io_u.c revision ce35b1ece06ce27ebff111e47c2a6610c311a92b
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13#include "lib/axmap.h"
14
15struct io_completion_data {
16	int nr;				/* input */
17
18	int error;			/* output */
19	uint64_t bytes_done[DDIR_RWDIR_CNT];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const uint64_t block)
28{
29	return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37	unsigned int min_bs = td->o.rw_min_bs;
38	struct fio_file *f = io_u->file;
39	unsigned int nr_blocks;
40	uint64_t block;
41
42	block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
43	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45	if (!(io_u->flags & IO_U_F_BUSY_OK))
46		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47
48	if ((nr_blocks * min_bs) < io_u->buflen)
49		io_u->buflen = nr_blocks * min_bs;
50}
51
52static uint64_t last_block(struct thread_data *td, struct fio_file *f,
53			   enum fio_ddir ddir)
54{
55	uint64_t max_blocks;
56	uint64_t max_size;
57
58	assert(ddir_rw(ddir));
59
60	/*
61	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
62	 */
63	max_size = f->io_size;
64	if (max_size > f->real_file_size)
65		max_size = f->real_file_size;
66
67	if (td->o.zone_range)
68		max_size = td->o.zone_range;
69
70	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
71	if (!max_blocks)
72		return 0;
73
74	return max_blocks;
75}
76
77struct rand_off {
78	struct flist_head list;
79	uint64_t off;
80};
81
82static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
83				  enum fio_ddir ddir, uint64_t *b)
84{
85	uint64_t r, lastb;
86
87	lastb = last_block(td, f, ddir);
88	if (!lastb)
89		return 1;
90
91	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
92		uint64_t rmax;
93
94		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
95
96		if (td->o.use_os_rand) {
97			rmax = OS_RAND_MAX;
98			r = os_random_long(&td->random_state);
99		} else {
100			rmax = FRAND_MAX;
101			r = __rand(&td->__random_state);
102		}
103
104		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
105
106		*b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
107	} else {
108		uint64_t off = 0;
109
110		if (lfsr_next(&f->lfsr, &off, lastb))
111			return 1;
112
113		*b = off;
114	}
115
116	/*
117	 * if we are not maintaining a random map, we are done.
118	 */
119	if (!file_randommap(td, f))
120		goto ret;
121
122	/*
123	 * calculate map offset and check if it's free
124	 */
125	if (random_map_free(f, *b))
126		goto ret;
127
128	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
129						(unsigned long long) *b);
130
131	*b = axmap_next_free(f->io_axmap, *b);
132	if (*b == (uint64_t) -1ULL)
133		return 1;
134ret:
135	return 0;
136}
137
138static int __get_next_rand_offset_zipf(struct thread_data *td,
139				       struct fio_file *f, enum fio_ddir ddir,
140				       uint64_t *b)
141{
142	*b = zipf_next(&f->zipf);
143	return 0;
144}
145
146static int __get_next_rand_offset_pareto(struct thread_data *td,
147					 struct fio_file *f, enum fio_ddir ddir,
148					 uint64_t *b)
149{
150	*b = pareto_next(&f->zipf);
151	return 0;
152}
153
154static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
155{
156	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
157	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
158
159	return r1->off - r2->off;
160}
161
162static int get_off_from_method(struct thread_data *td, struct fio_file *f,
163			       enum fio_ddir ddir, uint64_t *b)
164{
165	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
166		return __get_next_rand_offset(td, f, ddir, b);
167	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
168		return __get_next_rand_offset_zipf(td, f, ddir, b);
169	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
170		return __get_next_rand_offset_pareto(td, f, ddir, b);
171
172	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
173	return 1;
174}
175
176/*
177 * Sort the reads for a verify phase in batches of verifysort_nr, if
178 * specified.
179 */
180static inline int should_sort_io(struct thread_data *td)
181{
182	if (!td->o.verifysort_nr || !td->o.do_verify)
183		return 0;
184	if (!td_random(td))
185		return 0;
186	if (td->runstate != TD_VERIFYING)
187		return 0;
188	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
189		return 0;
190
191	return 1;
192}
193
194static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
195{
196	unsigned int v;
197	unsigned long r;
198
199	if (td->o.perc_rand[ddir] == 100)
200		return 1;
201
202	if (td->o.use_os_rand) {
203		r = os_random_long(&td->seq_rand_state[ddir]);
204		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
205	} else {
206		r = __rand(&td->__seq_rand_state[ddir]);
207		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
208	}
209
210	return v <= td->o.perc_rand[ddir];
211}
212
213static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
214				enum fio_ddir ddir, uint64_t *b)
215{
216	struct rand_off *r;
217	int i, ret = 1;
218
219	if (!should_sort_io(td))
220		return get_off_from_method(td, f, ddir, b);
221
222	if (!flist_empty(&td->next_rand_list)) {
223		struct rand_off *r;
224fetch:
225		r = flist_entry(td->next_rand_list.next, struct rand_off, list);
226		flist_del(&r->list);
227		*b = r->off;
228		free(r);
229		return 0;
230	}
231
232	for (i = 0; i < td->o.verifysort_nr; i++) {
233		r = malloc(sizeof(*r));
234
235		ret = get_off_from_method(td, f, ddir, &r->off);
236		if (ret) {
237			free(r);
238			break;
239		}
240
241		flist_add(&r->list, &td->next_rand_list);
242	}
243
244	if (ret && !i)
245		return ret;
246
247	assert(!flist_empty(&td->next_rand_list));
248	flist_sort(NULL, &td->next_rand_list, flist_cmp);
249	goto fetch;
250}
251
252static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
253			       enum fio_ddir ddir, uint64_t *b)
254{
255	if (!get_next_rand_offset(td, f, ddir, b))
256		return 0;
257
258	if (td->o.time_based) {
259		fio_file_reset(td, f);
260		if (!get_next_rand_offset(td, f, ddir, b))
261			return 0;
262	}
263
264	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
265			f->file_name, (unsigned long long) f->last_pos,
266			(unsigned long long) f->real_file_size);
267	return 1;
268}
269
270static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
271			       enum fio_ddir ddir, uint64_t *offset)
272{
273	assert(ddir_rw(ddir));
274
275	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
276		f->last_pos = f->last_pos - f->io_size;
277
278	if (f->last_pos < f->real_file_size) {
279		uint64_t pos;
280
281		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
282			f->last_pos = f->real_file_size;
283
284		pos = f->last_pos - f->file_offset;
285		if (pos)
286			pos += td->o.ddir_seq_add;
287
288		*offset = pos;
289		return 0;
290	}
291
292	return 1;
293}
294
295static int get_next_block(struct thread_data *td, struct io_u *io_u,
296			  enum fio_ddir ddir, int rw_seq,
297			  unsigned int *is_random)
298{
299	struct fio_file *f = io_u->file;
300	uint64_t b, offset;
301	int ret;
302
303	assert(ddir_rw(ddir));
304
305	b = offset = -1ULL;
306
307	if (rw_seq) {
308		if (td_random(td)) {
309			if (should_do_random(td, ddir)) {
310				ret = get_next_rand_block(td, f, ddir, &b);
311				*is_random = 1;
312			} else {
313				*is_random = 0;
314				io_u->flags |= IO_U_F_BUSY_OK;
315				ret = get_next_seq_offset(td, f, ddir, &offset);
316				if (ret)
317					ret = get_next_rand_block(td, f, ddir, &b);
318			}
319		} else {
320			*is_random = 0;
321			ret = get_next_seq_offset(td, f, ddir, &offset);
322		}
323	} else {
324		io_u->flags |= IO_U_F_BUSY_OK;
325		*is_random = 0;
326
327		if (td->o.rw_seq == RW_SEQ_SEQ) {
328			ret = get_next_seq_offset(td, f, ddir, &offset);
329			if (ret) {
330				ret = get_next_rand_block(td, f, ddir, &b);
331				*is_random = 0;
332			}
333		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
334			if (f->last_start != -1ULL)
335				offset = f->last_start - f->file_offset;
336			else
337				offset = 0;
338			ret = 0;
339		} else {
340			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
341			ret = 1;
342		}
343	}
344
345	if (!ret) {
346		if (offset != -1ULL)
347			io_u->offset = offset;
348		else if (b != -1ULL)
349			io_u->offset = b * td->o.ba[ddir];
350		else {
351			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
352			ret = 1;
353		}
354	}
355
356	return ret;
357}
358
359/*
360 * For random io, generate a random new block and see if it's used. Repeat
361 * until we find a free one. For sequential io, just return the end of
362 * the last io issued.
363 */
364static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
365			     unsigned int *is_random)
366{
367	struct fio_file *f = io_u->file;
368	enum fio_ddir ddir = io_u->ddir;
369	int rw_seq_hit = 0;
370
371	assert(ddir_rw(ddir));
372
373	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
374		rw_seq_hit = 1;
375		td->ddir_seq_nr = td->o.ddir_seq_nr;
376	}
377
378	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
379		return 1;
380
381	if (io_u->offset >= f->io_size) {
382		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
383					(unsigned long long) io_u->offset,
384					(unsigned long long) f->io_size);
385		return 1;
386	}
387
388	io_u->offset += f->file_offset;
389	if (io_u->offset >= f->real_file_size) {
390		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
391					(unsigned long long) io_u->offset,
392					(unsigned long long) f->real_file_size);
393		return 1;
394	}
395
396	return 0;
397}
398
399static int get_next_offset(struct thread_data *td, struct io_u *io_u,
400			   unsigned int *is_random)
401{
402	if (td->flags & TD_F_PROFILE_OPS) {
403		struct prof_io_ops *ops = &td->prof_io_ops;
404
405		if (ops->fill_io_u_off)
406			return ops->fill_io_u_off(td, io_u, is_random);
407	}
408
409	return __get_next_offset(td, io_u, is_random);
410}
411
412static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
413			    unsigned int buflen)
414{
415	struct fio_file *f = io_u->file;
416
417	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
418}
419
420static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
421				      unsigned int is_random)
422{
423	int ddir = io_u->ddir;
424	unsigned int buflen = 0;
425	unsigned int minbs, maxbs;
426	unsigned long r, rand_max;
427
428	assert(ddir_rw(io_u->ddir));
429
430	if (td->o.bs_is_seq_rand)
431		ddir = is_random ? DDIR_WRITE: DDIR_READ;
432	else
433		ddir = io_u->ddir;
434
435	minbs = td->o.min_bs[ddir];
436	maxbs = td->o.max_bs[ddir];
437
438	if (minbs == maxbs)
439		return minbs;
440
441	/*
442	 * If we can't satisfy the min block size from here, then fail
443	 */
444	if (!io_u_fits(td, io_u, minbs))
445		return 0;
446
447	if (td->o.use_os_rand)
448		rand_max = OS_RAND_MAX;
449	else
450		rand_max = FRAND_MAX;
451
452	do {
453		if (td->o.use_os_rand)
454			r = os_random_long(&td->bsrange_state);
455		else
456			r = __rand(&td->__bsrange_state);
457
458		if (!td->o.bssplit_nr[ddir]) {
459			buflen = 1 + (unsigned int) ((double) maxbs *
460					(r / (rand_max + 1.0)));
461			if (buflen < minbs)
462				buflen = minbs;
463		} else {
464			long perc = 0;
465			unsigned int i;
466
467			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
468				struct bssplit *bsp = &td->o.bssplit[ddir][i];
469
470				buflen = bsp->bs;
471				perc += bsp->perc;
472				if ((r <= ((rand_max / 100L) * perc)) &&
473				    io_u_fits(td, io_u, buflen))
474					break;
475			}
476		}
477
478		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
479			buflen = (buflen + td->o.verify_interval - 1) &
480				~(td->o.verify_interval - 1);
481
482		if (!td->o.bs_unaligned && is_power_of_2(minbs))
483			buflen = (buflen + minbs - 1) & ~(minbs - 1);
484
485	} while (!io_u_fits(td, io_u, buflen));
486
487	return buflen;
488}
489
490static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
491				    unsigned int is_random)
492{
493	if (td->flags & TD_F_PROFILE_OPS) {
494		struct prof_io_ops *ops = &td->prof_io_ops;
495
496		if (ops->fill_io_u_size)
497			return ops->fill_io_u_size(td, io_u, is_random);
498	}
499
500	return __get_next_buflen(td, io_u, is_random);
501}
502
503static void set_rwmix_bytes(struct thread_data *td)
504{
505	unsigned int diff;
506
507	/*
508	 * we do time or byte based switch. this is needed because
509	 * buffered writes may issue a lot quicker than they complete,
510	 * whereas reads do not.
511	 */
512	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
513	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
514}
515
516static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
517{
518	unsigned int v;
519	unsigned long r;
520
521	if (td->o.use_os_rand) {
522		r = os_random_long(&td->rwmix_state);
523		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
524	} else {
525		r = __rand(&td->__rwmix_state);
526		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
527	}
528
529	if (v <= td->o.rwmix[DDIR_READ])
530		return DDIR_READ;
531
532	return DDIR_WRITE;
533}
534
535void io_u_quiesce(struct thread_data *td)
536{
537	/*
538	 * We are going to sleep, ensure that we flush anything pending as
539	 * not to skew our latency numbers.
540	 *
541	 * Changed to only monitor 'in flight' requests here instead of the
542	 * td->cur_depth, b/c td->cur_depth does not accurately represent
543	 * io's that have been actually submitted to an async engine,
544	 * and cur_depth is meaningless for sync engines.
545	 */
546	while (td->io_u_in_flight) {
547		int fio_unused ret;
548
549		ret = io_u_queued_complete(td, 1, NULL);
550	}
551}
552
553static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
554{
555	enum fio_ddir odir = ddir ^ 1;
556	struct timeval t;
557	long usec;
558
559	assert(ddir_rw(ddir));
560
561	if (td->rate_pending_usleep[ddir] <= 0)
562		return ddir;
563
564	/*
565	 * We have too much pending sleep in this direction. See if we
566	 * should switch.
567	 */
568	if (td_rw(td) && td->o.rwmix[odir]) {
569		/*
570		 * Other direction does not have too much pending, switch
571		 */
572		if (td->rate_pending_usleep[odir] < 100000)
573			return odir;
574
575		/*
576		 * Both directions have pending sleep. Sleep the minimum time
577		 * and deduct from both.
578		 */
579		if (td->rate_pending_usleep[ddir] <=
580			td->rate_pending_usleep[odir]) {
581			usec = td->rate_pending_usleep[ddir];
582		} else {
583			usec = td->rate_pending_usleep[odir];
584			ddir = odir;
585		}
586	} else
587		usec = td->rate_pending_usleep[ddir];
588
589	io_u_quiesce(td);
590
591	fio_gettime(&t, NULL);
592	usec_sleep(td, usec);
593	usec = utime_since_now(&t);
594
595	td->rate_pending_usleep[ddir] -= usec;
596
597	odir = ddir ^ 1;
598	if (td_rw(td) && __should_check_rate(td, odir))
599		td->rate_pending_usleep[odir] -= usec;
600
601	if (ddir_trim(ddir))
602		return ddir;
603
604	return ddir;
605}
606
607/*
608 * Return the data direction for the next io_u. If the job is a
609 * mixed read/write workload, check the rwmix cycle and switch if
610 * necessary.
611 */
612static enum fio_ddir get_rw_ddir(struct thread_data *td)
613{
614	enum fio_ddir ddir;
615
616	/*
617	 * see if it's time to fsync
618	 */
619	if (td->o.fsync_blocks &&
620	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
621	     td->io_issues[DDIR_WRITE] && should_fsync(td))
622		return DDIR_SYNC;
623
624	/*
625	 * see if it's time to fdatasync
626	 */
627	if (td->o.fdatasync_blocks &&
628	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
629	     td->io_issues[DDIR_WRITE] && should_fsync(td))
630		return DDIR_DATASYNC;
631
632	/*
633	 * see if it's time to sync_file_range
634	 */
635	if (td->sync_file_range_nr &&
636	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
637	     td->io_issues[DDIR_WRITE] && should_fsync(td))
638		return DDIR_SYNC_FILE_RANGE;
639
640	if (td_rw(td)) {
641		/*
642		 * Check if it's time to seed a new data direction.
643		 */
644		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
645			/*
646			 * Put a top limit on how many bytes we do for
647			 * one data direction, to avoid overflowing the
648			 * ranges too much
649			 */
650			ddir = get_rand_ddir(td);
651
652			if (ddir != td->rwmix_ddir)
653				set_rwmix_bytes(td);
654
655			td->rwmix_ddir = ddir;
656		}
657		ddir = td->rwmix_ddir;
658	} else if (td_read(td))
659		ddir = DDIR_READ;
660	else if (td_write(td))
661		ddir = DDIR_WRITE;
662	else
663		ddir = DDIR_TRIM;
664
665	td->rwmix_ddir = rate_ddir(td, ddir);
666	return td->rwmix_ddir;
667}
668
669static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
670{
671	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
672
673	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
674	    td->o.barrier_blocks &&
675	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
676	     td->io_issues[DDIR_WRITE])
677		io_u->flags |= IO_U_F_BARRIER;
678}
679
680void put_file_log(struct thread_data *td, struct fio_file *f)
681{
682	int ret = put_file(td, f);
683
684	if (ret)
685		td_verror(td, ret, "file close");
686}
687
688void put_io_u(struct thread_data *td, struct io_u *io_u)
689{
690	td_io_u_lock(td);
691
692	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
693		put_file_log(td, io_u->file);
694	io_u->file = NULL;
695	io_u->flags &= ~IO_U_F_FREE_DEF;
696	io_u->flags |= IO_U_F_FREE;
697
698	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
699		td->cur_depth--;
700	io_u_qpush(&td->io_u_freelist, io_u);
701	td_io_u_unlock(td);
702	td_io_u_free_notify(td);
703}
704
705void clear_io_u(struct thread_data *td, struct io_u *io_u)
706{
707	io_u->flags &= ~IO_U_F_FLIGHT;
708	put_io_u(td, io_u);
709}
710
711void requeue_io_u(struct thread_data *td, struct io_u **io_u)
712{
713	struct io_u *__io_u = *io_u;
714	enum fio_ddir ddir = acct_ddir(__io_u);
715
716	dprint(FD_IO, "requeue %p\n", __io_u);
717
718	td_io_u_lock(td);
719
720	__io_u->flags |= IO_U_F_FREE;
721	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
722		td->io_issues[ddir]--;
723
724	__io_u->flags &= ~IO_U_F_FLIGHT;
725	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
726		td->cur_depth--;
727
728	io_u_rpush(&td->io_u_requeues, __io_u);
729	td_io_u_unlock(td);
730	*io_u = NULL;
731}
732
733static int fill_io_u(struct thread_data *td, struct io_u *io_u)
734{
735	unsigned int is_random;
736
737	if (td->io_ops->flags & FIO_NOIO)
738		goto out;
739
740	set_rw_ddir(td, io_u);
741
742	/*
743	 * fsync() or fdatasync() or trim etc, we are done
744	 */
745	if (!ddir_rw(io_u->ddir))
746		goto out;
747
748	/*
749	 * See if it's time to switch to a new zone
750	 */
751	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
752		td->zone_bytes = 0;
753		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
754		io_u->file->last_pos = io_u->file->file_offset;
755		td->io_skip_bytes += td->o.zone_skip;
756	}
757
758	/*
759	 * No log, let the seq/rand engine retrieve the next buflen and
760	 * position.
761	 */
762	if (get_next_offset(td, io_u, &is_random)) {
763		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
764		return 1;
765	}
766
767	io_u->buflen = get_next_buflen(td, io_u, is_random);
768	if (!io_u->buflen) {
769		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
770		return 1;
771	}
772
773	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
774		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
775		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
776			(unsigned long long) io_u->offset, io_u->buflen,
777			(unsigned long long) io_u->file->real_file_size);
778		return 1;
779	}
780
781	/*
782	 * mark entry before potentially trimming io_u
783	 */
784	if (td_random(td) && file_randommap(td, io_u->file))
785		mark_random_map(td, io_u);
786
787out:
788	dprint_io_u(io_u, "fill_io_u");
789	td->zone_bytes += io_u->buflen;
790	return 0;
791}
792
793static void __io_u_mark_map(unsigned int *map, unsigned int nr)
794{
795	int idx = 0;
796
797	switch (nr) {
798	default:
799		idx = 6;
800		break;
801	case 33 ... 64:
802		idx = 5;
803		break;
804	case 17 ... 32:
805		idx = 4;
806		break;
807	case 9 ... 16:
808		idx = 3;
809		break;
810	case 5 ... 8:
811		idx = 2;
812		break;
813	case 1 ... 4:
814		idx = 1;
815	case 0:
816		break;
817	}
818
819	map[idx]++;
820}
821
822void io_u_mark_submit(struct thread_data *td, unsigned int nr)
823{
824	__io_u_mark_map(td->ts.io_u_submit, nr);
825	td->ts.total_submit++;
826}
827
828void io_u_mark_complete(struct thread_data *td, unsigned int nr)
829{
830	__io_u_mark_map(td->ts.io_u_complete, nr);
831	td->ts.total_complete++;
832}
833
834void io_u_mark_depth(struct thread_data *td, unsigned int nr)
835{
836	int idx = 0;
837
838	switch (td->cur_depth) {
839	default:
840		idx = 6;
841		break;
842	case 32 ... 63:
843		idx = 5;
844		break;
845	case 16 ... 31:
846		idx = 4;
847		break;
848	case 8 ... 15:
849		idx = 3;
850		break;
851	case 4 ... 7:
852		idx = 2;
853		break;
854	case 2 ... 3:
855		idx = 1;
856	case 1:
857		break;
858	}
859
860	td->ts.io_u_map[idx] += nr;
861}
862
863static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
864{
865	int idx = 0;
866
867	assert(usec < 1000);
868
869	switch (usec) {
870	case 750 ... 999:
871		idx = 9;
872		break;
873	case 500 ... 749:
874		idx = 8;
875		break;
876	case 250 ... 499:
877		idx = 7;
878		break;
879	case 100 ... 249:
880		idx = 6;
881		break;
882	case 50 ... 99:
883		idx = 5;
884		break;
885	case 20 ... 49:
886		idx = 4;
887		break;
888	case 10 ... 19:
889		idx = 3;
890		break;
891	case 4 ... 9:
892		idx = 2;
893		break;
894	case 2 ... 3:
895		idx = 1;
896	case 0 ... 1:
897		break;
898	}
899
900	assert(idx < FIO_IO_U_LAT_U_NR);
901	td->ts.io_u_lat_u[idx]++;
902}
903
904static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
905{
906	int idx = 0;
907
908	switch (msec) {
909	default:
910		idx = 11;
911		break;
912	case 1000 ... 1999:
913		idx = 10;
914		break;
915	case 750 ... 999:
916		idx = 9;
917		break;
918	case 500 ... 749:
919		idx = 8;
920		break;
921	case 250 ... 499:
922		idx = 7;
923		break;
924	case 100 ... 249:
925		idx = 6;
926		break;
927	case 50 ... 99:
928		idx = 5;
929		break;
930	case 20 ... 49:
931		idx = 4;
932		break;
933	case 10 ... 19:
934		idx = 3;
935		break;
936	case 4 ... 9:
937		idx = 2;
938		break;
939	case 2 ... 3:
940		idx = 1;
941	case 0 ... 1:
942		break;
943	}
944
945	assert(idx < FIO_IO_U_LAT_M_NR);
946	td->ts.io_u_lat_m[idx]++;
947}
948
949static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
950{
951	if (usec < 1000)
952		io_u_mark_lat_usec(td, usec);
953	else
954		io_u_mark_lat_msec(td, usec / 1000);
955}
956
957/*
958 * Get next file to service by choosing one at random
959 */
960static struct fio_file *get_next_file_rand(struct thread_data *td,
961					   enum fio_file_flags goodf,
962					   enum fio_file_flags badf)
963{
964	struct fio_file *f;
965	int fno;
966
967	do {
968		int opened = 0;
969		unsigned long r;
970
971		if (td->o.use_os_rand) {
972			r = os_random_long(&td->next_file_state);
973			fno = (unsigned int) ((double) td->o.nr_files
974				* (r / (OS_RAND_MAX + 1.0)));
975		} else {
976			r = __rand(&td->__next_file_state);
977			fno = (unsigned int) ((double) td->o.nr_files
978				* (r / (FRAND_MAX + 1.0)));
979		}
980
981		f = td->files[fno];
982		if (fio_file_done(f))
983			continue;
984
985		if (!fio_file_open(f)) {
986			int err;
987
988			err = td_io_open_file(td, f);
989			if (err)
990				continue;
991			opened = 1;
992		}
993
994		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
995			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
996			return f;
997		}
998		if (opened)
999			td_io_close_file(td, f);
1000	} while (1);
1001}
1002
1003/*
1004 * Get next file to service by doing round robin between all available ones
1005 */
1006static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1007					 int badf)
1008{
1009	unsigned int old_next_file = td->next_file;
1010	struct fio_file *f;
1011
1012	do {
1013		int opened = 0;
1014
1015		f = td->files[td->next_file];
1016
1017		td->next_file++;
1018		if (td->next_file >= td->o.nr_files)
1019			td->next_file = 0;
1020
1021		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1022		if (fio_file_done(f)) {
1023			f = NULL;
1024			continue;
1025		}
1026
1027		if (!fio_file_open(f)) {
1028			int err;
1029
1030			err = td_io_open_file(td, f);
1031			if (err) {
1032				dprint(FD_FILE, "error %d on open of %s\n",
1033					err, f->file_name);
1034				f = NULL;
1035				continue;
1036			}
1037			opened = 1;
1038		}
1039
1040		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1041								f->flags);
1042		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1043			break;
1044
1045		if (opened)
1046			td_io_close_file(td, f);
1047
1048		f = NULL;
1049	} while (td->next_file != old_next_file);
1050
1051	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1052	return f;
1053}
1054
1055static struct fio_file *__get_next_file(struct thread_data *td)
1056{
1057	struct fio_file *f;
1058
1059	assert(td->o.nr_files <= td->files_index);
1060
1061	if (td->nr_done_files >= td->o.nr_files) {
1062		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1063				" nr_files=%d\n", td->nr_open_files,
1064						  td->nr_done_files,
1065						  td->o.nr_files);
1066		return NULL;
1067	}
1068
1069	f = td->file_service_file;
1070	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1071		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1072			goto out;
1073		if (td->file_service_left--)
1074			goto out;
1075	}
1076
1077	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1078	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1079		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1080	else
1081		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1082
1083	td->file_service_file = f;
1084	td->file_service_left = td->file_service_nr - 1;
1085out:
1086	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1087	return f;
1088}
1089
1090static struct fio_file *get_next_file(struct thread_data *td)
1091{
1092	if (!(td->flags & TD_F_PROFILE_OPS)) {
1093		struct prof_io_ops *ops = &td->prof_io_ops;
1094
1095		if (ops->get_next_file)
1096			return ops->get_next_file(td);
1097	}
1098
1099	return __get_next_file(td);
1100}
1101
1102static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1103{
1104	struct fio_file *f;
1105
1106	do {
1107		f = get_next_file(td);
1108		if (!f)
1109			return 1;
1110
1111		io_u->file = f;
1112		get_file(f);
1113
1114		if (!fill_io_u(td, io_u))
1115			break;
1116
1117		put_file_log(td, f);
1118		td_io_close_file(td, f);
1119		io_u->file = NULL;
1120		fio_file_set_done(f);
1121		td->nr_done_files++;
1122		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1123					td->nr_done_files, td->o.nr_files);
1124	} while (1);
1125
1126	return 0;
1127}
1128
1129static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1130		      unsigned long tusec, unsigned long max_usec)
1131{
1132	if (!td->error)
1133		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1134	td_verror(td, ETIMEDOUT, "max latency exceeded");
1135	icd->error = ETIMEDOUT;
1136}
1137
1138static void lat_new_cycle(struct thread_data *td)
1139{
1140	fio_gettime(&td->latency_ts, NULL);
1141	td->latency_ios = ddir_rw_sum(td->io_blocks);
1142	td->latency_failed = 0;
1143}
1144
1145/*
1146 * We had an IO outside the latency target. Reduce the queue depth. If we
1147 * are at QD=1, then it's time to give up.
1148 */
1149static int __lat_target_failed(struct thread_data *td)
1150{
1151	if (td->latency_qd == 1)
1152		return 1;
1153
1154	td->latency_qd_high = td->latency_qd;
1155	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1156
1157	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1158
1159	/*
1160	 * When we ramp QD down, quiesce existing IO to prevent
1161	 * a storm of ramp downs due to pending higher depth.
1162	 */
1163	io_u_quiesce(td);
1164	lat_new_cycle(td);
1165	return 0;
1166}
1167
1168static int lat_target_failed(struct thread_data *td)
1169{
1170	if (td->o.latency_percentile.u.f == 100.0)
1171		return __lat_target_failed(td);
1172
1173	td->latency_failed++;
1174	return 0;
1175}
1176
1177void lat_target_init(struct thread_data *td)
1178{
1179	if (td->o.latency_target) {
1180		dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1181		fio_gettime(&td->latency_ts, NULL);
1182		td->latency_qd = 1;
1183		td->latency_qd_high = td->o.iodepth;
1184		td->latency_qd_low = 1;
1185		td->latency_ios = ddir_rw_sum(td->io_blocks);
1186	} else
1187		td->latency_qd = td->o.iodepth;
1188}
1189
1190static void lat_target_success(struct thread_data *td)
1191{
1192	const unsigned int qd = td->latency_qd;
1193
1194	td->latency_qd_low = td->latency_qd;
1195
1196	/*
1197	 * If we haven't failed yet, we double up to a failing value instead
1198	 * of bisecting from highest possible queue depth. If we have set
1199	 * a limit other than td->o.iodepth, bisect between that.
1200	 */
1201	if (td->latency_qd_high != td->o.iodepth)
1202		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1203	else
1204		td->latency_qd *= 2;
1205
1206	if (td->latency_qd > td->o.iodepth)
1207		td->latency_qd = td->o.iodepth;
1208
1209	dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1210	/*
1211	 * Same as last one, we are done
1212	 */
1213	if (td->latency_qd == qd)
1214		td->done = 1;
1215
1216	lat_new_cycle(td);
1217}
1218
1219/*
1220 * Check if we can bump the queue depth
1221 */
1222void lat_target_check(struct thread_data *td)
1223{
1224	uint64_t usec_window;
1225	uint64_t ios;
1226	double success_ios;
1227
1228	usec_window = utime_since_now(&td->latency_ts);
1229	if (usec_window < td->o.latency_window)
1230		return;
1231
1232	ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1233	success_ios = (double) (ios - td->latency_failed) / (double) ios;
1234	success_ios *= 100.0;
1235
1236	dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1237
1238	if (success_ios >= td->o.latency_percentile.u.f)
1239		lat_target_success(td);
1240	else
1241		__lat_target_failed(td);
1242}
1243
1244/*
1245 * If latency target is enabled, we might be ramping up or down and not
1246 * using the full queue depth available.
1247 */
1248int queue_full(struct thread_data *td)
1249{
1250	const int qempty = io_u_qempty(&td->io_u_freelist);
1251
1252	if (qempty)
1253		return 1;
1254	if (!td->o.latency_target)
1255		return 0;
1256
1257	return td->cur_depth >= td->latency_qd;
1258}
1259
1260struct io_u *__get_io_u(struct thread_data *td)
1261{
1262	struct io_u *io_u;
1263
1264	td_io_u_lock(td);
1265
1266again:
1267	if (!io_u_rempty(&td->io_u_requeues))
1268		io_u = io_u_rpop(&td->io_u_requeues);
1269	else if (!queue_full(td)) {
1270		io_u = io_u_qpop(&td->io_u_freelist);
1271
1272		io_u->buflen = 0;
1273		io_u->resid = 0;
1274		io_u->file = NULL;
1275		io_u->end_io = NULL;
1276	}
1277
1278	if (io_u) {
1279		assert(io_u->flags & IO_U_F_FREE);
1280		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1281		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1282		io_u->flags &= ~IO_U_F_VER_LIST;
1283
1284		io_u->error = 0;
1285		io_u->acct_ddir = -1;
1286		td->cur_depth++;
1287		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1288	} else if (td->o.verify_async) {
1289		/*
1290		 * We ran out, wait for async verify threads to finish and
1291		 * return one
1292		 */
1293		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1294		goto again;
1295	}
1296
1297	td_io_u_unlock(td);
1298	return io_u;
1299}
1300
1301static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1302{
1303	if (!(td->flags & TD_F_TRIM_BACKLOG))
1304		return 0;
1305
1306	if (td->trim_entries) {
1307		int get_trim = 0;
1308
1309		if (td->trim_batch) {
1310			td->trim_batch--;
1311			get_trim = 1;
1312		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1313			 td->last_ddir != DDIR_READ) {
1314			td->trim_batch = td->o.trim_batch;
1315			if (!td->trim_batch)
1316				td->trim_batch = td->o.trim_backlog;
1317			get_trim = 1;
1318		}
1319
1320		if (get_trim && !get_next_trim(td, io_u))
1321			return 1;
1322	}
1323
1324	return 0;
1325}
1326
1327static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1328{
1329	if (!(td->flags & TD_F_VER_BACKLOG))
1330		return 0;
1331
1332	if (td->io_hist_len) {
1333		int get_verify = 0;
1334
1335		if (td->verify_batch)
1336			get_verify = 1;
1337		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1338			 td->last_ddir != DDIR_READ) {
1339			td->verify_batch = td->o.verify_batch;
1340			if (!td->verify_batch)
1341				td->verify_batch = td->o.verify_backlog;
1342			get_verify = 1;
1343		}
1344
1345		if (get_verify && !get_next_verify(td, io_u)) {
1346			td->verify_batch--;
1347			return 1;
1348		}
1349	}
1350
1351	return 0;
1352}
1353
1354/*
1355 * Fill offset and start time into the buffer content, to prevent too
1356 * easy compressible data for simple de-dupe attempts. Do this for every
1357 * 512b block in the range, since that should be the smallest block size
1358 * we can expect from a device.
1359 */
1360static void small_content_scramble(struct io_u *io_u)
1361{
1362	unsigned int i, nr_blocks = io_u->buflen / 512;
1363	uint64_t boffset;
1364	unsigned int offset;
1365	void *p, *end;
1366
1367	if (!nr_blocks)
1368		return;
1369
1370	p = io_u->xfer_buf;
1371	boffset = io_u->offset;
1372	io_u->buf_filled_len = 0;
1373
1374	for (i = 0; i < nr_blocks; i++) {
1375		/*
1376		 * Fill the byte offset into a "random" start offset of
1377		 * the buffer, given by the product of the usec time
1378		 * and the actual offset.
1379		 */
1380		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1381		offset &= ~(sizeof(uint64_t) - 1);
1382		if (offset >= 512 - sizeof(uint64_t))
1383			offset -= sizeof(uint64_t);
1384		memcpy(p + offset, &boffset, sizeof(boffset));
1385
1386		end = p + 512 - sizeof(io_u->start_time);
1387		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1388		p += 512;
1389		boffset += 512;
1390	}
1391}
1392
1393/*
1394 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1395 * etc. The returned io_u is fully ready to be prepped and submitted.
1396 */
1397struct io_u *get_io_u(struct thread_data *td)
1398{
1399	struct fio_file *f;
1400	struct io_u *io_u;
1401	int do_scramble = 0;
1402
1403	io_u = __get_io_u(td);
1404	if (!io_u) {
1405		dprint(FD_IO, "__get_io_u failed\n");
1406		return NULL;
1407	}
1408
1409	if (check_get_verify(td, io_u))
1410		goto out;
1411	if (check_get_trim(td, io_u))
1412		goto out;
1413
1414	/*
1415	 * from a requeue, io_u already setup
1416	 */
1417	if (io_u->file)
1418		goto out;
1419
1420	/*
1421	 * If using an iolog, grab next piece if any available.
1422	 */
1423	if (td->flags & TD_F_READ_IOLOG) {
1424		if (read_iolog_get(td, io_u))
1425			goto err_put;
1426	} else if (set_io_u_file(td, io_u)) {
1427		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1428		goto err_put;
1429	}
1430
1431	f = io_u->file;
1432	assert(fio_file_open(f));
1433
1434	if (ddir_rw(io_u->ddir)) {
1435		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1436			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1437			goto err_put;
1438		}
1439
1440		f->last_start = io_u->offset;
1441		f->last_pos = io_u->offset + io_u->buflen;
1442
1443		if (io_u->ddir == DDIR_WRITE) {
1444			if (td->flags & TD_F_REFILL_BUFFERS) {
1445				io_u_fill_buffer(td, io_u,
1446					io_u->xfer_buflen, io_u->xfer_buflen);
1447			} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1448				do_scramble = 1;
1449			if (td->flags & TD_F_VER_NONE) {
1450				populate_verify_io_u(td, io_u);
1451				do_scramble = 0;
1452			}
1453		} else if (io_u->ddir == DDIR_READ) {
1454			/*
1455			 * Reset the buf_filled parameters so next time if the
1456			 * buffer is used for writes it is refilled.
1457			 */
1458			io_u->buf_filled_len = 0;
1459		}
1460	}
1461
1462	/*
1463	 * Set io data pointers.
1464	 */
1465	io_u->xfer_buf = io_u->buf;
1466	io_u->xfer_buflen = io_u->buflen;
1467
1468out:
1469	assert(io_u->file);
1470	if (!td_io_prep(td, io_u)) {
1471		if (!td->o.disable_slat)
1472			fio_gettime(&io_u->start_time, NULL);
1473		if (do_scramble)
1474			small_content_scramble(io_u);
1475		return io_u;
1476	}
1477err_put:
1478	dprint(FD_IO, "get_io_u failed\n");
1479	put_io_u(td, io_u);
1480	return NULL;
1481}
1482
1483void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1484{
1485	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1486	const char *msg[] = { "read", "write", "sync", "datasync",
1487				"sync_file_range", "wait", "trim" };
1488
1489	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1490		return;
1491
1492	log_err("fio: io_u error");
1493
1494	if (io_u->file)
1495		log_err(" on file %s", io_u->file->file_name);
1496
1497	log_err(": %s\n", strerror(io_u->error));
1498
1499	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1500					io_u->offset, io_u->xfer_buflen);
1501
1502	if (!td->error)
1503		td_verror(td, io_u->error, "io_u error");
1504}
1505
1506static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1507				  struct io_completion_data *icd,
1508				  const enum fio_ddir idx, unsigned int bytes)
1509{
1510	unsigned long lusec = 0;
1511
1512	if (!td->o.disable_clat || !td->o.disable_bw)
1513		lusec = utime_since(&io_u->issue_time, &icd->time);
1514
1515	if (!td->o.disable_lat) {
1516		unsigned long tusec;
1517
1518		tusec = utime_since(&io_u->start_time, &icd->time);
1519		add_lat_sample(td, idx, tusec, bytes);
1520
1521		if (td->flags & TD_F_PROFILE_OPS) {
1522			struct prof_io_ops *ops = &td->prof_io_ops;
1523
1524			if (ops->io_u_lat)
1525				icd->error = ops->io_u_lat(td, tusec);
1526		}
1527
1528		if (td->o.max_latency && tusec > td->o.max_latency)
1529			lat_fatal(td, icd, tusec, td->o.max_latency);
1530		if (td->o.latency_target && tusec > td->o.latency_target) {
1531			if (lat_target_failed(td))
1532				lat_fatal(td, icd, tusec, td->o.latency_target);
1533		}
1534	}
1535
1536	if (!td->o.disable_clat) {
1537		add_clat_sample(td, idx, lusec, bytes);
1538		io_u_mark_latency(td, lusec);
1539	}
1540
1541	if (!td->o.disable_bw)
1542		add_bw_sample(td, idx, bytes, &icd->time);
1543
1544	add_iops_sample(td, idx, bytes, &icd->time);
1545
1546	if (td->o.number_ios && !--td->o.number_ios)
1547		td->done = 1;
1548}
1549
1550static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1551{
1552	uint64_t secs, remainder, bps, bytes;
1553
1554	bytes = td->this_io_bytes[ddir];
1555	bps = td->rate_bps[ddir];
1556	secs = bytes / bps;
1557	remainder = bytes % bps;
1558	return remainder * 1000000 / bps + secs * 1000000;
1559}
1560
1561static void io_completed(struct thread_data *td, struct io_u *io_u,
1562			 struct io_completion_data *icd)
1563{
1564	struct fio_file *f;
1565
1566	dprint_io_u(io_u, "io complete");
1567
1568	td_io_u_lock(td);
1569	assert(io_u->flags & IO_U_F_FLIGHT);
1570	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1571	td_io_u_unlock(td);
1572
1573	if (ddir_sync(io_u->ddir)) {
1574		td->last_was_sync = 1;
1575		f = io_u->file;
1576		if (f) {
1577			f->first_write = -1ULL;
1578			f->last_write = -1ULL;
1579		}
1580		return;
1581	}
1582
1583	td->last_was_sync = 0;
1584	td->last_ddir = io_u->ddir;
1585
1586	if (!io_u->error && ddir_rw(io_u->ddir)) {
1587		unsigned int bytes = io_u->buflen - io_u->resid;
1588		const enum fio_ddir idx = io_u->ddir;
1589		const enum fio_ddir odx = io_u->ddir ^ 1;
1590		int ret;
1591
1592		td->io_blocks[idx]++;
1593		td->this_io_blocks[idx]++;
1594		td->io_bytes[idx] += bytes;
1595
1596		if (!(io_u->flags & IO_U_F_VER_LIST))
1597			td->this_io_bytes[idx] += bytes;
1598
1599		if (idx == DDIR_WRITE) {
1600			f = io_u->file;
1601			if (f) {
1602				if (f->first_write == -1ULL ||
1603				    io_u->offset < f->first_write)
1604					f->first_write = io_u->offset;
1605				if (f->last_write == -1ULL ||
1606				    ((io_u->offset + bytes) > f->last_write))
1607					f->last_write = io_u->offset + bytes;
1608			}
1609		}
1610
1611		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1612					   td->runstate == TD_VERIFYING)) {
1613			account_io_completion(td, io_u, icd, idx, bytes);
1614
1615			if (__should_check_rate(td, idx)) {
1616				td->rate_pending_usleep[idx] =
1617					(usec_for_io(td, idx) -
1618					 utime_since_now(&td->start));
1619			}
1620			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1621				td->rate_pending_usleep[odx] =
1622					(usec_for_io(td, odx) -
1623					 utime_since_now(&td->start));
1624		}
1625
1626		if (td_write(td) && idx == DDIR_WRITE &&
1627		    td->o.do_verify &&
1628		    td->o.verify != VERIFY_NONE &&
1629		    !td->o.experimental_verify)
1630			log_io_piece(td, io_u);
1631
1632		icd->bytes_done[idx] += bytes;
1633
1634		if (io_u->end_io) {
1635			ret = io_u->end_io(td, io_u);
1636			if (ret && !icd->error)
1637				icd->error = ret;
1638		}
1639	} else if (io_u->error) {
1640		icd->error = io_u->error;
1641		io_u_log_error(td, io_u);
1642	}
1643	if (icd->error) {
1644		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1645		if (!td_non_fatal_error(td, eb, icd->error))
1646			return;
1647		/*
1648		 * If there is a non_fatal error, then add to the error count
1649		 * and clear all the errors.
1650		 */
1651		update_error_count(td, icd->error);
1652		td_clear_error(td);
1653		icd->error = 0;
1654		io_u->error = 0;
1655	}
1656}
1657
1658static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1659		     int nr)
1660{
1661	int ddir;
1662	if (!td->o.disable_clat || !td->o.disable_bw)
1663		fio_gettime(&icd->time, NULL);
1664
1665	icd->nr = nr;
1666
1667	icd->error = 0;
1668	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1669		icd->bytes_done[ddir] = 0;
1670}
1671
1672static void ios_completed(struct thread_data *td,
1673			  struct io_completion_data *icd)
1674{
1675	struct io_u *io_u;
1676	int i;
1677
1678	for (i = 0; i < icd->nr; i++) {
1679		io_u = td->io_ops->event(td, i);
1680
1681		io_completed(td, io_u, icd);
1682
1683		if (!(io_u->flags & IO_U_F_FREE_DEF))
1684			put_io_u(td, io_u);
1685	}
1686}
1687
1688/*
1689 * Complete a single io_u for the sync engines.
1690 */
1691int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1692		       uint64_t *bytes)
1693{
1694	struct io_completion_data icd;
1695
1696	init_icd(td, &icd, 1);
1697	io_completed(td, io_u, &icd);
1698
1699	if (!(io_u->flags & IO_U_F_FREE_DEF))
1700		put_io_u(td, io_u);
1701
1702	if (icd.error) {
1703		td_verror(td, icd.error, "io_u_sync_complete");
1704		return -1;
1705	}
1706
1707	if (bytes) {
1708		int ddir;
1709
1710		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1711			bytes[ddir] += icd.bytes_done[ddir];
1712	}
1713
1714	return 0;
1715}
1716
1717/*
1718 * Called to complete min_events number of io for the async engines.
1719 */
1720int io_u_queued_complete(struct thread_data *td, int min_evts,
1721			 uint64_t *bytes)
1722{
1723	struct io_completion_data icd;
1724	struct timespec *tvp = NULL;
1725	int ret;
1726	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1727
1728	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1729
1730	if (!min_evts)
1731		tvp = &ts;
1732
1733	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1734	if (ret < 0) {
1735		td_verror(td, -ret, "td_io_getevents");
1736		return ret;
1737	} else if (!ret)
1738		return ret;
1739
1740	init_icd(td, &icd, ret);
1741	ios_completed(td, &icd);
1742	if (icd.error) {
1743		td_verror(td, icd.error, "io_u_queued_complete");
1744		return -1;
1745	}
1746
1747	if (bytes) {
1748		int ddir;
1749
1750		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1751			bytes[ddir] += icd.bytes_done[ddir];
1752	}
1753
1754	return 0;
1755}
1756
1757/*
1758 * Call when io_u is really queued, to update the submission latency.
1759 */
1760void io_u_queued(struct thread_data *td, struct io_u *io_u)
1761{
1762	if (!td->o.disable_slat) {
1763		unsigned long slat_time;
1764
1765		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1766		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1767	}
1768}
1769
1770void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1771		    unsigned int max_bs)
1772{
1773	if (td->o.buffer_pattern_bytes)
1774		fill_buffer_pattern(td, buf, max_bs);
1775	else if (!td->o.zero_buffers) {
1776		unsigned int perc = td->o.compress_percentage;
1777
1778		if (perc) {
1779			unsigned int seg = min_write;
1780
1781			seg = min(min_write, td->o.compress_chunk);
1782			if (!seg)
1783				seg = min_write;
1784
1785			fill_random_buf_percentage(&td->buf_state, buf,
1786						perc, seg, max_bs);
1787		} else
1788			fill_random_buf(&td->buf_state, buf, max_bs);
1789	} else
1790		memset(buf, 0, max_bs);
1791}
1792
1793/*
1794 * "randomly" fill the buffer contents
1795 */
1796void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1797		      unsigned int min_write, unsigned int max_bs)
1798{
1799	io_u->buf_filled_len = 0;
1800	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1801}
1802