Merge branch 'samples-colnames' of https://github.com/parallel-fs-utils/fio
[fio.git] / backend.c
index 033d5a707f2a896f161eab84299e2b1bf6a23de0..cc3c4e78e9d8ae6b8fd9cdc245272fc395369f34 100644 (file)
--- a/backend.c
+++ b/backend.c
@@ -29,6 +29,7 @@
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <math.h>
+#include <pthread.h>
 
 #include "fio.h"
 #include "smalloc.h"
 #include "rate-submit.h"
 #include "helper_thread.h"
 #include "pshared.h"
+#include "zone-dist.h"
 
 static struct fio_sem *startup_sem;
 static struct flist_head *cgroup_list;
-static char *cgroup_mnt;
+static struct cgroup_mnt *cgroup_mnt;
 static int exit_value;
-static volatile int fio_abort;
+static volatile bool fio_abort;
 static unsigned int nr_process = 0;
 static unsigned int nr_thread = 0;
 
@@ -64,6 +66,7 @@ unsigned int stat_number = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
+pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
 
 #define JOB_START_TIMEOUT      (5 * 1000)
 
@@ -432,9 +435,7 @@ static int wait_for_completions(struct thread_data *td, struct timespec *time)
        if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
                min_evts = 1;
 
-       if (time && (__should_check_rate(td, DDIR_READ) ||
-           __should_check_rate(td, DDIR_WRITE) ||
-           __should_check_rate(td, DDIR_TRIM)))
+       if (time && __should_check_rate(td))
                fio_gettime(time, NULL);
 
        do {
@@ -456,14 +457,14 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                        *ret = -io_u->error;
                        clear_io_u(td, io_u);
                } else if (io_u->resid) {
-                       int bytes = io_u->xfer_buflen - io_u->resid;
+                       long long bytes = io_u->xfer_buflen - io_u->resid;
                        struct fio_file *f = io_u->file;
 
                        if (bytes_issued)
                                *bytes_issued += bytes;
 
                        if (!from_verify)
-                               trim_io_piece(td, io_u);
+                               trim_io_piece(io_u);
 
                        /*
                         * zero read, fail
@@ -489,9 +490,7 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                        requeue_io_u(td, &io_u);
                } else {
 sync_done:
-                       if (comp_time && (__should_check_rate(td, DDIR_READ) ||
-                           __should_check_rate(td, DDIR_WRITE) ||
-                           __should_check_rate(td, DDIR_TRIM)))
+                       if (comp_time && __should_check_rate(td))
                                fio_gettime(comp_time, NULL);
 
                        *ret = io_u_sync_complete(td, io_u);
@@ -570,7 +569,7 @@ static int unlink_all_files(struct thread_data *td)
 /*
  * Check if io_u will overlap an in-flight IO in the queue
  */
-static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
+bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 {
        bool overlap;
        struct io_u *check_io_u;
@@ -587,7 +586,7 @@ static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
 
                        if (x1 < y2 && y1 < x2) {
                                overlap = true;
-                               dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
+                               dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
                                                x1, io_u->buflen,
                                                y1, check_io_u->buflen);
                                break;
@@ -892,6 +891,8 @@ static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
                        over = (usperop - total) / usperop * -bs;
 
                td->rate_io_issue_bytes[ddir] += (missed - over);
+               /* adjust for rate_process=poisson */
+               td->last_usec[ddir] += total;
        }
 }
 
@@ -968,8 +969,10 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                 * 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.
+                * Jobs read from iolog do not use this stop condition.
                 */
                if (bytes_issued >= total_bytes &&
+                   !td->o.read_iolog_file &&
                    (!td->o.time_based ||
                     (td->o.time_based && td->o.verify != VERIFY_NONE)))
                        break;
@@ -1035,8 +1038,8 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                        log_io_piece(td, io_u);
 
                if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
-                       const unsigned long blen = io_u->xfer_buflen;
-                       const enum fio_ddir ddir = acct_ddir(io_u);
+                       const unsigned long long blen = io_u->xfer_buflen;
+                       const enum fio_ddir __ddir = acct_ddir(io_u);
 
                        if (td->error)
                                break;
@@ -1044,14 +1047,14 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                        workqueue_enqueue(&td->io_wq, &io_u->work);
                        ret = FIO_Q_QUEUED;
 
-                       if (ddir_rw(ddir)) {
-                               td->io_issues[ddir]++;
-                               td->io_issue_bytes[ddir] += blen;
-                               td->rate_io_issue_bytes[ddir] += blen;
+                       if (ddir_rw(__ddir)) {
+                               td->io_issues[__ddir]++;
+                               td->io_issue_bytes[__ddir] += blen;
+                               td->rate_io_issue_bytes[__ddir] += blen;
                        }
 
                        if (should_check_rate(td))
-                               td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
+                               td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
 
                } else {
                        ret = io_u_submit(td, io_u);
@@ -1201,19 +1204,10 @@ static void cleanup_io_u(struct thread_data *td)
 static int init_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
-       unsigned int max_bs, min_write;
        int cl_align, i, max_units;
-       int data_xfer = 1, err;
-       char *p;
+       int err;
 
        max_units = td->o.iodepth;
-       max_bs = td_max_bs(td);
-       min_write = td->o.min_bs[DDIR_WRITE];
-       td->orig_buffer_size = (unsigned long long) max_bs
-                                       * (unsigned long long) max_units;
-
-       if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
-               data_xfer = 0;
 
        err = 0;
        err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
@@ -1225,6 +1219,70 @@ static int init_io_u(struct thread_data *td)
                return 1;
        }
 
+       cl_align = os_cache_line_size();
+
+       for (i = 0; i < max_units; i++) {
+               void *ptr;
+
+               if (td->terminate)
+                       return 1;
+
+               ptr = fio_memalign(cl_align, sizeof(*io_u));
+               if (!ptr) {
+                       log_err("fio: unable to allocate aligned memory\n");
+                       break;
+               }
+
+               io_u = ptr;
+               memset(io_u, 0, sizeof(*io_u));
+               INIT_FLIST_HEAD(&io_u->verify_list);
+               dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
+
+               io_u->index = i;
+               io_u->flags = IO_U_F_FREE;
+               io_u_qpush(&td->io_u_freelist, io_u);
+
+               /*
+                * io_u never leaves this stack, used for iteration of all
+                * io_u buffers.
+                */
+               io_u_qpush(&td->io_u_all, io_u);
+
+               if (td->io_ops->io_u_init) {
+                       int ret = td->io_ops->io_u_init(td, io_u);
+
+                       if (ret) {
+                               log_err("fio: failed to init engine data: %d\n", ret);
+                               return 1;
+                       }
+               }
+       }
+
+       init_io_u_buffers(td);
+
+       if (init_file_completion_logging(td, max_units))
+               return 1;
+
+       return 0;
+}
+
+int init_io_u_buffers(struct thread_data *td)
+{
+       struct io_u *io_u;
+       unsigned long long max_bs, min_write;
+       int i, max_units;
+       int data_xfer = 1;
+       char *p;
+
+       max_units = td->o.iodepth;
+       max_bs = td_max_bs(td);
+       min_write = td->o.min_bs[DDIR_WRITE];
+       td->orig_buffer_size = (unsigned long long) max_bs
+                                       * (unsigned long long) max_units;
+
+       if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
+               data_xfer = 0;
+
        /*
         * if we may later need to do address alignment, then add any
         * possible adjustment here so that we don't cause a buffer
@@ -1236,7 +1294,7 @@ static int init_io_u(struct thread_data *td)
                td->orig_buffer_size += page_mask + td->o.mem_align;
 
        if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
-               unsigned long bs;
+               unsigned long long bs;
 
                bs = td->orig_buffer_size + td->o.hugepage_size - 1;
                td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
@@ -1256,23 +1314,8 @@ static int init_io_u(struct thread_data *td)
        else
                p = td->orig_buffer;
 
-       cl_align = os_cache_line_size();
-
        for (i = 0; i < max_units; i++) {
-               void *ptr;
-
-               if (td->terminate)
-                       return 1;
-
-               ptr = fio_memalign(cl_align, sizeof(*io_u));
-               if (!ptr) {
-                       log_err("fio: unable to allocate aligned memory\n");
-                       break;
-               }
-
-               io_u = ptr;
-               memset(io_u, 0, sizeof(*io_u));
-               INIT_FLIST_HEAD(&io_u->verify_list);
+               io_u = td->io_u_all.io_us[i];
                dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
                if (data_xfer) {
@@ -1289,32 +1332,9 @@ static int init_io_u(struct thread_data *td)
                                fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
                        }
                }
-
-               io_u->index = i;
-               io_u->flags = IO_U_F_FREE;
-               io_u_qpush(&td->io_u_freelist, io_u);
-
-               /*
-                * io_u never leaves this stack, used for iteration of all
-                * io_u buffers.
-                */
-               io_u_qpush(&td->io_u_all, io_u);
-
-               if (td->io_ops->io_u_init) {
-                       int ret = td->io_ops->io_u_init(td, io_u);
-
-                       if (ret) {
-                               log_err("fio: failed to init engine data: %d\n", ret);
-                               return 1;
-                       }
-               }
-
                p += max_bs;
        }
 
-       if (init_file_completion_logging(td, max_units))
-               return 1;
-
        return 0;
 }
 
@@ -1531,7 +1551,7 @@ static void *thread_main(void *data)
        } else
                td->pid = gettid();
 
-       fio_local_clock_init(o->use_thread);
+       fio_local_clock_init();
 
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
@@ -1575,6 +1595,8 @@ static void *thread_main(void *data)
                goto err;
        }
 
+       td_zone_gen_index(td);
+
        /*
         * 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
@@ -1852,7 +1874,11 @@ static void *thread_main(void *data)
                         "perhaps try --debug=io option for details?\n",
                         td->o.name, td->io_ops->name);
 
+       if (td->o.serialize_overlap && td->o.io_submit_mode == IO_MODE_OFFLOAD)
+               pthread_mutex_lock(&overlap_check);
        td_set_runstate(td, TD_FINISHING);
+       if (td->o.serialize_overlap && td->o.io_submit_mode == IO_MODE_OFFLOAD)
+               pthread_mutex_unlock(&overlap_check);
 
        update_rusage_stat(td);
        td->ts.total_run_time = mtime_since_now(&td->epoch);
@@ -1888,17 +1914,9 @@ err:
        close_and_free_files(td);
        cleanup_io_u(td);
        close_ioengine(td);
-       cgroup_shutdown(td, &cgroup_mnt);
+       cgroup_shutdown(td, cgroup_mnt);
        verify_free_state(td);
-
-       if (td->zone_state_index) {
-               int i;
-
-               for (i = 0; i < DDIR_RWDIR_CNT; i++)
-                       free(td->zone_state_index[i]);
-               free(td->zone_state_index);
-               td->zone_state_index = NULL;
-       }
+       td_zone_free_index(td);
 
        if (fio_option_is_set(o, cpumask)) {
                ret = fio_cpuset_exit(&o->cpumask);
@@ -1911,6 +1929,8 @@ err:
         */
        if (o->write_iolog_file)
                write_iolog_close(td);
+       if (td->io_log_rfile)
+               fclose(td->io_log_rfile);
 
        td_set_runstate(td, TD_EXITED);
 
@@ -2199,18 +2219,22 @@ static void run_threads(struct sk_out *sk_out)
        }
 
        if (output_format & FIO_OUTPUT_NORMAL) {
-               log_info("Starting ");
+               struct buf_output out;
+
+               buf_output_init(&out);
+               __log_buf(&out, "Starting ");
                if (nr_thread)
-                       log_info("%d thread%s", nr_thread,
+                       __log_buf(&out, "%d thread%s", nr_thread,
                                                nr_thread > 1 ? "s" : "");
                if (nr_process) {
                        if (nr_thread)
-                               log_info(" and ");
-                       log_info("%d process%s", nr_process,
+                               __log_buf(&out, " and ");
+                       __log_buf(&out, "%d process%s", nr_process,
                                                nr_process > 1 ? "es" : "");
                }
-               log_info("\n");
-               log_info_flush();
+               __log_buf(&out, "\n");
+               log_info_buf(out.buf, out.buflen);
+               buf_output_free(&out);
        }
 
        todo = thread_number;
@@ -2353,7 +2377,7 @@ reap:
                        if (fio_sem_down_timeout(startup_sem, 10000)) {
                                log_err("fio: job startup hung? exiting.\n");
                                fio_terminate_threads(TERMINATE_ALL);
-                               fio_abort = 1;
+                               fio_abort = true;
                                nr_started--;
                                free(fd);
                                break;
@@ -2467,6 +2491,8 @@ int fio_backend(struct sk_out *sk_out)
        }
 
        startup_sem = fio_sem_init(FIO_SEM_LOCKED);
+       if (!sk_out)
+               is_local_backend = true;
        if (startup_sem == NULL)
                return 1;
 
@@ -2510,7 +2536,6 @@ int fio_backend(struct sk_out *sk_out)
                cgroup_kill(cgroup_list);
                sfree(cgroup_list);
        }
-       sfree(cgroup_mnt);
 
        fio_sem_remove(startup_sem);
        stat_exit();