Merge branch 'master' into gfio
[fio.git] / backend.c
index 507faa9992e7325d4c80de368e43f2a779893739..49d6bc7cb2cc7da2f9842cca78150a732ad07c6a 100644 (file)
--- a/backend.c
+++ b/backend.c
@@ -50,6 +50,8 @@
 #include "lib/rand.h"
 #include "memalign.h"
 #include "server.h"
+#include "lib/getrusage.h"
+#include "idletime.h"
 
 static pthread_t disk_util_thread;
 static struct fio_mutex *disk_thread_mutex;
@@ -59,14 +61,14 @@ static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
 static int exit_value;
 static volatile int fio_abort;
+static unsigned int nr_process = 0;
+static unsigned int nr_thread = 0;
 
 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
 
 int groupid = 0;
 unsigned int thread_number = 0;
 unsigned int stat_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;
@@ -216,7 +218,7 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 }
 
 static int check_min_rate(struct thread_data *td, struct timeval *now,
-                         unsigned long *bytes_done)
+                         uint64_t *bytes_done)
 {
        int ret = 0;
 
@@ -395,10 +397,10 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
  */
 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 {
+       uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
        struct fio_file *f;
        struct io_u *io_u;
        int ret, min_events;
-       uint64_t io_bytes;
        unsigned int i;
 
        dprint(FD_VERIFY, "starting loop\n");
@@ -422,7 +424,6 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
        td_set_runstate(td, TD_VERIFYING);
 
        io_u = NULL;
-       io_bytes = 0;
        while (!td->terminate) {
                enum fio_ddir ddir;
                int ret2, full;
@@ -455,11 +456,37 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                break;
                        }
                } else {
-                       io_u = get_io_u(td);
-                       if (!io_u)
+                       if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
                                break;
 
-                       if (io_u->buflen + io_bytes > verify_bytes)
+                       while ((io_u = get_io_u(td)) != NULL) {
+                               /*
+                                * We are only interested in the places where
+                                * we wrote or trimmed IOs. Turn those into
+                                * reads for verification purposes.
+                                */
+                               if (io_u->ddir == DDIR_READ) {
+                                       /*
+                                        * Pretend we issued it for rwmix
+                                        * accounting
+                                        */
+                                       td->io_issues[DDIR_READ]++;
+                                       put_io_u(td, io_u);
+                                       continue;
+                               } else if (io_u->ddir == DDIR_TRIM) {
+                                       io_u->ddir = DDIR_READ;
+                                       io_u->flags |= IO_U_F_TRIMMED;
+                                       break;
+                               } else if (io_u->ddir == DDIR_WRITE) {
+                                       io_u->ddir = DDIR_READ;
+                                       break;
+                               } else {
+                                       put_io_u(td, io_u);
+                                       continue;
+                               }
+                       }
+
+                       if (!io_u)
                                break;
                }
 
@@ -491,7 +518,6 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
-                               io_bytes += bytes;
 
                                if (ddir_rw(io_u->ddir))
                                        td->ts.short_io_u[io_u->ddir]++;
@@ -503,11 +529,10 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               ret = io_u_sync_complete(td, io_u, NULL);
+                               ret = io_u_sync_complete(td, io_u, bytes_done);
                                if (ret < 0)
                                        break;
                        }
-                       io_bytes += io_u->xfer_buflen;
                        continue;
                case FIO_Q_QUEUED:
                        break;
@@ -542,18 +567,15 @@ sync_done:
                                min_events = 1;
 
                        do {
-                               unsigned long bytes = 0;
-
                                /*
                                 * Reap required number of io units, if any,
                                 * and do the verification on them through
                                 * the callback handler
                                 */
-                               if (io_u_queued_complete(td, min_events, &bytes) < 0) {
+                               if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
                                        ret = -1;
                                        break;
                                }
-                               io_bytes += bytes;
                        } while (full && (td->cur_depth > td->o.iodepth_low));
                }
                if (ret < 0)
@@ -592,9 +614,12 @@ static int io_bytes_exceeded(struct thread_data *td)
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
+ *
+ * Returns number of bytes written and trimmed.
  */
-static void do_io(struct thread_data *td)
+static uint64_t do_io(struct thread_data *td)
 {
+       uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
        unsigned int i;
        int ret = 0;
 
@@ -607,7 +632,6 @@ static void do_io(struct thread_data *td)
                (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
                td->o.time_based) {
                struct timeval comp_time;
-               unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
                int min_evts = 0;
                struct io_u *io_u;
                int ret2, full;
@@ -811,6 +835,8 @@ sync_done:
         */
        if (!ddir_rw_sum(td->this_io_bytes))
                td->done = 1;
+
+       return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -1021,10 +1047,12 @@ static void *thread_main(void *data)
 {
        unsigned long long elapsed;
        struct thread_data *td = data;
+       struct thread_options *o = &td->o;
        pthread_condattr_t attr;
        int clear_state;
+       int ret;
 
-       if (!td->o.use_thread) {
+       if (!o->use_thread) {
                setsid();
                td->pid = getpid();
        } else
@@ -1034,6 +1062,9 @@ static void *thread_main(void *data)
 
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
+       if (is_backend)
+               fio_server_send_start(td);
+
        INIT_FLIST_HEAD(&td->io_u_freelist);
        INIT_FLIST_HEAD(&td->io_u_busylist);
        INIT_FLIST_HEAD(&td->io_u_requeues);
@@ -1061,16 +1092,17 @@ static void *thread_main(void *data)
         * eating a file descriptor
         */
        fio_mutex_remove(td->mutex);
+       td->mutex = NULL;
 
        /*
         * A new gid requires privilege, so we need to do this before setting
         * the uid.
         */
-       if (td->o.gid != -1U && setgid(td->o.gid)) {
+       if (o->gid != -1U && setgid(o->gid)) {
                td_verror(td, errno, "setgid");
                goto err;
        }
-       if (td->o.uid != -1U && setuid(td->o.uid)) {
+       if (o->uid != -1U && setuid(o->uid)) {
                td_verror(td, errno, "setuid");
                goto err;
        }
@@ -1079,16 +1111,19 @@ static void *thread_main(void *data)
         * 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 (o->gtod_cpu)
+               fio_cpu_clear(&o->cpumask, 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;
+       if (o->cpumask_set) {
+               ret = fio_setaffinity(td->pid, o->cpumask);
+               if (ret == -1) {
+                       td_verror(td, errno, "cpu_set_affinity");
+                       goto err;
+               }
        }
 
 #ifdef CONFIG_LIBNUMA
@@ -1144,11 +1179,12 @@ static void *thread_main(void *data)
        if (init_io_u(td))
                goto err;
 
-       if (td->o.verify_async && verify_async_init(td))
+       if (o->verify_async && verify_async_init(td))
                goto err;
 
-       if (td->ioprio_set) {
-               if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
+       if (o->ioprio) {
+               ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
+               if (ret == -1) {
                        td_verror(td, errno, "ioprio_set");
                        goto err;
                }
@@ -1158,15 +1194,15 @@ static void *thread_main(void *data)
                goto err;
 
        errno = 0;
-       if (nice(td->o.nice) == -1 && errno != 0) {
+       if (nice(o->nice) == -1 && errno != 0) {
                td_verror(td, errno, "nice");
                goto err;
        }
 
-       if (td->o.ioscheduler && switch_ioscheduler(td))
+       if (o->ioscheduler && switch_ioscheduler(td))
                goto err;
 
-       if (!td->o.create_serialize && setup_files(td))
+       if (!o->create_serialize && setup_files(td))
                goto err;
 
        if (td_io_init(td))
@@ -1175,22 +1211,21 @@ static void *thread_main(void *data)
        if (init_random_map(td))
                goto err;
 
-       if (td->o.exec_prerun) {
-               if (exec_string(td->o.exec_prerun))
-                       goto err;
-       }
+       if (o->exec_prerun && exec_string(o->exec_prerun))
+               goto err;
 
-       if (td->o.pre_read) {
+       if (o->pre_read) {
                if (pre_read_files(td) < 0)
                        goto err;
        }
 
-       fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ru_start);
+       fio_verify_init(td);
 
+       fio_gettime(&td->epoch, NULL);
+       fio_getrusage(&td->ru_start);
        clear_state = 0;
        while (keep_running(td)) {
-               uint64_t write_bytes;
+               uint64_t verify_bytes;
 
                fio_gettime(&td->start, NULL);
                memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
@@ -1212,9 +1247,7 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
-               write_bytes = td->io_bytes[DDIR_WRITE];
-               do_io(td);
-               write_bytes = td->io_bytes[DDIR_WRITE] - write_bytes;
+               verify_bytes = do_io(td);
 
                clear_state = 1;
 
@@ -1243,7 +1276,7 @@ static void *thread_main(void *data)
 
                fio_gettime(&td->start, NULL);
 
-               do_verify(td, write_bytes);
+               do_verify(td, verify_bytes);
 
                td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
 
@@ -1260,6 +1293,8 @@ static void *thread_main(void *data)
        td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
        td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
 
+       fio_unpin_memory(td);
+
        fio_mutex_down(writeout_mutex);
        if (td->bw_log) {
                if (td->o.bw_log_file) {
@@ -1317,8 +1352,8 @@ err:
        close_ioengine(td);
        cgroup_shutdown(td, &cgroup_mnt);
 
-       if (td->o.cpumask_set) {
-               int ret = fio_cpuset_exit(&td->o.cpumask);
+       if (o->cpumask_set) {
+               int ret = fio_cpuset_exit(&o->cpumask);
 
                td_verror(td, ret, "fio_cpuset_exit");
        }
@@ -1472,14 +1507,21 @@ static void run_threads(void)
        unsigned long spent;
        unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
 
-       if (fio_pin_memory())
-               return;
-
        if (fio_gtod_offload && fio_start_gtod_thread())
                return;
+       
+       fio_idle_prof_init();
 
        set_sig_handlers();
 
+       nr_thread = nr_process = 0;
+       for_each_td(td, i) {
+               if (td->o.use_thread)
+                       nr_thread++;
+               else
+                       nr_process++;
+       }
+
        if (output_format == FIO_OUTPUT_NORMAL) {
                log_info("Starting ");
                if (nr_thread)
@@ -1534,6 +1576,9 @@ static void run_threads(void)
                }
        }
 
+       /* start idle threads before io threads start to run */
+       fio_idle_prof_start();
+
        set_genesis_time();
 
        while (todo) {
@@ -1679,25 +1724,18 @@ static void run_threads(void)
 
                reap_threads(&nr_running, &t_rate, &m_rate);
 
-               if (todo) {
-                       if (is_backend)
-                               fio_server_idle_loop();
-                       else
-                               usleep(100000);
-               }
+               if (todo)
+                       usleep(100000);
        }
 
        while (nr_running) {
                reap_threads(&nr_running, &t_rate, &m_rate);
-
-               if (is_backend)
-                       fio_server_idle_loop();
-               else
-                       usleep(10000);
+               usleep(10000);
        }
 
+       fio_idle_prof_stop();
+
        update_io_ticks();
-       fio_unpin_memory();
 }
 
 void wait_for_disk_thread_exit(void)
@@ -1775,9 +1813,9 @@ int fio_backend(void)
                return 0;
 
        if (write_bw_log) {
-               setup_log(&agg_io_log[DDIR_READ], 0);
-               setup_log(&agg_io_log[DDIR_WRITE], 0);
-               setup_log(&agg_io_log[DDIR_TRIM], 0);
+               setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
+               setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
+               setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
        }
 
        startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);