io_u.c revision 8055e41d0ecc54770a2653427532b3e2c5fabdad
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/bitmap.h"
14
15struct io_completion_data {
16	int nr;				/* input */
17
18	int error;			/* output */
19	unsigned long bytes_done[DDIR_RWDIR_CNT];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->io_bitmap 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	return !bitmap_isset(f->io_bitmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37	unsigned int min_bs = td->o.rw_min_bs;
38	struct fio_file *f = io_u->file;
39	unsigned long long block;
40	unsigned int nr_blocks;
41
42	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
43	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45	if (!(io_u->flags & IO_U_F_BUSY_OK))
46		nr_blocks = bitmap_set_nr(f->io_bitmap, block, nr_blocks);
47
48	if ((nr_blocks * min_bs) < io_u->buflen)
49		io_u->buflen = nr_blocks * min_bs;
50}
51
52static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
53				     enum fio_ddir ddir)
54{
55	unsigned long long max_blocks;
56	unsigned long long max_size;
57
58	assert(ddir_rw(ddir));
59
60	/*
61	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
62	 */
63	max_size = f->io_size;
64	if (max_size > f->real_file_size)
65		max_size = f->real_file_size;
66
67	if (td->o.zone_range)
68		max_size = td->o.zone_range;
69
70	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
71	if (!max_blocks)
72		return 0;
73
74	return max_blocks;
75}
76
77static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
78				  enum fio_ddir ddir, unsigned long long *b)
79{
80	unsigned long long r;
81
82	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
83		unsigned long long rmax, lastb;
84
85		lastb = last_block(td, f, ddir);
86		if (!lastb)
87			return 1;
88
89		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
90
91		if (td->o.use_os_rand) {
92			rmax = OS_RAND_MAX;
93			r = os_random_long(&td->random_state);
94		} else {
95			rmax = FRAND_MAX;
96			r = __rand(&td->__random_state);
97		}
98
99		dprint(FD_RANDOM, "off rand %llu\n", r);
100
101		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
102	} else {
103		uint64_t off = 0;
104
105		if (lfsr_next(&f->lfsr, &off))
106			return 1;
107
108		*b = off;
109	}
110
111	/*
112	 * if we are not maintaining a random map, we are done.
113	 */
114	if (!file_randommap(td, f))
115		goto ret;
116
117	/*
118	 * calculate map offset and check if it's free
119	 */
120	if (random_map_free(f, *b))
121		goto ret;
122
123	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", *b);
124
125	*b = bitmap_next_free(f->io_bitmap, *b);
126	if (*b == (uint64_t) -1ULL)
127		return 1;
128ret:
129	return 0;
130}
131
132static int __get_next_rand_offset_zipf(struct thread_data *td,
133				       struct fio_file *f, enum fio_ddir ddir,
134				       unsigned long long *b)
135{
136	*b = zipf_next(&f->zipf);
137	return 0;
138}
139
140static int __get_next_rand_offset_pareto(struct thread_data *td,
141					 struct fio_file *f, enum fio_ddir ddir,
142					 unsigned long long *b)
143{
144	*b = pareto_next(&f->zipf);
145	return 0;
146}
147
148static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
149				enum fio_ddir ddir, unsigned long long *b)
150{
151	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
152		return __get_next_rand_offset(td, f, ddir, b);
153	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
154		return __get_next_rand_offset_zipf(td, f, ddir, b);
155	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
156		return __get_next_rand_offset_pareto(td, f, ddir, b);
157
158	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
159	return 1;
160}
161
162static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
163			       enum fio_ddir ddir, unsigned long long *b)
164{
165	if (!get_next_rand_offset(td, f, ddir, b))
166		return 0;
167
168	if (td->o.time_based) {
169		fio_file_reset(f);
170		if (!get_next_rand_offset(td, f, ddir, b))
171			return 0;
172	}
173
174	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
175			f->file_name, f->last_pos, f->real_file_size);
176	return 1;
177}
178
179static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
180			       enum fio_ddir ddir, unsigned long long *offset)
181{
182	assert(ddir_rw(ddir));
183
184	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
185		f->last_pos = f->last_pos - f->io_size;
186
187	if (f->last_pos < f->real_file_size) {
188		unsigned long long pos;
189
190		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
191			f->last_pos = f->real_file_size;
192
193		pos = f->last_pos - f->file_offset;
194		if (pos)
195			pos += td->o.ddir_seq_add;
196
197		*offset = pos;
198		return 0;
199	}
200
201	return 1;
202}
203
204static int get_next_block(struct thread_data *td, struct io_u *io_u,
205			  enum fio_ddir ddir, int rw_seq)
206{
207	struct fio_file *f = io_u->file;
208	unsigned long long b, offset;
209	int ret;
210
211	assert(ddir_rw(ddir));
212
213	b = offset = -1ULL;
214
215	if (rw_seq) {
216		if (td_random(td))
217			ret = get_next_rand_block(td, f, ddir, &b);
218		else
219			ret = get_next_seq_offset(td, f, ddir, &offset);
220	} else {
221		io_u->flags |= IO_U_F_BUSY_OK;
222
223		if (td->o.rw_seq == RW_SEQ_SEQ) {
224			ret = get_next_seq_offset(td, f, ddir, &offset);
225			if (ret)
226				ret = get_next_rand_block(td, f, ddir, &b);
227		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
228			if (f->last_start != -1ULL)
229				offset = f->last_start - f->file_offset;
230			else
231				offset = 0;
232			ret = 0;
233		} else {
234			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
235			ret = 1;
236		}
237	}
238
239	if (!ret) {
240		if (offset != -1ULL)
241			io_u->offset = offset;
242		else if (b != -1ULL)
243			io_u->offset = b * td->o.ba[ddir];
244		else {
245			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
246								offset, b);
247			ret = 1;
248		}
249	}
250
251	return ret;
252}
253
254/*
255 * For random io, generate a random new block and see if it's used. Repeat
256 * until we find a free one. For sequential io, just return the end of
257 * the last io issued.
258 */
259static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
260{
261	struct fio_file *f = io_u->file;
262	enum fio_ddir ddir = io_u->ddir;
263	int rw_seq_hit = 0;
264
265	assert(ddir_rw(ddir));
266
267	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
268		rw_seq_hit = 1;
269		td->ddir_seq_nr = td->o.ddir_seq_nr;
270	}
271
272	if (get_next_block(td, io_u, ddir, rw_seq_hit))
273		return 1;
274
275	if (io_u->offset >= f->io_size) {
276		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
277					io_u->offset, f->io_size);
278		return 1;
279	}
280
281	io_u->offset += f->file_offset;
282	if (io_u->offset >= f->real_file_size) {
283		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
284					io_u->offset, f->real_file_size);
285		return 1;
286	}
287
288	return 0;
289}
290
291static int get_next_offset(struct thread_data *td, struct io_u *io_u)
292{
293	struct prof_io_ops *ops = &td->prof_io_ops;
294
295	if (ops->fill_io_u_off)
296		return ops->fill_io_u_off(td, io_u);
297
298	return __get_next_offset(td, io_u);
299}
300
301static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
302			    unsigned int buflen)
303{
304	struct fio_file *f = io_u->file;
305
306	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
307}
308
309static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
310{
311	const int ddir = io_u->ddir;
312	unsigned int buflen = 0;
313	unsigned int minbs, maxbs;
314	unsigned long r, rand_max;
315
316	assert(ddir_rw(ddir));
317
318	minbs = td->o.min_bs[ddir];
319	maxbs = td->o.max_bs[ddir];
320
321	if (minbs == maxbs)
322		return minbs;
323
324	/*
325	 * If we can't satisfy the min block size from here, then fail
326	 */
327	if (!io_u_fits(td, io_u, minbs))
328		return 0;
329
330	if (td->o.use_os_rand)
331		rand_max = OS_RAND_MAX;
332	else
333		rand_max = FRAND_MAX;
334
335	do {
336		if (td->o.use_os_rand)
337			r = os_random_long(&td->bsrange_state);
338		else
339			r = __rand(&td->__bsrange_state);
340
341		if (!td->o.bssplit_nr[ddir]) {
342			buflen = 1 + (unsigned int) ((double) maxbs *
343					(r / (rand_max + 1.0)));
344			if (buflen < minbs)
345				buflen = minbs;
346		} else {
347			long perc = 0;
348			unsigned int i;
349
350			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
351				struct bssplit *bsp = &td->o.bssplit[ddir][i];
352
353				buflen = bsp->bs;
354				perc += bsp->perc;
355				if ((r <= ((rand_max / 100L) * perc)) &&
356				    io_u_fits(td, io_u, buflen))
357					break;
358			}
359		}
360
361		if (!td->o.bs_unaligned && is_power_of_2(minbs))
362			buflen = (buflen + minbs - 1) & ~(minbs - 1);
363
364	} while (!io_u_fits(td, io_u, buflen));
365
366	return buflen;
367}
368
369static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
370{
371	struct prof_io_ops *ops = &td->prof_io_ops;
372
373	if (ops->fill_io_u_size)
374		return ops->fill_io_u_size(td, io_u);
375
376	return __get_next_buflen(td, io_u);
377}
378
379static void set_rwmix_bytes(struct thread_data *td)
380{
381	unsigned int diff;
382
383	/*
384	 * we do time or byte based switch. this is needed because
385	 * buffered writes may issue a lot quicker than they complete,
386	 * whereas reads do not.
387	 */
388	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
389	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
390}
391
392static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
393{
394	unsigned int v;
395	unsigned long r;
396
397	if (td->o.use_os_rand) {
398		r = os_random_long(&td->rwmix_state);
399		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
400	} else {
401		r = __rand(&td->__rwmix_state);
402		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
403	}
404
405	if (v <= td->o.rwmix[DDIR_READ])
406		return DDIR_READ;
407
408	return DDIR_WRITE;
409}
410
411static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
412{
413	enum fio_ddir odir = ddir ^ 1;
414	struct timeval t;
415	long usec;
416
417	assert(ddir_rw(ddir));
418
419	if (td->rate_pending_usleep[ddir] <= 0)
420		return ddir;
421
422	/*
423	 * We have too much pending sleep in this direction. See if we
424	 * should switch.
425	 */
426	if (td_rw(td)) {
427		/*
428		 * Other direction does not have too much pending, switch
429		 */
430		if (td->rate_pending_usleep[odir] < 100000)
431			return odir;
432
433		/*
434		 * Both directions have pending sleep. Sleep the minimum time
435		 * and deduct from both.
436		 */
437		if (td->rate_pending_usleep[ddir] <=
438			td->rate_pending_usleep[odir]) {
439			usec = td->rate_pending_usleep[ddir];
440		} else {
441			usec = td->rate_pending_usleep[odir];
442			ddir = odir;
443		}
444	} else
445		usec = td->rate_pending_usleep[ddir];
446
447	/*
448	 * We are going to sleep, ensure that we flush anything pending as
449	 * not to skew our latency numbers.
450	 *
451	 * Changed to only monitor 'in flight' requests here instead of the
452	 * td->cur_depth, b/c td->cur_depth does not accurately represent
453	 * io's that have been actually submitted to an async engine,
454	 * and cur_depth is meaningless for sync engines.
455	 */
456	if (td->io_u_in_flight) {
457		int fio_unused ret;
458
459		ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
460	}
461
462	fio_gettime(&t, NULL);
463	usec_sleep(td, usec);
464	usec = utime_since_now(&t);
465
466	td->rate_pending_usleep[ddir] -= usec;
467
468	odir = ddir ^ 1;
469	if (td_rw(td) && __should_check_rate(td, odir))
470		td->rate_pending_usleep[odir] -= usec;
471
472	if (ddir_trim(ddir))
473		return ddir;
474	return ddir;
475}
476
477/*
478 * Return the data direction for the next io_u. If the job is a
479 * mixed read/write workload, check the rwmix cycle and switch if
480 * necessary.
481 */
482static enum fio_ddir get_rw_ddir(struct thread_data *td)
483{
484	enum fio_ddir ddir;
485
486	/*
487	 * see if it's time to fsync
488	 */
489	if (td->o.fsync_blocks &&
490	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
491	     td->io_issues[DDIR_WRITE] && should_fsync(td))
492		return DDIR_SYNC;
493
494	/*
495	 * see if it's time to fdatasync
496	 */
497	if (td->o.fdatasync_blocks &&
498	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
499	     td->io_issues[DDIR_WRITE] && should_fsync(td))
500		return DDIR_DATASYNC;
501
502	/*
503	 * see if it's time to sync_file_range
504	 */
505	if (td->sync_file_range_nr &&
506	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
507	     td->io_issues[DDIR_WRITE] && should_fsync(td))
508		return DDIR_SYNC_FILE_RANGE;
509
510	if (td_rw(td)) {
511		/*
512		 * Check if it's time to seed a new data direction.
513		 */
514		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
515			/*
516			 * Put a top limit on how many bytes we do for
517			 * one data direction, to avoid overflowing the
518			 * ranges too much
519			 */
520			ddir = get_rand_ddir(td);
521
522			if (ddir != td->rwmix_ddir)
523				set_rwmix_bytes(td);
524
525			td->rwmix_ddir = ddir;
526		}
527		ddir = td->rwmix_ddir;
528	} else if (td_read(td))
529		ddir = DDIR_READ;
530	else if (td_write(td))
531		ddir = DDIR_WRITE;
532	else
533		ddir = DDIR_TRIM;
534
535	td->rwmix_ddir = rate_ddir(td, ddir);
536	return td->rwmix_ddir;
537}
538
539static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
540{
541	io_u->ddir = get_rw_ddir(td);
542
543	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
544	    td->o.barrier_blocks &&
545	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
546	     td->io_issues[DDIR_WRITE])
547		io_u->flags |= IO_U_F_BARRIER;
548}
549
550void put_file_log(struct thread_data *td, struct fio_file *f)
551{
552	int ret = put_file(td, f);
553
554	if (ret)
555		td_verror(td, ret, "file close");
556}
557
558void put_io_u(struct thread_data *td, struct io_u *io_u)
559{
560	td_io_u_lock(td);
561
562	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
563		put_file_log(td, io_u->file);
564	io_u->file = NULL;
565	io_u->flags &= ~IO_U_F_FREE_DEF;
566	io_u->flags |= IO_U_F_FREE;
567
568	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
569		td->cur_depth--;
570	flist_del_init(&io_u->list);
571	flist_add(&io_u->list, &td->io_u_freelist);
572	td_io_u_unlock(td);
573	td_io_u_free_notify(td);
574}
575
576void clear_io_u(struct thread_data *td, struct io_u *io_u)
577{
578	io_u->flags &= ~IO_U_F_FLIGHT;
579	put_io_u(td, io_u);
580}
581
582void requeue_io_u(struct thread_data *td, struct io_u **io_u)
583{
584	struct io_u *__io_u = *io_u;
585
586	dprint(FD_IO, "requeue %p\n", __io_u);
587
588	td_io_u_lock(td);
589
590	__io_u->flags |= IO_U_F_FREE;
591	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
592		td->io_issues[__io_u->ddir]--;
593
594	__io_u->flags &= ~IO_U_F_FLIGHT;
595	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
596		td->cur_depth--;
597	flist_del(&__io_u->list);
598	flist_add_tail(&__io_u->list, &td->io_u_requeues);
599	td_io_u_unlock(td);
600	*io_u = NULL;
601}
602
603static int fill_io_u(struct thread_data *td, struct io_u *io_u)
604{
605	if (td->io_ops->flags & FIO_NOIO)
606		goto out;
607
608	set_rw_ddir(td, io_u);
609
610	/*
611	 * fsync() or fdatasync() or trim etc, we are done
612	 */
613	if (!ddir_rw(io_u->ddir))
614		goto out;
615
616	/*
617	 * See if it's time to switch to a new zone
618	 */
619	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
620		td->zone_bytes = 0;
621		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
622		io_u->file->last_pos = io_u->file->file_offset;
623		td->io_skip_bytes += td->o.zone_skip;
624	}
625
626	/*
627	 * No log, let the seq/rand engine retrieve the next buflen and
628	 * position.
629	 */
630	if (get_next_offset(td, io_u)) {
631		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
632		return 1;
633	}
634
635	io_u->buflen = get_next_buflen(td, io_u);
636	if (!io_u->buflen) {
637		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
638		return 1;
639	}
640
641	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
642		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
643		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
644				io_u->buflen, io_u->file->real_file_size);
645		return 1;
646	}
647
648	/*
649	 * mark entry before potentially trimming io_u
650	 */
651	if (td_random(td) && file_randommap(td, io_u->file))
652		mark_random_map(td, io_u);
653
654	/*
655	 * If using a write iolog, store this entry.
656	 */
657out:
658	dprint_io_u(io_u, "fill_io_u");
659	td->zone_bytes += io_u->buflen;
660	log_io_u(td, io_u);
661	return 0;
662}
663
664static void __io_u_mark_map(unsigned int *map, unsigned int nr)
665{
666	int idx = 0;
667
668	switch (nr) {
669	default:
670		idx = 6;
671		break;
672	case 33 ... 64:
673		idx = 5;
674		break;
675	case 17 ... 32:
676		idx = 4;
677		break;
678	case 9 ... 16:
679		idx = 3;
680		break;
681	case 5 ... 8:
682		idx = 2;
683		break;
684	case 1 ... 4:
685		idx = 1;
686	case 0:
687		break;
688	}
689
690	map[idx]++;
691}
692
693void io_u_mark_submit(struct thread_data *td, unsigned int nr)
694{
695	__io_u_mark_map(td->ts.io_u_submit, nr);
696	td->ts.total_submit++;
697}
698
699void io_u_mark_complete(struct thread_data *td, unsigned int nr)
700{
701	__io_u_mark_map(td->ts.io_u_complete, nr);
702	td->ts.total_complete++;
703}
704
705void io_u_mark_depth(struct thread_data *td, unsigned int nr)
706{
707	int idx = 0;
708
709	switch (td->cur_depth) {
710	default:
711		idx = 6;
712		break;
713	case 32 ... 63:
714		idx = 5;
715		break;
716	case 16 ... 31:
717		idx = 4;
718		break;
719	case 8 ... 15:
720		idx = 3;
721		break;
722	case 4 ... 7:
723		idx = 2;
724		break;
725	case 2 ... 3:
726		idx = 1;
727	case 1:
728		break;
729	}
730
731	td->ts.io_u_map[idx] += nr;
732}
733
734static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
735{
736	int idx = 0;
737
738	assert(usec < 1000);
739
740	switch (usec) {
741	case 750 ... 999:
742		idx = 9;
743		break;
744	case 500 ... 749:
745		idx = 8;
746		break;
747	case 250 ... 499:
748		idx = 7;
749		break;
750	case 100 ... 249:
751		idx = 6;
752		break;
753	case 50 ... 99:
754		idx = 5;
755		break;
756	case 20 ... 49:
757		idx = 4;
758		break;
759	case 10 ... 19:
760		idx = 3;
761		break;
762	case 4 ... 9:
763		idx = 2;
764		break;
765	case 2 ... 3:
766		idx = 1;
767	case 0 ... 1:
768		break;
769	}
770
771	assert(idx < FIO_IO_U_LAT_U_NR);
772	td->ts.io_u_lat_u[idx]++;
773}
774
775static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
776{
777	int idx = 0;
778
779	switch (msec) {
780	default:
781		idx = 11;
782		break;
783	case 1000 ... 1999:
784		idx = 10;
785		break;
786	case 750 ... 999:
787		idx = 9;
788		break;
789	case 500 ... 749:
790		idx = 8;
791		break;
792	case 250 ... 499:
793		idx = 7;
794		break;
795	case 100 ... 249:
796		idx = 6;
797		break;
798	case 50 ... 99:
799		idx = 5;
800		break;
801	case 20 ... 49:
802		idx = 4;
803		break;
804	case 10 ... 19:
805		idx = 3;
806		break;
807	case 4 ... 9:
808		idx = 2;
809		break;
810	case 2 ... 3:
811		idx = 1;
812	case 0 ... 1:
813		break;
814	}
815
816	assert(idx < FIO_IO_U_LAT_M_NR);
817	td->ts.io_u_lat_m[idx]++;
818}
819
820static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
821{
822	if (usec < 1000)
823		io_u_mark_lat_usec(td, usec);
824	else
825		io_u_mark_lat_msec(td, usec / 1000);
826}
827
828/*
829 * Get next file to service by choosing one at random
830 */
831static struct fio_file *get_next_file_rand(struct thread_data *td,
832					   enum fio_file_flags goodf,
833					   enum fio_file_flags badf)
834{
835	struct fio_file *f;
836	int fno;
837
838	do {
839		int opened = 0;
840		unsigned long r;
841
842		if (td->o.use_os_rand) {
843			r = os_random_long(&td->next_file_state);
844			fno = (unsigned int) ((double) td->o.nr_files
845				* (r / (OS_RAND_MAX + 1.0)));
846		} else {
847			r = __rand(&td->__next_file_state);
848			fno = (unsigned int) ((double) td->o.nr_files
849				* (r / (FRAND_MAX + 1.0)));
850		}
851
852		f = td->files[fno];
853		if (fio_file_done(f))
854			continue;
855
856		if (!fio_file_open(f)) {
857			int err;
858
859			err = td_io_open_file(td, f);
860			if (err)
861				continue;
862			opened = 1;
863		}
864
865		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
866			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
867			return f;
868		}
869		if (opened)
870			td_io_close_file(td, f);
871	} while (1);
872}
873
874/*
875 * Get next file to service by doing round robin between all available ones
876 */
877static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
878					 int badf)
879{
880	unsigned int old_next_file = td->next_file;
881	struct fio_file *f;
882
883	do {
884		int opened = 0;
885
886		f = td->files[td->next_file];
887
888		td->next_file++;
889		if (td->next_file >= td->o.nr_files)
890			td->next_file = 0;
891
892		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
893		if (fio_file_done(f)) {
894			f = NULL;
895			continue;
896		}
897
898		if (!fio_file_open(f)) {
899			int err;
900
901			err = td_io_open_file(td, f);
902			if (err) {
903				dprint(FD_FILE, "error %d on open of %s\n",
904					err, f->file_name);
905				f = NULL;
906				continue;
907			}
908			opened = 1;
909		}
910
911		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
912								f->flags);
913		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
914			break;
915
916		if (opened)
917			td_io_close_file(td, f);
918
919		f = NULL;
920	} while (td->next_file != old_next_file);
921
922	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
923	return f;
924}
925
926static struct fio_file *__get_next_file(struct thread_data *td)
927{
928	struct fio_file *f;
929
930	assert(td->o.nr_files <= td->files_index);
931
932	if (td->nr_done_files >= td->o.nr_files) {
933		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
934				" nr_files=%d\n", td->nr_open_files,
935						  td->nr_done_files,
936						  td->o.nr_files);
937		return NULL;
938	}
939
940	f = td->file_service_file;
941	if (f && fio_file_open(f) && !fio_file_closing(f)) {
942		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
943			goto out;
944		if (td->file_service_left--)
945			goto out;
946	}
947
948	if (td->o.file_service_type == FIO_FSERVICE_RR ||
949	    td->o.file_service_type == FIO_FSERVICE_SEQ)
950		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
951	else
952		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
953
954	td->file_service_file = f;
955	td->file_service_left = td->file_service_nr - 1;
956out:
957	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
958	return f;
959}
960
961static struct fio_file *get_next_file(struct thread_data *td)
962{
963	struct prof_io_ops *ops = &td->prof_io_ops;
964
965	if (ops->get_next_file)
966		return ops->get_next_file(td);
967
968	return __get_next_file(td);
969}
970
971static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
972{
973	struct fio_file *f;
974
975	do {
976		f = get_next_file(td);
977		if (!f)
978			return 1;
979
980		io_u->file = f;
981		get_file(f);
982
983		if (!fill_io_u(td, io_u))
984			break;
985
986		put_file_log(td, f);
987		td_io_close_file(td, f);
988		io_u->file = NULL;
989		fio_file_set_done(f);
990		td->nr_done_files++;
991		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
992					td->nr_done_files, td->o.nr_files);
993	} while (1);
994
995	return 0;
996}
997
998
999struct io_u *__get_io_u(struct thread_data *td)
1000{
1001	struct io_u *io_u = NULL;
1002
1003	td_io_u_lock(td);
1004
1005again:
1006	if (!flist_empty(&td->io_u_requeues))
1007		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1008	else if (!queue_full(td)) {
1009		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1010
1011		io_u->buflen = 0;
1012		io_u->resid = 0;
1013		io_u->file = NULL;
1014		io_u->end_io = NULL;
1015	}
1016
1017	if (io_u) {
1018		assert(io_u->flags & IO_U_F_FREE);
1019		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1020		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1021		io_u->flags &= ~IO_U_F_VER_LIST;
1022
1023		io_u->error = 0;
1024		flist_del(&io_u->list);
1025		flist_add_tail(&io_u->list, &td->io_u_busylist);
1026		td->cur_depth++;
1027		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1028	} else if (td->o.verify_async) {
1029		/*
1030		 * We ran out, wait for async verify threads to finish and
1031		 * return one
1032		 */
1033		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1034		goto again;
1035	}
1036
1037	td_io_u_unlock(td);
1038	return io_u;
1039}
1040
1041static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1042{
1043	if (td->o.trim_backlog && td->trim_entries) {
1044		int get_trim = 0;
1045
1046		if (td->trim_batch) {
1047			td->trim_batch--;
1048			get_trim = 1;
1049		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1050			 td->last_ddir != DDIR_READ) {
1051			td->trim_batch = td->o.trim_batch;
1052			if (!td->trim_batch)
1053				td->trim_batch = td->o.trim_backlog;
1054			get_trim = 1;
1055		}
1056
1057		if (get_trim && !get_next_trim(td, io_u))
1058			return 1;
1059	}
1060
1061	return 0;
1062}
1063
1064static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1065{
1066	if (td->o.verify_backlog && td->io_hist_len) {
1067		int get_verify = 0;
1068
1069		if (td->verify_batch)
1070			get_verify = 1;
1071		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1072			 td->last_ddir != DDIR_READ) {
1073			td->verify_batch = td->o.verify_batch;
1074			if (!td->verify_batch)
1075				td->verify_batch = td->o.verify_backlog;
1076			get_verify = 1;
1077		}
1078
1079		if (get_verify && !get_next_verify(td, io_u)) {
1080			td->verify_batch--;
1081			return 1;
1082		}
1083	}
1084
1085	return 0;
1086}
1087
1088/*
1089 * Fill offset and start time into the buffer content, to prevent too
1090 * easy compressible data for simple de-dupe attempts. Do this for every
1091 * 512b block in the range, since that should be the smallest block size
1092 * we can expect from a device.
1093 */
1094static void small_content_scramble(struct io_u *io_u)
1095{
1096	unsigned int i, nr_blocks = io_u->buflen / 512;
1097	unsigned long long boffset;
1098	unsigned int offset;
1099	void *p, *end;
1100
1101	if (!nr_blocks)
1102		return;
1103
1104	p = io_u->xfer_buf;
1105	boffset = io_u->offset;
1106	io_u->buf_filled_len = 0;
1107
1108	for (i = 0; i < nr_blocks; i++) {
1109		/*
1110		 * Fill the byte offset into a "random" start offset of
1111		 * the buffer, given by the product of the usec time
1112		 * and the actual offset.
1113		 */
1114		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1115		offset &= ~(sizeof(unsigned long long) - 1);
1116		if (offset >= 512 - sizeof(unsigned long long))
1117			offset -= sizeof(unsigned long long);
1118		memcpy(p + offset, &boffset, sizeof(boffset));
1119
1120		end = p + 512 - sizeof(io_u->start_time);
1121		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1122		p += 512;
1123		boffset += 512;
1124	}
1125}
1126
1127/*
1128 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1129 * etc. The returned io_u is fully ready to be prepped and submitted.
1130 */
1131struct io_u *get_io_u(struct thread_data *td)
1132{
1133	struct fio_file *f;
1134	struct io_u *io_u;
1135	int do_scramble = 0;
1136
1137	io_u = __get_io_u(td);
1138	if (!io_u) {
1139		dprint(FD_IO, "__get_io_u failed\n");
1140		return NULL;
1141	}
1142
1143	if (check_get_verify(td, io_u))
1144		goto out;
1145	if (check_get_trim(td, io_u))
1146		goto out;
1147
1148	/*
1149	 * from a requeue, io_u already setup
1150	 */
1151	if (io_u->file)
1152		goto out;
1153
1154	/*
1155	 * If using an iolog, grab next piece if any available.
1156	 */
1157	if (td->o.read_iolog_file) {
1158		if (read_iolog_get(td, io_u))
1159			goto err_put;
1160	} else if (set_io_u_file(td, io_u)) {
1161		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1162		goto err_put;
1163	}
1164
1165	f = io_u->file;
1166	assert(fio_file_open(f));
1167
1168	if (ddir_rw(io_u->ddir)) {
1169		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1170			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1171			goto err_put;
1172		}
1173
1174		f->last_start = io_u->offset;
1175		f->last_pos = io_u->offset + io_u->buflen;
1176
1177		if (io_u->ddir == DDIR_WRITE) {
1178			if (td->o.refill_buffers) {
1179				io_u_fill_buffer(td, io_u,
1180					io_u->xfer_buflen, io_u->xfer_buflen);
1181			} else if (td->o.scramble_buffers)
1182				do_scramble = 1;
1183			if (td->o.verify != VERIFY_NONE) {
1184				populate_verify_io_u(td, io_u);
1185				do_scramble = 0;
1186			}
1187		} else if (io_u->ddir == DDIR_READ) {
1188			/*
1189			 * Reset the buf_filled parameters so next time if the
1190			 * buffer is used for writes it is refilled.
1191			 */
1192			io_u->buf_filled_len = 0;
1193		}
1194	}
1195
1196	/*
1197	 * Set io data pointers.
1198	 */
1199	io_u->xfer_buf = io_u->buf;
1200	io_u->xfer_buflen = io_u->buflen;
1201
1202out:
1203	assert(io_u->file);
1204	if (!td_io_prep(td, io_u)) {
1205		if (!td->o.disable_slat)
1206			fio_gettime(&io_u->start_time, NULL);
1207		if (do_scramble)
1208			small_content_scramble(io_u);
1209		return io_u;
1210	}
1211err_put:
1212	dprint(FD_IO, "get_io_u failed\n");
1213	put_io_u(td, io_u);
1214	return NULL;
1215}
1216
1217void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1218{
1219	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1220	const char *msg[] = { "read", "write", "sync", "datasync",
1221				"sync_file_range", "wait", "trim" };
1222
1223	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1224		return;
1225
1226	log_err("fio: io_u error");
1227
1228	if (io_u->file)
1229		log_err(" on file %s", io_u->file->file_name);
1230
1231	log_err(": %s\n", strerror(io_u->error));
1232
1233	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1234					io_u->offset, io_u->xfer_buflen);
1235
1236	if (!td->error)
1237		td_verror(td, io_u->error, "io_u error");
1238}
1239
1240static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1241				  struct io_completion_data *icd,
1242				  const enum fio_ddir idx, unsigned int bytes)
1243{
1244	unsigned long lusec = 0;
1245
1246	if (!td->o.disable_clat || !td->o.disable_bw)
1247		lusec = utime_since(&io_u->issue_time, &icd->time);
1248
1249	if (!td->o.disable_lat) {
1250		unsigned long tusec;
1251
1252		tusec = utime_since(&io_u->start_time, &icd->time);
1253		add_lat_sample(td, idx, tusec, bytes);
1254
1255		if (td->o.max_latency && tusec > td->o.max_latency) {
1256			if (!td->error)
1257				log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1258			td_verror(td, ETIMEDOUT, "max latency exceeded");
1259			icd->error = ETIMEDOUT;
1260		}
1261	}
1262
1263	if (!td->o.disable_clat) {
1264		add_clat_sample(td, idx, lusec, bytes);
1265		io_u_mark_latency(td, lusec);
1266	}
1267
1268	if (!td->o.disable_bw)
1269		add_bw_sample(td, idx, bytes, &icd->time);
1270
1271	add_iops_sample(td, idx, &icd->time);
1272}
1273
1274static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1275{
1276	unsigned long long secs, remainder, bps, bytes;
1277	bytes = td->this_io_bytes[ddir];
1278	bps = td->rate_bps[ddir];
1279	secs = bytes / bps;
1280	remainder = bytes % bps;
1281	return remainder * 1000000 / bps + secs * 1000000;
1282}
1283
1284static void io_completed(struct thread_data *td, struct io_u *io_u,
1285			 struct io_completion_data *icd)
1286{
1287	struct fio_file *f;
1288
1289	dprint_io_u(io_u, "io complete");
1290
1291	td_io_u_lock(td);
1292	assert(io_u->flags & IO_U_F_FLIGHT);
1293	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1294	td_io_u_unlock(td);
1295
1296	if (ddir_sync(io_u->ddir)) {
1297		td->last_was_sync = 1;
1298		f = io_u->file;
1299		if (f) {
1300			f->first_write = -1ULL;
1301			f->last_write = -1ULL;
1302		}
1303		return;
1304	}
1305
1306	td->last_was_sync = 0;
1307	td->last_ddir = io_u->ddir;
1308
1309	if (!io_u->error && ddir_rw(io_u->ddir)) {
1310		unsigned int bytes = io_u->buflen - io_u->resid;
1311		const enum fio_ddir idx = io_u->ddir;
1312		const enum fio_ddir odx = io_u->ddir ^ 1;
1313		int ret;
1314
1315		td->io_blocks[idx]++;
1316		td->this_io_blocks[idx]++;
1317		td->io_bytes[idx] += bytes;
1318
1319		if (!(io_u->flags & IO_U_F_VER_LIST))
1320			td->this_io_bytes[idx] += bytes;
1321
1322		if (idx == DDIR_WRITE) {
1323			f = io_u->file;
1324			if (f) {
1325				if (f->first_write == -1ULL ||
1326				    io_u->offset < f->first_write)
1327					f->first_write = io_u->offset;
1328				if (f->last_write == -1ULL ||
1329				    ((io_u->offset + bytes) > f->last_write))
1330					f->last_write = io_u->offset + bytes;
1331			}
1332		}
1333
1334		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1335					   td->runstate == TD_VERIFYING)) {
1336			account_io_completion(td, io_u, icd, idx, bytes);
1337
1338			if (__should_check_rate(td, idx)) {
1339				td->rate_pending_usleep[idx] =
1340					(usec_for_io(td, idx) -
1341					 utime_since_now(&td->start));
1342			}
1343			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1344				td->rate_pending_usleep[odx] =
1345					(usec_for_io(td, odx) -
1346					 utime_since_now(&td->start));
1347		}
1348
1349		if (td_write(td) && idx == DDIR_WRITE &&
1350		    td->o.do_verify &&
1351		    td->o.verify != VERIFY_NONE)
1352			log_io_piece(td, io_u);
1353
1354		icd->bytes_done[idx] += bytes;
1355
1356		if (io_u->end_io) {
1357			ret = io_u->end_io(td, io_u);
1358			if (ret && !icd->error)
1359				icd->error = ret;
1360		}
1361	} else if (io_u->error) {
1362		icd->error = io_u->error;
1363		io_u_log_error(td, io_u);
1364	}
1365	if (icd->error) {
1366		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1367		if (!td_non_fatal_error(td, eb, icd->error))
1368			return;
1369		/*
1370		 * If there is a non_fatal error, then add to the error count
1371		 * and clear all the errors.
1372		 */
1373		update_error_count(td, icd->error);
1374		td_clear_error(td);
1375		icd->error = 0;
1376		io_u->error = 0;
1377	}
1378}
1379
1380static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1381		     int nr)
1382{
1383	int ddir;
1384	if (!td->o.disable_clat || !td->o.disable_bw)
1385		fio_gettime(&icd->time, NULL);
1386
1387	icd->nr = nr;
1388
1389	icd->error = 0;
1390	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1391		icd->bytes_done[ddir] = 0;
1392}
1393
1394static void ios_completed(struct thread_data *td,
1395			  struct io_completion_data *icd)
1396{
1397	struct io_u *io_u;
1398	int i;
1399
1400	for (i = 0; i < icd->nr; i++) {
1401		io_u = td->io_ops->event(td, i);
1402
1403		io_completed(td, io_u, icd);
1404
1405		if (!(io_u->flags & IO_U_F_FREE_DEF))
1406			put_io_u(td, io_u);
1407	}
1408}
1409
1410/*
1411 * Complete a single io_u for the sync engines.
1412 */
1413int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1414		       unsigned long *bytes)
1415{
1416	struct io_completion_data icd;
1417
1418	init_icd(td, &icd, 1);
1419	io_completed(td, io_u, &icd);
1420
1421	if (!(io_u->flags & IO_U_F_FREE_DEF))
1422		put_io_u(td, io_u);
1423
1424	if (icd.error) {
1425		td_verror(td, icd.error, "io_u_sync_complete");
1426		return -1;
1427	}
1428
1429	if (bytes) {
1430		int ddir;
1431
1432		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1433			bytes[ddir] += icd.bytes_done[ddir];
1434	}
1435
1436	return 0;
1437}
1438
1439/*
1440 * Called to complete min_events number of io for the async engines.
1441 */
1442int io_u_queued_complete(struct thread_data *td, int min_evts,
1443			 unsigned long *bytes)
1444{
1445	struct io_completion_data icd;
1446	struct timespec *tvp = NULL;
1447	int ret;
1448	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1449
1450	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1451
1452	if (!min_evts)
1453		tvp = &ts;
1454
1455	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1456	if (ret < 0) {
1457		td_verror(td, -ret, "td_io_getevents");
1458		return ret;
1459	} else if (!ret)
1460		return ret;
1461
1462	init_icd(td, &icd, ret);
1463	ios_completed(td, &icd);
1464	if (icd.error) {
1465		td_verror(td, icd.error, "io_u_queued_complete");
1466		return -1;
1467	}
1468
1469	if (bytes) {
1470		int ddir;
1471
1472		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1473			bytes[ddir] += icd.bytes_done[ddir];
1474	}
1475
1476	return 0;
1477}
1478
1479/*
1480 * Call when io_u is really queued, to update the submission latency.
1481 */
1482void io_u_queued(struct thread_data *td, struct io_u *io_u)
1483{
1484	if (!td->o.disable_slat) {
1485		unsigned long slat_time;
1486
1487		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1488		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1489	}
1490}
1491
1492/*
1493 * "randomly" fill the buffer contents
1494 */
1495void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1496		      unsigned int min_write, unsigned int max_bs)
1497{
1498	io_u->buf_filled_len = 0;
1499
1500	if (!td->o.zero_buffers) {
1501		unsigned int perc = td->o.compress_percentage;
1502
1503		if (perc) {
1504			unsigned int seg = min_write;
1505
1506			seg = min(min_write, td->o.compress_chunk);
1507			fill_random_buf_percentage(&td->buf_state, io_u->buf,
1508						perc, seg, max_bs);
1509		} else
1510			fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1511	} else
1512		memset(io_u->buf, 0, max_bs);
1513}
1514