io_u.c revision 05074830620ea0811576e3565cc4c4bd48c89077
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#include "err.h"
15
16struct io_completion_data {
17	int nr;				/* input */
18
19	int error;			/* output */
20	uint64_t bytes_done[DDIR_RWDIR_CNT];	/* output */
21	struct timeval time;		/* output */
22};
23
24/*
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
28static int random_map_free(struct fio_file *f, const uint64_t block)
29{
30	return !axmap_isset(f->io_axmap, block);
31}
32
33/*
34 * Mark a given offset as used in the map.
35 */
36static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37{
38	unsigned int min_bs = td->o.rw_min_bs;
39	struct fio_file *f = io_u->file;
40	unsigned int nr_blocks;
41	uint64_t block;
42
43	block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45
46	if (!(io_u->flags & IO_U_F_BUSY_OK))
47		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48
49	if ((nr_blocks * min_bs) < io_u->buflen)
50		io_u->buflen = nr_blocks * min_bs;
51}
52
53static uint64_t last_block(struct thread_data *td, struct fio_file *f,
54			   enum fio_ddir ddir)
55{
56	uint64_t max_blocks;
57	uint64_t max_size;
58
59	assert(ddir_rw(ddir));
60
61	/*
62	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
63	 */
64	max_size = f->io_size;
65	if (max_size > f->real_file_size)
66		max_size = f->real_file_size;
67
68	if (td->o.zone_range)
69		max_size = td->o.zone_range;
70
71	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
72	if (!max_blocks)
73		return 0;
74
75	return max_blocks;
76}
77
78struct rand_off {
79	struct flist_head list;
80	uint64_t off;
81};
82
83static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84				  enum fio_ddir ddir, uint64_t *b)
85{
86	uint64_t r, lastb;
87
88	lastb = last_block(td, f, ddir);
89	if (!lastb)
90		return 1;
91
92	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
93		uint64_t rmax;
94
95		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
96
97		if (td->o.use_os_rand) {
98			rmax = OS_RAND_MAX;
99			r = os_random_long(&td->random_state);
100		} else {
101			rmax = FRAND_MAX;
102			r = __rand(&td->__random_state);
103		}
104
105		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
106
107		*b = lastb * (r / ((uint64_t) rmax + 1.0));
108	} else {
109		uint64_t off = 0;
110
111		if (lfsr_next(&f->lfsr, &off, lastb))
112			return 1;
113
114		*b = off;
115	}
116
117	/*
118	 * if we are not maintaining a random map, we are done.
119	 */
120	if (!file_randommap(td, f))
121		goto ret;
122
123	/*
124	 * calculate map offset and check if it's free
125	 */
126	if (random_map_free(f, *b))
127		goto ret;
128
129	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
130						(unsigned long long) *b);
131
132	*b = axmap_next_free(f->io_axmap, *b);
133	if (*b == (uint64_t) -1ULL)
134		return 1;
135ret:
136	return 0;
137}
138
139static int __get_next_rand_offset_zipf(struct thread_data *td,
140				       struct fio_file *f, enum fio_ddir ddir,
141				       uint64_t *b)
142{
143	*b = zipf_next(&f->zipf);
144	return 0;
145}
146
147static int __get_next_rand_offset_pareto(struct thread_data *td,
148					 struct fio_file *f, enum fio_ddir ddir,
149					 uint64_t *b)
150{
151	*b = pareto_next(&f->zipf);
152	return 0;
153}
154
155static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
156{
157	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
158	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
159
160	return r1->off - r2->off;
161}
162
163static int get_off_from_method(struct thread_data *td, struct fio_file *f,
164			       enum fio_ddir ddir, uint64_t *b)
165{
166	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
167		return __get_next_rand_offset(td, f, ddir, b);
168	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
169		return __get_next_rand_offset_zipf(td, f, ddir, b);
170	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
171		return __get_next_rand_offset_pareto(td, f, ddir, b);
172
173	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
174	return 1;
175}
176
177/*
178 * Sort the reads for a verify phase in batches of verifysort_nr, if
179 * specified.
180 */
181static inline int should_sort_io(struct thread_data *td)
182{
183	if (!td->o.verifysort_nr || !td->o.do_verify)
184		return 0;
185	if (!td_random(td))
186		return 0;
187	if (td->runstate != TD_VERIFYING)
188		return 0;
189	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
190		return 0;
191
192	return 1;
193}
194
195static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
196{
197	unsigned int v;
198	unsigned long r;
199
200	if (td->o.perc_rand[ddir] == 100)
201		return 1;
202
203	if (td->o.use_os_rand) {
204		r = os_random_long(&td->seq_rand_state[ddir]);
205		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
206	} else {
207		r = __rand(&td->__seq_rand_state[ddir]);
208		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
209	}
210
211	return v <= td->o.perc_rand[ddir];
212}
213
214static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
215				enum fio_ddir ddir, uint64_t *b)
216{
217	struct rand_off *r;
218	int i, ret = 1;
219
220	if (!should_sort_io(td))
221		return get_off_from_method(td, f, ddir, b);
222
223	if (!flist_empty(&td->next_rand_list)) {
224		struct rand_off *r;
225fetch:
226		r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
227		flist_del(&r->list);
228		*b = r->off;
229		free(r);
230		return 0;
231	}
232
233	for (i = 0; i < td->o.verifysort_nr; i++) {
234		r = malloc(sizeof(*r));
235
236		ret = get_off_from_method(td, f, ddir, &r->off);
237		if (ret) {
238			free(r);
239			break;
240		}
241
242		flist_add(&r->list, &td->next_rand_list);
243	}
244
245	if (ret && !i)
246		return ret;
247
248	assert(!flist_empty(&td->next_rand_list));
249	flist_sort(NULL, &td->next_rand_list, flist_cmp);
250	goto fetch;
251}
252
253static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
254			       enum fio_ddir ddir, uint64_t *b)
255{
256	if (!get_next_rand_offset(td, f, ddir, b))
257		return 0;
258
259	if (td->o.time_based) {
260		fio_file_reset(td, f);
261		if (!get_next_rand_offset(td, f, ddir, b))
262			return 0;
263	}
264
265	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
266			f->file_name, (unsigned long long) f->last_pos,
267			(unsigned long long) f->real_file_size);
268	return 1;
269}
270
271static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
272			       enum fio_ddir ddir, uint64_t *offset)
273{
274	assert(ddir_rw(ddir));
275
276	if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->o.time_based)
277		f->last_pos = f->last_pos - f->io_size;
278
279	if (f->last_pos < f->real_file_size) {
280		uint64_t pos;
281
282		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
283			f->last_pos = f->real_file_size;
284
285		pos = f->last_pos - f->file_offset;
286		if (pos)
287			pos += td->o.ddir_seq_add;
288
289		*offset = pos;
290		return 0;
291	}
292
293	return 1;
294}
295
296static int get_next_block(struct thread_data *td, struct io_u *io_u,
297			  enum fio_ddir ddir, int rw_seq,
298			  unsigned int *is_random)
299{
300	struct fio_file *f = io_u->file;
301	uint64_t b, offset;
302	int ret;
303
304	assert(ddir_rw(ddir));
305
306	b = offset = -1ULL;
307
308	if (rw_seq) {
309		if (td_random(td)) {
310			if (should_do_random(td, ddir)) {
311				ret = get_next_rand_block(td, f, ddir, &b);
312				*is_random = 1;
313			} else {
314				*is_random = 0;
315				io_u->flags |= IO_U_F_BUSY_OK;
316				ret = get_next_seq_offset(td, f, ddir, &offset);
317				if (ret)
318					ret = get_next_rand_block(td, f, ddir, &b);
319			}
320		} else {
321			*is_random = 0;
322			ret = get_next_seq_offset(td, f, ddir, &offset);
323		}
324	} else {
325		io_u->flags |= IO_U_F_BUSY_OK;
326		*is_random = 0;
327
328		if (td->o.rw_seq == RW_SEQ_SEQ) {
329			ret = get_next_seq_offset(td, f, ddir, &offset);
330			if (ret) {
331				ret = get_next_rand_block(td, f, ddir, &b);
332				*is_random = 0;
333			}
334		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
335			if (f->last_start != -1ULL)
336				offset = f->last_start - f->file_offset;
337			else
338				offset = 0;
339			ret = 0;
340		} else {
341			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
342			ret = 1;
343		}
344	}
345
346	if (!ret) {
347		if (offset != -1ULL)
348			io_u->offset = offset;
349		else if (b != -1ULL)
350			io_u->offset = b * td->o.ba[ddir];
351		else {
352			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
353			ret = 1;
354		}
355	}
356
357	return ret;
358}
359
360/*
361 * For random io, generate a random new block and see if it's used. Repeat
362 * until we find a free one. For sequential io, just return the end of
363 * the last io issued.
364 */
365static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
366			     unsigned int *is_random)
367{
368	struct fio_file *f = io_u->file;
369	enum fio_ddir ddir = io_u->ddir;
370	int rw_seq_hit = 0;
371
372	assert(ddir_rw(ddir));
373
374	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
375		rw_seq_hit = 1;
376		td->ddir_seq_nr = td->o.ddir_seq_nr;
377	}
378
379	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
380		return 1;
381
382	if (io_u->offset >= f->io_size) {
383		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
384					(unsigned long long) io_u->offset,
385					(unsigned long long) f->io_size);
386		return 1;
387	}
388
389	io_u->offset += f->file_offset;
390	if (io_u->offset >= f->real_file_size) {
391		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
392					(unsigned long long) io_u->offset,
393					(unsigned long long) f->real_file_size);
394		return 1;
395	}
396
397	return 0;
398}
399
400static int get_next_offset(struct thread_data *td, struct io_u *io_u,
401			   unsigned int *is_random)
402{
403	if (td->flags & TD_F_PROFILE_OPS) {
404		struct prof_io_ops *ops = &td->prof_io_ops;
405
406		if (ops->fill_io_u_off)
407			return ops->fill_io_u_off(td, io_u, is_random);
408	}
409
410	return __get_next_offset(td, io_u, is_random);
411}
412
413static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
414			    unsigned int buflen)
415{
416	struct fio_file *f = io_u->file;
417
418	return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
419}
420
421static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
422				      unsigned int is_random)
423{
424	int ddir = io_u->ddir;
425	unsigned int buflen = 0;
426	unsigned int minbs, maxbs;
427	unsigned long r, rand_max;
428
429	assert(ddir_rw(ddir));
430
431	if (td->o.bs_is_seq_rand)
432		ddir = is_random ? DDIR_WRITE: DDIR_READ;
433
434	minbs = td->o.min_bs[ddir];
435	maxbs = td->o.max_bs[ddir];
436
437	if (minbs == maxbs)
438		return minbs;
439
440	/*
441	 * If we can't satisfy the min block size from here, then fail
442	 */
443	if (!io_u_fits(td, io_u, minbs))
444		return 0;
445
446	if (td->o.use_os_rand)
447		rand_max = OS_RAND_MAX;
448	else
449		rand_max = FRAND_MAX;
450
451	do {
452		if (td->o.use_os_rand)
453			r = os_random_long(&td->bsrange_state);
454		else
455			r = __rand(&td->__bsrange_state);
456
457		if (!td->o.bssplit_nr[ddir]) {
458			buflen = 1 + (unsigned int) ((double) maxbs *
459					(r / (rand_max + 1.0)));
460			if (buflen < minbs)
461				buflen = minbs;
462		} else {
463			long perc = 0;
464			unsigned int i;
465
466			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
467				struct bssplit *bsp = &td->o.bssplit[ddir][i];
468
469				buflen = bsp->bs;
470				perc += bsp->perc;
471				if ((r <= ((rand_max / 100L) * perc)) &&
472				    io_u_fits(td, io_u, buflen))
473					break;
474			}
475		}
476
477		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
478			buflen = (buflen + td->o.verify_interval - 1) &
479				~(td->o.verify_interval - 1);
480
481		if (!td->o.bs_unaligned && is_power_of_2(minbs))
482			buflen = (buflen + minbs - 1) & ~(minbs - 1);
483
484	} while (!io_u_fits(td, io_u, buflen));
485
486	return buflen;
487}
488
489static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
490				    unsigned int is_random)
491{
492	if (td->flags & TD_F_PROFILE_OPS) {
493		struct prof_io_ops *ops = &td->prof_io_ops;
494
495		if (ops->fill_io_u_size)
496			return ops->fill_io_u_size(td, io_u, is_random);
497	}
498
499	return __get_next_buflen(td, io_u, is_random);
500}
501
502static void set_rwmix_bytes(struct thread_data *td)
503{
504	unsigned int diff;
505
506	/*
507	 * we do time or byte based switch. this is needed because
508	 * buffered writes may issue a lot quicker than they complete,
509	 * whereas reads do not.
510	 */
511	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
512	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
513}
514
515static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
516{
517	unsigned int v;
518	unsigned long r;
519
520	if (td->o.use_os_rand) {
521		r = os_random_long(&td->rwmix_state);
522		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
523	} else {
524		r = __rand(&td->__rwmix_state);
525		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
526	}
527
528	if (v <= td->o.rwmix[DDIR_READ])
529		return DDIR_READ;
530
531	return DDIR_WRITE;
532}
533
534void io_u_quiesce(struct thread_data *td)
535{
536	/*
537	 * We are going to sleep, ensure that we flush anything pending as
538	 * not to skew our latency numbers.
539	 *
540	 * Changed to only monitor 'in flight' requests here instead of the
541	 * td->cur_depth, b/c td->cur_depth does not accurately represent
542	 * io's that have been actually submitted to an async engine,
543	 * and cur_depth is meaningless for sync engines.
544	 */
545	while (td->io_u_in_flight) {
546		int fio_unused ret;
547
548		ret = io_u_queued_complete(td, 1, NULL);
549	}
550}
551
552static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
553{
554	enum fio_ddir odir = ddir ^ 1;
555	struct timeval t;
556	long usec;
557
558	assert(ddir_rw(ddir));
559
560	if (td->rate_pending_usleep[ddir] <= 0)
561		return ddir;
562
563	/*
564	 * We have too much pending sleep in this direction. See if we
565	 * should switch.
566	 */
567	if (td_rw(td) && td->o.rwmix[odir]) {
568		/*
569		 * Other direction does not have too much pending, switch
570		 */
571		if (td->rate_pending_usleep[odir] < 100000)
572			return odir;
573
574		/*
575		 * Both directions have pending sleep. Sleep the minimum time
576		 * and deduct from both.
577		 */
578		if (td->rate_pending_usleep[ddir] <=
579			td->rate_pending_usleep[odir]) {
580			usec = td->rate_pending_usleep[ddir];
581		} else {
582			usec = td->rate_pending_usleep[odir];
583			ddir = odir;
584		}
585	} else
586		usec = td->rate_pending_usleep[ddir];
587
588	io_u_quiesce(td);
589
590	fio_gettime(&t, NULL);
591	usec_sleep(td, usec);
592	usec = utime_since_now(&t);
593
594	td->rate_pending_usleep[ddir] -= usec;
595
596	odir = ddir ^ 1;
597	if (td_rw(td) && __should_check_rate(td, odir))
598		td->rate_pending_usleep[odir] -= usec;
599
600	if (ddir_trim(ddir))
601		return ddir;
602
603	return ddir;
604}
605
606/*
607 * Return the data direction for the next io_u. If the job is a
608 * mixed read/write workload, check the rwmix cycle and switch if
609 * necessary.
610 */
611static enum fio_ddir get_rw_ddir(struct thread_data *td)
612{
613	enum fio_ddir ddir;
614
615	/*
616	 * see if it's time to fsync
617	 */
618	if (td->o.fsync_blocks &&
619	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
620	     td->io_issues[DDIR_WRITE] && should_fsync(td))
621		return DDIR_SYNC;
622
623	/*
624	 * see if it's time to fdatasync
625	 */
626	if (td->o.fdatasync_blocks &&
627	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
628	     td->io_issues[DDIR_WRITE] && should_fsync(td))
629		return DDIR_DATASYNC;
630
631	/*
632	 * see if it's time to sync_file_range
633	 */
634	if (td->sync_file_range_nr &&
635	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
636	     td->io_issues[DDIR_WRITE] && should_fsync(td))
637		return DDIR_SYNC_FILE_RANGE;
638
639	if (td_rw(td)) {
640		/*
641		 * Check if it's time to seed a new data direction.
642		 */
643		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
644			/*
645			 * Put a top limit on how many bytes we do for
646			 * one data direction, to avoid overflowing the
647			 * ranges too much
648			 */
649			ddir = get_rand_ddir(td);
650
651			if (ddir != td->rwmix_ddir)
652				set_rwmix_bytes(td);
653
654			td->rwmix_ddir = ddir;
655		}
656		ddir = td->rwmix_ddir;
657	} else if (td_read(td))
658		ddir = DDIR_READ;
659	else if (td_write(td))
660		ddir = DDIR_WRITE;
661	else
662		ddir = DDIR_TRIM;
663
664	td->rwmix_ddir = rate_ddir(td, ddir);
665	return td->rwmix_ddir;
666}
667
668static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
669{
670	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
671
672	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
673	    td->o.barrier_blocks &&
674	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
675	     td->io_issues[DDIR_WRITE])
676		io_u->flags |= IO_U_F_BARRIER;
677}
678
679void put_file_log(struct thread_data *td, struct fio_file *f)
680{
681	unsigned int ret = put_file(td, f);
682
683	if (ret)
684		td_verror(td, ret, "file close");
685}
686
687void put_io_u(struct thread_data *td, struct io_u *io_u)
688{
689	td_io_u_lock(td);
690
691	if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
692		put_file_log(td, io_u->file);
693
694	io_u->file = NULL;
695	io_u->flags |= IO_U_F_FREE;
696
697	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
698		td->cur_depth--;
699	io_u_qpush(&td->io_u_freelist, io_u);
700	td_io_u_unlock(td);
701	td_io_u_free_notify(td);
702}
703
704void clear_io_u(struct thread_data *td, struct io_u *io_u)
705{
706	io_u->flags &= ~IO_U_F_FLIGHT;
707	put_io_u(td, io_u);
708}
709
710void requeue_io_u(struct thread_data *td, struct io_u **io_u)
711{
712	struct io_u *__io_u = *io_u;
713	enum fio_ddir ddir = acct_ddir(__io_u);
714
715	dprint(FD_IO, "requeue %p\n", __io_u);
716
717	td_io_u_lock(td);
718
719	__io_u->flags |= IO_U_F_FREE;
720	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
721		td->io_issues[ddir]--;
722
723	__io_u->flags &= ~IO_U_F_FLIGHT;
724	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
725		td->cur_depth--;
726
727	io_u_rpush(&td->io_u_requeues, __io_u);
728	td_io_u_unlock(td);
729	*io_u = NULL;
730}
731
732static int fill_io_u(struct thread_data *td, struct io_u *io_u)
733{
734	unsigned int is_random;
735
736	if (td->io_ops->flags & FIO_NOIO)
737		goto out;
738
739	set_rw_ddir(td, io_u);
740
741	/*
742	 * fsync() or fdatasync() or trim etc, we are done
743	 */
744	if (!ddir_rw(io_u->ddir))
745		goto out;
746
747	/*
748	 * See if it's time to switch to a new zone
749	 */
750	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
751		td->zone_bytes = 0;
752		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
753		io_u->file->last_pos = io_u->file->file_offset;
754		td->io_skip_bytes += td->o.zone_skip;
755	}
756
757	/*
758	 * No log, let the seq/rand engine retrieve the next buflen and
759	 * position.
760	 */
761	if (get_next_offset(td, io_u, &is_random)) {
762		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
763		return 1;
764	}
765
766	io_u->buflen = get_next_buflen(td, io_u, is_random);
767	if (!io_u->buflen) {
768		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
769		return 1;
770	}
771
772	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
773		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
774		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
775			(unsigned long long) io_u->offset, io_u->buflen,
776			(unsigned long long) io_u->file->real_file_size);
777		return 1;
778	}
779
780	/*
781	 * mark entry before potentially trimming io_u
782	 */
783	if (td_random(td) && file_randommap(td, io_u->file))
784		mark_random_map(td, io_u);
785
786out:
787	dprint_io_u(io_u, "fill_io_u");
788	td->zone_bytes += io_u->buflen;
789	return 0;
790}
791
792static void __io_u_mark_map(unsigned int *map, unsigned int nr)
793{
794	int idx = 0;
795
796	switch (nr) {
797	default:
798		idx = 6;
799		break;
800	case 33 ... 64:
801		idx = 5;
802		break;
803	case 17 ... 32:
804		idx = 4;
805		break;
806	case 9 ... 16:
807		idx = 3;
808		break;
809	case 5 ... 8:
810		idx = 2;
811		break;
812	case 1 ... 4:
813		idx = 1;
814	case 0:
815		break;
816	}
817
818	map[idx]++;
819}
820
821void io_u_mark_submit(struct thread_data *td, unsigned int nr)
822{
823	__io_u_mark_map(td->ts.io_u_submit, nr);
824	td->ts.total_submit++;
825}
826
827void io_u_mark_complete(struct thread_data *td, unsigned int nr)
828{
829	__io_u_mark_map(td->ts.io_u_complete, nr);
830	td->ts.total_complete++;
831}
832
833void io_u_mark_depth(struct thread_data *td, unsigned int nr)
834{
835	int idx = 0;
836
837	switch (td->cur_depth) {
838	default:
839		idx = 6;
840		break;
841	case 32 ... 63:
842		idx = 5;
843		break;
844	case 16 ... 31:
845		idx = 4;
846		break;
847	case 8 ... 15:
848		idx = 3;
849		break;
850	case 4 ... 7:
851		idx = 2;
852		break;
853	case 2 ... 3:
854		idx = 1;
855	case 1:
856		break;
857	}
858
859	td->ts.io_u_map[idx] += nr;
860}
861
862static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
863{
864	int idx = 0;
865
866	assert(usec < 1000);
867
868	switch (usec) {
869	case 750 ... 999:
870		idx = 9;
871		break;
872	case 500 ... 749:
873		idx = 8;
874		break;
875	case 250 ... 499:
876		idx = 7;
877		break;
878	case 100 ... 249:
879		idx = 6;
880		break;
881	case 50 ... 99:
882		idx = 5;
883		break;
884	case 20 ... 49:
885		idx = 4;
886		break;
887	case 10 ... 19:
888		idx = 3;
889		break;
890	case 4 ... 9:
891		idx = 2;
892		break;
893	case 2 ... 3:
894		idx = 1;
895	case 0 ... 1:
896		break;
897	}
898
899	assert(idx < FIO_IO_U_LAT_U_NR);
900	td->ts.io_u_lat_u[idx]++;
901}
902
903static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
904{
905	int idx = 0;
906
907	switch (msec) {
908	default:
909		idx = 11;
910		break;
911	case 1000 ... 1999:
912		idx = 10;
913		break;
914	case 750 ... 999:
915		idx = 9;
916		break;
917	case 500 ... 749:
918		idx = 8;
919		break;
920	case 250 ... 499:
921		idx = 7;
922		break;
923	case 100 ... 249:
924		idx = 6;
925		break;
926	case 50 ... 99:
927		idx = 5;
928		break;
929	case 20 ... 49:
930		idx = 4;
931		break;
932	case 10 ... 19:
933		idx = 3;
934		break;
935	case 4 ... 9:
936		idx = 2;
937		break;
938	case 2 ... 3:
939		idx = 1;
940	case 0 ... 1:
941		break;
942	}
943
944	assert(idx < FIO_IO_U_LAT_M_NR);
945	td->ts.io_u_lat_m[idx]++;
946}
947
948static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
949{
950	if (usec < 1000)
951		io_u_mark_lat_usec(td, usec);
952	else
953		io_u_mark_lat_msec(td, usec / 1000);
954}
955
956/*
957 * Get next file to service by choosing one at random
958 */
959static struct fio_file *get_next_file_rand(struct thread_data *td,
960					   enum fio_file_flags goodf,
961					   enum fio_file_flags badf)
962{
963	struct fio_file *f;
964	int fno;
965
966	do {
967		int opened = 0;
968		unsigned long r;
969
970		if (td->o.use_os_rand) {
971			r = os_random_long(&td->next_file_state);
972			fno = (unsigned int) ((double) td->o.nr_files
973				* (r / (OS_RAND_MAX + 1.0)));
974		} else {
975			r = __rand(&td->__next_file_state);
976			fno = (unsigned int) ((double) td->o.nr_files
977				* (r / (FRAND_MAX + 1.0)));
978		}
979
980		f = td->files[fno];
981		if (fio_file_done(f))
982			continue;
983
984		if (!fio_file_open(f)) {
985			int err;
986
987			if (td->nr_open_files >= td->o.open_files)
988				return ERR_PTR(-EBUSY);
989
990			err = td_io_open_file(td, f);
991			if (err)
992				continue;
993			opened = 1;
994		}
995
996		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
997			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
998			return f;
999		}
1000		if (opened)
1001			td_io_close_file(td, f);
1002	} while (1);
1003}
1004
1005/*
1006 * Get next file to service by doing round robin between all available ones
1007 */
1008static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1009					 int badf)
1010{
1011	unsigned int old_next_file = td->next_file;
1012	struct fio_file *f;
1013
1014	do {
1015		int opened = 0;
1016
1017		f = td->files[td->next_file];
1018
1019		td->next_file++;
1020		if (td->next_file >= td->o.nr_files)
1021			td->next_file = 0;
1022
1023		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1024		if (fio_file_done(f)) {
1025			f = NULL;
1026			continue;
1027		}
1028
1029		if (!fio_file_open(f)) {
1030			int err;
1031
1032			if (td->nr_open_files >= td->o.open_files)
1033				return ERR_PTR(-EBUSY);
1034
1035			err = td_io_open_file(td, f);
1036			if (err) {
1037				dprint(FD_FILE, "error %d on open of %s\n",
1038					err, f->file_name);
1039				f = NULL;
1040				continue;
1041			}
1042			opened = 1;
1043		}
1044
1045		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1046								f->flags);
1047		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1048			break;
1049
1050		if (opened)
1051			td_io_close_file(td, f);
1052
1053		f = NULL;
1054	} while (td->next_file != old_next_file);
1055
1056	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1057	return f;
1058}
1059
1060static struct fio_file *__get_next_file(struct thread_data *td)
1061{
1062	struct fio_file *f;
1063
1064	assert(td->o.nr_files <= td->files_index);
1065
1066	if (td->nr_done_files >= td->o.nr_files) {
1067		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1068				" nr_files=%d\n", td->nr_open_files,
1069						  td->nr_done_files,
1070						  td->o.nr_files);
1071		return NULL;
1072	}
1073
1074	f = td->file_service_file;
1075	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1076		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1077			goto out;
1078		if (td->file_service_left--)
1079			goto out;
1080	}
1081
1082	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1083	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1084		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1085	else
1086		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1087
1088	if (IS_ERR(f))
1089		return f;
1090
1091	td->file_service_file = f;
1092	td->file_service_left = td->file_service_nr - 1;
1093out:
1094	if (f)
1095		dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1096	else
1097		dprint(FD_FILE, "get_next_file: NULL\n");
1098	return f;
1099}
1100
1101static struct fio_file *get_next_file(struct thread_data *td)
1102{
1103	if (td->flags & TD_F_PROFILE_OPS) {
1104		struct prof_io_ops *ops = &td->prof_io_ops;
1105
1106		if (ops->get_next_file)
1107			return ops->get_next_file(td);
1108	}
1109
1110	return __get_next_file(td);
1111}
1112
1113static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1114{
1115	struct fio_file *f;
1116
1117	do {
1118		f = get_next_file(td);
1119		if (IS_ERR_OR_NULL(f))
1120			return PTR_ERR(f);
1121
1122		io_u->file = f;
1123		get_file(f);
1124
1125		if (!fill_io_u(td, io_u))
1126			break;
1127
1128		put_file_log(td, f);
1129		td_io_close_file(td, f);
1130		io_u->file = NULL;
1131		fio_file_set_done(f);
1132		td->nr_done_files++;
1133		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1134					td->nr_done_files, td->o.nr_files);
1135	} while (1);
1136
1137	return 0;
1138}
1139
1140static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1141		      unsigned long tusec, unsigned long max_usec)
1142{
1143	if (!td->error)
1144		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1145	td_verror(td, ETIMEDOUT, "max latency exceeded");
1146	icd->error = ETIMEDOUT;
1147}
1148
1149static void lat_new_cycle(struct thread_data *td)
1150{
1151	fio_gettime(&td->latency_ts, NULL);
1152	td->latency_ios = ddir_rw_sum(td->io_blocks);
1153	td->latency_failed = 0;
1154}
1155
1156/*
1157 * We had an IO outside the latency target. Reduce the queue depth. If we
1158 * are at QD=1, then it's time to give up.
1159 */
1160static int __lat_target_failed(struct thread_data *td)
1161{
1162	if (td->latency_qd == 1)
1163		return 1;
1164
1165	td->latency_qd_high = td->latency_qd;
1166
1167	if (td->latency_qd == td->latency_qd_low)
1168		td->latency_qd_low--;
1169
1170	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1171
1172	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1173
1174	/*
1175	 * When we ramp QD down, quiesce existing IO to prevent
1176	 * a storm of ramp downs due to pending higher depth.
1177	 */
1178	io_u_quiesce(td);
1179	lat_new_cycle(td);
1180	return 0;
1181}
1182
1183static int lat_target_failed(struct thread_data *td)
1184{
1185	if (td->o.latency_percentile.u.f == 100.0)
1186		return __lat_target_failed(td);
1187
1188	td->latency_failed++;
1189	return 0;
1190}
1191
1192void lat_target_init(struct thread_data *td)
1193{
1194	td->latency_end_run = 0;
1195
1196	if (td->o.latency_target) {
1197		dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1198		fio_gettime(&td->latency_ts, NULL);
1199		td->latency_qd = 1;
1200		td->latency_qd_high = td->o.iodepth;
1201		td->latency_qd_low = 1;
1202		td->latency_ios = ddir_rw_sum(td->io_blocks);
1203	} else
1204		td->latency_qd = td->o.iodepth;
1205}
1206
1207void lat_target_reset(struct thread_data *td)
1208{
1209	if (!td->latency_end_run)
1210		lat_target_init(td);
1211}
1212
1213static void lat_target_success(struct thread_data *td)
1214{
1215	const unsigned int qd = td->latency_qd;
1216	struct thread_options *o = &td->o;
1217
1218	td->latency_qd_low = td->latency_qd;
1219
1220	/*
1221	 * If we haven't failed yet, we double up to a failing value instead
1222	 * of bisecting from highest possible queue depth. If we have set
1223	 * a limit other than td->o.iodepth, bisect between that.
1224	 */
1225	if (td->latency_qd_high != o->iodepth)
1226		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1227	else
1228		td->latency_qd *= 2;
1229
1230	if (td->latency_qd > o->iodepth)
1231		td->latency_qd = o->iodepth;
1232
1233	dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1234
1235	/*
1236	 * Same as last one, we are done. Let it run a latency cycle, so
1237	 * we get only the results from the targeted depth.
1238	 */
1239	if (td->latency_qd == qd) {
1240		if (td->latency_end_run) {
1241			dprint(FD_RATE, "We are done\n");
1242			td->done = 1;
1243		} else {
1244			dprint(FD_RATE, "Quiesce and final run\n");
1245			io_u_quiesce(td);
1246			td->latency_end_run = 1;
1247			reset_all_stats(td);
1248			reset_io_stats(td);
1249		}
1250	}
1251
1252	lat_new_cycle(td);
1253}
1254
1255/*
1256 * Check if we can bump the queue depth
1257 */
1258void lat_target_check(struct thread_data *td)
1259{
1260	uint64_t usec_window;
1261	uint64_t ios;
1262	double success_ios;
1263
1264	usec_window = utime_since_now(&td->latency_ts);
1265	if (usec_window < td->o.latency_window)
1266		return;
1267
1268	ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1269	success_ios = (double) (ios - td->latency_failed) / (double) ios;
1270	success_ios *= 100.0;
1271
1272	dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1273
1274	if (success_ios >= td->o.latency_percentile.u.f)
1275		lat_target_success(td);
1276	else
1277		__lat_target_failed(td);
1278}
1279
1280/*
1281 * If latency target is enabled, we might be ramping up or down and not
1282 * using the full queue depth available.
1283 */
1284int queue_full(struct thread_data *td)
1285{
1286	const int qempty = io_u_qempty(&td->io_u_freelist);
1287
1288	if (qempty)
1289		return 1;
1290	if (!td->o.latency_target)
1291		return 0;
1292
1293	return td->cur_depth >= td->latency_qd;
1294}
1295
1296struct io_u *__get_io_u(struct thread_data *td)
1297{
1298	struct io_u *io_u = NULL;
1299
1300	td_io_u_lock(td);
1301
1302again:
1303	if (!io_u_rempty(&td->io_u_requeues))
1304		io_u = io_u_rpop(&td->io_u_requeues);
1305	else if (!queue_full(td)) {
1306		io_u = io_u_qpop(&td->io_u_freelist);
1307
1308		io_u->file = NULL;
1309		io_u->buflen = 0;
1310		io_u->resid = 0;
1311		io_u->end_io = NULL;
1312	}
1313
1314	if (io_u) {
1315		assert(io_u->flags & IO_U_F_FREE);
1316		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1317				 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1318				 IO_U_F_VER_LIST);
1319
1320		io_u->error = 0;
1321		io_u->acct_ddir = -1;
1322		td->cur_depth++;
1323		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1324		io_u->ipo = NULL;
1325	} else if (td->o.verify_async) {
1326		/*
1327		 * We ran out, wait for async verify threads to finish and
1328		 * return one
1329		 */
1330		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1331		goto again;
1332	}
1333
1334	td_io_u_unlock(td);
1335	return io_u;
1336}
1337
1338static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1339{
1340	if (!(td->flags & TD_F_TRIM_BACKLOG))
1341		return 0;
1342
1343	if (td->trim_entries) {
1344		int get_trim = 0;
1345
1346		if (td->trim_batch) {
1347			td->trim_batch--;
1348			get_trim = 1;
1349		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1350			 td->last_ddir != DDIR_READ) {
1351			td->trim_batch = td->o.trim_batch;
1352			if (!td->trim_batch)
1353				td->trim_batch = td->o.trim_backlog;
1354			get_trim = 1;
1355		}
1356
1357		if (get_trim && !get_next_trim(td, io_u))
1358			return 1;
1359	}
1360
1361	return 0;
1362}
1363
1364static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1365{
1366	if (!(td->flags & TD_F_VER_BACKLOG))
1367		return 0;
1368
1369	if (td->io_hist_len) {
1370		int get_verify = 0;
1371
1372		if (td->verify_batch)
1373			get_verify = 1;
1374		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1375			 td->last_ddir != DDIR_READ) {
1376			td->verify_batch = td->o.verify_batch;
1377			if (!td->verify_batch)
1378				td->verify_batch = td->o.verify_backlog;
1379			get_verify = 1;
1380		}
1381
1382		if (get_verify && !get_next_verify(td, io_u)) {
1383			td->verify_batch--;
1384			return 1;
1385		}
1386	}
1387
1388	return 0;
1389}
1390
1391/*
1392 * Fill offset and start time into the buffer content, to prevent too
1393 * easy compressible data for simple de-dupe attempts. Do this for every
1394 * 512b block in the range, since that should be the smallest block size
1395 * we can expect from a device.
1396 */
1397static void small_content_scramble(struct io_u *io_u)
1398{
1399	unsigned int i, nr_blocks = io_u->buflen / 512;
1400	uint64_t boffset;
1401	unsigned int offset;
1402	void *p, *end;
1403
1404	if (!nr_blocks)
1405		return;
1406
1407	p = io_u->xfer_buf;
1408	boffset = io_u->offset;
1409	io_u->buf_filled_len = 0;
1410
1411	for (i = 0; i < nr_blocks; i++) {
1412		/*
1413		 * Fill the byte offset into a "random" start offset of
1414		 * the buffer, given by the product of the usec time
1415		 * and the actual offset.
1416		 */
1417		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1418		offset &= ~(sizeof(uint64_t) - 1);
1419		if (offset >= 512 - sizeof(uint64_t))
1420			offset -= sizeof(uint64_t);
1421		memcpy(p + offset, &boffset, sizeof(boffset));
1422
1423		end = p + 512 - sizeof(io_u->start_time);
1424		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1425		p += 512;
1426		boffset += 512;
1427	}
1428}
1429
1430/*
1431 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1432 * etc. The returned io_u is fully ready to be prepped and submitted.
1433 */
1434struct io_u *get_io_u(struct thread_data *td)
1435{
1436	struct fio_file *f;
1437	struct io_u *io_u;
1438	int do_scramble = 0;
1439	long ret = 0;
1440
1441	io_u = __get_io_u(td);
1442	if (!io_u) {
1443		dprint(FD_IO, "__get_io_u failed\n");
1444		return NULL;
1445	}
1446
1447	if (check_get_verify(td, io_u))
1448		goto out;
1449	if (check_get_trim(td, io_u))
1450		goto out;
1451
1452	/*
1453	 * from a requeue, io_u already setup
1454	 */
1455	if (io_u->file)
1456		goto out;
1457
1458	/*
1459	 * If using an iolog, grab next piece if any available.
1460	 */
1461	if (td->flags & TD_F_READ_IOLOG) {
1462		if (read_iolog_get(td, io_u))
1463			goto err_put;
1464	} else if (set_io_u_file(td, io_u)) {
1465		ret = -EBUSY;
1466		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1467		goto err_put;
1468	}
1469
1470	f = io_u->file;
1471	if (!f) {
1472		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1473		goto err_put;
1474	}
1475
1476	assert(fio_file_open(f));
1477
1478	if (ddir_rw(io_u->ddir)) {
1479		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1480			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1481			goto err_put;
1482		}
1483
1484		f->last_start = io_u->offset;
1485		f->last_pos = io_u->offset + io_u->buflen;
1486
1487		if (io_u->ddir == DDIR_WRITE) {
1488			if (td->flags & TD_F_REFILL_BUFFERS) {
1489				io_u_fill_buffer(td, io_u,
1490					io_u->xfer_buflen, io_u->xfer_buflen);
1491			} else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1492				   !(td->flags & TD_F_COMPRESS))
1493				do_scramble = 1;
1494			if (td->flags & TD_F_VER_NONE) {
1495				populate_verify_io_u(td, io_u);
1496				do_scramble = 0;
1497			}
1498		} else if (io_u->ddir == DDIR_READ) {
1499			/*
1500			 * Reset the buf_filled parameters so next time if the
1501			 * buffer is used for writes it is refilled.
1502			 */
1503			io_u->buf_filled_len = 0;
1504		}
1505	}
1506
1507	/*
1508	 * Set io data pointers.
1509	 */
1510	io_u->xfer_buf = io_u->buf;
1511	io_u->xfer_buflen = io_u->buflen;
1512
1513out:
1514	assert(io_u->file);
1515	if (!td_io_prep(td, io_u)) {
1516		if (!td->o.disable_slat)
1517			fio_gettime(&io_u->start_time, NULL);
1518		if (do_scramble)
1519			small_content_scramble(io_u);
1520		return io_u;
1521	}
1522err_put:
1523	dprint(FD_IO, "get_io_u failed\n");
1524	put_io_u(td, io_u);
1525	return ERR_PTR(ret);
1526}
1527
1528void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1529{
1530	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1531	const char *msg[] = { "read", "write", "sync", "datasync",
1532				"sync_file_range", "wait", "trim" };
1533
1534	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1535		return;
1536
1537	log_err("fio: io_u error");
1538
1539	if (io_u->file)
1540		log_err(" on file %s", io_u->file->file_name);
1541
1542	log_err(": %s\n", strerror(io_u->error));
1543
1544	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1545					io_u->offset, io_u->xfer_buflen);
1546
1547	if (!td->error)
1548		td_verror(td, io_u->error, "io_u error");
1549}
1550
1551static inline int gtod_reduce(struct thread_data *td)
1552{
1553	return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1554		&& td->o.disable_bw;
1555}
1556
1557static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1558				  struct io_completion_data *icd,
1559				  const enum fio_ddir idx, unsigned int bytes)
1560{
1561	unsigned long lusec = 0;
1562
1563	if (!gtod_reduce(td))
1564		lusec = utime_since(&io_u->issue_time, &icd->time);
1565
1566	if (!td->o.disable_lat) {
1567		unsigned long tusec;
1568
1569		tusec = utime_since(&io_u->start_time, &icd->time);
1570		add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1571
1572		if (td->flags & TD_F_PROFILE_OPS) {
1573			struct prof_io_ops *ops = &td->prof_io_ops;
1574
1575			if (ops->io_u_lat)
1576				icd->error = ops->io_u_lat(td, tusec);
1577		}
1578
1579		if (td->o.max_latency && tusec > td->o.max_latency)
1580			lat_fatal(td, icd, tusec, td->o.max_latency);
1581		if (td->o.latency_target && tusec > td->o.latency_target) {
1582			if (lat_target_failed(td))
1583				lat_fatal(td, icd, tusec, td->o.latency_target);
1584		}
1585	}
1586
1587	if (!td->o.disable_clat) {
1588		add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1589		io_u_mark_latency(td, lusec);
1590	}
1591
1592	if (!td->o.disable_bw)
1593		add_bw_sample(td, idx, bytes, &icd->time);
1594
1595	if (!gtod_reduce(td))
1596		add_iops_sample(td, idx, bytes, &icd->time);
1597}
1598
1599static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1600{
1601	uint64_t secs, remainder, bps, bytes;
1602
1603	bytes = td->this_io_bytes[ddir];
1604	bps = td->rate_bps[ddir];
1605	secs = bytes / bps;
1606	remainder = bytes % bps;
1607	return remainder * 1000000 / bps + secs * 1000000;
1608}
1609
1610static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1611			 struct io_completion_data *icd)
1612{
1613	struct io_u *io_u = *io_u_ptr;
1614	enum fio_ddir ddir = io_u->ddir;
1615	struct fio_file *f = io_u->file;
1616
1617	dprint_io_u(io_u, "io complete");
1618
1619	td_io_u_lock(td);
1620	assert(io_u->flags & IO_U_F_FLIGHT);
1621	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1622
1623	/*
1624	 * Mark IO ok to verify
1625	 */
1626	if (io_u->ipo) {
1627		/*
1628		 * Remove errored entry from the verification list
1629		 */
1630		if (io_u->error)
1631			unlog_io_piece(td, io_u);
1632		else {
1633			io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1634			write_barrier();
1635		}
1636	}
1637
1638	td_io_u_unlock(td);
1639
1640	if (ddir_sync(ddir)) {
1641		td->last_was_sync = 1;
1642		if (f) {
1643			f->first_write = -1ULL;
1644			f->last_write = -1ULL;
1645		}
1646		return;
1647	}
1648
1649	td->last_was_sync = 0;
1650	td->last_ddir = ddir;
1651
1652	if (!io_u->error && ddir_rw(ddir)) {
1653		unsigned int bytes = io_u->buflen - io_u->resid;
1654		const enum fio_ddir oddir = ddir ^ 1;
1655		int ret;
1656
1657		td->io_blocks[ddir]++;
1658		td->this_io_blocks[ddir]++;
1659		td->io_bytes[ddir] += bytes;
1660
1661		if (!(io_u->flags & IO_U_F_VER_LIST))
1662			td->this_io_bytes[ddir] += bytes;
1663
1664		if (ddir == DDIR_WRITE && f) {
1665			if (f->first_write == -1ULL ||
1666			    io_u->offset < f->first_write)
1667				f->first_write = io_u->offset;
1668			if (f->last_write == -1ULL ||
1669			    ((io_u->offset + bytes) > f->last_write))
1670				f->last_write = io_u->offset + bytes;
1671		}
1672
1673		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1674					   td->runstate == TD_VERIFYING)) {
1675			account_io_completion(td, io_u, icd, ddir, bytes);
1676
1677			if (__should_check_rate(td, ddir)) {
1678				td->rate_pending_usleep[ddir] =
1679					(usec_for_io(td, ddir) -
1680					 utime_since_now(&td->start));
1681			}
1682			if (ddir != DDIR_TRIM &&
1683			    __should_check_rate(td, oddir)) {
1684				td->rate_pending_usleep[oddir] =
1685					(usec_for_io(td, oddir) -
1686					 utime_since_now(&td->start));
1687			}
1688		}
1689
1690		icd->bytes_done[ddir] += bytes;
1691
1692		if (io_u->end_io) {
1693			ret = io_u->end_io(td, io_u_ptr);
1694			io_u = *io_u_ptr;
1695			if (ret && !icd->error)
1696				icd->error = ret;
1697		}
1698	} else if (io_u->error) {
1699		icd->error = io_u->error;
1700		io_u_log_error(td, io_u);
1701	}
1702	if (icd->error) {
1703		enum error_type_bit eb = td_error_type(ddir, icd->error);
1704
1705		if (!td_non_fatal_error(td, eb, icd->error))
1706			return;
1707
1708		/*
1709		 * If there is a non_fatal error, then add to the error count
1710		 * and clear all the errors.
1711		 */
1712		update_error_count(td, icd->error);
1713		td_clear_error(td);
1714		icd->error = 0;
1715		if (io_u)
1716			io_u->error = 0;
1717	}
1718}
1719
1720static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1721		     int nr)
1722{
1723	int ddir;
1724
1725	if (!gtod_reduce(td))
1726		fio_gettime(&icd->time, NULL);
1727
1728	icd->nr = nr;
1729
1730	icd->error = 0;
1731	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1732		icd->bytes_done[ddir] = 0;
1733}
1734
1735static void ios_completed(struct thread_data *td,
1736			  struct io_completion_data *icd)
1737{
1738	struct io_u *io_u;
1739	int i;
1740
1741	for (i = 0; i < icd->nr; i++) {
1742		io_u = td->io_ops->event(td, i);
1743
1744		io_completed(td, &io_u, icd);
1745
1746		if (io_u)
1747			put_io_u(td, io_u);
1748	}
1749}
1750
1751/*
1752 * Complete a single io_u for the sync engines.
1753 */
1754int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1755		       uint64_t *bytes)
1756{
1757	struct io_completion_data icd;
1758
1759	init_icd(td, &icd, 1);
1760	io_completed(td, &io_u, &icd);
1761
1762	if (io_u)
1763		put_io_u(td, io_u);
1764
1765	if (icd.error) {
1766		td_verror(td, icd.error, "io_u_sync_complete");
1767		return -1;
1768	}
1769
1770	if (bytes) {
1771		int ddir;
1772
1773		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1774			bytes[ddir] += icd.bytes_done[ddir];
1775	}
1776
1777	return 0;
1778}
1779
1780/*
1781 * Called to complete min_events number of io for the async engines.
1782 */
1783int io_u_queued_complete(struct thread_data *td, int min_evts,
1784			 uint64_t *bytes)
1785{
1786	struct io_completion_data icd;
1787	struct timespec *tvp = NULL;
1788	int ret;
1789	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1790
1791	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1792
1793	if (!min_evts)
1794		tvp = &ts;
1795	else if (min_evts > td->cur_depth)
1796		min_evts = td->cur_depth;
1797
1798	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1799	if (ret < 0) {
1800		td_verror(td, -ret, "td_io_getevents");
1801		return ret;
1802	} else if (!ret)
1803		return ret;
1804
1805	init_icd(td, &icd, ret);
1806	ios_completed(td, &icd);
1807	if (icd.error) {
1808		td_verror(td, icd.error, "io_u_queued_complete");
1809		return -1;
1810	}
1811
1812	if (bytes) {
1813		int ddir;
1814
1815		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1816			bytes[ddir] += icd.bytes_done[ddir];
1817	}
1818
1819	return 0;
1820}
1821
1822/*
1823 * Call when io_u is really queued, to update the submission latency.
1824 */
1825void io_u_queued(struct thread_data *td, struct io_u *io_u)
1826{
1827	if (!td->o.disable_slat) {
1828		unsigned long slat_time;
1829
1830		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1831		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1832				io_u->offset);
1833	}
1834}
1835
1836void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1837		    unsigned int max_bs)
1838{
1839	if (td->o.buffer_pattern_bytes)
1840		fill_buffer_pattern(td, buf, max_bs);
1841	else if (!td->o.zero_buffers) {
1842		unsigned int perc = td->o.compress_percentage;
1843
1844		if (perc) {
1845			unsigned int seg = min_write;
1846
1847			seg = min(min_write, td->o.compress_chunk);
1848			if (!seg)
1849				seg = min_write;
1850
1851			fill_random_buf_percentage(&td->buf_state, buf,
1852						perc, seg, max_bs);
1853		} else
1854			fill_random_buf(&td->buf_state, buf, max_bs);
1855	} else
1856		memset(buf, 0, max_bs);
1857}
1858
1859/*
1860 * "randomly" fill the buffer contents
1861 */
1862void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1863		      unsigned int min_write, unsigned int max_bs)
1864{
1865	io_u->buf_filled_len = 0;
1866	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1867}
1868