Improve submission latency calculation

[fio.git] / io_u.c

diff --git a/io_u.c b/io_u.c
index ebfa0d45a324111a3cda91a087909cf57e063cbe..6234c42b3a91b32305110f96e9e216472a9f537c 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -3,11 +3,19 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
 #include <string.h>
 #include <signal.h>
 #include <time.h>

-#include <assert.h>

 
 #include "fio.h"
 #include "os.h"
 
 
 #include "fio.h"
 #include "os.h"
 

+struct io_completion_data {
+       int nr;                         /* input */
+       endio_handler *handler;         /* input */
+
+       int error;                      /* output */
+       unsigned long bytes_done[2];    /* output */
+       struct timeval time;            /* output */
+};
+

 /*
  * 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.
 /*
  * 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.


@@ -27,13 +35,16 @@ static int random_map_free(struct thread_data *td, struct fio_file *f,
 static void mark_random_map(struct thread_data *td, struct fio_file *f,
                            struct io_u *io_u)
 {
 static void mark_random_map(struct thread_data *td, struct fio_file *f,
                            struct io_u *io_u)
 {

-       unsigned int min_bs = td->min_bs[io_u->ddir];
+       unsigned int min_bs = td->rw_min_bs;

        unsigned long long block;
        unsigned int blocks;
        unsigned long long block;
        unsigned int blocks;

+       unsigned int nr_blocks;

 
        block = io_u->offset / (unsigned long long) min_bs;
        blocks = 0;
 
        block = io_u->offset / (unsigned long long) min_bs;
        blocks = 0;

-       while (blocks < (io_u->buflen / min_bs)) {
+       nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+
+       while (blocks < nr_blocks) {

                unsigned int idx, bit;
 
                if (!random_map_free(td, f, block))
                unsigned int idx, bit;
 
                if (!random_map_free(td, f, block))


@@ -42,7 +53,7 @@ static void mark_random_map(struct thread_data *td, struct fio_file *f,
                idx = RAND_MAP_IDX(td, f, block);
                bit = RAND_MAP_BIT(td, f, block);
 
                idx = RAND_MAP_IDX(td, f, block);
                bit = RAND_MAP_BIT(td, f, block);
 

-               assert(idx < f->num_maps);
+               fio_assert(td, idx < f->num_maps);

 
                f->file_map[idx] |= (1UL << bit);
                block++;
 
                f->file_map[idx] |= (1UL << bit);
                block++;


@@ -61,11 +72,12 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 {
        int i;
 
 {
        int i;
 

-       *b = 0;
-       i = 0;
+       i = f->last_free_lookup;
+       *b = (i * BLOCKS_PER_MAP);

        while ((*b) * td->rw_min_bs < f->real_file_size) {
                if (f->file_map[i] != -1UL) {
                        *b += ffz(f->file_map[i]);
        while ((*b) * td->rw_min_bs < f->real_file_size) {
                if (f->file_map[i] != -1UL) {
                        *b += ffz(f->file_map[i]);

+                       f->last_free_lookup = i;

                        return 0;
                }
 
                        return 0;
                }
 


@@ -82,14 +94,15 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
  * the last io issued.
  */
 static int get_next_offset(struct thread_data *td, struct fio_file *f,
  * the last io issued.
  */
 static int get_next_offset(struct thread_data *td, struct fio_file *f,

-                          unsigned long long *offset, int ddir)
+                          struct io_u *io_u)

 {
 {

+       const int ddir = io_u->ddir;

        unsigned long long b, rb;
        long r;
 
        if (!td->sequential) {
                unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
        unsigned long long b, rb;
        long r;
 
        if (!td->sequential) {
                unsigned long long max_blocks = f->file_size / td->min_bs[ddir];

-               int loops = 50;
+               int loops = 5;

 
                do {
                        r = os_random_long(&td->random_state);
 
                do {
                        r = os_random_long(&td->random_state);


@@ -100,22 +113,26 @@ static int get_next_offset(struct thread_data *td, struct fio_file *f,
                        loops--;
                } while (!random_map_free(td, f, rb) && loops);
 
                        loops--;
                } while (!random_map_free(td, f, rb) && loops);
 

-               if (!loops) {
-                       if (get_next_free_block(td, f, &b))
-                               return 1;
-               }
+               /*
+                * if we failed to retrieve a truly random offset within
+                * the loops assigned, see if there are free ones left at all
+                */
+               if (!loops && get_next_free_block(td, f, &b))
+                       return 1;

        } else
                b = f->last_pos / td->min_bs[ddir];
 
        } else
                b = f->last_pos / td->min_bs[ddir];
 

-       *offset = (b * td->min_bs[ddir]) + f->file_offset;
-       if (*offset > f->real_file_size)
+       io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
+       if (io_u->offset >= f->real_file_size)

                return 1;
 
        return 0;
 }
 
                return 1;
 
        return 0;
 }
 

-static unsigned int get_next_buflen(struct thread_data *td, int ddir)
+static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
+                                   struct io_u *io_u)

 {
 {

+       const int ddir = io_u->ddir;

        unsigned int buflen;
        long r;
 
        unsigned int buflen;
        long r;
 


@@ -128,15 +145,11 @@ static unsigned int get_next_buflen(struct thread_data *td, int ddir)
                        buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
        }
 
                        buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
        }
 

-       if (buflen > td->io_size - td->this_io_bytes[ddir]) {
-               /*
-                * if using direct/raw io, we may not be able to
-                * shrink the size. so just fail it.
-                */
-               if (td->io_ops->flags & FIO_RAWIO)
+       while (buflen + io_u->offset > f->real_file_size) {
+               if (buflen == td->min_bs[ddir])

                        return 0;
 
                        return 0;
 

-               buflen = td->io_size - td->this_io_bytes[ddir];
+               buflen = td->min_bs[ddir];

        }
 
        return buflen;
        }
 
        return buflen;


@@ -186,6 +199,16 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
        td->cur_depth--;
 }
 
        td->cur_depth--;
 }
 

+void requeue_io_u(struct thread_data *td, struct io_u **io_u)
+{
+       struct io_u *__io_u = *io_u;
+
+       list_del(&__io_u->list);
+       list_add_tail(&__io_u->list, &td->io_u_requeues);
+       td->cur_depth--;
+       *io_u = NULL;
+}
+

 static int fill_io_u(struct thread_data *td, struct fio_file *f,
                     struct io_u *io_u)
 {
 static int fill_io_u(struct thread_data *td, struct fio_file *f,
                     struct io_u *io_u)
 {


@@ -198,8 +221,8 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
        /*
         * see if it's time to sync
         */
        /*
         * see if it's time to sync
         */

-       if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
-           && should_fsync(td)) {
+       if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
+           && td->io_issues[DDIR_WRITE] && should_fsync(td)) {

                io_u->ddir = DDIR_SYNC;
                io_u->file = f;
                return 0;
                io_u->ddir = DDIR_SYNC;
                io_u->file = f;
                return 0;


@@ -208,23 +231,30 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f,
        io_u->ddir = get_rw_ddir(td);
 
        /*
        io_u->ddir = get_rw_ddir(td);
 
        /*

-        * No log, let the seq/rand engine retrieve the next position.
+        * No log, let the seq/rand engine retrieve the next buflen and
+        * position.

         */
         */

-       if (!get_next_offset(td, f, &io_u->offset, io_u->ddir)) {
-               io_u->buflen = get_next_buflen(td, io_u->ddir);
-               if (io_u->buflen) {
-                       /*
-                        * If using a write iolog, store this entry.
-                        */
-                       if (td->write_iolog_file)
-                               write_iolog_put(td, io_u);
-
-                       io_u->file = f;
-                       return 0;
-               }
-       }
+       if (get_next_offset(td, f, io_u))
+               return 1;

 
 

-       return 1;
+       io_u->buflen = get_next_buflen(td, f, io_u);
+       if (!io_u->buflen)
+               return 1;
+
+       /*
+        * mark entry before potentially trimming io_u
+        */
+       if (!td->read_iolog && !td->sequential && !td->norandommap)
+               mark_random_map(td, f, io_u);
+
+       /*
+        * If using a write iolog, store this entry.
+        */
+       if (td->write_iolog_file)
+               write_iolog_put(td, io_u);
+
+       io_u->file = f;
+       return 0;

 }
 
 static void io_u_mark_depth(struct thread_data *td)
 }
 
 static void io_u_mark_depth(struct thread_data *td)


@@ -252,16 +282,77 @@ static void io_u_mark_depth(struct thread_data *td)
        td->total_io_u++;
 }
 
        td->total_io_u++;
 }
 

+static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
+{
+       int index = 0;
+
+       switch (msec) {
+       default:
+               index++;
+       case 1024 ... 2047:
+               index++;
+       case 512 ... 1023:
+               index++;
+       case 256 ... 511:
+               index++;
+       case 128 ... 255:
+               index++;
+       case 64 ... 127:
+               index++;
+       case 32 ... 63:
+               index++;
+       case 16 ... 31:
+               index++;
+       case 8 ... 15:
+               index++;
+       case 4 ... 7:
+               index++;
+       case 2 ... 3:
+               index++;
+       case 0 ... 1:
+               break;
+       }
+
+       td->io_u_lat[index]++;
+}
+
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+       unsigned int old_next_file = td->next_file;
+       struct fio_file *f;
+
+       do {
+               f = &td->files[td->next_file];
+
+               td->next_file++;
+               if (td->next_file >= td->nr_files)
+                       td->next_file = 0;
+
+               if (f->fd != -1)
+                       break;
+
+               f = NULL;
+       } while (td->next_file != old_next_file);
+
+       return f;
+}
+

 struct io_u *__get_io_u(struct thread_data *td)
 {
        struct io_u *io_u = NULL;
 
 struct io_u *__get_io_u(struct thread_data *td)
 {
        struct io_u *io_u = NULL;
 

-       if (!queue_full(td)) {
+       if (!list_empty(&td->io_u_requeues))
+               io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
+       else if (!queue_full(td)) {

                io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
 
                io_u->buflen = 0;
                io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
 
                io_u->buflen = 0;

-               io_u->error = 0;

                io_u->resid = 0;
                io_u->resid = 0;

+               io_u->file = NULL;
+       }
+
+       if (io_u) {
+               io_u->error = 0;

                list_del(&io_u->list);
                list_add(&io_u->list, &td->io_u_busylist);
                td->cur_depth++;
                list_del(&io_u->list);
                list_add(&io_u->list, &td->io_u_busylist);
                td->cur_depth++;


@@ -275,14 +366,29 @@ struct io_u *__get_io_u(struct thread_data *td)
  * 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.
  */
  * 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.
  */

-struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
+struct io_u *get_io_u(struct thread_data *td)

 {
 {

+       struct fio_file *f;

        struct io_u *io_u;
 
        io_u = __get_io_u(td);
        if (!io_u)
                return NULL;
 
        struct io_u *io_u;
 
        io_u = __get_io_u(td);
        if (!io_u)
                return NULL;
 

+       /*
+        * from a requeue, io_u already setup
+        */
+       if (io_u->file)
+               goto out;
+
+       f = get_next_file(td);
+       if (!f) {
+               put_io_u(td, io_u);
+               return NULL;
+       }
+
+       io_u->file = f;
+

        if (td->zone_bytes >= td->zone_size) {
                td->zone_bytes = 0;
                f->last_pos += td->zone_skip;
        if (td->zone_bytes >= td->zone_size) {
                td->zone_bytes = 0;
                f->last_pos += td->zone_skip;


@@ -308,32 +414,30 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
                        return NULL;
                }
 
                        return NULL;
                }
 

-               if (!td->read_iolog && !td->sequential && !td->norandommap)
-                       mark_random_map(td, f, io_u);
-
-               f->last_pos += io_u->buflen;
+               f->last_pos = io_u->offset + io_u->buflen;

 
                if (td->verify != VERIFY_NONE)
                        populate_verify_io_u(td, io_u);
        }
 
 
                if (td->verify != VERIFY_NONE)
                        populate_verify_io_u(td, io_u);
        }
 

-       if (td_io_prep(td, io_u)) {
-               put_io_u(td, io_u);
-               return NULL;
-       }
-

        /*
         * Set io data pointers.
         */
        /*
         * Set io data pointers.
         */

+out:

        io_u->xfer_buf = io_u->buf;
        io_u->xfer_buflen = io_u->buflen;
 
        io_u->xfer_buf = io_u->buf;
        io_u->xfer_buflen = io_u->buflen;
 

+       if (td_io_prep(td, io_u)) {
+               put_io_u(td, io_u);
+               return NULL;
+       }
+

        fio_gettime(&io_u->start_time, NULL);
        return io_u;
 }
 
        fio_gettime(&io_u->start_time, NULL);
        return io_u;
 }
 

-void io_completed(struct thread_data *td, struct io_u *io_u,
-                 struct io_completion_data *icd)
+static void io_completed(struct thread_data *td, struct io_u *io_u,
+                        struct io_completion_data *icd)

 {
        unsigned long msec;
 
 {
        unsigned long msec;
 


@@ -347,6 +451,7 @@ void io_completed(struct thread_data *td, struct io_u *io_u,
        if (!io_u->error) {
                unsigned int bytes = io_u->buflen - io_u->resid;
                const enum fio_ddir idx = io_u->ddir;
        if (!io_u->error) {
                unsigned int bytes = io_u->buflen - io_u->resid;
                const enum fio_ddir idx = io_u->ddir;

+               int ret;

 
                td->io_blocks[idx]++;
                td->io_bytes[idx] += bytes;
 
                td->io_blocks[idx]++;
                td->io_bytes[idx] += bytes;


@@ -359,24 +464,39 @@ void io_completed(struct thread_data *td, struct io_u *io_u,
 
                add_clat_sample(td, idx, msec);
                add_bw_sample(td, idx, &icd->time);
 
                add_clat_sample(td, idx, msec);
                add_bw_sample(td, idx, &icd->time);

+               io_u_mark_latency(td, msec);

 
                if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
                        log_io_piece(td, io_u);
 
                icd->bytes_done[idx] += bytes;
 
                if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
                        log_io_piece(td, io_u);
 
                icd->bytes_done[idx] += bytes;

+
+               if (icd->handler) {
+                       ret = icd->handler(io_u);
+                       if (ret && !icd->error)
+                               icd->error = ret;
+               }

        } else
                icd->error = io_u->error;
 }
 
        } else
                icd->error = io_u->error;
 }
 

-void ios_completed(struct thread_data *td, struct io_completion_data *icd)
+static void init_icd(struct io_completion_data *icd, endio_handler *handler,
+                    int nr)

 {
 {

-       struct io_u *io_u;
-       int i;
-

        fio_gettime(&icd->time, NULL);
 
        fio_gettime(&icd->time, NULL);
 

+       icd->handler = handler;
+       icd->nr = nr;
+

        icd->error = 0;
        icd->bytes_done[0] = icd->bytes_done[1] = 0;
        icd->error = 0;
        icd->bytes_done[0] = icd->bytes_done[1] = 0;

+}
+
+static void ios_completed(struct thread_data *td,
+                         struct io_completion_data *icd)
+{
+       struct io_u *io_u;
+       int i;

 
        for (i = 0; i < icd->nr; i++) {
                io_u = td->io_ops->event(td, i);
 
        for (i = 0; i < icd->nr; i++) {
                io_u = td->io_ops->event(td, i);


@@ -385,3 +505,64 @@ void ios_completed(struct thread_data *td, struct io_completion_data *icd)
                put_io_u(td, io_u);
        }
 }
                put_io_u(td, io_u);
        }
 }

+
+long io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
+                       endio_handler *handler)
+{
+       struct io_completion_data icd;
+
+       init_icd(&icd, handler, 1);
+       io_completed(td, io_u, &icd);
+       put_io_u(td, io_u);
+
+       if (!icd.error)
+               return icd.bytes_done[0] + icd.bytes_done[1];
+
+       return -1;
+}
+
+long io_u_queued_complete(struct thread_data *td, int min_events,
+                         endio_handler *handler)
+
+{
+       struct io_completion_data icd;
+       struct timespec *tvp = NULL;
+       int ret;
+
+       if (min_events > 0) {
+               ret = td_io_commit(td);
+               if (ret < 0) {
+                       td_verror(td, -ret);
+                       return ret;
+               }
+       } else {
+               struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
+
+               tvp = &ts;
+       }
+
+       ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
+       if (ret < 0) {
+               td_verror(td, -ret);
+               return ret;
+       } else if (!ret)
+               return ret;
+
+       init_icd(&icd, handler, ret);
+       ios_completed(td, &icd);
+       if (!icd.error)
+               return icd.bytes_done[0] + icd.bytes_done[1];
+
+       return -1;
+}
+
+/*
+ * Call when io_u is really queued, to update the submission latency.
+ */
+void io_u_queued(struct thread_data *td, struct io_u *io_u)
+{
+       unsigned long slat_time;
+
+       slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
+       add_slat_sample(td, io_u->ddir, slat_time);
+}