while (blocks < nr_blocks) {
unsigned int idx, bit;
- if (!random_map_free(td, f, block))
+ /*
+ * If we have a mixed random workload, we may
+ * encounter blocks we already did IO to.
+ */
+ if (!td->o.ddir_nr && !random_map_free(td, f, block))
break;
idx = RAND_MAP_IDX(td, f, block);
return 1;
}
+static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+ int ddir, unsigned long long *b)
+{
+ unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
+ unsigned long long r, rb;
+ int loops = 5;
+
+ do {
+ r = os_random_long(&td->random_state);
+ if (!max_blocks)
+ *b = 0;
+ else
+ *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
+ if (td->o.norandommap)
+ break;
+ rb = *b + (f->file_offset / td->o.min_bs[ddir]);
+ loops--;
+ } while (!random_map_free(td, f, rb) && loops);
+
+ /*
+ * 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;
+
+ return 0;
+}
+
/*
* 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
{
struct fio_file *f = io_u->file;
const int ddir = io_u->ddir;
- unsigned long long b, rb;
- long r;
+ unsigned long long b;
- if (td_random(td)) {
- unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir];
- int loops = 5;
-
- do {
- r = os_random_long(&td->random_state);
- if (!max_blocks)
- b = 0;
- else
- b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
- if (td->o.norandommap)
- break;
- rb = b + (f->file_offset / td->o.min_bs[ddir]);
- loops--;
- } while (!random_map_free(td, f, rb) && loops);
+ if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
+ td->ddir_nr = td->o.ddir_nr;
- /*
- * 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))
+ if (get_next_rand_offset(td, f, ddir, &b))
+ return 1;
+ } else {
+ if (f->last_pos >= f->real_file_size)
return 1;
- } else
+
b = f->last_pos / td->o.min_bs[ddir];
+ }
io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
if (io_u->offset >= f->real_file_size)
static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
{
- struct fio_file *f = io_u->file;
const int ddir = io_u->ddir;
unsigned int buflen;
long r;
buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
}
- while (buflen + io_u->offset > f->real_file_size) {
- if (buflen == td->o.min_bs[ddir]) {
- if (!td->o.odirect) {
- assert(io_u->offset <= f->real_file_size);
- buflen = f->real_file_size - io_u->offset;
- return buflen;
- }
- return 0;
- }
+ return buflen;
+}
- buflen = td->o.min_bs[ddir];
- }
+static void set_rwmix_bytes(struct thread_data *td)
+{
+ unsigned long long rbytes;
+ unsigned int diff;
- return buflen;
+ /*
+ * we do time or byte based switch. this is needed because
+ * buffered writes may issue a lot quicker than they complete,
+ * whereas reads do not.
+ */
+ rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
+ diff = td->o.rwmix[td->rwmix_ddir ^ 1];
+
+ td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
+}
+
+static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
+{
+ 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->o.rwmix[DDIR_READ])
+ return DDIR_READ;
+
+ return DDIR_WRITE;
}
/*
if (td_rw(td)) {
struct timeval now;
unsigned long elapsed;
+ unsigned int cycle;
fio_gettime(&now, NULL);
elapsed = mtime_since_now(&td->rwmix_switch);
+ /*
+ * if this is the first cycle, make it shorter
+ */
+ cycle = td->o.rwmixcycle;
+ if (!td->rwmix_bytes)
+ cycle /= 10;
+
/*
* Check if it's time to seed a new data direction.
*/
- if (elapsed >= td->o.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->o.rwmixread)
- td->rwmix_ddir = DDIR_READ;
- else
- td->rwmix_ddir = DDIR_WRITE;
+ if (elapsed >= cycle ||
+ td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
+ unsigned long long max_bytes;
+ enum fio_ddir ddir;
+
+ /*
+ * Put a top limit on how many bytes we do for
+ * one data direction, to avoid overflowing the
+ * ranges too much
+ */
+ ddir = get_rand_ddir(td);
+ max_bytes = td->this_io_bytes[ddir];
+ if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
+ if (!td->rw_end_set[ddir]) {
+ td->rw_end_set[ddir] = 1;
+ memcpy(&td->rw_end[ddir], &now, sizeof(now));
+ }
+ ddir ^= 1;
+ }
+
+ if (ddir != td->rwmix_ddir)
+ set_rwmix_bytes(td);
+
+ td->rwmix_ddir = ddir;
memcpy(&td->rwmix_switch, &now, sizeof(now));
}
return td->rwmix_ddir;
!(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
td->io_issues[DDIR_WRITE] && should_fsync(td)) {
io_u->ddir = DDIR_SYNC;
- return 0;
+ goto out;
}
io_u->ddir = get_rw_ddir(td);
/*
* mark entry before potentially trimming io_u
*/
- if (!td->o.read_iolog && td_random(td) && !td->o.norandommap)
+ if (td_random(td) && !td->o.norandommap)
mark_random_map(td, io_u);
/*
* If using a write iolog, store this entry.
*/
+out:
if (td->o.write_iolog_file)
write_iolog_put(td, io_u);
fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
f = &td->files[fno];
+ if (f->flags & FIO_FILE_DONE)
+ continue;
if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
return f;
if (td->next_file >= td->o.nr_files)
td->next_file = 0;
+ if (f->flags & FIO_FILE_DONE) {
+ f = NULL;
+ continue;
+ }
+
if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
break;
*/
io_u->file = NULL;
td_io_close_file(td, f);
+ f->flags |= FIO_FILE_DONE;
/*
* probably not the right place to do this, but see
f->last_pos += td->o.zone_skip;
}
- if (io_u->buflen + io_u->offset > f->real_file_size) {
- if (td->io_ops->flags & FIO_RAWIO) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- io_u->buflen = f->real_file_size - io_u->offset;
- }
-
if (io_u->ddir != DDIR_SYNC) {
if (!io_u->buflen) {
put_io_u(td, io_u);
add_bw_sample(td, idx, &icd->time);
io_u_mark_latency(td, msec);
- if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
+ if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
+ td->o.verify != VERIFY_NONE)
log_io_piece(td, io_u);
icd->bytes_done[idx] += bytes;
if (io_u->end_io) {
- ret = io_u->end_io(io_u);
+ ret = io_u->end_io(td, io_u);
if (ret && !icd->error)
icd->error = ret;
}
struct io_completion_data icd;
struct timespec *tvp = NULL;
int ret;
+ struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
- if (!min_events) {
- struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
-
+ if (!min_events)
tvp = &ts;
- }
ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
if (ret < 0) {