Call path below SIGALRM isn't safe
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 6ac8ef057b57ed9f98f35f8dfef5ace8a1af196e..91bf6b613cac30403ff9c7ec96865a2b56051103 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -28,6 +28,7 @@
 #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"
 
 unsigned long page_mask;
 unsigned long page_size;
-#define ALIGN(buf)     \
+
+#define PAGE_ALIGN(buf)        \
        (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
@@ -57,8 +63,10 @@ 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 timer_t ival_timer;
 static pthread_t gtod_thread;
+static struct flist_head *cgroup_list;
+static char *cgroup_mnt;
 
 struct io_log *agg_io_log[2];
 
@@ -106,17 +114,21 @@ static void terminate_threads(int group_id)
 
 static void status_timer_arm(void)
 {
-       itimer.it_value.tv_sec = 0;
-       itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
-       setitimer(ITIMER_REAL, &itimer, NULL);
+       struct itimerspec value;
+
+       value.it_value.tv_sec = 0;
+       value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
+       value.it_interval.tv_sec = 0;
+       value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
+
+       timer_settime(ival_timer, 0, &value, NULL);
 }
 
-static void sig_alrm(int fio_unused sig)
+static void ival_fn(union sigval sig)
 {
        if (threads) {
                update_io_ticks();
                print_thread_status();
-               status_timer_arm();
        }
 }
 
@@ -130,70 +142,62 @@ static void sig_quit(int sig)
 static void sig_int(int sig)
 {
        if (threads) {
-               printf("\nfio: terminating on signal %d\n", sig);
+               log_info("\nfio: terminating on signal %d\n", sig);
                fflush(stdout);
                terminate_threads(TERMINATE_ALL);
        }
 }
 
-static void sig_ill(int fio_unused sig)
+static void posix_timer_teardown(void)
 {
-       if (!threads)
-               return;
+       timer_delete(ival_timer);
+}
+
+static void posix_timer_setup(void)
+{
+       struct sigevent evt;
+
+       memset(&evt, 0, sizeof(evt));
+       evt.sigev_notify = SIGEV_THREAD;
+       evt.sigev_notify_function = ival_fn;
 
-       log_err("fio: illegal instruction. your cpu does not support "
-               "the sse4.2 instruction for crc32c\n");
-       terminate_threads(TERMINATE_ALL);
-       exit(4);
+       if (timer_create(CLOCK_MONOTONIC, &evt, &ival_timer) < 0)
+               perror("timer_create");
 }
 
 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_RESTART;
-       sigaction(SIGILL, &act, NULL);
-
        memset(&act, 0, sizeof(act));
        act.sa_handler = sig_quit;
        act.sa_flags = SA_RESTART;
        sigaction(SIGQUIT, &act, NULL);
 }
 
-static inline int should_check_rate(struct thread_data *td)
-{
-       struct thread_options *o = &td->o;
-
-       /*
-        * If some rate setting was given, we need to check it
-        */
-       if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
-               return 1;
-
-       return 0;
-}
-
 /*
  * Check if we are above the minimum rate given.
  */
-static int check_min_rate(struct thread_data *td, struct timeval *now)
+static int __check_min_rate(struct thread_data *td, struct timeval *now,
+                           enum td_ddir ddir)
 {
        unsigned long long bytes = 0;
        unsigned long iops = 0;
        unsigned long spent;
        unsigned long rate;
+       unsigned int ratemin = 0;
+       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;
 
        /*
         * allow a 2 second settle period in the beginning
@@ -201,38 +205,35 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
-       if (td_read(td)) {
-               iops += td->io_blocks[DDIR_READ];
-               bytes += td->this_io_bytes[DDIR_READ];
-       }
-       if (td_write(td)) {
-               iops += td->io_blocks[DDIR_WRITE];
-               bytes += td->this_io_bytes[DDIR_WRITE];
-       }
+       iops += td->io_blocks[ddir];
+       bytes += td->this_io_bytes[ddir];
+       ratemin += td->o.ratemin[ddir];
+       rate_iops += td->o.rate_iops[ddir];
+       rate_iops_min += td->o.rate_iops_min[ddir];
 
        /*
         * if rate blocks is set, sample is running
         */
-       if (td->rate_bytes || td->rate_blocks) {
-               spent = mtime_since(&td->lastrate, now);
+       if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
+               spent = mtime_since(&td->lastrate[ddir], now);
                if (spent < td->o.ratecycle)
                        return 0;
 
-               if (td->o.rate) {
+               if (td->o.rate[ddir]) {
                        /*
                         * check bandwidth specified rate
                         */
-                       if (bytes < td->rate_bytes) {
+                       if (bytes < td->rate_bytes[ddir]) {
                                log_err("%s: min rate %u not met\n", td->o.name,
-                                                               td->o.ratemin);
+                                                               ratemin);
                                return 1;
                        } else {
-                               rate = (bytes - td->rate_bytes) / spent;
-                               if (rate < td->o.ratemin ||
-                                   bytes < td->rate_bytes) {
+                               rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               if (rate < ratemin ||
+                                   bytes < td->rate_bytes[ddir]) {
                                        log_err("%s: min rate %u not met, got"
-                                               " %luKiB/sec\n", td->o.name,
-                                                       td->o.ratemin, rate);
+                                               " %luKB/sec\n", td->o.name,
+                                                       ratemin, rate);
                                        return 1;
                                }
                        }
@@ -240,29 +241,41 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                        /*
                         * checks iops specified rate
                         */
-                       if (iops < td->o.rate_iops) {
+                       if (iops < rate_iops) {
                                log_err("%s: min iops rate %u not met\n",
-                                               td->o.name, td->o.rate_iops);
+                                               td->o.name, rate_iops);
                                return 1;
                        } else {
-                               rate = (iops - td->rate_blocks) / spent;
-                               if (rate < td->o.rate_iops_min ||
-                                   iops < td->rate_blocks) {
+                               rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+                               if (rate < rate_iops_min ||
+                                   iops < td->rate_blocks[ddir]) {
                                        log_err("%s: min iops rate %u not met,"
                                                " got %lu\n", td->o.name,
-                                                       td->o.rate_iops_min,
-                                                       rate);
+                                                       rate_iops_min, rate);
                                }
                        }
                }
        }
 
-       td->rate_bytes = bytes;
-       td->rate_blocks = iops;
-       memcpy(&td->lastrate, now, sizeof(*now));
+       td->rate_bytes[ddir] = bytes;
+       td->rate_blocks[ddir] = iops;
+       memcpy(&td->lastrate[ddir], now, sizeof(*now));
        return 0;
 }
 
+static int check_min_rate(struct thread_data *td, struct timeval *now,
+                         unsigned long *bytes_done)
+{
+       int ret = 0;
+
+       if (bytes_done[0])
+               ret |= __check_min_rate(td, now, 0);
+       if (bytes_done[1])
+               ret |= __check_min_rate(td, now, 1);
+
+       return ret;
+}
+
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
        if (!td->o.timeout)
@@ -286,7 +299,7 @@ static void cleanup_pending_aio(struct thread_data *td)
        /*
         * get immediately available events, if any
         */
-       r = io_u_queued_complete(td, 0);
+       r = io_u_queued_complete(td, 0, NULL);
        if (r < 0)
                return;
 
@@ -314,7 +327,7 @@ static void cleanup_pending_aio(struct thread_data *td)
        }
 
        if (td->cur_depth)
-               r = io_u_queued_complete(td, td->cur_depth);
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -344,7 +357,7 @@ requeue:
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
-               if (io_u_queued_complete(td, 1) < 0)
+               if (io_u_queued_complete(td, 1, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
@@ -352,7 +365,7 @@ requeue:
                        return 1;
                }
 
-               if (io_u_sync_complete(td, io_u) < 0)
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_BUSY) {
                if (td_io_commit(td))
@@ -369,6 +382,51 @@ static inline void update_tv_cache(struct thread_data *td)
                fio_gettime(&td->tv_cache, NULL);
 }
 
+static int break_on_this_error(struct thread_data *td, int *retptr)
+{
+       int ret = *retptr;
+
+       if (ret < 0 || td->error) {
+               int err;
+
+               if (!td->o.continue_on_error)
+                       return 1;
+
+               if (ret < 0)
+                       err = -ret;
+               else
+                       err = td->error;
+
+               if (td_non_fatal_error(err)) {
+                       /*
+                        * Continue with the I/Os in case of
+                        * a non fatal error.
+                        */
+                       update_error_count(td, err);
+                       td_clear_error(td);
+                       *retptr = 0;
+                       return 0;
+               } else if (td->o.fill_device && err == ENOSPC) {
+                       /*
+                        * We expect to hit this error if
+                        * fill_device option is set.
+                        */
+                       td_clear_error(td);
+                       td->terminate = 1;
+                       return 1;
+               } else {
+                       /*
+                        * Stop the I/O in case of a fatal
+                        * error.
+                        */
+                       update_error_count(td, err);
+                       return 1;
+               }
+       }
+
+       return 0;
+}
+
 /*
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -380,6 +438,8 @@ static void do_verify(struct thread_data *td)
        int ret, min_events;
        unsigned int i;
 
+       dprint(FD_VERIFY, "starting loop\n");
+
        /*
         * sync io first and invalidate cache, to make sure we really
         * read from disk.
@@ -402,18 +462,17 @@ static void do_verify(struct thread_data *td)
        while (!td->terminate) {
                int ret2, full;
 
-               io_u = __get_io_u(td);
-               if (!io_u)
-                       break;
-
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
                }
 
+               io_u = __get_io_u(td);
+               if (!io_u)
+                       break;
+
                if (get_next_verify(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
@@ -424,14 +483,18 @@ static void do_verify(struct thread_data *td)
                        break;
                }
 
-               io_u->end_io = verify_io_u;
+               if (td->o.verify_async)
+                       io_u->end_io = verify_io_u_async;
+               else
+                       io_u->end_io = verify_io_u;
 
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
-                       if (io_u->error)
+                       if (io_u->error) {
                                ret = -io_u->error;
-                       else if (io_u->resid) {
+                               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;
 
@@ -448,7 +511,8 @@ 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]++;
 
                                if (io_u->offset == f->real_file_size)
                                        goto sync_done;
@@ -456,7 +520,7 @@ static void do_verify(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               ret = io_u_sync_complete(td, io_u);
+                               ret = io_u_sync_complete(td, io_u, NULL);
                                if (ret < 0)
                                        break;
                        }
@@ -475,7 +539,7 @@ sync_done:
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
@@ -484,7 +548,8 @@ sync_done:
                 */
                full = queue_full(td) || ret == FIO_Q_BUSY;
                if (full || !td->o.iodepth_batch_complete) {
-                       min_events = td->o.iodepth_batch_complete;
+                       min_events = min(td->o.iodepth_batch_complete,
+                                        td->cur_depth);
                        if (full && !min_events)
                                min_events = 1;
 
@@ -494,7 +559,7 @@ sync_done:
                                 * and do the verification on them through
                                 * the callback handler
                                 */
-                               if (io_u_queued_complete(td, min_events) < 0) {
+                               if (io_u_queued_complete(td, min_events, NULL) < 0) {
                                        ret = -1;
                                        break;
                                }
@@ -508,11 +573,13 @@ sync_done:
                min_events = td->cur_depth;
 
                if (min_events)
-                       ret = io_u_queued_complete(td, min_events);
+                       ret = io_u_queued_complete(td, min_events, NULL);
        } else
                cleanup_pending_aio(td);
 
        td_set_runstate(td, TD_RUNNING);
+
+       dprint(FD_VERIFY, "exiting loop\n");
 }
 
 /*
@@ -521,7 +588,6 @@ sync_done:
  */
 static void do_io(struct thread_data *td)
 {
-       unsigned long usec;
        unsigned int i;
        int ret = 0;
 
@@ -530,9 +596,11 @@ static void do_io(struct thread_data *td)
        else
                td_set_runstate(td, TD_RUNNING);
 
-       while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
+       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;
-               long bytes_done = 0;
+               unsigned long bytes_done[2] = { 0, 0 };
                int min_evts = 0;
                struct io_u *io_u;
                int ret2, full;
@@ -540,24 +608,27 @@ static void do_io(struct thread_data *td)
                if (td->terminate)
                        break;
 
-               io_u = get_io_u(td);
-               if (!io_u)
-                       break;
-
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
                }
 
+               io_u = get_io_u(td);
+               if (!io_u)
+                       break;
+
                /*
                 * Add verification end_io handler, if asked to verify
                 * a previously written file.
                 */
-               if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
-                       io_u->end_io = verify_io_u;
+               if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
+                   !td_rw(td)) {
+                       if (td->o.verify_async)
+                               io_u->end_io = verify_io_u_async;
+                       else
+                               io_u->end_io = verify_io_u;
                        td_set_runstate(td, TD_VERIFYING);
                } else if (in_ramp_time(td))
                        td_set_runstate(td, TD_RAMP);
@@ -567,9 +638,10 @@ static void do_io(struct thread_data *td)
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
-                       if (io_u->error)
+                       if (io_u->error) {
                                ret = -io_u->error;
-                       else if (io_u->resid) {
+                               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;
 
@@ -586,7 +658,8 @@ 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;
@@ -594,12 +667,13 @@ static void do_io(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               if (should_check_rate(td))
+                               if (__should_check_rate(td, 0) ||
+                                   __should_check_rate(td, 1))
                                        fio_gettime(&comp_time, NULL);
 
-                               bytes_done = io_u_sync_complete(td, io_u);
-                               if (bytes_done < 0)
-                                       ret = bytes_done;
+                               ret = io_u_sync_complete(td, io_u, bytes_done);
+                               if (ret < 0)
+                                       break;
                        }
                        break;
                case FIO_Q_QUEUED:
@@ -623,7 +697,7 @@ sync_done:
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
@@ -631,38 +705,30 @@ sync_done:
                 */
                full = queue_full(td) || ret == FIO_Q_BUSY;
                if (full || !td->o.iodepth_batch_complete) {
-                       min_evts = td->o.iodepth_batch_complete;
+                       min_evts = min(td->o.iodepth_batch_complete,
+                                       td->cur_depth);
                        if (full && !min_evts)
                                min_evts = 1;
 
-                       if (should_check_rate(td))
+                       if (__should_check_rate(td, 0) ||
+                           __should_check_rate(td, 1))
                                fio_gettime(&comp_time, NULL);
 
                        do {
-                               ret = io_u_queued_complete(td, min_evts);
-                               if (ret <= 0)
+                               ret = io_u_queued_complete(td, min_evts, bytes_done);
+                               if (ret < 0)
                                        break;
 
-                               bytes_done += ret;
                        } while (full && (td->cur_depth > td->o.iodepth_low));
                }
 
                if (ret < 0)
                        break;
-               if (!bytes_done)
+               if (!(bytes_done[0] + bytes_done[1]))
                        continue;
 
-               /*
-                * the rate is batched for now, it should work for batches
-                * of completions except the very first one which may look
-                * a little bursty
-                */
-               if (!in_ramp_time(td) && should_check_rate(td)) {
-                       usec = utime_since(&td->tv_cache, &comp_time);
-
-                       rate_throttle(td, usec, bytes_done);
-
-                       if (check_min_rate(td, &comp_time)) {
+               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);
                                td_verror(td, EIO, "check_min_rate");
@@ -687,6 +753,9 @@ 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;
@@ -696,7 +765,7 @@ sync_done:
 
                i = td->cur_depth;
                if (i)
-                       ret = io_u_queued_complete(td, i);
+                       ret = io_u_queued_complete(td, i, NULL);
 
                if (should_fsync(td) && td->o.end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
@@ -759,8 +828,9 @@ static int init_io_u(struct thread_data *td)
        if (allocate_io_mem(td))
                return 1;
 
-       if (td->o.odirect)
-               p = ALIGN(td->orig_buffer);
+       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;
 
@@ -780,12 +850,21 @@ static int init_io_u(struct thread_data *td)
                io_u = ptr;
                memset(io_u, 0, sizeof(*io_u));
                INIT_FLIST_HEAD(&io_u->list);
+               dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
                if (!(td->io_ops->flags & FIO_NOIO)) {
                        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)
                                io_u_fill_buffer(td, io_u, max_bs);
+                       else 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);
+                       }
                }
 
                io_u->index = i;
@@ -878,8 +957,8 @@ static void reset_io_counters(struct thread_data *td)
        td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
-       td->rate_bytes = 0;
-       td->rate_blocks = 0;
+       td->rate_bytes[0] = td->rate_bytes[1] = 0;
+       td->rate_blocks[0] = td->rate_blocks[1] = 0;
 
        td->last_was_sync = 0;
 
@@ -950,6 +1029,7 @@ static void *thread_main(void *data)
 {
        unsigned long long runtime[2], elapsed;
        struct thread_data *td = data;
+       pthread_condattr_t attr;
        int clear_state;
 
        if (!td->o.use_thread)
@@ -964,8 +1044,15 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->io_u_requeues);
        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;
 
+       pthread_condattr_init(&attr);
+       pthread_cond_init(&td->verify_cond, &attr);
+       pthread_cond_init(&td->free_cond, &attr);
+
        td_set_runstate(td, TD_INITIALIZED);
        dprint(FD_MUTEX, "up startup_mutex\n");
        fio_mutex_up(startup_mutex);
@@ -979,6 +1066,15 @@ static void *thread_main(void *data)
         */
        fio_mutex_remove(td->mutex);
 
+       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");
+               goto err;
+       }
+
        /*
         * May alter parameters that init_io_u() will use, so we need to
         * do this first.
@@ -989,7 +1085,10 @@ static void *thread_main(void *data)
        if (init_io_u(td))
                goto err;
 
-       if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
+       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;
        }
@@ -1000,7 +1099,7 @@ static void *thread_main(void *data)
         */
        if (td->o.gtod_cpu) {
                fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
-               if (fio_setaffinity(td) == -1) {
+               if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
                        td_verror(td, errno, "cpu_set_affinity");
                        goto err;
                }
@@ -1013,6 +1112,9 @@ static void *thread_main(void *data)
                }
        }
 
+       if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
+               goto err;
+
        if (nice(td->o.nice) == -1) {
                td_verror(td, errno, "nice");
                goto err;
@@ -1047,10 +1149,13 @@ static void *thread_main(void *data)
        clear_state = 0;
        while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+               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->tv_cache, &td->start, sizeof(td->start));
 
-               if (td->o.ratemin)
+               if (td->o.ratemin[0] || td->o.ratemin[1])
                        memcpy(&td->lastrate, &td->ts.stat_sample_time,
                                                        sizeof(td->lastrate));
 
@@ -1107,6 +1212,13 @@ static void *thread_main(void *data)
                } else
                        finish_log(td, td->ts.bw_log, "bw");
        }
+       if (td->ts.lat_log) {
+               if (td->o.lat_log_file) {
+                       finish_log_named(td, td->ts.lat_log,
+                                               td->o.lat_log_file, "lat");
+               } else
+                       finish_log(td, td->ts.lat_log, "lat");
+       }
        if (td->ts.slat_log) {
                if (td->o.lat_log_file) {
                        finish_log_named(td, td->ts.slat_log,
@@ -1130,11 +1242,16 @@ static void *thread_main(void *data)
 
 err:
        if (td->error)
-               printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
+               log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
                                                        td->verror);
+
+       if (td->o.verify_async)
+               verify_async_exit(td);
+
        close_and_free_files(td);
        close_ioengine(td);
        cleanup_io_u(td);
+       cgroup_shutdown(td, &cgroup_mnt);
 
        if (td->o.cpumask_set) {
                int ret = fio_cpuset_exit(&td->o.cpumask);
@@ -1256,8 +1373,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                continue;
 reaped:
                (*nr_running)--;
-               (*m_rate) -= td->o.ratemin;
-               (*t_rate) -= td->o.rate;
+               (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
+               (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
                if (!td->pid)
                        pending--;
 
@@ -1291,9 +1408,13 @@ 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;
@@ -1327,17 +1448,17 @@ static void run_threads(void)
                return;
 
        if (!terse_output) {
-               printf("Starting ");
+               log_info("Starting ");
                if (nr_thread)
-                       printf("%d thread%s", nr_thread,
+                       log_info("%d thread%s", nr_thread,
                                                nr_thread > 1 ? "s" : "");
                if (nr_process) {
                        if (nr_thread)
                                printf(" and ");
-                       printf("%d process%s", nr_process,
+                       log_info("%d process%s", nr_process,
                                                nr_process > 1 ? "es" : "");
                }
-               printf("\n");
+               log_info("\n");
                fflush(stdout);
        }
 
@@ -1377,8 +1498,10 @@ static void run_threads(void)
                         * its own files. so close them, if we opened them
                         * for creation
                         */
-                       for_each_file(td, f, i)
-                               td_io_close_file(td, f);
+                       for_each_file(td, f, i) {
+                               if (fio_file_open(f))
+                                       td_io_close_file(td, f);
+                       }
                }
 
                init_disk_util(td);
@@ -1456,7 +1579,13 @@ static void run_threads(void)
                                        *fio_debug_jobp = pid;
                        }
                        dprint(FD_MUTEX, "wait on startup_mutex\n");
-                       fio_mutex_down(startup_mutex);
+                       if (fio_mutex_down_timeout(startup_mutex, 10)) {
+                               log_err("fio: job startup hung? exiting.\n");
+                               terminate_threads(TERMINATE_ALL);
+                               fio_abort = 1;
+                               nr_started--;
+                               break;
+                       }
                        dprint(FD_MUTEX, "done waiting on startup_mutex\n");
                }
 
@@ -1512,8 +1641,8 @@ static void run_threads(void)
                                td_set_runstate(td, TD_RUNNING);
                        nr_running++;
                        nr_started--;
-                       m_rate += td->o.ratemin;
-                       t_rate += td->o.rate;
+                       m_rate += td->o.ratemin[0] + td->o.ratemin[1];
+                       t_rate += td->o.rate[0] + td->o.rate[1];
                        todo--;
                        fio_mutex_up(td->mutex);
                }
@@ -1538,6 +1667,7 @@ int main(int argc, char *argv[])
        long ps;
 
        sinit();
+       init_rand(&__fio_rand_state);
 
        /*
         * We need locale for number printing, if it isn't set then just
@@ -1546,12 +1676,6 @@ int main(int argc, char *argv[])
        if (!getenv("LC_NUMERIC"))
                setlocale(LC_NUMERIC, "en_US");
 
-       if (parse_options(argc, argv))
-               return 1;
-
-       if (!thread_number)
-               return 0;
-
        ps = sysconf(_SC_PAGESIZE);
        if (ps < 0) {
                log_err("Failed to get page size\n");
@@ -1561,6 +1685,17 @@ int main(int argc, char *argv[])
        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;
+
        if (write_bw_log) {
                setup_log(&agg_io_log[DDIR_READ]);
                setup_log(&agg_io_log[DDIR_WRITE]);
@@ -1571,8 +1706,12 @@ int main(int argc, char *argv[])
 
        set_genesis_time();
 
+       posix_timer_setup();
        status_timer_arm();
 
+       cgroup_list = smalloc(sizeof(*cgroup_list));
+       INIT_FLIST_HEAD(cgroup_list);
+
        run_threads();
 
        if (!fio_abort) {
@@ -1584,6 +1723,11 @@ int main(int argc, char *argv[])
                }
        }
 
+       cgroup_kill(cgroup_list);
+       sfree(cgroup_list);
+       sfree(cgroup_mnt);
+
+       posix_timer_teardown();
        fio_mutex_remove(startup_mutex);
        fio_mutex_remove(writeout_mutex);
        return exit_value;