backend: move freeing of td->mutex to main thread
[fio.git] / backend.c
index b43486dc7b2ff016b6d76d281dc479692e90b231..38ef3482f85139143aca01dfea12a49ec232db19 100644 (file)
--- a/backend.c
+++ b/backend.c
@@ -76,9 +76,6 @@ int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
 
-#define PAGE_ALIGN(buf)        \
-       (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
-
 #define JOB_START_TIMEOUT      (5 * 1000)
 
 static void sig_int(int sig)
@@ -180,8 +177,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
                         * check bandwidth specified rate
                         */
                        if (bytes < td->rate_bytes[ddir]) {
-                               log_err("%s: min rate %u not met\n", td->o.name,
-                                                               ratemin);
+                               log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
+                                       td->o.name, ratemin, bytes);
                                return true;
                        } else {
                                if (spent)
@@ -191,9 +188,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 
                                if (rate < ratemin ||
                                    bytes < td->rate_bytes[ddir]) {
-                                       log_err("%s: min rate %u not met, got"
-                                               " %luKB/sec\n", td->o.name,
-                                                       ratemin, rate);
+                                       log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
+                                               td->o.name, ratemin, rate);
                                        return true;
                                }
                        }
@@ -202,8 +198,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
                         * checks iops specified rate
                         */
                        if (iops < rate_iops) {
-                               log_err("%s: min iops rate %u not met\n",
-                                               td->o.name, rate_iops);
+                               log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
+                                               td->o.name, rate_iops, iops);
                                return true;
                        } else {
                                if (spent)
@@ -213,9 +209,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 
                                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,
-                                                       rate_iops_min, rate);
+                                       log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
+                                               td->o.name, rate_iops_min, rate);
                                        return true;
                                }
                        }
@@ -441,11 +436,8 @@ 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 (td->flags & TD_F_REGROW_LOGS)
+               return io_u_quiesce(td);
 
        /*
         * if the queue is full, we MUST reap at least 1 event
@@ -771,21 +763,21 @@ static bool exceeds_number_ios(struct thread_data *td)
        return number_ios >= (td->o.number_ios * td->loops);
 }
 
-static bool io_issue_bytes_exceeded(struct thread_data *td)
+static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
 {
        unsigned long long bytes, limit;
 
        if (td_rw(td))
-               bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
+               bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
        else if (td_write(td))
-               bytes = td->io_issue_bytes[DDIR_WRITE];
+               bytes = this_bytes[DDIR_WRITE];
        else if (td_read(td))
-               bytes = td->io_issue_bytes[DDIR_READ];
+               bytes = this_bytes[DDIR_READ];
        else
-               bytes = td->io_issue_bytes[DDIR_TRIM];
+               bytes = this_bytes[DDIR_TRIM];
 
-       if (td->o.io_limit)
-               limit = td->o.io_limit;
+       if (td->o.io_size)
+               limit = td->o.io_size;
        else
                limit = td->o.size;
 
@@ -793,26 +785,14 @@ static bool io_issue_bytes_exceeded(struct thread_data *td)
        return bytes >= limit || exceeds_number_ios(td);
 }
 
-static bool io_complete_bytes_exceeded(struct thread_data *td)
+static bool io_issue_bytes_exceeded(struct thread_data *td)
 {
-       unsigned long long bytes, limit;
-
-       if (td_rw(td))
-               bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
-       else if (td_write(td))
-               bytes = td->this_io_bytes[DDIR_WRITE];
-       else if (td_read(td))
-               bytes = td->this_io_bytes[DDIR_READ];
-       else
-               bytes = td->this_io_bytes[DDIR_TRIM];
-
-       if (td->o.io_limit)
-               limit = td->o.io_limit;
-       else
-               limit = td->o.size;
+       return io_bytes_exceeded(td, td->io_issue_bytes);
+}
 
-       limit *= td->loops;
-       return bytes >= limit || exceeds_number_ios(td);
+static bool io_complete_bytes_exceeded(struct thread_data *td)
+{
+       return io_bytes_exceeded(td, td->this_io_bytes);
 }
 
 /*
@@ -871,11 +851,11 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
 
        total_bytes = td->o.size;
        /*
-       * Allow random overwrite workloads to write up to io_limit
+       * Allow random overwrite workloads to write up to io_size
        * before starting verification phase as 'size' doesn't apply.
        */
        if (td_write(td) && td_random(td) && td->o.norandommap)
-               total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
+               total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
        /*
         * If verify_backlog is enabled, we'll run the verify in this
         * handler as well. For that case, we may need up to twice the
@@ -1215,7 +1195,7 @@ static int init_io_u(struct thread_data *td)
 
        if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
            td_ioengine_flagged(td, FIO_RAWIO))
-               p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
+               p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
        else
                p = td->orig_buffer;
 
@@ -1281,6 +1261,10 @@ static int init_io_u(struct thread_data *td)
        return 0;
 }
 
+/*
+ * This function is Linux specific.
+ * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
+ */
 static int switch_ioscheduler(struct thread_data *td)
 {
 #ifdef FIO_HAVE_IOSCHED_SWITCH
@@ -1291,7 +1275,8 @@ static int switch_ioscheduler(struct thread_data *td)
        if (td_ioengine_flagged(td, FIO_DISKLESSIO))
                return 0;
 
-       sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
+       assert(td->files && td->files[0]);
+       sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
 
        f = fopen(tmp, "r+");
        if (!f) {
@@ -1370,8 +1355,8 @@ static bool keep_running(struct thread_data *td)
        if (exceeds_number_ios(td))
                return false;
 
-       if (td->o.io_limit)
-               limit = td->o.io_limit;
+       if (td->o.io_size)
+               limit = td->o.io_size;
        else
                limit = td->o.size;
 
@@ -1379,14 +1364,14 @@ static bool keep_running(struct thread_data *td)
                uint64_t diff;
 
                /*
-                * If the difference is less than the minimum IO size, we
+                * If the difference is less than the maximum IO size, we
                 * are done.
                 */
                diff = limit - ddir_rw_sum(td->io_bytes);
                if (diff < td_max_bs(td))
                        return false;
 
-               if (fio_files_done(td) && !td->o.io_limit)
+               if (fio_files_done(td) && !td->o.io_size)
                        return false;
 
                return true;
@@ -1471,6 +1456,7 @@ 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;
+       int deadlock_loop_cnt;
        int clear_state;
        int ret;
 
@@ -1675,10 +1661,11 @@ static void *thread_main(void *data)
        if (rate_submit_init(td, sk_out))
                goto err;
 
-       fio_gettime(&td->epoch, NULL);
+       set_epoch_time(td, o->log_unix_epoch);
        fio_getrusage(&td->ru_start);
        memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
        memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
+       memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
 
        if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
                        o->ratemin[DDIR_TRIM]) {
@@ -1706,7 +1693,7 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
-               if (td->o.verify_only && (td_write(td) || td_rw(td)))
+               if (td->o.verify_only && td_write(td))
                        verify_bytes = do_dry_run(td);
                else {
                        uint64_t bytes_done[DDIR_RWDIR_CNT];
@@ -1722,6 +1709,14 @@ static void *thread_main(void *data)
                        }
                }
 
+               /*
+                * If we took too long to shut down, the main thread could
+                * already consider us reaped/exited. If that happens, break
+                * out and clean up.
+                */
+               if (td->runstate >= TD_EXITED)
+                       break;
+
                clear_state = 1;
 
                /*
@@ -1731,9 +1726,19 @@ static void *thread_main(void *data)
                 * the rusage_sem, which would never get upped because
                 * this thread is waiting for the stat mutex.
                 */
-               check_update_rusage(td);
+               deadlock_loop_cnt = 0;
+               do {
+                       check_update_rusage(td);
+                       if (!fio_mutex_down_trylock(stat_mutex))
+                               break;
+                       usleep(1000);
+                       if (deadlock_loop_cnt++ > 5000) {
+                               log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
+                               td->error = EDEADLK;
+                               goto err;
+                       }
+               } while (1);
 
-               fio_mutex_down(stat_mutex);
                if (td_read(td) && td->io_bytes[DDIR_READ])
                        update_runtime(td, elapsed_us, DDIR_READ);
                if (td_write(td) && td->io_bytes[DDIR_WRITE])
@@ -1831,9 +1836,6 @@ err:
        if (o->write_iolog_file)
                write_iolog_close(td);
 
-       fio_mutex_remove(td->mutex);
-       td->mutex = NULL;
-
        td_set_runstate(td, TD_EXITED);
 
        /*
@@ -1857,8 +1859,8 @@ static void dump_td_info(struct thread_data *td)
 /*
  * Run over the job map and reap the threads that have exited, if any.
  */
-static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
-                        unsigned int *m_rate)
+static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
+                        uint64_t *m_rate)
 {
        struct thread_data *td;
        unsigned int cputhreads, realthreads, pending;
@@ -2051,8 +2053,16 @@ static bool check_mount_writes(struct thread_data *td)
        if (!td_write(td) || td->o.allow_mounted_write)
                return false;
 
+       /*
+        * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
+        * are mkfs'd and mounted.
+        */
        for_each_file(td, f, i) {
-               if (f->filetype != FIO_TYPE_BD)
+#ifdef FIO_HAVE_CHARDEV_SIZE
+               if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
+#else
+               if (f->filetype != FIO_TYPE_BLOCK)
+#endif
                        continue;
                if (device_is_mounted(f->file_name))
                        goto mounted;
@@ -2060,7 +2070,7 @@ static bool check_mount_writes(struct thread_data *td)
 
        return false;
 mounted:
-       log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
+       log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
        return true;
 }
 
@@ -2096,7 +2106,8 @@ static bool waitee_running(struct thread_data *me)
 static void run_threads(struct sk_out *sk_out)
 {
        struct thread_data *td;
-       unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
+       unsigned int i, todo, nr_running, nr_started;
+       uint64_t m_rate, t_rate;
        uint64_t spent;
 
        if (fio_gtod_offload && fio_start_gtod_thread())
@@ -2410,11 +2421,19 @@ int fio_backend(struct sk_out *sk_out)
        }
 
        for_each_td(td, i) {
+               if (td->ss.dur) {
+                       if (td->ss.iops_data != NULL) {
+                               free(td->ss.iops_data);
+                               free(td->ss.bw_data);
+                       }
+               }
                fio_options_free(td);
                if (td->rusage_sem) {
                        fio_mutex_remove(td->rusage_sem);
                        td->rusage_sem = NULL;
                }
+               fio_mutex_remove(td->mutex);
+               td->mutex = NULL;
        }
 
        free_disk_util();