Indicate 'V' verify state always
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index d85f956f7c552a7e78b5fc18fdec1e6f9a17f372..6a72ce773820d0a19cad1497dd41e0a64ae31128 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -36,6 +36,7 @@
 
 #include "fio.h"
 #include "hash.h"
+#include "smalloc.h"
 
 unsigned long page_mask;
 unsigned long page_size;
@@ -49,7 +50,7 @@ int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
 
-static struct fio_sem *startup_sem;
+static struct fio_mutex *startup_mutex;
 static volatile int fio_abort;
 static int exit_value;
 
@@ -60,6 +61,11 @@ struct io_log *agg_io_log[2];
 
 static inline void td_set_runstate(struct thread_data *td, int runstate)
 {
+       if (td->runstate == runstate)
+               return;
+
+       dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
+                                                               runstate);
        td->runstate = runstate;
 }
 
@@ -70,10 +76,22 @@ static void terminate_threads(int group_id)
 
        for_each_td(td, i) {
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
-                       if (td->runstate <= TD_RUNNING)
-                               kill(td->pid, SIGQUIT);
+                       dprint(FD_PROCESS, "setting terminate on %d\n",td->pid);
+
                        td->terminate = 1;
                        td->o.start_delay = 0;
+
+                       /*
+                        * if the thread is running, just let it exit
+                        */
+                       if (td->runstate < TD_RUNNING)
+                               kill(td->pid, SIGQUIT);
+                       else {
+                               struct ioengine_ops *ops = td->io_ops;
+
+                               if (ops && (ops->flags & FIO_SIGQUIT))
+                                       kill(td->pid, SIGQUIT);
+                       }
                }
        }
 }
@@ -308,6 +326,7 @@ static void do_verify(struct thread_data *td)
 
                if (runtime_exceeded(td, &io_u->start_time)) {
                        put_io_u(td, io_u);
+                       td->terminate = 1;
                        break;
                }
 
@@ -344,7 +363,6 @@ static void do_verify(struct thread_data *td)
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
-                               f->last_completed_pos = io_u->offset;
 
                                td->ts.short_io_u[io_u->ddir]++;
 
@@ -438,9 +456,20 @@ static void do_io(struct thread_data *td)
 
                if (runtime_exceeded(td, &s)) {
                        put_io_u(td, io_u);
+                       td->terminate = 1;
                        break;
                }
 
+               /*
+                * Add verification end_io handler, if asked to verify
+                * a previously written file.
+                */
+               if (td->o.verify != VERIFY_NONE) {
+                       io_u->end_io = verify_io_u;
+                       td_set_runstate(td, TD_VERIFYING);
+               } else
+                       td_set_runstate(td, TD_RUNNING);
+
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
@@ -462,7 +491,6 @@ static void do_io(struct thread_data *td)
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
-                               f->last_completed_pos = io_u->offset;
 
                                td->ts.short_io_u[io_u->ddir]++;
 
@@ -556,6 +584,10 @@ sync_done:
                }
        }
 
+       if (td->o.fill_device && td->error == ENOSPC) {
+               td->error = 0;
+               td->terminate = 1;
+       }
        if (!td->error) {
                struct fio_file *f;
 
@@ -574,6 +606,12 @@ sync_done:
                }
        } else
                cleanup_pending_aio(td);
+
+       /*
+        * stop job if we failed doing any IO
+        */
+       if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
+               td->done = 1;
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -609,45 +647,44 @@ static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 
 static int init_io_u(struct thread_data *td)
 {
-       unsigned long long buf_size;
        struct io_u *io_u;
        unsigned int max_bs;
        int i, max_units;
        char *p;
 
-       if (td->io_ops->flags & FIO_SYNCIO)
-               max_units = 1;
-       else
-               max_units = td->o.iodepth;
-
+       max_units = td->o.iodepth;
        max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
-       buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
-       buf_size += page_mask;
-       if (buf_size != (size_t) buf_size) {
-               log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
-               return 1;
-       }
-
-       td->orig_buffer_size = buf_size;
+       td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
 
        if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
                td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
-       else if (td->orig_buffer_size & page_mask)
-               td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
+
+       if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
+               log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
+               return 1;
+       }
 
        if (allocate_io_mem(td))
                return 1;
 
-       p = ALIGN(td->orig_buffer);
+       if (td->o.odirect)
+               p = ALIGN(td->orig_buffer);
+       else
+               p = td->orig_buffer;
+
        for (i = 0; i < max_units; i++) {
+               if (td->terminate)
+                       return 1;
                io_u = malloc(sizeof(*io_u));
                memset(io_u, 0, sizeof(*io_u));
                INIT_LIST_HEAD(&io_u->list);
 
-               io_u->buf = p + max_bs * i;
+               if (!(td->io_ops->flags & FIO_NOIO)) {
+                       io_u->buf = p + max_bs * i;
 
-               if (td_write(td))
-                       fill_io_buf(td, io_u, max_bs);
+                       if (td_write(td))
+                               fill_io_buf(td, io_u, max_bs);
+               }
 
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
@@ -714,6 +751,26 @@ static int switch_ioscheduler(struct thread_data *td)
        return 0;
 }
 
+static int keep_running(struct thread_data *td)
+{
+       unsigned long long io_done;
+
+       if (td->done)
+               return 0;
+       if (td->o.time_based)
+               return 1;
+       if (td->o.loops) {
+               td->o.loops--;
+               return 1;
+       }
+
+       io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
+       if (io_done < td->o.size)
+               return 1;
+
+       return 0;
+}
+
 static int clear_io_state(struct thread_data *td)
 {
        struct fio_file *f;
@@ -729,11 +786,17 @@ static int clear_io_state(struct thread_data *td)
 
        td->last_was_sync = 0;
 
-       for_each_file(td, f, i)
-               td_io_close_file(td, f);
+       /*
+        * reset file done count if we are to start over
+        */
+       if (td->o.time_based || td->o.loops)
+               td->nr_done_files = 0;
+
+       close_files(td);
 
        ret = 0;
        for_each_file(td, f, i) {
+               f->flags &= ~FIO_FILE_DONE;
                ret = td_io_open_file(td, f);
                if (ret)
                        break;
@@ -748,9 +811,8 @@ static int clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
-       unsigned long long runtime[2];
+       unsigned long long runtime[2], elapsed;
        struct thread_data *td = data;
-       unsigned long elapsed;
        int clear_state;
 
        if (!td->o.use_thread)
@@ -758,6 +820,8 @@ static void *thread_main(void *data)
 
        td->pid = getpid();
 
+       dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
+
        INIT_LIST_HEAD(&td->io_u_freelist);
        INIT_LIST_HEAD(&td->io_u_busylist);
        INIT_LIST_HEAD(&td->io_u_requeues);
@@ -765,41 +829,45 @@ static void *thread_main(void *data)
        INIT_LIST_HEAD(&td->io_hist_list);
        td->io_hist_tree = RB_ROOT;
 
+       td_set_runstate(td, TD_INITIALIZED);
+       fio_mutex_up(startup_mutex);
+       fio_mutex_down(td->mutex);
+
+       /*
+        * the ->mutex mutex is now no longer used, close it to avoid
+        * eating a file descriptor
+        */
+       fio_mutex_remove(td->mutex);
+
+       /*
+        * May alter parameters that init_io_u() will use, so we need to
+        * do this first.
+        */
+       if (init_iolog(td))
+               goto err;
+
        if (init_io_u(td))
-               goto err_sem;
+               goto err;
 
-       if (fio_setaffinity(td) == -1) {
+       if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
                td_verror(td, errno, "cpu_set_affinity");
-               goto err_sem;
+               goto err;
        }
 
-       if (init_iolog(td))
-               goto err_sem;
-
-       if (td->ioprio) {
+       if (td->ioprio_set) {
                if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
                        td_verror(td, errno, "ioprio_set");
-                       goto err_sem;
+                       goto err;
                }
        }
 
        if (nice(td->o.nice) == -1) {
                td_verror(td, errno, "nice");
-               goto err_sem;
+               goto err;
        }
 
        if (td->o.ioscheduler && switch_ioscheduler(td))
-               goto err_sem;
-
-       td_set_runstate(td, TD_INITIALIZED);
-       fio_sem_up(startup_sem);
-       fio_sem_down(td->mutex);
-
-       /*
-        * the ->mutex semaphore is now no longer used, close it to avoid
-        * eating a file descriptor
-        */
-       fio_sem_remove(td->mutex);
+               goto err;
 
        if (!td->o.create_serialize && setup_files(td))
                goto err;
@@ -824,7 +892,7 @@ static void *thread_main(void *data)
 
        runtime[0] = runtime[1] = 0;
        clear_state = 0;
-       while (td->o.loops--) {
+       while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
                memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
@@ -860,7 +928,9 @@ static void *thread_main(void *data)
                if (td->error || td->terminate)
                        break;
 
-               if (td->o.verify == VERIFY_NONE)
+               if (!td->o.do_verify ||
+                   td->o.verify == VERIFY_NONE ||
+                   (td->io_ops->flags & FIO_UNIDIR))
                        continue;
 
                if (clear_io_state(td))
@@ -877,8 +947,8 @@ static void *thread_main(void *data)
        }
 
        update_rusage_stat(td);
-       td->ts.runtime[0] = runtime[0] / 1000;
-       td->ts.runtime[1] = runtime[1] / 1000;
+       td->ts.runtime[0] = (runtime[0] + 999) / 1000;
+       td->ts.runtime[1] = (runtime[1] + 999) / 1000;
        td->ts.total_run_time = mtime_since_now(&td->epoch);
        td->ts.io_bytes[0] = td->io_bytes[0];
        td->ts.io_bytes[1] = td->io_bytes[1];
@@ -889,8 +959,6 @@ static void *thread_main(void *data)
                finish_log(td, td->ts.slat_log, "slat");
        if (td->ts.clat_log)
                finish_log(td, td->ts.clat_log, "clat");
-       if (td->o.write_iolog_file)
-               write_iolog_close(td);
        if (td->o.exec_postrun) {
                if (system(td->o.exec_postrun) < 0)
                        log_err("fio: postrun %s failed\n", td->o.exec_postrun);
@@ -902,15 +970,19 @@ static void *thread_main(void *data)
 err:
        if (td->error)
                printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
-       close_files(td);
+       close_and_free_files(td);
        close_ioengine(td);
        cleanup_io_u(td);
+
+       /*
+        * do this very late, it will log file closing as well
+        */
+       if (td->o.write_iolog_file)
+               write_iolog_close(td);
+
        options_mem_free(td);
        td_set_runstate(td, TD_EXITED);
        return (void *) (unsigned long) td->error;
-err_sem:
-       fio_sem_up(startup_sem);
-       goto err;
 }
 
 /*
@@ -942,12 +1014,12 @@ 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, status, ret;
+       int i, cputhreads, realthreads, pending, status, ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
         */
-       pending = cputhreads = 0;
+       realthreads = pending = cputhreads = 0;
        for_each_td(td, i) {
                int flags = 0;
 
@@ -957,8 +1029,14 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                 */
                if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
                        cputhreads++;
+               else
+                       realthreads++;
 
-               if (!td->pid || td->runstate == TD_REAPED)
+               if (!td->pid) {
+                       pending++;
+                       continue;
+               }
+               if (td->runstate == TD_REAPED)
                        continue;
                if (td->o.use_thread) {
                        if (td->runstate == TD_EXITED) {
@@ -1010,8 +1088,11 @@ reaped:
                if (td->o.use_thread) {
                        long ret;
 
-                       if (pthread_join(td->thread, (void *) &ret))
+                       dprint(FD_PROCESS, "joining tread %d\n", td->pid);
+                       if (pthread_join(td->thread, (void *) &ret)) {
+                               dprint(FD_PROCESS, "join failed %ld\n", ret);
                                perror("pthread_join");
+                       }
                }
 
                (*nr_running)--;
@@ -1023,7 +1104,7 @@ reaped:
                        exit_value++;
        }
 
-       if (*nr_running == cputhreads && !pending)
+       if (*nr_running == cputhreads && !pending && realthreads)
                terminate_threads(TERMINATE_ALL);
 }
 
@@ -1079,7 +1160,18 @@ static void run_threads(void)
                                log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
-               }
+               } else {
+                       struct fio_file *f;
+                       unsigned int i;
+
+                       /*
+                        * for sharing to work, each job must always open
+                        * its own files. so close them, if we opened them
+                        * for creation
+                        */
+                       for_each_file(td, f, i)
+                               td_io_close_file(td, f);
+               }
 
                init_disk_util(td);
        }
@@ -1126,19 +1218,21 @@ static void run_threads(void)
                        nr_started++;
 
                        if (td->o.use_thread) {
+                               dprint(FD_PROCESS, "will pthread_create\n");
                                if (pthread_create(&td->thread, NULL, thread_main, td)) {
                                        perror("thread_create");
                                        nr_started--;
                                        break;
                                }
                        } else {
+                               dprint(FD_PROCESS, "will fork\n");
                                if (!fork()) {
                                        int ret = fork_main(shm_id, i);
 
                                        exit(ret);
                                }
                        }
-                       fio_sem_down(startup_sem);
+                       fio_mutex_down(startup_mutex);
                }
 
                /*
@@ -1193,7 +1287,7 @@ static void run_threads(void)
                        m_rate += td->o.ratemin;
                        t_rate += td->o.rate;
                        todo--;
-                       fio_sem_up(td->mutex);
+                       fio_mutex_up(td->mutex);
                }
 
                reap_threads(&nr_running, &t_rate, &m_rate);
@@ -1215,6 +1309,8 @@ int main(int argc, char *argv[])
 {
        long ps;
 
+       sinit();
+
        /*
         * We need locale for number printing, if it isn't set then just
         * go with the US format.
@@ -1242,7 +1338,7 @@ int main(int argc, char *argv[])
                setup_log(&agg_io_log[DDIR_WRITE]);
        }
 
-       startup_sem = fio_sem_init(0);
+       startup_mutex = fio_mutex_init(0);
 
        set_genesis_time();
 
@@ -1258,6 +1354,6 @@ int main(int argc, char *argv[])
                }
        }
 
-       fio_sem_remove(startup_sem);
+       fio_mutex_remove(startup_mutex);
        return exit_value;
 }