[PATCH] More 32-bit fixes (unsigned long -> unsigned long long conversions)
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 09f40afe4d8cbfac612aeb79a03f037a55a4e40a..51cd153552596558d5d879f5136eb18eb1fc9436 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -65,6 +65,7 @@ extern unsigned long long mlock_size;
 enum {
        TD_NOT_CREATED = 0,
        TD_CREATED,
+       TD_INITIALIZED,
        TD_RUNNING,
        TD_VERIFYING,
        TD_EXITED,
@@ -76,6 +77,7 @@ enum {
 static sem_t startup_sem;
 
 #define TERMINATE_ALL          (-1)
+#define JOB_START_TIMEOUT      (5 * 1000)
 
 static void terminate_threads(int group_id)
 {
@@ -195,7 +197,7 @@ static int get_next_free_block(struct thread_data *td, unsigned long long *b)
 
 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 {
-       unsigned long block = io_u->offset / td->min_bs;
+       unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
        unsigned int blocks = 0;
 
        while (blocks < (io_u->buflen / td->min_bs)) {
@@ -286,12 +288,12 @@ static int get_next_offset(struct thread_data *td, unsigned long long *offset)
        long r;
 
        if (!td->sequential) {
-               unsigned long max_blocks = td->io_size / td->min_bs;
+               unsigned long long max_blocks = td->io_size / td->min_bs;
                int loops = 50;
 
                do {
                        lrand48_r(&td->random_state, &r);
-                       b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
+                       b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
                        rb = b + (td->file_offset / td->min_bs);
                        loops--;
                } while (!random_map_free(td, rb) && loops);
@@ -1739,23 +1741,19 @@ static void init_disk_util(struct thread_data *td)
                return;
 
        /*
-        * for md/dm, there's no queue dir. we already have the right place
+        * If there's a ../queue/ directory there, we are inside a partition.
+        * Check if that is the case and jump back. For loop/md/dm etc we
+        * are already in the right spot.
         */
-       sprintf(tmp, "%s/stat", foo);
-       if (stat(tmp, &st)) {
-               /*
-                * if this is inside a partition dir, jump back to parent
-                */
-               sprintf(tmp, "%s/queue", foo);
+       sprintf(tmp, "%s/../queue", foo);
+       if (!stat(tmp, &st)) {
+               p = dirname(foo);
+               sprintf(tmp, "%s/queue", p);
                if (stat(tmp, &st)) {
-                       p = dirname(foo);
-                       sprintf(tmp, "%s/queue", p);
-                       if (stat(tmp, &st)) {
-                               fprintf(stderr, "unknown sysfs layout\n");
-                               return;
-                       }
-                       sprintf(foo, "%s", p);
+                       fprintf(stderr, "unknown sysfs layout\n");
+                       return;
                }
+               sprintf(foo, "%s", p);
        }
 
        disk_util_add(dev, foo);
@@ -1833,15 +1831,16 @@ static void *thread_main(void *data)
                }
        }
 
+       if (init_random_state(td))
+               goto err;
+
+       td_set_runstate(td, TD_INITIALIZED);
        sem_post(&startup_sem);
        sem_wait(&td->mutex);
 
        if (!td->create_serialize && setup_file(td))
                goto err;
 
-       if (init_random_state(td))
-               goto err;
-
        gettimeofday(&td->epoch, NULL);
 
        while (td->loops--) {
@@ -1951,14 +1950,15 @@ static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
                             int ddir)
 {
        char *ddir_str[] = { "read ", "write" };
-       unsigned long min, max, bw;
+       unsigned long min, max;
+       unsigned long long bw;
        double mean, dev;
 
        if (!td->runtime[ddir])
                return;
 
        bw = td->io_bytes[ddir] / td->runtime[ddir];
-       printf("  %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
+       printf("  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
 
        if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
                printf("    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
@@ -2039,6 +2039,9 @@ static void check_str_update(struct thread_data *td)
                case TD_CREATED:
                        c = 'C';
                        break;
+               case TD_INITIALIZED:
+                       c = 'I';
+                       break;
                case TD_NOT_CREATED:
                        c = 'P';
                        break;
@@ -2110,7 +2113,8 @@ static int thread_eta(struct thread_data *td, unsigned long elapsed)
 
                if (td->timeout && eta_sec > (td->timeout - elapsed))
                        eta_sec = td->timeout - elapsed;
-       } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
+       } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
+                       || td->runstate == TD_INITIALIZED) {
                int t_eta = 0, r_eta = 0;
 
                /*
@@ -2231,28 +2235,62 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
        }
 }
 
+static void fio_unpin_memory(void *pinned)
+{
+       if (pinned) {
+               if (munlock(pinned, mlock_size) < 0)
+                       perror("munlock");
+               munmap(pinned, mlock_size);
+       }
+}
+
+static void *fio_pin_memory(void)
+{
+       long pagesize, pages;
+       void *ptr;
+
+       if (!mlock_size)
+               return NULL;
+
+       /*
+        * Don't allow mlock of more than real_mem-128MB
+        */
+       pagesize = sysconf(_SC_PAGESIZE);
+       pages = sysconf(_SC_PHYS_PAGES);
+       if (pages != -1 && pagesize != -1) {
+               unsigned long long real_mem = pages * pagesize;
+
+               if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
+                       mlock_size = real_mem - 128 * 1024 * 1024;
+                       printf("fio: limiting mlocked memory to %lluMiB\n",
+                                                       mlock_size >> 20);
+               }
+       }
+
+       ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+       if (!ptr) {
+               perror("malloc locked mem");
+               return NULL;
+       }
+       if (mlock(ptr, mlock_size) < 0) {
+               munmap(ptr, mlock_size);
+               perror("mlock");
+               return NULL;
+       }
+
+       return ptr;
+}
+
 static void run_threads(void)
 {
        struct thread_data *td;
        unsigned long spent;
        int i, todo, nr_running, m_rate, t_rate, nr_started;
-       void *mlocked_mem = NULL;
+       void *mlocked_mem;
 
-       printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
-
-       if (mlock_size) {
-               mlocked_mem = malloc(mlock_size);
-               if (!mlocked_mem) {
-                       perror("malloc locked mem");
-                       return;
-               }
-               if (mlock(mlocked_mem, mlock_size) < 0) {
-                       free(mlocked_mem);
-                       perror("mlock");
-                       return;
-               }
-       }
+       mlocked_mem = fio_pin_memory();
 
+       printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
        fflush(stdout);
 
        signal(SIGINT, sig_handler);
@@ -2287,6 +2325,10 @@ static void run_threads(void)
        gettimeofday(&genesis, NULL);
 
        while (todo) {
+               struct thread_data *map[MAX_JOBS];
+               struct timeval this_start;
+               int this_jobs = 0, left;
+
                /*
                 * create threads (TD_NOT_CREATED -> TD_CREATED)
                 */
@@ -2315,9 +2357,13 @@ static void run_threads(void)
                        if (td->stonewall && (nr_started || nr_running))
                                break;
 
+                       /*
+                        * Set state to created. Thread will transition
+                        * to TD_INITIALIZED when it's done setting up.
+                        */
                        td_set_runstate(td, TD_CREATED);
+                       map[this_jobs++] = td;
                        sem_init(&startup_sem, 0, 1);
-                       todo--;
                        nr_started++;
 
                        if (td->use_thread) {
@@ -2336,12 +2382,50 @@ static void run_threads(void)
                }
 
                /*
-                * start created threads (TD_CREATED -> TD_RUNNING)
+                * Wait for the started threads to transition to
+                * TD_INITIALIZED.
                 */
+               printf("fio: Waiting for threads to initialize...\n");
+               gettimeofday(&this_start, NULL);
+               left = this_jobs;
+               while (left) {
+                       if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
+                               break;
+
+                       usleep(100000);
+
+                       for (i = 0; i < this_jobs; i++) {
+                               td = map[i];
+                               if (!td)
+                                       continue;
+                               if (td->runstate == TD_INITIALIZED ||
+                                   td->runstate >= TD_EXITED) {
+                                       map[i] = NULL;
+                                       left--;
+                                       continue;
+                               }
+                       }
+               }
+
+               if (left) {
+                       fprintf(stderr, "fio: %d jobs failed to start\n", left);
+                       for (i = 0; i < this_jobs; i++) {
+                               td = map[i];
+                               if (!td)
+                                       continue;
+                               kill(td->pid, SIGTERM);
+                       }
+                       break;
+               }
+
+               /*
+                * start created threads (TD_INITIALIZED -> TD_RUNNING)
+                */
+               printf("fio: Go for launch\n");
                for (i = 0; i < thread_number; i++) {
                        td = &threads[i];
 
-                       if (td->runstate != TD_CREATED)
+                       if (td->runstate != TD_INITIALIZED)
                                continue;
 
                        td_set_runstate(td, TD_RUNNING);
@@ -2349,6 +2433,7 @@ static void run_threads(void)
                        nr_started--;
                        m_rate += td->ratemin;
                        t_rate += td->rate;
+                       todo--;
                        sem_post(&td->mutex);
                }
 
@@ -2364,12 +2449,7 @@ static void run_threads(void)
        }
 
        update_io_ticks();
-
-       if (mlocked_mem) {
-               if (munlock(mlocked_mem, mlock_size) < 0)
-                       perror("munlock");
-               free(mlocked_mem);
-       }
+       fio_unpin_memory(mlocked_mem);
 }
 
 static void show_group_stats(struct group_run_stats *rs, int id)
@@ -2420,7 +2500,7 @@ static void show_run_stats(void)
        }
 
        for (i = 0; i < thread_number; i++) {
-               unsigned long rbw, wbw;
+               unsigned long long rbw, wbw;
 
                td = &threads[i];
 
@@ -2442,9 +2522,9 @@ static void show_run_stats(void)
 
                rbw = wbw = 0;
                if (td->runtime[0])
-                       rbw = td->io_bytes[0] / td->runtime[0];
+                       rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
                if (td->runtime[1])
-                       wbw = td->io_bytes[1] / td->runtime[1];
+                       wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
 
                if (rbw < rs->min_bw[0])
                        rs->min_bw[0] = rbw;