io_u.c revision ce95d651568bd5494e4c4ddc95832715c6760d56
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
14struct io_completion_data {
15	int nr;				/* input */
16
17	int error;			/* output */
18	unsigned long bytes_done[2];	/* output */
19	struct timeval time;		/* output */
20};
21
22/*
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
25 */
26static int random_map_free(struct fio_file *f, const unsigned long long block)
27{
28	unsigned int idx = RAND_MAP_IDX(f, block);
29	unsigned int bit = RAND_MAP_BIT(f, block);
30
31	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
32
33	return (f->file_map[idx] & (1UL << bit)) == 0;
34}
35
36/*
37 * Mark a given offset as used in the map.
38 */
39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40{
41	unsigned int min_bs = td->o.rw_min_bs;
42	struct fio_file *f = io_u->file;
43	unsigned long long block;
44	unsigned int blocks, nr_blocks;
45	int busy_check;
46
47	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
48	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
49	blocks = 0;
50	busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
51
52	while (nr_blocks) {
53		unsigned int idx, bit;
54		unsigned long mask, this_blocks;
55
56		/*
57		 * If we have a mixed random workload, we may
58		 * encounter blocks we already did IO to.
59		 */
60		if (!busy_check) {
61			blocks = nr_blocks;
62			break;
63		}
64		if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
65			break;
66
67		idx = RAND_MAP_IDX(f, block);
68		bit = RAND_MAP_BIT(f, block);
69
70		fio_assert(td, idx < f->num_maps);
71
72		this_blocks = nr_blocks;
73		if (this_blocks + bit > BLOCKS_PER_MAP)
74			this_blocks = BLOCKS_PER_MAP - bit;
75
76		do {
77			if (this_blocks == BLOCKS_PER_MAP)
78				mask = -1UL;
79			else
80				mask = ((1UL << this_blocks) - 1) << bit;
81
82			if (!(f->file_map[idx] & mask))
83				break;
84
85			this_blocks--;
86		} while (this_blocks);
87
88		if (!this_blocks)
89			break;
90
91		f->file_map[idx] |= mask;
92		nr_blocks -= this_blocks;
93		blocks += this_blocks;
94		block += this_blocks;
95	}
96
97	if ((blocks * min_bs) < io_u->buflen)
98		io_u->buflen = blocks * min_bs;
99}
100
101static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
102				     enum fio_ddir ddir)
103{
104	unsigned long long max_blocks;
105	unsigned long long max_size;
106
107	assert(ddir_rw(ddir));
108
109	/*
110	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
111	 */
112	max_size = f->io_size;
113	if (max_size > f->real_file_size)
114		max_size = f->real_file_size;
115
116	if (td->o.zone_range)
117		max_size = td->o.zone_range;
118
119	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
120	if (!max_blocks)
121		return 0;
122
123	return max_blocks;
124}
125
126/*
127 * Return the next free block in the map.
128 */
129static int get_next_free_block(struct thread_data *td, struct fio_file *f,
130			       enum fio_ddir ddir, unsigned long long *b)
131{
132	unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
133	int i;
134
135	lastb = last_block(td, f, ddir);
136	if (!lastb)
137		return 1;
138
139	i = f->last_free_lookup;
140	block = i * BLOCKS_PER_MAP;
141	while (block * min_bs < f->real_file_size &&
142		block * min_bs < f->io_size) {
143		if (f->file_map[i] != -1UL) {
144			block += ffz(f->file_map[i]);
145			if (block > lastb)
146				break;
147			f->last_free_lookup = i;
148			*b = block;
149			return 0;
150		}
151
152		block += BLOCKS_PER_MAP;
153		i++;
154	}
155
156	dprint(FD_IO, "failed finding a free block\n");
157	return 1;
158}
159
160static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
161				enum fio_ddir ddir, unsigned long long *b)
162{
163	unsigned long long rmax, r, lastb;
164	int loops = 5;
165
166	lastb = last_block(td, f, ddir);
167	if (!lastb)
168		return 1;
169
170	if (f->failed_rands >= 200)
171		goto ffz;
172
173	rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
174	do {
175		if (td->o.use_os_rand)
176			r = os_random_long(&td->random_state);
177		else
178			r = __rand(&td->__random_state);
179
180		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
181
182		dprint(FD_RANDOM, "off rand %llu\n", r);
183
184
185		/*
186		 * if we are not maintaining a random map, we are done.
187		 */
188		if (!file_randommap(td, f))
189			goto ret_good;
190
191		/*
192		 * calculate map offset and check if it's free
193		 */
194		if (random_map_free(f, *b))
195			goto ret_good;
196
197		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
198									*b);
199	} while (--loops);
200
201	if (!f->failed_rands++)
202		f->last_free_lookup = 0;
203
204	/*
205	 * we get here, if we didn't suceed in looking up a block. generate
206	 * a random start offset into the filemap, and find the first free
207	 * block from there.
208	 */
209	loops = 10;
210	do {
211		f->last_free_lookup = (f->num_maps - 1) *
212					(r / ((unsigned long long) rmax + 1.0));
213		if (!get_next_free_block(td, f, ddir, b))
214			goto ret;
215
216		if (td->o.use_os_rand)
217			r = os_random_long(&td->random_state);
218		else
219			r = __rand(&td->__random_state);
220	} while (--loops);
221
222	/*
223	 * that didn't work either, try exhaustive search from the start
224	 */
225	f->last_free_lookup = 0;
226ffz:
227	if (!get_next_free_block(td, f, ddir, b))
228		return 0;
229	f->last_free_lookup = 0;
230	return get_next_free_block(td, f, ddir, b);
231ret_good:
232	f->failed_rands = 0;
233ret:
234	return 0;
235}
236
237static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238			       enum fio_ddir ddir, unsigned long long *b)
239{
240	if (!get_next_rand_offset(td, f, ddir, b))
241		return 0;
242
243	if (td->o.time_based) {
244		fio_file_reset(f);
245		if (!get_next_rand_offset(td, f, ddir, b))
246			return 0;
247	}
248
249	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250			f->file_name, f->last_pos, f->real_file_size);
251	return 1;
252}
253
254static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
255			       enum fio_ddir ddir, unsigned long long *offset)
256{
257	assert(ddir_rw(ddir));
258
259	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
260		f->last_pos = f->last_pos - f->io_size;
261
262	if (f->last_pos < f->real_file_size) {
263		unsigned long long pos;
264
265		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
266			f->last_pos = f->real_file_size;
267
268		pos = f->last_pos - f->file_offset;
269		if (pos)
270			pos += td->o.ddir_seq_add;
271
272		*offset = pos;
273		return 0;
274	}
275
276	return 1;
277}
278
279static int get_next_block(struct thread_data *td, struct io_u *io_u,
280			  enum fio_ddir ddir, int rw_seq)
281{
282	struct fio_file *f = io_u->file;
283	unsigned long long b, offset;
284	int ret;
285
286	assert(ddir_rw(ddir));
287
288	b = offset = -1ULL;
289
290	if (rw_seq) {
291		if (td_random(td))
292			ret = get_next_rand_block(td, f, ddir, &b);
293		else
294			ret = get_next_seq_offset(td, f, ddir, &offset);
295	} else {
296		io_u->flags |= IO_U_F_BUSY_OK;
297
298		if (td->o.rw_seq == RW_SEQ_SEQ) {
299			ret = get_next_seq_offset(td, f, ddir, &offset);
300			if (ret)
301				ret = get_next_rand_block(td, f, ddir, &b);
302		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
303			if (f->last_start != -1ULL)
304				offset = f->last_start - f->file_offset;
305			else
306				offset = 0;
307			ret = 0;
308		} else {
309			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
310			ret = 1;
311		}
312	}
313
314	if (!ret) {
315		if (offset != -1ULL)
316			io_u->offset = offset;
317		else if (b != -1ULL)
318			io_u->offset = b * td->o.ba[ddir];
319		else {
320			log_err("fio: bug in offset generation\n");
321			ret = 1;
322		}
323	}
324
325	return ret;
326}
327
328/*
329 * For random io, generate a random new block and see if it's used. Repeat
330 * until we find a free one. For sequential io, just return the end of
331 * the last io issued.
332 */
333static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
334{
335	struct fio_file *f = io_u->file;
336	enum fio_ddir ddir = io_u->ddir;
337	int rw_seq_hit = 0;
338
339	assert(ddir_rw(ddir));
340
341	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
342		rw_seq_hit = 1;
343		td->ddir_seq_nr = td->o.ddir_seq_nr;
344	}
345
346	if (get_next_block(td, io_u, ddir, rw_seq_hit))
347		return 1;
348
349	if (io_u->offset >= f->io_size) {
350		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
351					io_u->offset, f->io_size);
352		return 1;
353	}
354
355	io_u->offset += f->file_offset;
356	if (io_u->offset >= f->real_file_size) {
357		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
358					io_u->offset, f->real_file_size);
359		return 1;
360	}
361
362	return 0;
363}
364
365static int get_next_offset(struct thread_data *td, struct io_u *io_u)
366{
367	struct prof_io_ops *ops = &td->prof_io_ops;
368
369	if (ops->fill_io_u_off)
370		return ops->fill_io_u_off(td, io_u);
371
372	return __get_next_offset(td, io_u);
373}
374
375static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
376			    unsigned int buflen)
377{
378	struct fio_file *f = io_u->file;
379
380	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
381}
382
383static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
384{
385	const int ddir = io_u->ddir;
386	unsigned int uninitialized_var(buflen);
387	unsigned int minbs, maxbs;
388	unsigned long r, rand_max;
389
390	assert(ddir_rw(ddir));
391
392	minbs = td->o.min_bs[ddir];
393	maxbs = td->o.max_bs[ddir];
394
395	if (minbs == maxbs)
396		return minbs;
397
398	/*
399	 * If we can't satisfy the min block size from here, then fail
400	 */
401	if (!io_u_fits(td, io_u, minbs))
402		return 0;
403
404	if (td->o.use_os_rand)
405		rand_max = OS_RAND_MAX;
406	else
407		rand_max = FRAND_MAX;
408
409	do {
410		if (td->o.use_os_rand)
411			r = os_random_long(&td->bsrange_state);
412		else
413			r = __rand(&td->__bsrange_state);
414
415		if (!td->o.bssplit_nr[ddir]) {
416			buflen = 1 + (unsigned int) ((double) maxbs *
417					(r / (rand_max + 1.0)));
418			if (buflen < minbs)
419				buflen = minbs;
420		} else {
421			long perc = 0;
422			unsigned int i;
423
424			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
425				struct bssplit *bsp = &td->o.bssplit[ddir][i];
426
427				buflen = bsp->bs;
428				perc += bsp->perc;
429				if ((r <= ((rand_max / 100L) * perc)) &&
430				    io_u_fits(td, io_u, buflen))
431					break;
432			}
433		}
434
435		if (!td->o.bs_unaligned && is_power_of_2(minbs))
436			buflen = (buflen + minbs - 1) & ~(minbs - 1);
437
438	} while (!io_u_fits(td, io_u, buflen));
439
440	return buflen;
441}
442
443static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
444{
445	struct prof_io_ops *ops = &td->prof_io_ops;
446
447	if (ops->fill_io_u_size)
448		return ops->fill_io_u_size(td, io_u);
449
450	return __get_next_buflen(td, io_u);
451}
452
453static void set_rwmix_bytes(struct thread_data *td)
454{
455	unsigned int diff;
456
457	/*
458	 * we do time or byte based switch. this is needed because
459	 * buffered writes may issue a lot quicker than they complete,
460	 * whereas reads do not.
461	 */
462	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
463	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
464}
465
466static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
467{
468	unsigned int v;
469	unsigned long r;
470
471	if (td->o.use_os_rand) {
472		r = os_random_long(&td->rwmix_state);
473		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
474	} else {
475		r = __rand(&td->__rwmix_state);
476		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
477	}
478
479	if (v <= td->o.rwmix[DDIR_READ])
480		return DDIR_READ;
481
482	return DDIR_WRITE;
483}
484
485static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
486{
487	enum fio_ddir odir = ddir ^ 1;
488	struct timeval t;
489	long usec;
490
491	assert(ddir_rw(ddir));
492
493	if (td->rate_pending_usleep[ddir] <= 0)
494		return ddir;
495
496	/*
497	 * We have too much pending sleep in this direction. See if we
498	 * should switch.
499	 */
500	if (td_rw(td)) {
501		/*
502		 * Other direction does not have too much pending, switch
503		 */
504		if (td->rate_pending_usleep[odir] < 100000)
505			return odir;
506
507		/*
508		 * Both directions have pending sleep. Sleep the minimum time
509		 * and deduct from both.
510		 */
511		if (td->rate_pending_usleep[ddir] <=
512			td->rate_pending_usleep[odir]) {
513			usec = td->rate_pending_usleep[ddir];
514		} else {
515			usec = td->rate_pending_usleep[odir];
516			ddir = odir;
517		}
518	} else
519		usec = td->rate_pending_usleep[ddir];
520
521	/*
522	 * We are going to sleep, ensure that we flush anything pending as
523	 * not to skew our latency numbers.
524	 *
525	 * Changed to only monitor 'in flight' requests here instead of the
526	 * td->cur_depth, b/c td->cur_depth does not accurately represent
527	 * io's that have been actually submitted to an async engine,
528	 * and cur_depth is meaningless for sync engines.
529	 */
530	if (td->io_u_in_flight) {
531		int fio_unused ret;
532
533		ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
534	}
535
536	fio_gettime(&t, NULL);
537	usec_sleep(td, usec);
538	usec = utime_since_now(&t);
539
540	td->rate_pending_usleep[ddir] -= usec;
541
542	odir = ddir ^ 1;
543	if (td_rw(td) && __should_check_rate(td, odir))
544		td->rate_pending_usleep[odir] -= usec;
545
546	return ddir;
547}
548
549/*
550 * Return the data direction for the next io_u. If the job is a
551 * mixed read/write workload, check the rwmix cycle and switch if
552 * necessary.
553 */
554static enum fio_ddir get_rw_ddir(struct thread_data *td)
555{
556	enum fio_ddir ddir;
557
558	/*
559	 * see if it's time to fsync
560	 */
561	if (td->o.fsync_blocks &&
562	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
563	     td->io_issues[DDIR_WRITE] && should_fsync(td))
564		return DDIR_SYNC;
565
566	/*
567	 * see if it's time to fdatasync
568	 */
569	if (td->o.fdatasync_blocks &&
570	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
571	     td->io_issues[DDIR_WRITE] && should_fsync(td))
572		return DDIR_DATASYNC;
573
574	/*
575	 * see if it's time to sync_file_range
576	 */
577	if (td->sync_file_range_nr &&
578	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
579	     td->io_issues[DDIR_WRITE] && should_fsync(td))
580		return DDIR_SYNC_FILE_RANGE;
581
582	if (td_rw(td)) {
583		/*
584		 * Check if it's time to seed a new data direction.
585		 */
586		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
587			/*
588			 * Put a top limit on how many bytes we do for
589			 * one data direction, to avoid overflowing the
590			 * ranges too much
591			 */
592			ddir = get_rand_ddir(td);
593
594			if (ddir != td->rwmix_ddir)
595				set_rwmix_bytes(td);
596
597			td->rwmix_ddir = ddir;
598		}
599		ddir = td->rwmix_ddir;
600	} else if (td_read(td))
601		ddir = DDIR_READ;
602	else
603		ddir = DDIR_WRITE;
604
605	td->rwmix_ddir = rate_ddir(td, ddir);
606	return td->rwmix_ddir;
607}
608
609static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
610{
611	io_u->ddir = get_rw_ddir(td);
612
613	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
614	    td->o.barrier_blocks &&
615	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
616	     td->io_issues[DDIR_WRITE])
617		io_u->flags |= IO_U_F_BARRIER;
618}
619
620void put_file_log(struct thread_data *td, struct fio_file *f)
621{
622	int ret = put_file(td, f);
623
624	if (ret)
625		td_verror(td, ret, "file close");
626}
627
628void put_io_u(struct thread_data *td, struct io_u *io_u)
629{
630	td_io_u_lock(td);
631
632	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
633		put_file_log(td, io_u->file);
634	io_u->file = NULL;
635	io_u->flags &= ~IO_U_F_FREE_DEF;
636	io_u->flags |= IO_U_F_FREE;
637
638	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
639		td->cur_depth--;
640	flist_del_init(&io_u->list);
641	flist_add(&io_u->list, &td->io_u_freelist);
642	td_io_u_unlock(td);
643	td_io_u_free_notify(td);
644}
645
646void clear_io_u(struct thread_data *td, struct io_u *io_u)
647{
648	io_u->flags &= ~IO_U_F_FLIGHT;
649	put_io_u(td, io_u);
650}
651
652void requeue_io_u(struct thread_data *td, struct io_u **io_u)
653{
654	struct io_u *__io_u = *io_u;
655
656	dprint(FD_IO, "requeue %p\n", __io_u);
657
658	td_io_u_lock(td);
659
660	__io_u->flags |= IO_U_F_FREE;
661	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
662		td->io_issues[__io_u->ddir]--;
663
664	__io_u->flags &= ~IO_U_F_FLIGHT;
665	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
666		td->cur_depth--;
667	flist_del(&__io_u->list);
668	flist_add_tail(&__io_u->list, &td->io_u_requeues);
669	td_io_u_unlock(td);
670	*io_u = NULL;
671}
672
673static int fill_io_u(struct thread_data *td, struct io_u *io_u)
674{
675	if (td->io_ops->flags & FIO_NOIO)
676		goto out;
677
678	set_rw_ddir(td, io_u);
679
680	/*
681	 * fsync() or fdatasync() or trim etc, we are done
682	 */
683	if (!ddir_rw(io_u->ddir))
684		goto out;
685
686	/*
687	 * See if it's time to switch to a new zone
688	 */
689	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
690		td->zone_bytes = 0;
691		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
692		io_u->file->last_pos = io_u->file->file_offset;
693		td->io_skip_bytes += td->o.zone_skip;
694	}
695
696	/*
697	 * No log, let the seq/rand engine retrieve the next buflen and
698	 * position.
699	 */
700	if (get_next_offset(td, io_u)) {
701		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
702		return 1;
703	}
704
705	io_u->buflen = get_next_buflen(td, io_u);
706	if (!io_u->buflen) {
707		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
708		return 1;
709	}
710
711	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
712		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
713		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
714				io_u->buflen, io_u->file->real_file_size);
715		return 1;
716	}
717
718	/*
719	 * mark entry before potentially trimming io_u
720	 */
721	if (td_random(td) && file_randommap(td, io_u->file))
722		mark_random_map(td, io_u);
723
724	/*
725	 * If using a write iolog, store this entry.
726	 */
727out:
728	dprint_io_u(io_u, "fill_io_u");
729	td->zone_bytes += io_u->buflen;
730	log_io_u(td, io_u);
731	return 0;
732}
733
734static void __io_u_mark_map(unsigned int *map, unsigned int nr)
735{
736	int idx = 0;
737
738	switch (nr) {
739	default:
740		idx = 6;
741		break;
742	case 33 ... 64:
743		idx = 5;
744		break;
745	case 17 ... 32:
746		idx = 4;
747		break;
748	case 9 ... 16:
749		idx = 3;
750		break;
751	case 5 ... 8:
752		idx = 2;
753		break;
754	case 1 ... 4:
755		idx = 1;
756	case 0:
757		break;
758	}
759
760	map[idx]++;
761}
762
763void io_u_mark_submit(struct thread_data *td, unsigned int nr)
764{
765	__io_u_mark_map(td->ts.io_u_submit, nr);
766	td->ts.total_submit++;
767}
768
769void io_u_mark_complete(struct thread_data *td, unsigned int nr)
770{
771	__io_u_mark_map(td->ts.io_u_complete, nr);
772	td->ts.total_complete++;
773}
774
775void io_u_mark_depth(struct thread_data *td, unsigned int nr)
776{
777	int idx = 0;
778
779	switch (td->cur_depth) {
780	default:
781		idx = 6;
782		break;
783	case 32 ... 63:
784		idx = 5;
785		break;
786	case 16 ... 31:
787		idx = 4;
788		break;
789	case 8 ... 15:
790		idx = 3;
791		break;
792	case 4 ... 7:
793		idx = 2;
794		break;
795	case 2 ... 3:
796		idx = 1;
797	case 1:
798		break;
799	}
800
801	td->ts.io_u_map[idx] += nr;
802}
803
804static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
805{
806	int idx = 0;
807
808	assert(usec < 1000);
809
810	switch (usec) {
811	case 750 ... 999:
812		idx = 9;
813		break;
814	case 500 ... 749:
815		idx = 8;
816		break;
817	case 250 ... 499:
818		idx = 7;
819		break;
820	case 100 ... 249:
821		idx = 6;
822		break;
823	case 50 ... 99:
824		idx = 5;
825		break;
826	case 20 ... 49:
827		idx = 4;
828		break;
829	case 10 ... 19:
830		idx = 3;
831		break;
832	case 4 ... 9:
833		idx = 2;
834		break;
835	case 2 ... 3:
836		idx = 1;
837	case 0 ... 1:
838		break;
839	}
840
841	assert(idx < FIO_IO_U_LAT_U_NR);
842	td->ts.io_u_lat_u[idx]++;
843}
844
845static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
846{
847	int idx = 0;
848
849	switch (msec) {
850	default:
851		idx = 11;
852		break;
853	case 1000 ... 1999:
854		idx = 10;
855		break;
856	case 750 ... 999:
857		idx = 9;
858		break;
859	case 500 ... 749:
860		idx = 8;
861		break;
862	case 250 ... 499:
863		idx = 7;
864		break;
865	case 100 ... 249:
866		idx = 6;
867		break;
868	case 50 ... 99:
869		idx = 5;
870		break;
871	case 20 ... 49:
872		idx = 4;
873		break;
874	case 10 ... 19:
875		idx = 3;
876		break;
877	case 4 ... 9:
878		idx = 2;
879		break;
880	case 2 ... 3:
881		idx = 1;
882	case 0 ... 1:
883		break;
884	}
885
886	assert(idx < FIO_IO_U_LAT_M_NR);
887	td->ts.io_u_lat_m[idx]++;
888}
889
890static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
891{
892	if (usec < 1000)
893		io_u_mark_lat_usec(td, usec);
894	else
895		io_u_mark_lat_msec(td, usec / 1000);
896}
897
898/*
899 * Get next file to service by choosing one at random
900 */
901static struct fio_file *get_next_file_rand(struct thread_data *td,
902					   enum fio_file_flags goodf,
903					   enum fio_file_flags badf)
904{
905	struct fio_file *f;
906	int fno;
907
908	do {
909		int opened = 0;
910		unsigned long r;
911
912		if (td->o.use_os_rand) {
913			r = os_random_long(&td->next_file_state);
914			fno = (unsigned int) ((double) td->o.nr_files
915				* (r / (OS_RAND_MAX + 1.0)));
916		} else {
917			r = __rand(&td->__next_file_state);
918			fno = (unsigned int) ((double) td->o.nr_files
919				* (r / (FRAND_MAX + 1.0)));
920		}
921
922		f = td->files[fno];
923		if (fio_file_done(f))
924			continue;
925
926		if (!fio_file_open(f)) {
927			int err;
928
929			err = td_io_open_file(td, f);
930			if (err)
931				continue;
932			opened = 1;
933		}
934
935		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
936			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
937			return f;
938		}
939		if (opened)
940			td_io_close_file(td, f);
941	} while (1);
942}
943
944/*
945 * Get next file to service by doing round robin between all available ones
946 */
947static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
948					 int badf)
949{
950	unsigned int old_next_file = td->next_file;
951	struct fio_file *f;
952
953	do {
954		int opened = 0;
955
956		f = td->files[td->next_file];
957
958		td->next_file++;
959		if (td->next_file >= td->o.nr_files)
960			td->next_file = 0;
961
962		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
963		if (fio_file_done(f)) {
964			f = NULL;
965			continue;
966		}
967
968		if (!fio_file_open(f)) {
969			int err;
970
971			err = td_io_open_file(td, f);
972			if (err) {
973				dprint(FD_FILE, "error %d on open of %s\n",
974					err, f->file_name);
975				f = NULL;
976				continue;
977			}
978			opened = 1;
979		}
980
981		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
982								f->flags);
983		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
984			break;
985
986		if (opened)
987			td_io_close_file(td, f);
988
989		f = NULL;
990	} while (td->next_file != old_next_file);
991
992	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
993	return f;
994}
995
996static struct fio_file *__get_next_file(struct thread_data *td)
997{
998	struct fio_file *f;
999
1000	assert(td->o.nr_files <= td->files_index);
1001
1002	if (td->nr_done_files >= td->o.nr_files) {
1003		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1004				" nr_files=%d\n", td->nr_open_files,
1005						  td->nr_done_files,
1006						  td->o.nr_files);
1007		return NULL;
1008	}
1009
1010	f = td->file_service_file;
1011	if (f && fio_file_open(f) && !fio_file_closing(f)) {
1012		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1013			goto out;
1014		if (td->file_service_left--)
1015			goto out;
1016	}
1017
1018	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1019	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1020		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1021	else
1022		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1023
1024	td->file_service_file = f;
1025	td->file_service_left = td->file_service_nr - 1;
1026out:
1027	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1028	return f;
1029}
1030
1031static struct fio_file *get_next_file(struct thread_data *td)
1032{
1033	struct prof_io_ops *ops = &td->prof_io_ops;
1034
1035	if (ops->get_next_file)
1036		return ops->get_next_file(td);
1037
1038	return __get_next_file(td);
1039}
1040
1041static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1042{
1043	struct fio_file *f;
1044
1045	do {
1046		f = get_next_file(td);
1047		if (!f)
1048			return 1;
1049
1050		io_u->file = f;
1051		get_file(f);
1052
1053		if (!fill_io_u(td, io_u))
1054			break;
1055
1056		put_file_log(td, f);
1057		td_io_close_file(td, f);
1058		io_u->file = NULL;
1059		fio_file_set_done(f);
1060		td->nr_done_files++;
1061		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1062					td->nr_done_files, td->o.nr_files);
1063	} while (1);
1064
1065	return 0;
1066}
1067
1068
1069struct io_u *__get_io_u(struct thread_data *td)
1070{
1071	struct io_u *io_u = NULL;
1072
1073	td_io_u_lock(td);
1074
1075again:
1076	if (!flist_empty(&td->io_u_requeues))
1077		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1078	else if (!queue_full(td)) {
1079		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1080
1081		io_u->buflen = 0;
1082		io_u->resid = 0;
1083		io_u->file = NULL;
1084		io_u->end_io = NULL;
1085	}
1086
1087	if (io_u) {
1088		assert(io_u->flags & IO_U_F_FREE);
1089		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1090		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1091		io_u->flags &= ~IO_U_F_VER_LIST;
1092
1093		io_u->error = 0;
1094		flist_del(&io_u->list);
1095		flist_add(&io_u->list, &td->io_u_busylist);
1096		td->cur_depth++;
1097		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1098	} else if (td->o.verify_async) {
1099		/*
1100		 * We ran out, wait for async verify threads to finish and
1101		 * return one
1102		 */
1103		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1104		goto again;
1105	}
1106
1107	td_io_u_unlock(td);
1108	return io_u;
1109}
1110
1111static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1112{
1113	if (td->o.trim_backlog && td->trim_entries) {
1114		int get_trim = 0;
1115
1116		if (td->trim_batch) {
1117			td->trim_batch--;
1118			get_trim = 1;
1119		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1120			 td->last_ddir != DDIR_READ) {
1121			td->trim_batch = td->o.trim_batch;
1122			if (!td->trim_batch)
1123				td->trim_batch = td->o.trim_backlog;
1124			get_trim = 1;
1125		}
1126
1127		if (get_trim && !get_next_trim(td, io_u))
1128			return 1;
1129	}
1130
1131	return 0;
1132}
1133
1134static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1135{
1136	if (td->o.verify_backlog && td->io_hist_len) {
1137		int get_verify = 0;
1138
1139		if (td->verify_batch)
1140			get_verify = 1;
1141		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1142			 td->last_ddir != DDIR_READ) {
1143			td->verify_batch = td->o.verify_batch;
1144			if (!td->verify_batch)
1145				td->verify_batch = td->o.verify_backlog;
1146			get_verify = 1;
1147		}
1148
1149		if (get_verify && !get_next_verify(td, io_u)) {
1150			td->verify_batch--;
1151			return 1;
1152		}
1153	}
1154
1155	return 0;
1156}
1157
1158/*
1159 * Fill offset and start time into the buffer content, to prevent too
1160 * easy compressible data for simple de-dupe attempts. Do this for every
1161 * 512b block in the range, since that should be the smallest block size
1162 * we can expect from a device.
1163 */
1164static void small_content_scramble(struct io_u *io_u)
1165{
1166	unsigned int i, nr_blocks = io_u->buflen / 512;
1167	unsigned long long boffset;
1168	unsigned int offset;
1169	void *p, *end;
1170
1171	if (!nr_blocks)
1172		return;
1173
1174	p = io_u->xfer_buf;
1175	boffset = io_u->offset;
1176	io_u->buf_filled_len = 0;
1177
1178	for (i = 0; i < nr_blocks; i++) {
1179		/*
1180		 * Fill the byte offset into a "random" start offset of
1181		 * the buffer, given by the product of the usec time
1182		 * and the actual offset.
1183		 */
1184		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1185		offset &= ~(sizeof(unsigned long long) - 1);
1186		if (offset >= 512 - sizeof(unsigned long long))
1187			offset -= sizeof(unsigned long long);
1188		memcpy(p + offset, &boffset, sizeof(boffset));
1189
1190		end = p + 512 - sizeof(io_u->start_time);
1191		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1192		p += 512;
1193		boffset += 512;
1194	}
1195}
1196
1197/*
1198 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1199 * etc. The returned io_u is fully ready to be prepped and submitted.
1200 */
1201struct io_u *get_io_u(struct thread_data *td)
1202{
1203	struct fio_file *f;
1204	struct io_u *io_u;
1205	int do_scramble = 0;
1206
1207	io_u = __get_io_u(td);
1208	if (!io_u) {
1209		dprint(FD_IO, "__get_io_u failed\n");
1210		return NULL;
1211	}
1212
1213	if (check_get_verify(td, io_u))
1214		goto out;
1215	if (check_get_trim(td, io_u))
1216		goto out;
1217
1218	/*
1219	 * from a requeue, io_u already setup
1220	 */
1221	if (io_u->file)
1222		goto out;
1223
1224	/*
1225	 * If using an iolog, grab next piece if any available.
1226	 */
1227	if (td->o.read_iolog_file) {
1228		if (read_iolog_get(td, io_u))
1229			goto err_put;
1230	} else if (set_io_u_file(td, io_u)) {
1231		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1232		goto err_put;
1233	}
1234
1235	f = io_u->file;
1236	assert(fio_file_open(f));
1237
1238	if (ddir_rw(io_u->ddir)) {
1239		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1240			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1241			goto err_put;
1242		}
1243
1244		f->last_start = io_u->offset;
1245		f->last_pos = io_u->offset + io_u->buflen;
1246
1247		if (io_u->ddir == DDIR_WRITE) {
1248			if (td->o.refill_buffers) {
1249				io_u_fill_buffer(td, io_u,
1250					io_u->xfer_buflen, io_u->xfer_buflen);
1251			} else if (td->o.scramble_buffers)
1252				do_scramble = 1;
1253			if (td->o.verify != VERIFY_NONE) {
1254				populate_verify_io_u(td, io_u);
1255				do_scramble = 0;
1256			}
1257		} else if (io_u->ddir == DDIR_READ) {
1258			/*
1259			 * Reset the buf_filled parameters so next time if the
1260			 * buffer is used for writes it is refilled.
1261			 */
1262			io_u->buf_filled_len = 0;
1263		}
1264	}
1265
1266	/*
1267	 * Set io data pointers.
1268	 */
1269	io_u->xfer_buf = io_u->buf;
1270	io_u->xfer_buflen = io_u->buflen;
1271
1272out:
1273	assert(io_u->file);
1274	if (!td_io_prep(td, io_u)) {
1275		if (!td->o.disable_slat)
1276			fio_gettime(&io_u->start_time, NULL);
1277		if (do_scramble)
1278			small_content_scramble(io_u);
1279		return io_u;
1280	}
1281err_put:
1282	dprint(FD_IO, "get_io_u failed\n");
1283	put_io_u(td, io_u);
1284	return NULL;
1285}
1286
1287void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1288{
1289	const char *msg[] = { "read", "write", "sync", "datasync",
1290				"sync_file_range", "wait", "trim" };
1291
1292
1293
1294	log_err("fio: io_u error");
1295
1296	if (io_u->file)
1297		log_err(" on file %s", io_u->file->file_name);
1298
1299	log_err(": %s\n", strerror(io_u->error));
1300
1301	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1302					io_u->offset, io_u->xfer_buflen);
1303
1304	if (!td->error)
1305		td_verror(td, io_u->error, "io_u error");
1306}
1307
1308static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1309				  struct io_completion_data *icd,
1310				  const enum fio_ddir idx, unsigned int bytes)
1311{
1312	unsigned long uninitialized_var(lusec);
1313
1314	if (!td->o.disable_clat || !td->o.disable_bw)
1315		lusec = utime_since(&io_u->issue_time, &icd->time);
1316
1317	if (!td->o.disable_lat) {
1318		unsigned long tusec;
1319
1320		tusec = utime_since(&io_u->start_time, &icd->time);
1321		add_lat_sample(td, idx, tusec, bytes);
1322	}
1323
1324	if (!td->o.disable_clat) {
1325		add_clat_sample(td, idx, lusec, bytes);
1326		io_u_mark_latency(td, lusec);
1327	}
1328
1329	if (!td->o.disable_bw)
1330		add_bw_sample(td, idx, bytes, &icd->time);
1331
1332	add_iops_sample(td, idx, &icd->time);
1333}
1334
1335static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1336{
1337	unsigned long long secs, remainder, bps, bytes;
1338	bytes = td->this_io_bytes[ddir];
1339	bps = td->rate_bps[ddir];
1340	secs = bytes / bps;
1341	remainder = bytes % bps;
1342	return remainder * 1000000 / bps + secs * 1000000;
1343}
1344
1345static void io_completed(struct thread_data *td, struct io_u *io_u,
1346			 struct io_completion_data *icd)
1347{
1348	/*
1349	 * Older gcc's are too dumb to realize that usec is always used
1350	 * initialized, silence that warning.
1351	 */
1352	unsigned long uninitialized_var(usec);
1353	struct fio_file *f;
1354
1355	dprint_io_u(io_u, "io complete");
1356
1357	td_io_u_lock(td);
1358	assert(io_u->flags & IO_U_F_FLIGHT);
1359	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1360	td_io_u_unlock(td);
1361
1362	if (ddir_sync(io_u->ddir)) {
1363		td->last_was_sync = 1;
1364		f = io_u->file;
1365		if (f) {
1366			f->first_write = -1ULL;
1367			f->last_write = -1ULL;
1368		}
1369		return;
1370	}
1371
1372	td->last_was_sync = 0;
1373	td->last_ddir = io_u->ddir;
1374
1375	if (!io_u->error && ddir_rw(io_u->ddir)) {
1376		unsigned int bytes = io_u->buflen - io_u->resid;
1377		const enum fio_ddir idx = io_u->ddir;
1378		const enum fio_ddir odx = io_u->ddir ^ 1;
1379		int ret;
1380
1381		td->io_blocks[idx]++;
1382		td->this_io_blocks[idx]++;
1383		td->io_bytes[idx] += bytes;
1384
1385		if (!(io_u->flags & IO_U_F_VER_LIST))
1386			td->this_io_bytes[idx] += bytes;
1387
1388		if (idx == DDIR_WRITE) {
1389			f = io_u->file;
1390			if (f) {
1391				if (f->first_write == -1ULL ||
1392				    io_u->offset < f->first_write)
1393					f->first_write = io_u->offset;
1394				if (f->last_write == -1ULL ||
1395				    ((io_u->offset + bytes) > f->last_write))
1396					f->last_write = io_u->offset + bytes;
1397			}
1398		}
1399
1400		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1401					   td->runstate == TD_VERIFYING)) {
1402			account_io_completion(td, io_u, icd, idx, bytes);
1403
1404			if (__should_check_rate(td, idx)) {
1405				td->rate_pending_usleep[idx] =
1406					(usec_for_io(td, idx) -
1407					 utime_since_now(&td->start));
1408			}
1409			if (__should_check_rate(td, odx))
1410				td->rate_pending_usleep[odx] =
1411					(usec_for_io(td, odx) -
1412					 utime_since_now(&td->start));
1413		}
1414
1415		if (td_write(td) && idx == DDIR_WRITE &&
1416		    td->o.do_verify &&
1417		    td->o.verify != VERIFY_NONE)
1418			log_io_piece(td, io_u);
1419
1420		icd->bytes_done[idx] += bytes;
1421
1422		if (io_u->end_io) {
1423			ret = io_u->end_io(td, io_u);
1424			if (ret && !icd->error)
1425				icd->error = ret;
1426		}
1427	} else if (io_u->error) {
1428		icd->error = io_u->error;
1429		io_u_log_error(td, io_u);
1430	}
1431	if (icd->error && td_non_fatal_error(icd->error) &&
1432           (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1433		/*
1434		 * If there is a non_fatal error, then add to the error count
1435		 * and clear all the errors.
1436		 */
1437		update_error_count(td, icd->error);
1438		td_clear_error(td);
1439		icd->error = 0;
1440		io_u->error = 0;
1441	}
1442}
1443
1444static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1445		     int nr)
1446{
1447	if (!td->o.disable_clat || !td->o.disable_bw)
1448		fio_gettime(&icd->time, NULL);
1449
1450	icd->nr = nr;
1451
1452	icd->error = 0;
1453	icd->bytes_done[0] = icd->bytes_done[1] = 0;
1454}
1455
1456static void ios_completed(struct thread_data *td,
1457			  struct io_completion_data *icd)
1458{
1459	struct io_u *io_u;
1460	int i;
1461
1462	for (i = 0; i < icd->nr; i++) {
1463		io_u = td->io_ops->event(td, i);
1464
1465		io_completed(td, io_u, icd);
1466
1467		if (!(io_u->flags & IO_U_F_FREE_DEF))
1468			put_io_u(td, io_u);
1469	}
1470}
1471
1472/*
1473 * Complete a single io_u for the sync engines.
1474 */
1475int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1476		       unsigned long *bytes)
1477{
1478	struct io_completion_data icd;
1479
1480	init_icd(td, &icd, 1);
1481	io_completed(td, io_u, &icd);
1482
1483	if (!(io_u->flags & IO_U_F_FREE_DEF))
1484		put_io_u(td, io_u);
1485
1486	if (icd.error) {
1487		td_verror(td, icd.error, "io_u_sync_complete");
1488		return -1;
1489	}
1490
1491	if (bytes) {
1492		bytes[0] += icd.bytes_done[0];
1493		bytes[1] += icd.bytes_done[1];
1494	}
1495
1496	return 0;
1497}
1498
1499/*
1500 * Called to complete min_events number of io for the async engines.
1501 */
1502int io_u_queued_complete(struct thread_data *td, int min_evts,
1503			 unsigned long *bytes)
1504{
1505	struct io_completion_data icd;
1506	struct timespec *tvp = NULL;
1507	int ret;
1508	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1509
1510	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1511
1512	if (!min_evts)
1513		tvp = &ts;
1514
1515	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1516	if (ret < 0) {
1517		td_verror(td, -ret, "td_io_getevents");
1518		return ret;
1519	} else if (!ret)
1520		return ret;
1521
1522	init_icd(td, &icd, ret);
1523	ios_completed(td, &icd);
1524	if (icd.error) {
1525		td_verror(td, icd.error, "io_u_queued_complete");
1526		return -1;
1527	}
1528
1529	if (bytes) {
1530		bytes[0] += icd.bytes_done[0];
1531		bytes[1] += icd.bytes_done[1];
1532	}
1533
1534	return 0;
1535}
1536
1537/*
1538 * Call when io_u is really queued, to update the submission latency.
1539 */
1540void io_u_queued(struct thread_data *td, struct io_u *io_u)
1541{
1542	if (!td->o.disable_slat) {
1543		unsigned long slat_time;
1544
1545		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1546		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1547	}
1548}
1549
1550/*
1551 * "randomly" fill the buffer contents
1552 */
1553void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1554		      unsigned int min_write, unsigned int max_bs)
1555{
1556	io_u->buf_filled_len = 0;
1557
1558	if (!td->o.zero_buffers) {
1559		unsigned int perc = td->o.compress_percentage;
1560
1561		if (perc) {
1562			unsigned int seg = min_write;
1563
1564			seg = min(min_write, td->o.compress_chunk);
1565			fill_random_buf_percentage(&td->buf_state, io_u->buf,
1566						perc, seg, max_bs);
1567		} else
1568			fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1569	} else
1570		memset(io_u->buf, 0, max_bs);
1571}
1572