Ubstat system call
This article describes an interesting system call which was designed to pick beancounters statistics.
Contents
Overview
The system call appeared in the very first version of the OpenVZ. Its API is rather sloppy, but this is something we have to live with due to backward compatibility reasons.
The main intention of this system call is to allow a user space process get the beancounters statistics periodically. This statistics includes the fields observed in the /proc/user_beancounters file and one more field - the so called minheld value which is opposite to the maxheld one. As long as exporting the statistics the system call also notifies the task about the desired period has elapsed. The notification is performed by sending a signal to a process and this notification is one-shot.
How it works
The typical usage of this call is in performing the following steps.
- Request the amount of resources
- Get the IDs of all the living beancounters
- Setup a handler for some signal (e.g. USR1)
- Perform a system call to setup the notification
- Go do something (or sleep for ever)
In a signal handler one should just perform a respective system call to get the stats and schedule the next notification (yes, they are performed in one go; see below for more details).
API
The system call description is
long ubstat(int func, unsigned long luid, void *notif, void *buf, int size);
The macros and data typed used are declared in include/ub/ub_stat.h file.
Arguments description
func is like cmd in the ioctl system call. It can be one of
UBSTAT_READ_ONEto read basic stats for one resource. The desired resource itself should beor-ed with thefuncUBSTAT_READ_ALLto read basic stats about all the resourcesUBSTAT_READ_FULLto read extended stats about all the resourcesUBSTAT_UBLISTto get the ids of the beancountersUBSTAT_UBPARMNUMto get the number of resources used by the kernelUBSTAT_GETTIME
See below for what basic and extended stats mean.
luid is the desired beancounter ID. Only one beancounter can be checked at one call.
notif is the pointer to a ubnotifrq_t structure which describes the notification details (see below).
buf is the pointer to a chunk of memory, which will contain the data requested.
size is the buf memory size.
The statistics format
The format of data rerurned into the buffer depends on the function requested.
1. UBSTAT_READ_ONE format is
typedef unsigned long ubstattime_t;
typedef struct {
unsigned long maxheld;
unsigned long failcnt;
} ubstatparm_t;
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparm_t param[1];
}
It contains the time period for which the stats are returned and the maxheld and failcnt for the resource.
2. UBSTAT_READ_ALL format is
typedef unsigned long ubstattime_t;
typedef struct {
unsigned long maxheld;
unsigned long failcnt;
} ubstatparm_t;
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparm_t param[UB_RESOURCES];
}
It contains the same info as the UBSTAT_READ_ONE does, but for all the resources.
3. UBSTAT_READ_FULL format is
typedef unsigned long ubstattime_t;
typedef struct {
unsigned long barrier;
unsigned long limit;
unsigned long held;
unsigned long maxheld;
unsigned long minheld;
unsigned long failcnt;
unsigned long __unused1;
unsigned long __unused2;
} ubstatparmf_t;
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparmf_t param[UB_RESOURCES];
}
It contains the extended info for all the resources.
4. UBSTAT_UBLIST treats the buf to point to the unsigned long array.
5. UBSTAT_UBPARMNUM ignores the buf.
6. UBSTAT_GETTIME format is
typedef unsigned long ubstattime_t;
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstattime_t cur_time;
};
It returns the time interval within which the stats are collected and the current time.
All the times used are in seconds.
Notification
The notification info is passed via the notif argument and is being set up for all the functions except the UBLIST and the UBPARNUM. The notification is one-shot, but note that once you requested the statistics the next shot is scheduled at the same time.
The notif should point to
typedef struct {
long maxinterval;
int signum;
} ubnotifrq_t;
The maxinterval is the time after which the notification will be delivered. It should be more than 1 (second).
The signum is the signal that will be sent to notify.
Return value
The system call returns -1 in case error has occurred. In case of UBSTAT_UBPARMNUM</code it returns UB_RESOURCES and in all other cases it returns the amount of bytes written to the buf.
Demo
The following program demonstrates how you can (but not should) use the described API. This example is deliberately made very stupid and simple to demonstrate the main idea and will only work on x86_64.
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <ub_stat.h>
#include <signal.h>
#define UBSTAT_BUFSIZE 4096
#define UBSTAT_NOTIFSIG SIGUSR1
#define UB_RESOURCES 24
static char ubstat_buf[UBSTAT_BUFSIZE];
static int luid, func;
static ubnotifrq_t notif;
static void (*print_stat)(void *buf);
#define sys_ubstat(args...) syscall(503, ## args)
static void usage(void)
{
printf("ubstat <id> <period> <resource>\n");
printf("resource can be either num or:\n");
printf("-1 for all maxheld and failcnt\n");
printf("-2 for all at all\n");
}
static void print_stat_one(void *buf)
{
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparm_t param[1];
} *data;
data = buf;
printf("maxheld: %lu\n", data->param[0].maxheld);
printf("failcnt: %lu\n", data->param[0].failcnt);
}
static void print_stat_all(void *buf)
{
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparm_t param[UB_RESOURCES];
} *data;
int res;
data = buf;
for (res = 0; res < UB_RESOURCES; res++) {
printf("res %d\n", res);
printf("\tmaxheld: %lu\n", data->param[res].maxheld);
printf("\tfailcnt: %lu\n", data->param[res].failcnt);
}
}
static void print_stat_full(void *buf)
{
struct {
ubstattime_t start_time;
ubstattime_t end_time;
ubstatparmf_t param[UB_RESOURCES];
} *data;
int res;
data = buf;
for (res = 0; res < UB_RESOURCES; res++) {
printf("res %d\n", res);
printf("minheld: %lu\n", data->param[res].minheld);
printf("maxheld: %lu\n", data->param[res].maxheld);
printf("failcnt: %lu\n", data->param[res].failcnt);
}
}
static inline int res2func(int resource)
{
if (resource >= 0) {
print_stat = print_stat_one;
return UBSTAT_READ_ONE | resource;
}
if (resource == -1) {
print_stat = print_stat_all;
return UBSTAT_READ_ALL;
}
if (resource == -2) {
print_stat = print_stat_full;
return UBSTAT_READ_FULL;
}
printf("Bad resource %d\n", resource);
exit(1);
}
static void do_notify(int x)
{
int err;
err = sys_ubstat(func, luid, (unsigned long)¬if,
ubstat_buf, UBSTAT_BUFSIZE);
if (err < 0) {
perror("Can't set stat");
exit(0);
}
print_stat(ubstat_buf);
}
static int do_ubstat(int id, int period, int resource)
{
luid = id;
func = res2func(resource);
notif.maxinterval = period;
notif.signum = UBSTAT_NOTIFSIG;
signal(UBSTAT_NOTIFSIG, do_notify);
do_notify(0);
}
int main(int argc, char **argv)
{
int id, period, res;
if (argc == 1) {
usage();
return 0;
}
id = atoi(argv[1]);
period = atoi(argv[2]);
res = atoi(argv[3]);
do_ubstat(id, period, res);
while (1)
sleep(10);
return 0;
}
Implementation constraints
Unfortunately the API is not architecture independent and thus 32-bit application will simply not work on x86_64.
