Track io_u state (free or in-flight)
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index e386f53..0ccd5d8 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -27,6 +27,7 @@
 #include <signal.h>
 #include <time.h>
 #include <locale.h>
+#include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
@@ -135,47 +136,22 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
-{
-       unsigned int old_next_file = td->next_file;
-       struct fio_file *f;
-
-       do {
-               f = &td->files[td->next_file];
-
-               td->next_file++;
-               if (td->next_file >= td->nr_files)
-                       td->next_file = 0;
-
-               if (f->fd != -1)
-                       break;
-
-               f = NULL;
-       } while (td->next_file != old_next_file);
-
-       return f;
-}
-
 /*
  * When job exits, we can cancel the in-flight IO if we are using async
  * io. Attempt to do so.
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
        struct list_head *entry, *n;
-       struct io_completion_data icd;
        struct io_u *io_u;
        int r;
 
        /*
         * get immediately available events, if any
         */
-       r = td_io_getevents(td, 0, td->cur_depth, &ts);
-       if (r > 0) {
-               icd.nr = r;
-               ios_completed(td, &icd);
-       }
+       r = io_u_queued_complete(td, 0, NULL);
+       if (r < 0)
+               return;
 
        /*
         * now cancel remaining active events
@@ -184,19 +160,24 @@ static void cleanup_pending_aio(struct thread_data *td)
                list_for_each_safe(entry, n, &td->io_u_busylist) {
                        io_u = list_entry(entry, struct io_u, list);
 
-                       r = td->io_ops->cancel(td, io_u);
-                       if (!r)
+                       /*
+                        * if the io_u isn't in flight, then that generally
+                        * means someone leaked an io_u. complain but fix
+                        * it up, so we don't stall here.
+                        */
+                       if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+                               log_err("fio: non-busy IO on busy list\n");
                                put_io_u(td, io_u);
+                       } else {
+                               r = td->io_ops->cancel(td, io_u);
+                               if (!r)
+                                       put_io_u(td, io_u);
+                       }
                }
        }
 
-       if (td->cur_depth) {
-               r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
-               if (r > 0) {
-                       icd.nr = r;
-                       ios_completed(td, &icd);
-               }
-       }
+       if (td->cur_depth)
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -206,7 +187,6 @@ static void cleanup_pending_aio(struct thread_data *td)
 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
        struct io_u *io_u = __get_io_u(td);
-       struct io_completion_data icd;
        int ret;
 
        if (!io_u)
@@ -220,24 +200,27 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
                return 1;
        }
 
+requeue:
        ret = td_io_queue(td, io_u);
-       if (ret) {
+       if (ret < 0) {
                td_verror(td, io_u->error);
                put_io_u(td, io_u);
                return 1;
-       }
-
-       ret = td_io_getevents(td, 1, td->cur_depth, NULL);
-       if (ret < 0) {
-               td_verror(td, ret);
-               return 1;
-       }
+       } else if (ret == FIO_Q_QUEUED) {
+               if (io_u_queued_complete(td, 1, NULL) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_COMPLETED) {
+               if (io_u->error) {
+                       td_verror(td, io_u->error);
+                       return 1;
+               }
 
-       icd.nr = ret;
-       ios_completed(td, &icd);
-       if (icd.error) {
-               td_verror(td, icd.error);
-               return 1;
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_BUSY) {
+               if (td_io_commit(td))
+                       return 1;
+               goto requeue;
        }
 
        return 0;
@@ -249,26 +232,28 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
  */
 static void do_verify(struct thread_data *td)
 {
-       struct io_u *io_u, *v_io_u = NULL;
-       struct io_completion_data icd;
        struct fio_file *f;
-       int ret, i;
+       struct io_u *io_u;
+       int ret, i, min_events;
 
        /*
         * sync io first and invalidate cache, to make sure we really
         * read from disk.
         */
        for_each_file(td, f, i) {
-               fio_io_sync(td, f);
-               file_invalidate_cache(td, f);
+               if (fio_io_sync(td, f))
+                       break;
+               if (file_invalidate_cache(td, f))
+                       break;
        }
 
-       td_set_runstate(td, TD_VERIFYING);
+       if (td->error)
+               return;
 
-       do {
-               if (td->terminate)
-                       break;
+       td_set_runstate(td, TD_VERIFYING);
 
+       io_u = NULL;
+       while (!td->terminate) {
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
@@ -283,60 +268,62 @@ static void do_verify(struct thread_data *td)
                        break;
                }
 
-               f = get_next_file(td);
-               if (!f)
-                       break;
-
-               io_u->file = f;
-
                if (td_io_prep(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
                }
-
+requeue:
                ret = td_io_queue(td, io_u);
-               if (ret) {
-                       td_verror(td, io_u->error);
-                       put_io_u(td, io_u);
-                       break;
-               }
 
-               /*
-                * we have one pending to verify, do that while
-                * we are doing io on the next one
-                */
-               if (do_io_u_verify(td, &v_io_u))
-                       break;
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (io_u->error)
+                               ret = -io_u->error;
+                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
 
-               ret = td_io_getevents(td, 1, 1, NULL);
-               if (ret != 1) {
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       ret = io_u_sync_complete(td, io_u, verify_io_u);
                        if (ret < 0)
-                               td_verror(td, ret);
+                               break;
+                       continue;
+               case FIO_Q_QUEUED:
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       td_verror(td, -ret);
                        break;
                }
 
-               v_io_u = td->io_ops->event(td, 0);
-               icd.nr = 1;
-               icd.error = 0;
-               fio_gettime(&icd.time, NULL);
-               io_completed(td, v_io_u, &icd);
-
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       put_io_u(td, v_io_u);
-                       v_io_u = NULL;
+               if (ret < 0 || td->error)
                        break;
-               }
 
                /*
-                * if we can't submit more io, we need to verify now
+                * if we can queue more, do so. but check if there are
+                * completed io_u's first.
                 */
-               if (queue_full(td) && do_io_u_verify(td, &v_io_u))
-                       break;
+               min_events = 0;
+               if (queue_full(td) || ret == FIO_Q_BUSY) {
+                       min_events = 1;
 
-       } while (1);
+                       if (td->cur_depth > td->iodepth_low)
+                               min_events = td->cur_depth - td->iodepth_low;
+               }
 
-       do_io_u_verify(td, &v_io_u);
+               /*
+                * 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, verify_io_u) < 0)
+                       break;
+       }
 
        if (td->cur_depth)
                cleanup_pending_aio(td);
@@ -375,103 +362,131 @@ static void do_cpuio(struct thread_data *td)
  */
 static void do_io(struct thread_data *td)
 {
-       struct io_completion_data icd;
        struct timeval s;
        unsigned long usec;
-       struct fio_file *f;
        int i, ret = 0;
 
        td_set_runstate(td, TD_RUNNING);
 
        while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
-               struct timespec *timeout;
+               struct timeval comp_time;
+               long bytes_done = 0;
                int min_evts = 0;
                struct io_u *io_u;
 
                if (td->terminate)
                        break;
 
-               f = get_next_file(td);
-               if (!f)
-                       break;
-
-               io_u = get_io_u(td, f);
+               io_u = get_io_u(td);
                if (!io_u)
                        break;
 
                memcpy(&s, &io_u->start_time, sizeof(s));
+
+               if (runtime_exceeded(td, &s)) {
+                       put_io_u(td, io_u);
+                       break;
+               }
 requeue:
                ret = td_io_queue(td, io_u);
-               if (ret) {
-                       if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
-                           io_u->resid) {
-                               /*
-                                * short read/write. requeue.
-                                */
+
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (io_u->error) {
+                               ret = io_u->error;
+                               break;
+                       }
+                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
+
                                io_u->xfer_buflen = io_u->resid;
-                               io_u->xfer_buf += ret;
+                               io_u->xfer_buf += bytes;
                                goto requeue;
-                       } else {
-                               put_io_u(td, io_u);
-                               break;
                        }
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_sync_complete(td, io_u, NULL);
+                       if (bytes_done < 0)
+                               ret = bytes_done;
+                       break;
+               case FIO_Q_QUEUED:
+                       /*
+                        * if the engine doesn't have a commit hook,
+                        * the io_u is really queued. if it does have such
+                        * a hook, it has to call io_u_queued() itself.
+                        */
+                       if (td->io_ops->commit == NULL)
+                               io_u_queued(td, io_u);
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       put_io_u(td, io_u);
+                       break;
                }
 
-               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
-               if (td->cur_depth < td->iodepth) {
-                       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+               if (ret < 0 || td->error)
+                       break;
 
-                       timeout = &ts;
+               /*
+                * See if we need to complete some commands
+                */
+               if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
                        min_evts = 0;
-               } else {
-                       timeout = NULL;
-                       min_evts = 1;
-               }
+                       if (queue_full(td) || ret == FIO_Q_BUSY) {
+                               min_evts = 1;
 
-               ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-               if (ret < 0) {
-                       td_verror(td, ret);
-                       break;
-               } else if (!ret)
-                       continue;
+                               if (td->cur_depth > td->iodepth_low)
+                                       min_evts = td->cur_depth - td->iodepth_low;
+                       }
 
-               icd.nr = ret;
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       break;
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
+                       if (bytes_done < 0)
+                               break;
                }
 
+               if (!bytes_done)
+                       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
                 */
-               usec = utime_since(&s, &icd.time);
+               usec = utime_since(&s, &comp_time);
 
-               rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
+               rate_throttle(td, usec, bytes_done, td->ddir);
 
-               if (check_min_rate(td, &icd.time)) {
+               if (check_min_rate(td, &comp_time)) {
                        if (exitall_on_terminate)
                                terminate_threads(td->groupid, 0);
                        td_verror(td, ENODATA);
                        break;
                }
 
-               if (runtime_exceeded(td, &icd.time))
-                       break;
-
                if (td->thinktime) {
                        unsigned long long b;
 
                        b = td->io_blocks[0] + td->io_blocks[1];
-                       if (!(b % td->thinktime_blocks))
-                               usec_sleep(td, td->thinktime);
+                       if (!(b % td->thinktime_blocks)) {
+                               int left;
+
+                               if (td->thinktime_spin)
+                                       __usec_sleep(td->thinktime_spin);
+
+                               left = td->thinktime - td->thinktime_spin;
+                               if (left)
+                                       usec_sleep(td, left);
+                       }
                }
        }
 
        if (!td->error) {
+               struct fio_file *f;
+
                if (td->cur_depth)
                        cleanup_pending_aio(td);
 
@@ -548,6 +563,7 @@ static int init_io_u(struct thread_data *td)
                        fill_rand_buf(io_u, max_bs);
 
                io_u->index = i;
+               io_u->flags = IO_U_F_FREE;
                list_add(&io_u->list, &td->io_u_freelist);
        }
 
@@ -610,11 +626,15 @@ static void clear_io_state(struct thread_data *td)
        struct fio_file *f;
        int i;
 
-       td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+       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->last_was_sync = 0;
+
        for_each_file(td, f, i) {
+               f->last_completed_pos = 0;
+
                f->last_pos = 0;
                if (td->io_ops->flags & FIO_SYNCIO)
                        lseek(f->fd, SEEK_SET, 0);
@@ -640,6 +660,7 @@ static void *thread_main(void *data)
 
        INIT_LIST_HEAD(&td->io_u_freelist);
        INIT_LIST_HEAD(&td->io_u_busylist);
+       INIT_LIST_HEAD(&td->io_u_requeues);
        INIT_LIST_HEAD(&td->io_hist_list);
        INIT_LIST_HEAD(&td->io_log_list);
 
@@ -694,15 +715,15 @@ static void *thread_main(void *data)
        }
 
        fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ru_start);
+       getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
        runtime[0] = runtime[1] = 0;
        while (td->loops--) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
-                       memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
                clear_io_state(td);
                prune_io_piece_log(td);
@@ -738,12 +759,12 @@ static void *thread_main(void *data)
        td->runtime[0] = runtime[0] / 1000;
        td->runtime[1] = runtime[1] / 1000;
 
-       if (td->bw_log)
-               finish_log(td, td->bw_log, "bw");
-       if (td->slat_log)
-               finish_log(td, td->slat_log, "slat");
-       if (td->clat_log)
-               finish_log(td, td->clat_log, "clat");
+       if (td->ts.bw_log)
+               finish_log(td, td->ts.bw_log, "bw");
+       if (td->ts.slat_log)
+               finish_log(td, td->ts.slat_log, "slat");
+       if (td->ts.clat_log)
+               finish_log(td, td->ts.clat_log, "clat");
        if (td->write_iolog_file)
                write_iolog_close(td);
        if (td->exec_postrun) {
@@ -761,7 +782,7 @@ err:
        close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return (void *) td->error;
+       return (void *) (unsigned long) td->error;
 }
 
 /*
@@ -784,7 +805,7 @@ static int fork_main(int shmid, int offset)
        td = data + offset * sizeof(struct thread_data);
        ret = thread_main(td);
        shmdt(data);
-       return (int) ret;
+       return (int) (unsigned long) ret;
 }
 
 /*
@@ -793,7 +814,7 @@ static int fork_main(int shmid, int offset)
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
        struct thread_data *td;
-       int i, cputhreads, pending;
+       int i, cputhreads, pending, status, ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
@@ -807,6 +828,22 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
                        cputhreads++;
 
+               if (td->runstate < TD_EXITED) {
+                       /*
+                        * check if someone quit or got killed in an unusual way
+                        */
+                       ret = waitpid(td->pid, &status, WNOHANG);
+                       if (ret < 0)
+                               perror("waitpid");
+                       else if ((ret == td->pid) && WIFSIGNALED(status)) {
+                               int sig = WTERMSIG(status);
+
+                               log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
+               }
+
                if (td->runstate != TD_EXITED) {
                        if (td->runstate < TD_RUNNING)
                                pending++;
@@ -827,13 +864,16 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                } else {
                        int status;
 
-                       waitpid(td->pid, &status, 0);
-                       if (WIFEXITED(status) && WEXITSTATUS(status)) {
+                       ret = waitpid(td->pid, &status, 0);
+                       if (ret < 0)
+                               perror("waitpid");
+                       else if (WIFEXITED(status) && WEXITSTATUS(status)) {
                                if (!exit_value)
                                        exit_value++;
                        }
                }
 
+reaped:
                (*nr_running)--;
                (*m_rate) -= td->ratemin;
                (*t_rate) -= td->rate;