io_u.c revision cc86c395fd9dd2002ec1edc0967b7c9453debdfb
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13#include "lib/axmap.h"
14
15struct io_completion_data {
16	int nr;				/* input */
17
18	int error;			/* output */
19	uint64_t bytes_done[DDIR_RWDIR_CNT];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const uint64_t block)
28{
29	return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37	unsigned int min_bs = td->o.rw_min_bs;
38	struct fio_file *f = io_u->file;
39	unsigned int nr_blocks;
40	uint64_t block;
41
42	block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
43	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45	if (!(io_u->flags & IO_U_F_BUSY_OK))
46		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47
48	if ((nr_blocks * min_bs) < io_u->buflen)
49		io_u->buflen = nr_blocks * min_bs;
50}
51
52static uint64_t last_block(struct thread_data *td, struct fio_file *f,
53			   enum fio_ddir ddir)
54{
55	uint64_t max_blocks;
56	uint64_t max_size;
57
58	assert(ddir_rw(ddir));
59
60	/*
61	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
62	 */
63	max_size = f->io_size;
64	if (max_size > f->real_file_size)
65		max_size = f->real_file_size;
66
67	if (td->o.zone_range)
68		max_size = td->o.zone_range;
69
70	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
71	if (!max_blocks)
72		return 0;
73
74	return max_blocks;
75}
76
77struct rand_off {
78	struct flist_head list;
79	uint64_t off;
80};
81
82static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
83				  enum fio_ddir ddir, uint64_t *b)
84{
85	uint64_t r, lastb;
86
87	lastb = last_block(td, f, ddir);
88	if (!lastb)
89		return 1;
90
91	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
92		uint64_t rmax;
93
94		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
95
96		if (td->o.use_os_rand) {
97			rmax = OS_RAND_MAX;
98			r = os_random_long(&td->random_state);
99		} else {
100			rmax = FRAND_MAX;
101			r = __rand(&td->__random_state);
102		}
103
104		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
105
106		*b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
107	} else {
108		uint64_t off = 0;
109
110		if (lfsr_next(&f->lfsr, &off, lastb))
111			return 1;
112
113		*b = off;
114	}
115
116	/*
117	 * if we are not maintaining a random map, we are done.
118	 */
119	if (!file_randommap(td, f))
120		goto ret;
121
122	/*
123	 * calculate map offset and check if it's free
124	 */
125	if (random_map_free(f, *b))
126		goto ret;
127
128	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
129						(unsigned long long) *b);
130
131	*b = axmap_next_free(f->io_axmap, *b);
132	if (*b == (uint64_t) -1ULL)
133		return 1;
134ret:
135	return 0;
136}
137
138static int __get_next_rand_offset_zipf(struct thread_data *td,
139				       struct fio_file *f, enum fio_ddir ddir,
140				       uint64_t *b)
141{
142	*b = zipf_next(&f->zipf);
143	return 0;
144}
145
146static int __get_next_rand_offset_pareto(struct thread_data *td,
147					 struct fio_file *f, enum fio_ddir ddir,
148					 uint64_t *b)
149{
150	*b = pareto_next(&f->zipf);
151	return 0;
152}
153
154static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
155{
156	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
157	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
158
159	return r1->off - r2->off;
160}
161
162static int get_off_from_method(struct thread_data *td, struct fio_file *f,
163			       enum fio_ddir ddir, uint64_t *b)
164{
165	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
166		return __get_next_rand_offset(td, f, ddir, b);
167	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
168		return __get_next_rand_offset_zipf(td, f, ddir, b);
169	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
170		return __get_next_rand_offset_pareto(td, f, ddir, b);
171
172	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
173	return 1;
174}
175
176/*
177 * Sort the reads for a verify phase in batches of verifysort_nr, if
178 * specified.
179 */
180static inline int should_sort_io(struct thread_data *td)
181{
182	if (!td->o.verifysort_nr || !td->o.do_verify)
183		return 0;
184	if (!td_random(td))
185		return 0;
186	if (td->runstate != TD_VERIFYING)
187		return 0;
188	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
189		return 0;
190
191	return 1;
192}
193
194static int should_do_random(struct thread_data *td)
195{
196	unsigned int v;
197	unsigned long r;
198
199	if (td->o.perc_rand == 100)
200		return 1;
201
202	if (td->o.use_os_rand) {
203		r = os_random_long(&td->seq_rand_state);
204		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
205	} else {
206		r = __rand(&td->__seq_rand_state);
207		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
208	}
209
210	return v <= td->o.perc_rand;
211}
212
213static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
214				enum fio_ddir ddir, uint64_t *b)
215{
216	struct rand_off *r;
217	int i, ret = 1;
218
219	if (!should_sort_io(td))
220		return get_off_from_method(td, f, ddir, b);
221
222	if (!flist_empty(&td->next_rand_list)) {
223		struct rand_off *r;
224fetch:
225		r = flist_entry(td->next_rand_list.next, struct rand_off, list);
226		flist_del(&r->list);
227		*b = r->off;
228		free(r);
229		return 0;
230	}
231
232	for (i = 0; i < td->o.verifysort_nr; i++) {
233		r = malloc(sizeof(*r));
234
235		ret = get_off_from_method(td, f, ddir, &r->off);
236		if (ret) {
237			free(r);
238			break;
239		}
240
241		flist_add(&r->list, &td->next_rand_list);
242	}
243
244	if (ret && !i)
245		return ret;
246
247	assert(!flist_empty(&td->next_rand_list));
248	flist_sort(NULL, &td->next_rand_list, flist_cmp);
249	goto fetch;
250}
251
252static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
253			       enum fio_ddir ddir, uint64_t *b)
254{
255	if (!get_next_rand_offset(td, f, ddir, b))
256		return 0;
257
258	if (td->o.time_based) {
259		fio_file_reset(td, f);
260		if (!get_next_rand_offset(td, f, ddir, b))
261			return 0;
262	}
263
264	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
265			f->file_name, (unsigned long long) f->last_pos,
266			(unsigned long long) f->real_file_size);
267	return 1;
268}
269
270static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
271			       enum fio_ddir ddir, uint64_t *offset)
272{
273	assert(ddir_rw(ddir));
274
275	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
276		f->last_pos = f->last_pos - f->io_size;
277
278	if (f->last_pos < f->real_file_size) {
279		uint64_t pos;
280
281		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
282			f->last_pos = f->real_file_size;
283
284		pos = f->last_pos - f->file_offset;
285		if (pos)
286			pos += td->o.ddir_seq_add;
287
288		*offset = pos;
289		return 0;
290	}
291
292	return 1;
293}
294
295static int get_next_block(struct thread_data *td, struct io_u *io_u,
296			  enum fio_ddir ddir, int rw_seq)
297{
298	struct fio_file *f = io_u->file;
299	uint64_t b, offset;
300	int ret;
301
302	assert(ddir_rw(ddir));
303
304	b = offset = -1ULL;
305
306	if (rw_seq) {
307		if (td_random(td)) {
308			if (should_do_random(td))
309				ret = get_next_rand_block(td, f, ddir, &b);
310			else {
311				io_u->flags |= IO_U_F_BUSY_OK;
312				ret = get_next_seq_offset(td, f, ddir, &offset);
313				if (ret)
314					ret = get_next_rand_block(td, f, ddir, &b);
315			}
316		} else
317			ret = get_next_seq_offset(td, f, ddir, &offset);
318	} else {
319		io_u->flags |= IO_U_F_BUSY_OK;
320
321		if (td->o.rw_seq == RW_SEQ_SEQ) {
322			ret = get_next_seq_offset(td, f, ddir, &offset);
323			if (ret)
324				ret = get_next_rand_block(td, f, ddir, &b);
325		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
326			if (f->last_start != -1ULL)
327				offset = f->last_start - f->file_offset;
328			else
329				offset = 0;
330			ret = 0;
331		} else {
332			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
333			ret = 1;
334		}
335	}
336
337	if (!ret) {
338		if (offset != -1ULL)
339			io_u->offset = offset;
340		else if (b != -1ULL)
341			io_u->offset = b * td->o.ba[ddir];
342		else {
343			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
344			ret = 1;
345		}
346	}
347
348	return ret;
349}
350
351/*
352 * For random io, generate a random new block and see if it's used. Repeat
353 * until we find a free one. For sequential io, just return the end of
354 * the last io issued.
355 */
356static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
357{
358	struct fio_file *f = io_u->file;
359	enum fio_ddir ddir = io_u->ddir;
360	int rw_seq_hit = 0;
361
362	assert(ddir_rw(ddir));
363
364	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
365		rw_seq_hit = 1;
366		td->ddir_seq_nr = td->o.ddir_seq_nr;
367	}
368
369	if (get_next_block(td, io_u, ddir, rw_seq_hit))
370		return 1;
371
372	if (io_u->offset >= f->io_size) {
373		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
374					(unsigned long long) io_u->offset,
375					(unsigned long long) f->io_size);
376		return 1;
377	}
378
379	io_u->offset += f->file_offset;
380	if (io_u->offset >= f->real_file_size) {
381		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
382					(unsigned long long) io_u->offset,
383					(unsigned long long) f->real_file_size);
384		return 1;
385	}
386
387	return 0;
388}
389
390static int get_next_offset(struct thread_data *td, struct io_u *io_u)
391{
392	if (td->flags & TD_F_PROFILE_OPS) {
393		struct prof_io_ops *ops = &td->prof_io_ops;
394
395		if (ops->fill_io_u_off)
396			return ops->fill_io_u_off(td, io_u);
397	}
398
399	return __get_next_offset(td, io_u);
400}
401
402static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
403			    unsigned int buflen)
404{
405	struct fio_file *f = io_u->file;
406
407	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
408}
409
410static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
411{
412	const int ddir = io_u->ddir;
413	unsigned int buflen = 0;
414	unsigned int minbs, maxbs;
415	unsigned long r, rand_max;
416
417	assert(ddir_rw(ddir));
418
419	minbs = td->o.min_bs[ddir];
420	maxbs = td->o.max_bs[ddir];
421
422	if (minbs == maxbs)
423		return minbs;
424
425	/*
426	 * If we can't satisfy the min block size from here, then fail
427	 */
428	if (!io_u_fits(td, io_u, minbs))
429		return 0;
430
431	if (td->o.use_os_rand)
432		rand_max = OS_RAND_MAX;
433	else
434		rand_max = FRAND_MAX;
435
436	do {
437		if (td->o.use_os_rand)
438			r = os_random_long(&td->bsrange_state);
439		else
440			r = __rand(&td->__bsrange_state);
441
442		if (!td->o.bssplit_nr[ddir]) {
443			buflen = 1 + (unsigned int) ((double) maxbs *
444					(r / (rand_max + 1.0)));
445			if (buflen < minbs)
446				buflen = minbs;
447		} else {
448			long perc = 0;
449			unsigned int i;
450
451			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
452				struct bssplit *bsp = &td->o.bssplit[ddir][i];
453
454				buflen = bsp->bs;
455				perc += bsp->perc;
456				if ((r <= ((rand_max / 100L) * perc)) &&
457				    io_u_fits(td, io_u, buflen))
458					break;
459			}
460		}
461
462		if (!td->o.bs_unaligned && is_power_of_2(minbs))
463			buflen = (buflen + minbs - 1) & ~(minbs - 1);
464
465	} while (!io_u_fits(td, io_u, buflen));
466
467	return buflen;
468}
469
470static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
471{
472	if (td->flags & TD_F_PROFILE_OPS) {
473		struct prof_io_ops *ops = &td->prof_io_ops;
474
475		if (ops->fill_io_u_size)
476			return ops->fill_io_u_size(td, io_u);
477	}
478
479	return __get_next_buflen(td, io_u);
480}
481
482static void set_rwmix_bytes(struct thread_data *td)
483{
484	unsigned int diff;
485
486	/*
487	 * we do time or byte based switch. this is needed because
488	 * buffered writes may issue a lot quicker than they complete,
489	 * whereas reads do not.
490	 */
491	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
492	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
493}
494
495static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
496{
497	unsigned int v;
498	unsigned long r;
499
500	if (td->o.use_os_rand) {
501		r = os_random_long(&td->rwmix_state);
502		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
503	} else {
504		r = __rand(&td->__rwmix_state);
505		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
506	}
507
508	if (v <= td->o.rwmix[DDIR_READ])
509		return DDIR_READ;
510
511	return DDIR_WRITE;
512}
513
514static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
515{
516	enum fio_ddir odir = ddir ^ 1;
517	struct timeval t;
518	long usec;
519
520	assert(ddir_rw(ddir));
521
522	if (td->rate_pending_usleep[ddir] <= 0)
523		return ddir;
524
525	/*
526	 * We have too much pending sleep in this direction. See if we
527	 * should switch.
528	 */
529	if (td_rw(td) && td->o.rwmix[odir]) {
530		/*
531		 * Other direction does not have too much pending, switch
532		 */
533		if (td->rate_pending_usleep[odir] < 100000)
534			return odir;
535
536		/*
537		 * Both directions have pending sleep. Sleep the minimum time
538		 * and deduct from both.
539		 */
540		if (td->rate_pending_usleep[ddir] <=
541			td->rate_pending_usleep[odir]) {
542			usec = td->rate_pending_usleep[ddir];
543		} else {
544			usec = td->rate_pending_usleep[odir];
545			ddir = odir;
546		}
547	} else
548		usec = td->rate_pending_usleep[ddir];
549
550	/*
551	 * We are going to sleep, ensure that we flush anything pending as
552	 * not to skew our latency numbers.
553	 *
554	 * Changed to only monitor 'in flight' requests here instead of the
555	 * td->cur_depth, b/c td->cur_depth does not accurately represent
556	 * io's that have been actually submitted to an async engine,
557	 * and cur_depth is meaningless for sync engines.
558	 */
559	while (td->io_u_in_flight) {
560		int fio_unused ret;
561
562		ret = io_u_queued_complete(td, 1, NULL);
563	}
564
565	fio_gettime(&t, NULL);
566	usec_sleep(td, usec);
567	usec = utime_since_now(&t);
568
569	td->rate_pending_usleep[ddir] -= usec;
570
571	odir = ddir ^ 1;
572	if (td_rw(td) && __should_check_rate(td, odir))
573		td->rate_pending_usleep[odir] -= usec;
574
575	if (ddir_trim(ddir))
576		return ddir;
577
578	return ddir;
579}
580
581/*
582 * Return the data direction for the next io_u. If the job is a
583 * mixed read/write workload, check the rwmix cycle and switch if
584 * necessary.
585 */
586static enum fio_ddir get_rw_ddir(struct thread_data *td)
587{
588	enum fio_ddir ddir;
589
590	/*
591	 * see if it's time to fsync
592	 */
593	if (td->o.fsync_blocks &&
594	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
595	     td->io_issues[DDIR_WRITE] && should_fsync(td))
596		return DDIR_SYNC;
597
598	/*
599	 * see if it's time to fdatasync
600	 */
601	if (td->o.fdatasync_blocks &&
602	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
603	     td->io_issues[DDIR_WRITE] && should_fsync(td))
604		return DDIR_DATASYNC;
605
606	/*
607	 * see if it's time to sync_file_range
608	 */
609	if (td->sync_file_range_nr &&
610	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
611	     td->io_issues[DDIR_WRITE] && should_fsync(td))
612		return DDIR_SYNC_FILE_RANGE;
613
614	if (td_rw(td)) {
615		/*
616		 * Check if it's time to seed a new data direction.
617		 */
618		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
619			/*
620			 * Put a top limit on how many bytes we do for
621			 * one data direction, to avoid overflowing the
622			 * ranges too much
623			 */
624			ddir = get_rand_ddir(td);
625
626			if (ddir != td->rwmix_ddir)
627				set_rwmix_bytes(td);
628
629			td->rwmix_ddir = ddir;
630		}
631		ddir = td->rwmix_ddir;
632	} else if (td_read(td))
633		ddir = DDIR_READ;
634	else if (td_write(td))
635		ddir = DDIR_WRITE;
636	else
637		ddir = DDIR_TRIM;
638
639	td->rwmix_ddir = rate_ddir(td, ddir);
640	return td->rwmix_ddir;
641}
642
643static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
644{
645	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
646
647	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
648	    td->o.barrier_blocks &&
649	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
650	     td->io_issues[DDIR_WRITE])
651		io_u->flags |= IO_U_F_BARRIER;
652}
653
654void put_file_log(struct thread_data *td, struct fio_file *f)
655{
656	int ret = put_file(td, f);
657
658	if (ret)
659		td_verror(td, ret, "file close");
660}
661
662void put_io_u(struct thread_data *td, struct io_u *io_u)
663{
664	td_io_u_lock(td);
665
666	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
667		put_file_log(td, io_u->file);
668	io_u->file = NULL;
669	io_u->flags &= ~IO_U_F_FREE_DEF;
670	io_u->flags |= IO_U_F_FREE;
671
672	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
673		td->cur_depth--;
674	flist_del_init(&io_u->list);
675	flist_add(&io_u->list, &td->io_u_freelist);
676	td_io_u_unlock(td);
677	td_io_u_free_notify(td);
678}
679
680void clear_io_u(struct thread_data *td, struct io_u *io_u)
681{
682	io_u->flags &= ~IO_U_F_FLIGHT;
683	put_io_u(td, io_u);
684}
685
686void requeue_io_u(struct thread_data *td, struct io_u **io_u)
687{
688	struct io_u *__io_u = *io_u;
689	enum fio_ddir ddir = acct_ddir(__io_u);
690
691	dprint(FD_IO, "requeue %p\n", __io_u);
692
693	td_io_u_lock(td);
694
695	__io_u->flags |= IO_U_F_FREE;
696	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
697		td->io_issues[ddir]--;
698
699	__io_u->flags &= ~IO_U_F_FLIGHT;
700	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
701		td->cur_depth--;
702	flist_del(&__io_u->list);
703	flist_add_tail(&__io_u->list, &td->io_u_requeues);
704	td_io_u_unlock(td);
705	*io_u = NULL;
706}
707
708static int fill_io_u(struct thread_data *td, struct io_u *io_u)
709{
710	if (td->io_ops->flags & FIO_NOIO)
711		goto out;
712
713	set_rw_ddir(td, io_u);
714
715	/*
716	 * fsync() or fdatasync() or trim etc, we are done
717	 */
718	if (!ddir_rw(io_u->ddir))
719		goto out;
720
721	/*
722	 * See if it's time to switch to a new zone
723	 */
724	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
725		td->zone_bytes = 0;
726		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
727		io_u->file->last_pos = io_u->file->file_offset;
728		td->io_skip_bytes += td->o.zone_skip;
729	}
730
731	/*
732	 * No log, let the seq/rand engine retrieve the next buflen and
733	 * position.
734	 */
735	if (get_next_offset(td, io_u)) {
736		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
737		return 1;
738	}
739
740	io_u->buflen = get_next_buflen(td, io_u);
741	if (!io_u->buflen) {
742		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
743		return 1;
744	}
745
746	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
747		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
748		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
749			(unsigned long long) io_u->offset, io_u->buflen,
750			(unsigned long long) io_u->file->real_file_size);
751		return 1;
752	}
753
754	/*
755	 * mark entry before potentially trimming io_u
756	 */
757	if (td_random(td) && file_randommap(td, io_u->file))
758		mark_random_map(td, io_u);
759
760out:
761	dprint_io_u(io_u, "fill_io_u");
762	td->zone_bytes += io_u->buflen;
763	return 0;
764}
765
766static void __io_u_mark_map(unsigned int *map, unsigned int nr)
767{
768	int idx = 0;
769
770	switch (nr) {
771	default:
772		idx = 6;
773		break;
774	case 33 ... 64:
775		idx = 5;
776		break;
777	case 17 ... 32:
778		idx = 4;
779		break;
780	case 9 ... 16:
781		idx = 3;
782		break;
783	case 5 ... 8:
784		idx = 2;
785		break;
786	case 1 ... 4:
787		idx = 1;
788	case 0:
789		break;
790	}
791
792	map[idx]++;
793}
794
795void io_u_mark_submit(struct thread_data *td, unsigned int nr)
796{
797	__io_u_mark_map(td->ts.io_u_submit, nr);
798	td->ts.total_submit++;
799}
800
801void io_u_mark_complete(struct thread_data *td, unsigned int nr)
802{
803	__io_u_mark_map(td->ts.io_u_complete, nr);
804	td->ts.total_complete++;
805}
806
807void io_u_mark_depth(struct thread_data *td, unsigned int nr)
808{
809	int idx = 0;
810
811	switch (td->cur_depth) {
812	default:
813		idx = 6;
814		break;
815	case 32 ... 63:
816		idx = 5;
817		break;
818	case 16 ... 31:
819		idx = 4;
820		break;
821	case 8 ... 15:
822		idx = 3;
823		break;
824	case 4 ... 7:
825		idx = 2;
826		break;
827	case 2 ... 3:
828		idx = 1;
829	case 1:
830		break;
831	}
832
833	td->ts.io_u_map[idx] += nr;
834}
835
836static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
837{
838	int idx = 0;
839
840	assert(usec < 1000);
841
842	switch (usec) {
843	case 750 ... 999:
844		idx = 9;
845		break;
846	case 500 ... 749:
847		idx = 8;
848		break;
849	case 250 ... 499:
850		idx = 7;
851		break;
852	case 100 ... 249:
853		idx = 6;
854		break;
855	case 50 ... 99:
856		idx = 5;
857		break;
858	case 20 ... 49:
859		idx = 4;
860		break;
861	case 10 ... 19:
862		idx = 3;
863		break;
864	case 4 ... 9:
865		idx = 2;
866		break;
867	case 2 ... 3:
868		idx = 1;
869	case 0 ... 1:
870		break;
871	}
872
873	assert(idx < FIO_IO_U_LAT_U_NR);
874	td->ts.io_u_lat_u[idx]++;
875}
876
877static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
878{
879	int idx = 0;
880
881	switch (msec) {
882	default:
883		idx = 11;
884		break;
885	case 1000 ... 1999:
886		idx = 10;
887		break;
888	case 750 ... 999:
889		idx = 9;
890		break;
891	case 500 ... 749:
892		idx = 8;
893		break;
894	case 250 ... 499:
895		idx = 7;
896		break;
897	case 100 ... 249:
898		idx = 6;
899		break;
900	case 50 ... 99:
901		idx = 5;
902		break;
903	case 20 ... 49:
904		idx = 4;
905		break;
906	case 10 ... 19:
907		idx = 3;
908		break;
909	case 4 ... 9:
910		idx = 2;
911		break;
912	case 2 ... 3:
913		idx = 1;
914	case 0 ... 1:
915		break;
916	}
917
918	assert(idx < FIO_IO_U_LAT_M_NR);
919	td->ts.io_u_lat_m[idx]++;
920}
921
922static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
923{
924	if (usec < 1000)
925		io_u_mark_lat_usec(td, usec);
926	else
927		io_u_mark_lat_msec(td, usec / 1000);
928}
929
930/*
931 * Get next file to service by choosing one at random
932 */
933static struct fio_file *get_next_file_rand(struct thread_data *td,
934					   enum fio_file_flags goodf,
935					   enum fio_file_flags badf)
936{
937	struct fio_file *f;
938	int fno;
939
940	do {
941		int opened = 0;
942		unsigned long r;
943
944		if (td->o.use_os_rand) {
945			r = os_random_long(&td->next_file_state);
946			fno = (unsigned int) ((double) td->o.nr_files
947				* (r / (OS_RAND_MAX + 1.0)));
948		} else {
949			r = __rand(&td->__next_file_state);
950			fno = (unsigned int) ((double) td->o.nr_files
951				* (r / (FRAND_MAX + 1.0)));
952		}
953
954		f = td->files[fno];
955		if (fio_file_done(f))
956			continue;
957
958		if (!fio_file_open(f)) {
959			int err;
960
961			err = td_io_open_file(td, f);
962			if (err)
963				continue;
964			opened = 1;
965		}
966
967		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
968			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
969			return f;
970		}
971		if (opened)
972			td_io_close_file(td, f);
973	} while (1);
974}
975
976/*
977 * Get next file to service by doing round robin between all available ones
978 */
979static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
980					 int badf)
981{
982	unsigned int old_next_file = td->next_file;
983	struct fio_file *f;
984
985	do {
986		int opened = 0;
987
988		f = td->files[td->next_file];
989
990		td->next_file++;
991		if (td->next_file >= td->o.nr_files)
992			td->next_file = 0;
993
994		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
995		if (fio_file_done(f)) {
996			f = NULL;
997			continue;
998		}
999
1000		if (!fio_file_open(f)) {
1001			int err;
1002
1003			err = td_io_open_file(td, f);
1004			if (err) {
1005				dprint(FD_FILE, "error %d on open of %s\n",
1006					err, f->file_name);
1007				f = NULL;
1008				continue;
1009			}
1010			opened = 1;
1011		}
1012
1013		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1014								f->flags);
1015		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1016			break;
1017
1018		if (opened)
1019			td_io_close_file(td, f);
1020
1021		f = NULL;
1022	} while (td->next_file != old_next_file);
1023
1024	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1025	return f;
1026}
1027
1028static struct fio_file *__get_next_file(struct thread_data *td)
1029{
1030	struct fio_file *f;
1031
1032	assert(td->o.nr_files <= td->files_index);
1033
1034	if (td->nr_done_files >= td->o.nr_files) {
1035		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1036				" nr_files=%d\n", td->nr_open_files,
1037						  td->nr_done_files,
1038						  td->o.nr_files);
1039		return NULL;
1040	}
1041
1042	f = td->file_service_file;
1043	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1044		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1045			goto out;
1046		if (td->file_service_left--)
1047			goto out;
1048	}
1049
1050	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1051	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1052		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1053	else
1054		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1055
1056	td->file_service_file = f;
1057	td->file_service_left = td->file_service_nr - 1;
1058out:
1059	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1060	return f;
1061}
1062
1063static struct fio_file *get_next_file(struct thread_data *td)
1064{
1065	if (!(td->flags & TD_F_PROFILE_OPS)) {
1066		struct prof_io_ops *ops = &td->prof_io_ops;
1067
1068		if (ops->get_next_file)
1069			return ops->get_next_file(td);
1070	}
1071
1072	return __get_next_file(td);
1073}
1074
1075static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1076{
1077	struct fio_file *f;
1078
1079	do {
1080		f = get_next_file(td);
1081		if (!f)
1082			return 1;
1083
1084		io_u->file = f;
1085		get_file(f);
1086
1087		if (!fill_io_u(td, io_u))
1088			break;
1089
1090		put_file_log(td, f);
1091		td_io_close_file(td, f);
1092		io_u->file = NULL;
1093		fio_file_set_done(f);
1094		td->nr_done_files++;
1095		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1096					td->nr_done_files, td->o.nr_files);
1097	} while (1);
1098
1099	return 0;
1100}
1101
1102
1103struct io_u *__get_io_u(struct thread_data *td)
1104{
1105	struct io_u *io_u = NULL;
1106
1107	td_io_u_lock(td);
1108
1109again:
1110	if (!flist_empty(&td->io_u_requeues))
1111		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1112	else if (!queue_full(td)) {
1113		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1114
1115		io_u->buflen = 0;
1116		io_u->resid = 0;
1117		io_u->file = NULL;
1118		io_u->end_io = NULL;
1119	}
1120
1121	if (io_u) {
1122		assert(io_u->flags & IO_U_F_FREE);
1123		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1124		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1125		io_u->flags &= ~IO_U_F_VER_LIST;
1126
1127		io_u->error = 0;
1128		io_u->acct_ddir = -1;
1129		flist_del(&io_u->list);
1130		flist_add_tail(&io_u->list, &td->io_u_busylist);
1131		td->cur_depth++;
1132		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1133	} else if (td->o.verify_async) {
1134		/*
1135		 * We ran out, wait for async verify threads to finish and
1136		 * return one
1137		 */
1138		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1139		goto again;
1140	}
1141
1142	td_io_u_unlock(td);
1143	return io_u;
1144}
1145
1146static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1147{
1148	if (!(td->flags & TD_F_TRIM_BACKLOG))
1149		return 0;
1150
1151	if (td->trim_entries) {
1152		int get_trim = 0;
1153
1154		if (td->trim_batch) {
1155			td->trim_batch--;
1156			get_trim = 1;
1157		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1158			 td->last_ddir != DDIR_READ) {
1159			td->trim_batch = td->o.trim_batch;
1160			if (!td->trim_batch)
1161				td->trim_batch = td->o.trim_backlog;
1162			get_trim = 1;
1163		}
1164
1165		if (get_trim && !get_next_trim(td, io_u))
1166			return 1;
1167	}
1168
1169	return 0;
1170}
1171
1172static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1173{
1174	if (!(td->flags & TD_F_VER_BACKLOG))
1175		return 0;
1176
1177	if (td->io_hist_len) {
1178		int get_verify = 0;
1179
1180		if (td->verify_batch)
1181			get_verify = 1;
1182		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1183			 td->last_ddir != DDIR_READ) {
1184			td->verify_batch = td->o.verify_batch;
1185			if (!td->verify_batch)
1186				td->verify_batch = td->o.verify_backlog;
1187			get_verify = 1;
1188		}
1189
1190		if (get_verify && !get_next_verify(td, io_u)) {
1191			td->verify_batch--;
1192			return 1;
1193		}
1194	}
1195
1196	return 0;
1197}
1198
1199/*
1200 * Fill offset and start time into the buffer content, to prevent too
1201 * easy compressible data for simple de-dupe attempts. Do this for every
1202 * 512b block in the range, since that should be the smallest block size
1203 * we can expect from a device.
1204 */
1205static void small_content_scramble(struct io_u *io_u)
1206{
1207	unsigned int i, nr_blocks = io_u->buflen / 512;
1208	uint64_t boffset;
1209	unsigned int offset;
1210	void *p, *end;
1211
1212	if (!nr_blocks)
1213		return;
1214
1215	p = io_u->xfer_buf;
1216	boffset = io_u->offset;
1217	io_u->buf_filled_len = 0;
1218
1219	for (i = 0; i < nr_blocks; i++) {
1220		/*
1221		 * Fill the byte offset into a "random" start offset of
1222		 * the buffer, given by the product of the usec time
1223		 * and the actual offset.
1224		 */
1225		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1226		offset &= ~(sizeof(uint64_t) - 1);
1227		if (offset >= 512 - sizeof(uint64_t))
1228			offset -= sizeof(uint64_t);
1229		memcpy(p + offset, &boffset, sizeof(boffset));
1230
1231		end = p + 512 - sizeof(io_u->start_time);
1232		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1233		p += 512;
1234		boffset += 512;
1235	}
1236}
1237
1238/*
1239 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1240 * etc. The returned io_u is fully ready to be prepped and submitted.
1241 */
1242struct io_u *get_io_u(struct thread_data *td)
1243{
1244	struct fio_file *f;
1245	struct io_u *io_u;
1246	int do_scramble = 0;
1247
1248	io_u = __get_io_u(td);
1249	if (!io_u) {
1250		dprint(FD_IO, "__get_io_u failed\n");
1251		return NULL;
1252	}
1253
1254	if (check_get_verify(td, io_u))
1255		goto out;
1256	if (check_get_trim(td, io_u))
1257		goto out;
1258
1259	/*
1260	 * from a requeue, io_u already setup
1261	 */
1262	if (io_u->file)
1263		goto out;
1264
1265	/*
1266	 * If using an iolog, grab next piece if any available.
1267	 */
1268	if (td->flags & TD_F_READ_IOLOG) {
1269		if (read_iolog_get(td, io_u))
1270			goto err_put;
1271	} else if (set_io_u_file(td, io_u)) {
1272		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1273		goto err_put;
1274	}
1275
1276	f = io_u->file;
1277	assert(fio_file_open(f));
1278
1279	if (ddir_rw(io_u->ddir)) {
1280		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1281			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1282			goto err_put;
1283		}
1284
1285		f->last_start = io_u->offset;
1286		f->last_pos = io_u->offset + io_u->buflen;
1287
1288		if (io_u->ddir == DDIR_WRITE) {
1289			if (td->flags & TD_F_REFILL_BUFFERS) {
1290				io_u_fill_buffer(td, io_u,
1291					io_u->xfer_buflen, io_u->xfer_buflen);
1292			} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1293				do_scramble = 1;
1294			if (td->flags & TD_F_VER_NONE) {
1295				populate_verify_io_u(td, io_u);
1296				do_scramble = 0;
1297			}
1298		} else if (io_u->ddir == DDIR_READ) {
1299			/*
1300			 * Reset the buf_filled parameters so next time if the
1301			 * buffer is used for writes it is refilled.
1302			 */
1303			io_u->buf_filled_len = 0;
1304		}
1305	}
1306
1307	/*
1308	 * Set io data pointers.
1309	 */
1310	io_u->xfer_buf = io_u->buf;
1311	io_u->xfer_buflen = io_u->buflen;
1312
1313out:
1314	assert(io_u->file);
1315	if (!td_io_prep(td, io_u)) {
1316		if (!td->o.disable_slat)
1317			fio_gettime(&io_u->start_time, NULL);
1318		if (do_scramble)
1319			small_content_scramble(io_u);
1320		return io_u;
1321	}
1322err_put:
1323	dprint(FD_IO, "get_io_u failed\n");
1324	put_io_u(td, io_u);
1325	return NULL;
1326}
1327
1328void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1329{
1330	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1331	const char *msg[] = { "read", "write", "sync", "datasync",
1332				"sync_file_range", "wait", "trim" };
1333
1334	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1335		return;
1336
1337	log_err("fio: io_u error");
1338
1339	if (io_u->file)
1340		log_err(" on file %s", io_u->file->file_name);
1341
1342	log_err(": %s\n", strerror(io_u->error));
1343
1344	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1345					io_u->offset, io_u->xfer_buflen);
1346
1347	if (!td->error)
1348		td_verror(td, io_u->error, "io_u error");
1349}
1350
1351static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1352				  struct io_completion_data *icd,
1353				  const enum fio_ddir idx, unsigned int bytes)
1354{
1355	unsigned long lusec = 0;
1356
1357	if (!td->o.disable_clat || !td->o.disable_bw)
1358		lusec = utime_since(&io_u->issue_time, &icd->time);
1359
1360	if (!td->o.disable_lat) {
1361		unsigned long tusec;
1362
1363		tusec = utime_since(&io_u->start_time, &icd->time);
1364		add_lat_sample(td, idx, tusec, bytes);
1365
1366		if (td->o.max_latency && tusec > td->o.max_latency) {
1367			if (!td->error)
1368				log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1369			td_verror(td, ETIMEDOUT, "max latency exceeded");
1370			icd->error = ETIMEDOUT;
1371		}
1372	}
1373
1374	if (!td->o.disable_clat) {
1375		add_clat_sample(td, idx, lusec, bytes);
1376		io_u_mark_latency(td, lusec);
1377	}
1378
1379	if (!td->o.disable_bw)
1380		add_bw_sample(td, idx, bytes, &icd->time);
1381
1382	add_iops_sample(td, idx, &icd->time);
1383}
1384
1385static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1386{
1387	uint64_t secs, remainder, bps, bytes;
1388
1389	bytes = td->this_io_bytes[ddir];
1390	bps = td->rate_bps[ddir];
1391	secs = bytes / bps;
1392	remainder = bytes % bps;
1393	return remainder * 1000000 / bps + secs * 1000000;
1394}
1395
1396static void io_completed(struct thread_data *td, struct io_u *io_u,
1397			 struct io_completion_data *icd)
1398{
1399	struct fio_file *f;
1400
1401	dprint_io_u(io_u, "io complete");
1402
1403	td_io_u_lock(td);
1404	assert(io_u->flags & IO_U_F_FLIGHT);
1405	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1406	td_io_u_unlock(td);
1407
1408	if (ddir_sync(io_u->ddir)) {
1409		td->last_was_sync = 1;
1410		f = io_u->file;
1411		if (f) {
1412			f->first_write = -1ULL;
1413			f->last_write = -1ULL;
1414		}
1415		return;
1416	}
1417
1418	td->last_was_sync = 0;
1419	td->last_ddir = io_u->ddir;
1420
1421	if (!io_u->error && ddir_rw(io_u->ddir)) {
1422		unsigned int bytes = io_u->buflen - io_u->resid;
1423		const enum fio_ddir idx = io_u->ddir;
1424		const enum fio_ddir odx = io_u->ddir ^ 1;
1425		int ret;
1426
1427		td->io_blocks[idx]++;
1428		td->this_io_blocks[idx]++;
1429		td->io_bytes[idx] += bytes;
1430
1431		if (!(io_u->flags & IO_U_F_VER_LIST))
1432			td->this_io_bytes[idx] += bytes;
1433
1434		if (idx == DDIR_WRITE) {
1435			f = io_u->file;
1436			if (f) {
1437				if (f->first_write == -1ULL ||
1438				    io_u->offset < f->first_write)
1439					f->first_write = io_u->offset;
1440				if (f->last_write == -1ULL ||
1441				    ((io_u->offset + bytes) > f->last_write))
1442					f->last_write = io_u->offset + bytes;
1443			}
1444		}
1445
1446		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1447					   td->runstate == TD_VERIFYING)) {
1448			account_io_completion(td, io_u, icd, idx, bytes);
1449
1450			if (__should_check_rate(td, idx)) {
1451				td->rate_pending_usleep[idx] =
1452					(usec_for_io(td, idx) -
1453					 utime_since_now(&td->start));
1454			}
1455			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1456				td->rate_pending_usleep[odx] =
1457					(usec_for_io(td, odx) -
1458					 utime_since_now(&td->start));
1459		}
1460
1461		if (td_write(td) && idx == DDIR_WRITE &&
1462		    td->o.do_verify &&
1463		    td->o.verify != VERIFY_NONE &&
1464		    !td->o.experimental_verify)
1465			log_io_piece(td, io_u);
1466
1467		icd->bytes_done[idx] += bytes;
1468
1469		if (io_u->end_io) {
1470			ret = io_u->end_io(td, io_u);
1471			if (ret && !icd->error)
1472				icd->error = ret;
1473		}
1474	} else if (io_u->error) {
1475		icd->error = io_u->error;
1476		io_u_log_error(td, io_u);
1477	}
1478	if (icd->error) {
1479		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1480		if (!td_non_fatal_error(td, eb, icd->error))
1481			return;
1482		/*
1483		 * If there is a non_fatal error, then add to the error count
1484		 * and clear all the errors.
1485		 */
1486		update_error_count(td, icd->error);
1487		td_clear_error(td);
1488		icd->error = 0;
1489		io_u->error = 0;
1490	}
1491}
1492
1493static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1494		     int nr)
1495{
1496	int ddir;
1497	if (!td->o.disable_clat || !td->o.disable_bw)
1498		fio_gettime(&icd->time, NULL);
1499
1500	icd->nr = nr;
1501
1502	icd->error = 0;
1503	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1504		icd->bytes_done[ddir] = 0;
1505}
1506
1507static void ios_completed(struct thread_data *td,
1508			  struct io_completion_data *icd)
1509{
1510	struct io_u *io_u;
1511	int i;
1512
1513	for (i = 0; i < icd->nr; i++) {
1514		io_u = td->io_ops->event(td, i);
1515
1516		io_completed(td, io_u, icd);
1517
1518		if (!(io_u->flags & IO_U_F_FREE_DEF))
1519			put_io_u(td, io_u);
1520	}
1521}
1522
1523/*
1524 * Complete a single io_u for the sync engines.
1525 */
1526int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1527		       uint64_t *bytes)
1528{
1529	struct io_completion_data icd;
1530
1531	init_icd(td, &icd, 1);
1532	io_completed(td, io_u, &icd);
1533
1534	if (!(io_u->flags & IO_U_F_FREE_DEF))
1535		put_io_u(td, io_u);
1536
1537	if (icd.error) {
1538		td_verror(td, icd.error, "io_u_sync_complete");
1539		return -1;
1540	}
1541
1542	if (bytes) {
1543		int ddir;
1544
1545		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1546			bytes[ddir] += icd.bytes_done[ddir];
1547	}
1548
1549	return 0;
1550}
1551
1552/*
1553 * Called to complete min_events number of io for the async engines.
1554 */
1555int io_u_queued_complete(struct thread_data *td, int min_evts,
1556			 uint64_t *bytes)
1557{
1558	struct io_completion_data icd;
1559	struct timespec *tvp = NULL;
1560	int ret;
1561	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1562
1563	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1564
1565	if (!min_evts)
1566		tvp = &ts;
1567
1568	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1569	if (ret < 0) {
1570		td_verror(td, -ret, "td_io_getevents");
1571		return ret;
1572	} else if (!ret)
1573		return ret;
1574
1575	init_icd(td, &icd, ret);
1576	ios_completed(td, &icd);
1577	if (icd.error) {
1578		td_verror(td, icd.error, "io_u_queued_complete");
1579		return -1;
1580	}
1581
1582	if (bytes) {
1583		int ddir;
1584
1585		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1586			bytes[ddir] += icd.bytes_done[ddir];
1587	}
1588
1589	return 0;
1590}
1591
1592/*
1593 * Call when io_u is really queued, to update the submission latency.
1594 */
1595void io_u_queued(struct thread_data *td, struct io_u *io_u)
1596{
1597	if (!td->o.disable_slat) {
1598		unsigned long slat_time;
1599
1600		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1601		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1602	}
1603}
1604
1605void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1606		    unsigned int max_bs)
1607{
1608	if (!td->o.zero_buffers) {
1609		unsigned int perc = td->o.compress_percentage;
1610
1611		if (perc) {
1612			unsigned int seg = min_write;
1613
1614			seg = min(min_write, td->o.compress_chunk);
1615			if (!seg)
1616				seg = min_write;
1617
1618			fill_random_buf_percentage(&td->buf_state, buf,
1619						perc, seg, max_bs);
1620		} else
1621			fill_random_buf(&td->buf_state, buf, max_bs);
1622	} else
1623		memset(buf, 0, max_bs);
1624}
1625
1626/*
1627 * "randomly" fill the buffer contents
1628 */
1629void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1630		      unsigned int min_write, unsigned int max_bs)
1631{
1632	io_u->buf_filled_len = 0;
1633	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1634}
1635