client: defer local trigger execute until after state is received
[fio.git] / backend.c
index 3016e9866b966a0c134db161002ac1c744d34104..3c2697dec3edb71740cf51d3eddd9cd7fe7b2828 100644 (file)
--- a/backend.c
+++ b/backend.c
 #include "err.h"
 #include "lib/tp.h"
 
-static pthread_t disk_util_thread;
-static struct fio_mutex *disk_thread_mutex;
+static pthread_t helper_thread;
+static pthread_mutex_t helper_lock;
+pthread_cond_t helper_cond;
+int helper_do_stat = 0;
+
 static struct fio_mutex *startup_mutex;
 static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
@@ -73,7 +76,7 @@ unsigned int stat_number = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
-volatile int disk_util_exit = 0;
+volatile int helper_exit = 0;
 
 #define PAGE_ALIGN(buf)        \
        (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
@@ -523,6 +526,11 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                break;
                }
 
+               if (verify_state_should_stop(td, io_u)) {
+                       put_io_u(td, io_u);
+                       break;
+               }
+
                if (td->o.verify_async)
                        io_u->end_io = verify_io_u_async;
                else
@@ -750,9 +758,14 @@ static uint64_t do_io(struct thread_data *td)
                    ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
 
                        if (!td->o.verify_pattern_bytes) {
-                               io_u->rand_seed = __rand(&td->__verify_state);
+                               io_u->rand_seed = __rand(&td->verify_state);
                                if (sizeof(int) != sizeof(long *))
-                                       io_u->rand_seed *= __rand(&td->__verify_state);
+                                       io_u->rand_seed *= __rand(&td->verify_state);
+                       }
+
+                       if (verify_state_should_stop(td, io_u)) {
+                               put_io_u(td, io_u);
+                               break;
                        }
 
                        if (td->o.verify_async)
@@ -974,6 +987,9 @@ static void cleanup_io_u(struct thread_data *td)
        io_u_rexit(&td->io_u_requeues);
        io_u_qexit(&td->io_u_freelist);
        io_u_qexit(&td->io_u_all);
+
+       if (td->last_write_comp)
+               sfree(td->last_write_comp);
 }
 
 static int init_io_u(struct thread_data *td)
@@ -1090,6 +1106,14 @@ static int init_io_u(struct thread_data *td)
                p += max_bs;
        }
 
+       if (td->o.verify != VERIFY_NONE) {
+               td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
+               if (!td->last_write_comp) {
+                       log_err("fio: failed to alloc write comp data\n");
+                       return 1;
+               }
+       }
+
        return 0;
 }
 
@@ -1342,7 +1366,6 @@ static void *thread_main(void *data)
        /* numa node setup */
        if (o->numa_cpumask_set || o->numa_memmask_set) {
                struct bitmask *mask;
-               int ret;
 
                if (numa_available() < 0) {
                        td_verror(td, errno, "Does not support NUMA API\n");
@@ -1482,18 +1505,21 @@ static void *thread_main(void *data)
 
                clear_state = 1;
 
+               fio_mutex_down(stat_mutex);
                if (td_read(td) && td->io_bytes[DDIR_READ]) {
-                       elapsed = utime_since_now(&td->start);
+                       elapsed = mtime_since_now(&td->start);
                        td->ts.runtime[DDIR_READ] += elapsed;
                }
                if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
-                       elapsed = utime_since_now(&td->start);
+                       elapsed = mtime_since_now(&td->start);
                        td->ts.runtime[DDIR_WRITE] += elapsed;
                }
                if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
-                       elapsed = utime_since_now(&td->start);
+                       elapsed = mtime_since_now(&td->start);
                        td->ts.runtime[DDIR_TRIM] += elapsed;
                }
+               fio_gettime(&td->start, NULL);
+               fio_mutex_up(stat_mutex);
 
                if (td->error || td->terminate)
                        break;
@@ -1509,21 +1535,33 @@ static void *thread_main(void *data)
 
                do_verify(td, verify_bytes);
 
-               td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
+               fio_mutex_down(stat_mutex);
+               td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
+               fio_gettime(&td->start, NULL);
+               fio_mutex_up(stat_mutex);
 
                if (td->error || td->terminate)
                        break;
        }
 
        update_rusage_stat(td);
-       td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
-       td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
-       td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
        td->ts.total_run_time = mtime_since_now(&td->epoch);
        td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
        td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
        td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
 
+       if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
+           (td->o.verify != VERIFY_NONE && td_write(td))) {
+               struct all_io_list *state;
+               size_t sz;
+
+               state = get_all_io_list(td->thread_number, &sz);
+               if (state) {
+                       __verify_save_state(state, "local");
+                       free(state);
+               }
+       }
+
        fio_unpin_memory(td);
 
        fio_writeout_logs(td);
@@ -1549,11 +1587,12 @@ err:
        cleanup_io_u(td);
        close_ioengine(td);
        cgroup_shutdown(td, &cgroup_mnt);
+       verify_free_state(td);
 
        if (o->cpumask_set) {
-               int ret = fio_cpuset_exit(&o->cpumask);
-
-               td_verror(td, ret, "fio_cpuset_exit");
+               ret = fio_cpuset_exit(&o->cpumask);
+               if (ret)
+                       td_verror(td, ret, "fio_cpuset_exit");
        }
 
        /*
@@ -1562,13 +1601,17 @@ err:
        if (o->write_iolog_file)
                write_iolog_close(td);
 
-       fio_mutex_remove(td->rusage_sem);
-       td->rusage_sem = NULL;
-
        fio_mutex_remove(td->mutex);
        td->mutex = NULL;
 
        td_set_runstate(td, TD_EXITED);
+
+       /*
+        * Do this last after setting our runstate to exited, so we
+        * know that the stat thread is signaled.
+        */
+       check_update_rusage(td);
+
        return (void *) (uintptr_t) td->error;
 }
 
@@ -1721,9 +1764,81 @@ reaped:
                fio_terminate_threads(TERMINATE_ALL);
 }
 
+static int __check_trigger_file(void)
+{
+       struct stat sb;
+
+       if (!trigger_file)
+               return 0;
+
+       if (stat(trigger_file, &sb))
+               return 0;
+
+       if (unlink(trigger_file) < 0)
+               log_err("fio: failed to unlink %s: %s\n", trigger_file,
+                                                       strerror(errno));
+
+       return 1;
+}
+
+static int trigger_timedout(void)
+{
+       if (trigger_timeout)
+               return time_since_genesis() >= trigger_timeout;
+
+       return 0;
+}
+
+void exec_trigger(const char *cmd)
+{
+       int ret;
+
+       if (!cmd)
+               return;
+
+       ret = system(cmd);
+       if (ret == -1)
+               log_err("fio: failed executing %s trigger\n", cmd);
+}
+
+void check_trigger_file(void)
+{
+       if (__check_trigger_file() || trigger_timedout()) {
+               if (nr_clients) {
+                       if (trigger_remote_cmd)
+                               fio_clients_send_trigger(trigger_remote_cmd);
+               } else {
+                       verify_save_state();
+                       fio_terminate_threads(TERMINATE_ALL);
+                       exec_trigger(trigger_cmd);
+               }
+       }
+}
+
+static int fio_verify_load_state(struct thread_data *td)
+{
+       int ret;
+
+       if (!td->o.verify_state)
+               return 0;
+
+       if (is_backend) {
+               void *data;
+
+               ret = fio_server_get_verify_state(td->o.name,
+                                       td->thread_number - 1, &data);
+               if (!ret)
+                       verify_convert_assign_state(td, data);
+       } else
+               ret = verify_load_state(td, "local");
+
+       return ret;
+}
+
 static void do_usleep(unsigned int usecs)
 {
        check_for_running_stats();
+       check_trigger_file();
        usleep(usecs);
 }
 
@@ -1777,12 +1892,16 @@ static void run_threads(void)
                if (!td->o.create_serialize)
                        continue;
 
+               if (fio_verify_load_state(td))
+                       goto reap;
+
                /*
                 * do file setup here so it happens sequentially,
                 * we don't want X number of threads getting their
                 * client data interspersed on disk
                 */
                if (setup_files(td)) {
+reap:
                        exit_value++;
                        if (td->error)
                                log_err("fio: pid=%d, err=%d/%s\n",
@@ -1970,57 +2089,70 @@ static void run_threads(void)
        update_io_ticks();
 }
 
-void wait_for_disk_thread_exit(void)
+static void wait_for_helper_thread_exit(void)
 {
-       fio_mutex_down(disk_thread_mutex);
+       void *ret;
+
+       helper_exit = 1;
+       pthread_cond_signal(&helper_cond);
+       pthread_join(helper_thread, &ret);
 }
 
 static void free_disk_util(void)
 {
-       disk_util_start_exit();
-       wait_for_disk_thread_exit();
        disk_util_prune_entries();
+
+       pthread_cond_destroy(&helper_cond);
 }
 
-static void *disk_thread_main(void *data)
+static void *helper_thread_main(void *data)
 {
        int ret = 0;
 
        fio_mutex_up(startup_mutex);
 
-       while (threads && !ret) {
-               usleep(DISK_UTIL_MSEC * 1000);
-               if (!threads)
-                       break;
+       while (!ret) {
+               uint64_t sec = DISK_UTIL_MSEC / 1000;
+               uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
+               struct timespec ts;
+               struct timeval tv;
+
+               gettimeofday(&tv, NULL);
+               ts.tv_sec = tv.tv_sec + sec;
+               ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
+               if (ts.tv_nsec > 1000000000ULL) {
+                       ts.tv_nsec -= 1000000000ULL;
+                       ts.tv_sec++;
+               }
+
+               pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
+
                ret = update_io_ticks();
 
+               if (helper_do_stat) {
+                       helper_do_stat = 0;
+                       __show_running_run_stats();
+               }
+
                if (!is_backend)
                        print_thread_status();
        }
 
-       fio_mutex_up(disk_thread_mutex);
        return NULL;
 }
 
-static int create_disk_util_thread(void)
+static int create_helper_thread(void)
 {
        int ret;
 
        setup_disk_util();
 
-       disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
-
-       ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
-       if (ret) {
-               fio_mutex_remove(disk_thread_mutex);
-               log_err("Can't create disk util thread: %s\n", strerror(ret));
-               return 1;
-       }
+       pthread_cond_init(&helper_cond, NULL);
+       pthread_mutex_init(&helper_lock, NULL);
 
-       ret = pthread_detach(disk_util_thread);
+       ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
        if (ret) {
-               fio_mutex_remove(disk_thread_mutex);
-               log_err("Can't detatch disk util thread: %s\n", strerror(ret));
+               log_err("Can't create helper thread: %s\n", strerror(ret));
                return 1;
        }
 
@@ -2060,18 +2192,18 @@ int fio_backend(void)
 
        set_genesis_time();
        stat_init();
-       create_disk_util_thread();
+       create_helper_thread();
 
        cgroup_list = smalloc(sizeof(*cgroup_list));
        INIT_FLIST_HEAD(cgroup_list);
 
        run_threads();
 
+       wait_for_helper_thread_exit();
+
        if (!fio_abort) {
                __show_run_stats();
                if (write_bw_log) {
-                       int i;
-
                        for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                                struct io_log *log = agg_io_log[i];
 
@@ -2081,8 +2213,13 @@ int fio_backend(void)
                }
        }
 
-       for_each_td(td, i)
+       for_each_td(td, i) {
                fio_options_free(td);
+               if (td->rusage_sem) {
+                       fio_mutex_remove(td->rusage_sem);
+                       td->rusage_sem = NULL;
+               }
+       }
 
        free_disk_util();
        cgroup_kill(cgroup_list);
@@ -2090,7 +2227,6 @@ int fio_backend(void)
        sfree(cgroup_mnt);
 
        fio_mutex_remove(startup_mutex);
-       fio_mutex_remove(disk_thread_mutex);
        stat_exit();
        return exit_value;
 }