#include <assert.h>
#include "fio.h"
-
-/*
- * Change this define to play with the timeout handling
- */
-#undef FIO_USE_TIMEOUT
+#include "hash.h"
struct io_completion_data {
int nr; /* input */
* 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,
- const unsigned long long block)
+static int random_map_free(struct fio_file *f, const unsigned long long block)
{
- unsigned int idx = RAND_MAP_IDX(td, f, block);
- unsigned int bit = RAND_MAP_BIT(td, f, block);
+ unsigned int idx = RAND_MAP_IDX(f, block);
+ unsigned int bit = RAND_MAP_BIT(f, block);
dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
- return (f->file_map[idx] & (1UL << bit)) == 0;
+ return (f->file_map[idx] & (1 << bit)) == 0;
}
/*
unsigned int min_bs = td->o.rw_min_bs;
struct fio_file *f = io_u->file;
unsigned long long block;
- unsigned int blocks;
- unsigned int nr_blocks;
+ unsigned int blocks, nr_blocks;
block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
- blocks = 0;
nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+ blocks = 0;
- while (blocks < nr_blocks) {
+ while (nr_blocks) {
+ unsigned int this_blocks, mask;
unsigned int idx, bit;
/*
* If we have a mixed random workload, we may
* encounter blocks we already did IO to.
*/
- if ((td->o.ddir_nr == 1) && !random_map_free(td, f, block))
+ if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
+ if (!blocks)
+ blocks = 1;
break;
+ }
- idx = RAND_MAP_IDX(td, f, block);
- bit = RAND_MAP_BIT(td, f, block);
+ idx = RAND_MAP_IDX(f, block);
+ bit = RAND_MAP_BIT(f, block);
fio_assert(td, idx < f->num_maps);
- f->file_map[idx] |= (1UL << bit);
- block++;
- blocks++;
+ this_blocks = nr_blocks;
+ if (this_blocks + bit > BLOCKS_PER_MAP)
+ this_blocks = BLOCKS_PER_MAP - bit;
+
+ if (this_blocks == BLOCKS_PER_MAP)
+ mask = -1U;
+ else
+ mask = ((1U << this_blocks) - 1) << bit;
+
+ f->file_map[idx] |= mask;
+ nr_blocks -= this_blocks;
+ blocks += this_blocks;
+ block += this_blocks;
}
if ((blocks * min_bs) < io_u->buflen)
io_u->buflen = blocks * min_bs;
}
-static inline unsigned long long last_block(struct thread_data *td,
- struct fio_file *f,
- enum fio_ddir ddir)
+static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
+ enum fio_ddir ddir)
{
unsigned long long max_blocks;
+ unsigned long long max_size;
+
+ /*
+ * Hmm, should we make sure that ->io_size <= ->real_file_size?
+ */
+ max_size = f->io_size;
+ if (max_size > f->real_file_size)
+ max_size = f->real_file_size;
- max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
+ max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
if (!max_blocks)
return 0;
- return max_blocks - 1;
+ return max_blocks;
}
/*
i = f->last_free_lookup;
*b = (i * BLOCKS_PER_MAP);
while ((*b) * min_bs < f->real_file_size) {
- if (f->file_map[i] != -1UL) {
- *b += fio_ffz(f->file_map[i]);
+ if (f->file_map[i] != (unsigned int) -1) {
+ *b += ffz(f->file_map[i]);
if (*b > last_block(td, f, ddir))
break;
f->last_free_lookup = i;
r = os_random_long(&td->random_state);
dprint(FD_RANDOM, "off rand %llu\n", r);
*b = (last_block(td, f, ddir) - 1)
- * (r / ((unsigned long long) RAND_MAX + 1.0));
+ * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
/*
* if we are not maintaining a random map, we are done.
/*
* calculate map offset and check if it's free
*/
- if (random_map_free(td, f, *b))
+ if (random_map_free(f, *b))
return 0;
dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
*/
loops = 10;
do {
- f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
+ f->last_free_lookup = (f->num_maps - 1) *
+ (r / (OS_RAND_MAX + 1.0));
if (!get_next_free_block(td, f, ddir, b))
return 0;
b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
}
- io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
+ io_u->offset = b * td->o.min_bs[ddir];
+ if (io_u->offset >= f->io_size) {
+ dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
+ io_u->offset, f->io_size);
+ return 1;
+ }
+
+ io_u->offset += f->file_offset;
if (io_u->offset >= f->real_file_size) {
dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
io_u->offset, f->real_file_size);
return 0;
}
+static inline int is_power_of_2(unsigned int val)
+{
+ return (val != 0 && ((val & (val - 1)) == 0));
+}
+
static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
{
const int ddir = io_u->ddir;
- unsigned int buflen;
+ unsigned int uninitialized_var(buflen);
+ unsigned int minbs, maxbs;
long r;
- if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
- buflen = td->o.min_bs[ddir];
+ minbs = td->o.min_bs[ddir];
+ maxbs = td->o.max_bs[ddir];
+
+ if (minbs == maxbs)
+ buflen = minbs;
else {
r = os_random_long(&td->bsrange_state);
if (!td->o.bssplit_nr) {
- buflen = (unsigned int)
- (1 + (double) (td->o.max_bs[ddir] - 1)
- * r / (RAND_MAX + 1.0));
+ buflen = 1 + (unsigned int) ((double) maxbs *
+ (r / (OS_RAND_MAX + 1.0)));
+ if (buflen < minbs)
+ buflen = minbs;
} else {
long perc = 0;
unsigned int i;
buflen = bsp->bs;
perc += bsp->perc;
- if (r <= ((LONG_MAX / 100L) * perc))
+ if (r <= ((OS_RAND_MAX / 100L) * perc))
break;
}
}
- if (!td->o.bs_unaligned) {
- buflen = (buflen + td->o.min_bs[ddir] - 1)
- & ~(td->o.min_bs[ddir] - 1);
- }
+ if (!td->o.bs_unaligned && is_power_of_2(minbs))
+ buflen = (buflen + minbs - 1) & ~(minbs - 1);
}
if (io_u->offset + buflen > io_u->file->real_file_size) {
dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
- td->o.min_bs[ddir], ddir);
- buflen = td->o.min_bs[ddir];
+ minbs, ddir);
+ buflen = minbs;
}
return buflen;
static void set_rwmix_bytes(struct thread_data *td)
{
- unsigned long issues;
unsigned int diff;
/*
* buffered writes may issue a lot quicker than they complete,
* whereas reads do not.
*/
- issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
diff = td->o.rwmix[td->rwmix_ddir ^ 1];
-
- td->rwmix_issues = td->io_issues[td->rwmix_ddir]
- + (issues * ((100 - diff)) / diff);
+ td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
}
static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
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])
+ v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+ if (v <= td->o.rwmix[DDIR_READ])
return DDIR_READ;
return DDIR_WRITE;
max_bytes = td->this_io_bytes[ddir];
if (max_bytes >=
(td->o.size * td->o.rwmix[ddir] / 100)) {
- if (!td->rw_end_set[ddir])
+ if (!td->rw_end_set[ddir]) {
td->rw_end_set[ddir] = 1;
+ fio_gettime(&td->rw_end[ddir], NULL);
+ }
ddir ^= 1;
}
return DDIR_WRITE;
}
+static void put_file_log(struct thread_data *td, struct fio_file *f)
+{
+ int ret = put_file(td, f);
+
+ if (ret)
+ td_verror(td, ret, "file close");
+}
+
void put_io_u(struct thread_data *td, struct io_u *io_u)
{
assert((io_u->flags & IO_U_F_FREE) == 0);
io_u->flags |= IO_U_F_FREE;
- if (io_u->file) {
- int ret = put_file(td, io_u->file);
-
- if (ret)
- td_verror(td, ret, "file close");
- }
+ if (io_u->file)
+ put_file_log(td, io_u->file);
io_u->file = NULL;
- list_del(&io_u->list);
- list_add(&io_u->list, &td->io_u_freelist);
+ flist_del(&io_u->list);
+ flist_add(&io_u->list, &td->io_u_freelist);
td->cur_depth--;
}
{
struct io_u *__io_u = *io_u;
+ dprint(FD_IO, "requeue %p\n", __io_u);
+
__io_u->flags |= IO_U_F_FREE;
if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
td->io_issues[__io_u->ddir]--;
__io_u->flags &= ~IO_U_F_FLIGHT;
- list_del(&__io_u->list);
- list_add_tail(&__io_u->list, &td->io_u_requeues);
+ flist_del(&__io_u->list);
+ flist_add_tail(&__io_u->list, &td->io_u_requeues);
td->cur_depth--;
*io_u = NULL;
}
return 0;
}
-void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
+static void __io_u_mark_map(unsigned int *map, unsigned int nr)
{
int index = 0;
- if (io_u->ddir == DDIR_SYNC)
- return;
+ switch (nr) {
+ default:
+ index = 6;
+ break;
+ case 33 ... 64:
+ index = 5;
+ break;
+ case 17 ... 32:
+ index = 4;
+ break;
+ case 9 ... 16:
+ index = 3;
+ break;
+ case 5 ... 8:
+ index = 2;
+ break;
+ case 1 ... 4:
+ index = 1;
+ case 0:
+ break;
+ }
+
+ map[index]++;
+}
+
+void io_u_mark_submit(struct thread_data *td, unsigned int nr)
+{
+ __io_u_mark_map(td->ts.io_u_submit, nr);
+ td->ts.total_submit++;
+}
+
+void io_u_mark_complete(struct thread_data *td, unsigned int nr)
+{
+ __io_u_mark_map(td->ts.io_u_complete, nr);
+ td->ts.total_complete++;
+}
+
+void io_u_mark_depth(struct thread_data *td, unsigned int nr)
+{
+ int index = 0;
switch (td->cur_depth) {
default:
break;
}
- td->ts.io_u_map[index]++;
- td->ts.total_io_u[io_u->ddir]++;
+ td->ts.io_u_map[index] += nr;
}
static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
long r = os_random_long(&td->next_file_state);
fno = (unsigned int) ((double) td->o.nr_files
- * (r / (RAND_MAX + 1.0)));
+ * (r / (OS_RAND_MAX + 1.0)));
f = td->files[fno];
if (f->flags & FIO_FILE_DONE)
continue;
if (!fill_io_u(td, io_u))
break;
+ /*
+ * optimization to prevent close/open of the same file. This
+ * way we preserve queueing etc.
+ */
+ if (td->o.nr_files == 1 && td->o.time_based) {
+ put_file_log(td, f);
+ fio_file_reset(f);
+ goto set_file;
+ }
+
/*
* td_io_close() does a put_file() as well, so no need to
* do that here.
{
struct io_u *io_u = NULL;
- if (!list_empty(&td->io_u_requeues))
- io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
+ if (!flist_empty(&td->io_u_requeues))
+ io_u = flist_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 = flist_entry(td->io_u_freelist.next, struct io_u, list);
io_u->buflen = 0;
io_u->resid = 0;
io_u->flags &= ~IO_U_F_FREE;
io_u->error = 0;
- list_del(&io_u->list);
- list_add(&io_u->list, &td->io_u_busylist);
+ flist_del(&io_u->list);
+ flist_add(&io_u->list, &td->io_u_busylist);
td->cur_depth++;
}
if (td->o.verify != VERIFY_NONE)
populate_verify_io_u(td, io_u);
+ else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
+ io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
}
/*
io_u->endpos = io_u->offset + io_u->buflen;
io_u->xfer_buf = io_u->buf;
io_u->xfer_buflen = io_u->buflen;
+
out:
if (!td_io_prep(td, io_u)) {
- fio_gettime(&io_u->start_time, NULL);
+ if (!td->o.disable_slat)
+ fio_gettime(&io_u->start_time, NULL);
return io_u;
}
err_put:
td->io_bytes[idx] += bytes;
td->this_io_bytes[idx] += bytes;
- usec = utime_since(&io_u->issue_time, &icd->time);
+ if (ramp_time_over(td)) {
+ if (!td->o.disable_clat || !td->o.disable_bw)
+ usec = utime_since(&io_u->issue_time,
+ &icd->time);
- add_clat_sample(td, idx, usec);
- add_bw_sample(td, idx, &icd->time);
- io_u_mark_latency(td, usec);
+ if (!td->o.disable_clat) {
+ add_clat_sample(td, idx, usec);
+ io_u_mark_latency(td, usec);
+ }
+ if (!td->o.disable_bw)
+ add_bw_sample(td, idx, &icd->time);
+ }
if (td_write(td) && idx == DDIR_WRITE &&
td->o.do_verify &&
}
}
-static void init_icd(struct io_completion_data *icd, int nr)
+static void init_icd(struct thread_data *td, struct io_completion_data *icd,
+ int nr)
{
- fio_gettime(&icd->time, NULL);
+ if (!td->o.disable_clat || !td->o.disable_bw)
+ fio_gettime(&icd->time, NULL);
icd->nr = nr;
{
struct io_completion_data icd;
- init_icd(&icd, 1);
+ init_icd(td, &icd, 1);
io_completed(td, io_u, &icd);
put_io_u(td, io_u);
/*
* Called to complete min_events number of io for the async engines.
*/
-long io_u_queued_complete(struct thread_data *td, int min_events)
+long io_u_queued_complete(struct thread_data *td, int min_evts)
{
struct io_completion_data icd;
struct timespec *tvp = NULL;
int ret;
struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
- dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
+ dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
- if (!min_events)
+ if (!min_evts)
tvp = &ts;
- ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
+ ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
if (ret < 0) {
td_verror(td, -ret, "td_io_getevents");
return ret;
} else if (!ret)
return ret;
- init_icd(&icd, ret);
+ init_icd(td, &icd, ret);
ios_completed(td, &icd);
if (!icd.error)
return icd.bytes_done[0] + icd.bytes_done[1];
*/
void io_u_queued(struct thread_data *td, struct io_u *io_u)
{
- unsigned long slat_time;
-
- slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
- add_slat_sample(td, io_u->ddir, slat_time);
-}
-
-#ifdef FIO_USE_TIMEOUT
-void io_u_set_timeout(struct thread_data *td)
-{
- assert(td->cur_depth);
+ if (!td->o.disable_slat) {
+ unsigned long slat_time;
- td->timer.it_interval.tv_sec = 0;
- td->timer.it_interval.tv_usec = 0;
- td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
- td->timer.it_value.tv_usec = 0;
- setitimer(ITIMER_REAL, &td->timer, NULL);
- fio_gettime(&td->timeout_end, NULL);
-}
-
-static void io_u_dump(struct io_u *io_u)
-{
- unsigned long t_start = mtime_since_now(&io_u->start_time);
- unsigned long t_issue = mtime_since_now(&io_u->issue_time);
-
- log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
- log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
- io_u->xfer_buf, io_u->buflen,
- io_u->xfer_buflen,
- io_u->offset);
- log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
-}
-#else
-void io_u_set_timeout(struct thread_data fio_unused *td)
-{
+ slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+ add_slat_sample(td, io_u->ddir, slat_time);
+ }
}
-#endif
-#ifdef FIO_USE_TIMEOUT
-static void io_u_timeout_handler(int fio_unused sig)
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+ unsigned int max_bs)
{
- struct thread_data *td, *__td;
- pid_t pid = getpid();
- struct list_head *entry;
- struct io_u *io_u;
- int i;
-
- log_err("fio: io_u timeout\n");
+ long *ptr = io_u->buf;
- /*
- * TLS would be nice...
- */
- td = NULL;
- for_each_td(__td, i) {
- if (__td->pid == pid) {
- td = __td;
- break;
+ if (!td->o.zero_buffers) {
+ while ((void *) ptr - io_u->buf < max_bs) {
+ *ptr = rand() * GOLDEN_RATIO_PRIME;
+ ptr++;
}
- }
-
- if (!td) {
- log_err("fio: io_u timeout, can't find job\n");
- exit(1);
- }
-
- if (!td->cur_depth) {
- log_err("fio: timeout without pending work?\n");
- return;
- }
-
- log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
-
- list_for_each(entry, &td->io_u_busylist) {
- io_u = list_entry(entry, struct io_u, list);
-
- io_u_dump(io_u);
- }
-
- td_verror(td, ETIMEDOUT, "io_u timeout");
- exit(1);
-}
-#endif
-
-void io_u_init_timeout(void)
-{
-#ifdef FIO_USE_TIMEOUT
- signal(SIGALRM, io_u_timeout_handler);
-#endif
+ } else
+ memset(ptr, 0, max_bs);
}
projects / fio.git / blobdiff