[PATCH] Aggregate bandwidth off-by 1.024
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 51cd153552596558d5d879f5136eb18eb1fc9436..d9f92b094636d6c652d08e4e907cbefc9b37f61a 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -545,44 +545,32 @@ static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
        memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
 }
 
-unsigned int hweight32(unsigned int w)
-{
-       unsigned int res = w - ((w >> 1) & 0x55555555);
-
-       res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
-       res = (res + (res >> 4)) & 0x0F0F0F0F;
-       res = res + (res >> 8);
-
-       return (res + (res >> 16)) & 0x000000FF;
-}
-
-unsigned long hweight64(unsigned long long w)
+static int get_rw_ddir(struct thread_data *td)
 {
-#if __WORDSIZE == 32
-        return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w);
-#elif __WORDSIZE == 64
-       unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul);
+       if (td_rw(td)) {
+               struct timeval now;
+               unsigned long elapsed;
 
-       v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul);
-       v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful;
-       v = v + (v >> 8);
-       v = v + (v >> 16);
+               gettimeofday(&now, NULL);
+               elapsed = mtime_since_now(&td->rwmix_switch);
 
-       return (v + (v >> 32)) & 0x00000000000000FFul;
-#else
-#error __WORDSIZE not defined
-#endif
-}
+               /*
+                * Check if it's time to seed a new data direction.
+                */
+               if (elapsed >= td->rwmixcycle) {
+                       unsigned long v;
+                       long r;
 
-static int get_rw_ddir(struct thread_data *td)
-{
-       /*
-        * perhaps cheasy, but use the hamming weight of the position
-        * as a randomizer for data direction.
-        */
-       if (td_rw(td))
-               return hweight64(td->last_pos) & 1;
-       else if (td_read(td))
+                       lrand48_r(&td->random_state, &r);
+                       v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
+                       if (v < td->rwmixread)
+                               td->rwmix_ddir = DDIR_READ;
+                       else
+                               td->rwmix_ddir = DDIR_WRITE;
+                       memcpy(&td->rwmix_switch, &now, sizeof(now));
+               }
+               return td->rwmix_ddir;
+       } else if (td_read(td))
                return DDIR_READ;
        else
                return DDIR_WRITE;
@@ -628,17 +616,16 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
        td->cur_depth--;
 }
 
-static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+static void write_iolog_put(struct thread_data *td, struct io_u *io_u)
 {
-       /*
-        * If using an iolog, grab next piece if any available.
-        */
-       if (td->iolog) {
-               struct io_piece *ipo;
+       fprintf(td->iolog_f, "%d,%llu,%u\n", io_u->ddir, io_u->offset, io_u->buflen);
+}
 
-               if (list_empty(&td->io_log_list))
-                       return 1;
+static int read_iolog_get(struct thread_data *td, struct io_u *io_u)
+{
+       struct io_piece *ipo;
 
+       if (!list_empty(&td->io_log_list)) {
                ipo = list_entry(td->io_log_list.next, struct io_piece, list);
                list_del(&ipo->list);
                io_u->offset = ipo->offset;
@@ -648,6 +635,17 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
                return 0;
        }
 
+       return 1;
+}
+
+static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+{
+       /*
+        * If using an iolog, grab next piece if any available.
+        */
+       if (td->read_iolog)
+               return read_iolog_get(td, io_u);
+
        /*
         * No log, let the seq/rand engine retrieve the next position.
         */
@@ -656,6 +654,13 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 
                if (io_u->buflen) {
                        io_u->ddir = get_rw_ddir(td);
+
+                       /*
+                        * If using a write iolog, store this entry.
+                        */
+                       if (td->write_iolog)
+                               write_iolog_put(td, io_u);
+
                        return 0;
                }
        }
@@ -707,7 +712,7 @@ static struct io_u *get_io_u(struct thread_data *td)
                return NULL;
        }
 
-       if (!td->iolog && !td->sequential)
+       if (!td->read_iolog && !td->sequential)
                mark_random_map(td, io_u);
 
        td->last_pos += io_u->buflen;
@@ -795,6 +800,13 @@ static void log_io_piece(struct thread_data *td, struct io_u *io_u)
        list_add(&ipo->list, entry);
 }
 
+static void write_iolog_close(struct thread_data *td)
+{
+       fflush(td->iolog_f);
+       fclose(td->iolog_f);
+       free(td->iolog_buf);
+}
+
 static int init_iolog(struct thread_data *td)
 {
        unsigned long long offset;
@@ -803,15 +815,34 @@ static int init_iolog(struct thread_data *td)
        FILE *f;
        int rw, i, reads, writes;
 
-       if (!td->iolog)
+       if (!td->read_iolog && !td->write_iolog)
                return 0;
 
-       f = fopen(td->iolog_file, "r");
+       if (td->read_iolog)
+               f = fopen(td->iolog_file, "r");
+       else
+               f = fopen(td->iolog_file, "w");
+
        if (!f) {
                perror("fopen iolog");
+               printf("file %s, %d/%d\n", td->iolog_file, td->read_iolog, td->write_iolog);
                return 1;
        }
 
+       /*
+        * That's it for writing, setup a log buffer and we're done.
+         */
+       if (td->write_iolog) {
+               td->iolog_f = f;
+               td->iolog_buf = malloc(8192);
+               setvbuf(f, td->iolog_buf, _IOFBF, 8192);
+               return 0;
+       }
+
+       /*
+        * Read in the read iolog and store it, reuse the infrastructure
+        * for doing verifications.
+        */
        str = malloc(4096);
        reads = writes = i = 0;
        while ((p = fgets(str, 4096, f)) != NULL) {
@@ -1251,6 +1282,22 @@ static int init_io_u(struct thread_data *td)
        return 0;
 }
 
+static void cleanup_allocs(struct thread_data *td)
+{
+       if (td->directory)
+               free(td->directory);
+       if (td->iolog_file)
+               free(td->iolog_file);
+       if (td->exec_prerun)
+               free(td->exec_prerun);
+       if (td->exec_postrun)
+               free(td->exec_postrun);
+       if (td->ioscheduler)
+               free(td->ioscheduler);
+       if (td->sysfs_root)
+               free(td->sysfs_root);
+}
+
 static int create_file(struct thread_data *td, unsigned long long size,
                       int extend)
 {
@@ -1756,9 +1803,58 @@ static void init_disk_util(struct thread_data *td)
                sprintf(foo, "%s", p);
        }
 
+       td->sysfs_root = strdup(foo);
        disk_util_add(dev, foo);
 }
 
+static int switch_ioscheduler(struct thread_data *td)
+{
+       char tmp[256], tmp2[128];
+       FILE *f;
+       int ret;
+
+       sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
+
+       f = fopen(tmp, "r+");
+       if (!f) {
+               td_verror(td, errno);
+               return 1;
+       }
+
+       /*
+        * Set io scheduler.
+        */
+       ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
+       if (ferror(f) || ret != 1) {
+               td_verror(td, errno);
+               fclose(f);
+               return 1;
+       }
+
+       rewind(f);
+
+       /*
+        * Read back and check that the selected scheduler is now the default.
+        */
+       ret = fread(tmp, 1, sizeof(tmp), f);
+       if (ferror(f) || ret < 0) {
+               td_verror(td, errno);
+               fclose(f);
+               return 1;
+       }
+
+       sprintf(tmp2, "[%s]", td->ioscheduler);
+       if (!strstr(tmp, tmp2)) {
+               fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
+               td_verror(td, EINVAL);
+               fclose(f);
+               return 1;
+       }
+
+       fclose(f);
+       return 0;
+}
+
 static void disk_util_timer_arm(void)
 {
        itimer.it_value.tv_sec = 0;
@@ -1798,7 +1894,6 @@ static void update_rusage_stat(struct thread_data *td)
 static void *thread_main(void *data)
 {
        struct thread_data *td = data;
-       int ret = 1;
 
        if (!td->use_thread)
                setsid();
@@ -1831,9 +1926,17 @@ static void *thread_main(void *data)
                }
        }
 
+       if (nice(td->nice) < 0) {
+               td_verror(td, errno);
+               goto err;
+       }
+
        if (init_random_state(td))
                goto err;
 
+       if (td->ioscheduler && switch_ioscheduler(td))
+               goto err;
+
        td_set_runstate(td, TD_INITIALIZED);
        sem_post(&startup_sem);
        sem_wait(&td->mutex);
@@ -1843,6 +1946,9 @@ static void *thread_main(void *data)
 
        gettimeofday(&td->epoch, NULL);
 
+       if (td->exec_prerun)
+               system(td->exec_prerun);
+
        while (td->loops--) {
                getrusage(RUSAGE_SELF, &td->ru_start);
                gettimeofday(&td->start, NULL);
@@ -1879,14 +1985,16 @@ static void *thread_main(void *data)
                        break;
        }
 
-       ret = 0;
-
        if (td->bw_log)
                finish_log(td, td->bw_log, "bw");
        if (td->slat_log)
                finish_log(td, td->slat_log, "slat");
        if (td->clat_log)
                finish_log(td, td->clat_log, "clat");
+       if (td->write_iolog)
+               write_iolog_close(td);
+       if (td->exec_postrun)
+               system(td->exec_postrun);
 
        if (exitall_on_terminate)
                terminate_threads(td->groupid);
@@ -1898,12 +2006,9 @@ err:
        }
        if (td->mmap)
                munmap(td->mmap, td->file_size);
+       cleanup_allocs(td);
        cleanup_io(td);
        cleanup_io_u(td);
-       if (ret) {
-               sem_post(&startup_sem);
-               sem_wait(&td->mutex);
-       }
        td_set_runstate(td, TD_EXITED);
        return NULL;
 
@@ -2398,11 +2503,14 @@ static void run_threads(void)
                                td = map[i];
                                if (!td)
                                        continue;
-                               if (td->runstate == TD_INITIALIZED ||
-                                   td->runstate >= TD_EXITED) {
+                               if (td->runstate == TD_INITIALIZED) {
                                        map[i] = NULL;
                                        left--;
-                                       continue;
+                               } else if (td->runstate >= TD_EXITED) {
+                                       map[i] = NULL;
+                                       left--;
+                                       todo--;
+                                       nr_running++; /* work-around... */
                                }
                        }
                }
@@ -2419,7 +2527,7 @@ static void run_threads(void)
                }
 
                /*
-                * start created threads (TD_INITIALIZED -> TD_RUNNING)
+                * start created threads (TD_INITIALIZED -> TD_RUNNING).
                 */
                printf("fio: Go for launch\n");
                for (i = 0; i < thread_number; i++) {
@@ -2543,9 +2651,9 @@ static void show_run_stats(void)
                rs = &runstats[i];
 
                if (rs->max_run[0])
-                       rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
+                       rs->agg[0] = (rs->io_mb[0]*1024*1024) / rs->max_run[0];
                if (rs->max_run[1])
-                       rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
+                       rs->agg[1] = (rs->io_mb[1]*1024*1024) / rs->max_run[1];
        }
 
        /*