io_u.c revision 2ae0b204743d6b4048c6fffd46c6280a70f2ecd1
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
514void io_u_quiesce(struct thread_data *td)
515{
516	/*
517	 * We are going to sleep, ensure that we flush anything pending as
518	 * not to skew our latency numbers.
519	 *
520	 * Changed to only monitor 'in flight' requests here instead of the
521	 * td->cur_depth, b/c td->cur_depth does not accurately represent
522	 * io's that have been actually submitted to an async engine,
523	 * and cur_depth is meaningless for sync engines.
524	 */
525	while (td->io_u_in_flight) {
526		int fio_unused ret;
527
528		ret = io_u_queued_complete(td, 1, NULL);
529	}
530}
531
532static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
533{
534	enum fio_ddir odir = ddir ^ 1;
535	struct timeval t;
536	long usec;
537
538	assert(ddir_rw(ddir));
539
540	if (td->rate_pending_usleep[ddir] <= 0)
541		return ddir;
542
543	/*
544	 * We have too much pending sleep in this direction. See if we
545	 * should switch.
546	 */
547	if (td_rw(td) && td->o.rwmix[odir]) {
548		/*
549		 * Other direction does not have too much pending, switch
550		 */
551		if (td->rate_pending_usleep[odir] < 100000)
552			return odir;
553
554		/*
555		 * Both directions have pending sleep. Sleep the minimum time
556		 * and deduct from both.
557		 */
558		if (td->rate_pending_usleep[ddir] <=
559			td->rate_pending_usleep[odir]) {
560			usec = td->rate_pending_usleep[ddir];
561		} else {
562			usec = td->rate_pending_usleep[odir];
563			ddir = odir;
564		}
565	} else
566		usec = td->rate_pending_usleep[ddir];
567
568	io_u_quiesce(td);
569
570	fio_gettime(&t, NULL);
571	usec_sleep(td, usec);
572	usec = utime_since_now(&t);
573
574	td->rate_pending_usleep[ddir] -= usec;
575
576	odir = ddir ^ 1;
577	if (td_rw(td) && __should_check_rate(td, odir))
578		td->rate_pending_usleep[odir] -= usec;
579
580	if (ddir_trim(ddir))
581		return ddir;
582
583	return ddir;
584}
585
586/*
587 * Return the data direction for the next io_u. If the job is a
588 * mixed read/write workload, check the rwmix cycle and switch if
589 * necessary.
590 */
591static enum fio_ddir get_rw_ddir(struct thread_data *td)
592{
593	enum fio_ddir ddir;
594
595	/*
596	 * see if it's time to fsync
597	 */
598	if (td->o.fsync_blocks &&
599	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
600	     td->io_issues[DDIR_WRITE] && should_fsync(td))
601		return DDIR_SYNC;
602
603	/*
604	 * see if it's time to fdatasync
605	 */
606	if (td->o.fdatasync_blocks &&
607	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
608	     td->io_issues[DDIR_WRITE] && should_fsync(td))
609		return DDIR_DATASYNC;
610
611	/*
612	 * see if it's time to sync_file_range
613	 */
614	if (td->sync_file_range_nr &&
615	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
616	     td->io_issues[DDIR_WRITE] && should_fsync(td))
617		return DDIR_SYNC_FILE_RANGE;
618
619	if (td_rw(td)) {
620		/*
621		 * Check if it's time to seed a new data direction.
622		 */
623		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
624			/*
625			 * Put a top limit on how many bytes we do for
626			 * one data direction, to avoid overflowing the
627			 * ranges too much
628			 */
629			ddir = get_rand_ddir(td);
630
631			if (ddir != td->rwmix_ddir)
632				set_rwmix_bytes(td);
633
634			td->rwmix_ddir = ddir;
635		}
636		ddir = td->rwmix_ddir;
637	} else if (td_read(td))
638		ddir = DDIR_READ;
639	else if (td_write(td))
640		ddir = DDIR_WRITE;
641	else
642		ddir = DDIR_TRIM;
643
644	td->rwmix_ddir = rate_ddir(td, ddir);
645	return td->rwmix_ddir;
646}
647
648static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
649{
650	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
651
652	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
653	    td->o.barrier_blocks &&
654	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
655	     td->io_issues[DDIR_WRITE])
656		io_u->flags |= IO_U_F_BARRIER;
657}
658
659void put_file_log(struct thread_data *td, struct fio_file *f)
660{
661	int ret = put_file(td, f);
662
663	if (ret)
664		td_verror(td, ret, "file close");
665}
666
667void put_io_u(struct thread_data *td, struct io_u *io_u)
668{
669	td_io_u_lock(td);
670
671	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
672		put_file_log(td, io_u->file);
673	io_u->file = NULL;
674	io_u->flags &= ~IO_U_F_FREE_DEF;
675	io_u->flags |= IO_U_F_FREE;
676
677	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
678		td->cur_depth--;
679	io_u_qpush(&td->io_u_freelist, io_u);
680	td_io_u_unlock(td);
681	td_io_u_free_notify(td);
682}
683
684void clear_io_u(struct thread_data *td, struct io_u *io_u)
685{
686	io_u->flags &= ~IO_U_F_FLIGHT;
687	put_io_u(td, io_u);
688}
689
690void requeue_io_u(struct thread_data *td, struct io_u **io_u)
691{
692	struct io_u *__io_u = *io_u;
693	enum fio_ddir ddir = acct_ddir(__io_u);
694
695	dprint(FD_IO, "requeue %p\n", __io_u);
696
697	td_io_u_lock(td);
698
699	__io_u->flags |= IO_U_F_FREE;
700	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
701		td->io_issues[ddir]--;
702
703	__io_u->flags &= ~IO_U_F_FLIGHT;
704	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
705		td->cur_depth--;
706
707	io_u_rpush(&td->io_u_requeues, __io_u);
708	td_io_u_unlock(td);
709	*io_u = NULL;
710}
711
712static int fill_io_u(struct thread_data *td, struct io_u *io_u)
713{
714	if (td->io_ops->flags & FIO_NOIO)
715		goto out;
716
717	set_rw_ddir(td, io_u);
718
719	/*
720	 * fsync() or fdatasync() or trim etc, we are done
721	 */
722	if (!ddir_rw(io_u->ddir))
723		goto out;
724
725	/*
726	 * See if it's time to switch to a new zone
727	 */
728	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
729		td->zone_bytes = 0;
730		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
731		io_u->file->last_pos = io_u->file->file_offset;
732		td->io_skip_bytes += td->o.zone_skip;
733	}
734
735	/*
736	 * No log, let the seq/rand engine retrieve the next buflen and
737	 * position.
738	 */
739	if (get_next_offset(td, io_u)) {
740		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
741		return 1;
742	}
743
744	io_u->buflen = get_next_buflen(td, io_u);
745	if (!io_u->buflen) {
746		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
747		return 1;
748	}
749
750	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
751		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
752		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
753			(unsigned long long) io_u->offset, io_u->buflen,
754			(unsigned long long) io_u->file->real_file_size);
755		return 1;
756	}
757
758	/*
759	 * mark entry before potentially trimming io_u
760	 */
761	if (td_random(td) && file_randommap(td, io_u->file))
762		mark_random_map(td, io_u);
763
764out:
765	dprint_io_u(io_u, "fill_io_u");
766	td->zone_bytes += io_u->buflen;
767	return 0;
768}
769
770static void __io_u_mark_map(unsigned int *map, unsigned int nr)
771{
772	int idx = 0;
773
774	switch (nr) {
775	default:
776		idx = 6;
777		break;
778	case 33 ... 64:
779		idx = 5;
780		break;
781	case 17 ... 32:
782		idx = 4;
783		break;
784	case 9 ... 16:
785		idx = 3;
786		break;
787	case 5 ... 8:
788		idx = 2;
789		break;
790	case 1 ... 4:
791		idx = 1;
792	case 0:
793		break;
794	}
795
796	map[idx]++;
797}
798
799void io_u_mark_submit(struct thread_data *td, unsigned int nr)
800{
801	__io_u_mark_map(td->ts.io_u_submit, nr);
802	td->ts.total_submit++;
803}
804
805void io_u_mark_complete(struct thread_data *td, unsigned int nr)
806{
807	__io_u_mark_map(td->ts.io_u_complete, nr);
808	td->ts.total_complete++;
809}
810
811void io_u_mark_depth(struct thread_data *td, unsigned int nr)
812{
813	int idx = 0;
814
815	switch (td->cur_depth) {
816	default:
817		idx = 6;
818		break;
819	case 32 ... 63:
820		idx = 5;
821		break;
822	case 16 ... 31:
823		idx = 4;
824		break;
825	case 8 ... 15:
826		idx = 3;
827		break;
828	case 4 ... 7:
829		idx = 2;
830		break;
831	case 2 ... 3:
832		idx = 1;
833	case 1:
834		break;
835	}
836
837	td->ts.io_u_map[idx] += nr;
838}
839
840static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
841{
842	int idx = 0;
843
844	assert(usec < 1000);
845
846	switch (usec) {
847	case 750 ... 999:
848		idx = 9;
849		break;
850	case 500 ... 749:
851		idx = 8;
852		break;
853	case 250 ... 499:
854		idx = 7;
855		break;
856	case 100 ... 249:
857		idx = 6;
858		break;
859	case 50 ... 99:
860		idx = 5;
861		break;
862	case 20 ... 49:
863		idx = 4;
864		break;
865	case 10 ... 19:
866		idx = 3;
867		break;
868	case 4 ... 9:
869		idx = 2;
870		break;
871	case 2 ... 3:
872		idx = 1;
873	case 0 ... 1:
874		break;
875	}
876
877	assert(idx < FIO_IO_U_LAT_U_NR);
878	td->ts.io_u_lat_u[idx]++;
879}
880
881static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
882{
883	int idx = 0;
884
885	switch (msec) {
886	default:
887		idx = 11;
888		break;
889	case 1000 ... 1999:
890		idx = 10;
891		break;
892	case 750 ... 999:
893		idx = 9;
894		break;
895	case 500 ... 749:
896		idx = 8;
897		break;
898	case 250 ... 499:
899		idx = 7;
900		break;
901	case 100 ... 249:
902		idx = 6;
903		break;
904	case 50 ... 99:
905		idx = 5;
906		break;
907	case 20 ... 49:
908		idx = 4;
909		break;
910	case 10 ... 19:
911		idx = 3;
912		break;
913	case 4 ... 9:
914		idx = 2;
915		break;
916	case 2 ... 3:
917		idx = 1;
918	case 0 ... 1:
919		break;
920	}
921
922	assert(idx < FIO_IO_U_LAT_M_NR);
923	td->ts.io_u_lat_m[idx]++;
924}
925
926static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
927{
928	if (usec < 1000)
929		io_u_mark_lat_usec(td, usec);
930	else
931		io_u_mark_lat_msec(td, usec / 1000);
932}
933
934/*
935 * Get next file to service by choosing one at random
936 */
937static struct fio_file *get_next_file_rand(struct thread_data *td,
938					   enum fio_file_flags goodf,
939					   enum fio_file_flags badf)
940{
941	struct fio_file *f;
942	int fno;
943
944	do {
945		int opened = 0;
946		unsigned long r;
947
948		if (td->o.use_os_rand) {
949			r = os_random_long(&td->next_file_state);
950			fno = (unsigned int) ((double) td->o.nr_files
951				* (r / (OS_RAND_MAX + 1.0)));
952		} else {
953			r = __rand(&td->__next_file_state);
954			fno = (unsigned int) ((double) td->o.nr_files
955				* (r / (FRAND_MAX + 1.0)));
956		}
957
958		f = td->files[fno];
959		if (fio_file_done(f))
960			continue;
961
962		if (!fio_file_open(f)) {
963			int err;
964
965			err = td_io_open_file(td, f);
966			if (err)
967				continue;
968			opened = 1;
969		}
970
971		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
972			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
973			return f;
974		}
975		if (opened)
976			td_io_close_file(td, f);
977	} while (1);
978}
979
980/*
981 * Get next file to service by doing round robin between all available ones
982 */
983static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
984					 int badf)
985{
986	unsigned int old_next_file = td->next_file;
987	struct fio_file *f;
988
989	do {
990		int opened = 0;
991
992		f = td->files[td->next_file];
993
994		td->next_file++;
995		if (td->next_file >= td->o.nr_files)
996			td->next_file = 0;
997
998		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
999		if (fio_file_done(f)) {
1000			f = NULL;
1001			continue;
1002		}
1003
1004		if (!fio_file_open(f)) {
1005			int err;
1006
1007			err = td_io_open_file(td, f);
1008			if (err) {
1009				dprint(FD_FILE, "error %d on open of %s\n",
1010					err, f->file_name);
1011				f = NULL;
1012				continue;
1013			}
1014			opened = 1;
1015		}
1016
1017		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1018								f->flags);
1019		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1020			break;
1021
1022		if (opened)
1023			td_io_close_file(td, f);
1024
1025		f = NULL;
1026	} while (td->next_file != old_next_file);
1027
1028	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1029	return f;
1030}
1031
1032static struct fio_file *__get_next_file(struct thread_data *td)
1033{
1034	struct fio_file *f;
1035
1036	assert(td->o.nr_files <= td->files_index);
1037
1038	if (td->nr_done_files >= td->o.nr_files) {
1039		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1040				" nr_files=%d\n", td->nr_open_files,
1041						  td->nr_done_files,
1042						  td->o.nr_files);
1043		return NULL;
1044	}
1045
1046	f = td->file_service_file;
1047	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1048		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1049			goto out;
1050		if (td->file_service_left--)
1051			goto out;
1052	}
1053
1054	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1055	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1056		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1057	else
1058		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1059
1060	td->file_service_file = f;
1061	td->file_service_left = td->file_service_nr - 1;
1062out:
1063	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1064	return f;
1065}
1066
1067static struct fio_file *get_next_file(struct thread_data *td)
1068{
1069	if (!(td->flags & TD_F_PROFILE_OPS)) {
1070		struct prof_io_ops *ops = &td->prof_io_ops;
1071
1072		if (ops->get_next_file)
1073			return ops->get_next_file(td);
1074	}
1075
1076	return __get_next_file(td);
1077}
1078
1079static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1080{
1081	struct fio_file *f;
1082
1083	do {
1084		f = get_next_file(td);
1085		if (!f)
1086			return 1;
1087
1088		io_u->file = f;
1089		get_file(f);
1090
1091		if (!fill_io_u(td, io_u))
1092			break;
1093
1094		put_file_log(td, f);
1095		td_io_close_file(td, f);
1096		io_u->file = NULL;
1097		fio_file_set_done(f);
1098		td->nr_done_files++;
1099		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1100					td->nr_done_files, td->o.nr_files);
1101	} while (1);
1102
1103	return 0;
1104}
1105
1106
1107struct io_u *__get_io_u(struct thread_data *td)
1108{
1109	struct io_u *io_u;
1110
1111	td_io_u_lock(td);
1112
1113again:
1114	if (!io_u_rempty(&td->io_u_requeues))
1115		io_u = io_u_rpop(&td->io_u_requeues);
1116	else if (!io_u_qempty(&td->io_u_freelist))
1117		io_u = io_u_qpop(&td->io_u_freelist);
1118
1119	if (io_u) {
1120		io_u->buflen = 0;
1121		io_u->resid = 0;
1122		io_u->file = NULL;
1123		io_u->end_io = NULL;
1124	}
1125
1126	if (io_u) {
1127		assert(io_u->flags & IO_U_F_FREE);
1128		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1129		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1130		io_u->flags &= ~IO_U_F_VER_LIST;
1131
1132		io_u->error = 0;
1133		io_u->acct_ddir = -1;
1134		td->cur_depth++;
1135		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1136	} else if (td->o.verify_async) {
1137		/*
1138		 * We ran out, wait for async verify threads to finish and
1139		 * return one
1140		 */
1141		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1142		goto again;
1143	}
1144
1145	td_io_u_unlock(td);
1146	return io_u;
1147}
1148
1149static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1150{
1151	if (!(td->flags & TD_F_TRIM_BACKLOG))
1152		return 0;
1153
1154	if (td->trim_entries) {
1155		int get_trim = 0;
1156
1157		if (td->trim_batch) {
1158			td->trim_batch--;
1159			get_trim = 1;
1160		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1161			 td->last_ddir != DDIR_READ) {
1162			td->trim_batch = td->o.trim_batch;
1163			if (!td->trim_batch)
1164				td->trim_batch = td->o.trim_backlog;
1165			get_trim = 1;
1166		}
1167
1168		if (get_trim && !get_next_trim(td, io_u))
1169			return 1;
1170	}
1171
1172	return 0;
1173}
1174
1175static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1176{
1177	if (!(td->flags & TD_F_VER_BACKLOG))
1178		return 0;
1179
1180	if (td->io_hist_len) {
1181		int get_verify = 0;
1182
1183		if (td->verify_batch)
1184			get_verify = 1;
1185		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1186			 td->last_ddir != DDIR_READ) {
1187			td->verify_batch = td->o.verify_batch;
1188			if (!td->verify_batch)
1189				td->verify_batch = td->o.verify_backlog;
1190			get_verify = 1;
1191		}
1192
1193		if (get_verify && !get_next_verify(td, io_u)) {
1194			td->verify_batch--;
1195			return 1;
1196		}
1197	}
1198
1199	return 0;
1200}
1201
1202/*
1203 * Fill offset and start time into the buffer content, to prevent too
1204 * easy compressible data for simple de-dupe attempts. Do this for every
1205 * 512b block in the range, since that should be the smallest block size
1206 * we can expect from a device.
1207 */
1208static void small_content_scramble(struct io_u *io_u)
1209{
1210	unsigned int i, nr_blocks = io_u->buflen / 512;
1211	uint64_t boffset;
1212	unsigned int offset;
1213	void *p, *end;
1214
1215	if (!nr_blocks)
1216		return;
1217
1218	p = io_u->xfer_buf;
1219	boffset = io_u->offset;
1220	io_u->buf_filled_len = 0;
1221
1222	for (i = 0; i < nr_blocks; i++) {
1223		/*
1224		 * Fill the byte offset into a "random" start offset of
1225		 * the buffer, given by the product of the usec time
1226		 * and the actual offset.
1227		 */
1228		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1229		offset &= ~(sizeof(uint64_t) - 1);
1230		if (offset >= 512 - sizeof(uint64_t))
1231			offset -= sizeof(uint64_t);
1232		memcpy(p + offset, &boffset, sizeof(boffset));
1233
1234		end = p + 512 - sizeof(io_u->start_time);
1235		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1236		p += 512;
1237		boffset += 512;
1238	}
1239}
1240
1241/*
1242 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1243 * etc. The returned io_u is fully ready to be prepped and submitted.
1244 */
1245struct io_u *get_io_u(struct thread_data *td)
1246{
1247	struct fio_file *f;
1248	struct io_u *io_u;
1249	int do_scramble = 0;
1250
1251	io_u = __get_io_u(td);
1252	if (!io_u) {
1253		dprint(FD_IO, "__get_io_u failed\n");
1254		return NULL;
1255	}
1256
1257	if (check_get_verify(td, io_u))
1258		goto out;
1259	if (check_get_trim(td, io_u))
1260		goto out;
1261
1262	/*
1263	 * from a requeue, io_u already setup
1264	 */
1265	if (io_u->file)
1266		goto out;
1267
1268	/*
1269	 * If using an iolog, grab next piece if any available.
1270	 */
1271	if (td->flags & TD_F_READ_IOLOG) {
1272		if (read_iolog_get(td, io_u))
1273			goto err_put;
1274	} else if (set_io_u_file(td, io_u)) {
1275		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1276		goto err_put;
1277	}
1278
1279	f = io_u->file;
1280	assert(fio_file_open(f));
1281
1282	if (ddir_rw(io_u->ddir)) {
1283		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1284			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1285			goto err_put;
1286		}
1287
1288		f->last_start = io_u->offset;
1289		f->last_pos = io_u->offset + io_u->buflen;
1290
1291		if (io_u->ddir == DDIR_WRITE) {
1292			if (td->flags & TD_F_REFILL_BUFFERS) {
1293				io_u_fill_buffer(td, io_u,
1294					io_u->xfer_buflen, io_u->xfer_buflen);
1295			} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1296				do_scramble = 1;
1297			if (td->flags & TD_F_VER_NONE) {
1298				populate_verify_io_u(td, io_u);
1299				do_scramble = 0;
1300			}
1301		} else if (io_u->ddir == DDIR_READ) {
1302			/*
1303			 * Reset the buf_filled parameters so next time if the
1304			 * buffer is used for writes it is refilled.
1305			 */
1306			io_u->buf_filled_len = 0;
1307		}
1308	}
1309
1310	/*
1311	 * Set io data pointers.
1312	 */
1313	io_u->xfer_buf = io_u->buf;
1314	io_u->xfer_buflen = io_u->buflen;
1315
1316out:
1317	assert(io_u->file);
1318	if (!td_io_prep(td, io_u)) {
1319		if (!td->o.disable_slat)
1320			fio_gettime(&io_u->start_time, NULL);
1321		if (do_scramble)
1322			small_content_scramble(io_u);
1323		return io_u;
1324	}
1325err_put:
1326	dprint(FD_IO, "get_io_u failed\n");
1327	put_io_u(td, io_u);
1328	return NULL;
1329}
1330
1331void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1332{
1333	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1334	const char *msg[] = { "read", "write", "sync", "datasync",
1335				"sync_file_range", "wait", "trim" };
1336
1337	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1338		return;
1339
1340	log_err("fio: io_u error");
1341
1342	if (io_u->file)
1343		log_err(" on file %s", io_u->file->file_name);
1344
1345	log_err(": %s\n", strerror(io_u->error));
1346
1347	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1348					io_u->offset, io_u->xfer_buflen);
1349
1350	if (!td->error)
1351		td_verror(td, io_u->error, "io_u error");
1352}
1353
1354static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1355				  struct io_completion_data *icd,
1356				  const enum fio_ddir idx, unsigned int bytes)
1357{
1358	unsigned long lusec = 0;
1359
1360	if (!td->o.disable_clat || !td->o.disable_bw)
1361		lusec = utime_since(&io_u->issue_time, &icd->time);
1362
1363	if (!td->o.disable_lat) {
1364		unsigned long tusec;
1365
1366		tusec = utime_since(&io_u->start_time, &icd->time);
1367		add_lat_sample(td, idx, tusec, bytes);
1368
1369		if (td->flags & TD_F_PROFILE_OPS) {
1370			struct prof_io_ops *ops = &td->prof_io_ops;
1371
1372			if (ops->io_u_lat)
1373				icd->error = ops->io_u_lat(td, tusec);
1374		}
1375
1376		if (td->o.max_latency && tusec > td->o.max_latency) {
1377			if (!td->error)
1378				log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1379			td_verror(td, ETIMEDOUT, "max latency exceeded");
1380			icd->error = ETIMEDOUT;
1381		}
1382	}
1383
1384	if (!td->o.disable_clat) {
1385		add_clat_sample(td, idx, lusec, bytes);
1386		io_u_mark_latency(td, lusec);
1387	}
1388
1389	if (!td->o.disable_bw)
1390		add_bw_sample(td, idx, bytes, &icd->time);
1391
1392	add_iops_sample(td, idx, &icd->time);
1393}
1394
1395static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1396{
1397	uint64_t secs, remainder, bps, bytes;
1398
1399	bytes = td->this_io_bytes[ddir];
1400	bps = td->rate_bps[ddir];
1401	secs = bytes / bps;
1402	remainder = bytes % bps;
1403	return remainder * 1000000 / bps + secs * 1000000;
1404}
1405
1406static void io_completed(struct thread_data *td, struct io_u *io_u,
1407			 struct io_completion_data *icd)
1408{
1409	struct fio_file *f;
1410
1411	dprint_io_u(io_u, "io complete");
1412
1413	td_io_u_lock(td);
1414	assert(io_u->flags & IO_U_F_FLIGHT);
1415	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1416	td_io_u_unlock(td);
1417
1418	if (ddir_sync(io_u->ddir)) {
1419		td->last_was_sync = 1;
1420		f = io_u->file;
1421		if (f) {
1422			f->first_write = -1ULL;
1423			f->last_write = -1ULL;
1424		}
1425		return;
1426	}
1427
1428	td->last_was_sync = 0;
1429	td->last_ddir = io_u->ddir;
1430
1431	if (!io_u->error && ddir_rw(io_u->ddir)) {
1432		unsigned int bytes = io_u->buflen - io_u->resid;
1433		const enum fio_ddir idx = io_u->ddir;
1434		const enum fio_ddir odx = io_u->ddir ^ 1;
1435		int ret;
1436
1437		td->io_blocks[idx]++;
1438		td->this_io_blocks[idx]++;
1439		td->io_bytes[idx] += bytes;
1440
1441		if (!(io_u->flags & IO_U_F_VER_LIST))
1442			td->this_io_bytes[idx] += bytes;
1443
1444		if (idx == DDIR_WRITE) {
1445			f = io_u->file;
1446			if (f) {
1447				if (f->first_write == -1ULL ||
1448				    io_u->offset < f->first_write)
1449					f->first_write = io_u->offset;
1450				if (f->last_write == -1ULL ||
1451				    ((io_u->offset + bytes) > f->last_write))
1452					f->last_write = io_u->offset + bytes;
1453			}
1454		}
1455
1456		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1457					   td->runstate == TD_VERIFYING)) {
1458			account_io_completion(td, io_u, icd, idx, bytes);
1459
1460			if (__should_check_rate(td, idx)) {
1461				td->rate_pending_usleep[idx] =
1462					(usec_for_io(td, idx) -
1463					 utime_since_now(&td->start));
1464			}
1465			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1466				td->rate_pending_usleep[odx] =
1467					(usec_for_io(td, odx) -
1468					 utime_since_now(&td->start));
1469		}
1470
1471		if (td_write(td) && idx == DDIR_WRITE &&
1472		    td->o.do_verify &&
1473		    td->o.verify != VERIFY_NONE &&
1474		    !td->o.experimental_verify)
1475			log_io_piece(td, io_u);
1476
1477		icd->bytes_done[idx] += bytes;
1478
1479		if (io_u->end_io) {
1480			ret = io_u->end_io(td, io_u);
1481			if (ret && !icd->error)
1482				icd->error = ret;
1483		}
1484	} else if (io_u->error) {
1485		icd->error = io_u->error;
1486		io_u_log_error(td, io_u);
1487	}
1488	if (icd->error) {
1489		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1490		if (!td_non_fatal_error(td, eb, icd->error))
1491			return;
1492		/*
1493		 * If there is a non_fatal error, then add to the error count
1494		 * and clear all the errors.
1495		 */
1496		update_error_count(td, icd->error);
1497		td_clear_error(td);
1498		icd->error = 0;
1499		io_u->error = 0;
1500	}
1501}
1502
1503static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1504		     int nr)
1505{
1506	int ddir;
1507	if (!td->o.disable_clat || !td->o.disable_bw)
1508		fio_gettime(&icd->time, NULL);
1509
1510	icd->nr = nr;
1511
1512	icd->error = 0;
1513	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1514		icd->bytes_done[ddir] = 0;
1515}
1516
1517static void ios_completed(struct thread_data *td,
1518			  struct io_completion_data *icd)
1519{
1520	struct io_u *io_u;
1521	int i;
1522
1523	for (i = 0; i < icd->nr; i++) {
1524		io_u = td->io_ops->event(td, i);
1525
1526		io_completed(td, io_u, icd);
1527
1528		if (!(io_u->flags & IO_U_F_FREE_DEF))
1529			put_io_u(td, io_u);
1530	}
1531}
1532
1533/*
1534 * Complete a single io_u for the sync engines.
1535 */
1536int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1537		       uint64_t *bytes)
1538{
1539	struct io_completion_data icd;
1540
1541	init_icd(td, &icd, 1);
1542	io_completed(td, io_u, &icd);
1543
1544	if (!(io_u->flags & IO_U_F_FREE_DEF))
1545		put_io_u(td, io_u);
1546
1547	if (icd.error) {
1548		td_verror(td, icd.error, "io_u_sync_complete");
1549		return -1;
1550	}
1551
1552	if (bytes) {
1553		int ddir;
1554
1555		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1556			bytes[ddir] += icd.bytes_done[ddir];
1557	}
1558
1559	return 0;
1560}
1561
1562/*
1563 * Called to complete min_events number of io for the async engines.
1564 */
1565int io_u_queued_complete(struct thread_data *td, int min_evts,
1566			 uint64_t *bytes)
1567{
1568	struct io_completion_data icd;
1569	struct timespec *tvp = NULL;
1570	int ret;
1571	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1572
1573	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1574
1575	if (!min_evts)
1576		tvp = &ts;
1577
1578	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1579	if (ret < 0) {
1580		td_verror(td, -ret, "td_io_getevents");
1581		return ret;
1582	} else if (!ret)
1583		return ret;
1584
1585	init_icd(td, &icd, ret);
1586	ios_completed(td, &icd);
1587	if (icd.error) {
1588		td_verror(td, icd.error, "io_u_queued_complete");
1589		return -1;
1590	}
1591
1592	if (bytes) {
1593		int ddir;
1594
1595		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1596			bytes[ddir] += icd.bytes_done[ddir];
1597	}
1598
1599	return 0;
1600}
1601
1602/*
1603 * Call when io_u is really queued, to update the submission latency.
1604 */
1605void io_u_queued(struct thread_data *td, struct io_u *io_u)
1606{
1607	if (!td->o.disable_slat) {
1608		unsigned long slat_time;
1609
1610		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1611		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1612	}
1613}
1614
1615void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1616		    unsigned int max_bs)
1617{
1618	if (!td->o.zero_buffers) {
1619		unsigned int perc = td->o.compress_percentage;
1620
1621		if (perc) {
1622			unsigned int seg = min_write;
1623
1624			seg = min(min_write, td->o.compress_chunk);
1625			if (!seg)
1626				seg = min_write;
1627
1628			fill_random_buf_percentage(&td->buf_state, buf,
1629						perc, seg, max_bs);
1630		} else
1631			fill_random_buf(&td->buf_state, buf, max_bs);
1632	} else
1633		memset(buf, 0, max_bs);
1634}
1635
1636/*
1637 * "randomly" fill the buffer contents
1638 */
1639void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1640		      unsigned int min_write, unsigned int max_bs)
1641{
1642	io_u->buf_filled_len = 0;
1643	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1644}
1645