io_u.c revision ed33585509a97c4a65e46967cb49419b6a4f8870
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	int account;			/* input */
17
18	int error;			/* output */
19	unsigned long bytes_done[2];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->file_map[] 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 unsigned long long block)
28{
29	unsigned int idx = RAND_MAP_IDX(f, block);
30	unsigned int bit = RAND_MAP_BIT(f, block);
31
32	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33
34	return (f->file_map[idx] & (1UL << bit)) == 0;
35}
36
37/*
38 * Mark a given offset as used in the map.
39 */
40static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41{
42	unsigned int min_bs = td->o.rw_min_bs;
43	struct fio_file *f = io_u->file;
44	unsigned long long block;
45	unsigned int blocks, nr_blocks;
46	int busy_check;
47
48	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
49	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50	blocks = 0;
51	busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
52
53	while (nr_blocks) {
54		unsigned int idx, bit;
55		unsigned long mask, this_blocks;
56
57		/*
58		 * If we have a mixed random workload, we may
59		 * encounter blocks we already did IO to.
60		 */
61		if (!busy_check) {
62			blocks = nr_blocks;
63			break;
64		}
65		if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
66			break;
67
68		idx = RAND_MAP_IDX(f, block);
69		bit = RAND_MAP_BIT(f, block);
70
71		fio_assert(td, idx < f->num_maps);
72
73		this_blocks = nr_blocks;
74		if (this_blocks + bit > BLOCKS_PER_MAP)
75			this_blocks = BLOCKS_PER_MAP - bit;
76
77		do {
78			if (this_blocks == BLOCKS_PER_MAP)
79				mask = -1UL;
80			else
81				mask = ((1UL << this_blocks) - 1) << bit;
82
83			if (!(f->file_map[idx] & mask))
84				break;
85
86			this_blocks--;
87		} while (this_blocks);
88
89		if (!this_blocks)
90			break;
91
92		f->file_map[idx] |= mask;
93		nr_blocks -= this_blocks;
94		blocks += this_blocks;
95		block += this_blocks;
96	}
97
98	if ((blocks * min_bs) < io_u->buflen)
99		io_u->buflen = blocks * min_bs;
100}
101
102static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
103				     enum fio_ddir ddir)
104{
105	unsigned long long max_blocks;
106	unsigned long long max_size;
107
108	assert(ddir_rw(ddir));
109
110	/*
111	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
112	 */
113	max_size = f->io_size;
114	if (max_size > f->real_file_size)
115		max_size = f->real_file_size;
116
117	if (td->o.zone_range)
118		max_size = td->o.zone_range;
119
120	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
121	if (!max_blocks)
122		return 0;
123
124	return max_blocks;
125}
126
127/*
128 * Return the next free block in the map.
129 */
130static int get_next_free_block(struct thread_data *td, struct fio_file *f,
131			       enum fio_ddir ddir, unsigned long long *b)
132{
133	unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
134	int i;
135
136	lastb = last_block(td, f, ddir);
137	if (!lastb)
138		return 1;
139
140	i = f->last_free_lookup;
141	block = i * BLOCKS_PER_MAP;
142	while (block * min_bs < f->real_file_size &&
143		block * min_bs < f->io_size) {
144		if (f->file_map[i] != -1UL) {
145			block += ffz(f->file_map[i]);
146			if (block > lastb)
147				break;
148			f->last_free_lookup = i;
149			*b = block;
150			return 0;
151		}
152
153		block += BLOCKS_PER_MAP;
154		i++;
155	}
156
157	dprint(FD_IO, "failed finding a free block\n");
158	return 1;
159}
160
161static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
162				enum fio_ddir ddir, unsigned long long *b)
163{
164	unsigned long long rmax, r, lastb;
165	int loops = 5;
166
167	lastb = last_block(td, f, ddir);
168	if (!lastb)
169		return 1;
170
171	if (f->failed_rands >= 200)
172		goto ffz;
173
174	rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
175	do {
176		if (td->o.use_os_rand)
177			r = os_random_long(&td->random_state);
178		else
179			r = __rand(&td->__random_state);
180
181		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
182
183		dprint(FD_RANDOM, "off rand %llu\n", r);
184
185
186		/*
187		 * if we are not maintaining a random map, we are done.
188		 */
189		if (!file_randommap(td, f))
190			goto ret_good;
191
192		/*
193		 * calculate map offset and check if it's free
194		 */
195		if (random_map_free(f, *b))
196			goto ret_good;
197
198		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
199									*b);
200	} while (--loops);
201
202	if (!f->failed_rands++)
203		f->last_free_lookup = 0;
204
205	/*
206	 * we get here, if we didn't suceed in looking up a block. generate
207	 * a random start offset into the filemap, and find the first free
208	 * block from there.
209	 */
210	loops = 10;
211	do {
212		f->last_free_lookup = (f->num_maps - 1) *
213					(r / ((unsigned long long) rmax + 1.0));
214		if (!get_next_free_block(td, f, ddir, b))
215			goto ret;
216
217		if (td->o.use_os_rand)
218			r = os_random_long(&td->random_state);
219		else
220			r = __rand(&td->__random_state);
221	} while (--loops);
222
223	/*
224	 * that didn't work either, try exhaustive search from the start
225	 */
226	f->last_free_lookup = 0;
227ffz:
228	if (!get_next_free_block(td, f, ddir, b))
229		return 0;
230	f->last_free_lookup = 0;
231	return get_next_free_block(td, f, ddir, b);
232ret_good:
233	f->failed_rands = 0;
234ret:
235	return 0;
236}
237
238static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239			       enum fio_ddir ddir, unsigned long long *b)
240{
241	if (get_next_rand_offset(td, f, ddir, b)) {
242		dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
243				f->file_name, f->last_pos, f->real_file_size);
244		return 1;
245	}
246
247	return 0;
248}
249
250static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
251			      enum fio_ddir ddir, unsigned long long *b)
252{
253	assert(ddir_rw(ddir));
254
255	if (f->last_pos < f->real_file_size) {
256		unsigned long long pos;
257
258		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
259			f->last_pos = f->real_file_size;
260
261		pos = f->last_pos - f->file_offset;
262		if (pos)
263			pos += td->o.ddir_seq_add;
264
265		*b = pos / td->o.min_bs[ddir];
266		return 0;
267	}
268
269	return 1;
270}
271
272static int get_next_block(struct thread_data *td, struct io_u *io_u,
273			  enum fio_ddir ddir, int rw_seq, unsigned long long *b)
274{
275	struct fio_file *f = io_u->file;
276	int ret;
277
278	assert(ddir_rw(ddir));
279
280	if (rw_seq) {
281		if (td_random(td))
282			ret = get_next_rand_block(td, f, ddir, b);
283		else
284			ret = get_next_seq_block(td, f, ddir, b);
285	} else {
286		io_u->flags |= IO_U_F_BUSY_OK;
287
288		if (td->o.rw_seq == RW_SEQ_SEQ) {
289			ret = get_next_seq_block(td, f, ddir, b);
290			if (ret)
291				ret = get_next_rand_block(td, f, ddir, b);
292		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
293			if (f->last_start != -1ULL)
294				*b = (f->last_start - f->file_offset)
295					/ td->o.min_bs[ddir];
296			else
297				*b = 0;
298			ret = 0;
299		} else {
300			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
301			ret = 1;
302		}
303	}
304
305	return ret;
306}
307
308/*
309 * For random io, generate a random new block and see if it's used. Repeat
310 * until we find a free one. For sequential io, just return the end of
311 * the last io issued.
312 */
313static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
314{
315	struct fio_file *f = io_u->file;
316	unsigned long long b;
317	enum fio_ddir ddir = io_u->ddir;
318	int rw_seq_hit = 0;
319
320	assert(ddir_rw(ddir));
321
322	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
323		rw_seq_hit = 1;
324		td->ddir_seq_nr = td->o.ddir_seq_nr;
325	}
326
327	if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
328		return 1;
329
330	io_u->offset = b * td->o.ba[ddir];
331	if (io_u->offset >= f->io_size) {
332		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
333					io_u->offset, f->io_size);
334		return 1;
335	}
336
337	io_u->offset += f->file_offset;
338	if (io_u->offset >= f->real_file_size) {
339		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
340					io_u->offset, f->real_file_size);
341		return 1;
342	}
343
344	return 0;
345}
346
347static int get_next_offset(struct thread_data *td, struct io_u *io_u)
348{
349	struct prof_io_ops *ops = &td->prof_io_ops;
350
351	if (ops->fill_io_u_off)
352		return ops->fill_io_u_off(td, io_u);
353
354	return __get_next_offset(td, io_u);
355}
356
357static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
358			    unsigned int buflen)
359{
360	struct fio_file *f = io_u->file;
361
362	return io_u->offset + buflen <= f->io_size + td->o.start_offset;
363}
364
365static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
366{
367	const int ddir = io_u->ddir;
368	unsigned int uninitialized_var(buflen);
369	unsigned int minbs, maxbs;
370	unsigned long r, rand_max;
371
372	assert(ddir_rw(ddir));
373
374	minbs = td->o.min_bs[ddir];
375	maxbs = td->o.max_bs[ddir];
376
377	if (minbs == maxbs)
378		return minbs;
379
380	if (td->o.use_os_rand)
381		rand_max = OS_RAND_MAX;
382	else
383		rand_max = FRAND_MAX;
384
385	do {
386		if (td->o.use_os_rand)
387			r = os_random_long(&td->bsrange_state);
388		else
389			r = __rand(&td->__bsrange_state);
390
391		if (!td->o.bssplit_nr[ddir]) {
392			buflen = 1 + (unsigned int) ((double) maxbs *
393					(r / (rand_max + 1.0)));
394			if (buflen < minbs)
395				buflen = minbs;
396		} else {
397			long perc = 0;
398			unsigned int i;
399
400			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
401				struct bssplit *bsp = &td->o.bssplit[ddir][i];
402
403				buflen = bsp->bs;
404				perc += bsp->perc;
405				if ((r <= ((rand_max / 100L) * perc)) &&
406				    io_u_fits(td, io_u, buflen))
407					break;
408			}
409		}
410
411		if (!td->o.bs_unaligned && is_power_of_2(minbs))
412			buflen = (buflen + minbs - 1) & ~(minbs - 1);
413
414	} while (!io_u_fits(td, io_u, buflen));
415
416	return buflen;
417}
418
419static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
420{
421	struct prof_io_ops *ops = &td->prof_io_ops;
422
423	if (ops->fill_io_u_size)
424		return ops->fill_io_u_size(td, io_u);
425
426	return __get_next_buflen(td, io_u);
427}
428
429static void set_rwmix_bytes(struct thread_data *td)
430{
431	unsigned int diff;
432
433	/*
434	 * we do time or byte based switch. this is needed because
435	 * buffered writes may issue a lot quicker than they complete,
436	 * whereas reads do not.
437	 */
438	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
439	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
440}
441
442static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
443{
444	unsigned int v;
445	unsigned long r;
446
447	if (td->o.use_os_rand) {
448		r = os_random_long(&td->rwmix_state);
449		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
450	} else {
451		r = __rand(&td->__rwmix_state);
452		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
453	}
454
455	if (v <= td->o.rwmix[DDIR_READ])
456		return DDIR_READ;
457
458	return DDIR_WRITE;
459}
460
461static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
462{
463	enum fio_ddir odir = ddir ^ 1;
464	struct timeval t;
465	long usec;
466
467	assert(ddir_rw(ddir));
468
469	if (td->rate_pending_usleep[ddir] <= 0)
470		return ddir;
471
472	/*
473	 * We have too much pending sleep in this direction. See if we
474	 * should switch.
475	 */
476	if (td_rw(td)) {
477		/*
478		 * Other direction does not have too much pending, switch
479		 */
480		if (td->rate_pending_usleep[odir] < 100000)
481			return odir;
482
483		/*
484		 * Both directions have pending sleep. Sleep the minimum time
485		 * and deduct from both.
486		 */
487		if (td->rate_pending_usleep[ddir] <=
488			td->rate_pending_usleep[odir]) {
489			usec = td->rate_pending_usleep[ddir];
490		} else {
491			usec = td->rate_pending_usleep[odir];
492			ddir = odir;
493		}
494	} else
495		usec = td->rate_pending_usleep[ddir];
496
497	/*
498	 * We are going to sleep, ensure that we flush anything pending as
499	 * not to skew our latency numbers
500	 */
501	if (td->cur_depth) {
502		int fio_unused ret;
503
504		ret = io_u_queued_complete(td, td->cur_depth, NULL);
505	}
506
507	fio_gettime(&t, NULL);
508	usec_sleep(td, usec);
509	usec = utime_since_now(&t);
510
511	td->rate_pending_usleep[ddir] -= usec;
512
513	odir = ddir ^ 1;
514	if (td_rw(td) && __should_check_rate(td, odir))
515		td->rate_pending_usleep[odir] -= usec;
516
517	return ddir;
518}
519
520/*
521 * Return the data direction for the next io_u. If the job is a
522 * mixed read/write workload, check the rwmix cycle and switch if
523 * necessary.
524 */
525static enum fio_ddir get_rw_ddir(struct thread_data *td)
526{
527	enum fio_ddir ddir;
528
529	/*
530	 * see if it's time to fsync
531	 */
532	if (td->o.fsync_blocks &&
533	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
534	     td->io_issues[DDIR_WRITE] && should_fsync(td))
535		return DDIR_SYNC;
536
537	/*
538	 * see if it's time to fdatasync
539	 */
540	if (td->o.fdatasync_blocks &&
541	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
542	     td->io_issues[DDIR_WRITE] && should_fsync(td))
543		return DDIR_DATASYNC;
544
545	/*
546	 * see if it's time to sync_file_range
547	 */
548	if (td->sync_file_range_nr &&
549	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
550	     td->io_issues[DDIR_WRITE] && should_fsync(td))
551		return DDIR_SYNC_FILE_RANGE;
552
553	if (td_rw(td)) {
554		/*
555		 * Check if it's time to seed a new data direction.
556		 */
557		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
558			/*
559			 * Put a top limit on how many bytes we do for
560			 * one data direction, to avoid overflowing the
561			 * ranges too much
562			 */
563			ddir = get_rand_ddir(td);
564
565			if (ddir != td->rwmix_ddir)
566				set_rwmix_bytes(td);
567
568			td->rwmix_ddir = ddir;
569		}
570		ddir = td->rwmix_ddir;
571	} else if (td_read(td))
572		ddir = DDIR_READ;
573	else
574		ddir = DDIR_WRITE;
575
576	td->rwmix_ddir = rate_ddir(td, ddir);
577	return td->rwmix_ddir;
578}
579
580static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
581{
582	io_u->ddir = get_rw_ddir(td);
583
584	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
585	    td->o.barrier_blocks &&
586	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
587	     td->io_issues[DDIR_WRITE])
588		io_u->flags |= IO_U_F_BARRIER;
589}
590
591void put_file_log(struct thread_data *td, struct fio_file *f)
592{
593	int ret = put_file(td, f);
594
595	if (ret)
596		td_verror(td, ret, "file close");
597}
598
599void put_io_u(struct thread_data *td, struct io_u *io_u)
600{
601	td_io_u_lock(td);
602
603	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
604		put_file_log(td, io_u->file);
605	io_u->file = NULL;
606	io_u->flags &= ~IO_U_F_FREE_DEF;
607	io_u->flags |= IO_U_F_FREE;
608
609	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
610		td->cur_depth--;
611	flist_del_init(&io_u->list);
612	flist_add(&io_u->list, &td->io_u_freelist);
613	td_io_u_unlock(td);
614	td_io_u_free_notify(td);
615}
616
617void clear_io_u(struct thread_data *td, struct io_u *io_u)
618{
619	io_u->flags &= ~IO_U_F_FLIGHT;
620	put_io_u(td, io_u);
621}
622
623void requeue_io_u(struct thread_data *td, struct io_u **io_u)
624{
625	struct io_u *__io_u = *io_u;
626
627	dprint(FD_IO, "requeue %p\n", __io_u);
628
629	td_io_u_lock(td);
630
631	__io_u->flags |= IO_U_F_FREE;
632	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
633		td->io_issues[__io_u->ddir]--;
634
635	__io_u->flags &= ~IO_U_F_FLIGHT;
636	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
637		td->cur_depth--;
638	flist_del(&__io_u->list);
639	flist_add_tail(&__io_u->list, &td->io_u_requeues);
640	td_io_u_unlock(td);
641	*io_u = NULL;
642}
643
644static int fill_io_u(struct thread_data *td, struct io_u *io_u)
645{
646	if (td->io_ops->flags & FIO_NOIO)
647		goto out;
648
649	set_rw_ddir(td, io_u);
650
651	/*
652	 * fsync() or fdatasync() or trim etc, we are done
653	 */
654	if (!ddir_rw(io_u->ddir))
655		goto out;
656
657	/*
658	 * See if it's time to switch to a new zone
659	 */
660	if (td->zone_bytes >= td->o.zone_size) {
661		td->zone_bytes = 0;
662		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
663		io_u->file->last_pos = io_u->file->file_offset;
664		td->io_skip_bytes += td->o.zone_skip;
665	}
666
667	/*
668	 * No log, let the seq/rand engine retrieve the next buflen and
669	 * position.
670	 */
671	if (get_next_offset(td, io_u)) {
672		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
673		return 1;
674	}
675
676	io_u->buflen = get_next_buflen(td, io_u);
677	if (!io_u->buflen) {
678		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
679		return 1;
680	}
681
682	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
683		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
684		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
685				io_u->buflen, io_u->file->real_file_size);
686		return 1;
687	}
688
689	/*
690	 * mark entry before potentially trimming io_u
691	 */
692	if (td_random(td) && file_randommap(td, io_u->file))
693		mark_random_map(td, io_u);
694
695	/*
696	 * If using a write iolog, store this entry.
697	 */
698out:
699	dprint_io_u(io_u, "fill_io_u");
700	td->zone_bytes += io_u->buflen;
701	log_io_u(td, io_u);
702	return 0;
703}
704
705static void __io_u_mark_map(unsigned int *map, unsigned int nr)
706{
707	int idx = 0;
708
709	switch (nr) {
710	default:
711		idx = 6;
712		break;
713	case 33 ... 64:
714		idx = 5;
715		break;
716	case 17 ... 32:
717		idx = 4;
718		break;
719	case 9 ... 16:
720		idx = 3;
721		break;
722	case 5 ... 8:
723		idx = 2;
724		break;
725	case 1 ... 4:
726		idx = 1;
727	case 0:
728		break;
729	}
730
731	map[idx]++;
732}
733
734void io_u_mark_submit(struct thread_data *td, unsigned int nr)
735{
736	__io_u_mark_map(td->ts.io_u_submit, nr);
737	td->ts.total_submit++;
738}
739
740void io_u_mark_complete(struct thread_data *td, unsigned int nr)
741{
742	__io_u_mark_map(td->ts.io_u_complete, nr);
743	td->ts.total_complete++;
744}
745
746void io_u_mark_depth(struct thread_data *td, unsigned int nr)
747{
748	int idx = 0;
749
750	switch (td->cur_depth) {
751	default:
752		idx = 6;
753		break;
754	case 32 ... 63:
755		idx = 5;
756		break;
757	case 16 ... 31:
758		idx = 4;
759		break;
760	case 8 ... 15:
761		idx = 3;
762		break;
763	case 4 ... 7:
764		idx = 2;
765		break;
766	case 2 ... 3:
767		idx = 1;
768	case 1:
769		break;
770	}
771
772	td->ts.io_u_map[idx] += nr;
773}
774
775static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
776{
777	int idx = 0;
778
779	assert(usec < 1000);
780
781	switch (usec) {
782	case 750 ... 999:
783		idx = 9;
784		break;
785	case 500 ... 749:
786		idx = 8;
787		break;
788	case 250 ... 499:
789		idx = 7;
790		break;
791	case 100 ... 249:
792		idx = 6;
793		break;
794	case 50 ... 99:
795		idx = 5;
796		break;
797	case 20 ... 49:
798		idx = 4;
799		break;
800	case 10 ... 19:
801		idx = 3;
802		break;
803	case 4 ... 9:
804		idx = 2;
805		break;
806	case 2 ... 3:
807		idx = 1;
808	case 0 ... 1:
809		break;
810	}
811
812	assert(idx < FIO_IO_U_LAT_U_NR);
813	td->ts.io_u_lat_u[idx]++;
814}
815
816static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
817{
818	int idx = 0;
819
820	switch (msec) {
821	default:
822		idx = 11;
823		break;
824	case 1000 ... 1999:
825		idx = 10;
826		break;
827	case 750 ... 999:
828		idx = 9;
829		break;
830	case 500 ... 749:
831		idx = 8;
832		break;
833	case 250 ... 499:
834		idx = 7;
835		break;
836	case 100 ... 249:
837		idx = 6;
838		break;
839	case 50 ... 99:
840		idx = 5;
841		break;
842	case 20 ... 49:
843		idx = 4;
844		break;
845	case 10 ... 19:
846		idx = 3;
847		break;
848	case 4 ... 9:
849		idx = 2;
850		break;
851	case 2 ... 3:
852		idx = 1;
853	case 0 ... 1:
854		break;
855	}
856
857	assert(idx < FIO_IO_U_LAT_M_NR);
858	td->ts.io_u_lat_m[idx]++;
859}
860
861static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
862{
863	if (usec < 1000)
864		io_u_mark_lat_usec(td, usec);
865	else
866		io_u_mark_lat_msec(td, usec / 1000);
867}
868
869/*
870 * Get next file to service by choosing one at random
871 */
872static struct fio_file *get_next_file_rand(struct thread_data *td,
873					   enum fio_file_flags goodf,
874					   enum fio_file_flags badf)
875{
876	struct fio_file *f;
877	int fno;
878
879	do {
880		int opened = 0;
881		unsigned long r;
882
883		if (td->o.use_os_rand) {
884			r = os_random_long(&td->next_file_state);
885			fno = (unsigned int) ((double) td->o.nr_files
886				* (r / (OS_RAND_MAX + 1.0)));
887		} else {
888			r = __rand(&td->__next_file_state);
889			fno = (unsigned int) ((double) td->o.nr_files
890				* (r / (FRAND_MAX + 1.0)));
891		}
892
893		f = td->files[fno];
894		if (fio_file_done(f))
895			continue;
896
897		if (!fio_file_open(f)) {
898			int err;
899
900			err = td_io_open_file(td, f);
901			if (err)
902				continue;
903			opened = 1;
904		}
905
906		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
907			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
908			return f;
909		}
910		if (opened)
911			td_io_close_file(td, f);
912	} while (1);
913}
914
915/*
916 * Get next file to service by doing round robin between all available ones
917 */
918static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
919					 int badf)
920{
921	unsigned int old_next_file = td->next_file;
922	struct fio_file *f;
923
924	do {
925		int opened = 0;
926
927		f = td->files[td->next_file];
928
929		td->next_file++;
930		if (td->next_file >= td->o.nr_files)
931			td->next_file = 0;
932
933		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
934		if (fio_file_done(f)) {
935			f = NULL;
936			continue;
937		}
938
939		if (!fio_file_open(f)) {
940			int err;
941
942			err = td_io_open_file(td, f);
943			if (err) {
944				dprint(FD_FILE, "error %d on open of %s\n",
945					err, f->file_name);
946				f = NULL;
947				continue;
948			}
949			opened = 1;
950		}
951
952		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
953								f->flags);
954		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
955			break;
956
957		if (opened)
958			td_io_close_file(td, f);
959
960		f = NULL;
961	} while (td->next_file != old_next_file);
962
963	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
964	return f;
965}
966
967static struct fio_file *__get_next_file(struct thread_data *td)
968{
969	struct fio_file *f;
970
971	assert(td->o.nr_files <= td->files_index);
972
973	if (td->nr_done_files >= td->o.nr_files) {
974		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
975				" nr_files=%d\n", td->nr_open_files,
976						  td->nr_done_files,
977						  td->o.nr_files);
978		return NULL;
979	}
980
981	f = td->file_service_file;
982	if (f && fio_file_open(f) && !fio_file_closing(f)) {
983		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
984			goto out;
985		if (td->file_service_left--)
986			goto out;
987	}
988
989	if (td->o.file_service_type == FIO_FSERVICE_RR ||
990	    td->o.file_service_type == FIO_FSERVICE_SEQ)
991		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
992	else
993		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
994
995	td->file_service_file = f;
996	td->file_service_left = td->file_service_nr - 1;
997out:
998	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
999	return f;
1000}
1001
1002static struct fio_file *get_next_file(struct thread_data *td)
1003{
1004	struct prof_io_ops *ops = &td->prof_io_ops;
1005
1006	if (ops->get_next_file)
1007		return ops->get_next_file(td);
1008
1009	return __get_next_file(td);
1010}
1011
1012static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1013{
1014	struct fio_file *f;
1015
1016	do {
1017		f = get_next_file(td);
1018		if (!f)
1019			return 1;
1020
1021		io_u->file = f;
1022		get_file(f);
1023
1024		if (!fill_io_u(td, io_u))
1025			break;
1026
1027		put_file_log(td, f);
1028		td_io_close_file(td, f);
1029		io_u->file = NULL;
1030		fio_file_set_done(f);
1031		td->nr_done_files++;
1032		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1033					td->nr_done_files, td->o.nr_files);
1034	} while (1);
1035
1036	return 0;
1037}
1038
1039
1040struct io_u *__get_io_u(struct thread_data *td)
1041{
1042	struct io_u *io_u = NULL;
1043
1044	td_io_u_lock(td);
1045
1046again:
1047	if (!flist_empty(&td->io_u_requeues))
1048		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1049	else if (!queue_full(td)) {
1050		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1051
1052		io_u->buflen = 0;
1053		io_u->resid = 0;
1054		io_u->file = NULL;
1055		io_u->end_io = NULL;
1056	}
1057
1058	if (io_u) {
1059		assert(io_u->flags & IO_U_F_FREE);
1060		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1061		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1062
1063		io_u->error = 0;
1064		flist_del(&io_u->list);
1065		flist_add(&io_u->list, &td->io_u_busylist);
1066		td->cur_depth++;
1067		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1068	} else if (td->o.verify_async) {
1069		/*
1070		 * We ran out, wait for async verify threads to finish and
1071		 * return one
1072		 */
1073		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1074		goto again;
1075	}
1076
1077	td_io_u_unlock(td);
1078	return io_u;
1079}
1080
1081static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1082{
1083	if (td->o.trim_backlog && td->trim_entries) {
1084		int get_trim = 0;
1085
1086		if (td->trim_batch) {
1087			td->trim_batch--;
1088			get_trim = 1;
1089		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1090			 td->last_ddir != DDIR_READ) {
1091			td->trim_batch = td->o.trim_batch;
1092			if (!td->trim_batch)
1093				td->trim_batch = td->o.trim_backlog;
1094			get_trim = 1;
1095		}
1096
1097		if (get_trim && !get_next_trim(td, io_u))
1098			return 1;
1099	}
1100
1101	return 0;
1102}
1103
1104static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1105{
1106	if (td->o.verify_backlog && td->io_hist_len) {
1107		int get_verify = 0;
1108
1109		if (td->verify_batch) {
1110			td->verify_batch--;
1111			get_verify = 1;
1112		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
1113			 td->last_ddir != DDIR_READ) {
1114			td->verify_batch = td->o.verify_batch;
1115			if (!td->verify_batch)
1116				td->verify_batch = td->o.verify_backlog;
1117			get_verify = 1;
1118		}
1119
1120		if (get_verify && !get_next_verify(td, io_u))
1121			return 1;
1122	}
1123
1124	return 0;
1125}
1126
1127/*
1128 * Fill offset and start time into the buffer content, to prevent too
1129 * easy compressible data for simple de-dupe attempts. Do this for every
1130 * 512b block in the range, since that should be the smallest block size
1131 * we can expect from a device.
1132 */
1133static void small_content_scramble(struct io_u *io_u)
1134{
1135	unsigned int i, nr_blocks = io_u->buflen / 512;
1136	unsigned long long boffset;
1137	unsigned int offset;
1138	void *p, *end;
1139
1140	if (!nr_blocks)
1141		return;
1142
1143	p = io_u->xfer_buf;
1144	boffset = io_u->offset;
1145
1146	for (i = 0; i < nr_blocks; i++) {
1147		/*
1148		 * Fill the byte offset into a "random" start offset of
1149		 * the buffer, given by the product of the usec time
1150		 * and the actual offset.
1151		 */
1152		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1153		offset &= ~(sizeof(unsigned long long) - 1);
1154		if (offset >= 512 - sizeof(unsigned long long))
1155			offset -= sizeof(unsigned long long);
1156		memcpy(p + offset, &boffset, sizeof(boffset));
1157
1158		end = p + 512 - sizeof(io_u->start_time);
1159		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1160		p += 512;
1161		boffset += 512;
1162	}
1163}
1164
1165/*
1166 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1167 * etc. The returned io_u is fully ready to be prepped and submitted.
1168 */
1169struct io_u *get_io_u(struct thread_data *td)
1170{
1171	struct fio_file *f;
1172	struct io_u *io_u;
1173	int do_scramble = 0;
1174
1175	io_u = __get_io_u(td);
1176	if (!io_u) {
1177		dprint(FD_IO, "__get_io_u failed\n");
1178		return NULL;
1179	}
1180
1181	if (check_get_verify(td, io_u))
1182		goto out;
1183	if (check_get_trim(td, io_u))
1184		goto out;
1185
1186	/*
1187	 * from a requeue, io_u already setup
1188	 */
1189	if (io_u->file)
1190		goto out;
1191
1192	/*
1193	 * If using an iolog, grab next piece if any available.
1194	 */
1195	if (td->o.read_iolog_file) {
1196		if (read_iolog_get(td, io_u))
1197			goto err_put;
1198	} else if (set_io_u_file(td, io_u)) {
1199		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1200		goto err_put;
1201	}
1202
1203	f = io_u->file;
1204	assert(fio_file_open(f));
1205
1206	if (ddir_rw(io_u->ddir)) {
1207		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1208			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1209			goto err_put;
1210		}
1211
1212		f->last_start = io_u->offset;
1213		f->last_pos = io_u->offset + io_u->buflen;
1214
1215		if (io_u->ddir == DDIR_WRITE) {
1216			if (td->o.verify != VERIFY_NONE)
1217				populate_verify_io_u(td, io_u);
1218			else if (td->o.refill_buffers)
1219				io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1220			else if (td->o.scramble_buffers)
1221				do_scramble = 1;
1222		} else if (io_u->ddir == DDIR_READ) {
1223			/*
1224			 * Reset the buf_filled parameters so next time if the
1225			 * buffer is used for writes it is refilled.
1226			 */
1227			io_u->buf_filled_len = 0;
1228		}
1229	}
1230
1231	/*
1232	 * Set io data pointers.
1233	 */
1234	io_u->xfer_buf = io_u->buf;
1235	io_u->xfer_buflen = io_u->buflen;
1236
1237out:
1238	assert(io_u->file);
1239	if (!td_io_prep(td, io_u)) {
1240		if (!td->o.disable_slat)
1241			fio_gettime(&io_u->start_time, NULL);
1242		if (do_scramble)
1243			small_content_scramble(io_u);
1244		return io_u;
1245	}
1246err_put:
1247	dprint(FD_IO, "get_io_u failed\n");
1248	put_io_u(td, io_u);
1249	return NULL;
1250}
1251
1252void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1253{
1254	const char *msg[] = { "read", "write", "sync", "datasync",
1255				"sync_file_range", "wait", "trim" };
1256
1257
1258
1259	log_err("fio: io_u error");
1260
1261	if (io_u->file)
1262		log_err(" on file %s", io_u->file->file_name);
1263
1264	log_err(": %s\n", strerror(io_u->error));
1265
1266	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1267					io_u->offset, io_u->xfer_buflen);
1268
1269	if (!td->error)
1270		td_verror(td, io_u->error, "io_u error");
1271}
1272
1273static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1274				  struct io_completion_data *icd,
1275				  const enum fio_ddir idx, unsigned int bytes)
1276{
1277	unsigned long uninitialized_var(lusec);
1278
1279	if (!icd->account)
1280		return;
1281
1282	if (!td->o.disable_clat || !td->o.disable_bw)
1283		lusec = utime_since(&io_u->issue_time, &icd->time);
1284
1285	if (!td->o.disable_lat) {
1286		unsigned long tusec;
1287
1288		tusec = utime_since(&io_u->start_time, &icd->time);
1289		add_lat_sample(td, idx, tusec, bytes);
1290	}
1291
1292	if (!td->o.disable_clat) {
1293		add_clat_sample(td, idx, lusec, bytes);
1294		io_u_mark_latency(td, lusec);
1295	}
1296
1297	if (!td->o.disable_bw)
1298		add_bw_sample(td, idx, bytes, &icd->time);
1299
1300	add_iops_sample(td, idx, &icd->time);
1301}
1302
1303static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1304{
1305	unsigned long long secs, remainder, bps, bytes;
1306	bytes = td->this_io_bytes[ddir];
1307	bps = td->rate_bps[ddir];
1308	secs = bytes / bps;
1309	remainder = bytes % bps;
1310	return remainder * 1000000 / bps + secs * 1000000;
1311}
1312
1313static void io_completed(struct thread_data *td, struct io_u *io_u,
1314			 struct io_completion_data *icd)
1315{
1316	/*
1317	 * Older gcc's are too dumb to realize that usec is always used
1318	 * initialized, silence that warning.
1319	 */
1320	unsigned long uninitialized_var(usec);
1321	struct fio_file *f;
1322
1323	dprint_io_u(io_u, "io complete");
1324
1325	td_io_u_lock(td);
1326	assert(io_u->flags & IO_U_F_FLIGHT);
1327	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1328	td_io_u_unlock(td);
1329
1330	if (ddir_sync(io_u->ddir)) {
1331		td->last_was_sync = 1;
1332		f = io_u->file;
1333		if (f) {
1334			f->first_write = -1ULL;
1335			f->last_write = -1ULL;
1336		}
1337		return;
1338	}
1339
1340	td->last_was_sync = 0;
1341	td->last_ddir = io_u->ddir;
1342
1343	if (!io_u->error && ddir_rw(io_u->ddir)) {
1344		unsigned int bytes = io_u->buflen - io_u->resid;
1345		const enum fio_ddir idx = io_u->ddir;
1346		const enum fio_ddir odx = io_u->ddir ^ 1;
1347		int ret;
1348
1349		td->io_blocks[idx]++;
1350		td->this_io_blocks[idx]++;
1351		td->io_bytes[idx] += bytes;
1352		td->this_io_bytes[idx] += bytes;
1353
1354		if (idx == DDIR_WRITE) {
1355			f = io_u->file;
1356			if (f) {
1357				if (f->first_write == -1ULL ||
1358				    io_u->offset < f->first_write)
1359					f->first_write = io_u->offset;
1360				if (f->last_write == -1ULL ||
1361				    ((io_u->offset + bytes) > f->last_write))
1362					f->last_write = io_u->offset + bytes;
1363			}
1364		}
1365
1366		if (ramp_time_over(td) && td->runstate == TD_RUNNING) {
1367			account_io_completion(td, io_u, icd, idx, bytes);
1368
1369			if (__should_check_rate(td, idx)) {
1370				td->rate_pending_usleep[idx] =
1371					(usec_for_io(td, idx) -
1372					 utime_since_now(&td->start));
1373			}
1374			if (__should_check_rate(td, odx))
1375				td->rate_pending_usleep[odx] =
1376					(usec_for_io(td, odx) -
1377					 utime_since_now(&td->start));
1378		}
1379
1380		if (td_write(td) && idx == DDIR_WRITE &&
1381		    td->o.do_verify &&
1382		    td->o.verify != VERIFY_NONE)
1383			log_io_piece(td, io_u);
1384
1385		icd->bytes_done[idx] += bytes;
1386
1387		if (io_u->end_io) {
1388			ret = io_u->end_io(td, io_u);
1389			if (ret && !icd->error)
1390				icd->error = ret;
1391		}
1392	} else if (io_u->error) {
1393		icd->error = io_u->error;
1394		io_u_log_error(td, io_u);
1395	}
1396	if (icd->error && td_non_fatal_error(icd->error) &&
1397           (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1398		/*
1399		 * If there is a non_fatal error, then add to the error count
1400		 * and clear all the errors.
1401		 */
1402		update_error_count(td, icd->error);
1403		td_clear_error(td);
1404		icd->error = 0;
1405		io_u->error = 0;
1406	}
1407}
1408
1409static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1410		     int nr)
1411{
1412	if (!td->o.disable_clat || !td->o.disable_bw)
1413		fio_gettime(&icd->time, NULL);
1414
1415	icd->nr = nr;
1416	icd->account = 1;
1417
1418	icd->error = 0;
1419	icd->bytes_done[0] = icd->bytes_done[1] = 0;
1420}
1421
1422static void ios_completed(struct thread_data *td,
1423			  struct io_completion_data *icd)
1424{
1425	struct io_u *io_u;
1426	int i;
1427
1428	for (i = 0; i < icd->nr; i++) {
1429		io_u = td->io_ops->event(td, i);
1430
1431		io_completed(td, io_u, icd);
1432
1433		if (!(io_u->flags & IO_U_F_FREE_DEF))
1434			put_io_u(td, io_u);
1435
1436		icd->account = 0;
1437	}
1438}
1439
1440/*
1441 * Complete a single io_u for the sync engines.
1442 */
1443int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1444		       unsigned long *bytes)
1445{
1446	struct io_completion_data icd;
1447
1448	init_icd(td, &icd, 1);
1449	io_completed(td, io_u, &icd);
1450
1451	if (!(io_u->flags & IO_U_F_FREE_DEF))
1452		put_io_u(td, io_u);
1453
1454	if (icd.error) {
1455		td_verror(td, icd.error, "io_u_sync_complete");
1456		return -1;
1457	}
1458
1459	if (bytes) {
1460		bytes[0] += icd.bytes_done[0];
1461		bytes[1] += icd.bytes_done[1];
1462	}
1463
1464	return 0;
1465}
1466
1467/*
1468 * Called to complete min_events number of io for the async engines.
1469 */
1470int io_u_queued_complete(struct thread_data *td, int min_evts,
1471			 unsigned long *bytes)
1472{
1473	struct io_completion_data icd;
1474	struct timespec *tvp = NULL;
1475	int ret;
1476	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1477
1478	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1479
1480	if (!min_evts)
1481		tvp = &ts;
1482
1483	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1484	if (ret < 0) {
1485		td_verror(td, -ret, "td_io_getevents");
1486		return ret;
1487	} else if (!ret)
1488		return ret;
1489
1490	init_icd(td, &icd, ret);
1491	ios_completed(td, &icd);
1492	if (icd.error) {
1493		td_verror(td, icd.error, "io_u_queued_complete");
1494		return -1;
1495	}
1496
1497	if (bytes) {
1498		bytes[0] += icd.bytes_done[0];
1499		bytes[1] += icd.bytes_done[1];
1500	}
1501
1502	return 0;
1503}
1504
1505/*
1506 * Call when io_u is really queued, to update the submission latency.
1507 */
1508void io_u_queued(struct thread_data *td, struct io_u *io_u)
1509{
1510	if (!td->o.disable_slat) {
1511		unsigned long slat_time;
1512
1513		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1514		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1515	}
1516}
1517
1518/*
1519 * "randomly" fill the buffer contents
1520 */
1521void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1522		      unsigned int max_bs)
1523{
1524	io_u->buf_filled_len = 0;
1525
1526	if (!td->o.zero_buffers)
1527		fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1528	else
1529		memset(io_u->buf, 0, max_bs);
1530}
1531