at wattapower.net>
**/
#include "collectd.h"
#include "common.h"
#include "plugin.h"
+
#include <sys/protosw.h>
#include <libperfstat.h>
-#include <sys/systemcfg.h>
#include <sys/utsname.h>
+/* XINTFRAC was defined in libperfstat.h somewhere between AIX 5.3 and 6.1 */
#ifndef XINTFRAC
+# include <sys/systemcfg.h>
# define XINTFRAC ((double)(_system_configuration.Xint) / \
(double)(_system_configuration.Xfrac))
#endif
-/* Max length of the type instance string */
-#define TYPE_INST_LEN (sizeof("lpar--total") + 2*sizeof(int) + 1)
+#define CLOCKTICKS_TO_TICKS(cticks) ((cticks) / XINTFRAC)
static const char *config_keys[] =
{
- "CpuPoolStats"
+ "CpuPoolStats",
+ "ReportBySerial"
};
static int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
-static int pool_stats = 0;
-
-/* As an LPAR can be moved transparently across physical systems
- * through Live Partition Mobility (LPM), and the resources we are
- * monitoring are tied to the underlying hardware, we need to keep
- * track on which physical server we are currently on. This is done
- * through the plugin instance which holds the chassis' serial.
- */
-static char plugin_inst[SYS_NMLN];
-
-static u_longlong_t last_time_base;
-static u_longlong_t last_pcpu_user,
- last_pcpu_sys,
- last_pcpu_idle,
- last_pcpu_wait;
-static u_longlong_t last_pool_idle_time = 0;
-static u_longlong_t last_idle_donated_purr = 0,
- last_busy_donated_purr = 0,
- last_busy_stolen_purr = 0,
- last_idle_stolen_purr = 0;
-static int donate_flag = 0;
-
-
-/* Save the current values for the next iteration */
-static void save_last_values (perfstat_partition_total_t *lparstats)
-{
- last_time_base = lparstats->timebase_last;
- last_pcpu_user = lparstats->puser;
- last_pcpu_sys = lparstats->psys;
- last_pcpu_idle = lparstats->pidle;
- last_pcpu_wait = lparstats->pwait;
-
- if (donate_flag)
- {
- last_idle_donated_purr = lparstats->idle_donated_purr;
- last_busy_donated_purr = lparstats->busy_donated_purr;
- last_busy_stolen_purr = lparstats->busy_stolen_purr;
- last_idle_stolen_purr = lparstats->idle_stolen_purr;
- }
+static _Bool pool_stats = 0;
+static _Bool report_by_serial = 0;
+#if PERFSTAT_SUPPORTS_DONATION
+static _Bool donate_flag = 0;
+#endif
+static char serial[SYS_NMLN];
- last_pool_idle_time = lparstats->pool_idle_time;
-}
+static perfstat_partition_total_t lparstats_old;
static int lpar_config (const char *key, const char *value)
{
else
pool_stats = 0;
}
+ else if (strcasecmp ("ReportBySerial", key) == 0)
+ {
+ if (IS_TRUE (value))
+ report_by_serial = 1;
+ else
+ report_by_serial = 0;
+ }
else
{
return (-1);
static int lpar_init (void)
{
- perfstat_partition_total_t lparstats;
+ int status;
- /* retrieve the initial metrics */
- if (!perfstat_partition_total (NULL, &lparstats,
- sizeof (perfstat_partition_total_t), 1))
+ /* Retrieve the initial metrics. Returns the number of structures filled. */
+ status = perfstat_partition_total (/* name = */ NULL, /* (must be NULL) */
+ &lparstats_old, sizeof (perfstat_partition_total_t),
+ /* number = */ 1 /* (must be 1) */);
+ if (status != 1)
{
- ERROR ("lpar plugin: perfstat_partition_total failed.");
+ char errbuf[1024];
+ ERROR ("lpar plugin: perfstat_partition_total failed: %s (%i)",
+ sstrerror (errno, errbuf, sizeof (errbuf)),
+ status);
return (-1);
}
- if (!lparstats.type.b.shared_enabled && lparstats.type.b.donate_enabled)
+#if PERFSTAT_SUPPORTS_DONATION
+ if (!lparstats_old.type.b.shared_enabled
+ && lparstats_old.type.b.donate_enabled)
{
donate_flag = 1;
}
+#endif
- /* save the initial data */
- save_last_values (&lparstats);
+ if (pool_stats && !lparstats_old.type.b.pool_util_authority)
+ {
+ WARNING ("lpar plugin: This partition does not have pool authority. "
+ "Disabling CPU pool statistics collection.");
+ pool_stats = 0;
+ }
return (0);
} /* int lpar_init */
vl.values = values;
vl.values_len = 1;
- sstrncpy (vl.host, hostname_g, sizeof (vl.host));
+ if (report_by_serial)
+ {
+ sstrncpy (vl.host, serial, sizeof (vl.host));
+ sstrncpy (vl.plugin_instance, hostname_g, sizeof (vl.plugin));
+ }
+ else
+ {
+ sstrncpy (vl.host, hostname_g, sizeof (vl.host));
+ }
sstrncpy (vl.plugin, "lpar", sizeof (vl.plugin));
- sstrncpy (vl.plugin_instance, plugin_inst, sizeof (vl.plugin));
- sstrncpy (vl.type, "lpar_pcpu", sizeof (vl.type));
+ sstrncpy (vl.type, "vcpu", sizeof (vl.type));
sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
plugin_dispatch_values (&vl);
-}
+} /* void lpar_submit */
static int lpar_read (void)
{
- u_longlong_t dlt_pcpu_user, dlt_pcpu_sys, dlt_pcpu_idle, dlt_pcpu_wait;
- u_longlong_t delta_time_base;
perfstat_partition_total_t lparstats;
+ int status;
struct utsname name;
-
- /* retrieve the current physical server's id and build the plugin
- instance's name */
+ u_longlong_t ticks;
+ u_longlong_t user_ticks, syst_ticks, wait_ticks, idle_ticks;
+ u_longlong_t consumed_ticks;
+ double entitled_proc_capacity;
+
+ /* An LPAR has the same serial number as the physical system it is currently
+ running on. It is a convenient way of tracking LPARs as they are moved
+ from chassis to chassis through Live Partition Mobility (LPM). */
if (uname (&name) != 0)
{
ERROR ("lpar plugin: uname failed.");
return (-1);
}
- sstrncpy (plugin_inst, name.machine, sizeof (plugin_inst));
+ sstrncpy (serial, name.machine, sizeof (serial));
- /* retrieve the current metrics */
- if (!perfstat_partition_total (NULL, &lparstats,
- sizeof (perfstat_partition_total_t), 1))
+ /* Retrieve the current metrics. Returns the number of structures filled. */
+ status = perfstat_partition_total (/* name = */ NULL, /* (must be NULL) */
+ &lparstats, sizeof (perfstat_partition_total_t),
+ /* number = */ 1 /* (must be 1) */);
+ if (status != 1)
{
- ERROR ("lpar plugin: perfstat_partition_total failed.");
+ char errbuf[1024];
+ ERROR ("lpar plugin: perfstat_partition_total failed: %s (%i)",
+ sstrerror (errno, errbuf, sizeof (errbuf)),
+ status);
return (-1);
}
- delta_time_base = lparstats.timebase_last - last_time_base;
- if (delta_time_base == 0)
+ /* Number of ticks since we last run. */
+ ticks = lparstats.timebase_last - lparstats_old.timebase_last;
+ if (ticks == 0)
{
- /* The system stats have not been updated since last time */
+ /* The stats have not been updated. Return now to avoid
+ * dividing by zero */
return (0);
}
- dlt_pcpu_user = lparstats.puser - last_pcpu_user;
- dlt_pcpu_sys = lparstats.psys - last_pcpu_sys;
- dlt_pcpu_idle = lparstats.pidle - last_pcpu_idle;
- dlt_pcpu_wait = lparstats.pwait - last_pcpu_wait;
-
- lpar_submit ("user", (double)dlt_pcpu_user / delta_time_base);
- lpar_submit ("sys", (double)dlt_pcpu_sys / delta_time_base);
- lpar_submit ("wait", (double)dlt_pcpu_wait / delta_time_base);
- lpar_submit ("idle", (double)dlt_pcpu_idle / delta_time_base);
- lpar_submit ("ent", (double)lparstats.entitled_proc_capacity / 100.0);
- lpar_submit ("max", (double)lparstats.max_proc_capacity / 100.0);
- lpar_submit ("min", (double)lparstats.min_proc_capacity / 100.0);
-
+ /*
+ * On a shared partition, we're "entitled" to a certain amount of
+ * processing power, for example 250/100 of a physical CPU. Processing
+ * capacity not used by the partition may be assigned to a different
+ * partition by the hypervisor, so "idle" is hopefully a very small
+ * number.
+ *
+ * A dedicated partition may donate its CPUs to another partition and
+ * may steal ticks from somewhere else (another partition or maybe the
+ * shared pool, I don't know --octo).
+ */
+
+ /* entitled_proc_capacity is in 1/100th of a CPU */
+ entitled_proc_capacity = 0.01 * ((double) lparstats.entitled_proc_capacity);
+ lpar_submit ("entitled", entitled_proc_capacity);
+
+ /* The number of ticks actually spent in the various states */
+ user_ticks = lparstats.puser - lparstats_old.puser;
+ syst_ticks = lparstats.psys - lparstats_old.psys;
+ wait_ticks = lparstats.pwait - lparstats_old.pwait;
+ idle_ticks = lparstats.pidle - lparstats_old.pidle;
+ consumed_ticks = user_ticks + syst_ticks + wait_ticks + idle_ticks;
+
+ lpar_submit ("user", (double) user_ticks / (double) ticks);
+ lpar_submit ("system", (double) syst_ticks / (double) ticks);
+ lpar_submit ("wait", (double) wait_ticks / (double) ticks);
+ lpar_submit ("idle", (double) idle_ticks / (double) ticks);
+
+#if PERFSTAT_SUPPORTS_DONATION
if (donate_flag)
{
- u_longlong_t dlt_busy_stolen, dlt_idle_stolen;
- u_longlong_t dlt_idle_donated, dlt_busy_donated;
-
- dlt_idle_donated = lparstats.idle_donated_purr - last_idle_donated_purr;
- dlt_busy_donated = lparstats.busy_donated_purr - last_busy_donated_purr;
- dlt_idle_stolen = lparstats.idle_stolen_purr - last_idle_stolen_purr;
- dlt_busy_stolen = lparstats.busy_stolen_purr - last_busy_stolen_purr;
-
- lpar_submit ("idle_donated", (double)dlt_idle_donated / delta_time_base);
- lpar_submit ("busy_donated", (double)dlt_busy_donated / delta_time_base);
- lpar_submit ("idle_stolen", (double)dlt_idle_stolen / delta_time_base);
- lpar_submit ("busy_stolen", (double)dlt_busy_stolen / delta_time_base);
+ /* donated => ticks given to another partition
+ * stolen => ticks received from another partition */
+ u_longlong_t idle_donated_ticks, busy_donated_ticks;
+ u_longlong_t idle_stolen_ticks, busy_stolen_ticks;
+
+ /* FYI: PURR == Processor Utilization of Resources Register
+ * SPURR == Scaled PURR */
+ idle_donated_ticks = lparstats.idle_donated_purr - lparstats_old.idle_donated_purr;
+ busy_donated_ticks = lparstats.busy_donated_purr - lparstats_old.busy_donated_purr;
+ idle_stolen_ticks = lparstats.idle_stolen_purr - lparstats_old.idle_stolen_purr;
+ busy_stolen_ticks = lparstats.busy_stolen_purr - lparstats_old.busy_stolen_purr;
+
+ lpar_submit ("idle_donated", (double) idle_donated_ticks / (double) ticks);
+ lpar_submit ("busy_donated", (double) busy_donated_ticks / (double) ticks);
+ lpar_submit ("idle_stolen", (double) idle_stolen_ticks / (double) ticks);
+ lpar_submit ("busy_stolen", (double) busy_stolen_ticks / (double) ticks);
+
+ /* Donated ticks will be accounted for as stolen ticks in other LPARs */
+ consumed_ticks += idle_stolen_ticks + busy_stolen_ticks;
}
+#endif
+
+ lpar_submit ("consumed", (double) consumed_ticks / (double) ticks);
if (pool_stats)
{
- if (!lparstats.type.b.pool_util_authority)
- {
- WARNING ("lpar plugin: this system does not have pool authority.");
- }
- else
- {
- u_longlong_t dlt_pit;
- double total, idle;
- char type[TYPE_INST_LEN];
-
- dlt_pit = lparstats.pool_idle_time - last_pool_idle_time;
- total = (double)lparstats.phys_cpus_pool;
- idle = (double)dlt_pit / XINTFRAC / (double)delta_time_base;
- ssnprintf (type, sizeof(type), "pool-%X-total", lparstats.pool_id);
- lpar_submit (type, total);
- ssnprintf (type, sizeof(type), "pool-%X-used", lparstats.pool_id);
- lpar_submit (type, total - idle);
- }
+ char typinst[DATA_MAX_NAME_LEN];
+ u_longlong_t pool_idle_cticks;
+ double pool_idle_cpus;
+ double pool_busy_cpus;
+
+ /* We're calculating "busy" from "idle" and the total number of
+ * CPUs, because the "busy" member didn't exist in early versions
+ * of libperfstat. It was added somewhere between AIX 5.3 ML5 and ML9. */
+ pool_idle_cticks = lparstats.pool_idle_time - lparstats_old.pool_idle_time;
+ pool_idle_cpus = CLOCKTICKS_TO_TICKS ((double) pool_idle_cticks) / (double) ticks;
+ pool_busy_cpus = ((double) lparstats.phys_cpus_pool) - pool_idle_cpus;
+ if (pool_busy_cpus < 0.0)
+ pool_busy_cpus = 0.0;
+
+ ssnprintf (typinst, sizeof (typinst), "pool-%X-busy", lparstats.pool_id);
+ lpar_submit (typinst, pool_busy_cpus);
+
+ ssnprintf (typinst, sizeof (typinst), "pool-%X-idle", lparstats.pool_id);
+ lpar_submit (typinst, pool_idle_cpus);
}
- save_last_values (&lparstats);
+ memcpy (&lparstats_old, &lparstats, sizeof (lparstats_old));
return (0);
} /* int lpar_read */
plugin_register_read ("lpar", lpar_read);
} /* void module_register */
-/* vim: set sw=2 sts=2 ts=8 : */
+/* vim: set sw=8 noet : */
projects / collectd.git / blobdiff