io_u.c revision 225ba9e3433cf27d8ff7b213d9f78b7ef2776c70
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 - 1) * (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_entry(td->next_rand_list.next, 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) && 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);
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(io_u->ddir));
430
431	if (td->o.bs_is_seq_rand)
432		ddir = is_random ? DDIR_WRITE: DDIR_READ;
433	else
434		ddir = io_u->ddir;
435
436	minbs = td->o.min_bs[ddir];
437	maxbs = td->o.max_bs[ddir];
438
439	if (minbs == maxbs)
440		return minbs;
441
442	/*
443	 * If we can't satisfy the min block size from here, then fail
444	 */
445	if (!io_u_fits(td, io_u, minbs))
446		return 0;
447
448	if (td->o.use_os_rand)
449		rand_max = OS_RAND_MAX;
450	else
451		rand_max = FRAND_MAX;
452
453	do {
454		if (td->o.use_os_rand)
455			r = os_random_long(&td->bsrange_state);
456		else
457			r = __rand(&td->__bsrange_state);
458
459		if (!td->o.bssplit_nr[ddir]) {
460			buflen = 1 + (unsigned int) ((double) maxbs *
461					(r / (rand_max + 1.0)));
462			if (buflen < minbs)
463				buflen = minbs;
464		} else {
465			long perc = 0;
466			unsigned int i;
467
468			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
469				struct bssplit *bsp = &td->o.bssplit[ddir][i];
470
471				buflen = bsp->bs;
472				perc += bsp->perc;
473				if ((r <= ((rand_max / 100L) * perc)) &&
474				    io_u_fits(td, io_u, buflen))
475					break;
476			}
477		}
478
479		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
480			buflen = (buflen + td->o.verify_interval - 1) &
481				~(td->o.verify_interval - 1);
482
483		if (!td->o.bs_unaligned && is_power_of_2(minbs))
484			buflen = (buflen + minbs - 1) & ~(minbs - 1);
485
486	} while (!io_u_fits(td, io_u, buflen));
487
488	return buflen;
489}
490
491static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
492				    unsigned int is_random)
493{
494	if (td->flags & TD_F_PROFILE_OPS) {
495		struct prof_io_ops *ops = &td->prof_io_ops;
496
497		if (ops->fill_io_u_size)
498			return ops->fill_io_u_size(td, io_u, is_random);
499	}
500
501	return __get_next_buflen(td, io_u, is_random);
502}
503
504static void set_rwmix_bytes(struct thread_data *td)
505{
506	unsigned int diff;
507
508	/*
509	 * we do time or byte based switch. this is needed because
510	 * buffered writes may issue a lot quicker than they complete,
511	 * whereas reads do not.
512	 */
513	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
514	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
515}
516
517static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
518{
519	unsigned int v;
520	unsigned long r;
521
522	if (td->o.use_os_rand) {
523		r = os_random_long(&td->rwmix_state);
524		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
525	} else {
526		r = __rand(&td->__rwmix_state);
527		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
528	}
529
530	if (v <= td->o.rwmix[DDIR_READ])
531		return DDIR_READ;
532
533	return DDIR_WRITE;
534}
535
536void io_u_quiesce(struct thread_data *td)
537{
538	/*
539	 * We are going to sleep, ensure that we flush anything pending as
540	 * not to skew our latency numbers.
541	 *
542	 * Changed to only monitor 'in flight' requests here instead of the
543	 * td->cur_depth, b/c td->cur_depth does not accurately represent
544	 * io's that have been actually submitted to an async engine,
545	 * and cur_depth is meaningless for sync engines.
546	 */
547	while (td->io_u_in_flight) {
548		int fio_unused ret;
549
550		ret = io_u_queued_complete(td, 1, NULL);
551	}
552}
553
554static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
555{
556	enum fio_ddir odir = ddir ^ 1;
557	struct timeval t;
558	long usec;
559
560	assert(ddir_rw(ddir));
561
562	if (td->rate_pending_usleep[ddir] <= 0)
563		return ddir;
564
565	/*
566	 * We have too much pending sleep in this direction. See if we
567	 * should switch.
568	 */
569	if (td_rw(td) && td->o.rwmix[odir]) {
570		/*
571		 * Other direction does not have too much pending, switch
572		 */
573		if (td->rate_pending_usleep[odir] < 100000)
574			return odir;
575
576		/*
577		 * Both directions have pending sleep. Sleep the minimum time
578		 * and deduct from both.
579		 */
580		if (td->rate_pending_usleep[ddir] <=
581			td->rate_pending_usleep[odir]) {
582			usec = td->rate_pending_usleep[ddir];
583		} else {
584			usec = td->rate_pending_usleep[odir];
585			ddir = odir;
586		}
587	} else
588		usec = td->rate_pending_usleep[ddir];
589
590	io_u_quiesce(td);
591
592	fio_gettime(&t, NULL);
593	usec_sleep(td, usec);
594	usec = utime_since_now(&t);
595
596	td->rate_pending_usleep[ddir] -= usec;
597
598	odir = ddir ^ 1;
599	if (td_rw(td) && __should_check_rate(td, odir))
600		td->rate_pending_usleep[odir] -= usec;
601
602	if (ddir_trim(ddir))
603		return ddir;
604
605	return ddir;
606}
607
608/*
609 * Return the data direction for the next io_u. If the job is a
610 * mixed read/write workload, check the rwmix cycle and switch if
611 * necessary.
612 */
613static enum fio_ddir get_rw_ddir(struct thread_data *td)
614{
615	enum fio_ddir ddir;
616
617	/*
618	 * see if it's time to fsync
619	 */
620	if (td->o.fsync_blocks &&
621	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
622	     td->io_issues[DDIR_WRITE] && should_fsync(td))
623		return DDIR_SYNC;
624
625	/*
626	 * see if it's time to fdatasync
627	 */
628	if (td->o.fdatasync_blocks &&
629	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
630	     td->io_issues[DDIR_WRITE] && should_fsync(td))
631		return DDIR_DATASYNC;
632
633	/*
634	 * see if it's time to sync_file_range
635	 */
636	if (td->sync_file_range_nr &&
637	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
638	     td->io_issues[DDIR_WRITE] && should_fsync(td))
639		return DDIR_SYNC_FILE_RANGE;
640
641	if (td_rw(td)) {
642		/*
643		 * Check if it's time to seed a new data direction.
644		 */
645		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
646			/*
647			 * Put a top limit on how many bytes we do for
648			 * one data direction, to avoid overflowing the
649			 * ranges too much
650			 */
651			ddir = get_rand_ddir(td);
652
653			if (ddir != td->rwmix_ddir)
654				set_rwmix_bytes(td);
655
656			td->rwmix_ddir = ddir;
657		}
658		ddir = td->rwmix_ddir;
659	} else if (td_read(td))
660		ddir = DDIR_READ;
661	else if (td_write(td))
662		ddir = DDIR_WRITE;
663	else
664		ddir = DDIR_TRIM;
665
666	td->rwmix_ddir = rate_ddir(td, ddir);
667	return td->rwmix_ddir;
668}
669
670static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
671{
672	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
673
674	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
675	    td->o.barrier_blocks &&
676	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
677	     td->io_issues[DDIR_WRITE])
678		io_u->flags |= IO_U_F_BARRIER;
679}
680
681void put_file_log(struct thread_data *td, struct fio_file *f)
682{
683	int ret = put_file(td, f);
684
685	if (ret)
686		td_verror(td, ret, "file close");
687}
688
689void put_io_u(struct thread_data *td, struct io_u *io_u)
690{
691	td_io_u_lock(td);
692
693	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
694		put_file_log(td, io_u->file);
695	io_u->file = NULL;
696	io_u->flags &= ~IO_U_F_FREE_DEF;
697	io_u->flags |= IO_U_F_FREE;
698
699	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
700		td->cur_depth--;
701	io_u_qpush(&td->io_u_freelist, io_u);
702	td_io_u_unlock(td);
703	td_io_u_free_notify(td);
704}
705
706void clear_io_u(struct thread_data *td, struct io_u *io_u)
707{
708	io_u->flags &= ~IO_U_F_FLIGHT;
709	put_io_u(td, io_u);
710}
711
712void requeue_io_u(struct thread_data *td, struct io_u **io_u)
713{
714	struct io_u *__io_u = *io_u;
715	enum fio_ddir ddir = acct_ddir(__io_u);
716
717	dprint(FD_IO, "requeue %p\n", __io_u);
718
719	td_io_u_lock(td);
720
721	__io_u->flags |= IO_U_F_FREE;
722	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
723		td->io_issues[ddir]--;
724
725	__io_u->flags &= ~IO_U_F_FLIGHT;
726	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
727		td->cur_depth--;
728
729	io_u_rpush(&td->io_u_requeues, __io_u);
730	td_io_u_unlock(td);
731	*io_u = NULL;
732}
733
734static int fill_io_u(struct thread_data *td, struct io_u *io_u)
735{
736	unsigned int is_random;
737
738	if (td->io_ops->flags & FIO_NOIO)
739		goto out;
740
741	set_rw_ddir(td, io_u);
742
743	/*
744	 * fsync() or fdatasync() or trim etc, we are done
745	 */
746	if (!ddir_rw(io_u->ddir))
747		goto out;
748
749	/*
750	 * See if it's time to switch to a new zone
751	 */
752	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
753		td->zone_bytes = 0;
754		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
755		io_u->file->last_pos = io_u->file->file_offset;
756		td->io_skip_bytes += td->o.zone_skip;
757	}
758
759	/*
760	 * No log, let the seq/rand engine retrieve the next buflen and
761	 * position.
762	 */
763	if (get_next_offset(td, io_u, &is_random)) {
764		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
765		return 1;
766	}
767
768	io_u->buflen = get_next_buflen(td, io_u, is_random);
769	if (!io_u->buflen) {
770		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
771		return 1;
772	}
773
774	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
775		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
776		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
777			(unsigned long long) io_u->offset, io_u->buflen,
778			(unsigned long long) io_u->file->real_file_size);
779		return 1;
780	}
781
782	/*
783	 * mark entry before potentially trimming io_u
784	 */
785	if (td_random(td) && file_randommap(td, io_u->file))
786		mark_random_map(td, io_u);
787
788out:
789	dprint_io_u(io_u, "fill_io_u");
790	td->zone_bytes += io_u->buflen;
791	return 0;
792}
793
794static void __io_u_mark_map(unsigned int *map, unsigned int nr)
795{
796	int idx = 0;
797
798	switch (nr) {
799	default:
800		idx = 6;
801		break;
802	case 33 ... 64:
803		idx = 5;
804		break;
805	case 17 ... 32:
806		idx = 4;
807		break;
808	case 9 ... 16:
809		idx = 3;
810		break;
811	case 5 ... 8:
812		idx = 2;
813		break;
814	case 1 ... 4:
815		idx = 1;
816	case 0:
817		break;
818	}
819
820	map[idx]++;
821}
822
823void io_u_mark_submit(struct thread_data *td, unsigned int nr)
824{
825	__io_u_mark_map(td->ts.io_u_submit, nr);
826	td->ts.total_submit++;
827}
828
829void io_u_mark_complete(struct thread_data *td, unsigned int nr)
830{
831	__io_u_mark_map(td->ts.io_u_complete, nr);
832	td->ts.total_complete++;
833}
834
835void io_u_mark_depth(struct thread_data *td, unsigned int nr)
836{
837	int idx = 0;
838
839	switch (td->cur_depth) {
840	default:
841		idx = 6;
842		break;
843	case 32 ... 63:
844		idx = 5;
845		break;
846	case 16 ... 31:
847		idx = 4;
848		break;
849	case 8 ... 15:
850		idx = 3;
851		break;
852	case 4 ... 7:
853		idx = 2;
854		break;
855	case 2 ... 3:
856		idx = 1;
857	case 1:
858		break;
859	}
860
861	td->ts.io_u_map[idx] += nr;
862}
863
864static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
865{
866	int idx = 0;
867
868	assert(usec < 1000);
869
870	switch (usec) {
871	case 750 ... 999:
872		idx = 9;
873		break;
874	case 500 ... 749:
875		idx = 8;
876		break;
877	case 250 ... 499:
878		idx = 7;
879		break;
880	case 100 ... 249:
881		idx = 6;
882		break;
883	case 50 ... 99:
884		idx = 5;
885		break;
886	case 20 ... 49:
887		idx = 4;
888		break;
889	case 10 ... 19:
890		idx = 3;
891		break;
892	case 4 ... 9:
893		idx = 2;
894		break;
895	case 2 ... 3:
896		idx = 1;
897	case 0 ... 1:
898		break;
899	}
900
901	assert(idx < FIO_IO_U_LAT_U_NR);
902	td->ts.io_u_lat_u[idx]++;
903}
904
905static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
906{
907	int idx = 0;
908
909	switch (msec) {
910	default:
911		idx = 11;
912		break;
913	case 1000 ... 1999:
914		idx = 10;
915		break;
916	case 750 ... 999:
917		idx = 9;
918		break;
919	case 500 ... 749:
920		idx = 8;
921		break;
922	case 250 ... 499:
923		idx = 7;
924		break;
925	case 100 ... 249:
926		idx = 6;
927		break;
928	case 50 ... 99:
929		idx = 5;
930		break;
931	case 20 ... 49:
932		idx = 4;
933		break;
934	case 10 ... 19:
935		idx = 3;
936		break;
937	case 4 ... 9:
938		idx = 2;
939		break;
940	case 2 ... 3:
941		idx = 1;
942	case 0 ... 1:
943		break;
944	}
945
946	assert(idx < FIO_IO_U_LAT_M_NR);
947	td->ts.io_u_lat_m[idx]++;
948}
949
950static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
951{
952	if (usec < 1000)
953		io_u_mark_lat_usec(td, usec);
954	else
955		io_u_mark_lat_msec(td, usec / 1000);
956}
957
958/*
959 * Get next file to service by choosing one at random
960 */
961static struct fio_file *get_next_file_rand(struct thread_data *td,
962					   enum fio_file_flags goodf,
963					   enum fio_file_flags badf)
964{
965	struct fio_file *f;
966	int fno;
967
968	do {
969		int opened = 0;
970		unsigned long r;
971
972		if (td->o.use_os_rand) {
973			r = os_random_long(&td->next_file_state);
974			fno = (unsigned int) ((double) td->o.nr_files
975				* (r / (OS_RAND_MAX + 1.0)));
976		} else {
977			r = __rand(&td->__next_file_state);
978			fno = (unsigned int) ((double) td->o.nr_files
979				* (r / (FRAND_MAX + 1.0)));
980		}
981
982		f = td->files[fno];
983		if (fio_file_done(f))
984			continue;
985
986		if (!fio_file_open(f)) {
987			int err;
988
989			if (td->nr_open_files >= td->o.open_files)
990				return ERR_PTR(-EBUSY);
991
992			err = td_io_open_file(td, f);
993			if (err)
994				continue;
995			opened = 1;
996		}
997
998		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
999			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1000			return f;
1001		}
1002		if (opened)
1003			td_io_close_file(td, f);
1004	} while (1);
1005}
1006
1007/*
1008 * Get next file to service by doing round robin between all available ones
1009 */
1010static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1011					 int badf)
1012{
1013	unsigned int old_next_file = td->next_file;
1014	struct fio_file *f;
1015
1016	do {
1017		int opened = 0;
1018
1019		f = td->files[td->next_file];
1020
1021		td->next_file++;
1022		if (td->next_file >= td->o.nr_files)
1023			td->next_file = 0;
1024
1025		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1026		if (fio_file_done(f)) {
1027			f = NULL;
1028			continue;
1029		}
1030
1031		if (!fio_file_open(f)) {
1032			int err;
1033
1034			if (td->nr_open_files >= td->o.open_files)
1035				return ERR_PTR(-EBUSY);
1036
1037			err = td_io_open_file(td, f);
1038			if (err) {
1039				dprint(FD_FILE, "error %d on open of %s\n",
1040					err, f->file_name);
1041				f = NULL;
1042				continue;
1043			}
1044			opened = 1;
1045		}
1046
1047		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1048								f->flags);
1049		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1050			break;
1051
1052		if (opened)
1053			td_io_close_file(td, f);
1054
1055		f = NULL;
1056	} while (td->next_file != old_next_file);
1057
1058	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1059	return f;
1060}
1061
1062static struct fio_file *__get_next_file(struct thread_data *td)
1063{
1064	struct fio_file *f;
1065
1066	assert(td->o.nr_files <= td->files_index);
1067
1068	if (td->nr_done_files >= td->o.nr_files) {
1069		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1070				" nr_files=%d\n", td->nr_open_files,
1071						  td->nr_done_files,
1072						  td->o.nr_files);
1073		return NULL;
1074	}
1075
1076	f = td->file_service_file;
1077	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1078		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1079			goto out;
1080		if (td->file_service_left--)
1081			goto out;
1082	}
1083
1084	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1085	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1086		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1087	else
1088		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1089
1090	if (IS_ERR(f))
1091		return f;
1092
1093	td->file_service_file = f;
1094	td->file_service_left = td->file_service_nr - 1;
1095out:
1096	if (f)
1097		dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1098	else
1099		dprint(FD_FILE, "get_next_file: NULL\n");
1100	return f;
1101}
1102
1103static struct fio_file *get_next_file(struct thread_data *td)
1104{
1105	if (!(td->flags & TD_F_PROFILE_OPS)) {
1106		struct prof_io_ops *ops = &td->prof_io_ops;
1107
1108		if (ops->get_next_file)
1109			return ops->get_next_file(td);
1110	}
1111
1112	return __get_next_file(td);
1113}
1114
1115static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1116{
1117	struct fio_file *f;
1118
1119	do {
1120		f = get_next_file(td);
1121		if (IS_ERR_OR_NULL(f))
1122			return PTR_ERR(f);
1123
1124		io_u->file = f;
1125		get_file(f);
1126
1127		if (!fill_io_u(td, io_u))
1128			break;
1129
1130		put_file_log(td, f);
1131		td_io_close_file(td, f);
1132		io_u->file = NULL;
1133		fio_file_set_done(f);
1134		td->nr_done_files++;
1135		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1136					td->nr_done_files, td->o.nr_files);
1137	} while (1);
1138
1139	return 0;
1140}
1141
1142static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1143		      unsigned long tusec, unsigned long max_usec)
1144{
1145	if (!td->error)
1146		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1147	td_verror(td, ETIMEDOUT, "max latency exceeded");
1148	icd->error = ETIMEDOUT;
1149}
1150
1151static void lat_new_cycle(struct thread_data *td)
1152{
1153	fio_gettime(&td->latency_ts, NULL);
1154	td->latency_ios = ddir_rw_sum(td->io_blocks);
1155	td->latency_failed = 0;
1156}
1157
1158/*
1159 * We had an IO outside the latency target. Reduce the queue depth. If we
1160 * are at QD=1, then it's time to give up.
1161 */
1162static int __lat_target_failed(struct thread_data *td)
1163{
1164	if (td->latency_qd == 1)
1165		return 1;
1166
1167	td->latency_qd_high = td->latency_qd;
1168
1169	if (td->latency_qd == td->latency_qd_low)
1170		td->latency_qd_low--;
1171
1172	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1173
1174	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1175
1176	/*
1177	 * When we ramp QD down, quiesce existing IO to prevent
1178	 * a storm of ramp downs due to pending higher depth.
1179	 */
1180	io_u_quiesce(td);
1181	lat_new_cycle(td);
1182	return 0;
1183}
1184
1185static int lat_target_failed(struct thread_data *td)
1186{
1187	if (td->o.latency_percentile.u.f == 100.0)
1188		return __lat_target_failed(td);
1189
1190	td->latency_failed++;
1191	return 0;
1192}
1193
1194void lat_target_init(struct thread_data *td)
1195{
1196	td->latency_end_run = 0;
1197
1198	if (td->o.latency_target) {
1199		dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1200		fio_gettime(&td->latency_ts, NULL);
1201		td->latency_qd = 1;
1202		td->latency_qd_high = td->o.iodepth;
1203		td->latency_qd_low = 1;
1204		td->latency_ios = ddir_rw_sum(td->io_blocks);
1205	} else
1206		td->latency_qd = td->o.iodepth;
1207}
1208
1209void lat_target_reset(struct thread_data *td)
1210{
1211	if (!td->latency_end_run)
1212		lat_target_init(td);
1213}
1214
1215static void lat_target_success(struct thread_data *td)
1216{
1217	const unsigned int qd = td->latency_qd;
1218	struct thread_options *o = &td->o;
1219
1220	td->latency_qd_low = td->latency_qd;
1221
1222	/*
1223	 * If we haven't failed yet, we double up to a failing value instead
1224	 * of bisecting from highest possible queue depth. If we have set
1225	 * a limit other than td->o.iodepth, bisect between that.
1226	 */
1227	if (td->latency_qd_high != o->iodepth)
1228		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1229	else
1230		td->latency_qd *= 2;
1231
1232	if (td->latency_qd > o->iodepth)
1233		td->latency_qd = o->iodepth;
1234
1235	dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1236
1237	/*
1238	 * Same as last one, we are done. Let it run a latency cycle, so
1239	 * we get only the results from the targeted depth.
1240	 */
1241	if (td->latency_qd == qd) {
1242		if (td->latency_end_run) {
1243			dprint(FD_RATE, "We are done\n");
1244			td->done = 1;
1245		} else {
1246			dprint(FD_RATE, "Quiesce and final run\n");
1247			io_u_quiesce(td);
1248			td->latency_end_run = 1;
1249			reset_all_stats(td);
1250			reset_io_stats(td);
1251		}
1252	}
1253
1254	lat_new_cycle(td);
1255}
1256
1257/*
1258 * Check if we can bump the queue depth
1259 */
1260void lat_target_check(struct thread_data *td)
1261{
1262	uint64_t usec_window;
1263	uint64_t ios;
1264	double success_ios;
1265
1266	usec_window = utime_since_now(&td->latency_ts);
1267	if (usec_window < td->o.latency_window)
1268		return;
1269
1270	ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1271	success_ios = (double) (ios - td->latency_failed) / (double) ios;
1272	success_ios *= 100.0;
1273
1274	dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1275
1276	if (success_ios >= td->o.latency_percentile.u.f)
1277		lat_target_success(td);
1278	else
1279		__lat_target_failed(td);
1280}
1281
1282/*
1283 * If latency target is enabled, we might be ramping up or down and not
1284 * using the full queue depth available.
1285 */
1286int queue_full(struct thread_data *td)
1287{
1288	const int qempty = io_u_qempty(&td->io_u_freelist);
1289
1290	if (qempty)
1291		return 1;
1292	if (!td->o.latency_target)
1293		return 0;
1294
1295	return td->cur_depth >= td->latency_qd;
1296}
1297
1298struct io_u *__get_io_u(struct thread_data *td)
1299{
1300	struct io_u *io_u;
1301
1302	td_io_u_lock(td);
1303
1304again:
1305	if (!io_u_rempty(&td->io_u_requeues))
1306		io_u = io_u_rpop(&td->io_u_requeues);
1307	else if (!queue_full(td)) {
1308		io_u = io_u_qpop(&td->io_u_freelist);
1309
1310		io_u->file = NULL;
1311		io_u->buflen = 0;
1312		io_u->resid = 0;
1313		io_u->end_io = NULL;
1314	}
1315
1316	if (io_u) {
1317		assert(io_u->flags & IO_U_F_FREE);
1318		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1319		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1320		io_u->flags &= ~IO_U_F_VER_LIST;
1321
1322		io_u->error = 0;
1323		io_u->acct_ddir = -1;
1324		td->cur_depth++;
1325		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1326		io_u->ipo = NULL;
1327	} else if (td->o.verify_async) {
1328		/*
1329		 * We ran out, wait for async verify threads to finish and
1330		 * return one
1331		 */
1332		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1333		goto again;
1334	}
1335
1336	td_io_u_unlock(td);
1337	return io_u;
1338}
1339
1340static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1341{
1342	if (!(td->flags & TD_F_TRIM_BACKLOG))
1343		return 0;
1344
1345	if (td->trim_entries) {
1346		int get_trim = 0;
1347
1348		if (td->trim_batch) {
1349			td->trim_batch--;
1350			get_trim = 1;
1351		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1352			 td->last_ddir != DDIR_READ) {
1353			td->trim_batch = td->o.trim_batch;
1354			if (!td->trim_batch)
1355				td->trim_batch = td->o.trim_backlog;
1356			get_trim = 1;
1357		}
1358
1359		if (get_trim && !get_next_trim(td, io_u))
1360			return 1;
1361	}
1362
1363	return 0;
1364}
1365
1366static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1367{
1368	if (!(td->flags & TD_F_VER_BACKLOG))
1369		return 0;
1370
1371	if (td->io_hist_len) {
1372		int get_verify = 0;
1373
1374		if (td->verify_batch)
1375			get_verify = 1;
1376		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1377			 td->last_ddir != DDIR_READ) {
1378			td->verify_batch = td->o.verify_batch;
1379			if (!td->verify_batch)
1380				td->verify_batch = td->o.verify_backlog;
1381			get_verify = 1;
1382		}
1383
1384		if (get_verify && !get_next_verify(td, io_u)) {
1385			td->verify_batch--;
1386			return 1;
1387		}
1388	}
1389
1390	return 0;
1391}
1392
1393/*
1394 * Fill offset and start time into the buffer content, to prevent too
1395 * easy compressible data for simple de-dupe attempts. Do this for every
1396 * 512b block in the range, since that should be the smallest block size
1397 * we can expect from a device.
1398 */
1399static void small_content_scramble(struct io_u *io_u)
1400{
1401	unsigned int i, nr_blocks = io_u->buflen / 512;
1402	uint64_t boffset;
1403	unsigned int offset;
1404	void *p, *end;
1405
1406	if (!nr_blocks)
1407		return;
1408
1409	p = io_u->xfer_buf;
1410	boffset = io_u->offset;
1411	io_u->buf_filled_len = 0;
1412
1413	for (i = 0; i < nr_blocks; i++) {
1414		/*
1415		 * Fill the byte offset into a "random" start offset of
1416		 * the buffer, given by the product of the usec time
1417		 * and the actual offset.
1418		 */
1419		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1420		offset &= ~(sizeof(uint64_t) - 1);
1421		if (offset >= 512 - sizeof(uint64_t))
1422			offset -= sizeof(uint64_t);
1423		memcpy(p + offset, &boffset, sizeof(boffset));
1424
1425		end = p + 512 - sizeof(io_u->start_time);
1426		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1427		p += 512;
1428		boffset += 512;
1429	}
1430}
1431
1432/*
1433 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1434 * etc. The returned io_u is fully ready to be prepped and submitted.
1435 */
1436struct io_u *get_io_u(struct thread_data *td)
1437{
1438	struct fio_file *f;
1439	struct io_u *io_u;
1440	int do_scramble = 0;
1441	long ret = 0;
1442
1443	io_u = __get_io_u(td);
1444	if (!io_u) {
1445		dprint(FD_IO, "__get_io_u failed\n");
1446		return NULL;
1447	}
1448
1449	if (check_get_verify(td, io_u))
1450		goto out;
1451	if (check_get_trim(td, io_u))
1452		goto out;
1453
1454	/*
1455	 * from a requeue, io_u already setup
1456	 */
1457	if (io_u->file)
1458		goto out;
1459
1460	/*
1461	 * If using an iolog, grab next piece if any available.
1462	 */
1463	if (td->flags & TD_F_READ_IOLOG) {
1464		if (read_iolog_get(td, io_u))
1465			goto err_put;
1466	} else if (set_io_u_file(td, io_u)) {
1467		ret = -EBUSY;
1468		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1469		goto err_put;
1470	}
1471
1472	f = io_u->file;
1473	if (!f) {
1474		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1475		goto err_put;
1476	}
1477
1478	assert(fio_file_open(f));
1479
1480	if (ddir_rw(io_u->ddir)) {
1481		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1482			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1483			goto err_put;
1484		}
1485
1486		f->last_start = io_u->offset;
1487		f->last_pos = io_u->offset + io_u->buflen;
1488
1489		if (io_u->ddir == DDIR_WRITE) {
1490			if (td->flags & TD_F_REFILL_BUFFERS) {
1491				io_u_fill_buffer(td, io_u,
1492					io_u->xfer_buflen, io_u->xfer_buflen);
1493			} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1494				do_scramble = 1;
1495			if (td->flags & TD_F_VER_NONE) {
1496				populate_verify_io_u(td, io_u);
1497				do_scramble = 0;
1498			}
1499		} else if (io_u->ddir == DDIR_READ) {
1500			/*
1501			 * Reset the buf_filled parameters so next time if the
1502			 * buffer is used for writes it is refilled.
1503			 */
1504			io_u->buf_filled_len = 0;
1505		}
1506	}
1507
1508	/*
1509	 * Set io data pointers.
1510	 */
1511	io_u->xfer_buf = io_u->buf;
1512	io_u->xfer_buflen = io_u->buflen;
1513
1514out:
1515	assert(io_u->file);
1516	if (!td_io_prep(td, io_u)) {
1517		if (!td->o.disable_slat)
1518			fio_gettime(&io_u->start_time, NULL);
1519		if (do_scramble)
1520			small_content_scramble(io_u);
1521		return io_u;
1522	}
1523err_put:
1524	dprint(FD_IO, "get_io_u failed\n");
1525	put_io_u(td, io_u);
1526	return ERR_PTR(ret);
1527}
1528
1529void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1530{
1531	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1532	const char *msg[] = { "read", "write", "sync", "datasync",
1533				"sync_file_range", "wait", "trim" };
1534
1535	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1536		return;
1537
1538	log_err("fio: io_u error");
1539
1540	if (io_u->file)
1541		log_err(" on file %s", io_u->file->file_name);
1542
1543	log_err(": %s\n", strerror(io_u->error));
1544
1545	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1546					io_u->offset, io_u->xfer_buflen);
1547
1548	if (!td->error)
1549		td_verror(td, io_u->error, "io_u error");
1550}
1551
1552static inline int gtod_reduce(struct thread_data *td)
1553{
1554	return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1555		&& td->o.disable_bw;
1556}
1557
1558static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1559				  struct io_completion_data *icd,
1560				  const enum fio_ddir idx, unsigned int bytes)
1561{
1562	unsigned long lusec = 0;
1563
1564	if (!gtod_reduce(td))
1565		lusec = utime_since(&io_u->issue_time, &icd->time);
1566
1567	if (!td->o.disable_lat) {
1568		unsigned long tusec;
1569
1570		tusec = utime_since(&io_u->start_time, &icd->time);
1571		add_lat_sample(td, idx, tusec, bytes);
1572
1573		if (td->flags & TD_F_PROFILE_OPS) {
1574			struct prof_io_ops *ops = &td->prof_io_ops;
1575
1576			if (ops->io_u_lat)
1577				icd->error = ops->io_u_lat(td, tusec);
1578		}
1579
1580		if (td->o.max_latency && tusec > td->o.max_latency)
1581			lat_fatal(td, icd, tusec, td->o.max_latency);
1582		if (td->o.latency_target && tusec > td->o.latency_target) {
1583			if (lat_target_failed(td))
1584				lat_fatal(td, icd, tusec, td->o.latency_target);
1585		}
1586	}
1587
1588	if (!td->o.disable_clat) {
1589		add_clat_sample(td, idx, lusec, bytes);
1590		io_u_mark_latency(td, lusec);
1591	}
1592
1593	if (!td->o.disable_bw)
1594		add_bw_sample(td, idx, bytes, &icd->time);
1595
1596	if (!gtod_reduce(td))
1597		add_iops_sample(td, idx, bytes, &icd->time);
1598
1599	if (td->o.number_ios && !--td->o.number_ios)
1600		td->done = 1;
1601}
1602
1603static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1604{
1605	uint64_t secs, remainder, bps, bytes;
1606
1607	bytes = td->this_io_bytes[ddir];
1608	bps = td->rate_bps[ddir];
1609	secs = bytes / bps;
1610	remainder = bytes % bps;
1611	return remainder * 1000000 / bps + secs * 1000000;
1612}
1613
1614static void io_completed(struct thread_data *td, struct io_u *io_u,
1615			 struct io_completion_data *icd)
1616{
1617	struct fio_file *f;
1618
1619	dprint_io_u(io_u, "io complete");
1620
1621	td_io_u_lock(td);
1622	assert(io_u->flags & IO_U_F_FLIGHT);
1623	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1624
1625	/*
1626	 * Mark IO ok to verify
1627	 */
1628	if (io_u->ipo) {
1629		io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1630		write_barrier();
1631	}
1632
1633	td_io_u_unlock(td);
1634
1635	if (ddir_sync(io_u->ddir)) {
1636		td->last_was_sync = 1;
1637		f = io_u->file;
1638		if (f) {
1639			f->first_write = -1ULL;
1640			f->last_write = -1ULL;
1641		}
1642		return;
1643	}
1644
1645	td->last_was_sync = 0;
1646	td->last_ddir = io_u->ddir;
1647
1648	if (!io_u->error && ddir_rw(io_u->ddir)) {
1649		unsigned int bytes = io_u->buflen - io_u->resid;
1650		const enum fio_ddir idx = io_u->ddir;
1651		const enum fio_ddir odx = io_u->ddir ^ 1;
1652		int ret;
1653
1654		td->io_blocks[idx]++;
1655		td->this_io_blocks[idx]++;
1656		td->io_bytes[idx] += bytes;
1657
1658		if (!(io_u->flags & IO_U_F_VER_LIST))
1659			td->this_io_bytes[idx] += bytes;
1660
1661		if (idx == DDIR_WRITE) {
1662			f = io_u->file;
1663			if (f) {
1664				if (f->first_write == -1ULL ||
1665				    io_u->offset < f->first_write)
1666					f->first_write = io_u->offset;
1667				if (f->last_write == -1ULL ||
1668				    ((io_u->offset + bytes) > f->last_write))
1669					f->last_write = io_u->offset + bytes;
1670			}
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, idx, bytes);
1676
1677			if (__should_check_rate(td, idx)) {
1678				td->rate_pending_usleep[idx] =
1679					(usec_for_io(td, idx) -
1680					 utime_since_now(&td->start));
1681			}
1682			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1683				td->rate_pending_usleep[odx] =
1684					(usec_for_io(td, odx) -
1685					 utime_since_now(&td->start));
1686		}
1687
1688		icd->bytes_done[idx] += bytes;
1689
1690		if (io_u->end_io) {
1691			ret = io_u->end_io(td, io_u);
1692			if (ret && !icd->error)
1693				icd->error = ret;
1694		}
1695	} else if (io_u->error) {
1696		icd->error = io_u->error;
1697		io_u_log_error(td, io_u);
1698	}
1699	if (icd->error) {
1700		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1701		if (!td_non_fatal_error(td, eb, icd->error))
1702			return;
1703		/*
1704		 * If there is a non_fatal error, then add to the error count
1705		 * and clear all the errors.
1706		 */
1707		update_error_count(td, icd->error);
1708		td_clear_error(td);
1709		icd->error = 0;
1710		io_u->error = 0;
1711	}
1712}
1713
1714static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1715		     int nr)
1716{
1717	int ddir;
1718
1719	if (!gtod_reduce(td))
1720		fio_gettime(&icd->time, NULL);
1721
1722	icd->nr = nr;
1723
1724	icd->error = 0;
1725	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1726		icd->bytes_done[ddir] = 0;
1727}
1728
1729static void ios_completed(struct thread_data *td,
1730			  struct io_completion_data *icd)
1731{
1732	struct io_u *io_u;
1733	int i;
1734
1735	for (i = 0; i < icd->nr; i++) {
1736		io_u = td->io_ops->event(td, i);
1737
1738		io_completed(td, io_u, icd);
1739
1740		if (!(io_u->flags & IO_U_F_FREE_DEF))
1741			put_io_u(td, io_u);
1742	}
1743}
1744
1745/*
1746 * Complete a single io_u for the sync engines.
1747 */
1748int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1749		       uint64_t *bytes)
1750{
1751	struct io_completion_data icd;
1752
1753	init_icd(td, &icd, 1);
1754	io_completed(td, io_u, &icd);
1755
1756	if (!(io_u->flags & IO_U_F_FREE_DEF))
1757		put_io_u(td, io_u);
1758
1759	if (icd.error) {
1760		td_verror(td, icd.error, "io_u_sync_complete");
1761		return -1;
1762	}
1763
1764	if (bytes) {
1765		int ddir;
1766
1767		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1768			bytes[ddir] += icd.bytes_done[ddir];
1769	}
1770
1771	return 0;
1772}
1773
1774/*
1775 * Called to complete min_events number of io for the async engines.
1776 */
1777int io_u_queued_complete(struct thread_data *td, int min_evts,
1778			 uint64_t *bytes)
1779{
1780	struct io_completion_data icd;
1781	struct timespec *tvp = NULL;
1782	int ret;
1783	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1784
1785	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1786
1787	if (!min_evts)
1788		tvp = &ts;
1789
1790	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1791	if (ret < 0) {
1792		td_verror(td, -ret, "td_io_getevents");
1793		return ret;
1794	} else if (!ret)
1795		return ret;
1796
1797	init_icd(td, &icd, ret);
1798	ios_completed(td, &icd);
1799	if (icd.error) {
1800		td_verror(td, icd.error, "io_u_queued_complete");
1801		return -1;
1802	}
1803
1804	if (bytes) {
1805		int ddir;
1806
1807		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1808			bytes[ddir] += icd.bytes_done[ddir];
1809	}
1810
1811	return 0;
1812}
1813
1814/*
1815 * Call when io_u is really queued, to update the submission latency.
1816 */
1817void io_u_queued(struct thread_data *td, struct io_u *io_u)
1818{
1819	if (!td->o.disable_slat) {
1820		unsigned long slat_time;
1821
1822		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1823		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1824	}
1825}
1826
1827void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1828		    unsigned int max_bs)
1829{
1830	if (td->o.buffer_pattern_bytes)
1831		fill_buffer_pattern(td, buf, max_bs);
1832	else if (!td->o.zero_buffers) {
1833		unsigned int perc = td->o.compress_percentage;
1834
1835		if (perc) {
1836			unsigned int seg = min_write;
1837
1838			seg = min(min_write, td->o.compress_chunk);
1839			if (!seg)
1840				seg = min_write;
1841
1842			fill_random_buf_percentage(&td->buf_state, buf,
1843						perc, seg, max_bs);
1844		} else
1845			fill_random_buf(&td->buf_state, buf, max_bs);
1846	} else
1847		memset(buf, 0, max_bs);
1848}
1849
1850/*
1851 * "randomly" fill the buffer contents
1852 */
1853void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1854		      unsigned int min_write, unsigned int max_bs)
1855{
1856	io_u->buf_filled_len = 0;
1857	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1858}
1859