[PATCH] fio: show group stats after individual job stats
[disktools.git] / fio.c
diff --git a/fio.c b/fio.c
index 2abfb7e71fe8a5ec4dfe30ffafab8a1c682cb7af..aa45e3c2f95052c3a64ec48e8d8e4c395fe976e7 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include <errno.h>
 #include <signal.h>
 #include <time.h>
-#include <ctype.h>
-#include <sched.h>
-#include <libaio.h>
 #include <math.h>
-#include <limits.h>
 #include <assert.h>
-#include <sys/time.h>
+#include <pthread.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
-#include <semaphore.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
 #include <sys/ioctl.h>
 #include <asm/unistd.h>
-#include <asm/types.h>
-#include <asm/bitops.h>
 
-#include "arch.h"
-#include "list.h"
-#include "md5.h"
-
-#ifndef BLKGETSIZE64
-#define BLKGETSIZE64   _IOR(0x12,114,size_t)
-#endif
-
-#define MAX_JOBS       (1024)
-
-static int ioprio_set(int which, int who, int ioprio)
-{
-       return syscall(__NR_ioprio_set, which, who, ioprio);
-}
-
-/*
- * we want fadvise64 really, but it's so tangled... later
- */
-static int fadvise(int fd, loff_t offset, size_t len, int advice)
-{
-#if 0
-       return syscall(__NR_fadvise64, fd, offset, offset >> 32, len, advice);
-#else
-       return posix_fadvise(fd, (off_t) offset, len, advice);
-#endif
-}
-
-enum {
-       IOPRIO_WHO_PROCESS = 1,
-       IOPRIO_WHO_PGRP,
-       IOPRIO_WHO_USER,
-};
-
-#define IOPRIO_CLASS_SHIFT     13
+#include "fio.h"
 
 #define MASK   (4095)
 
-#define DEF_BS         (4096)
-#define DEF_TIMEOUT    (0)
-#define DEF_RATE_CYCLE (1000)
-#define DEF_ODIRECT    (1)
-#define DEF_SEQUENTIAL (1)
-#define DEF_RAND_REPEAT        (1)
-#define DEF_OVERWRITE  (0)
-#define DEF_CREATE     (1)
-#define DEF_INVALIDATE (1)
-#define DEF_SYNCIO     (0)
-#define DEF_RANDSEED   (0xb1899bedUL)
-#define DEF_BWAVGTIME  (500)
-#define DEF_CREATE_SER (1)
-#define DEF_CREATE_FSYNC       (1)
-#define DEF_LOOPS      (1)
-#define DEF_VERIFY     (0)
-#define DEF_STONEWALL  (0)
-
 #define ALIGN(buf)     (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
 
-static int repeatable = DEF_RAND_REPEAT;
-static int rate_quit = 1;
-static int write_lat_log;
-static int write_bw_log;
-static int exitall_on_terminate;
-
-static int thread_number;
-static char *ini_file;
-
-static int max_jobs = MAX_JOBS;
-
-static char run_str[MAX_JOBS + 1];
-
-static int shm_id;
-
-enum {
-       DDIR_READ = 0,
-       DDIR_WRITE,
-};
+int groupid = 0;
+int thread_number = 0;
+char run_str[MAX_JOBS + 1];
+int shm_id = 0;
 
 /*
  * thread life cycle
@@ -133,11 +60,6 @@ enum {
        TD_REAPED,
 };
 
-enum {
-       MEM_MALLOC,
-       MEM_SHM,
-};
-
 /*
  * The io unit
  */
@@ -153,194 +75,29 @@ struct io_u {
        struct list_head list;
 };
 
-struct io_stat {
-       unsigned long val;
-       unsigned long val_sq;
-       unsigned long max_val;
-       unsigned long min_val;
-       unsigned long samples;
-};
-
-struct io_sample {
-       unsigned long time;
-       unsigned long val;
-};
-
-struct io_log {
-       unsigned long nr_samples;
-       unsigned long max_samples;
-       struct io_sample *log;
-};
-
-struct io_piece {
-       struct list_head list;
-       unsigned long long offset;
-       unsigned int len;
-};
-
-#define FIO_HDR_MAGIC  0xf00baaef
-
-struct verify_header {
-       unsigned int fio_magic;
-       unsigned int len;
-       char md5_digest[MD5_HASH_WORDS * 4];
-};
-
-#define td_read(td)            ((td)->ddir == DDIR_READ)
-#define td_write(td)           ((td)->ddir == DDIR_WRITE)
 #define should_fsync(td)       (td_write(td) && !(td)->odirect)
 
-#define BLOCKS_PER_MAP         (8 * sizeof(long))
-#define TO_MAP_BLOCK(td, b)    ((b) - ((td)->file_offset / (td)->min_bs))
-#define RAND_MAP_IDX(td, b)    (TO_MAP_BLOCK(td, b) / BLOCKS_PER_MAP)
-#define RAND_MAP_BIT(td, b)    (TO_MAP_BLOCK(td, b) & (BLOCKS_PER_MAP - 1))
-
-struct thread_data {
-       char file_name[256];
-       int thread_number;
-       int error;
-       int fd;
-       pid_t pid;
-       char *orig_buffer;
-       volatile int terminate;
-       volatile int runstate;
-       volatile int old_runstate;
-       unsigned int ddir;
-       unsigned int ioprio;
-       unsigned int sequential;
-       unsigned int bs;
-       unsigned int min_bs;
-       unsigned int max_bs;
-       unsigned int odirect;
-       unsigned int thinktime;
-       unsigned int fsync_blocks;
-       unsigned int start_delay;
-       unsigned int timeout;
-       unsigned int use_aio;
-       unsigned int create_file;
-       unsigned int overwrite;
-       unsigned int invalidate_cache;
-       unsigned int bw_avg_time;
-       unsigned int create_serialize;
-       unsigned int create_fsync;
-       unsigned int loops;
-       unsigned long long file_size;
-       unsigned long long file_offset;
-       unsigned int sync_io;
-       unsigned int mem_type;
-       unsigned int verify;
-       unsigned int stonewall;
-       cpu_set_t cpumask;
-
-       struct drand48_data bsrange_state;
-       struct drand48_data verify_state;
-
-       int shm_id;
-
-       off_t cur_off;
-
-       io_context_t aio_ctx;
-       unsigned int aio_depth;
-       struct io_event *aio_events;
-
-       unsigned int cur_depth;
-       struct list_head io_u_freelist;
-       struct list_head io_u_busylist;
-
-       unsigned int rate;
-       unsigned int ratemin;
-       unsigned int ratecycle;
-       unsigned long rate_usec_cycle;
-       long rate_pending_usleep;
-       unsigned long rate_bytes;
-       struct timeval lastrate;
-
-       unsigned long runtime;          /* sec */
-       unsigned long long io_size;
-
-       unsigned long io_blocks;
-       unsigned long io_bytes;
-       unsigned long this_io_bytes;
-       unsigned long last_bytes;
-       sem_t mutex;
-
-       struct drand48_data random_state;
-       unsigned long *file_map;
-       unsigned int num_maps;
-
-       /*
-        * bandwidth and latency stats
-        */
-       struct io_stat clat_stat;               /* completion latency */
-       struct io_stat slat_stat;               /* submission latency */
-
-       struct io_stat bw_stat;                 /* bandwidth stats */
-       unsigned long stat_io_bytes;
-       struct timeval stat_sample_time;
-
-       struct io_log *lat_log;
-       struct io_log *bw_log;
-
-       struct timeval start;
-       struct rusage ru_start;
-       struct rusage ru_end;
-
-       struct list_head io_hist_list;
-};
-
-static struct thread_data *threads;
-static struct thread_data def_thread;
-
 static sem_t startup_sem;
 
-static void sig_handler(int sig)
+#define TERMINATE_ALL          (-1)
+
+static void terminate_threads(int groupid)
 {
        int i;
 
        for (i = 0; i < thread_number; i++) {
                struct thread_data *td = &threads[i];
 
-               td->terminate = 1;
-               td->start_delay = 0;
+               if (groupid == TERMINATE_ALL || groupid == td->groupid) {
+                       td->terminate = 1;
+                       td->start_delay = 0;
+               }
        }
 }
 
-static int init_random_state(struct thread_data *td)
+static void sig_handler(int sig)
 {
-       unsigned long seed;
-       int fd, num_maps, blocks;
-
-       fd = open("/dev/random", O_RDONLY);
-       if (fd == -1) {
-               td->error = errno;
-               return 1;
-       }
-
-       if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
-               td->error = EIO;
-               close(fd);
-               return 1;
-       }
-
-       close(fd);
-
-       srand48_r(seed, &td->bsrange_state);
-       srand48_r(seed, &td->verify_state);
-
-       if (td->sequential)
-               return 0;
-
-       if (repeatable)
-               seed = DEF_RANDSEED;
-
-       blocks = (td->io_size + td->min_bs - 1) / td->min_bs;
-       num_maps = blocks / BLOCKS_PER_MAP;
-       td->file_map = malloc(num_maps * sizeof(long));
-       td->num_maps = num_maps;
-       memset(td->file_map, 0, num_maps * sizeof(long));
-
-       srand48_r(seed, &td->random_state);
-       return 0;
+       terminate_threads(TERMINATE_ALL);
 }
 
 static unsigned long utime_since(struct timeval *s, struct timeval *e)
@@ -566,7 +323,7 @@ static void __usec_sleep(int usec)
 
        gettimeofday(&start, NULL);
        while (utime_since_now(&start) < usec)
-                __asm__ __volatile__("rep;nop": : :"memory");
+               nop;
 }
 
 static void usec_sleep(int usec)
@@ -637,7 +394,7 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                if (rate < td->ratemin) {
                        printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
                        if (rate_quit)
-                               sig_handler(0);
+                               terminate_threads(td->groupid);
                        return 1;
                }
        }
@@ -1355,37 +1112,6 @@ static int init_io_u(struct thread_data *td)
        return 0;
 }
 
-static void setup_log(struct io_log **log)
-{
-       struct io_log *l = malloc(sizeof(*l));
-
-       l->nr_samples = 0;
-       l->max_samples = 1024;
-       l->log = malloc(l->max_samples * sizeof(struct io_sample));
-       *log = l;
-}
-
-static void finish_log(struct thread_data *td, struct io_log *log, char *name)
-{
-       char file_name[128];
-       FILE *f;
-       unsigned int i;
-
-       sprintf(file_name, "client%d_%s.log", td->thread_number, name);
-       f = fopen(file_name, "w");
-       if (!f) {
-               perror("fopen log");
-               return;
-       }
-
-       for (i = 0; i < log->nr_samples; i++)
-               fprintf(f, "%lu, %lu\n", log->log[i].time, log->log[i].val);
-
-       fclose(f);
-       free(log->log);
-       free(log);
-}
-
 static int create_file(struct thread_data *td)
 {
        unsigned long long left;
@@ -1580,21 +1306,27 @@ static void clear_io_state(struct thread_data *td)
                memset(td->file_map, 0, td->num_maps * sizeof(long));
 }
 
-static void *thread_main(int shm_id, int offset, char *argv[])
+static void update_rusage_stat(struct thread_data *td)
 {
-       struct thread_data *td;
-       int ret = 1;
-       void *data;
+       if (!td->runtime)
+               return;
 
-       setsid();
+       getrusage(RUSAGE_SELF, &td->ru_end);
 
-       data = shmat(shm_id, NULL, 0);
-       if (data == (void *) -1) {
-               perror("shmat");
-               return NULL;
-       }
+       td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
+       td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
+       td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
 
-       td = data + offset * sizeof(struct thread_data);
+       
+       memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
+}
+
+static void *thread_main(void *data)
+{
+       struct thread_data *td = data;
+       int ret = 1;
+
+       setsid();
        td->pid = getpid();
 
        if (init_io_u(td))
@@ -1605,8 +1337,6 @@ static void *thread_main(int shm_id, int offset, char *argv[])
                goto err;
        }
 
-       sprintf(argv[0], "fio%d", offset);
-
        if (td->use_aio && init_aio(td))
                goto err;
 
@@ -1626,12 +1356,10 @@ static void *thread_main(int shm_id, int offset, char *argv[])
        if (init_random_state(td))
                goto err;
 
-       gettimeofday(&td->start, NULL);
-
-       getrusage(RUSAGE_SELF, &td->ru_start);
-
        while (td->loops--) {
-               gettimeofday(&td->stat_sample_time, NULL);
+               getrusage(RUSAGE_SELF, &td->ru_start);
+               gettimeofday(&td->start, NULL);
+               memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
                        memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
@@ -1647,6 +1375,9 @@ static void *thread_main(int shm_id, int offset, char *argv[])
                if (td->error)
                        break;
 
+               td->runtime += mtime_since_now(&td->start);
+               update_rusage_stat(td);
+
                if (!td->verify)
                        continue;
 
@@ -1661,8 +1392,6 @@ static void *thread_main(int shm_id, int offset, char *argv[])
                        break;
        }
 
-       td->runtime = mtime_since_now(&td->start);
-       getrusage(RUSAGE_SELF, &td->ru_end);
        ret = 0;
 
        if (td->bw_log)
@@ -1671,7 +1400,7 @@ static void *thread_main(int shm_id, int offset, char *argv[])
                finish_log(td, td->lat_log, "lat");
 
        if (exitall_on_terminate)
-               sig_handler(0);
+               terminate_threads(td->groupid);
 
 err:
        if (td->fd != -1) {
@@ -1686,19 +1415,25 @@ err:
                sem_wait(&td->mutex);
        }
        td_set_runstate(td, TD_EXITED);
-       shmdt(data);
        return NULL;
+
 }
 
-static void free_shm(void)
+static void *fork_main(int shm_id, int offset)
 {
-       struct shmid_ds sbuf;
+       struct thread_data *td;
+       void *data;
 
-       if (threads) {
-               shmdt(threads);
-               threads = NULL;
-               shmctl(shm_id, IPC_RMID, &sbuf);
+       data = shmat(shm_id, NULL, 0);
+       if (data == (void *) -1) {
+               perror("shmat");
+               return NULL;
        }
+
+       td = data + offset * sizeof(struct thread_data);
+       thread_main(td);
+       shmdt(data);
+       return NULL;
 }
 
 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
@@ -1718,10 +1453,11 @@ static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
        return 1;
 }
 
-static void show_thread_status(struct thread_data *td)
+static void show_thread_status(struct thread_data *td,
+                              struct group_run_stats *rs)
 {
        int prio, prio_class;
-       unsigned long min, max, bw = 0, ctx;
+       unsigned long min, max, bw = 0;
        double mean, dev, usr_cpu, sys_cpu;
 
        if (!td->io_bytes && !td->error)
@@ -1733,551 +1469,28 @@ static void show_thread_status(struct thread_data *td)
        prio = td->ioprio & 0xff;
        prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
 
-       printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_bytes >> 20, bw, td->runtime);
+       printf("Client%d (g=%d): err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->groupid, td->error, td->io_bytes >> 20, bw, td->runtime);
 
        if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
                printf("  slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
        if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev))
                printf("  clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
-       if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
-               printf("  bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
-
-       if (td->runtime) {
-               unsigned long t;
+       if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev)) {
+               double p_of_agg;
 
-               t = mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
-               usr_cpu = (double) t * 100 / (double) td->runtime;
+               p_of_agg = mean * 100 / (double) rs->agg[td->ddir];
+               printf("  bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
+       }
 
-               t = mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
-               sys_cpu = (double) t * 100 / (double) td->runtime;
+       if (td->runtime) {
+               usr_cpu = (double) td->usr_time * 100 / (double) td->runtime;
+               sys_cpu = (double) td->sys_time * 100 / (double) td->runtime;
        } else {
                usr_cpu = 0;
                sys_cpu = 0;
        }
 
-       ctx = td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
-
-       printf("  cpu        : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ctx);
-}
-
-static int setup_rate(struct thread_data *td)
-{
-       int nr_reads_per_sec;
-
-       if (!td->rate)
-               return 0;
-
-       if (td->rate < td->ratemin) {
-               fprintf(stderr, "min rate larger than nominal rate\n");
-               return -1;
-       }
-
-       nr_reads_per_sec = (td->rate * 1024) / td->min_bs;
-       td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
-       td->rate_pending_usleep = 0;
-       return 0;
-}
-
-static struct thread_data *get_new_job(int global)
-{
-       struct thread_data *td;
-
-       if (global)
-               return &def_thread;
-       if (thread_number >= max_jobs)
-               return NULL;
-
-       td = &threads[thread_number++];
-       memset(td, 0, sizeof(*td));
-
-       td->fd = -1;
-       td->thread_number = thread_number;
-
-       td->ddir = def_thread.ddir;
-       td->ioprio = def_thread.ioprio;
-       td->sequential = def_thread.sequential;
-       td->bs = def_thread.bs;
-       td->min_bs = def_thread.min_bs;
-       td->max_bs = def_thread.max_bs;
-       td->odirect = def_thread.odirect;
-       td->thinktime = def_thread.thinktime;
-       td->fsync_blocks = def_thread.fsync_blocks;
-       td->start_delay = def_thread.start_delay;
-       td->timeout = def_thread.timeout;
-       td->use_aio = def_thread.use_aio;
-       td->create_file = def_thread.create_file;
-       td->overwrite = def_thread.overwrite;
-       td->invalidate_cache = def_thread.invalidate_cache;
-       td->file_size = def_thread.file_size;
-       td->file_offset = def_thread.file_offset;
-       td->rate = def_thread.rate;
-       td->ratemin = def_thread.ratemin;
-       td->ratecycle = def_thread.ratecycle;
-       td->aio_depth = def_thread.aio_depth;
-       td->sync_io = def_thread.sync_io;
-       td->mem_type = def_thread.mem_type;
-       td->bw_avg_time = def_thread.bw_avg_time;
-       td->create_serialize = def_thread.create_serialize;
-       td->create_fsync = def_thread.create_fsync;
-       td->loops = def_thread.loops;
-       td->verify = def_thread.verify;
-       td->stonewall = def_thread.stonewall;
-       memcpy(&td->cpumask, &def_thread.cpumask, sizeof(td->cpumask));
-
-       return td;
-}
-
-static void put_job(struct thread_data *td)
-{
-       memset(&threads[td->thread_number - 1], 0, sizeof(*td));
-       thread_number--;
-}
-
-static int add_job(struct thread_data *td, const char *filename, int prioclass,
-                  int prio)
-{
-       if (td == &def_thread)
-               return 0;
-
-       strcpy(td->file_name, filename);
-       sem_init(&td->mutex, 1, 0);
-       td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
-
-       td->clat_stat.min_val = ULONG_MAX;
-       td->slat_stat.min_val = ULONG_MAX;
-       td->bw_stat.min_val = ULONG_MAX;
-
-       run_str[td->thread_number - 1] = 'P';
-
-       if (td->use_aio && !td->aio_depth)
-               td->aio_depth = 1;
-
-       if (td->min_bs == -1U)
-               td->min_bs = td->bs;
-       if (td->max_bs == -1U)
-               td->max_bs = td->bs;
-       if (td_read(td))
-               td->verify = 0;
-
-       if (setup_rate(td))
-               return -1;
-
-       if (write_lat_log)
-               setup_log(&td->lat_log);
-       if (write_bw_log)
-               setup_log(&td->bw_log);
-
-       printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
-       return 0;
-}
-
-static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
-{
-       unsigned int i;
-
-       CPU_ZERO(&cpumask);
-
-       for (i = 0; i < sizeof(int) * 8; i++) {
-               if ((1 << i) & cpu)
-                       CPU_SET(i, &cpumask);
-       }
-}
-
-unsigned long get_mult(char c)
-{
-       switch (c) {
-               case 'k':
-               case 'K':
-                       return 1024;
-               case 'm':
-               case 'M':
-                       return 1024 * 1024;
-               case 'g':
-               case 'G':
-                       return 1024 * 1024 * 1024;
-               default:
-                       return 1;
-       }
-}
-
-/*
- * convert string after '=' into decimal value, noting any size suffix
- */
-static int str_cnv(char *p, unsigned long long *val)
-{
-       char *str;
-       int len;
-
-       str = strstr(p, "=");
-       if (!str)
-               return 1;
-
-       str++;
-       len = strlen(str);
-
-       *val = strtoul(str, NULL, 10);
-       if (*val == ULONG_MAX && errno == ERANGE)
-               return 1;
-
-       *val *= get_mult(str[len - 2]);
-       return 0;
-}
-
-static int check_strcnv(char *p, char *name, unsigned long long *val)
-{
-       if (!strstr(p, name))
-               return 1;
-
-       return str_cnv(p, val);
-}
-
-static int check_str(char *p, char *name, char *option)
-{
-       char *s = strstr(p, name);
-
-       if (!s)
-               return 1;
-
-       s += strlen(name);
-       if (strstr(s, option))
-               return 0;
-
-       return 1;
-}
-
-static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
-{
-       char str[128];
-       char s1, s2;
-
-       sprintf(str, "%s=%%lu%%c-%%lu%%c", name);
-       if (sscanf(p, str, s, &s1, e, &s2) == 4) {
-               *s *= get_mult(s1);
-               *e *= get_mult(s2);
-               return 0;
-       }
-
-       sprintf(str, "%s = %%lu%%c-%%lu%%c", name);
-       if (sscanf(p, str, s, &s1, e, &s2) == 4) {
-               *s *= get_mult(s1);
-               *e *= get_mult(s2);
-               return 0;
-       }
-
-       sprintf(str, "%s=%%lu-%%lu", name);
-       if (sscanf(p, str, s, e) == 2)
-               return 0;
-
-       sprintf(str, "%s = %%lu-%%lu", name);
-       if (sscanf(p, str, s, e) == 2)
-               return 0;
-
-       return 1;
-
-}
-
-static int check_int(char *p, char *name, unsigned int *val)
-{
-       char str[128];
-
-       sprintf(str, "%s=%%d", name);
-       if (sscanf(p, str, val) == 1)
-               return 0;
-
-       sprintf(str, "%s = %%d", name);
-       if (sscanf(p, str, val) == 1)
-               return 0;
-
-       return 1;
-}
-
-static int is_empty_or_comment(char *line)
-{
-       unsigned int i;
-
-       for (i = 0; i < strlen(line); i++) {
-               if (line[i] == ';')
-                       return 1;
-               if (!isspace(line[i]) && !iscntrl(line[i]))
-                       return 0;
-       }
-
-       return 1;
-}
-
-static int parse_jobs_ini(char *file)
-{
-       unsigned int prioclass, prio, cpu, global;
-       unsigned long long ull;
-       unsigned long ul1, ul2;
-       struct thread_data *td;
-       char *string, *name;
-       fpos_t off;
-       FILE *f;
-       char *p;
-
-       f = fopen(file, "r");
-       if (!f) {
-               perror("fopen");
-               return 1;
-       }
-
-       string = malloc(4096);
-       name = malloc(256);
-
-       while ((p = fgets(string, 4096, f)) != NULL) {
-               if (is_empty_or_comment(p))
-                       continue;
-               if (sscanf(p, "[%s]", name) != 1)
-                       continue;
-
-               global = !strncmp(name, "global", 6);
-
-               name[strlen(name) - 1] = '\0';
-
-               td = get_new_job(global);
-               if (!td)
-                       break;
-
-               prioclass = 2;
-               prio = 4;
-
-               fgetpos(f, &off);
-               while ((p = fgets(string, 4096, f)) != NULL) {
-                       if (is_empty_or_comment(p))
-                               continue;
-                       if (strstr(p, "["))
-                               break;
-                       if (!check_int(p, "rw", &td->ddir)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "prio", &prio)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "prioclass", &prioclass)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "direct", &td->odirect)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "rate", &td->rate)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "ratemin", &td->ratemin)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "ratecycle", &td->ratecycle)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "thinktime", &td->thinktime)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "cpumask", &cpu)) {
-                               fill_cpu_mask(td->cpumask, cpu);
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "fsync", &td->fsync_blocks)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "startdelay", &td->start_delay)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "timeout", &td->timeout)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "invalidate",&td->invalidate_cache)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "aio_depth", &td->aio_depth)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "sync", &td->sync_io)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "bwavgtime", &td->bw_avg_time)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "create_serialize", &td->create_serialize)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "create_fsync", &td->create_fsync)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "loops", &td->loops)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_int(p, "verify", &td->verify)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_range(p, "bsrange", &ul1, &ul2)) {
-                               if (ul1 & 511)
-                                       printf("bad min block size, must be a multiple of 512\n");
-                               else
-                                       td->min_bs = ul1;
-                               if (ul2 & 511)
-                                       printf("bad max block size, must be a multiple of 512\n");
-                               else
-                                       td->max_bs = ul2;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_strcnv(p, "bs", &ull)) {
-                               if (ull & 511)
-                                       printf("bad block size, must be a multiple of 512\n");
-                               else
-                                       td->bs = ull;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_strcnv(p, "size", &td->file_size)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_strcnv(p, "offset", &td->file_offset)) {
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_str(p, "mem", "malloc")) {
-                               td->mem_type = MEM_MALLOC;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!check_str(p, "mem", "shm")) {
-                               td->mem_type = MEM_SHM;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "sequential", 10)) {
-                               td->sequential = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "random", 6)) {
-                               td->sequential = 0;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "aio", 3)) {
-                               td->use_aio = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "create", 6)) {
-                               td->create_file = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "overwrite", 9)) {
-                               td->overwrite = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "exitall", 7)) {
-                               exitall_on_terminate = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       if (!strncmp(p, "stonewall", 9)) {
-                               td->stonewall = 1;
-                               fgetpos(f, &off);
-                               continue;
-                       }
-                       printf("Client%d: bad option %s\n",td->thread_number,p);
-               }
-               fsetpos(f, &off);
-
-               if (add_job(td, name, prioclass, prio))
-                       put_job(td);
-       }
-
-       free(string);
-       free(name);
-       fclose(f);
-       return 0;
-}
-
-static int parse_options(int argc, char *argv[])
-{
-       int i;
-
-       for (i = 1; i < argc; i++) {
-               char *parm = argv[i];
-
-               if (parm[0] != '-')
-                       break;
-
-               parm++;
-               switch (*parm) {
-                       case 's':
-                               parm++;
-                               def_thread.sequential = !!atoi(parm);
-                               break;
-                       case 'b':
-                               parm++;
-                               def_thread.bs = atoi(parm);
-                               def_thread.bs <<= 10;
-                               if (!def_thread.bs) {
-                                       printf("bad block size\n");
-                                       def_thread.bs = DEF_BS;
-                               }
-                               break;
-                       case 't':
-                               parm++;
-                               def_thread.timeout = atoi(parm);
-                               break;
-                       case 'r':
-                               parm++;
-                               repeatable = !!atoi(parm);
-                               break;
-                       case 'R':
-                               parm++;
-                               rate_quit = !!atoi(parm);
-                               break;
-                       case 'o':
-                               parm++;
-                               def_thread.odirect = !!atoi(parm);
-                               break;
-                       case 'f':
-                               if (i + 1 >= argc) {
-                                       printf("-f needs file as arg\n");
-                                       break;
-                               }
-                               ini_file = strdup(argv[i+1]);
-                               i++;
-                               break;
-                       case 'l':
-                               write_lat_log = 1;
-                               break;
-                       case 'w':
-                               write_bw_log = 1;
-                               break;
-                       default:
-                               printf("bad option %s\n", argv[i]);
-                               break;
-               }
-       }
-
-       return i;
+       printf("  cpu        : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
 }
 
 static void print_thread_status(struct thread_data *td, int nr_running,
@@ -2351,14 +1564,19 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                        continue;
 
                td_set_runstate(td, TD_REAPED);
-               waitpid(td->pid, NULL, 0);
+
+               if (td->use_thread) {
+                       long ret;
+
+                       if (pthread_join(td->thread, (void *) &ret))
+                               perror("thread_join");
+               } else
+                       waitpid(td->pid, NULL, 0);
+
                (*nr_running)--;
                (*m_rate) -= td->ratemin;
                (*t_rate) -= td->rate;
                check_str_update(td, *nr_running, *t_rate, *m_rate);
-
-               if (td->terminate)
-                       continue;
        }
 }
 
@@ -2425,7 +1643,7 @@ static void run_threads(char *argv[])
                        }
 
                        if (td->stonewall && (nr_started || nr_running))
-                               continue;
+                               break;
 
                        td_set_runstate(td, TD_CREATED);
                        check_str_update(td, nr_running, t_rate, m_rate);
@@ -2433,11 +1651,18 @@ static void run_threads(char *argv[])
                        todo--;
                        nr_started++;
 
-                       if (fork())
-                               sem_wait(&startup_sem);
-                       else {
-                               thread_main(shm_id, i, argv);
-                               exit(0);
+                       if (td->use_thread) {
+                               if (pthread_create(&td->thread, NULL, thread_main, td)) {
+                                       perror("thread_create");
+                                       nr_started--;
+                               }
+                       } else {
+                               if (fork())
+                                       sem_wait(&startup_sem);
+                               else {
+                                       fork_main(shm_id, i);
+                                       exit(0);
+                               }
                        }
                }
 
@@ -2462,11 +1687,8 @@ static void run_threads(char *argv[])
                for (i = 0; i < thread_number; i++) {
                        struct thread_data *td = &threads[i];
 
-                       if (td->runstate == TD_RUNNING)
-                               run_str[td->thread_number - 1] = '+';
-                       else if (td->runstate == TD_VERIFYING)
-                               run_str[td->thread_number - 1] = 'V';
-                       else
+                       if (td->runstate != TD_RUNNING &&
+                           td->runstate != TD_VERIFYING)
                                continue;
 
                        check_str_update(td, nr_running, t_rate, m_rate);
@@ -2484,131 +1706,92 @@ static void run_threads(char *argv[])
        }
 }
 
-int setup_thread_area(void)
+static void show_group_stats(struct group_run_stats *rs, int id)
 {
-       /*
-        * 1024 is too much on some machines, scale max_jobs if
-        * we get a failure that looks like too large a shm segment
-        */
-       do {
-               int s = max_jobs * sizeof(struct thread_data);
+       printf("\nRun status group %d:\n", id);
 
-               shm_id = shmget(0, s, IPC_CREAT | 0600);
-               if (shm_id != -1)
-                       break;
-               if (errno != EINVAL) {
-                       perror("shmget");
-                       break;
-               }
+       if (rs->max_run[DDIR_READ])
+               printf("   READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
+       if (rs->max_run[DDIR_WRITE])
+               printf("  WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
+}
 
-               max_jobs >>= 1;
-       } while (max_jobs);
+static void show_run_stats(void)
+{
+       struct group_run_stats *runstats, *rs;
+       struct thread_data *td;
+       int i;
 
-       if (shm_id == -1)
-               return 1;
+       runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
-       threads = shmat(shm_id, NULL, 0);
-       if (threads == (void *) -1) {
-               perror("shmat");
-               return 1;
+       for (i = 0; i < groupid + 1; i++) {
+               rs = &runstats[i];
+
+               memset(rs, 0, sizeof(*rs));
+               rs->min_bw[0] = rs->min_run[0] = ~0UL;
+               rs->min_bw[1] = rs->min_run[1] = ~0UL;
        }
 
-       atexit(free_shm);
-       return 0;
-}
+       for (i = 0; i < thread_number; i++) {
+               unsigned long bw = 0;
 
-int main(int argc, char *argv[])
-{
-       static unsigned long max_run[2], min_run[2];
-       static unsigned long max_bw[2], min_bw[2];
-       static unsigned long io_mb[2], agg[2];
-       int i;
+               td = &threads[i];
 
-       if (setup_thread_area())
-               return 1;
+               if (td->error)
+                       continue;
 
-       if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
-               perror("sched_getaffinity");
-               return 1;
-       }
+               rs = &runstats[td->groupid];
 
-       /*
-        * fill globals
-        */
-       def_thread.ddir = DDIR_READ;
-       def_thread.bs = DEF_BS;
-       def_thread.min_bs = -1;
-       def_thread.max_bs = -1;
-       def_thread.odirect = DEF_ODIRECT;
-       def_thread.ratecycle = DEF_RATE_CYCLE;
-       def_thread.sequential = DEF_SEQUENTIAL;
-       def_thread.timeout = DEF_TIMEOUT;
-       def_thread.create_file = DEF_CREATE;
-       def_thread.overwrite = DEF_OVERWRITE;
-       def_thread.invalidate_cache = DEF_INVALIDATE;
-       def_thread.sync_io = DEF_SYNCIO;
-       def_thread.mem_type = MEM_MALLOC;
-       def_thread.bw_avg_time = DEF_BWAVGTIME;
-       def_thread.create_serialize = DEF_CREATE_SER;
-       def_thread.create_fsync = DEF_CREATE_FSYNC;
-       def_thread.loops = DEF_LOOPS;
-       def_thread.verify = DEF_VERIFY;
-       def_thread.stonewall = DEF_STONEWALL;
-
-       i = parse_options(argc, argv);
-
-       if (!ini_file) {
-               printf("Need job file\n");
-               return 1;
-       }
+               if (td->runtime < rs->min_run[td->ddir])
+                       rs->min_run[td->ddir] = td->runtime;
+               if (td->runtime > rs->max_run[td->ddir])
+                       rs->max_run[td->ddir] = td->runtime;
 
-       if (parse_jobs_ini(ini_file))
-               return 1;
+               if (td->runtime)
+                       bw = td->io_bytes / td->runtime;
+               if (bw < rs->min_bw[td->ddir])
+                       rs->min_bw[td->ddir] = bw;
+               if (bw > rs->max_bw[td->ddir])
+                       rs->max_bw[td->ddir] = bw;
 
-       if (!thread_number) {
-               printf("Nothing to do\n");
-               return 1;
+               rs->io_mb[td->ddir] += td->io_bytes >> 20;
        }
 
-       run_threads(argv);
+       for (i = 0; i < groupid + 1; i++) {
+               rs = &runstats[i];
+
+               if (rs->max_run[0])
+                       rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
+               if (rs->max_run[1])
+                       rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
+       }
 
-       min_bw[0] = min_run[0] = ~0UL;
-       min_bw[1] = min_run[1] = ~0UL;
-       io_mb[0] = io_mb[1] = 0;
-       agg[0] = agg[1] = 0;
        for (i = 0; i < thread_number; i++) {
-               struct thread_data *td = &threads[i];
-               unsigned long bw = 0;
+               td = &threads[i];
+               rs = &runstats[td->groupid];
 
-               if (!td->error) {
-                       if (td->runtime < min_run[td->ddir])
-                               min_run[td->ddir] = td->runtime;
-                       if (td->runtime > max_run[td->ddir])
-                               max_run[td->ddir] = td->runtime;
+               if (!td->error)
+                       show_thread_status(td, rs);
+       }
 
-                       if (td->runtime)
-                               bw = td->io_bytes / td->runtime;
-                       if (bw < min_bw[td->ddir])
-                               min_bw[td->ddir] = bw;
-                       if (bw > max_bw[td->ddir])
-                               max_bw[td->ddir] = bw;
+       for (i = 0; i < groupid + 1; i++)
+               show_group_stats(&runstats[i], i);
+}
 
-                       io_mb[td->ddir] += td->io_bytes >> 20;
-               }
+int main(int argc, char *argv[])
+{
+       memset(run_str, 0, sizeof(run_str));
+
+       if (parse_options(argc, argv))
+               return 1;
 
-               show_thread_status(td);
+       if (!thread_number) {
+               printf("Nothing to do\n");
+               return 1;
        }
-       
-       if (max_run[0])
-               agg[0] = (io_mb[0] * 1024 * 1000) / max_run[0];
-       if (max_run[1])
-               agg[1] = (io_mb[1] * 1024 * 1000) / max_run[1];
-
-       printf("\nRun status:\n");
-       if (max_run[DDIR_READ])
-               printf("   READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[0], agg[0], min_bw[0], max_bw[0], min_run[0], max_run[0]);
-       if (max_run[DDIR_WRITE])
-               printf("  WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[1], agg[1], min_bw[1], max_bw[1], min_run[1], max_run[1]);
+
+       run_threads(argv);
+       show_run_stats();
 
        return 0;
 }