io_u.c revision efd633f3ae8fd7aa81bd2a106629aba6f2a0bd76
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	if (td->o.min_bs[ddir] > td->o.ba[ddir])
72		max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
73
74	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
75	if (!max_blocks)
76		return 0;
77
78	return max_blocks;
79}
80
81struct rand_off {
82	struct flist_head list;
83	uint64_t off;
84};
85
86static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
87				  enum fio_ddir ddir, uint64_t *b)
88{
89	uint64_t r;
90
91	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
92		uint64_t lastb;
93
94		lastb = last_block(td, f, ddir);
95		if (!lastb)
96			return 1;
97
98		r = __rand(&td->random_state);
99
100		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
101
102		*b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
103	} else {
104		uint64_t off = 0;
105
106		assert(fio_file_lfsr(f));
107
108		if (lfsr_next(&f->lfsr, &off))
109			return 1;
110
111		*b = off;
112	}
113
114	/*
115	 * if we are not maintaining a random map, we are done.
116	 */
117	if (!file_randommap(td, f))
118		goto ret;
119
120	/*
121	 * calculate map offset and check if it's free
122	 */
123	if (random_map_free(f, *b))
124		goto ret;
125
126	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127						(unsigned long long) *b);
128
129	*b = axmap_next_free(f->io_axmap, *b);
130	if (*b == (uint64_t) -1ULL)
131		return 1;
132ret:
133	return 0;
134}
135
136static int __get_next_rand_offset_zipf(struct thread_data *td,
137				       struct fio_file *f, enum fio_ddir ddir,
138				       uint64_t *b)
139{
140	*b = zipf_next(&f->zipf);
141	return 0;
142}
143
144static int __get_next_rand_offset_pareto(struct thread_data *td,
145					 struct fio_file *f, enum fio_ddir ddir,
146					 uint64_t *b)
147{
148	*b = pareto_next(&f->zipf);
149	return 0;
150}
151
152static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
153{
154	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
155	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
156
157	return r1->off - r2->off;
158}
159
160static int get_off_from_method(struct thread_data *td, struct fio_file *f,
161			       enum fio_ddir ddir, uint64_t *b)
162{
163	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
164		return __get_next_rand_offset(td, f, ddir, b);
165	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
166		return __get_next_rand_offset_zipf(td, f, ddir, b);
167	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
168		return __get_next_rand_offset_pareto(td, f, ddir, b);
169
170	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
171	return 1;
172}
173
174/*
175 * Sort the reads for a verify phase in batches of verifysort_nr, if
176 * specified.
177 */
178static inline int should_sort_io(struct thread_data *td)
179{
180	if (!td->o.verifysort_nr || !td->o.do_verify)
181		return 0;
182	if (!td_random(td))
183		return 0;
184	if (td->runstate != TD_VERIFYING)
185		return 0;
186	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
187		return 0;
188
189	return 1;
190}
191
192static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
193{
194	unsigned int v;
195	unsigned long r;
196
197	if (td->o.perc_rand[ddir] == 100)
198		return 1;
199
200	r = __rand(&td->seq_rand_state[ddir]);
201	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
202
203	return v <= td->o.perc_rand[ddir];
204}
205
206static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
207				enum fio_ddir ddir, uint64_t *b)
208{
209	struct rand_off *r;
210	int i, ret = 1;
211
212	if (!should_sort_io(td))
213		return get_off_from_method(td, f, ddir, b);
214
215	if (!flist_empty(&td->next_rand_list)) {
216fetch:
217		r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
218		flist_del(&r->list);
219		*b = r->off;
220		free(r);
221		return 0;
222	}
223
224	for (i = 0; i < td->o.verifysort_nr; i++) {
225		r = malloc(sizeof(*r));
226
227		ret = get_off_from_method(td, f, ddir, &r->off);
228		if (ret) {
229			free(r);
230			break;
231		}
232
233		flist_add(&r->list, &td->next_rand_list);
234	}
235
236	if (ret && !i)
237		return ret;
238
239	assert(!flist_empty(&td->next_rand_list));
240	flist_sort(NULL, &td->next_rand_list, flist_cmp);
241	goto fetch;
242}
243
244static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
245			       enum fio_ddir ddir, uint64_t *b)
246{
247	if (!get_next_rand_offset(td, f, ddir, b))
248		return 0;
249
250	if (td->o.time_based) {
251		fio_file_reset(td, f);
252		if (!get_next_rand_offset(td, f, ddir, b))
253			return 0;
254	}
255
256	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
257			f->file_name, (unsigned long long) f->last_pos[ddir],
258			(unsigned long long) f->real_file_size);
259	return 1;
260}
261
262static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
263			       enum fio_ddir ddir, uint64_t *offset)
264{
265	struct thread_options *o = &td->o;
266
267	assert(ddir_rw(ddir));
268
269	if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
270	    o->time_based)
271		f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
272
273	if (f->last_pos[ddir] < f->real_file_size) {
274		uint64_t pos;
275
276		if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
277			f->last_pos[ddir] = f->real_file_size;
278
279		pos = f->last_pos[ddir] - f->file_offset;
280		if (pos && o->ddir_seq_add) {
281			pos += o->ddir_seq_add;
282
283			/*
284			 * If we reach beyond the end of the file
285			 * with holed IO, wrap around to the
286			 * beginning again.
287			 */
288			if (pos >= f->real_file_size)
289				pos = f->file_offset;
290		}
291
292		*offset = pos;
293		return 0;
294	}
295
296	return 1;
297}
298
299static int get_next_block(struct thread_data *td, struct io_u *io_u,
300			  enum fio_ddir ddir, int rw_seq,
301			  unsigned int *is_random)
302{
303	struct fio_file *f = io_u->file;
304	uint64_t b, offset;
305	int ret;
306
307	assert(ddir_rw(ddir));
308
309	b = offset = -1ULL;
310
311	if (rw_seq) {
312		if (td_random(td)) {
313			if (should_do_random(td, ddir)) {
314				ret = get_next_rand_block(td, f, ddir, &b);
315				*is_random = 1;
316			} else {
317				*is_random = 0;
318				io_u->flags |= IO_U_F_BUSY_OK;
319				ret = get_next_seq_offset(td, f, ddir, &offset);
320				if (ret)
321					ret = get_next_rand_block(td, f, ddir, &b);
322			}
323		} else {
324			*is_random = 0;
325			ret = get_next_seq_offset(td, f, ddir, &offset);
326		}
327	} else {
328		io_u->flags |= IO_U_F_BUSY_OK;
329		*is_random = 0;
330
331		if (td->o.rw_seq == RW_SEQ_SEQ) {
332			ret = get_next_seq_offset(td, f, ddir, &offset);
333			if (ret) {
334				ret = get_next_rand_block(td, f, ddir, &b);
335				*is_random = 0;
336			}
337		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
338			if (f->last_start[ddir] != -1ULL)
339				offset = f->last_start[ddir] - f->file_offset;
340			else
341				offset = 0;
342			ret = 0;
343		} else {
344			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
345			ret = 1;
346		}
347	}
348
349	if (!ret) {
350		if (offset != -1ULL)
351			io_u->offset = offset;
352		else if (b != -1ULL)
353			io_u->offset = b * td->o.ba[ddir];
354		else {
355			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
356			ret = 1;
357		}
358	}
359
360	return ret;
361}
362
363/*
364 * For random io, generate a random new block and see if it's used. Repeat
365 * until we find a free one. For sequential io, just return the end of
366 * the last io issued.
367 */
368static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
369			     unsigned int *is_random)
370{
371	struct fio_file *f = io_u->file;
372	enum fio_ddir ddir = io_u->ddir;
373	int rw_seq_hit = 0;
374
375	assert(ddir_rw(ddir));
376
377	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
378		rw_seq_hit = 1;
379		td->ddir_seq_nr = td->o.ddir_seq_nr;
380	}
381
382	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
383		return 1;
384
385	if (io_u->offset >= f->io_size) {
386		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
387					(unsigned long long) io_u->offset,
388					(unsigned long long) f->io_size);
389		return 1;
390	}
391
392	io_u->offset += f->file_offset;
393	if (io_u->offset >= f->real_file_size) {
394		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
395					(unsigned long long) io_u->offset,
396					(unsigned long long) f->real_file_size);
397		return 1;
398	}
399
400	return 0;
401}
402
403static int get_next_offset(struct thread_data *td, struct io_u *io_u,
404			   unsigned int *is_random)
405{
406	if (td->flags & TD_F_PROFILE_OPS) {
407		struct prof_io_ops *ops = &td->prof_io_ops;
408
409		if (ops->fill_io_u_off)
410			return ops->fill_io_u_off(td, io_u, is_random);
411	}
412
413	return __get_next_offset(td, io_u, is_random);
414}
415
416static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
417			    unsigned int buflen)
418{
419	struct fio_file *f = io_u->file;
420
421	return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
422}
423
424static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
425				      unsigned int is_random)
426{
427	int ddir = io_u->ddir;
428	unsigned int buflen = 0;
429	unsigned int minbs, maxbs;
430	unsigned long r;
431
432	assert(ddir_rw(ddir));
433
434	if (td->o.bs_is_seq_rand)
435		ddir = is_random ? DDIR_WRITE: DDIR_READ;
436
437	minbs = td->o.min_bs[ddir];
438	maxbs = td->o.max_bs[ddir];
439
440	if (minbs == maxbs)
441		return minbs;
442
443	/*
444	 * If we can't satisfy the min block size from here, then fail
445	 */
446	if (!io_u_fits(td, io_u, minbs))
447		return 0;
448
449	do {
450		r = __rand(&td->bsrange_state);
451
452		if (!td->o.bssplit_nr[ddir]) {
453			buflen = 1 + (unsigned int) ((double) maxbs *
454					(r / (FRAND_MAX + 1.0)));
455			if (buflen < minbs)
456				buflen = minbs;
457		} else {
458			long perc = 0;
459			unsigned int i;
460
461			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
462				struct bssplit *bsp = &td->o.bssplit[ddir][i];
463
464				buflen = bsp->bs;
465				perc += bsp->perc;
466				if ((r <= ((FRAND_MAX / 100L) * perc)) &&
467				    io_u_fits(td, io_u, buflen))
468					break;
469			}
470		}
471
472		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
473			buflen = (buflen + td->o.verify_interval - 1) &
474				~(td->o.verify_interval - 1);
475
476		if (!td->o.bs_unaligned && is_power_of_2(minbs))
477			buflen = (buflen + minbs - 1) & ~(minbs - 1);
478
479	} while (!io_u_fits(td, io_u, buflen));
480
481	return buflen;
482}
483
484static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
485				    unsigned int is_random)
486{
487	if (td->flags & TD_F_PROFILE_OPS) {
488		struct prof_io_ops *ops = &td->prof_io_ops;
489
490		if (ops->fill_io_u_size)
491			return ops->fill_io_u_size(td, io_u, is_random);
492	}
493
494	return __get_next_buflen(td, io_u, is_random);
495}
496
497static void set_rwmix_bytes(struct thread_data *td)
498{
499	unsigned int diff;
500
501	/*
502	 * we do time or byte based switch. this is needed because
503	 * buffered writes may issue a lot quicker than they complete,
504	 * whereas reads do not.
505	 */
506	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
507	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
508}
509
510static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
511{
512	unsigned int v;
513	unsigned long r;
514
515	r = __rand(&td->rwmix_state);
516	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
517
518	if (v <= td->o.rwmix[DDIR_READ])
519		return DDIR_READ;
520
521	return DDIR_WRITE;
522}
523
524void io_u_quiesce(struct thread_data *td)
525{
526	/*
527	 * We are going to sleep, ensure that we flush anything pending as
528	 * not to skew our latency numbers.
529	 *
530	 * Changed to only monitor 'in flight' requests here instead of the
531	 * td->cur_depth, b/c td->cur_depth does not accurately represent
532	 * io's that have been actually submitted to an async engine,
533	 * and cur_depth is meaningless for sync engines.
534	 */
535	if (td->io_u_queued || td->cur_depth) {
536		int fio_unused ret;
537
538		ret = td_io_commit(td);
539	}
540
541	while (td->io_u_in_flight) {
542		int fio_unused ret;
543
544		ret = io_u_queued_complete(td, 1, NULL);
545	}
546}
547
548static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
549{
550	enum fio_ddir odir = ddir ^ 1;
551	long usec;
552
553	assert(ddir_rw(ddir));
554
555	if (td->rate_pending_usleep[ddir] <= 0)
556		return ddir;
557
558	/*
559	 * We have too much pending sleep in this direction. See if we
560	 * should switch.
561	 */
562	if (td_rw(td) && td->o.rwmix[odir]) {
563		/*
564		 * Other direction does not have too much pending, switch
565		 */
566		if (td->rate_pending_usleep[odir] < 100000)
567			return odir;
568
569		/*
570		 * Both directions have pending sleep. Sleep the minimum time
571		 * and deduct from both.
572		 */
573		if (td->rate_pending_usleep[ddir] <=
574			td->rate_pending_usleep[odir]) {
575			usec = td->rate_pending_usleep[ddir];
576		} else {
577			usec = td->rate_pending_usleep[odir];
578			ddir = odir;
579		}
580	} else
581		usec = td->rate_pending_usleep[ddir];
582
583	io_u_quiesce(td);
584
585	usec = usec_sleep(td, usec);
586
587	td->rate_pending_usleep[ddir] -= usec;
588
589	odir = ddir ^ 1;
590	if (td_rw(td) && __should_check_rate(td, odir))
591		td->rate_pending_usleep[odir] -= usec;
592
593	if (ddir == DDIR_TRIM)
594		return DDIR_TRIM;
595
596	return ddir;
597}
598
599/*
600 * Return the data direction for the next io_u. If the job is a
601 * mixed read/write workload, check the rwmix cycle and switch if
602 * necessary.
603 */
604static enum fio_ddir get_rw_ddir(struct thread_data *td)
605{
606	enum fio_ddir ddir;
607
608	/*
609	 * see if it's time to fsync
610	 */
611	if (td->o.fsync_blocks &&
612	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
613	     td->io_issues[DDIR_WRITE] && should_fsync(td))
614		return DDIR_SYNC;
615
616	/*
617	 * see if it's time to fdatasync
618	 */
619	if (td->o.fdatasync_blocks &&
620	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
621	     td->io_issues[DDIR_WRITE] && should_fsync(td))
622		return DDIR_DATASYNC;
623
624	/*
625	 * see if it's time to sync_file_range
626	 */
627	if (td->sync_file_range_nr &&
628	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
629	     td->io_issues[DDIR_WRITE] && should_fsync(td))
630		return DDIR_SYNC_FILE_RANGE;
631
632	if (td_rw(td)) {
633		/*
634		 * Check if it's time to seed a new data direction.
635		 */
636		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
637			/*
638			 * Put a top limit on how many bytes we do for
639			 * one data direction, to avoid overflowing the
640			 * ranges too much
641			 */
642			ddir = get_rand_ddir(td);
643
644			if (ddir != td->rwmix_ddir)
645				set_rwmix_bytes(td);
646
647			td->rwmix_ddir = ddir;
648		}
649		ddir = td->rwmix_ddir;
650	} else if (td_read(td))
651		ddir = DDIR_READ;
652	else if (td_write(td))
653		ddir = DDIR_WRITE;
654	else
655		ddir = DDIR_TRIM;
656
657	td->rwmix_ddir = rate_ddir(td, ddir);
658	return td->rwmix_ddir;
659}
660
661static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
662{
663	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
664
665	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
666	    td->o.barrier_blocks &&
667	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
668	     td->io_issues[DDIR_WRITE])
669		io_u->flags |= IO_U_F_BARRIER;
670}
671
672void put_file_log(struct thread_data *td, struct fio_file *f)
673{
674	unsigned int ret = put_file(td, f);
675
676	if (ret)
677		td_verror(td, ret, "file close");
678}
679
680void put_io_u(struct thread_data *td, struct io_u *io_u)
681{
682	td_io_u_lock(td);
683
684	if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
685		put_file_log(td, io_u->file);
686
687	io_u->file = NULL;
688	io_u->flags |= IO_U_F_FREE;
689
690	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
691		td->cur_depth--;
692	io_u_qpush(&td->io_u_freelist, io_u);
693	td_io_u_unlock(td);
694	td_io_u_free_notify(td);
695}
696
697void clear_io_u(struct thread_data *td, struct io_u *io_u)
698{
699	io_u->flags &= ~IO_U_F_FLIGHT;
700	put_io_u(td, io_u);
701}
702
703void requeue_io_u(struct thread_data *td, struct io_u **io_u)
704{
705	struct io_u *__io_u = *io_u;
706	enum fio_ddir ddir = acct_ddir(__io_u);
707
708	dprint(FD_IO, "requeue %p\n", __io_u);
709
710	td_io_u_lock(td);
711
712	__io_u->flags |= IO_U_F_FREE;
713	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
714		td->io_issues[ddir]--;
715
716	__io_u->flags &= ~IO_U_F_FLIGHT;
717	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
718		td->cur_depth--;
719
720	io_u_rpush(&td->io_u_requeues, __io_u);
721	td_io_u_unlock(td);
722	*io_u = NULL;
723}
724
725static int fill_io_u(struct thread_data *td, struct io_u *io_u)
726{
727	unsigned int is_random;
728
729	if (td->io_ops->flags & FIO_NOIO)
730		goto out;
731
732	set_rw_ddir(td, io_u);
733
734	/*
735	 * fsync() or fdatasync() or trim etc, we are done
736	 */
737	if (!ddir_rw(io_u->ddir))
738		goto out;
739
740	/*
741	 * See if it's time to switch to a new zone
742	 */
743	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
744		struct fio_file *f = io_u->file;
745
746		td->zone_bytes = 0;
747		f->file_offset += td->o.zone_range + td->o.zone_skip;
748
749		/*
750		 * Wrap from the beginning, if we exceed the file size
751		 */
752		if (f->file_offset >= f->real_file_size)
753			f->file_offset = f->real_file_size - f->file_offset;
754		f->last_pos[io_u->ddir] = f->file_offset;
755		td->io_skip_bytes += td->o.zone_skip;
756	}
757
758	/*
759	 * No log, let the seq/rand engine retrieve the next buflen and
760	 * position.
761	 */
762	if (get_next_offset(td, io_u, &is_random)) {
763		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
764		return 1;
765	}
766
767	io_u->buflen = get_next_buflen(td, io_u, is_random);
768	if (!io_u->buflen) {
769		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
770		return 1;
771	}
772
773	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
774		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
775		dprint(FD_IO, "  off=%llu/%lu > %llu\n",
776			(unsigned long long) io_u->offset, io_u->buflen,
777			(unsigned long long) io_u->file->real_file_size);
778		return 1;
779	}
780
781	/*
782	 * mark entry before potentially trimming io_u
783	 */
784	if (td_random(td) && file_randommap(td, io_u->file))
785		mark_random_map(td, io_u);
786
787out:
788	dprint_io_u(io_u, "fill_io_u");
789	td->zone_bytes += io_u->buflen;
790	return 0;
791}
792
793static void __io_u_mark_map(unsigned int *map, unsigned int nr)
794{
795	int idx = 0;
796
797	switch (nr) {
798	default:
799		idx = 6;
800		break;
801	case 33 ... 64:
802		idx = 5;
803		break;
804	case 17 ... 32:
805		idx = 4;
806		break;
807	case 9 ... 16:
808		idx = 3;
809		break;
810	case 5 ... 8:
811		idx = 2;
812		break;
813	case 1 ... 4:
814		idx = 1;
815	case 0:
816		break;
817	}
818
819	map[idx]++;
820}
821
822void io_u_mark_submit(struct thread_data *td, unsigned int nr)
823{
824	__io_u_mark_map(td->ts.io_u_submit, nr);
825	td->ts.total_submit++;
826}
827
828void io_u_mark_complete(struct thread_data *td, unsigned int nr)
829{
830	__io_u_mark_map(td->ts.io_u_complete, nr);
831	td->ts.total_complete++;
832}
833
834void io_u_mark_depth(struct thread_data *td, unsigned int nr)
835{
836	int idx = 0;
837
838	switch (td->cur_depth) {
839	default:
840		idx = 6;
841		break;
842	case 32 ... 63:
843		idx = 5;
844		break;
845	case 16 ... 31:
846		idx = 4;
847		break;
848	case 8 ... 15:
849		idx = 3;
850		break;
851	case 4 ... 7:
852		idx = 2;
853		break;
854	case 2 ... 3:
855		idx = 1;
856	case 1:
857		break;
858	}
859
860	td->ts.io_u_map[idx] += nr;
861}
862
863static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
864{
865	int idx = 0;
866
867	assert(usec < 1000);
868
869	switch (usec) {
870	case 750 ... 999:
871		idx = 9;
872		break;
873	case 500 ... 749:
874		idx = 8;
875		break;
876	case 250 ... 499:
877		idx = 7;
878		break;
879	case 100 ... 249:
880		idx = 6;
881		break;
882	case 50 ... 99:
883		idx = 5;
884		break;
885	case 20 ... 49:
886		idx = 4;
887		break;
888	case 10 ... 19:
889		idx = 3;
890		break;
891	case 4 ... 9:
892		idx = 2;
893		break;
894	case 2 ... 3:
895		idx = 1;
896	case 0 ... 1:
897		break;
898	}
899
900	assert(idx < FIO_IO_U_LAT_U_NR);
901	td->ts.io_u_lat_u[idx]++;
902}
903
904static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
905{
906	int idx = 0;
907
908	switch (msec) {
909	default:
910		idx = 11;
911		break;
912	case 1000 ... 1999:
913		idx = 10;
914		break;
915	case 750 ... 999:
916		idx = 9;
917		break;
918	case 500 ... 749:
919		idx = 8;
920		break;
921	case 250 ... 499:
922		idx = 7;
923		break;
924	case 100 ... 249:
925		idx = 6;
926		break;
927	case 50 ... 99:
928		idx = 5;
929		break;
930	case 20 ... 49:
931		idx = 4;
932		break;
933	case 10 ... 19:
934		idx = 3;
935		break;
936	case 4 ... 9:
937		idx = 2;
938		break;
939	case 2 ... 3:
940		idx = 1;
941	case 0 ... 1:
942		break;
943	}
944
945	assert(idx < FIO_IO_U_LAT_M_NR);
946	td->ts.io_u_lat_m[idx]++;
947}
948
949static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
950{
951	if (usec < 1000)
952		io_u_mark_lat_usec(td, usec);
953	else
954		io_u_mark_lat_msec(td, usec / 1000);
955}
956
957/*
958 * Get next file to service by choosing one at random
959 */
960static struct fio_file *get_next_file_rand(struct thread_data *td,
961					   enum fio_file_flags goodf,
962					   enum fio_file_flags badf)
963{
964	struct fio_file *f;
965	int fno;
966
967	do {
968		int opened = 0;
969		unsigned long r;
970
971		r = __rand(&td->next_file_state);
972		fno = (unsigned int) ((double) td->o.nr_files
973				* (r / (FRAND_MAX + 1.0)));
974
975		f = td->files[fno];
976		if (fio_file_done(f))
977			continue;
978
979		if (!fio_file_open(f)) {
980			int err;
981
982			if (td->nr_open_files >= td->o.open_files)
983				return ERR_PTR(-EBUSY);
984
985			err = td_io_open_file(td, f);
986			if (err)
987				continue;
988			opened = 1;
989		}
990
991		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
992			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
993			return f;
994		}
995		if (opened)
996			td_io_close_file(td, f);
997	} while (1);
998}
999
1000/*
1001 * Get next file to service by doing round robin between all available ones
1002 */
1003static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1004					 int badf)
1005{
1006	unsigned int old_next_file = td->next_file;
1007	struct fio_file *f;
1008
1009	do {
1010		int opened = 0;
1011
1012		f = td->files[td->next_file];
1013
1014		td->next_file++;
1015		if (td->next_file >= td->o.nr_files)
1016			td->next_file = 0;
1017
1018		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1019		if (fio_file_done(f)) {
1020			f = NULL;
1021			continue;
1022		}
1023
1024		if (!fio_file_open(f)) {
1025			int err;
1026
1027			if (td->nr_open_files >= td->o.open_files)
1028				return ERR_PTR(-EBUSY);
1029
1030			err = td_io_open_file(td, f);
1031			if (err) {
1032				dprint(FD_FILE, "error %d on open of %s\n",
1033					err, f->file_name);
1034				f = NULL;
1035				continue;
1036			}
1037			opened = 1;
1038		}
1039
1040		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1041								f->flags);
1042		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1043			break;
1044
1045		if (opened)
1046			td_io_close_file(td, f);
1047
1048		f = NULL;
1049	} while (td->next_file != old_next_file);
1050
1051	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1052	return f;
1053}
1054
1055static struct fio_file *__get_next_file(struct thread_data *td)
1056{
1057	struct fio_file *f;
1058
1059	assert(td->o.nr_files <= td->files_index);
1060
1061	if (td->nr_done_files >= td->o.nr_files) {
1062		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1063				" nr_files=%d\n", td->nr_open_files,
1064						  td->nr_done_files,
1065						  td->o.nr_files);
1066		return NULL;
1067	}
1068
1069	f = td->file_service_file;
1070	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1071		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1072			goto out;
1073		if (td->file_service_left--)
1074			goto out;
1075	}
1076
1077	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1078	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1079		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1080	else
1081		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1082
1083	if (IS_ERR(f))
1084		return f;
1085
1086	td->file_service_file = f;
1087	td->file_service_left = td->file_service_nr - 1;
1088out:
1089	if (f)
1090		dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1091	else
1092		dprint(FD_FILE, "get_next_file: NULL\n");
1093	return f;
1094}
1095
1096static struct fio_file *get_next_file(struct thread_data *td)
1097{
1098	if (td->flags & TD_F_PROFILE_OPS) {
1099		struct prof_io_ops *ops = &td->prof_io_ops;
1100
1101		if (ops->get_next_file)
1102			return ops->get_next_file(td);
1103	}
1104
1105	return __get_next_file(td);
1106}
1107
1108static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1109{
1110	struct fio_file *f;
1111
1112	do {
1113		f = get_next_file(td);
1114		if (IS_ERR_OR_NULL(f))
1115			return PTR_ERR(f);
1116
1117		io_u->file = f;
1118		get_file(f);
1119
1120		if (!fill_io_u(td, io_u))
1121			break;
1122
1123		put_file_log(td, f);
1124		td_io_close_file(td, f);
1125		io_u->file = NULL;
1126		fio_file_set_done(f);
1127		td->nr_done_files++;
1128		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1129					td->nr_done_files, td->o.nr_files);
1130	} while (1);
1131
1132	return 0;
1133}
1134
1135static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1136		      unsigned long tusec, unsigned long max_usec)
1137{
1138	if (!td->error)
1139		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1140	td_verror(td, ETIMEDOUT, "max latency exceeded");
1141	icd->error = ETIMEDOUT;
1142}
1143
1144static void lat_new_cycle(struct thread_data *td)
1145{
1146	fio_gettime(&td->latency_ts, NULL);
1147	td->latency_ios = ddir_rw_sum(td->io_blocks);
1148	td->latency_failed = 0;
1149}
1150
1151/*
1152 * We had an IO outside the latency target. Reduce the queue depth. If we
1153 * are at QD=1, then it's time to give up.
1154 */
1155static int __lat_target_failed(struct thread_data *td)
1156{
1157	if (td->latency_qd == 1)
1158		return 1;
1159
1160	td->latency_qd_high = td->latency_qd;
1161
1162	if (td->latency_qd == td->latency_qd_low)
1163		td->latency_qd_low--;
1164
1165	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1166
1167	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1168
1169	/*
1170	 * When we ramp QD down, quiesce existing IO to prevent
1171	 * a storm of ramp downs due to pending higher depth.
1172	 */
1173	io_u_quiesce(td);
1174	lat_new_cycle(td);
1175	return 0;
1176}
1177
1178static int lat_target_failed(struct thread_data *td)
1179{
1180	if (td->o.latency_percentile.u.f == 100.0)
1181		return __lat_target_failed(td);
1182
1183	td->latency_failed++;
1184	return 0;
1185}
1186
1187void lat_target_init(struct thread_data *td)
1188{
1189	td->latency_end_run = 0;
1190
1191	if (td->o.latency_target) {
1192		dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1193		fio_gettime(&td->latency_ts, NULL);
1194		td->latency_qd = 1;
1195		td->latency_qd_high = td->o.iodepth;
1196		td->latency_qd_low = 1;
1197		td->latency_ios = ddir_rw_sum(td->io_blocks);
1198	} else
1199		td->latency_qd = td->o.iodepth;
1200}
1201
1202void lat_target_reset(struct thread_data *td)
1203{
1204	if (!td->latency_end_run)
1205		lat_target_init(td);
1206}
1207
1208static void lat_target_success(struct thread_data *td)
1209{
1210	const unsigned int qd = td->latency_qd;
1211	struct thread_options *o = &td->o;
1212
1213	td->latency_qd_low = td->latency_qd;
1214
1215	/*
1216	 * If we haven't failed yet, we double up to a failing value instead
1217	 * of bisecting from highest possible queue depth. If we have set
1218	 * a limit other than td->o.iodepth, bisect between that.
1219	 */
1220	if (td->latency_qd_high != o->iodepth)
1221		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1222	else
1223		td->latency_qd *= 2;
1224
1225	if (td->latency_qd > o->iodepth)
1226		td->latency_qd = o->iodepth;
1227
1228	dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1229
1230	/*
1231	 * Same as last one, we are done. Let it run a latency cycle, so
1232	 * we get only the results from the targeted depth.
1233	 */
1234	if (td->latency_qd == qd) {
1235		if (td->latency_end_run) {
1236			dprint(FD_RATE, "We are done\n");
1237			td->done = 1;
1238		} else {
1239			dprint(FD_RATE, "Quiesce and final run\n");
1240			io_u_quiesce(td);
1241			td->latency_end_run = 1;
1242			reset_all_stats(td);
1243			reset_io_stats(td);
1244		}
1245	}
1246
1247	lat_new_cycle(td);
1248}
1249
1250/*
1251 * Check if we can bump the queue depth
1252 */
1253void lat_target_check(struct thread_data *td)
1254{
1255	uint64_t usec_window;
1256	uint64_t ios;
1257	double success_ios;
1258
1259	usec_window = utime_since_now(&td->latency_ts);
1260	if (usec_window < td->o.latency_window)
1261		return;
1262
1263	ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1264	success_ios = (double) (ios - td->latency_failed) / (double) ios;
1265	success_ios *= 100.0;
1266
1267	dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1268
1269	if (success_ios >= td->o.latency_percentile.u.f)
1270		lat_target_success(td);
1271	else
1272		__lat_target_failed(td);
1273}
1274
1275/*
1276 * If latency target is enabled, we might be ramping up or down and not
1277 * using the full queue depth available.
1278 */
1279int queue_full(const struct thread_data *td)
1280{
1281	const int qempty = io_u_qempty(&td->io_u_freelist);
1282
1283	if (qempty)
1284		return 1;
1285	if (!td->o.latency_target)
1286		return 0;
1287
1288	return td->cur_depth >= td->latency_qd;
1289}
1290
1291struct io_u *__get_io_u(struct thread_data *td)
1292{
1293	struct io_u *io_u = NULL;
1294
1295	if (td->stop_io)
1296		return NULL;
1297
1298	td_io_u_lock(td);
1299
1300again:
1301	if (!io_u_rempty(&td->io_u_requeues))
1302		io_u = io_u_rpop(&td->io_u_requeues);
1303	else if (!queue_full(td)) {
1304		io_u = io_u_qpop(&td->io_u_freelist);
1305
1306		io_u->file = NULL;
1307		io_u->buflen = 0;
1308		io_u->resid = 0;
1309		io_u->end_io = NULL;
1310	}
1311
1312	if (io_u) {
1313		assert(io_u->flags & IO_U_F_FREE);
1314		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1315				 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1316				 IO_U_F_VER_LIST);
1317
1318		io_u->error = 0;
1319		io_u->acct_ddir = -1;
1320		td->cur_depth++;
1321		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1322		io_u->ipo = NULL;
1323	} else if (td->o.verify_async) {
1324		/*
1325		 * We ran out, wait for async verify threads to finish and
1326		 * return one
1327		 */
1328		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1329		goto again;
1330	}
1331
1332	td_io_u_unlock(td);
1333	return io_u;
1334}
1335
1336static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1337{
1338	if (!(td->flags & TD_F_TRIM_BACKLOG))
1339		return 0;
1340
1341	if (td->trim_entries) {
1342		int get_trim = 0;
1343
1344		if (td->trim_batch) {
1345			td->trim_batch--;
1346			get_trim = 1;
1347		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1348			 td->last_ddir != DDIR_READ) {
1349			td->trim_batch = td->o.trim_batch;
1350			if (!td->trim_batch)
1351				td->trim_batch = td->o.trim_backlog;
1352			get_trim = 1;
1353		}
1354
1355		if (get_trim && !get_next_trim(td, io_u))
1356			return 1;
1357	}
1358
1359	return 0;
1360}
1361
1362static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1363{
1364	if (!(td->flags & TD_F_VER_BACKLOG))
1365		return 0;
1366
1367	if (td->io_hist_len) {
1368		int get_verify = 0;
1369
1370		if (td->verify_batch)
1371			get_verify = 1;
1372		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1373			 td->last_ddir != DDIR_READ) {
1374			td->verify_batch = td->o.verify_batch;
1375			if (!td->verify_batch)
1376				td->verify_batch = td->o.verify_backlog;
1377			get_verify = 1;
1378		}
1379
1380		if (get_verify && !get_next_verify(td, io_u)) {
1381			td->verify_batch--;
1382			return 1;
1383		}
1384	}
1385
1386	return 0;
1387}
1388
1389/*
1390 * Fill offset and start time into the buffer content, to prevent too
1391 * easy compressible data for simple de-dupe attempts. Do this for every
1392 * 512b block in the range, since that should be the smallest block size
1393 * we can expect from a device.
1394 */
1395static void small_content_scramble(struct io_u *io_u)
1396{
1397	unsigned int i, nr_blocks = io_u->buflen / 512;
1398	uint64_t boffset;
1399	unsigned int offset;
1400	void *p, *end;
1401
1402	if (!nr_blocks)
1403		return;
1404
1405	p = io_u->xfer_buf;
1406	boffset = io_u->offset;
1407	io_u->buf_filled_len = 0;
1408
1409	for (i = 0; i < nr_blocks; i++) {
1410		/*
1411		 * Fill the byte offset into a "random" start offset of
1412		 * the buffer, given by the product of the usec time
1413		 * and the actual offset.
1414		 */
1415		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1416		offset &= ~(sizeof(uint64_t) - 1);
1417		if (offset >= 512 - sizeof(uint64_t))
1418			offset -= sizeof(uint64_t);
1419		memcpy(p + offset, &boffset, sizeof(boffset));
1420
1421		end = p + 512 - sizeof(io_u->start_time);
1422		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1423		p += 512;
1424		boffset += 512;
1425	}
1426}
1427
1428/*
1429 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1430 * etc. The returned io_u is fully ready to be prepped and submitted.
1431 */
1432struct io_u *get_io_u(struct thread_data *td)
1433{
1434	struct fio_file *f;
1435	struct io_u *io_u;
1436	int do_scramble = 0;
1437	long ret = 0;
1438
1439	io_u = __get_io_u(td);
1440	if (!io_u) {
1441		dprint(FD_IO, "__get_io_u failed\n");
1442		return NULL;
1443	}
1444
1445	if (check_get_verify(td, io_u))
1446		goto out;
1447	if (check_get_trim(td, io_u))
1448		goto out;
1449
1450	/*
1451	 * from a requeue, io_u already setup
1452	 */
1453	if (io_u->file)
1454		goto out;
1455
1456	/*
1457	 * If using an iolog, grab next piece if any available.
1458	 */
1459	if (td->flags & TD_F_READ_IOLOG) {
1460		if (read_iolog_get(td, io_u))
1461			goto err_put;
1462	} else if (set_io_u_file(td, io_u)) {
1463		ret = -EBUSY;
1464		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1465		goto err_put;
1466	}
1467
1468	f = io_u->file;
1469	if (!f) {
1470		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1471		goto err_put;
1472	}
1473
1474	assert(fio_file_open(f));
1475
1476	if (ddir_rw(io_u->ddir)) {
1477		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1478			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1479			goto err_put;
1480		}
1481
1482		f->last_start[io_u->ddir] = io_u->offset;
1483		f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1484
1485		if (io_u->ddir == DDIR_WRITE) {
1486			if (td->flags & TD_F_REFILL_BUFFERS) {
1487				io_u_fill_buffer(td, io_u,
1488					td->o.min_bs[DDIR_WRITE],
1489					io_u->xfer_buflen);
1490			} else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1491				   !(td->flags & TD_F_COMPRESS))
1492				do_scramble = 1;
1493			if (td->flags & TD_F_VER_NONE) {
1494				populate_verify_io_u(td, io_u);
1495				do_scramble = 0;
1496			}
1497		} else if (io_u->ddir == DDIR_READ) {
1498			/*
1499			 * Reset the buf_filled parameters so next time if the
1500			 * buffer is used for writes it is refilled.
1501			 */
1502			io_u->buf_filled_len = 0;
1503		}
1504	}
1505
1506	/*
1507	 * Set io data pointers.
1508	 */
1509	io_u->xfer_buf = io_u->buf;
1510	io_u->xfer_buflen = io_u->buflen;
1511
1512out:
1513	assert(io_u->file);
1514	if (!td_io_prep(td, io_u)) {
1515		if (!td->o.disable_slat)
1516			fio_gettime(&io_u->start_time, NULL);
1517		if (do_scramble)
1518			small_content_scramble(io_u);
1519		return io_u;
1520	}
1521err_put:
1522	dprint(FD_IO, "get_io_u failed\n");
1523	put_io_u(td, io_u);
1524	return ERR_PTR(ret);
1525}
1526
1527void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1528{
1529	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1530
1531	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1532		return;
1533
1534	log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1535		io_u->file ? " on file " : "",
1536		io_u->file ? io_u->file->file_name : "",
1537		strerror(io_u->error),
1538		io_ddir_name(io_u->ddir),
1539		io_u->offset, io_u->xfer_buflen);
1540
1541	if (!td->error)
1542		td_verror(td, io_u->error, "io_u error");
1543}
1544
1545static inline int gtod_reduce(struct thread_data *td)
1546{
1547	return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1548		&& td->o.disable_bw;
1549}
1550
1551static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1552				  struct io_completion_data *icd,
1553				  const enum fio_ddir idx, unsigned int bytes)
1554{
1555	unsigned long lusec = 0;
1556
1557	if (!gtod_reduce(td))
1558		lusec = utime_since(&io_u->issue_time, &icd->time);
1559
1560	if (!td->o.disable_lat) {
1561		unsigned long tusec;
1562
1563		tusec = utime_since(&io_u->start_time, &icd->time);
1564		add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1565
1566		if (td->flags & TD_F_PROFILE_OPS) {
1567			struct prof_io_ops *ops = &td->prof_io_ops;
1568
1569			if (ops->io_u_lat)
1570				icd->error = ops->io_u_lat(td, tusec);
1571		}
1572
1573		if (td->o.max_latency && tusec > td->o.max_latency)
1574			lat_fatal(td, icd, tusec, td->o.max_latency);
1575		if (td->o.latency_target && tusec > td->o.latency_target) {
1576			if (lat_target_failed(td))
1577				lat_fatal(td, icd, tusec, td->o.latency_target);
1578		}
1579	}
1580
1581	if (!td->o.disable_clat) {
1582		add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1583		io_u_mark_latency(td, lusec);
1584	}
1585
1586	if (!td->o.disable_bw)
1587		add_bw_sample(td, idx, bytes, &icd->time);
1588
1589	if (!gtod_reduce(td))
1590		add_iops_sample(td, idx, bytes, &icd->time);
1591}
1592
1593static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1594{
1595	uint64_t secs, remainder, bps, bytes;
1596
1597	bytes = td->this_io_bytes[ddir];
1598	bps = td->rate_bps[ddir];
1599	secs = bytes / bps;
1600	remainder = bytes % bps;
1601	return remainder * 1000000 / bps + secs * 1000000;
1602}
1603
1604static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1605			 struct io_completion_data *icd)
1606{
1607	struct io_u *io_u = *io_u_ptr;
1608	enum fio_ddir ddir = io_u->ddir;
1609	struct fio_file *f = io_u->file;
1610
1611	dprint_io_u(io_u, "io complete");
1612
1613	td_io_u_lock(td);
1614	assert(io_u->flags & IO_U_F_FLIGHT);
1615	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1616
1617	/*
1618	 * Mark IO ok to verify
1619	 */
1620	if (io_u->ipo) {
1621		/*
1622		 * Remove errored entry from the verification list
1623		 */
1624		if (io_u->error)
1625			unlog_io_piece(td, io_u);
1626		else {
1627			io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1628			write_barrier();
1629		}
1630	}
1631
1632	td_io_u_unlock(td);
1633
1634	if (ddir_sync(ddir)) {
1635		td->last_was_sync = 1;
1636		if (f) {
1637			f->first_write = -1ULL;
1638			f->last_write = -1ULL;
1639		}
1640		return;
1641	}
1642
1643	td->last_was_sync = 0;
1644	td->last_ddir = ddir;
1645
1646	if (!io_u->error && ddir_rw(ddir)) {
1647		unsigned int bytes = io_u->buflen - io_u->resid;
1648		const enum fio_ddir oddir = ddir ^ 1;
1649		int ret;
1650
1651		td->io_blocks[ddir]++;
1652		td->this_io_blocks[ddir]++;
1653		td->io_bytes[ddir] += bytes;
1654
1655		if (!(io_u->flags & IO_U_F_VER_LIST))
1656			td->this_io_bytes[ddir] += bytes;
1657
1658		if (ddir == DDIR_WRITE) {
1659			if (f) {
1660				if (f->first_write == -1ULL ||
1661				    io_u->offset < f->first_write)
1662					f->first_write = io_u->offset;
1663				if (f->last_write == -1ULL ||
1664				    ((io_u->offset + bytes) > f->last_write))
1665					f->last_write = io_u->offset + bytes;
1666			}
1667			if (td->last_write_comp) {
1668				int idx = td->last_write_idx++;
1669
1670				td->last_write_comp[idx] = io_u->offset;
1671				if (td->last_write_idx == td->o.iodepth)
1672					td->last_write_idx = 0;
1673			}
1674		}
1675
1676		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1677					   td->runstate == TD_VERIFYING)) {
1678			account_io_completion(td, io_u, icd, ddir, bytes);
1679
1680			if (__should_check_rate(td, ddir)) {
1681				td->rate_pending_usleep[ddir] =
1682					(usec_for_io(td, ddir) -
1683					 utime_since_now(&td->start));
1684			}
1685			if (ddir != DDIR_TRIM &&
1686			    __should_check_rate(td, oddir)) {
1687				td->rate_pending_usleep[oddir] =
1688					(usec_for_io(td, oddir) -
1689					 utime_since_now(&td->start));
1690			}
1691		}
1692
1693		icd->bytes_done[ddir] += bytes;
1694
1695		if (io_u->end_io) {
1696			ret = io_u->end_io(td, io_u_ptr);
1697			io_u = *io_u_ptr;
1698			if (ret && !icd->error)
1699				icd->error = ret;
1700		}
1701	} else if (io_u->error) {
1702		icd->error = io_u->error;
1703		io_u_log_error(td, io_u);
1704	}
1705	if (icd->error) {
1706		enum error_type_bit eb = td_error_type(ddir, icd->error);
1707
1708		if (!td_non_fatal_error(td, eb, icd->error))
1709			return;
1710
1711		/*
1712		 * If there is a non_fatal error, then add to the error count
1713		 * and clear all the errors.
1714		 */
1715		update_error_count(td, icd->error);
1716		td_clear_error(td);
1717		icd->error = 0;
1718		if (io_u)
1719			io_u->error = 0;
1720	}
1721}
1722
1723static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1724		     int nr)
1725{
1726	int ddir;
1727
1728	if (!gtod_reduce(td))
1729		fio_gettime(&icd->time, NULL);
1730
1731	icd->nr = nr;
1732
1733	icd->error = 0;
1734	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1735		icd->bytes_done[ddir] = 0;
1736}
1737
1738static void ios_completed(struct thread_data *td,
1739			  struct io_completion_data *icd)
1740{
1741	struct io_u *io_u;
1742	int i;
1743
1744	for (i = 0; i < icd->nr; i++) {
1745		io_u = td->io_ops->event(td, i);
1746
1747		io_completed(td, &io_u, icd);
1748
1749		if (io_u)
1750			put_io_u(td, io_u);
1751	}
1752}
1753
1754/*
1755 * Complete a single io_u for the sync engines.
1756 */
1757int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1758		       uint64_t *bytes)
1759{
1760	struct io_completion_data icd;
1761
1762	init_icd(td, &icd, 1);
1763	io_completed(td, &io_u, &icd);
1764
1765	if (io_u)
1766		put_io_u(td, io_u);
1767
1768	if (icd.error) {
1769		td_verror(td, icd.error, "io_u_sync_complete");
1770		return -1;
1771	}
1772
1773	if (bytes) {
1774		int ddir;
1775
1776		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1777			bytes[ddir] += icd.bytes_done[ddir];
1778	}
1779
1780	return 0;
1781}
1782
1783/*
1784 * Called to complete min_events number of io for the async engines.
1785 */
1786int io_u_queued_complete(struct thread_data *td, int min_evts,
1787			 uint64_t *bytes)
1788{
1789	struct io_completion_data icd;
1790	struct timespec *tvp = NULL;
1791	int ret;
1792	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1793
1794	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1795
1796	if (!min_evts)
1797		tvp = &ts;
1798	else if (min_evts > td->cur_depth)
1799		min_evts = td->cur_depth;
1800
1801	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1802	if (ret < 0) {
1803		td_verror(td, -ret, "td_io_getevents");
1804		return ret;
1805	} else if (!ret)
1806		return ret;
1807
1808	init_icd(td, &icd, ret);
1809	ios_completed(td, &icd);
1810	if (icd.error) {
1811		td_verror(td, icd.error, "io_u_queued_complete");
1812		return -1;
1813	}
1814
1815	if (bytes) {
1816		int ddir;
1817
1818		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1819			bytes[ddir] += icd.bytes_done[ddir];
1820	}
1821
1822	return 0;
1823}
1824
1825/*
1826 * Call when io_u is really queued, to update the submission latency.
1827 */
1828void io_u_queued(struct thread_data *td, struct io_u *io_u)
1829{
1830	if (!td->o.disable_slat) {
1831		unsigned long slat_time;
1832
1833		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1834		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1835				io_u->offset);
1836	}
1837}
1838
1839/*
1840 * See if we should reuse the last seed, if dedupe is enabled
1841 */
1842static struct frand_state *get_buf_state(struct thread_data *td)
1843{
1844	unsigned int v;
1845	unsigned long r;
1846
1847	if (!td->o.dedupe_percentage)
1848		return &td->buf_state;
1849	else if (td->o.dedupe_percentage == 100)
1850		return &td->buf_state_prev;
1851
1852	r = __rand(&td->dedupe_state);
1853	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1854
1855	if (v <= td->o.dedupe_percentage)
1856		return &td->buf_state_prev;
1857
1858	return &td->buf_state;
1859}
1860
1861static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1862{
1863	if (rs == &td->buf_state)
1864		frand_copy(&td->buf_state_prev, rs);
1865}
1866
1867void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1868		    unsigned int max_bs)
1869{
1870	struct thread_options *o = &td->o;
1871
1872	if (o->compress_percentage || o->dedupe_percentage) {
1873		unsigned int perc = td->o.compress_percentage;
1874		struct frand_state *rs;
1875		unsigned int left = max_bs;
1876
1877		do {
1878			rs = get_buf_state(td);
1879
1880			min_write = min(min_write, left);
1881
1882			if (perc) {
1883				unsigned int seg = min_write;
1884
1885				seg = min(min_write, td->o.compress_chunk);
1886				if (!seg)
1887					seg = min_write;
1888
1889				fill_random_buf_percentage(rs, buf, perc, seg,
1890					min_write, o->buffer_pattern,
1891						   o->buffer_pattern_bytes);
1892			} else
1893				fill_random_buf(rs, buf, min_write);
1894
1895			buf += min_write;
1896			left -= min_write;
1897			save_buf_state(td, rs);
1898		} while (left);
1899	} else if (o->buffer_pattern_bytes)
1900		fill_buffer_pattern(td, buf, max_bs);
1901	else
1902		memset(buf, 0, max_bs);
1903}
1904
1905/*
1906 * "randomly" fill the buffer contents
1907 */
1908void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1909		      unsigned int min_write, unsigned int max_bs)
1910{
1911	io_u->buf_filled_len = 0;
1912	fill_io_buffer(td, io_u->buf, min_write, max_bs);
1913}
1914