[PATCH] Split the io_u io handling out of fio.c
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
diff --git a/Makefile b/Makefile
index 111d5cd..312f8de 100644
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,7 @@
PROGS = fio
SCRIPTS = fio_generate_plots
OBJS = fio.o ioengines.o init.o stat.o log.o time.o md5.o crc32.o \
- filesetup.o eta.o verify.o memory.o
+ filesetup.o eta.o verify.o memory.o io_u.o
INSTALL = install
prefix = /usr/local
diff --git a/fio.c b/fio.c
index 8d2a8e5..818394a 100644
--- a/fio.c
+++ b/fio.c
@@ -82,134 +82,6 @@
}
/*
- * 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, struct fio_file *f,
- unsigned long long block)
-{
- unsigned int idx = RAND_MAP_IDX(td, f, block);
- unsigned int bit = RAND_MAP_BIT(td, f, block);
-
- return (f->file_map[idx] & (1UL << bit)) == 0;
-}
-
-/*
- * Return the next free block in the map.
- */
-static int get_next_free_block(struct thread_data *td, struct fio_file *f,
- unsigned long long *b)
-{
- int i;
-
- *b = 0;
- i = 0;
- while ((*b) * td->min_bs < f->file_size) {
- if (f->file_map[i] != -1UL) {
- *b += ffz(f->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 fio_file *f,
- 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, f, block))
- break;
-
- idx = RAND_MAP_IDX(td, f, block);
- bit = RAND_MAP_BIT(td, f, block);
-
- assert(idx < f->num_maps);
-
- f->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, struct fio_file *f,
- 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 + (f->file_offset / td->min_bs);
- loops--;
- } while (!random_map_free(td, f, rb) && loops);
-
- if (!loops) {
- if (get_next_free_block(td, f, &b))
- return 1;
- }
- } else
- b = f->last_pos / td->min_bs;
-
- *offset = (b * td->min_bs) + f->file_offset;
- if (*offset > f->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]) {
- /*
- * 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)
- return 0;
-
- buflen = td->io_size - td->this_io_bytes[td->ddir];
- }
-
- return buflen;
-}
-
-/*
* Check if we are above the minimum rate given.
*/
static int check_min_rate(struct thread_data *td, struct timeval *now)
@@ -254,162 +126,6 @@
return 0;
}
-/*
- * 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) {
- unsigned 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;
-}
-
-static int td_io_prep(struct thread_data *td, struct io_u *io_u)
-{
- if (td->io_ops->prep && td->io_ops->prep(td, io_u))
- return 1;
-
- return 0;
-}
-
-void put_io_u(struct thread_data *td, struct io_u *io_u)
-{
- io_u->file = NULL;
- list_del(&io_u->list);
- list_add(&io_u->list, &td->io_u_freelist);
- td->cur_depth--;
-}
-
-static int fill_io_u(struct thread_data *td, struct fio_file *f,
- struct io_u *io_u)
-{
- /*
- * If using an iolog, grab next piece if any available.
- */
- if (td->read_iolog)
- return read_iolog_get(td, io_u);
-
- /*
- * No log, let the seq/rand engine retrieve the next position.
- */
- if (!get_next_offset(td, f, &io_u->offset)) {
- io_u->buflen = get_next_buflen(td);
-
- if (io_u->buflen) {
- io_u->ddir = get_rw_ddir(td);
-
- /*
- * If using a write iolog, store this entry.
- */
- if (td->write_iolog)
- write_iolog_put(td, io_u);
-
- io_u->file = f;
- return 0;
- }
- }
-
- 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;
-}
-
-/*
- * 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.
- */
-static 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;
-
- if (td->zone_bytes >= td->zone_size) {
- td->zone_bytes = 0;
- f->last_pos += td->zone_skip;
- }
-
- if (fill_io_u(td, f, io_u)) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- if (io_u->buflen + io_u->offset > f->file_size) {
- if (td->io_ops->flags & FIO_RAWIO) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- io_u->buflen = f->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, f, io_u);
-
- f->last_pos += io_u->buflen;
-
- 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;
- }
-
- gettimeofday(&io_u->start_time, NULL);
- return io_u;
-}
-
static inline void td_set_runstate(struct thread_data *td, int runstate)
{
td->runstate = runstate;
@@ -436,75 +152,6 @@
return f;
}
-static int td_io_sync(struct thread_data *td, struct fio_file *f)
-{
- if (td->io_ops->sync)
- return td->io_ops->sync(td, f);
-
- return 0;
-}
-
-static int td_io_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- return td->io_ops->getevents(td, min, max, t);
-}
-
-static int td_io_queue(struct thread_data *td, struct io_u *io_u)
-{
- gettimeofday(&io_u->issue_time, NULL);
-
- return td->io_ops->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_ops->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.
@@ -723,7 +370,6 @@
min_evts = 1;
}
-
ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
if (ret < 0) {
td_verror(td, -ret);
diff --git a/fio.h b/fio.h
index b9ae15b..03a2378 100644
--- a/fio.h
+++ b/fio.h
@@ -335,9 +335,6 @@
#define td_verror(td, err) __td_verror((td), (err), strerror((err)))
#define td_vmsg(td, err, msg) __td_verror((td), (err), (msg))
-extern struct io_u *__get_io_u(struct thread_data *);
-extern void put_io_u(struct thread_data *, struct io_u *);
-
extern int rate_quit;
extern int exitall_on_terminate;
extern int thread_number;
@@ -488,6 +485,24 @@
extern void free_io_mem(struct thread_data *);
/*
+ * io unit handling
+ */
+#define queue_full(td) list_empty(&(td)->io_u_freelist)
+extern struct io_u *__get_io_u(struct thread_data *);
+extern struct io_u *get_io_u(struct thread_data *, struct fio_file *);
+extern void put_io_u(struct thread_data *, struct io_u *);
+extern void ios_completed(struct thread_data *, struct io_completion_data *);
+extern void io_completed(struct thread_data *, struct io_u *, struct io_completion_data *);
+
+/*
+ * io engine entry points
+ */
+extern int td_io_prep(struct thread_data *, struct io_u *);
+extern int td_io_queue(struct thread_data *, struct io_u *);
+extern int td_io_sync(struct thread_data *, struct fio_file *);
+extern int td_io_getevents(struct thread_data *, int, int, struct timespec *);
+
+/*
* This is a pretty crappy semaphore implementation, but with the use that fio
* has (just signalling start/go conditions), it doesn't have to be better.
* Naturally this would not work for any type of contended semaphore or
diff --git a/io_u.c b/io_u.c
new file mode 100644
index 0000000..baf5d47
--- /dev/null
+++ b/io_u.c
@@ -0,0 +1,331 @@
+#include <unistd.h>
+#include <fcntl.h>
+#include <string.h>
+#include <signal.h>
+#include <time.h>
+#include <assert.h>
+
+#include "fio.h"
+#include "os.h"
+
+/*
+ * 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, struct fio_file *f,
+ unsigned long long block)
+{
+ unsigned int idx = RAND_MAP_IDX(td, f, block);
+ unsigned int bit = RAND_MAP_BIT(td, f, block);
+
+ return (f->file_map[idx] & (1UL << bit)) == 0;
+}
+
+/*
+ * Return the next free block in the map.
+ */
+static int get_next_free_block(struct thread_data *td, struct fio_file *f,
+ unsigned long long *b)
+{
+ int i;
+
+ *b = 0;
+ i = 0;
+ while ((*b) * td->min_bs < f->file_size) {
+ if (f->file_map[i] != -1UL) {
+ *b += ffz(f->file_map[i]);
+ return 0;
+ }
+
+ *b += BLOCKS_PER_MAP;
+ i++;
+ }
+
+ return 1;
+}
+
+/*
+ * 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, struct fio_file *f,
+ 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 + (f->file_offset / td->min_bs);
+ loops--;
+ } while (!random_map_free(td, f, rb) && loops);
+
+ if (!loops) {
+ if (get_next_free_block(td, f, &b))
+ return 1;
+ }
+ } else
+ b = f->last_pos / td->min_bs;
+
+ *offset = (b * td->min_bs) + f->file_offset;
+ if (*offset > f->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]) {
+ /*
+ * 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)
+ return 0;
+
+ buflen = td->io_size - td->this_io_bytes[td->ddir];
+ }
+
+ return buflen;
+}
+
+/*
+ * 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) {
+ unsigned 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;
+}
+
+/*
+ * Mark a given offset as used in the map.
+ */
+static void mark_random_map(struct thread_data *td, struct fio_file *f,
+ 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, f, block))
+ break;
+
+ idx = RAND_MAP_IDX(td, f, block);
+ bit = RAND_MAP_BIT(td, f, block);
+
+ assert(idx < f->num_maps);
+
+ f->file_map[idx] |= (1UL << bit);
+ block++;
+ blocks++;
+ }
+
+ if ((blocks * td->min_bs) < io_u->buflen)
+ io_u->buflen = blocks * td->min_bs;
+}
+
+void put_io_u(struct thread_data *td, struct io_u *io_u)
+{
+ io_u->file = NULL;
+ list_del(&io_u->list);
+ list_add(&io_u->list, &td->io_u_freelist);
+ td->cur_depth--;
+}
+
+static int fill_io_u(struct thread_data *td, struct fio_file *f,
+ struct io_u *io_u)
+{
+ /*
+ * If using an iolog, grab next piece if any available.
+ */
+ if (td->read_iolog)
+ return read_iolog_get(td, io_u);
+
+ /*
+ * No log, let the seq/rand engine retrieve the next position.
+ */
+ if (!get_next_offset(td, f, &io_u->offset)) {
+ io_u->buflen = get_next_buflen(td);
+
+ if (io_u->buflen) {
+ io_u->ddir = get_rw_ddir(td);
+
+ /*
+ * If using a write iolog, store this entry.
+ */
+ if (td->write_iolog)
+ write_iolog_put(td, io_u);
+
+ io_u->file = f;
+ return 0;
+ }
+ }
+
+ return 1;
+}
+
+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;
+}
+
+/*
+ * 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 *io_u;
+
+ io_u = __get_io_u(td);
+ if (!io_u)
+ return NULL;
+
+ if (td->zone_bytes >= td->zone_size) {
+ td->zone_bytes = 0;
+ f->last_pos += td->zone_skip;
+ }
+
+ if (fill_io_u(td, f, io_u)) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
+ if (io_u->buflen + io_u->offset > f->file_size) {
+ if (td->io_ops->flags & FIO_RAWIO) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
+ io_u->buflen = f->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, f, io_u);
+
+ f->last_pos += io_u->buflen;
+
+ 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;
+ }
+
+ gettimeofday(&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)
+{
+ 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;
+}
+
+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_ops->event(td, i);
+
+ io_completed(td, io_u, icd);
+ put_io_u(td, io_u);
+ }
+}
+
+
diff --git a/ioengines.c b/ioengines.c
index abc3853..5a32165 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -69,3 +69,32 @@
free(td->io_ops);
td->io_ops = NULL;
}
+
+int td_io_prep(struct thread_data *td, struct io_u *io_u)
+{
+ if (td->io_ops->prep && td->io_ops->prep(td, io_u))
+ return 1;
+
+ return 0;
+}
+
+int td_io_sync(struct thread_data *td, struct fio_file *f)
+{
+ if (td->io_ops->sync)
+ return td->io_ops->sync(td, f);
+
+ return 0;
+}
+
+int td_io_getevents(struct thread_data *td, int min, int max,
+ struct timespec *t)
+{
+ return td->io_ops->getevents(td, min, max, t);
+}
+
+int td_io_queue(struct thread_data *td, struct io_u *io_u)
+{
+ gettimeofday(&io_u->issue_time, NULL);
+
+ return td->io_ops->queue(td, io_u);
+}