Track io_u state (free or in-flight)
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 445babd..0ccd5d8 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -4,10 +4,12 @@
  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
  *
  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
  *
+ * The license below covers all files distributed with fio unless otherwise
+ * noted in the file itself.
+ *
  *  This program is free software; you can redistribute it and/or modify
  *  This program is free software; you can redistribute it and/or modify
- *  it under the terms of the GNU General Public License as published by
- *  the Free Software Foundation; either version 2 of the License, or
- *  (at your option) any later version.
+ *  it under the terms of the GNU General Public License version 2 as
+ *  published by the Free Software Foundation.
  *
  *  This program is distributed in the hope that it will be useful,
  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  *
  *  This program is distributed in the hope that it will be useful,
  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 #include <string.h>
 #include <signal.h>
 #include <time.h>
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <locale.h>
 #include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
 #include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
-#include <sys/ioctl.h>
 #include <sys/mman.h>
 
 #include "fio.h"
 #include "os.h"
 
 #include <sys/mman.h>
 
 #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;
 
 int groupid = 0;
 int thread_number = 0;
-static char run_str[MAX_JOBS + 1];
 int shm_id = 0;
 int shm_id = 0;
-static struct timeval genesis;
-static int temp_stall_ts;
-
-static void print_thread_status(void);
-
-extern unsigned long long mlock_size;
-
-/*
- * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
- * will never back again. It may cycle between running/verififying/fsyncing.
- * Once the thread reaches TD_EXITED, it is just waiting for the core to
- * reap it.
- */
-enum {
-       TD_NOT_CREATED = 0,
-       TD_CREATED,
-       TD_INITIALIZED,
-       TD_RUNNING,
-       TD_VERIFYING,
-       TD_FSYNCING,
-       TD_EXITED,
-       TD_REAPED,
-};
-
-#define should_fsync(td)       ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
+int temp_stall_ts;
 
 static volatile int startup_sem;
 
 static volatile int startup_sem;
+static volatile int fio_abort;
+static int exit_value;
+
+struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
 
 #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)
 {
 {
-       int i;
+       td->runstate = runstate;
+}
 
 
-       for (i = 0; i < thread_number; i++) {
-               struct thread_data *td = &threads[i];
+static void terminate_threads(int group_id, int forced_kill)
+{
+       struct thread_data *td;
+       int i;
 
 
+       for_each_td(td, i) {
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        td->terminate = 1;
                        td->start_delay = 0;
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        td->terminate = 1;
                        td->start_delay = 0;
+                       if (forced_kill)
+                               td_set_runstate(td, TD_EXITED);
                }
        }
 }
                }
        }
 }
@@ -97,127 +84,11 @@ static void sig_handler(int sig)
                        print_thread_status();
                        break;
                default:
                        print_thread_status();
                        break;
                default:
-                       printf("\nfio: terminating on signal\n");
+                       printf("\nfio: terminating on signal %d\n", sig);
                        fflush(stdout);
                        fflush(stdout);
-                       terminate_threads(TERMINATE_ALL);
-                       break;
-       }
-}
-
-/*
- * The ->file_map[] contains a map of blocks we have or have not done io
- * to yet. Used to make sure we cover the entire range in a fair fashion.
- */
-static int random_map_free(struct thread_data *td, unsigned long long block)
-{
-       unsigned int idx = RAND_MAP_IDX(td, block);
-       unsigned int bit = RAND_MAP_BIT(td, block);
-
-       return (td->file_map[idx] & (1UL << bit)) == 0;
-}
-
-/*
- * Return the next free block in the map.
- */
-static int get_next_free_block(struct thread_data *td, unsigned long long *b)
-{
-       int i;
-
-       *b = 0;
-       i = 0;
-       while ((*b) * td->min_bs < td->io_size) {
-               if (td->file_map[i] != -1UL) {
-                       *b += ffz(td->file_map[i]);
-                       return 0;
-               }
-
-               *b += BLOCKS_PER_MAP;
-               i++;
-       }
-
-       return 1;
-}
-
-/*
- * Mark a given offset as used in the map.
- */
-static void mark_random_map(struct thread_data *td, struct io_u *io_u)
-{
-       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)) {
-               unsigned int idx, bit;
-
-               if (!random_map_free(td, block))
+                       terminate_threads(TERMINATE_ALL, 0);
                        break;
                        break;
-
-               idx = RAND_MAP_IDX(td, block);
-               bit = RAND_MAP_BIT(td, block);
-
-               assert(idx < td->num_maps);
-
-               td->file_map[idx] |= (1UL << bit);
-               block++;
-               blocks++;
        }
        }
-
-       if ((blocks * td->min_bs) < io_u->buflen)
-               io_u->buflen = blocks * td->min_bs;
-}
-
-/*
- * For random io, generate a random new block and see if it's used. Repeat
- * until we find a free one. For sequential io, just return the end of
- * the last io issued.
- */
-static int get_next_offset(struct thread_data *td, unsigned long long *offset)
-{
-       unsigned long long b, rb;
-       long r;
-
-       if (!td->sequential) {
-               unsigned long long max_blocks = td->io_size / td->min_bs;
-               int loops = 50;
-
-               do {
-                       r = os_random_long(&td->random_state);
-                       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);
-
-               if (!loops) {
-                       if (get_next_free_block(td, &b))
-                               return 1;
-               }
-       } else
-               b = td->last_pos / td->min_bs;
-
-       *offset = (b * td->min_bs) + td->file_offset;
-       if (*offset > td->real_file_size)
-               return 1;
-
-       return 0;
-}
-
-static unsigned int get_next_buflen(struct thread_data *td)
-{
-       unsigned int buflen;
-       long r;
-
-       if (td->min_bs == td->max_bs)
-               buflen = td->min_bs;
-       else {
-               r = os_random_long(&td->bsrange_state);
-               buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
-               buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
-       }
-
-       if (buflen > td->io_size - td->this_io_bytes[td->ddir])
-               buflen = td->io_size - td->this_io_bytes[td->ddir];
-
-       return buflen;
 }
 
 /*
 }
 
 /*
@@ -245,9 +116,7 @@ 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) {
 
                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);
-                       if (rate_quit)
-                               terminate_threads(td->groupid);
+                       fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
                        return 1;
                }
        }
                        return 1;
                }
        }
@@ -267,509 +136,194 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
        return 0;
 }
 
-static void fill_random_bytes(struct thread_data *td,
-                             unsigned char *p, unsigned int len)
-{
-       unsigned int todo;
-       double r;
-
-       while (len) {
-               r = os_random_double(&td->verify_state);
-
-               /*
-                * lrand48_r seems to be broken and only fill the bottom
-                * 32-bits, even on 64-bit archs with 64-bit longs
-                */
-               todo = sizeof(r);
-               if (todo > len)
-                       todo = len;
-
-               memcpy(p, &r, todo);
-
-               len -= todo;
-               p += todo;
-       }
-}
-
-static void hexdump(void *buffer, int len)
-{
-       unsigned char *p = buffer;
-       int i;
-
-       for (i = 0; i < len; i++)
-               fprintf(f_out, "%02x", p[i]);
-       fprintf(f_out, "\n");
-}
-
-static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
-{
-       unsigned char *p = (unsigned char *) io_u->buf;
-       unsigned long c;
-
-       p += sizeof(*hdr);
-       c = crc32(p, hdr->len - sizeof(*hdr));
-
-       if (c != hdr->crc32) {
-               log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
-               log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c);
-               return 1;
-       }
-
-       return 0;
-}
-
-static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
-{
-       unsigned char *p = (unsigned char *) io_u->buf;
-       struct md5_ctx md5_ctx;
-
-       memset(&md5_ctx, 0, sizeof(md5_ctx));
-       p += sizeof(*hdr);
-       md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
-
-       if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
-               log_err("md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
-               hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
-               hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
-               return 1;
-       }
-
-       return 0;
-}
-
-static int verify_io_u(struct io_u *io_u)
-{
-       struct verify_header *hdr = (struct verify_header *) io_u->buf;
-       int ret;
-
-       if (hdr->fio_magic != FIO_HDR_MAGIC)
-               return 1;
-
-       if (hdr->verify_type == VERIFY_MD5)
-               ret = verify_io_u_md5(hdr, io_u);
-       else if (hdr->verify_type == VERIFY_CRC32)
-               ret = verify_io_u_crc32(hdr, io_u);
-       else {
-               log_err("Bad verify type %d\n", hdr->verify_type);
-               ret = 1;
-       }
-
-       return ret;
-}
-
-static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
-{
-       hdr->crc32 = crc32(p, len);
-}
-
-static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
-{
-       struct md5_ctx md5_ctx;
-
-       memset(&md5_ctx, 0, sizeof(md5_ctx));
-       md5_update(&md5_ctx, p, len);
-       memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
-}
-
-/*
- * Return the data direction for the next io_u. If the job is a
- * mixed read/write workload, check the rwmix cycle and switch if
- * necessary.
- */
-static int get_rw_ddir(struct thread_data *td)
-{
-       if (td_rw(td)) {
-               struct timeval now;
-               unsigned long elapsed;
-
-               gettimeofday(&now, NULL);
-               elapsed = mtime_since_now(&td->rwmix_switch);
-
-               /*
-                * Check if it's time to seed a new data direction.
-                */
-               if (elapsed >= td->rwmixcycle) {
-                       int v;
-                       long r;
-
-                       r = os_random_long(&td->rwmix_state);
-                       v = 1 + (int) (100.0 * (r / (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;
-}
-
 /*
 /*
- * fill body of io_u->buf with random data and add a header with the
- * crc32 or md5 sum of that data.
+ * When job exits, we can cancel the in-flight IO if we are using async
+ * io. Attempt to do so.
  */
  */
-static void populate_io_u(struct thread_data *td, struct io_u *io_u)
-{
-       unsigned char *p = (unsigned char *) io_u->buf;
-       struct verify_header hdr;
-
-       hdr.fio_magic = FIO_HDR_MAGIC;
-       hdr.len = io_u->buflen;
-       p += sizeof(hdr);
-       fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
-
-       if (td->verify == VERIFY_MD5) {
-               fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
-               hdr.verify_type = VERIFY_MD5;
-       } else {
-               fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
-               hdr.verify_type = VERIFY_CRC32;
-       }
-
-       memcpy(io_u->buf, &hdr, sizeof(hdr));
-}
-
-static int td_io_prep(struct thread_data *td, struct io_u *io_u)
-{
-       if (td->io_prep && td->io_prep(td, io_u))
-               return 1;
-
-       return 0;
-}
-
-void put_io_u(struct thread_data *td, struct io_u *io_u)
+static void cleanup_pending_aio(struct thread_data *td)
 {
 {
-       list_del(&io_u->list);
-       list_add(&io_u->list, &td->io_u_freelist);
-       td->cur_depth--;
-}
+       struct list_head *entry, *n;
+       struct io_u *io_u;
+       int r;
 
 
-static int fill_io_u(struct thread_data *td, struct io_u *io_u)
-{
        /*
        /*
-        * If using an iolog, grab next piece if any available.
+        * get immediately available events, if any
         */
         */
-       if (td->read_iolog)
-               return read_iolog_get(td, io_u);
+       r = io_u_queued_complete(td, 0, NULL);
+       if (r < 0)
+               return;
 
        /*
 
        /*
-        * No log, let the seq/rand engine retrieve the next position.
+        * now cancel remaining active events
         */
         */
-       if (!get_next_offset(td, &io_u->offset)) {
-               io_u->buflen = get_next_buflen(td);
-
-               if (io_u->buflen) {
-                       io_u->ddir = get_rw_ddir(td);
+       if (td->io_ops->cancel) {
+               list_for_each_safe(entry, n, &td->io_u_busylist) {
+                       io_u = list_entry(entry, struct io_u, list);
 
                        /*
 
                        /*
-                        * If using a write iolog, store this entry.
+                        * if the io_u isn't in flight, then that generally
+                        * means someone leaked an io_u. complain but fix
+                        * it up, so we don't stall here.
                         */
                         */
-                       if (td->write_iolog)
-                               write_iolog_put(td, io_u);
-
-                       return 0;
+                       if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+                               log_err("fio: non-busy IO on busy list\n");
+                               put_io_u(td, io_u);
+                       } else {
+                               r = td->io_ops->cancel(td, io_u);
+                               if (!r)
+                                       put_io_u(td, io_u);
+                       }
                }
        }
 
                }
        }
 
-       return 1;
-}
-
-#define queue_full(td) list_empty(&(td)->io_u_freelist)
-
-struct io_u *__get_io_u(struct thread_data *td)
-{
-       struct io_u *io_u = NULL;
-
-       if (!queue_full(td)) {
-               io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
-
-               io_u->error = 0;
-               io_u->resid = 0;
-               list_del(&io_u->list);
-               list_add(&io_u->list, &td->io_u_busylist);
-               td->cur_depth++;
-       }
-
-       return io_u;
+       if (td->cur_depth)
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
 }
 
 /*
- * Return an io_u to be processed. Gets a buflen and offset, sets direction,
- * etc. The returned io_u is fully ready to be prepped and submitted.
+ * Helper to handle the final sync of a file. Works just like the normal
+ * io path, just does everything sync.
  */
  */
-static struct io_u *get_io_u(struct thread_data *td)
+static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
 {
-       struct io_u *io_u;
+       struct io_u *io_u = __get_io_u(td);
+       int ret;
 
 
-       io_u = __get_io_u(td);
        if (!io_u)
        if (!io_u)
-               return NULL;
-
-       if (td->zone_bytes >= td->zone_size) {
-               td->zone_bytes = 0;
-               td->last_pos += td->zone_skip;
-       }
-
-       if (fill_io_u(td, io_u)) {
-               put_io_u(td, io_u);
-               return NULL;
-       }
-
-       if (io_u->buflen + io_u->offset > td->real_file_size)
-               io_u->buflen = td->real_file_size - io_u->offset;
-
-       if (!io_u->buflen) {
-               put_io_u(td, io_u);
-               return NULL;
-       }
-
-       if (!td->read_iolog && !td->sequential)
-               mark_random_map(td, io_u);
-
-       td->last_pos += io_u->buflen;
+               return 1;
 
 
-       if (td->verify != VERIFY_NONE)
-               populate_io_u(td, io_u);
+       io_u->ddir = DDIR_SYNC;
+       io_u->file = f;
 
        if (td_io_prep(td, io_u)) {
                put_io_u(td, io_u);
 
        if (td_io_prep(td, io_u)) {
                put_io_u(td, io_u);
-               return NULL;
+               return 1;
        }
 
        }
 
-       gettimeofday(&io_u->start_time, NULL);
-       return io_u;
-}
-
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-       td->runstate = runstate;
-}
-
-static int get_next_verify(struct thread_data *td, struct io_u *io_u)
-{
-       struct io_piece *ipo;
-
-       if (!list_empty(&td->io_hist_list)) {
-               ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
-
-               list_del(&ipo->list);
+requeue:
+       ret = td_io_queue(td, io_u);
+       if (ret < 0) {
+               td_verror(td, io_u->error);
+               put_io_u(td, io_u);
+               return 1;
+       } else if (ret == FIO_Q_QUEUED) {
+               if (io_u_queued_complete(td, 1, NULL) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_COMPLETED) {
+               if (io_u->error) {
+                       td_verror(td, io_u->error);
+                       return 1;
+               }
 
 
-               io_u->offset = ipo->offset;
-               io_u->buflen = ipo->len;
-               io_u->ddir = DDIR_READ;
-               free(ipo);
-               return 0;
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_BUSY) {
+               if (td_io_commit(td))
+                       return 1;
+               goto requeue;
        }
 
        }
 
-       return 1;
-}
-
-static int sync_td(struct thread_data *td)
-{
-       if (td->io_sync)
-               return td->io_sync(td);
-
        return 0;
 }
 
        return 0;
 }
 
-static int io_u_getevents(struct thread_data *td, int min, int max,
-                         struct timespec *t)
-{
-       return td->io_getevents(td, min, max, t);
-}
-
-static int io_u_queue(struct thread_data *td, struct io_u *io_u)
-{
-       gettimeofday(&io_u->issue_time, NULL);
-
-       return td->io_queue(td, io_u);
-}
-
-#define iocb_time(iocb)        ((unsigned long) (iocb)->data)
-
-static void io_completed(struct thread_data *td, struct io_u *io_u,
-                        struct io_completion_data *icd)
-{
-       struct timeval e;
-       unsigned long msec;
-
-       gettimeofday(&e, NULL);
-
-       if (!io_u->error) {
-               unsigned int bytes = io_u->buflen - io_u->resid;
-               const int idx = io_u->ddir;
-
-               td->io_blocks[idx]++;
-               td->io_bytes[idx] += bytes;
-               td->zone_bytes += bytes;
-               td->this_io_bytes[idx] += bytes;
-
-               msec = mtime_since(&io_u->issue_time, &e);
-
-               add_clat_sample(td, idx, msec);
-               add_bw_sample(td, idx);
-
-               if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
-                       log_io_piece(td, io_u);
-
-               icd->bytes_done[idx] += bytes;
-       } else
-               icd->error = io_u->error;
-}
-
-static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
-{
-       struct io_u *io_u;
-       int i;
-
-       icd->error = 0;
-       icd->bytes_done[0] = icd->bytes_done[1] = 0;
-
-       for (i = 0; i < icd->nr; i++) {
-               io_u = td->io_event(td, i);
-
-               io_completed(td, io_u, icd);
-               put_io_u(td, io_u);
-       }
-}
-
 /*
 /*
- * When job exits, we can cancel the in-flight IO if we are using async
- * io. Attempt to do so.
+ * The main verify engine. Runs over the writes we previusly submitted,
+ * reads the blocks back in, and checks the crc/md5 of the data.
  */
  */
-static void cleanup_pending_aio(struct thread_data *td)
+static void do_verify(struct thread_data *td)
 {
 {
-       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
-       struct list_head *entry, *n;
-       struct io_completion_data icd;
+       struct fio_file *f;
        struct io_u *io_u;
        struct io_u *io_u;
-       int r;
-
-       /*
-        * get immediately available events, if any
-        */
-       r = io_u_getevents(td, 0, td->cur_depth, &ts);
-       if (r > 0) {
-               icd.nr = r;
-               ios_completed(td, &icd);
-       }
+       int ret, i, min_events;
 
        /*
 
        /*
-        * now cancel remaining active events
+        * sync io first and invalidate cache, to make sure we really
+        * read from disk.
         */
         */
-       if (td->io_cancel) {
-               list_for_each_safe(entry, n, &td->io_u_busylist) {
-                       io_u = list_entry(entry, struct io_u, list);
-
-                       r = td->io_cancel(td, io_u);
-                       if (!r)
-                               put_io_u(td, io_u);
-               }
-       }
-
-       if (td->cur_depth) {
-               r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
-               if (r > 0) {
-                       icd.nr = r;
-                       ios_completed(td, &icd);
-               }
-       }
-}
-
-static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
-{
-       struct io_u *v_io_u = *io_u;
-       int ret = 0;
-
-       if (v_io_u) {
-               ret = verify_io_u(v_io_u);
-               put_io_u(td, v_io_u);
-               *io_u = NULL;
+       for_each_file(td, f, i) {
+               if (fio_io_sync(td, f))
+                       break;
+               if (file_invalidate_cache(td, f))
+                       break;
        }
 
        }
 
-       return ret;
-}
-
-/*
- * The main verify engine. Runs over the writes we previusly submitted,
- * reads the blocks back in, and checks the crc/md5 of the data.
- */
-static void do_verify(struct thread_data *td)
-{
-       struct timeval t;
-       struct io_u *io_u, *v_io_u = NULL;
-       struct io_completion_data icd;
-       int ret;
+       if (td->error)
+               return;
 
        td_set_runstate(td, TD_VERIFYING);
 
 
        td_set_runstate(td, TD_VERIFYING);
 
-       do {
-               if (td->terminate)
-                       break;
-
-               gettimeofday(&t, NULL);
-               if (runtime_exceeded(td, &t))
-                       break;
-
+       io_u = NULL;
+       while (!td->terminate) {
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
-               if (get_next_verify(td, io_u)) {
+               if (runtime_exceeded(td, &io_u->start_time)) {
                        put_io_u(td, io_u);
                        break;
                }
 
                        put_io_u(td, io_u);
                        break;
                }
 
-               if (td_io_prep(td, io_u)) {
+               if (get_next_verify(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
                }
 
                        put_io_u(td, io_u);
                        break;
                }
 
-               ret = io_u_queue(td, io_u);
-               if (ret) {
+               if (td_io_prep(td, io_u)) {
                        put_io_u(td, io_u);
                        put_io_u(td, io_u);
-                       td_verror(td, ret);
                        break;
                }
                        break;
                }
-
-               /*
-                * we have one pending to verify, do that while
-                * we are doing io on the next one
-                */
-               if (do_io_u_verify(td, &v_io_u))
-                       break;
-
-               ret = io_u_getevents(td, 1, 1, NULL);
-               if (ret != 1) {
+requeue:
+               ret = td_io_queue(td, io_u);
+
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (io_u->error)
+                               ret = -io_u->error;
+                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
+
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       ret = io_u_sync_complete(td, io_u, verify_io_u);
                        if (ret < 0)
                        if (ret < 0)
-                               td_verror(td, ret);
+                               break;
+                       continue;
+               case FIO_Q_QUEUED:
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       td_verror(td, -ret);
                        break;
                }
 
                        break;
                }
 
-               v_io_u = td->io_event(td, 0);
-               icd.nr = 1;
-               icd.error = 0;
-               io_completed(td, v_io_u, &icd);
-
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       put_io_u(td, v_io_u);
-                       v_io_u = NULL;
+               if (ret < 0 || td->error)
                        break;
                        break;
-               }
 
                /*
 
                /*
-                * if we can't submit more io, we need to verify now
+                * if we can queue more, do so. but check if there are
+                * completed io_u's first.
                 */
                 */
-               if (queue_full(td) && do_io_u_verify(td, &v_io_u))
-                       break;
+               min_events = 0;
+               if (queue_full(td) || ret == FIO_Q_BUSY) {
+                       min_events = 1;
 
 
-       } while (1);
+                       if (td->cur_depth > td->iodepth_low)
+                               min_events = td->cur_depth - td->iodepth_low;
+               }
 
 
-       do_io_u_verify(td, &v_io_u);
+               /*
+                * Reap required number of io units, if any, and do the
+                * verification on them through the callback handler
+                */
+               if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
+                       break;
+       }
 
        if (td->cur_depth)
                cleanup_pending_aio(td);
 
        if (td->cur_depth)
                cleanup_pending_aio(td);
@@ -788,7 +342,7 @@ static void do_cpuio(struct thread_data *td)
        int i = 0;
 
        while (!td->terminate) {
        int i = 0;
 
        while (!td->terminate) {
-               gettimeofday(&e, NULL);
+               fio_gettime(&e, NULL);
 
                if (runtime_exceeded(td, &e))
                        break;
 
                if (runtime_exceeded(td, &e))
                        break;
@@ -808,16 +362,16 @@ static void do_cpuio(struct thread_data *td)
  */
 static void do_io(struct thread_data *td)
 {
  */
 static void do_io(struct thread_data *td)
 {
-       struct io_completion_data icd;
-       struct timeval s, e;
+       struct timeval s;
        unsigned long usec;
        unsigned long usec;
+       int i, ret = 0;
 
        td_set_runstate(td, TD_RUNNING);
 
 
        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 ret, min_evts = 0;
+       while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
+               struct timeval comp_time;
+               long bytes_done = 0;
+               int min_evts = 0;
                struct io_u *io_u;
 
                if (td->terminate)
                struct io_u *io_u;
 
                if (td->terminate)
@@ -829,463 +383,191 @@ static void do_io(struct thread_data *td)
 
                memcpy(&s, &io_u->start_time, sizeof(s));
 
 
                memcpy(&s, &io_u->start_time, sizeof(s));
 
-               ret = io_u_queue(td, io_u);
-               if (ret) {
+               if (runtime_exceeded(td, &s)) {
                        put_io_u(td, io_u);
                        put_io_u(td, io_u);
-                       td_verror(td, ret);
                        break;
                }
                        break;
                }
+requeue:
+               ret = td_io_queue(td, io_u);
 
 
-               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
-
-               if (td->cur_depth < td->iodepth) {
-                       timeout = &ts;
-                       min_evts = 0;
-               } else {
-                       timeout = NULL;
-                       min_evts = 1;
-               }
+               switch (ret) {
+               case FIO_Q_COMPLETED:
+                       if (io_u->error) {
+                               ret = io_u->error;
+                               break;
+                       }
+                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
 
 
-               ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
-               if (ret < 0) {
-                       td_verror(td, ret);
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_sync_complete(td, io_u, NULL);
+                       if (bytes_done < 0)
+                               ret = bytes_done;
                        break;
                        break;
-               } else if (!ret)
-                       continue;
-
-               icd.nr = ret;
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
+               case FIO_Q_QUEUED:
+                       /*
+                        * if the engine doesn't have a commit hook,
+                        * the io_u is really queued. if it does have such
+                        * a hook, it has to call io_u_queued() itself.
+                        */
+                       if (td->io_ops->commit == NULL)
+                               io_u_queued(td, io_u);
                        break;
                        break;
-               }
-
-               /*
-                * the rate is batched for now, it should work for batches
-                * of completions except the very first one which may look
-                * a little bursty
-                */
-               gettimeofday(&e, NULL);
-               usec = utime_since(&s, &e);
-
-               rate_throttle(td, usec, icd.bytes_done[td->ddir]);
-
-               if (check_min_rate(td, &e)) {
-                       td_verror(td, ENOMEM);
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
                        break;
                        break;
-               }
-
-               if (runtime_exceeded(td, &e))
+               default:
+                       assert(ret < 0);
+                       put_io_u(td, io_u);
                        break;
                        break;
-
-               if (td->thinktime)
-                       usec_sleep(td, td->thinktime);
-
-               if (should_fsync(td) && td->fsync_blocks &&
-                   (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
-                       sync_td(td);
-       }
-
-       if (td->cur_depth)
-               cleanup_pending_aio(td);
-
-       if (should_fsync(td) && td->end_fsync) {
-               td_set_runstate(td, TD_FSYNCING);
-               sync_td(td);
-       }
-}
-
-static void cleanup_io(struct thread_data *td)
-{
-       if (td->io_cleanup)
-               td->io_cleanup(td);
-}
-
-static int init_io(struct thread_data *td)
-{
-       if (td->io_engine == FIO_SYNCIO)
-               return fio_syncio_init(td);
-       else if (td->io_engine == FIO_MMAPIO)
-               return fio_mmapio_init(td);
-       else if (td->io_engine == FIO_LIBAIO)
-               return fio_libaio_init(td);
-       else if (td->io_engine == FIO_POSIXAIO)
-               return fio_posixaio_init(td);
-       else if (td->io_engine == FIO_SGIO)
-               return fio_sgio_init(td);
-       else if (td->io_engine == FIO_SPLICEIO)
-               return fio_spliceio_init(td);
-       else if (td->io_engine == FIO_CPUIO)
-               return fio_cpuio_init(td);
-       else {
-               log_err("bad io_engine %d\n", td->io_engine);
-               return 1;
-       }
-}
-
-static void cleanup_io_u(struct thread_data *td)
-{
-       struct list_head *entry, *n;
-       struct io_u *io_u;
-
-       list_for_each_safe(entry, n, &td->io_u_freelist) {
-               io_u = list_entry(entry, struct io_u, list);
-
-               list_del(&io_u->list);
-               free(io_u);
-       }
-
-       if (td->mem_type == MEM_MALLOC)
-               free(td->orig_buffer);
-       else if (td->mem_type == MEM_SHM) {
-               struct shmid_ds sbuf;
-
-               shmdt(td->orig_buffer);
-               shmctl(td->shm_id, IPC_RMID, &sbuf);
-       } else if (td->mem_type == MEM_MMAP)
-               munmap(td->orig_buffer, td->orig_buffer_size);
-       else
-               log_err("Bad memory type %d\n", td->mem_type);
-
-       td->orig_buffer = NULL;
-}
-
-static int init_io_u(struct thread_data *td)
-{
-       struct io_u *io_u;
-       int i, max_units;
-       char *p;
-
-       if (td->io_engine == FIO_CPUIO)
-               return 0;
-
-       if (td->io_engine & FIO_SYNCIO)
-               max_units = 1;
-       else
-               max_units = td->iodepth;
-
-       td->orig_buffer_size = td->max_bs * max_units + MASK;
-
-       if (td->mem_type == MEM_MALLOC)
-               td->orig_buffer = malloc(td->orig_buffer_size);
-       else if (td->mem_type == MEM_SHM) {
-               td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
-               if (td->shm_id < 0) {
-                       td_verror(td, errno);
-                       perror("shmget");
-                       return 1;
                }
 
                }
 
-               td->orig_buffer = shmat(td->shm_id, NULL, 0);
-               if (td->orig_buffer == (void *) -1) {
-                       td_verror(td, errno);
-                       perror("shmat");
-                       td->orig_buffer = NULL;
-                       return 1;
-               }
-       } else if (td->mem_type == MEM_MMAP) {
-               td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
-               if (td->orig_buffer == MAP_FAILED) {
-                       td_verror(td, errno);
-                       perror("mmap");
-                       td->orig_buffer = NULL;
-                       return 1;
-               }
-       }
-
-       p = ALIGN(td->orig_buffer);
-       for (i = 0; i < max_units; i++) {
-               io_u = malloc(sizeof(*io_u));
-               memset(io_u, 0, sizeof(*io_u));
-               INIT_LIST_HEAD(&io_u->list);
-
-               io_u->buf = p + td->max_bs * i;
-               io_u->index = i;
-               list_add(&io_u->list, &td->io_u_freelist);
-       }
-
-       return 0;
-}
-
-static int create_file(struct thread_data *td, unsigned long long size)
-{
-       unsigned long long left;
-       unsigned int bs;
-       char *b;
-       int r;
-
-       /*
-        * unless specifically asked for overwrite, let normal io extend it
-        */
-       if (!td->overwrite) {
-               td->real_file_size = size;
-               return 0;
-       }
-
-       if (!size) {
-               log_err("Need size for create\n");
-               td_verror(td, EINVAL);
-               return 1;
-       }
-
-       temp_stall_ts = 1;
-       fprintf(f_out, "%s: Laying out IO file (%LuMiB)\n",td->name,size >> 20);
-
-       td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
-       if (td->fd < 0) {
-               td_verror(td, errno);
-               goto done_noclose;
-       }
-
-       if (ftruncate(td->fd, td->file_size) == -1) {
-               td_verror(td, errno);
-               goto done;
-       }
-
-       td->io_size = td->file_size;
-       b = malloc(td->max_bs);
-       memset(b, 0, td->max_bs);
-
-       left = size;
-       while (left && !td->terminate) {
-               bs = td->max_bs;
-               if (bs > left)
-                       bs = left;
-
-               r = write(td->fd, b, bs);
-
-               if (r == (int) bs) {
-                       left -= bs;
-                       continue;
-               } else {
-                       if (r < 0)
-                               td_verror(td, errno);
-                       else
-                               td_verror(td, EIO);
-
+               if (ret < 0 || td->error)
                        break;
                        break;
-               }
-       }
 
 
-       if (td->terminate)
-               unlink(td->file_name);
-       else if (td->create_fsync)
-               fsync(td->fd);
-
-       free(b);
-done:
-       close(td->fd);
-       td->fd = -1;
-done_noclose:
-       temp_stall_ts = 0;
-       return 0;
-}
+               /*
+                * See if we need to complete some commands
+                */
+               if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
+                       min_evts = 0;
+                       if (queue_full(td) || ret == FIO_Q_BUSY) {
+                               min_evts = 1;
 
 
-static int file_size(struct thread_data *td)
-{
-       struct stat st;
+                               if (td->cur_depth > td->iodepth_low)
+                                       min_evts = td->cur_depth - td->iodepth_low;
+                       }
 
 
-       if (td->overwrite) {
-               if (fstat(td->fd, &st) == -1) {
-                       td_verror(td, errno);
-                       return 1;
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
+                       if (bytes_done < 0)
+                               break;
                }
 
                }
 
-               td->real_file_size = st.st_size;
-
-               if (!td->file_size || td->file_size > td->real_file_size)
-                       td->file_size = td->real_file_size;
-       }
-
-       td->file_size -= td->file_offset;
-       return 0;
-}
-
-static int bdev_size(struct thread_data *td)
-{
-       unsigned long long bytes;
-       int r;
-
-       r = blockdev_size(td->fd, &bytes);
-       if (r) {
-               td_verror(td, r);
-               return 1;
-       }
+               if (!bytes_done)
+                       continue;
 
 
-       td->real_file_size = bytes;
+               /*
+                * the rate is batched for now, it should work for batches
+                * of completions except the very first one which may look
+                * a little bursty
+                */
+               usec = utime_since(&s, &comp_time);
 
 
-       /*
-        * no extend possibilities, so limit size to device size if too large
-        */
-       if (!td->file_size || td->file_size > td->real_file_size)
-               td->file_size = td->real_file_size;
+               rate_throttle(td, usec, bytes_done, td->ddir);
 
 
-       td->file_size -= td->file_offset;
-       return 0;
-}
+               if (check_min_rate(td, &comp_time)) {
+                       if (exitall_on_terminate)
+                               terminate_threads(td->groupid, 0);
+                       td_verror(td, ENODATA);
+                       break;
+               }
 
 
-static int get_file_size(struct thread_data *td)
-{
-       int ret = 0;
+               if (td->thinktime) {
+                       unsigned long long b;
 
 
-       if (td->filetype == FIO_TYPE_FILE)
-               ret = file_size(td);
-       else if (td->filetype == FIO_TYPE_BD)
-               ret = bdev_size(td);
-       else
-               td->real_file_size = -1;
+                       b = td->io_blocks[0] + td->io_blocks[1];
+                       if (!(b % td->thinktime_blocks)) {
+                               int left;
 
 
-       if (ret)
-               return ret;
+                               if (td->thinktime_spin)
+                                       __usec_sleep(td->thinktime_spin);
 
 
-       if (td->file_offset > td->real_file_size) {
-               log_err("%s: offset extends end (%Lu > %Lu)\n", td->name, td->file_offset, td->real_file_size);
-               return 1;
+                               left = td->thinktime - td->thinktime_spin;
+                               if (left)
+                                       usec_sleep(td, left);
+                       }
+               }
        }
 
        }
 
-       td->io_size = td->file_size;
-       if (td->io_size == 0) {
-               log_err("%s: no io blocks\n", td->name);
-               td_verror(td, EINVAL);
-               return 1;
-       }
+       if (!td->error) {
+               struct fio_file *f;
 
 
-       if (!td->zone_size)
-               td->zone_size = td->io_size;
+               if (td->cur_depth)
+                       cleanup_pending_aio(td);
 
 
-       td->total_io_size = td->io_size * td->loops;
-       return 0;
+               if (should_fsync(td) && td->end_fsync) {
+                       td_set_runstate(td, TD_FSYNCING);
+                       for_each_file(td, f, i)
+                               fio_io_sync(td, f);
+               }
+       }
 }
 
 }
 
-static int setup_file_mmap(struct thread_data *td)
+static void cleanup_io_u(struct thread_data *td)
 {
 {
-       int flags;
-
-       if (td_rw(td))
-               flags = PROT_READ | PROT_WRITE;
-       else if (td_write(td)) {
-               flags = PROT_WRITE;
-
-               if (td->verify != VERIFY_NONE)
-                       flags |= PROT_READ;
-       } else
-               flags = PROT_READ;
-
-       td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
-       if (td->mmap == MAP_FAILED) {
-               td->mmap = NULL;
-               td_verror(td, errno);
-               return 1;
-       }
+       struct list_head *entry, *n;
+       struct io_u *io_u;
 
 
-       if (td->invalidate_cache) {
-               if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       }
+       list_for_each_safe(entry, n, &td->io_u_freelist) {
+               io_u = list_entry(entry, struct io_u, list);
 
 
-       if (td->sequential) {
-               if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       } else {
-               if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
+               list_del(&io_u->list);
+               free(io_u);
        }
 
        }
 
-       return 0;
+       free_io_mem(td);
 }
 
 }
 
-static int setup_file_plain(struct thread_data *td)
+/*
+ * "randomly" fill the buffer contents
+ */
+static void fill_rand_buf(struct io_u *io_u, int max_bs)
 {
 {
-       if (td->invalidate_cache) {
-               if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       }
+       int *ptr = io_u->buf;
 
 
-       if (td->sequential) {
-               if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       } else {
-               if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
+       while ((void *) ptr - io_u->buf < max_bs) {
+               *ptr = rand() * 0x9e370001;
+               ptr++;
        }
        }
-
-       return 0;
 }
 
 }
 
-static int setup_file(struct thread_data *td)
+static int init_io_u(struct thread_data *td)
 {
 {
-       struct stat st;
-       int flags = 0;
+       struct io_u *io_u;
+       unsigned int max_bs;
+       int i, max_units;
+       char *p;
 
 
-       if (td->io_engine == FIO_CPUIO)
+       if (td->io_ops->flags & FIO_CPUIO)
                return 0;
 
                return 0;
 
-       if (stat(td->file_name, &st) == -1) {
-               if (errno != ENOENT) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-               if (!td->create_file) {
-                       td_verror(td, ENOENT);
-                       return 1;
-               }
-               if (create_file(td, td->file_size))
-                       return 1;
-       } else if (td->filetype == FIO_TYPE_FILE &&
-                  st.st_size < (off_t) td->file_size) {
-               if (create_file(td, td->file_size))
-                       return 1;
-       }
-
-       if (td->odirect)
-               flags |= OS_O_DIRECT;
+       if (td->io_ops->flags & FIO_SYNCIO)
+               max_units = 1;
+       else
+               max_units = td->iodepth;
 
 
-       if (td_write(td) || td_rw(td)) {
-               if (td->filetype == FIO_TYPE_FILE) {
-                       if (!td->overwrite)
-                               flags |= O_TRUNC;
+       max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
+       td->orig_buffer_size = max_bs * max_units;
 
 
-                       flags |= O_CREAT;
-               }
-               if (td->sync_io)
-                       flags |= O_SYNC;
+       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;
 
 
-               flags |= O_RDWR;
+       if (allocate_io_mem(td))
+               return 1;
 
 
-               td->fd = open(td->file_name, flags, 0600);
-       } else {
-               if (td->filetype == FIO_TYPE_CHAR)
-                       flags |= O_RDWR;
-               else
-                       flags |= O_RDONLY;
+       p = ALIGN(td->orig_buffer);
+       for (i = 0; i < max_units; i++) {
+               io_u = malloc(sizeof(*io_u));
+               memset(io_u, 0, sizeof(*io_u));
+               INIT_LIST_HEAD(&io_u->list);
 
 
-               td->fd = open(td->file_name, flags);
-       }
+               io_u->buf = p + max_bs * i;
+               if (td_write(td) || td_rw(td))
+                       fill_rand_buf(io_u, max_bs);
 
 
-       if (td->fd == -1) {
-               td_verror(td, errno);
-               return 1;
+               io_u->index = i;
+               io_u->flags = IO_U_F_FREE;
+               list_add(&io_u->list, &td->io_u_freelist);
        }
 
        }
 
-       if (get_file_size(td))
-               return 1;
-
-       if (td->io_engine != FIO_MMAPIO)
-               return setup_file_plain(td);
-       else
-               return setup_file_mmap(td);
+       return 0;
 }
 
 static int switch_ioscheduler(struct thread_data *td)
 }
 
 static int switch_ioscheduler(struct thread_data *td)
@@ -1294,6 +576,9 @@ static int switch_ioscheduler(struct thread_data *td)
        FILE *f;
        int ret;
 
        FILE *f;
        int ret;
 
+       if (td->io_ops->flags & FIO_CPUIO)
+               return 0;
+
        sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
        f = fopen(tmp, "r+");
        sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
        f = fopen(tmp, "r+");
@@ -1338,16 +623,25 @@ static int switch_ioscheduler(struct thread_data *td)
 
 static void clear_io_state(struct thread_data *td)
 {
 
 static void clear_io_state(struct thread_data *td)
 {
-       if (td->io_engine == FIO_SYNCIO)
-               lseek(td->fd, SEEK_SET, 0);
+       struct fio_file *f;
+       int i;
 
 
-       td->last_pos = 0;
-       td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+       td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
 
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
 
-       if (td->file_map)
-               memset(td->file_map, 0, td->num_maps * sizeof(long));
+       td->last_was_sync = 0;
+
+       for_each_file(td, f, i) {
+               f->last_completed_pos = 0;
+
+               f->last_pos = 0;
+               if (td->io_ops->flags & FIO_SYNCIO)
+                       lseek(f->fd, SEEK_SET, 0);
+
+               if (f->file_map)
+                       memset(f->file_map, 0, f->num_maps * sizeof(long));
+       }
 }
 
 /*
 }
 
 /*
@@ -1356,6 +650,7 @@ static void clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
  */
 static void *thread_main(void *data)
 {
+       unsigned long long runtime[2];
        struct thread_data *td = data;
 
        if (!td->use_thread)
        struct thread_data *td = data;
 
        if (!td->use_thread)
@@ -1365,6 +660,7 @@ static void *thread_main(void *data)
 
        INIT_LIST_HEAD(&td->io_u_freelist);
        INIT_LIST_HEAD(&td->io_u_busylist);
 
        INIT_LIST_HEAD(&td->io_u_freelist);
        INIT_LIST_HEAD(&td->io_u_busylist);
+       INIT_LIST_HEAD(&td->io_u_requeues);
        INIT_LIST_HEAD(&td->io_hist_list);
        INIT_LIST_HEAD(&td->io_log_list);
 
        INIT_LIST_HEAD(&td->io_hist_list);
        INIT_LIST_HEAD(&td->io_log_list);
 
@@ -1376,9 +672,6 @@ static void *thread_main(void *data)
                goto err;
        }
 
                goto err;
        }
 
-       if (init_io(td))
-               goto err;
-
        if (init_iolog(td))
                goto err;
 
        if (init_iolog(td))
                goto err;
 
@@ -1404,35 +697,45 @@ static void *thread_main(void *data)
        fio_sem_up(&startup_sem);
        fio_sem_down(&td->mutex);
 
        fio_sem_up(&startup_sem);
        fio_sem_down(&td->mutex);
 
-       if (!td->create_serialize && setup_file(td))
+       if (!td->create_serialize && setup_files(td))
+               goto err;
+       if (open_files(td))
+               goto err;
+
+       /*
+        * 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;
 
                goto err;
 
-       gettimeofday(&td->epoch, NULL);
+       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->ts.ru_start);
 
 
+       runtime[0] = runtime[1] = 0;
        while (td->loops--) {
        while (td->loops--) {
-               getrusage(RUSAGE_SELF, &td->ru_start);
-               gettimeofday(&td->start, NULL);
-               memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+               fio_gettime(&td->start, NULL);
+               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
 
                if (td->ratemin)
-                       memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
                clear_io_state(td);
                prune_io_piece_log(td);
 
 
                clear_io_state(td);
                prune_io_piece_log(td);
 
-               if (td->io_engine == FIO_CPUIO)
+               if (td->io_ops->flags & FIO_CPUIO)
                        do_cpuio(td);
                else
                        do_io(td);
 
                        do_cpuio(td);
                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])
                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;
 
                if (td->error || td->terminate)
                        break;
@@ -1441,308 +744,115 @@ static void *thread_main(void *data)
                        continue;
 
                clear_io_state(td);
                        continue;
 
                clear_io_state(td);
-               gettimeofday(&td->start, NULL);
+               fio_gettime(&td->start, NULL);
 
                do_verify(td);
 
 
                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;
        }
 
 
                if (td->error || td->terminate)
                        break;
        }
 
-       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)
+       update_rusage_stat(td);
+       fio_gettime(&td->end_time, NULL);
+       td->runtime[0] = runtime[0] / 1000;
+       td->runtime[1] = runtime[1] / 1000;
+
+       if (td->ts.bw_log)
+               finish_log(td, td->ts.bw_log, "bw");
+       if (td->ts.slat_log)
+               finish_log(td, td->ts.slat_log, "slat");
+       if (td->ts.clat_log)
+               finish_log(td, td->ts.clat_log, "clat");
+       if (td->write_iolog_file)
                write_iolog_close(td);
                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)
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid);
+               terminate_threads(td->groupid, 0);
 
 err:
 
 err:
-       if (td->fd != -1) {
-               close(td->fd);
-               td->fd = -1;
-       }
-       if (td->mmap)
-               munmap(td->mmap, td->file_size);
-       cleanup_io(td);
+       if (td->error)
+               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
+       close_files(td);
+       close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return NULL;
-
+       return (void *) (unsigned long) td->error;
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
-static void *fork_main(int shmid, int offset)
+static int fork_main(int shmid, int offset)
 {
        struct thread_data *td;
 {
        struct thread_data *td;
-       void *data;
+       void *data, *ret;
 
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
 
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
+               int __err = errno;
+
                perror("shmat");
                perror("shmat");
-               return NULL;
+               return __err;
        }
 
        td = data + offset * sizeof(struct thread_data);
        }
 
        td = data + offset * sizeof(struct thread_data);
-       thread_main(td);
+       ret = thread_main(td);
        shmdt(data);
        shmdt(data);
-       return NULL;
-}
-
-/*
- * Sets the status of the 'td' in the printed status map.
- */
-static void check_str_update(struct thread_data *td)
-{
-       char c = run_str[td->thread_number - 1];
-
-       switch (td->runstate) {
-               case TD_REAPED:
-                       c = '_';
-                       break;
-               case TD_EXITED:
-                       c = 'E';
-                       break;
-               case TD_RUNNING:
-                       if (td_rw(td)) {
-                               if (td->sequential)
-                                       c = 'M';
-                               else
-                                       c = 'm';
-                       } else if (td_read(td)) {
-                               if (td->sequential)
-                                       c = 'R';
-                               else
-                                       c = 'r';
-                       } else {
-                               if (td->sequential)
-                                       c = 'W';
-                               else
-                                       c = 'w';
-                       }
-                       break;
-               case TD_VERIFYING:
-                       c = 'V';
-                       break;
-               case TD_FSYNCING:
-                       c = 'F';
-                       break;
-               case TD_CREATED:
-                       c = 'C';
-                       break;
-               case TD_INITIALIZED:
-                       c = 'I';
-                       break;
-               case TD_NOT_CREATED:
-                       c = 'P';
-                       break;
-               default:
-                       log_err("state %d\n", td->runstate);
-       }
-
-       run_str[td->thread_number - 1] = c;
-}
-
-/*
- * Convert seconds to a printable string.
- */
-static void eta_to_str(char *str, int eta_sec)
-{
-       unsigned int d, h, m, s;
-       static int always_d, always_h;
-
-       d = h = m = s = 0;
-
-       s = eta_sec % 60;
-       eta_sec /= 60;
-       m = eta_sec % 60;
-       eta_sec /= 60;
-       h = eta_sec % 24;
-       eta_sec /= 24;
-       d = eta_sec;
-
-       if (d || always_d) {
-               always_d = 1;
-               str += sprintf(str, "%02dd:", d);
-       }
-       if (h || always_h) {
-               always_h = 1;
-               str += sprintf(str, "%02dh:", h);
-       }
-
-       str += sprintf(str, "%02dm:", m);
-       str += sprintf(str, "%02ds", s);
+       return (int) (unsigned long) ret;
 }
 
 /*
 }
 
 /*
- * Best effort calculation of the estimated pending runtime of a job.
+ * Run over the job map and reap the threads that have exited, if any.
  */
  */
-static int thread_eta(struct thread_data *td, unsigned long elapsed)
+static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
 {
-       unsigned long long bytes_total, bytes_done;
-       unsigned int eta_sec = 0;
-
-       bytes_total = td->total_io_size;
+       struct thread_data *td;
+       int i, cputhreads, pending, status, ret;
 
        /*
 
        /*
-        * if writing, bytes_total will be twice the size. If mixing,
-        * assume a 50/50 split and thus bytes_total will be 50% larger.
+        * reap exited threads (TD_EXITED -> TD_REAPED)
         */
         */
-       if (td->verify) {
-               if (td_rw(td))
-                       bytes_total = bytes_total * 3 / 2;
-               else
-                       bytes_total <<= 1;
-       }
-       if (td->zone_size && td->zone_skip)
-               bytes_total /= (td->zone_skip / td->zone_size);
-
-       if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
-               double perc;
-
-               bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
-               perc = (double) bytes_done / (double) bytes_total;
-               if (perc > 1.0)
-                       perc = 1.0;
-
-               eta_sec = (elapsed * (1.0 / perc)) - 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
-                       || td->runstate == TD_INITIALIZED) {
-               int t_eta = 0, r_eta = 0;
-
-               /*
-                * We can only guess - assume it'll run the full timeout
-                * if given, otherwise assume it'll run at the specified rate.
-                */
-               if (td->timeout)
-                       t_eta = td->timeout + td->start_delay - elapsed;
-               if (td->rate) {
-                       r_eta = (bytes_total / 1024) / td->rate;
-                       r_eta += td->start_delay - elapsed;
-               }
-
-               if (r_eta && t_eta)
-                       eta_sec = min(r_eta, t_eta);
-               else if (r_eta)
-                       eta_sec = r_eta;
-               else if (t_eta)
-                       eta_sec = t_eta;
-               else
-                       eta_sec = 0;
-       } else {
+       pending = cputhreads = 0;
+       for_each_td(td, i) {
                /*
                /*
-                * thread is already done or waiting for fsync
+                * ->io_ops is NULL for a thread that has closed its
+                * io engine
                 */
                 */
-               eta_sec = 0;
-       }
-
-       return eta_sec;
-}
-
-/*
- * Print status of the jobs we know about. This includes rate estimates,
- * ETA, thread state, etc.
- */
-static void print_thread_status(void)
-{
-       unsigned long elapsed = time_since_now(&genesis);
-       int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec;
-       char eta_str[32];
-       double perc = 0.0;
-
-       if (temp_stall_ts || terse_output)
-               return;
-
-       eta_secs = malloc(thread_number * sizeof(int));
-       memset(eta_secs, 0, thread_number * sizeof(int));
-
-       nr_pending = nr_running = t_rate = m_rate = 0;
-       for (i = 0; i < thread_number; i++) {
-               struct thread_data *td = &threads[i];
-
-               if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING||
-                   td->runstate == TD_FSYNCING) {
-                       nr_running++;
-                       t_rate += td->rate;
-                       m_rate += td->ratemin;
-               } else if (td->runstate < TD_RUNNING)
-                       nr_pending++;
-
-               if (elapsed >= 3)
-                       eta_secs[i] = thread_eta(td, elapsed);
-               else
-                       eta_secs[i] = INT_MAX;
-
-               check_str_update(td);
-       }
+               if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
+                       cputhreads++;
 
 
-       if (exitall_on_terminate)
-               eta_sec = INT_MAX;
-       else
-               eta_sec = 0;
+               if (td->runstate < TD_EXITED) {
+                       /*
+                        * check if someone quit or got killed in an unusual way
+                        */
+                       ret = waitpid(td->pid, &status, WNOHANG);
+                       if (ret < 0)
+                               perror("waitpid");
+                       else if ((ret == td->pid) && WIFSIGNALED(status)) {
+                               int sig = WTERMSIG(status);
 
 
-       for (i = 0; i < thread_number; i++) {
-               if (exitall_on_terminate) {
-                       if (eta_secs[i] < eta_sec)
-                               eta_sec = eta_secs[i];
-               } else {
-                       if (eta_secs[i] > eta_sec)
-                               eta_sec = eta_secs[i];
+                               log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
                }
                }
-       }
-
-       if (eta_sec != INT_MAX && elapsed) {
-               perc = (double) elapsed / (double) (elapsed + eta_sec);
-               eta_to_str(eta_str, eta_sec);
-       }
-
-       if (!nr_running && !nr_pending)
-               return;
-
-       printf("Threads running: %d", nr_running);
-       if (m_rate || t_rate)
-               printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
-       if (eta_sec != INT_MAX && nr_running) {
-               perc *= 100.0;
-               printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
-       }
-       printf("\r");
-       fflush(stdout);
-       free(eta_secs);
-}
-
-/*
- * Run over the job map and reap the threads that have exited, if any.
- */
-static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
-{
-       int i, cputhreads;
-
-       /*
-        * reap exited threads (TD_EXITED -> TD_REAPED)
-        */
-       for (i = 0, cputhreads = 0; i < thread_number; i++) {
-               struct thread_data *td = &threads[i];
 
 
-               if (td->io_engine == FIO_CPUIO)
-                       cputhreads++;
+               if (td->runstate != TD_EXITED) {
+                       if (td->runstate < TD_RUNNING)
+                               pending++;
 
 
-               if (td->runstate != TD_EXITED)
                        continue;
                        continue;
+               }
+
+               if (td->error)
+                       exit_value++;
 
                td_set_runstate(td, TD_REAPED);
 
 
                td_set_runstate(td, TD_REAPED);
 
@@ -1751,58 +861,26 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 
                        if (pthread_join(td->thread, (void *) &ret))
                                perror("thread_join");
 
                        if (pthread_join(td->thread, (void *) &ret))
                                perror("thread_join");
-               } else
-                       waitpid(td->pid, NULL, 0);
+               } else {
+                       int status;
+
+                       ret = waitpid(td->pid, &status, 0);
+                       if (ret < 0)
+                               perror("waitpid");
+                       else if (WIFEXITED(status) && WEXITSTATUS(status)) {
+                               if (!exit_value)
+                                       exit_value++;
+                       }
+               }
 
 
+reaped:
                (*nr_running)--;
                (*m_rate) -= td->ratemin;
                (*t_rate) -= td->rate;
        }
 
                (*nr_running)--;
                (*m_rate) -= td->ratemin;
                (*t_rate) -= td->rate;
        }
 
-       if (*nr_running == cputhreads)
-               terminate_threads(TERMINATE_ALL);
-}
-
-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)
-{
-       unsigned long long phys_mem;
-       void *ptr;
-
-       if (!mlock_size)
-               return NULL;
-
-       /*
-        * Don't allow mlock of more than real_mem-128MB
-        */
-       phys_mem = os_phys_mem();
-       if (phys_mem) {
-               if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
-                       mlock_size = phys_mem - 128 * 1024 * 1024;
-                       fprintf(f_out, "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;
+       if (*nr_running == cputhreads && !pending)
+               terminate_threads(TERMINATE_ALL, 0);
 }
 
 /*
 }
 
 /*
@@ -1813,9 +891,9 @@ static void run_threads(void)
        struct thread_data *td;
        unsigned long spent;
        int i, todo, nr_running, m_rate, t_rate, nr_started;
        struct thread_data *td;
        unsigned long spent;
        int i, todo, nr_running, m_rate, t_rate, nr_started;
-       void *mlocked_mem;
 
 
-       mlocked_mem = fio_pin_memory();
+       if (fio_pin_memory())
+               return;
 
        if (!terse_output) {
                printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
 
        if (!terse_output) {
                printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
@@ -1830,28 +908,29 @@ static void run_threads(void)
        nr_started = 0;
        m_rate = t_rate = 0;
 
        nr_started = 0;
        m_rate = t_rate = 0;
 
-       for (i = 0; i < thread_number; i++) {
-               td = &threads[i];
+       for_each_td(td, i) {
+               print_status_init(td->thread_number - 1);
 
 
-               run_str[td->thread_number - 1] = 'P';
-
-               init_disk_util(td);
-
-               if (!td->create_serialize)
+               if (!td->create_serialize) {
+                       init_disk_util(td);
                        continue;
                        continue;
+               }
 
                /*
                 * do file setup here so it happens sequentially,
                 * we don't want X number of threads getting their
                 * client data interspersed on disk
                 */
 
                /*
                 * do file setup here so it happens sequentially,
                 * we don't want X number of threads getting their
                 * client data interspersed on disk
                 */
-               if (setup_file(td)) {
+               if (setup_files(td)) {
+                       exit_value++;
+                       if (td->error)
+                               log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
-       }
 
 
-       gettimeofday(&genesis, NULL);
+               init_disk_util(td);
+       }
 
        while (todo) {
                struct thread_data *map[MAX_JOBS];
 
        while (todo) {
                struct thread_data *map[MAX_JOBS];
@@ -1861,9 +940,7 @@ static void run_threads(void)
                /*
                 * create threads (TD_NOT_CREATED -> TD_CREATED)
                 */
                /*
                 * create threads (TD_NOT_CREATED -> TD_CREATED)
                 */
-               for (i = 0; i < thread_number; i++) {
-                       td = &threads[i];
-
+               for_each_td(td, i) {
                        if (td->runstate != TD_NOT_CREATED)
                                continue;
 
                        if (td->runstate != TD_NOT_CREATED)
                                continue;
 
@@ -1877,7 +954,7 @@ static void run_threads(void)
                        }
 
                        if (td->start_delay) {
                        }
 
                        if (td->start_delay) {
-                               spent = mtime_since_now(&genesis);
+                               spent = mtime_since_genesis();
 
                                if (td->start_delay * 1000 > spent)
                                        continue;
 
                                if (td->start_delay * 1000 > spent)
                                        continue;
@@ -1904,8 +981,9 @@ static void run_threads(void)
                                if (fork())
                                        fio_sem_down(&startup_sem);
                                else {
                                if (fork())
                                        fio_sem_down(&startup_sem);
                                else {
-                                       fork_main(shm_id, i);
-                                       exit(0);
+                                       int ret = fork_main(shm_id, i);
+
+                                       exit(ret);
                                }
                        }
                }
                                }
                        }
                }
@@ -1914,9 +992,9 @@ static void run_threads(void)
                 * Wait for the started threads to transition to
                 * TD_INITIALIZED.
                 */
                 * Wait for the started threads to transition to
                 * TD_INITIALIZED.
                 */
-               gettimeofday(&this_start, NULL);
+               fio_gettime(&this_start, NULL);
                left = this_jobs;
                left = this_jobs;
-               while (left) {
+               while (left && !fio_abort) {
                        if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
                                break;
 
                        if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
                                break;
 
@@ -1952,9 +1030,7 @@ static void run_threads(void)
                /*
                 * start created threads (TD_INITIALIZED -> TD_RUNNING).
                 */
                /*
                 * start created threads (TD_INITIALIZED -> TD_RUNNING).
                 */
-               for (i = 0; i < thread_number; i++) {
-                       td = &threads[i];
-
+               for_each_td(td, i) {
                        if (td->runstate != TD_INITIALIZED)
                                continue;
 
                        if (td->runstate != TD_INITIALIZED)
                                continue;
 
@@ -1979,11 +1055,20 @@ static void run_threads(void)
        }
 
        update_io_ticks();
        }
 
        update_io_ticks();
-       fio_unpin_memory(mlocked_mem);
+       fio_unpin_memory();
 }
 
 int main(int argc, char *argv[])
 {
 }
 
 int main(int argc, char *argv[])
 {
+       long ps;
+
+       /*
+        * We need locale for number printing, if it isn't set then just
+        * go with the US format.
+        */
+       if (!getenv("LC_NUMERIC"))
+               setlocale(LC_NUMERIC, "en_US");
+
        if (parse_options(argc, argv))
                return 1;
 
        if (parse_options(argc, argv))
                return 1;
 
@@ -1992,10 +1077,30 @@ int main(int argc, char *argv[])
                return 1;
        }
 
                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();
        disk_util_timer_arm();
 
        run_threads();
-       show_run_stats();
 
 
-       return 0;
+       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 exit_value;
 }
 }