Fix printf() being used instead of log_info()
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 4d260600056ba7ee99c1d06365195f962e29f4b3..3f6a586b79143adcac8726c4a33697e9c6a63837 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include <unistd.h>
 #include <fcntl.h>
 #include <string.h>
+#include <limits.h>
 #include <signal.h>
 #include <time.h>
 #include <locale.h>
 #include <assert.h>
+#include <time.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include "hash.h"
 #include "smalloc.h"
 #include "verify.h"
+#include "trim.h"
 #include "diskutil.h"
 #include "cgroup.h"
 #include "profile.h"
 #include "lib/rand.h"
+#include "memalign.h"
+#include "server.h"
 
 unsigned long page_mask;
 unsigned long page_size;
@@ -50,27 +55,79 @@ unsigned long page_size;
        (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
-int thread_number = 0;
-int nr_process = 0;
-int nr_thread = 0;
+unsigned int thread_number = 0;
+unsigned int nr_process = 0;
+unsigned int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
 
+/*
+ * Just expose an empty list, if the OS does not support disk util stats
+ */
+#ifndef FIO_HAVE_DISK_UTIL
+FLIST_HEAD(disk_list);
+#endif
+
 static struct fio_mutex *startup_mutex;
 static struct fio_mutex *writeout_mutex;
 static volatile int fio_abort;
 static int exit_value;
-static struct itimerval itimer;
 static pthread_t gtod_thread;
+static pthread_t disk_util_thread;
 static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
 
+unsigned long arch_flags = 0;
+
 struct io_log *agg_io_log[2];
 
-#define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
+static const char *fio_os_strings[os_nr] = {
+       "Invalid",
+       "Linux",
+       "AIX",
+       "FreeBSD",
+       "HP-UX",
+       "OSX",
+       "NetBSD",
+       "Solaris",
+       "Windows"
+};
+
+static const char *fio_arch_strings[arch_nr] = {
+       "Invalid",
+       "x86-64",
+       "x86",
+       "ppc",
+       "ia64",
+       "s390",
+       "alpha",
+       "sparc",
+       "sparc64",
+       "arm",
+       "sh",
+       "hppa",
+       "generic"
+};
+
+const char *fio_get_os_string(int nr)
+{
+       if (nr < os_nr)
+               return fio_os_strings[nr];
+
+       return NULL;
+}
+
+const char *fio_get_arch_string(int nr)
+{
+       if (nr < arch_nr)
+               return fio_arch_strings[nr];
+
+       return NULL;
+}
+
 void td_set_runstate(struct thread_data *td, int runstate)
 {
        if (td->runstate == runstate)
@@ -81,7 +138,7 @@ void td_set_runstate(struct thread_data *td, int runstate)
        td->runstate = runstate;
 }
 
-static void terminate_threads(int group_id)
+void fio_terminate_threads(int group_id)
 {
        struct thread_data *td;
        int i;
@@ -98,89 +155,101 @@ static void terminate_threads(int group_id)
                        /*
                         * if the thread is running, just let it exit
                         */
-                       if (td->runstate < TD_RUNNING)
-                               kill(td->pid, SIGQUIT);
+                       if (!td->pid)
+                               continue;
+                       else if (td->runstate < TD_RAMP)
+                               kill(td->pid, SIGTERM);
                        else {
                                struct ioengine_ops *ops = td->io_ops;
 
-                               if (ops && (ops->flags & FIO_SIGQUIT))
-                                       kill(td->pid, SIGQUIT);
+                               if (ops && (ops->flags & FIO_SIGTERM))
+                                       kill(td->pid, SIGTERM);
                        }
                }
        }
 }
 
-static void status_timer_arm(void)
+static void sig_int(int sig)
 {
-       itimer.it_value.tv_sec = 0;
-       itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
-       setitimer(ITIMER_REAL, &itimer, NULL);
+       if (threads) {
+               if (is_backend)
+                       fio_server_got_signal(sig);
+               else {
+                       log_info("\nfio: terminating on signal %d\n", sig);
+                       fflush(stdout);
+                       exit_value = 128;
+               }
+
+               fio_terminate_threads(TERMINATE_ALL);
+       }
 }
 
-static void sig_alrm(int fio_unused sig)
+static void *disk_thread_main(void *data)
 {
-       if (threads) {
+       fio_mutex_up(startup_mutex);
+
+       while (threads) {
+               usleep(DISK_UTIL_MSEC * 1000);
+               if (!threads)
+                       break;
                update_io_ticks();
-               print_thread_status();
-               status_timer_arm();
+
+               if (!is_backend)
+                       print_thread_status();
        }
-}
 
-/*
- * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
- */
-static void sig_quit(int sig)
-{
+       return NULL;
 }
 
-static void sig_int(int sig)
+static int create_disk_util_thread(void)
 {
-       if (threads) {
-               log_info("\nfio: terminating on signal %d\n", sig);
-               fflush(stdout);
-               terminate_threads(TERMINATE_ALL);
+       int ret;
+
+       ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
+       if (ret) {
+               log_err("Can't create disk util thread: %s\n", strerror(ret));
+               return 1;
        }
-}
 
-static void sig_ill(int fio_unused sig)
-{
-       if (!threads)
-               return;
+       ret = pthread_detach(disk_util_thread);
+       if (ret) {
+               log_err("Can't detatch disk util thread: %s\n", strerror(ret));
+               return 1;
+       }
 
-       log_err("fio: system does not support the sse4.2 instruction for crc32c-intel.\nUse crc32c instead.\n");
-       terminate_threads(TERMINATE_ALL);
+       dprint(FD_MUTEX, "wait on startup_mutex\n");
+       fio_mutex_down(startup_mutex);
+       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+       return 0;
 }
 
 static void set_sig_handlers(void)
 {
        struct sigaction act;
 
-       memset(&act, 0, sizeof(act));
-       act.sa_handler = sig_alrm;
-       act.sa_flags = SA_RESTART;
-       sigaction(SIGALRM, &act, NULL);
-
        memset(&act, 0, sizeof(act));
        act.sa_handler = sig_int;
        act.sa_flags = SA_RESTART;
        sigaction(SIGINT, &act, NULL);
 
        memset(&act, 0, sizeof(act));
-       act.sa_handler = sig_ill;
-       act.sa_flags = SA_RESETHAND;
-       sigaction(SIGILL, &act, NULL);
-
-       memset(&act, 0, sizeof(act));
-       act.sa_handler = sig_quit;
+       act.sa_handler = sig_int;
        act.sa_flags = SA_RESTART;
-       sigaction(SIGQUIT, &act, NULL);
+       sigaction(SIGTERM, &act, NULL);
+
+       if (is_backend) {
+               memset(&act, 0, sizeof(act));
+               act.sa_handler = sig_int;
+               act.sa_flags = SA_RESTART;
+               sigaction(SIGPIPE, &act, NULL);
+       }
 }
 
 /*
  * Check if we are above the minimum rate given.
  */
 static int __check_min_rate(struct thread_data *td, struct timeval *now,
-                           enum td_ddir ddir)
+                           enum fio_ddir ddir)
 {
        unsigned long long bytes = 0;
        unsigned long iops = 0;
@@ -190,6 +259,8 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
        unsigned int rate_iops = 0;
        unsigned int rate_iops_min = 0;
 
+       assert(ddir_rw(ddir));
+
        if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
                return 0;
 
@@ -199,7 +270,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
-       iops += td->io_blocks[ddir];
+       iops += td->this_io_blocks[ddir];
        bytes += td->this_io_bytes[ddir];
        ratemin += td->o.ratemin[ddir];
        rate_iops += td->o.rate_iops[ddir];
@@ -370,10 +441,15 @@ requeue:
        return 0;
 }
 
+static inline void __update_tv_cache(struct thread_data *td)
+{
+       fio_gettime(&td->tv_cache, NULL);
+}
+
 static inline void update_tv_cache(struct thread_data *td)
 {
        if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
-               fio_gettime(&td->tv_cache, NULL);
+               __update_tv_cache(td);
 }
 
 static int break_on_this_error(struct thread_data *td, int *retptr)
@@ -459,8 +535,11 @@ static void do_verify(struct thread_data *td)
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       td->terminate = 1;
-                       break;
+                       __update_tv_cache(td);
+                       if (runtime_exceeded(td, &td->tv_cache)) {
+                               td->terminate = 1;
+                               break;
+                       }
                }
 
                io_u = __get_io_u(td);
@@ -490,7 +569,6 @@ static void do_verify(struct thread_data *td)
                                clear_io_u(td, io_u);
                        } else if (io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
-                               struct fio_file *f = io_u->file;
 
                                /*
                                 * zero read, fail
@@ -505,8 +583,10 @@ static void do_verify(struct thread_data *td)
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
 
-                               td->ts.short_io_u[io_u->ddir]++;
+                               if (ddir_rw(io_u->ddir))
+                                       td->ts.short_io_u[io_u->ddir]++;
 
+                               f = io_u->file;
                                if (io_u->offset == f->real_file_size)
                                        goto sync_done;
 
@@ -537,13 +617,14 @@ sync_done:
 
                /*
                 * if we can queue more, do so. but check if there are
-                * completed io_u's first.
+                * completed io_u's first. Note that we can get BUSY even
+                * without IO queued, if the system is resource starved.
                 */
-               full = queue_full(td) || ret == FIO_Q_BUSY;
+               full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
                if (full || !td->o.iodepth_batch_complete) {
                        min_events = min(td->o.iodepth_batch_complete,
                                         td->cur_depth);
-                       if (full && !min_events)
+                       if (full && !min_events && td->o.iodepth_batch_complete != 0)
                                min_events = 1;
 
                        do {
@@ -590,6 +671,7 @@ static void do_io(struct thread_data *td)
                td_set_runstate(td, TD_RUNNING);
 
        while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
+               (!flist_empty(&td->trim_list)) ||
                ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
                struct timeval comp_time;
                unsigned long bytes_done[2] = { 0, 0 };
@@ -603,8 +685,11 @@ static void do_io(struct thread_data *td)
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       td->terminate = 1;
-                       break;
+                       __update_tv_cache(td);
+                       if (runtime_exceeded(td, &td->tv_cache)) {
+                               td->terminate = 1;
+                               break;
+                       }
                }
 
                io_u = get_io_u(td);
@@ -650,7 +735,8 @@ static void do_io(struct thread_data *td)
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
 
-                               td->ts.short_io_u[io_u->ddir]++;
+                               if (ddir_rw(io_u->ddir))
+                                       td->ts.short_io_u[io_u->ddir]++;
 
                                if (io_u->offset == f->real_file_size)
                                        goto sync_done;
@@ -692,13 +778,15 @@ sync_done:
                        break;
 
                /*
-                * See if we need to complete some commands
+                * See if we need to complete some commands. Note that we
+                * can get BUSY even without IO queued, if the system is
+                * resource starved.
                 */
-               full = queue_full(td) || ret == FIO_Q_BUSY;
+               full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
                if (full || !td->o.iodepth_batch_complete) {
                        min_evts = min(td->o.iodepth_batch_complete,
                                        td->cur_depth);
-                       if (full && !min_evts)
+                       if (full && !min_evts && td->o.iodepth_batch_complete != 0)
                                min_evts = 1;
 
                        if (__should_check_rate(td, 0) ||
@@ -721,7 +809,7 @@ sync_done:
                if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
                        if (check_min_rate(td, &comp_time, bytes_done)) {
                                if (exitall_on_terminate)
-                                       terminate_threads(td->groupid);
+                                       fio_terminate_threads(td->groupid);
                                td_verror(td, EIO, "check_min_rate");
                                break;
                        }
@@ -744,6 +832,9 @@ sync_done:
                }
        }
 
+       if (td->trim_entries)
+               printf("trim entries %ld\n", td->trim_entries);
+
        if (td->o.fill_device && td->error == ENOSPC) {
                td->error = 0;
                td->terminate = 1;
@@ -752,8 +843,11 @@ sync_done:
                struct fio_file *f;
 
                i = td->cur_depth;
-               if (i)
+               if (i) {
                        ret = io_u_queued_complete(td, i, NULL);
+                       if (td->o.fill_device && td->error == ENOSPC)
+                               td->error = 0;
+               }
 
                if (should_fsync(td) && td->o.end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
@@ -783,7 +877,7 @@ static void cleanup_io_u(struct thread_data *td)
                io_u = flist_entry(entry, struct io_u, list);
 
                flist_del(&io_u->list);
-               free(io_u);
+               fio_memfree(io_u, sizeof(*io_u));
        }
 
        free_io_mem(td);
@@ -816,7 +910,8 @@ static int init_io_u(struct thread_data *td)
        if (allocate_io_mem(td))
                return 1;
 
-       if (td->o.odirect || td->o.mem_align)
+       if (td->o.odirect || td->o.mem_align ||
+           (td->io_ops->flags & FIO_RAWIO))
                p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
        else
                p = td->orig_buffer;
@@ -829,8 +924,9 @@ static int init_io_u(struct thread_data *td)
                if (td->terminate)
                        return 1;
 
-               if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
-                       log_err("fio: posix_memalign=%s\n", strerror(errno));
+               ptr = fio_memalign(cl_align, sizeof(*io_u));
+               if (!ptr) {
+                       log_err("fio: unable to allocate aligned memory\n");
                        break;
                }
 
@@ -843,8 +939,15 @@ static int init_io_u(struct thread_data *td)
                        io_u->buf = p + max_bs * i;
                        dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 
-                       if (td_write(td) && !td->o.refill_buffers)
+                       if (td_write(td))
                                io_u_fill_buffer(td, io_u, max_bs);
+                       if (td_write(td) && td->o.verify_pattern_bytes) {
+                               /*
+                                * Fill the buffer with the pattern if we are
+                                * going to be doing writes.
+                                */
+                               fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
+                       }
                }
 
                io_u->index = i;
@@ -934,8 +1037,10 @@ static int keep_running(struct thread_data *td)
 
 static void reset_io_counters(struct thread_data *td)
 {
-       td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
+       td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
+       td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
+       td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
        td->zone_bytes = 0;
        td->rate_bytes[0] = td->rate_bytes[1] = 0;
        td->rate_blocks[0] = td->rate_blocks[1] = 0;
@@ -947,11 +1052,6 @@ static void reset_io_counters(struct thread_data *td)
         */
        if (td->o.time_based || td->o.loops)
                td->nr_done_files = 0;
-
-       /*
-        * Set the same seed to get repeatable runs
-        */
-       td_fill_rand_seeds(td);
 }
 
 void reset_all_stats(struct thread_data *td)
@@ -967,8 +1067,10 @@ void reset_all_stats(struct thread_data *td)
                td->io_issues[i] = 0;
                td->ts.total_io_u[i] = 0;
        }
-       
+
        fio_gettime(&tv, NULL);
+       td->ts.runtime[0] = 0;
+       td->ts.runtime[1] = 0;
        memcpy(&td->epoch, &tv, sizeof(tv));
        memcpy(&td->start, &tv, sizeof(tv));
 }
@@ -983,6 +1085,11 @@ static void clear_io_state(struct thread_data *td)
        close_files(td);
        for_each_file(td, f, i)
                fio_file_clear_done(f);
+
+       /*
+        * Set the same seed to get repeatable runs
+        */
+       td_fill_rand_seeds(td);
 }
 
 static int exec_string(const char *string)
@@ -1007,15 +1114,16 @@ static int exec_string(const char *string)
  */
 static void *thread_main(void *data)
 {
-       unsigned long long runtime[2], elapsed;
+       unsigned long long elapsed;
        struct thread_data *td = data;
        pthread_condattr_t attr;
        int clear_state;
 
-       if (!td->o.use_thread)
+       if (!td->o.use_thread) {
                setsid();
-
-       td->pid = getpid();
+               td->pid = getpid();
+       } else
+               td->pid = gettid();
 
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
@@ -1025,6 +1133,7 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->io_log_list);
        INIT_FLIST_HEAD(&td->io_hist_list);
        INIT_FLIST_HEAD(&td->verify_list);
+       INIT_FLIST_HEAD(&td->trim_list);
        pthread_mutex_init(&td->io_u_lock, NULL);
        td->io_hist_tree = RB_ROOT;
 
@@ -1045,12 +1154,32 @@ static void *thread_main(void *data)
         */
        fio_mutex_remove(td->mutex);
 
+       /*
+        * A new gid requires privilege, so we need to do this before setting
+        * the uid.
+        */
+       if (td->o.gid != -1U && setgid(td->o.gid)) {
+               td_verror(td, errno, "setgid");
+               goto err;
+       }
        if (td->o.uid != -1U && setuid(td->o.uid)) {
                td_verror(td, errno, "setuid");
                goto err;
        }
-       if (td->o.gid != -1U && setgid(td->o.gid)) {
-               td_verror(td, errno, "setgid");
+
+       /*
+        * If we have a gettimeofday() thread, make sure we exclude that
+        * thread from this job
+        */
+       if (td->o.gtod_cpu)
+               fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
+
+       /*
+        * Set affinity first, in case it has an impact on the memory
+        * allocations.
+        */
+       if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
+               td_verror(td, errno, "cpu_set_affinity");
                goto err;
        }
 
@@ -1067,23 +1196,6 @@ static void *thread_main(void *data)
        if (td->o.verify_async && verify_async_init(td))
                goto err;
 
-       if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
-               td_verror(td, errno, "cpu_set_affinity");
-               goto err;
-       }
-
-       /*
-        * If we have a gettimeofday() thread, make sure we exclude that
-        * thread from this job
-        */
-       if (td->o.gtod_cpu) {
-               fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
-               if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
-                       td_verror(td, errno, "cpu_set_affinity");
-                       goto err;
-               }
-       }
-
        if (td->ioprio_set) {
                if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
                        td_verror(td, errno, "ioprio_set");
@@ -1122,20 +1234,17 @@ static void *thread_main(void *data)
        }
 
        fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ts.ru_start);
+       getrusage(RUSAGE_SELF, &td->ru_start);
 
-       runtime[0] = runtime[1] = 0;
        clear_state = 0;
        while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->ts.stat_sample_time[0], &td->start,
-                               sizeof(td->start));
-               memcpy(&td->ts.stat_sample_time[1], &td->start,
-                               sizeof(td->start));
+               memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
                memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
                if (td->o.ratemin[0] || td->o.ratemin[1])
-                       memcpy(&td->lastrate, &td->ts.stat_sample_time,
+                       memcpy(&td->lastrate, &td->bw_sample_time,
                                                        sizeof(td->lastrate));
 
                if (clear_state)
@@ -1149,11 +1258,11 @@ static void *thread_main(void *data)
 
                if (td_read(td) && td->io_bytes[DDIR_READ]) {
                        elapsed = utime_since_now(&td->start);
-                       runtime[DDIR_READ] += elapsed;
+                       td->ts.runtime[DDIR_READ] += elapsed;
                }
                if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
                        elapsed = utime_since_now(&td->start);
-                       runtime[DDIR_WRITE] += elapsed;
+                       td->ts.runtime[DDIR_WRITE] += elapsed;
                }
 
                if (td->error || td->terminate)
@@ -1170,47 +1279,62 @@ static void *thread_main(void *data)
 
                do_verify(td);
 
-               runtime[DDIR_READ] += utime_since_now(&td->start);
+               td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
 
                if (td->error || td->terminate)
                        break;
        }
 
        update_rusage_stat(td);
-       td->ts.runtime[0] = (runtime[0] + 999) / 1000;
-       td->ts.runtime[1] = (runtime[1] + 999) / 1000;
+       td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
+       td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
        td->ts.total_run_time = mtime_since_now(&td->epoch);
        td->ts.io_bytes[0] = td->io_bytes[0];
        td->ts.io_bytes[1] = td->io_bytes[1];
 
        fio_mutex_down(writeout_mutex);
-       if (td->ts.bw_log) {
+       if (td->bw_log) {
                if (td->o.bw_log_file) {
-                       finish_log_named(td, td->ts.bw_log,
+                       finish_log_named(td, td->bw_log,
                                                td->o.bw_log_file, "bw");
                } else
-                       finish_log(td, td->ts.bw_log, "bw");
+                       finish_log(td, td->bw_log, "bw");
+       }
+       if (td->lat_log) {
+               if (td->o.lat_log_file) {
+                       finish_log_named(td, td->lat_log,
+                                               td->o.lat_log_file, "lat");
+               } else
+                       finish_log(td, td->lat_log, "lat");
        }
-       if (td->ts.slat_log) {
+       if (td->slat_log) {
                if (td->o.lat_log_file) {
-                       finish_log_named(td, td->ts.slat_log,
+                       finish_log_named(td, td->slat_log,
                                                td->o.lat_log_file, "slat");
                } else
-                       finish_log(td, td->ts.slat_log, "slat");
+                       finish_log(td, td->slat_log, "slat");
        }
-       if (td->ts.clat_log) {
+       if (td->clat_log) {
                if (td->o.lat_log_file) {
-                       finish_log_named(td, td->ts.clat_log,
+                       finish_log_named(td, td->clat_log,
                                                td->o.lat_log_file, "clat");
                } else
-                       finish_log(td, td->ts.clat_log, "clat");
+                       finish_log(td, td->clat_log, "clat");
        }
+       if (td->iops_log) {
+               if (td->o.iops_log_file) {
+                       finish_log_named(td, td->iops_log,
+                                               td->o.iops_log_file, "iops");
+               } else
+                       finish_log(td, td->iops_log, "iops");
+       }
+
        fio_mutex_up(writeout_mutex);
        if (td->o.exec_postrun)
                exec_string(td->o.exec_postrun);
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid);
+               fio_terminate_threads(td->groupid);
 
 err:
        if (td->error)
@@ -1237,7 +1361,6 @@ err:
        if (td->o.write_iolog_file)
                write_iolog_close(td);
 
-       options_mem_free(td);
        td_set_runstate(td, TD_EXITED);
        return (void *) (unsigned long) td->error;
 }
@@ -1251,6 +1374,7 @@ static int fork_main(int shmid, int offset)
        struct thread_data *td;
        void *data, *ret;
 
+#ifndef __hpux
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
                int __err = errno;
@@ -1258,6 +1382,12 @@ static int fork_main(int shmid, int offset)
                perror("shmat");
                return __err;
        }
+#else
+       /*
+        * HP-UX inherits shm mappings?
+        */
+       data = threads;
+#endif
 
        td = data + offset * sizeof(struct thread_data);
        ret = thread_main(td);
@@ -1268,10 +1398,12 @@ static int fork_main(int shmid, int offset)
 /*
  * Run over the job map and reap the threads that have exited, if any.
  */
-static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
+static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
+                        unsigned int *m_rate)
 {
        struct thread_data *td;
-       int i, cputhreads, realthreads, pending, status, ret;
+       unsigned int cputhreads, realthreads, pending;
+       int i, status, ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
@@ -1323,7 +1455,7 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                        if (WIFSIGNALED(status)) {
                                int sig = WTERMSIG(status);
 
-                               if (sig != SIGQUIT)
+                               if (sig != SIGTERM)
                                        log_err("fio: pid=%d, got signal=%d\n",
                                                        (int) td->pid, sig);
                                td_set_runstate(td, TD_REAPED);
@@ -1357,7 +1489,7 @@ reaped:
        }
 
        if (*nr_running == cputhreads && !pending && realthreads)
-               terminate_threads(TERMINATE_ALL);
+               fio_terminate_threads(TERMINATE_ALL);
 }
 
 static void *gtod_thread_main(void *data)
@@ -1380,9 +1512,13 @@ static void *gtod_thread_main(void *data)
 
 static int fio_start_gtod_thread(void)
 {
+       pthread_attr_t attr;
        int ret;
 
-       ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
+       pthread_attr_init(&attr);
+       pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
+       ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
+       pthread_attr_destroy(&attr);
        if (ret) {
                log_err("Can't create gtod thread: %s\n", strerror(ret));
                return 1;
@@ -1407,7 +1543,7 @@ static void run_threads(void)
 {
        struct thread_data *td;
        unsigned long spent;
-       int i, todo, nr_running, m_rate, t_rate, nr_started;
+       unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
 
        if (fio_pin_memory())
                return;
@@ -1415,6 +1551,8 @@ static void run_threads(void)
        if (fio_gtod_offload && fio_start_gtod_thread())
                return;
 
+       set_sig_handlers();
+
        if (!terse_output) {
                log_info("Starting ");
                if (nr_thread)
@@ -1422,7 +1560,7 @@ static void run_threads(void)
                                                nr_thread > 1 ? "s" : "");
                if (nr_process) {
                        if (nr_thread)
-                               printf(" and ");
+                               log_info(" and ");
                        log_info("%d process%s", nr_process,
                                                nr_process > 1 ? "es" : "");
                }
@@ -1430,8 +1568,6 @@ static void run_threads(void)
                fflush(stdout);
        }
 
-       set_sig_handlers();
-
        todo = thread_number;
        nr_running = 0;
        nr_started = 0;
@@ -1440,10 +1576,8 @@ static void run_threads(void)
        for_each_td(td, i) {
                print_status_init(td->thread_number - 1);
 
-               if (!td->o.create_serialize) {
-                       init_disk_util(td);
+               if (!td->o.create_serialize)
                        continue;
-               }
 
                /*
                 * do file setup here so it happens sequentially,
@@ -1459,26 +1593,24 @@ static void run_threads(void)
                        todo--;
                } else {
                        struct fio_file *f;
-                       unsigned int i;
+                       unsigned int j;
 
                        /*
                         * for sharing to work, each job must always open
                         * its own files. so close them, if we opened them
                         * for creation
                         */
-                       for_each_file(td, f, i) {
+                       for_each_file(td, f, j) {
                                if (fio_file_open(f))
                                        td_io_close_file(td, f);
                        }
                }
-
-               init_disk_util(td);
        }
 
        set_genesis_time();
 
        while (todo) {
-               struct thread_data *map[MAX_JOBS];
+               struct thread_data *map[REAL_MAX_JOBS];
                struct timeval this_start;
                int this_jobs = 0, left;
 
@@ -1511,6 +1643,8 @@ static void run_threads(void)
                                break;
                        }
 
+                       init_disk_util(td);
+
                        /*
                         * Set state to created. Thread will transition
                         * to TD_INITIALIZED when it's done setting up.
@@ -1549,7 +1683,7 @@ static void run_threads(void)
                        dprint(FD_MUTEX, "wait on startup_mutex\n");
                        if (fio_mutex_down_timeout(startup_mutex, 10)) {
                                log_err("fio: job startup hung? exiting.\n");
-                               terminate_threads(TERMINATE_ALL);
+                               fio_terminate_threads(TERMINATE_ALL);
                                fio_abort = 1;
                                nr_started--;
                                break;
@@ -1617,50 +1751,40 @@ static void run_threads(void)
 
                reap_threads(&nr_running, &t_rate, &m_rate);
 
-               if (todo)
-                       usleep(100000);
+               if (todo) {
+                       if (is_backend)
+                               fio_server_idle_loop();
+                       else
+                               usleep(100000);
+               }
        }
 
        while (nr_running) {
                reap_threads(&nr_running, &t_rate, &m_rate);
-               usleep(10000);
+
+               if (is_backend)
+                       fio_server_idle_loop();
+               else
+                       usleep(10000);
        }
 
        update_io_ticks();
        fio_unpin_memory();
 }
 
-int main(int argc, char *argv[])
+int exec_run(void)
 {
-       long ps;
-
-       sinit();
-       init_rand(&__fio_rand_state);
-
-       /*
-        * We need locale for number printing, if it isn't set then just
-        * go with the US format.
-        */
-       if (!getenv("LC_NUMERIC"))
-               setlocale(LC_NUMERIC, "en_US");
+       struct thread_data *td;
+       int i;
 
-       ps = sysconf(_SC_PAGESIZE);
-       if (ps < 0) {
-               log_err("Failed to get page size\n");
-               return 1;
+       if (nr_clients)
+               return fio_handle_clients();
+       if (exec_profile) {
+               if (load_profile(exec_profile))
+                       return 1;
+               free(exec_profile);
+               exec_profile = NULL;
        }
-
-       page_size = ps;
-       page_mask = ps - 1;
-
-       fio_keywords_init();
-
-       if (parse_options(argc, argv))
-               return 1;
-
-       if (exec_profile && load_profile(exec_profile))
-               return 1;
-
        if (!thread_number)
                return 0;
 
@@ -1670,11 +1794,14 @@ int main(int argc, char *argv[])
        }
 
        startup_mutex = fio_mutex_init(0);
+       if (startup_mutex == NULL)
+               return 1;
        writeout_mutex = fio_mutex_init(1);
+       if (writeout_mutex == NULL)
+               return 1;
 
        set_genesis_time();
-
-       status_timer_arm();
+       create_disk_util_thread();
 
        cgroup_list = smalloc(sizeof(*cgroup_list));
        INIT_FLIST_HEAD(cgroup_list);
@@ -1690,6 +1817,9 @@ int main(int argc, char *argv[])
                }
        }
 
+       for_each_td(td, i)
+               fio_options_free(td);
+
        cgroup_kill(cgroup_list);
        sfree(cgroup_list);
        sfree(cgroup_mnt);
@@ -1698,3 +1828,80 @@ int main(int argc, char *argv[])
        fio_mutex_remove(writeout_mutex);
        return exit_value;
 }
+
+void reset_fio_state(void)
+{
+       groupid = 0;
+       thread_number = 0;
+       nr_process = 0;
+       nr_thread = 0;
+       done_secs = 0;
+}
+
+static int endian_check(void)
+{
+       union {
+               uint8_t c[8];
+               uint64_t v;
+       } u;
+       int le = 0, be = 0;
+
+       u.v = 0x12;
+       if (u.c[7] == 0x12)
+               be = 1;
+       else if (u.c[0] == 0x12)
+               le = 1;
+
+#if defined(FIO_LITTLE_ENDIAN)
+       if (be)
+               return 1;
+#elif defined(FIO_BIG_ENDIAN)
+       if (le)
+               return 1;
+#else
+       return 1;
+#endif
+
+       if (!le && !be)
+               return 1;
+
+       return 0;
+}
+
+int main(int argc, char *argv[], char *envp[])
+{
+       long ps;
+
+       if (endian_check()) {
+               log_err("fio: endianness settings appear wrong.\n");
+               log_err("fio: please report this to fio@vger.kernel.org\n");
+               return 1;
+       }
+
+       arch_init(envp);
+
+       sinit();
+
+       /*
+        * We need locale for number printing, if it isn't set then just
+        * go with the US format.
+        */
+       if (!getenv("LC_NUMERIC"))
+               setlocale(LC_NUMERIC, "en_US");
+
+       ps = sysconf(_SC_PAGESIZE);
+       if (ps < 0) {
+               log_err("Failed to get page size\n");
+               return 1;
+       }
+
+       page_size = ps;
+       page_mask = ps - 1;
+
+       fio_keywords_init();
+
+       if (parse_options(argc, argv))
+               return 1;
+
+       return exec_run();
+}