#include "fio.h"
#include "hash.h"
#include "verify.h"
+#include "trim.h"
#include "lib/rand.h"
struct io_completion_data {
dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
- return (f->file_map[idx] & (1 << bit)) == 0;
+ return (f->file_map[idx] & (1UL << bit)) == 0;
}
/*
struct fio_file *f = io_u->file;
unsigned long long block;
unsigned int blocks, nr_blocks;
+ int busy_check;
block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
blocks = 0;
+ busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
while (nr_blocks) {
- unsigned int this_blocks, mask;
unsigned int idx, bit;
+ unsigned long mask, this_blocks;
/*
* 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(f, block))
+ if (!busy_check) {
+ blocks = nr_blocks;
+ break;
+ }
+ if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
break;
idx = RAND_MAP_IDX(f, block);
do {
if (this_blocks == BLOCKS_PER_MAP)
- mask = -1U;
+ mask = -1UL;
else
- mask = ((1U << this_blocks) - 1) << bit;
+ mask = ((1UL << this_blocks) - 1) << bit;
if (!(f->file_map[idx] & mask))
break;
unsigned long long max_blocks;
unsigned long long max_size;
+ assert(ddir_rw(ddir));
+
/*
* Hmm, should we make sure that ->io_size <= ->real_file_size?
*/
static int get_next_free_block(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir, unsigned long long *b)
{
- unsigned long long min_bs = td->o.rw_min_bs;
+ unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
int i;
+ lastb = last_block(td, f, ddir);
+ if (!lastb)
+ return 1;
+
i = f->last_free_lookup;
- *b = (i * BLOCKS_PER_MAP);
- while ((*b) * min_bs < f->real_file_size &&
- (*b) * min_bs < f->io_size) {
- if (f->file_map[i] != (unsigned int) -1) {
- *b += ffz(f->file_map[i]);
- if (*b > last_block(td, f, ddir))
+ block = i * BLOCKS_PER_MAP;
+ while (block * min_bs < f->real_file_size &&
+ block * min_bs < f->io_size) {
+ if (f->file_map[i] != -1UL) {
+ block += ffz(f->file_map[i]);
+ if (block > lastb)
break;
f->last_free_lookup = i;
+ *b = block;
return 0;
}
- *b += BLOCKS_PER_MAP;
+ block += BLOCKS_PER_MAP;
i++;
}
static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir, unsigned long long *b)
{
- unsigned long long r;
+ unsigned long long r, lastb;
int loops = 5;
+ lastb = last_block(td, f, ddir);
+ if (!lastb)
+ return 1;
+
+ if (f->failed_rands >= 200)
+ goto ffz;
+
do {
- r = os_random_long(&td->random_state);
+ if (td->o.use_os_rand) {
+ r = os_random_long(&td->random_state);
+ *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
+ } else {
+ r = __rand(&td->__random_state);
+ *b = (lastb - 1) * (r / ((unsigned long long) FRAND_MAX + 1.0));
+ }
+
dprint(FD_RANDOM, "off rand %llu\n", r);
- *b = (last_block(td, f, ddir) - 1)
- * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
+
/*
* if we are not maintaining a random map, we are done.
*/
if (!file_randommap(td, f))
- return 0;
+ goto ret_good;
/*
* calculate map offset and check if it's free
*/
if (random_map_free(f, *b))
- return 0;
+ goto ret_good;
dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
*b);
} while (--loops);
+ if (!f->failed_rands++)
+ f->last_free_lookup = 0;
+
/*
* we get here, if we didn't suceed in looking up a block. generate
* a random start offset into the filemap, and find the first free
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;
+ goto ret;
- r = os_random_long(&td->random_state);
+ if (td->o.use_os_rand)
+ r = os_random_long(&td->random_state);
+ else
+ r = __rand(&td->__random_state);
} while (--loops);
/*
* that didn't work either, try exhaustive search from the start
*/
f->last_free_lookup = 0;
+ffz:
+ if (!get_next_free_block(td, f, ddir, b))
+ return 0;
+ f->last_free_lookup = 0;
return get_next_free_block(td, f, ddir, b);
+ret_good:
+ f->failed_rands = 0;
+ret:
+ return 0;
+}
+
+static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
+ enum fio_ddir ddir, unsigned long long *b)
+{
+ if (get_next_rand_offset(td, f, ddir, b)) {
+ dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
+ f->file_name, f->last_pos, f->real_file_size);
+ return 1;
+ }
+
+ return 0;
+}
+
+static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
+ enum fio_ddir ddir, unsigned long long *b)
+{
+ assert(ddir_rw(ddir));
+
+ if (f->last_pos < f->real_file_size) {
+ unsigned long long pos = f->last_pos - f->file_offset;
+
+ if (pos)
+ pos += td->o.ddir_seq_add;
+
+ *b = pos / td->o.min_bs[ddir];
+ return 0;
+ }
+
+ return 1;
+}
+
+static int get_next_block(struct thread_data *td, struct io_u *io_u,
+ enum fio_ddir ddir, int rw_seq, unsigned long long *b)
+{
+ struct fio_file *f = io_u->file;
+ int ret;
+
+ assert(ddir_rw(ddir));
+
+ if (rw_seq) {
+ if (td_random(td))
+ ret = get_next_rand_block(td, f, ddir, b);
+ else
+ ret = get_next_seq_block(td, f, ddir, b);
+ } else {
+ io_u->flags |= IO_U_F_BUSY_OK;
+
+ if (td->o.rw_seq == RW_SEQ_SEQ) {
+ ret = get_next_seq_block(td, f, ddir, b);
+ if (ret)
+ ret = get_next_rand_block(td, f, ddir, b);
+ } else if (td->o.rw_seq == RW_SEQ_IDENT) {
+ if (f->last_start != -1ULL)
+ *b = (f->last_start - f->file_offset)
+ / td->o.min_bs[ddir];
+ else
+ *b = 0;
+ ret = 0;
+ } else {
+ log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
+ ret = 1;
+ }
+ }
+
+ return ret;
}
/*
struct fio_file *f = io_u->file;
unsigned long long b;
enum fio_ddir ddir = io_u->ddir;
+ int rw_seq_hit = 0;
- if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
- td->ddir_nr = td->o.ddir_nr;
+ assert(ddir_rw(ddir));
- if (get_next_rand_offset(td, f, ddir, &b)) {
- dprint(FD_IO, "%s: getting rand offset failed\n",
- f->file_name);
- return 1;
- }
- } else {
- if (f->last_pos >= f->real_file_size) {
- if (!td_random(td) ||
- get_next_rand_offset(td, f, ddir, &b)) {
- dprint(FD_IO, "%s: pos %llu > size %llu\n",
- f->file_name, f->last_pos,
- f->real_file_size);
- return 1;
- }
- } else
- b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
+ if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
+ rw_seq_hit = 1;
+ td->ddir_seq_nr = td->o.ddir_seq_nr;
}
+ if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
+ return 1;
+
io_u->offset = b * td->o.ba[ddir];
if (io_u->offset >= f->io_size) {
dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
return __get_next_offset(td, io_u);
}
+static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
+ unsigned int buflen)
+{
+ struct fio_file *f = io_u->file;
+
+ return io_u->offset + buflen <= f->io_size + td->o.start_offset;
+}
+
static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
{
const int ddir = io_u->ddir;
unsigned int uninitialized_var(buflen);
unsigned int minbs, maxbs;
- long r;
+ unsigned long r, rand_max;
+
+ assert(ddir_rw(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);
+ return minbs;
+
+ if (td->o.use_os_rand)
+ rand_max = OS_RAND_MAX;
+ else
+ rand_max = FRAND_MAX;
+
+ do {
+ if (td->o.use_os_rand)
+ r = os_random_long(&td->bsrange_state);
+ else
+ r = __rand(&td->__bsrange_state);
+
if (!td->o.bssplit_nr[ddir]) {
buflen = 1 + (unsigned int) ((double) maxbs *
- (r / (OS_RAND_MAX + 1.0)));
+ (r / (rand_max + 1.0)));
if (buflen < minbs)
buflen = minbs;
} else {
buflen = bsp->bs;
perc += bsp->perc;
- if (r <= ((OS_RAND_MAX / 100L) * perc))
+ if ((r <= ((rand_max / 100L) * perc)) &&
+ io_u_fits(td, io_u, buflen))
break;
}
}
+
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,
- minbs, ddir);
- buflen = minbs;
- }
+ } while (!io_u_fits(td, io_u, buflen));
return buflen;
}
static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
{
unsigned int v;
- long r;
+ unsigned long r;
+
+ if (td->o.use_os_rand) {
+ r = os_random_long(&td->rwmix_state);
+ v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+ } else {
+ r = __rand(&td->__rwmix_state);
+ v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+ }
- r = os_random_long(&td->rwmix_state);
- v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
if (v <= td->o.rwmix[DDIR_READ])
return DDIR_READ;
struct timeval t;
long usec;
+ assert(ddir_rw(ddir));
+
if (td->rate_pending_usleep[ddir] <= 0)
return ddir;
} else
usec = td->rate_pending_usleep[ddir];
+ /*
+ * We are going to sleep, ensure that we flush anything pending as
+ * not to skew our latency numbers
+ */
+ if (td->cur_depth) {
+ int fio_unused ret;
+
+ ret = io_u_queued_complete(td, td->cur_depth, NULL);
+ }
+
fio_gettime(&t, NULL);
usec_sleep(td, usec);
usec = utime_since_now(&t);
return td->rwmix_ddir;
}
+static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
+{
+ io_u->ddir = get_rw_ddir(td);
+
+ if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+ td->o.barrier_blocks &&
+ !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
+ td->io_issues[DDIR_WRITE])
+ io_u->flags |= IO_U_F_BARRIER;
+}
+
void put_file_log(struct thread_data *td, struct fio_file *f)
{
int ret = put_file(td, f);
td_io_u_lock(td);
__io_u->flags |= IO_U_F_FREE;
- if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
+ if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
td->io_issues[__io_u->ddir]--;
__io_u->flags &= ~IO_U_F_FLIGHT;
if (td->io_ops->flags & FIO_NOIO)
goto out;
- io_u->ddir = get_rw_ddir(td);
+ set_rw_ddir(td, io_u);
/*
- * fsync() or fdatasync(), we are done
+ * fsync() or fdatasync() or trim etc, we are done
*/
- if (ddir_sync(io_u->ddir))
+ if (!ddir_rw(io_u->ddir))
goto out;
/*
static void __io_u_mark_map(unsigned int *map, unsigned int nr)
{
- int index = 0;
+ int idx = 0;
switch (nr) {
default:
- index = 6;
+ idx = 6;
break;
case 33 ... 64:
- index = 5;
+ idx = 5;
break;
case 17 ... 32:
- index = 4;
+ idx = 4;
break;
case 9 ... 16:
- index = 3;
+ idx = 3;
break;
case 5 ... 8:
- index = 2;
+ idx = 2;
break;
case 1 ... 4:
- index = 1;
+ idx = 1;
case 0:
break;
}
- map[index]++;
+ map[idx]++;
}
void io_u_mark_submit(struct thread_data *td, unsigned int nr)
void io_u_mark_depth(struct thread_data *td, unsigned int nr)
{
- int index = 0;
+ int idx = 0;
switch (td->cur_depth) {
default:
- index = 6;
+ idx = 6;
break;
case 32 ... 63:
- index = 5;
+ idx = 5;
break;
case 16 ... 31:
- index = 4;
+ idx = 4;
break;
case 8 ... 15:
- index = 3;
+ idx = 3;
break;
case 4 ... 7:
- index = 2;
+ idx = 2;
break;
case 2 ... 3:
- index = 1;
+ idx = 1;
case 1:
break;
}
- td->ts.io_u_map[index] += nr;
+ td->ts.io_u_map[idx] += nr;
}
static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
{
- int index = 0;
+ int idx = 0;
assert(usec < 1000);
switch (usec) {
case 750 ... 999:
- index = 9;
+ idx = 9;
break;
case 500 ... 749:
- index = 8;
+ idx = 8;
break;
case 250 ... 499:
- index = 7;
+ idx = 7;
break;
case 100 ... 249:
- index = 6;
+ idx = 6;
break;
case 50 ... 99:
- index = 5;
+ idx = 5;
break;
case 20 ... 49:
- index = 4;
+ idx = 4;
break;
case 10 ... 19:
- index = 3;
+ idx = 3;
break;
case 4 ... 9:
- index = 2;
+ idx = 2;
break;
case 2 ... 3:
- index = 1;
+ idx = 1;
case 0 ... 1:
break;
}
- assert(index < FIO_IO_U_LAT_U_NR);
- td->ts.io_u_lat_u[index]++;
+ assert(idx < FIO_IO_U_LAT_U_NR);
+ td->ts.io_u_lat_u[idx]++;
}
static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
{
- int index = 0;
+ int idx = 0;
switch (msec) {
default:
- index = 11;
+ idx = 11;
break;
case 1000 ... 1999:
- index = 10;
+ idx = 10;
break;
case 750 ... 999:
- index = 9;
+ idx = 9;
break;
case 500 ... 749:
- index = 8;
+ idx = 8;
break;
case 250 ... 499:
- index = 7;
+ idx = 7;
break;
case 100 ... 249:
- index = 6;
+ idx = 6;
break;
case 50 ... 99:
- index = 5;
+ idx = 5;
break;
case 20 ... 49:
- index = 4;
+ idx = 4;
break;
case 10 ... 19:
- index = 3;
+ idx = 3;
break;
case 4 ... 9:
- index = 2;
+ idx = 2;
break;
case 2 ... 3:
- index = 1;
+ idx = 1;
case 0 ... 1:
break;
}
- assert(index < FIO_IO_U_LAT_M_NR);
- td->ts.io_u_lat_m[index]++;
+ assert(idx < FIO_IO_U_LAT_M_NR);
+ td->ts.io_u_lat_m[idx]++;
}
static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
int fno;
do {
- long r = os_random_long(&td->next_file_state);
int opened = 0;
+ unsigned long r;
+
+ if (td->o.use_os_rand) {
+ r = os_random_long(&td->next_file_state);
+ fno = (unsigned int) ((double) td->o.nr_files
+ * (r / (OS_RAND_MAX + 1.0)));
+ } else {
+ r = __rand(&td->__next_file_state);
+ fno = (unsigned int) ((double) td->o.nr_files
+ * (r / (FRAND_MAX + 1.0)));
+ }
- fno = (unsigned int) ((double) td->o.nr_files
- * (r / (OS_RAND_MAX + 1.0)));
f = td->files[fno];
if (fio_file_done(f))
continue;
if (io_u) {
assert(io_u->flags & IO_U_F_FREE);
io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
+ io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
io_u->error = 0;
flist_del(&io_u->list);
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)
+static int check_get_trim(struct thread_data *td, struct io_u *io_u)
{
- struct fio_file *f;
- struct io_u *io_u;
+ if (td->o.trim_backlog && td->trim_entries) {
+ int get_trim = 0;
- io_u = __get_io_u(td);
- if (!io_u) {
- dprint(FD_IO, "__get_io_u failed\n");
- return NULL;
+ if (td->trim_batch) {
+ td->trim_batch--;
+ get_trim = 1;
+ } else if (!(td->io_hist_len % td->o.trim_backlog) &&
+ td->last_ddir != DDIR_READ) {
+ td->trim_batch = td->o.trim_batch;
+ if (!td->trim_batch)
+ td->trim_batch = td->o.trim_backlog;
+ get_trim = 1;
+ }
+
+ if (get_trim && !get_next_trim(td, io_u))
+ return 1;
}
+ return 0;
+}
+
+static int check_get_verify(struct thread_data *td, struct io_u *io_u)
+{
if (td->o.verify_backlog && td->io_hist_len) {
int get_verify = 0;
}
if (get_verify && !get_next_verify(td, io_u))
- goto out;
+ return 1;
+ }
+
+ return 0;
+}
+
+/*
+ * 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) {
+ dprint(FD_IO, "__get_io_u failed\n");
+ return NULL;
}
+ if (check_get_verify(td, io_u))
+ goto out;
+ if (check_get_trim(td, io_u))
+ goto out;
+
/*
* from a requeue, io_u already setup
*/
f = io_u->file;
assert(fio_file_open(f));
- if (!ddir_sync(io_u->ddir)) {
+ if (ddir_rw(io_u->ddir)) {
if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
goto err_put;
}
+ f->last_start = io_u->offset;
f->last_pos = io_u->offset + io_u->buflen;
if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
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);
+ else if (io_u->ddir == DDIR_READ) {
+ /*
+ * Reset the buf_filled parameters so next time if the
+ * buffer is used for writes it is refilled.
+ */
+ io_u->buf_filled_len = 0;
+ }
}
/*
io_u->xfer_buflen = io_u->buflen;
out:
+ assert(io_u->file);
if (!td_io_prep(td, io_u)) {
if (!td->o.disable_slat)
fio_gettime(&io_u->start_time, NULL);
void io_u_log_error(struct thread_data *td, struct io_u *io_u)
{
- const char *msg[] = { "read", "write", "sync" };
+ const char *msg[] = { "read", "write", "sync", "datasync",
+ "sync_file_range", "wait", "trim" };
+
+
log_err("fio: io_u error");
td_io_u_lock(td);
assert(io_u->flags & IO_U_F_FLIGHT);
- io_u->flags &= ~IO_U_F_FLIGHT;
+ io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
td_io_u_unlock(td);
if (ddir_sync(io_u->ddir)) {
td->last_was_sync = 0;
td->last_ddir = io_u->ddir;
- if (!io_u->error) {
+ if (!io_u->error && ddir_rw(io_u->ddir)) {
unsigned int bytes = io_u->buflen - io_u->resid;
const enum fio_ddir idx = io_u->ddir;
const enum fio_ddir odx = io_u->ddir ^ 1;
if (!td->o.disable_clat || !td->o.disable_bw)
lusec = utime_since(&io_u->issue_time,
&icd->time);
+ if (!td->o.disable_lat) {
+ unsigned long tusec;
+ tusec = utime_since(&io_u->start_time,
+ &icd->time);
+ add_lat_sample(td, idx, tusec, bytes);
+ }
if (!td->o.disable_clat) {
add_clat_sample(td, idx, lusec, bytes);
io_u_mark_latency(td, lusec);
if (ret && !icd->error)
icd->error = ret;
}
- } else {
+ } else if (io_u->error) {
icd->error = io_u->error;
io_u_log_error(td, io_u);
}
void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
unsigned int max_bs)
{
- long *ptr = io_u->buf;
-
- if (!td->o.zero_buffers) {
- unsigned long r = __rand(&__fio_rand_state);
+ io_u->buf_filled_len = 0;
- if (sizeof(int) != sizeof(*ptr))
- r *= (unsigned long) __rand(&__fio_rand_state);
-
- while ((void *) ptr - io_u->buf < max_bs) {
- *ptr = r;
- ptr++;
- r *= GOLDEN_RATIO_PRIME;
- r >>= 3;
- }
- } else
- memset(ptr, 0, max_bs);
+ if (!td->o.zero_buffers)
+ fill_random_buf(io_u->buf, max_bs);
+ else
+ memset(io_u->buf, 0, max_bs);
}
projects / fio.git / blobdiff