[PATCH] SG IO engine: reduce allocations and memory leaks
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index a175be7a1f6bc07c13107d5f8be9152e14c67aea..0b0a5e8e540851e032e0004e23175b7cfcd798e3 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -37,9 +37,9 @@
 #include "fio.h"
 #include "os.h"
 
-#define MASK   (4095)
-
-#define ALIGN(buf)     (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
+static unsigned long page_mask;
+#define ALIGN(buf)     \
+       (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
 int thread_number = 0;
@@ -47,11 +47,19 @@ int shm_id = 0;
 int temp_stall_ts;
 
 static volatile int startup_sem;
+static volatile int fio_abort;
+
+struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
-static void terminate_threads(int group_id)
+static inline void td_set_runstate(struct thread_data *td, int runstate)
+{
+       td->runstate = runstate;
+}
+
+static void terminate_threads(int group_id, int forced_kill)
 {
        struct thread_data *td;
        int i;
@@ -60,6 +68,8 @@ static void terminate_threads(int group_id)
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        td->terminate = 1;
                        td->start_delay = 0;
+                       if (forced_kill)
+                               td_set_runstate(td, TD_EXITED);
                }
        }
 }
@@ -73,9 +83,9 @@ static void sig_handler(int sig)
                        print_thread_status();
                        break;
                default:
-                       printf("\nfio: terminating on signal\n");
+                       printf("\nfio: terminating on signal %d\n", sig);
                        fflush(stdout);
-                       terminate_threads(TERMINATE_ALL);
+                       terminate_threads(TERMINATE_ALL, 0);
                        break;
        }
 }
@@ -105,7 +115,7 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 
                rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
                if (rate < td->ratemin) {
-                       fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
+                       fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
                        return 1;
                }
        }
@@ -125,11 +135,6 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-       td->runstate = runstate;
-}
-
 static struct fio_file *get_next_file(struct thread_data *td)
 {
        unsigned int old_next_file = td->next_file;
@@ -242,7 +247,7 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
  * The main verify engine. Runs over the writes we previusly submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
-void do_verify(struct thread_data *td)
+static void do_verify(struct thread_data *td)
 {
        struct io_u *io_u, *v_io_u = NULL;
        struct io_completion_data icd;
@@ -379,7 +384,6 @@ 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 ts = { .tv_sec = 0, .tv_nsec = 0};
                struct timespec *timeout;
                int min_evts = 0;
                struct io_u *io_u;
@@ -397,16 +401,29 @@ static void do_io(struct thread_data *td)
 
                memcpy(&s, &io_u->start_time, sizeof(s));
 
+requeue:
                ret = td_io_queue(td, io_u);
                if (ret) {
-                       td_verror(td, io_u->error);
-                       put_io_u(td, io_u);
-                       break;
+                       if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
+                           io_u->resid) {
+                               /*
+                                * short read/write. requeue.
+                                */
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += ret;
+                               goto requeue;
+                       } else {
+                               td_verror(td, io_u->error);
+                               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};
+
                        timeout = &ts;
                        min_evts = 0;
                } else {
@@ -439,16 +456,21 @@ static void do_io(struct thread_data *td)
 
                if (check_min_rate(td, &icd.time)) {
                        if (exitall_on_terminate)
-                               terminate_threads(td->groupid);
-                       td_verror(td, ENOMEM);
+                               terminate_threads(td->groupid, 0);
+                       td_verror(td, ENODATA);
                        break;
                }
 
                if (runtime_exceeded(td, &icd.time))
                        break;
 
-               if (td->thinktime)
-                       usec_sleep(td, td->thinktime);
+               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 (!td->error) {
@@ -507,7 +529,12 @@ static int init_io_u(struct thread_data *td)
                max_units = td->iodepth;
 
        max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
-       td->orig_buffer_size = max_bs * max_units + MASK;
+       td->orig_buffer_size = max_bs * max_units;
+
+       if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
+               td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
+       else
+               td->orig_buffer_size += page_mask;
 
        if (allocate_io_mem(td))
                return 1;
@@ -605,6 +632,7 @@ static void clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
+       unsigned long long runtime[2];
        struct thread_data *td = data;
 
        if (!td->use_thread)
@@ -625,9 +653,6 @@ static void *thread_main(void *data)
                goto err;
        }
 
-       if (td_io_init(td))
-               goto err;
-
        if (init_iolog(td))
                goto err;
 
@@ -658,13 +683,23 @@ static void *thread_main(void *data)
        if (open_files(td))
                goto err;
 
-       fio_gettime(&td->epoch, NULL);
+       /*
+        * Do this late, as some IO engines would like to have the
+        * files setup prior to initializing structures.
+        */
+       if (td_io_init(td))
+               goto err;
+
+       if (td->exec_prerun) {
+               if (system(td->exec_prerun) < 0)
+                       goto err;
+       }
 
-       if (td->exec_prerun)
-               system(td->exec_prerun);
+       fio_gettime(&td->epoch, NULL);
+       getrusage(RUSAGE_SELF, &td->ru_start);
 
+       runtime[0] = runtime[1] = 0;
        while (td->loops--) {
-               getrusage(RUSAGE_SELF, &td->ru_start);
                fio_gettime(&td->start, NULL);
                memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 
@@ -679,11 +714,9 @@ static void *thread_main(void *data)
                else
                        do_io(td);
 
-               td->runtime[td->ddir] += mtime_since_now(&td->start);
+               runtime[td->ddir] += utime_since_now(&td->start);
                if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
-                       td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
-
-               update_rusage_stat(td);
+                       runtime[td->ddir ^ 1] = runtime[td->ddir];
 
                if (td->error || td->terminate)
                        break;
@@ -696,12 +729,17 @@ static void *thread_main(void *data)
 
                do_verify(td);
 
-               td->runtime[DDIR_READ] += mtime_since_now(&td->start);
+               runtime[DDIR_READ] += utime_since_now(&td->start);
 
                if (td->error || td->terminate)
                        break;
        }
 
+       update_rusage_stat(td);
+       fio_gettime(&td->end_time, NULL);
+       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)
@@ -710,11 +748,13 @@ static void *thread_main(void *data)
                finish_log(td, td->clat_log, "clat");
        if (td->write_iolog_file)
                write_iolog_close(td);
-       if (td->exec_postrun)
-               system(td->exec_postrun);
+       if (td->exec_postrun) {
+               if (system(td->exec_postrun) < 0)
+                       log_err("fio: postrun %s failed\n", td->exec_postrun);
+       }
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid);
+               terminate_threads(td->groupid, 0);
 
 err:
        close_files(td);
@@ -789,7 +829,7 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
        }
 
        if (*nr_running == cputhreads && !pending)
-               terminate_threads(TERMINATE_ALL);
+               terminate_threads(TERMINATE_ALL, 0);
 }
 
 /*
@@ -820,10 +860,10 @@ static void run_threads(void)
        for_each_td(td, i) {
                print_status_init(td->thread_number - 1);
 
-               init_disk_util(td);
-
-               if (!td->create_serialize)
+               if (!td->create_serialize) {
+                       init_disk_util(td);
                        continue;
+               }
 
                /*
                 * do file setup here so it happens sequentially,
@@ -834,6 +874,8 @@ static void run_threads(void)
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
+
+               init_disk_util(td);
        }
 
        while (todo) {
@@ -897,7 +939,7 @@ static void run_threads(void)
                 */
                fio_gettime(&this_start, NULL);
                left = this_jobs;
-               while (left) {
+               while (left && !fio_abort) {
                        if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
                                break;
 
@@ -963,6 +1005,8 @@ static void run_threads(void)
 
 int main(int argc, char *argv[])
 {
+       long ps;
+
        if (parse_options(argc, argv))
                return 1;
 
@@ -971,10 +1015,30 @@ int main(int argc, char *argv[])
                return 1;
        }
 
+       ps = sysconf(_SC_PAGESIZE);
+       if (ps < 0) {
+               log_err("Failed to get page size\n");
+               return 1;
+       }
+
+       page_mask = ps - 1;
+
+       if (write_bw_log) {
+               setup_log(&agg_io_log[DDIR_READ]);
+               setup_log(&agg_io_log[DDIR_WRITE]);
+       }
+
        disk_util_timer_arm();
 
        run_threads();
-       show_run_stats();
+
+       if (!fio_abort) {
+               show_run_stats();
+               if (write_bw_log) {
+                       __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
+                       __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
+               }
+       }
 
        return 0;
 }