Cleanup end IO handling
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 10625f26afd25fd54d72bb896a17c075f0681401..78deaf66a732b6e92434cf0ccef52b66b62d5a71 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -26,6 +26,8 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <locale.h>
+#include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
@@ -46,6 +48,7 @@ int temp_stall_ts;
 
 static volatile int startup_sem;
 static volatile int fio_abort;
+static int exit_value;
 
 struct io_log *agg_io_log[2];
 
@@ -160,20 +163,14 @@ static struct fio_file *get_next_file(struct thread_data *td)
  */
 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);
-       }
+       io_u_queued_complete(td, 0, NULL);
 
        /*
         * now cancel remaining active events
@@ -188,13 +185,8 @@ static void cleanup_pending_aio(struct thread_data *td)
                }
        }
 
-       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)
+               io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -204,7 +196,6 @@ 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)
@@ -219,23 +210,20 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
        }
 
        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;
-       }
+       } else if (ret == FIO_Q_QUEUED) {
+               if (io_u_queued_complete(td, 1, NULL))
+                       return 1;
+       } else if (ret == FIO_Q_COMPLETED) {
+               if (io_u->error) {
+                       td_verror(td, io_u->error);
+                       return 1;
+               }
 
-       ret = td_io_getevents(td, 1, td->cur_depth, NULL);
-       if (ret < 0) {
-               td_verror(td, ret);
-               return 1;
-       }
-
-       icd.nr = ret;
-       ios_completed(td, &icd);
-       if (icd.error) {
-               td_verror(td, icd.error);
-               return 1;
+               io_u_sync_complete(td, io_u, NULL);
        }
 
        return 0;
@@ -247,10 +235,9 @@ 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
@@ -263,78 +250,67 @@ static void do_verify(struct thread_data *td)
 
        td_set_runstate(td, TD_VERIFYING);
 
-       do {
-               if (td->terminate)
-                       break;
-
+       io_u = NULL;
+       while (!td->terminate) {
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
-               if (runtime_exceeded(td, &io_u->start_time)) {
-                       put_io_u(td, io_u);
+               if (runtime_exceeded(td, &io_u->start_time))
                        break;
-               }
 
-               if (get_next_verify(td, io_u)) {
-                       put_io_u(td, io_u);
+               if (get_next_verify(td, io_u))
                        break;
-               }
 
-               f = get_next_file(td);
-               if (!f)
+               if (td_io_prep(td, io_u))
                        break;
 
-               io_u->file = f;
+requeue:
+               ret = td_io_queue(td, io_u);
 
-               if (td_io_prep(td, io_u)) {
-                       put_io_u(td, io_u);
+               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;
+
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       if (io_u_sync_complete(td, io_u, verify_io_u))
+                               break;
+                       continue;
+               case FIO_Q_QUEUED:
+                       break;
+               default:
+                       assert(ret < 0);
+                       td_verror(td, -ret);
                        break;
                }
 
-               ret = td_io_queue(td, io_u);
-               if (ret) {
-                       td_verror(td, io_u->error);
-                       put_io_u(td, io_u);
+               if (ret < 0)
                        break;
-               }
 
                /*
-                * we have one pending to verify, do that while
-                * we are doing io on the next one
+                * if we can queue more, do so. but check if there are
+                * completed io_u's first.
                 */
-               if (do_io_u_verify(td, &v_io_u))
-                       break;
-
-               ret = td_io_getevents(td, 1, 1, NULL);
-               if (ret != 1) {
-                       if (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;
-                       break;
-               }
+               min_events = 0;
+               if (queue_full(td))
+                       min_events = 1;
 
                /*
-                * if we can't submit more io, we need to verify now
+                * Reap required number of io units, if any, and do the
+                * verification on them through the callback handler
                 */
-               if (queue_full(td) && do_io_u_verify(td, &v_io_u))
+               if (io_u_queued_complete(td, min_events, verify_io_u))
                        break;
+       }
 
-       } while (1);
-
-       do_io_u_verify(td, &v_io_u);
+       if (io_u)
+               put_io_u(td, io_u);
 
        if (td->cur_depth)
                cleanup_pending_aio(td);
@@ -373,7 +349,6 @@ 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;
@@ -381,8 +356,9 @@ static void do_io(struct thread_data *td)
 
        td_set_runstate(td, TD_RUNNING);
 
-       while (td->this_io_bytes[td->ddir] < td->io_size) {
-               struct timespec *timeout;
+       while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
+               struct timeval comp_time;
+               long bytes_done = 0;
                int min_evts = 0;
                struct io_u *io_u;
 
@@ -399,75 +375,89 @@ static void do_io(struct thread_data *td)
 
                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 {
-                               td_verror(td, io_u->error);
-                               put_io_u(td, io_u);
-                               break;
                        }
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_sync_complete(td, io_u, NULL);
+                       break;
+               case FIO_Q_QUEUED:
+                       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 (ret < 0)
+                       break;
 
-               if (td->cur_depth < td->iodepth) {
-                       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 
-                       timeout = &ts;
+               /*
+                * See if we need to complete some commands
+                */
+               if (ret == FIO_Q_QUEUED) {
                        min_evts = 0;
-               } else {
-                       timeout = NULL;
-                       min_evts = 1;
+                       if (queue_full(td))
+                               min_evts = 1;
+
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
+                       if (bytes_done < 0)
+                               break;
                }
 
-               ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-               if (ret < 0) {
-                       td_verror(td, ret);
-                       break;
-               } else if (!ret)
+               if (!bytes_done)
                        continue;
 
-               icd.nr = ret;
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       break;
-               }
-
                /*
                 * 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);
+                       }
                }
        }
 
@@ -755,33 +745,36 @@ static void *thread_main(void *data)
                terminate_threads(td->groupid, 0);
 
 err:
+       if (td->error)
+               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
        close_files(td);
        close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return NULL;
-
+       return (void *) td->error;
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
-static void *fork_main(int shmid, int offset)
+static int fork_main(int shmid, int offset)
 {
        struct thread_data *td;
-       void *data;
+       void *data, *ret;
 
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
+               int __err = errno;
+
                perror("shmat");
-               return NULL;
+               return __err;
        }
 
        td = data + offset * sizeof(struct thread_data);
-       thread_main(td);
+       ret = thread_main(td);
        shmdt(data);
-       return NULL;
+       return (int) ret;
 }
 
 /*
@@ -790,7 +783,7 @@ static void *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)
@@ -804,6 +797,22 @@ 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++;
@@ -811,6 +820,9 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                        continue;
                }
 
+               if (td->error)
+                       exit_value++;
+
                td_set_runstate(td, TD_REAPED);
 
                if (td->use_thread) {
@@ -818,9 +830,19 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 
                        if (pthread_join(td->thread, (void *) &ret))
                                perror("thread_join");
-               } else
-                       waitpid(td->pid, NULL, 0);
+               } else {
+                       int 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;
@@ -869,6 +891,9 @@ static void run_threads(void)
                 * client data interspersed on disk
                 */
                if (setup_files(td)) {
+                       exit_value++;
+                       if (td->error)
+                               log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
@@ -925,8 +950,9 @@ static void run_threads(void)
                                if (fork())
                                        fio_sem_down(&startup_sem);
                                else {
-                                       fork_main(shm_id, i);
-                                       exit(0);
+                                       int ret = fork_main(shm_id, i);
+
+                                       exit(ret);
                                }
                        }
                }
@@ -1005,6 +1031,13 @@ int main(int argc, char *argv[])
 {
        long ps;
 
+       /*
+        * We need locale for number printing, if it isn't set then just
+        * go with the US format.
+        */
+       if (!getenv("LC_NUMERIC"))
+               setlocale(LC_NUMERIC, "en_US");
+
        if (parse_options(argc, argv))
                return 1;
 
@@ -1038,5 +1071,5 @@ int main(int argc, char *argv[])
                }
        }
 
-       return 0;
+       return exit_value;
 }