Make switch_ioscheduler() return 0 if FIO_HAVE_IOSCHED_SWITCH is undefined
[fio.git] / backend.c
index 7f57c65ae64678d10f351db12abaa3cfe287d098..2f290d2d7d1358429c1a66440f79de648bfd88c9 100644 (file)
--- a/backend.c
+++ b/backend.c
 #include "workqueue.h"
 #include "lib/mountcheck.h"
 #include "rate-submit.h"
-
-static pthread_t helper_thread;
-static pthread_mutex_t helper_lock;
-pthread_cond_t helper_cond;
-int helper_do_stat = 0;
+#include "helper_thread.h"
 
 static struct fio_mutex *startup_mutex;
 static struct flist_head *cgroup_list;
@@ -79,7 +75,6 @@ unsigned int stat_number = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
-volatile int helper_exit = 0;
 
 #define PAGE_ALIGN(buf)        \
        (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
@@ -446,6 +441,12 @@ static int wait_for_completions(struct thread_data *td, struct timeval *time)
        int min_evts = 0;
        int ret;
 
+       if (td->flags & TD_F_REGROW_LOGS) {
+               ret = io_u_quiesce(td);
+               regrow_logs(td);
+               return ret;
+       }
+
        /*
         * if the queue is full, we MUST reap at least 1 event
         */
@@ -523,6 +524,9 @@ sync_done:
                                break;
                }
 
+               if (td->flags & TD_F_REGROW_LOGS)
+                       regrow_logs(td);
+
                /*
                 * when doing I/O (not when verifying),
                 * check for any errors that are to be ignored
@@ -599,6 +603,15 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
        if (td->error)
                return;
 
+       /*
+        * verify_state needs to be reset before verification
+        * proceeds so that expected random seeds match actual
+        * random seeds in headers. The main loop will reset
+        * all random number generators if randrepeat is set.
+        */
+       if (!td->o.rand_repeatable)
+               td_fill_verify_state_seed(td);
+
        td_set_runstate(td, TD_VERIFYING);
 
        io_u = NULL;
@@ -881,7 +894,14 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                if (flow_threshold_exceeded(td))
                        continue;
 
-               if (!td->o.time_based && bytes_issued >= total_bytes)
+               /*
+                * Break if we exceeded the bytes. The exception is time
+                * based runs, but we still need to break out of the loop
+                * for those to run verification, if enabled.
+                */
+               if (bytes_issued >= total_bytes &&
+                   (!td->o.time_based ||
+                    (td->o.time_based && td->o.verify != VERIFY_NONE)))
                        break;
 
                io_u = get_io_u(td);
@@ -1063,6 +1083,41 @@ reap:
                bytes_done[i] = td->bytes_done[i] - bytes_done[i];
 }
 
+static void free_file_completion_logging(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       for_each_file(td, f, i) {
+               if (!f->last_write_comp)
+                       break;
+               sfree(f->last_write_comp);
+       }
+}
+
+static int init_file_completion_logging(struct thread_data *td,
+                                       unsigned int depth)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
+               return 0;
+
+       for_each_file(td, f, i) {
+               f->last_write_comp = scalloc(depth, sizeof(uint64_t));
+               if (!f->last_write_comp)
+                       goto cleanup;
+       }
+
+       return 0;
+
+cleanup:
+       free_file_completion_logging(td);
+       log_err("fio: failed to alloc write comp data\n");
+       return 1;
+}
+
 static void cleanup_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
@@ -1081,8 +1136,7 @@ static void cleanup_io_u(struct thread_data *td)
        io_u_qexit(&td->io_u_freelist);
        io_u_qexit(&td->io_u_all);
 
-       if (td->last_write_comp)
-               sfree(td->last_write_comp);
+       free_file_completion_logging(td);
 }
 
 static int init_io_u(struct thread_data *td)
@@ -1199,19 +1253,15 @@ 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;
-               }
-       }
+       if (init_file_completion_logging(td, max_units))
+               return 1;
 
        return 0;
 }
 
 static int switch_ioscheduler(struct thread_data *td)
 {
+#ifdef FIO_HAVE_IOSCHED_SWITCH
        char tmp[256], tmp2[128];
        FILE *f;
        int ret;
@@ -1270,6 +1320,9 @@ static int switch_ioscheduler(struct thread_data *td)
 
        fclose(f);
        return 0;
+#else
+       return 0;
+#endif
 }
 
 static bool keep_running(struct thread_data *td)
@@ -1388,7 +1441,6 @@ static void *thread_main(void *data)
        struct thread_data *td = fd->td;
        struct thread_options *o = &td->o;
        struct sk_out *sk_out = fd->sk_out;
-       pthread_condattr_t attr;
        int clear_state;
        int ret;
 
@@ -1413,12 +1465,18 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->verify_list);
        INIT_FLIST_HEAD(&td->trim_list);
        INIT_FLIST_HEAD(&td->next_rand_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);
+       ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
+       if (ret) {
+               td_verror(td, ret, "mutex_cond_init_pshared");
+               goto err;
+       }
+       ret = cond_init_pshared(&td->verify_cond);
+       if (ret) {
+               td_verror(td, ret, "mutex_cond_pshared");
+               goto err;
+       }
 
        td_set_runstate(td, TD_INITIALIZED);
        dprint(FD_MUTEX, "up startup_mutex\n");
@@ -1440,6 +1498,14 @@ static void *thread_main(void *data)
                goto err;
        }
 
+       /*
+        * Do this early, we don't want the compress threads to be limited
+        * to the same CPUs as the IO workers. So do this before we set
+        * any potential CPU affinity
+        */
+       if (iolog_compress_init(td, sk_out))
+               goto err;
+
        /*
         * If we have a gettimeofday() thread, make sure we exclude that
         * thread from this job
@@ -1574,9 +1640,6 @@ static void *thread_main(void *data)
                        goto err;
        }
 
-       if (iolog_compress_init(td, sk_out))
-               goto err;
-
        fio_verify_init(td);
 
        if (rate_submit_init(td, sk_out))
@@ -1674,6 +1737,8 @@ static void *thread_main(void *data)
                        break;
        }
 
+       td_set_runstate(td, TD_FINISHING);
+
        update_rusage_stat(td);
        td->ts.total_run_time = mtime_since_now(&td->epoch);
        td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
@@ -1686,7 +1751,7 @@ static void *thread_main(void *data)
 
        fio_unpin_memory(td);
 
-       fio_writeout_logs(td);
+       td_writeout_logs(td, true);
 
        iolog_compress_exit(td);
        rate_submit_exit(td);
@@ -1747,43 +1812,11 @@ err:
        return (void *) (uintptr_t) td->error;
 }
 
-
-/*
- * We cannot pass the td data into a forked process, so attach the td and
- * pass it to the thread worker.
- */
-static int fork_main(struct sk_out *sk_out, int shmid, int offset)
-{
-       struct fork_data *fd;
-       void *data, *ret;
-
-#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
-       data = shmat(shmid, NULL, 0);
-       if (data == (void *) -1) {
-               int __err = errno;
-
-               perror("shmat");
-               return __err;
-       }
-#else
-       /*
-        * HP-UX inherits shm mappings?
-        */
-       data = threads;
-#endif
-
-       fd = calloc(1, sizeof(*fd));
-       fd->td = data + offset * sizeof(struct thread_data);
-       fd->sk_out = sk_out;
-       ret = thread_main(fd);
-       shmdt(data);
-       return (int) (uintptr_t) ret;
-}
-
 static void dump_td_info(struct thread_data *td)
 {
-       log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
-               "be stuck. Doing forceful exit of this job.\n", td->o.name,
+       log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
+               "appears to be stuck. Doing forceful exit of this job.\n",
+                       td->o.name, td->runstate,
                        (unsigned long) time_since_now(&td->terminate_time));
 }
 
@@ -1869,6 +1902,7 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                 * move on.
                 */
                if (td->terminate &&
+                   td->runstate < TD_FSYNCING &&
                    time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
                        dump_td_info(td);
                        td_set_runstate(td, TD_REAPED);
@@ -1957,12 +1991,11 @@ static int fio_verify_load_state(struct thread_data *td)
 
        if (is_backend) {
                void *data;
-               int ver;
 
                ret = fio_server_get_verify_state(td->o.name,
-                                       td->thread_number - 1, &data, &ver);
+                                       td->thread_number - 1, &data);
                if (!ret)
-                       verify_convert_assign_state(td, data, ver);
+                       verify_assign_state(td, data);
        } else
                ret = verify_load_state(td, "local");
 
@@ -2116,6 +2149,7 @@ reap:
                struct thread_data *map[REAL_MAX_JOBS];
                struct timeval this_start;
                int this_jobs = 0, left;
+               struct fork_data *fd;
 
                /*
                 * create threads (TD_NOT_CREATED -> TD_CREATED)
@@ -2165,14 +2199,13 @@ reap:
                        map[this_jobs++] = td;
                        nr_started++;
 
+                       fd = calloc(1, sizeof(*fd));
+                       fd->td = td;
+                       fd->sk_out = sk_out;
+
                        if (td->o.use_thread) {
-                               struct fork_data *fd;
                                int ret;
 
-                               fd = calloc(1, sizeof(*fd));
-                               fd->td = td;
-                               fd->sk_out = sk_out;
-
                                dprint(FD_PROCESS, "will pthread_create\n");
                                ret = pthread_create(&td->thread, NULL,
                                                        thread_main, fd);
@@ -2192,8 +2225,9 @@ reap:
                                dprint(FD_PROCESS, "will fork\n");
                                pid = fork();
                                if (!pid) {
-                                       int ret = fork_main(sk_out, shm_id, i);
+                                       int ret;
 
+                                       ret = (int)(uintptr_t)thread_main(fd);
                                        _exit(ret);
                                } else if (i == fio_debug_jobno)
                                        *fio_debug_jobp = pid;
@@ -2284,82 +2318,10 @@ reap:
        update_io_ticks();
 }
 
-static void wait_for_helper_thread_exit(void)
-{
-       void *ret;
-
-       helper_exit = 1;
-       pthread_cond_signal(&helper_cond);
-       pthread_join(helper_thread, &ret);
-}
-
 static void free_disk_util(void)
 {
        disk_util_prune_entries();
-
-       pthread_cond_destroy(&helper_cond);
-}
-
-static void *helper_thread_main(void *data)
-{
-       struct sk_out *sk_out = data;
-       int ret = 0;
-
-       sk_out_assign(sk_out);
-
-       fio_mutex_up(startup_mutex);
-
-       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();
-       }
-
-       sk_out_drop();
-       return NULL;
-}
-
-static int create_helper_thread(struct sk_out *sk_out)
-{
-       int ret;
-
-       setup_disk_util();
-
-       pthread_cond_init(&helper_cond, NULL);
-       pthread_mutex_init(&helper_lock, NULL);
-
-       ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
-       if (ret) {
-               log_err("Can't create helper thread: %s\n", strerror(ret));
-               return 1;
-       }
-
-       dprint(FD_MUTEX, "wait on startup_mutex\n");
-       fio_mutex_down(startup_mutex);
-       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
-       return 0;
+       helper_thread_destroy();
 }
 
 int fio_backend(struct sk_out *sk_out)
@@ -2392,14 +2354,14 @@ int fio_backend(struct sk_out *sk_out)
 
        set_genesis_time();
        stat_init();
-       create_helper_thread(sk_out);
+       helper_thread_create(startup_mutex, sk_out);
 
        cgroup_list = smalloc(sizeof(*cgroup_list));
        INIT_FLIST_HEAD(cgroup_list);
 
        run_threads(sk_out);
 
-       wait_for_helper_thread_exit();
+       helper_thread_exit();
 
        if (!fio_abort) {
                __show_run_stats();
@@ -2407,7 +2369,7 @@ int fio_backend(struct sk_out *sk_out)
                        for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                                struct io_log *log = agg_io_log[i];
 
-                               flush_log(log, 0);
+                               flush_log(log, false);
                                free_log(log);
                        }
                }