fio: fix aio trim completion latencies
[fio.git] / blktrace.c
index 579a39f6775ba79cea1296beceabb0546efacfd3..efe9ce248a8c7f861a6aa5dbec9163f6d51ca90d 100644 (file)
@@ -3,14 +3,15 @@
  */
 #include <stdio.h>
 #include <stdlib.h>
-#include <sys/stat.h>
-#include <dirent.h>
+#include <unistd.h>
 
-#include "list.h"
+#include "flist.h"
 #include "fio.h"
+#include "blktrace.h"
 #include "blktrace_api.h"
+#include "oslib/linux-dev-lookup.h"
 
-#define TRACE_FIFO_SIZE        65536
+#define TRACE_FIFO_SIZE        8192
 
 /*
  * fifo refill frontend, to avoid reading data in trace sized bites
@@ -71,76 +72,41 @@ static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
  * Check if this is a blktrace binary data file. We read a single trace
  * into memory and check for the magic signature.
  */
-int is_blktrace(const char *filename)
+bool is_blktrace(const char *filename, int *need_swap)
 {
        struct blk_io_trace t;
        int fd, ret;
 
        fd = open(filename, O_RDONLY);
        if (fd < 0)
-               return 0;
+               return false;
 
        ret = read(fd, &t, sizeof(t));
        close(fd);
 
        if (ret < 0) {
                perror("read blktrace");
-               return 0;
+               return false;
        } else if (ret != sizeof(t)) {
                log_err("fio: short read on blktrace file\n");
-               return 0;
+               return false;
        }
 
-       if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
-               return 1;
-
-       return 0;
-}
-
-static int lookup_device(char *path, unsigned int maj, unsigned int min)
-{
-       struct dirent *dir;
-       struct stat st;
-       int found = 0;
-       DIR *D;
-
-       D = opendir(path);
-       if (!D)
-               return 0;
-
-       while ((dir = readdir(D)) != NULL) {
-               char full_path[256];
-
-               if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
-                       continue;
-
-               sprintf(full_path, "%s/%s", path, dir->d_name);
-               if (lstat(full_path, &st) == -1) {
-                       perror("lstat");
-                       break;
-               }
-
-               if (S_ISDIR(st.st_mode)) {
-                       found = lookup_device(full_path, maj, min);
-                       if (found) {
-                               strcpy(path, full_path);
-                               break;
-                       }
-               }
-
-               if (!S_ISBLK(st.st_mode))
-                       continue;
+       if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+               *need_swap = 0;
+               return true;
+       }
 
-               if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
-                       dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
-                       strcpy(path, full_path);
-                       found = 1;
-                       break;
-               }
+       /*
+        * Maybe it needs to be endian swapped...
+        */
+       t.magic = fio_swap32(t.magic);
+       if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+               *need_swap = 1;
+               return true;
        }
 
-       closedir(D);
-       return found;
+       return false;
 }
 
 #define FMINORBITS     20
@@ -148,21 +114,22 @@ static int lookup_device(char *path, unsigned int maj, unsigned int min)
 #define FMAJOR(dev)    ((unsigned int) ((dev) >> FMINORBITS))
 #define FMINOR(dev)    ((unsigned int) ((dev) & FMINORMASK))
 
-static void trace_add_open_event(struct thread_data *td, int fileno)
+static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
 {
        struct io_piece *ipo;
 
        ipo = calloc(1, sizeof(*ipo));
+       init_ipo(ipo);
 
        ipo->ddir = DDIR_INVAL;
        ipo->fileno = fileno;
-       ipo->file_action = FIO_LOG_OPEN_FILE;
-       list_add_tail(&ipo->list, &td->io_log_list);
+       ipo->file_action = action;
+       flist_add_tail(&ipo->list, &td->io_log_list);
 }
 
-static void trace_add_file(struct thread_data *td, __u32 device)
+static int trace_add_file(struct thread_data *td, __u32 device)
 {
-       static unsigned int last_maj, last_min;
+       static unsigned int last_maj, last_min, last_fileno;
        unsigned int maj = FMAJOR(device);
        unsigned int min = FMINOR(device);
        struct fio_file *f;
@@ -170,7 +137,7 @@ static void trace_add_file(struct thread_data *td, __u32 device)
        unsigned int i;
 
        if (last_maj == maj && last_min == min)
-               return;
+               return last_fileno;
 
        last_maj = maj;
        last_min = min;
@@ -179,39 +146,65 @@ static void trace_add_file(struct thread_data *td, __u32 device)
         * check for this file in our list
         */
        for_each_file(td, f, i)
-               if (f->major == maj && f->minor == min)
-                       return;
+               if (f->major == maj && f->minor == min) {
+                       last_fileno = f->fileno;
+                       return last_fileno;
+               }
 
        strcpy(dev, "/dev");
-       if (lookup_device(dev, maj, min)) {
+       if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
                int fileno;
 
+               if (td->o.replay_redirect)
+                       dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
+                                       " with: %s\n", maj, min,
+                                       td->o.replay_redirect);
+               else
+                       dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
+
                dprint(FD_BLKTRACE, "add devices %s\n", dev);
-               fileno = add_file(td, dev);
-               trace_add_open_event(td, fileno);
+               fileno = add_file_exclusive(td, dev);
+               td->o.open_files++;
+               td->files[fileno]->major = maj;
+               td->files[fileno]->minor = min;
+               trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
+               last_fileno = fileno;
        }
+
+       return last_fileno;
+}
+
+static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
+{
+       if (!o->replay_align)
+               return;
+
+       t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 }
 
 /*
  * Store blk_io_trace data in an ipo for later retrieval.
  */
 static void store_ipo(struct thread_data *td, unsigned long long offset,
-                     unsigned int bytes, int rw, unsigned long long ttime)
+                     unsigned int bytes, int rw, unsigned long long ttime,
+                     int fileno)
 {
-       struct io_piece *ipo = malloc(sizeof(*ipo));
+       struct io_piece *ipo;
+
+       ipo = calloc(1, sizeof(*ipo));
+       init_ipo(ipo);
 
-       memset(ipo, 0, sizeof(*ipo));
-       INIT_LIST_HEAD(&ipo->list);
-       /*
-        * the 512 is wrong here, it should be the hardware sector size...
-        */
        ipo->offset = offset * 512;
+       if (td->o.replay_scale)
+               ipo->offset = ipo->offset / td->o.replay_scale;
+       ipo_bytes_align(td->o.replay_align, ipo);
        ipo->len = bytes;
        ipo->delay = ttime / 1000;
        if (rw)
                ipo->ddir = DDIR_WRITE;
        else
                ipo->ddir = DDIR_READ;
+       ipo->fileno = fileno;
 
        dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
                                                        ipo->ddir, ipo->offset,
@@ -219,14 +212,18 @@ static void store_ipo(struct thread_data *td, unsigned long long offset,
        queue_io_piece(td, ipo);
 }
 
-static void handle_trace_notify(struct thread_data *td, struct blk_io_trace *t)
+static void handle_trace_notify(struct blk_io_trace *t)
 {
        switch (t->action) {
        case BLK_TN_PROCESS:
-               printf("got process notify: %x, %d\n", t->action, t->pid);
+               dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
+                               t->action, t->pid);
                break;
        case BLK_TN_TIMESTAMP:
-               printf("got timestamp notify: %x, %d\n", t->action, t->pid);
+               dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
+                               t->action, t->pid);
+               break;
+       case BLK_TN_MESSAGE:
                break;
        default:
                dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
@@ -234,22 +231,102 @@ static void handle_trace_notify(struct thread_data *td, struct blk_io_trace *t)
        }
 }
 
+static void handle_trace_discard(struct thread_data *td,
+                                struct blk_io_trace *t,
+                                unsigned long long ttime,
+                                unsigned long *ios, unsigned int *bs)
+{
+       struct io_piece *ipo;
+       int fileno;
+
+       if (td->o.replay_skip & (1u << DDIR_TRIM))
+               return;
+
+       ipo = calloc(1, sizeof(*ipo));
+       init_ipo(ipo);
+       fileno = trace_add_file(td, t->device);
+
+       ios[DDIR_TRIM]++;
+       if (t->bytes > bs[DDIR_TRIM])
+               bs[DDIR_TRIM] = t->bytes;
+
+       td->o.size += t->bytes;
+
+       INIT_FLIST_HEAD(&ipo->list);
+
+       ipo->offset = t->sector * 512;
+       if (td->o.replay_scale)
+               ipo->offset = ipo->offset / td->o.replay_scale;
+       ipo_bytes_align(td->o.replay_align, ipo);
+       ipo->len = t->bytes;
+       ipo->delay = ttime / 1000;
+       ipo->ddir = DDIR_TRIM;
+       ipo->fileno = fileno;
+
+       dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
+                                                       ipo->offset, ipo->len,
+                                                       ipo->delay);
+       queue_io_piece(td, ipo);
+}
+
+static void dump_trace(struct blk_io_trace *t)
+{
+       log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
+}
+
 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
                            unsigned long long ttime, unsigned long *ios,
                            unsigned int *bs)
 {
        int rw;
+       int fileno;
 
-       trace_add_file(td, t->device);
+       fileno = trace_add_file(td, t->device);
 
        rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 
+       if (rw) {
+               if (td->o.replay_skip & (1u << DDIR_WRITE))
+                       return;
+       } else {
+               if (td->o.replay_skip & (1u << DDIR_READ))
+                       return;
+       }
+
+       if (!t->bytes) {
+               if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
+                       dump_trace(t);
+               return;
+       }
+
        if (t->bytes > bs[rw])
                bs[rw] = t->bytes;
 
        ios[rw]++;
        td->o.size += t->bytes;
-       store_ipo(td, t->sector, t->bytes, rw, ttime);
+       store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
+}
+
+static void handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
+                              unsigned long long ttime, unsigned long *ios)
+{
+       struct io_piece *ipo;
+       int fileno;
+
+       if (td->o.replay_skip & (1u << DDIR_SYNC))
+               return;
+
+       ipo = calloc(1, sizeof(*ipo));
+       init_ipo(ipo);
+       fileno = trace_add_file(td, t->device);
+
+       ipo->delay = ttime / 1000;
+       ipo->ddir = DDIR_SYNC;
+       ipo->fileno = fileno;
+
+       ios[DDIR_SYNC]++;
+       dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
+       queue_io_piece(td, ipo);
 }
 
 /*
@@ -257,48 +334,130 @@ static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
  * due to internal workings of the block layer.
  */
 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
-                        unsigned long long ttime, unsigned long *ios,
-                        unsigned int *bs)
+                        unsigned long *ios, unsigned int *bs)
 {
+       static unsigned long long last_ttime;
+       unsigned long long delay = 0;
+
        if ((t->action & 0xffff) != __BLK_TA_QUEUE)
                return;
-       if (t->action & BLK_TC_ACT(BLK_TC_PC))
-               return;
+
+       if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
+               if (!last_ttime || td->o.no_stall)
+                       delay = 0;
+               else if (td->o.replay_time_scale == 100)
+                       delay = t->time - last_ttime;
+               else {
+                       double tmp = t->time - last_ttime;
+                       double scale;
+
+                       scale = (double) 100.0 / (double) td->o.replay_time_scale;
+                       tmp *= scale;
+                       delay = tmp;
+               }
+               last_ttime = t->time;
+       }
+
+       t_bytes_align(&td->o, t);
 
        if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
-               handle_trace_notify(td, t);
+               handle_trace_notify(t);
+       else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
+               handle_trace_discard(td, t, delay, ios, bs);
+       else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
+               handle_trace_flush(td, t, delay, ios);
        else
-               handle_trace_fs(td, t, ttime, ios, bs);
+               handle_trace_fs(td, t, delay, ios, bs);
+}
+
+static void byteswap_trace(struct blk_io_trace *t)
+{
+       t->magic = fio_swap32(t->magic);
+       t->sequence = fio_swap32(t->sequence);
+       t->time = fio_swap64(t->time);
+       t->sector = fio_swap64(t->sector);
+       t->bytes = fio_swap32(t->bytes);
+       t->action = fio_swap32(t->action);
+       t->pid = fio_swap32(t->pid);
+       t->device = fio_swap32(t->device);
+       t->cpu = fio_swap32(t->cpu);
+       t->error = fio_swap16(t->error);
+       t->pdu_len = fio_swap16(t->pdu_len);
+}
+
+static bool t_is_write(struct blk_io_trace *t)
+{
+       return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
+}
+
+static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
+{
+       if (t->action & BLK_TC_ACT(BLK_TC_READ))
+               return DDIR_READ;
+       else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
+               return DDIR_WRITE;
+       else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
+               return DDIR_TRIM;
+
+       return DDIR_INVAL;
+}
+
+static void depth_inc(struct blk_io_trace *t, int *depth)
+{
+       enum fio_ddir ddir;
+
+       ddir = t_get_ddir(t);
+       if (ddir != DDIR_INVAL)
+               depth[ddir]++;
+}
+
+static void depth_dec(struct blk_io_trace *t, int *depth)
+{
+       enum fio_ddir ddir;
+
+       ddir = t_get_ddir(t);
+       if (ddir != DDIR_INVAL)
+               depth[ddir]--;
+}
+
+static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
+{
+       enum fio_ddir ddir = DDIR_INVAL;
+
+       ddir = t_get_ddir(t);
+       if (ddir != DDIR_INVAL) {
+               depth[ddir] = max(depth[ddir], this_depth[ddir]);
+               this_depth[ddir] = 0;
+       }
 }
 
 /*
  * Load a blktrace file by reading all the blk_io_trace entries, and storing
  * them as io_pieces like the fio text version would do.
  */
-int load_blktrace(struct thread_data *td, const char *filename)
+bool load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 {
-       unsigned long long ttime, delay;
        struct blk_io_trace t;
-       unsigned long ios[2], skipped_writes;
-       unsigned int cpu;
-       unsigned int rw_bs[2];
+       unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
+       unsigned int rw_bs[DDIR_RWDIR_CNT] = { };
+       unsigned long skipped_writes;
        struct fifo *fifo;
-       int fd;
+       int fd, i, old_state, max_depth;
+       struct fio_file *f;
+       int this_depth[DDIR_RWDIR_CNT] = { };
+       int depth[DDIR_RWDIR_CNT] = { };
 
        fd = open(filename, O_RDONLY);
        if (fd < 0) {
                td_verror(td, errno, "open blktrace file");
-               return 1;
+               return false;
        }
 
        fifo = fifo_alloc(TRACE_FIFO_SIZE);
 
-       td->o.size = 0;
+       old_state = td_bump_runstate(td, TD_SETTING_UP);
 
-       cpu = 0;
-       ttime = 0;
-       ios[0] = ios[1] = 0;
-       rw_bs[0] = rw_bs[1] = 0;
+       td->o.size = 0;
        skipped_writes = 0;
        do {
                int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
@@ -312,6 +471,9 @@ int load_blktrace(struct thread_data *td, const char *filename)
                        break;
                }
 
+               if (need_swap)
+                       byteswap_trace(&t);
+
                if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
                        log_err("fio: bad magic in blktrace data: %x\n",
                                                                t.magic);
@@ -331,38 +493,58 @@ int load_blktrace(struct thread_data *td, const char *filename)
                        goto err;
                }
                if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
-                       if (!ttime) {
-                               ttime = t.time;
-                               cpu = t.cpu;
-                       }
-
-                       delay = 0;
-                       if (cpu == t.cpu)
-                               delay = t.time - ttime;
-                       if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
+                       if ((t.action & 0xffff) == __BLK_TA_QUEUE)
+                               depth_inc(&t, this_depth);
+                       else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
+                               ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
+                               depth_dec(&t, this_depth);
+                       else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
+                               depth_end(&t, this_depth, depth);
+
+                       if (t_is_write(&t) && read_only) {
                                skipped_writes++;
-                       else
-                               handle_trace(td, &t, delay, ios, rw_bs);
-
-                       ttime = t.time;
-                       cpu = t.cpu;
-               } else {
-                       delay = 0;
-                       handle_trace(td, &t, delay, ios, rw_bs);
+                               continue;
+                       }
                }
+
+               handle_trace(td, &t, ios, rw_bs);
        } while (1);
 
+       for_each_file(td, f, i)
+               trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
+
        fifo_free(fifo);
        close(fd);
 
+       td_restore_runstate(td, old_state);
+
+       if (!td->files_index) {
+               log_err("fio: did not find replay device(s)\n");
+               return false;
+       }
+
+       /*
+        * For stacked devices, we don't always get a COMPLETE event so
+        * the depth grows to insane values. Limit it to something sane(r).
+        */
+       max_depth = 0;
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               if (depth[i] > 1024)
+                       depth[i] = 1024;
+               else if (!depth[i] && ios[i])
+                       depth[i] = 1;
+               max_depth = max(depth[i], max_depth);
+       }
+
        if (skipped_writes)
                log_err("fio: %s skips replay of %lu writes due to read-only\n",
                                                td->o.name, skipped_writes);
 
-       if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
+       if (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
+           !ios[DDIR_SYNC]) {
                log_err("fio: found no ios in blktrace data\n");
-               return 1;
-       } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
+               return false;
+       } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
                td->o.td_ddir = TD_DDIR_READ;
                td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
        } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
@@ -372,17 +554,236 @@ int load_blktrace(struct thread_data *td, const char *filename)
                td->o.td_ddir = TD_DDIR_RW;
                td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
                td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
+               td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
        }
 
        /*
         * We need to do direct/raw ios to the device, to avoid getting
-        * read-ahead in our way.
+        * read-ahead in our way. But only do so if the minimum block size
+        * is a multiple of 4k, otherwise we don't know if it's safe to do so.
         */
-       td->o.odirect = 1;
+       if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
+               td->o.odirect = 1;
 
-       return 0;
+       /*
+        * If depth wasn't manually set, use probed depth
+        */
+       if (!fio_option_is_set(&td->o, iodepth))
+               td->o.iodepth = td->o.iodepth_low = max_depth;
+
+       return true;
 err:
        close(fd);
        fifo_free(fifo);
-       return 1;
+       return false;
+}
+
+static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
+                                int nr_logs, int def, size_t off)
+{
+       int i = 0, len = 0;
+
+       while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
+               len++;
+
+       if (len && len != nr_logs)
+               return len;
+
+       for (i = 0; i < nr_logs; i++) {
+               int *val = (int *)((char *)&bcs[i] + off);
+               *val = def;
+               if (len)
+                       *val = (int)vals[i].u.f;
+       }
+
+       return 0;
+
+}
+
+static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
+{
+       __u64 time = ~(__u64)0;
+       int idx = 0, i;
+
+       for (i = 0; i < nr_logs; i++) {
+               if (bcs[i].t.time < time) {
+                       time = bcs[i].t.time;
+                       idx = i;
+               }
+       }
+
+       return idx;
+}
+
+static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
+{
+       bcs[i].iter++;
+       if (bcs[i].iter < bcs[i].nr_iter) {
+               lseek(bcs[i].fd, 0, SEEK_SET);
+               return;
+       }
+
+       *nr_logs -= 1;
+
+       /* close file */
+       fifo_free(bcs[i].fifo);
+       close(bcs[i].fd);
+
+       /* keep active files contiguous */
+       memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
+}
+
+static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
+{
+       int ret = 0;
+       struct blk_io_trace *t = &bc->t;
+
+read_skip:
+       /* read an io trace */
+       ret = trace_fifo_get(td, bc->fifo, bc->fd, t, sizeof(*t));
+       if (ret < 0) {
+               return ret;
+       } else if (!ret) {
+               if (!bc->length)
+                       bc->length = bc->t.time;
+               return ret;
+       } else if (ret < (int) sizeof(*t)) {
+               log_err("fio: short fifo get\n");
+               return -1;
+       }
+
+       if (bc->swap)
+               byteswap_trace(t);
+
+       /* skip over actions that fio does not care about */
+       if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
+           t_get_ddir(t) == DDIR_INVAL) {
+               ret = discard_pdu(td, bc->fifo, bc->fd, t);
+               if (ret < 0) {
+                       td_verror(td, ret, "blktrace lseek");
+                       return ret;
+               } else if (t->pdu_len != ret) {
+                       log_err("fio: discarded %d of %d\n", ret,
+                               t->pdu_len);
+                       return -1;
+               }
+               goto read_skip;
+       }
+
+       t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
+
+       return ret;
+}
+
+static int write_trace(FILE *fp, struct blk_io_trace *t)
+{
+       /* pdu is not used so just write out only the io trace */
+       t->pdu_len = 0;
+       return fwrite((void *)t, sizeof(*t), 1, fp);
+}
+
+int merge_blktrace_iologs(struct thread_data *td)
+{
+       int nr_logs = get_max_str_idx(td->o.read_iolog_file);
+       struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
+                                            nr_logs);
+       struct blktrace_cursor *bc;
+       FILE *merge_fp;
+       char *str, *ptr, *name, *merge_buf;
+       int i, ret;
+
+       ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
+                                   100, offsetof(struct blktrace_cursor,
+                                                 scalar));
+       if (ret) {
+               log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
+                       ret, nr_logs);
+               goto err_param;
+       }
+
+       ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
+                                   1, offsetof(struct blktrace_cursor,
+                                               nr_iter));
+       if (ret) {
+               log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
+                       ret, nr_logs);
+               goto err_param;
+       }
+
+       /* setup output file */
+       merge_fp = fopen(td->o.merge_blktrace_file, "w");
+       merge_buf = malloc(128 * 1024);
+       ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
+       if (ret)
+               goto err_out_file;
+
+       /* setup input files */
+       str = ptr = strdup(td->o.read_iolog_file);
+       nr_logs = 0;
+       for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
+               bcs[i].fd = open(name, O_RDONLY);
+               if (bcs[i].fd < 0) {
+                       log_err("fio: could not open file: %s\n", name);
+                       ret = bcs[i].fd;
+                       goto err_file;
+               }
+               bcs[i].fifo = fifo_alloc(TRACE_FIFO_SIZE);
+               nr_logs++;
+
+               if (!is_blktrace(name, &bcs[i].swap)) {
+                       log_err("fio: file is not a blktrace: %s\n", name);
+                       goto err_file;
+               }
+
+               ret = read_trace(td, &bcs[i]);
+               if (ret < 0) {
+                       goto err_file;
+               } else if (!ret) {
+                       merge_finish_file(bcs, i, &nr_logs);
+                       i--;
+               }
+       }
+       free(str);
+
+       /* merge files */
+       while (nr_logs) {
+               i = find_earliest_io(bcs, nr_logs);
+               bc = &bcs[i];
+               /* skip over the pdu */
+               ret = discard_pdu(td, bc->fifo, bc->fd, &bc->t);
+               if (ret < 0) {
+                       td_verror(td, ret, "blktrace lseek");
+                       goto err_file;
+               } else if (bc->t.pdu_len != ret) {
+                       log_err("fio: discarded %d of %d\n", ret,
+                               bc->t.pdu_len);
+                       goto err_file;
+               }
+
+               ret = write_trace(merge_fp, &bc->t);
+               ret = read_trace(td, bc);
+               if (ret < 0)
+                       goto err_file;
+               else if (!ret)
+                       merge_finish_file(bcs, i, &nr_logs);
+       }
+
+       /* set iolog file to read from the newly merged file */
+       td->o.read_iolog_file = td->o.merge_blktrace_file;
+       ret = 0;
+
+err_file:
+       /* cleanup */
+       for (i = 0; i < nr_logs; i++) {
+               fifo_free(bcs[i].fifo);
+               close(bcs[i].fd);
+       }
+err_out_file:
+       fflush(merge_fp);
+       fclose(merge_fp);
+       free(merge_buf);
+err_param:
+       free(bcs);
+
+       return ret;
 }