io_u.c revision 068420271828b3b2426ffc3ccf64404cb9d340fb
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	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
118	if (!max_blocks)
119		return 0;
120
121	return max_blocks;
122}
123
124/*
125 * Return the next free block in the map.
126 */
127static int get_next_free_block(struct thread_data *td, struct fio_file *f,
128			       enum fio_ddir ddir, unsigned long long *b)
129{
130	unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
131	int i;
132
133	lastb = last_block(td, f, ddir);
134	if (!lastb)
135		return 1;
136
137	i = f->last_free_lookup;
138	block = i * BLOCKS_PER_MAP;
139	while (block * min_bs < f->real_file_size &&
140		block * min_bs < f->io_size) {
141		if (f->file_map[i] != -1UL) {
142			block += ffz(f->file_map[i]);
143			if (block > lastb)
144				break;
145			f->last_free_lookup = i;
146			*b = block;
147			return 0;
148		}
149
150		block += BLOCKS_PER_MAP;
151		i++;
152	}
153
154	dprint(FD_IO, "failed finding a free block\n");
155	return 1;
156}
157
158static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
159				enum fio_ddir ddir, unsigned long long *b)
160{
161	unsigned long long rmax, r, lastb;
162	int loops = 5;
163
164	lastb = last_block(td, f, ddir);
165	if (!lastb)
166		return 1;
167
168	if (f->failed_rands >= 200)
169		goto ffz;
170
171	rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
172	do {
173		if (td->o.use_os_rand)
174			r = os_random_long(&td->random_state);
175		else
176			r = __rand(&td->__random_state);
177
178		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
179
180		dprint(FD_RANDOM, "off rand %llu\n", r);
181
182
183		/*
184		 * if we are not maintaining a random map, we are done.
185		 */
186		if (!file_randommap(td, f))
187			goto ret_good;
188
189		/*
190		 * calculate map offset and check if it's free
191		 */
192		if (random_map_free(f, *b))
193			goto ret_good;
194
195		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
196									*b);
197	} while (--loops);
198
199	if (!f->failed_rands++)
200		f->last_free_lookup = 0;
201
202	/*
203	 * we get here, if we didn't suceed in looking up a block. generate
204	 * a random start offset into the filemap, and find the first free
205	 * block from there.
206	 */
207	loops = 10;
208	do {
209		f->last_free_lookup = (f->num_maps - 1) *
210					(r / ((unsigned long long) rmax + 1.0));
211		if (!get_next_free_block(td, f, ddir, b))
212			goto ret;
213
214		if (td->o.use_os_rand)
215			r = os_random_long(&td->random_state);
216		else
217			r = __rand(&td->__random_state);
218	} while (--loops);
219
220	/*
221	 * that didn't work either, try exhaustive search from the start
222	 */
223	f->last_free_lookup = 0;
224ffz:
225	if (!get_next_free_block(td, f, ddir, b))
226		return 0;
227	f->last_free_lookup = 0;
228	return get_next_free_block(td, f, ddir, b);
229ret_good:
230	f->failed_rands = 0;
231ret:
232	return 0;
233}
234
235static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
236			       enum fio_ddir ddir, unsigned long long *b)
237{
238	if (get_next_rand_offset(td, f, ddir, b)) {
239		dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
240				f->file_name, f->last_pos, f->real_file_size);
241		return 1;
242	}
243
244	return 0;
245}
246
247static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
248			      enum fio_ddir ddir, unsigned long long *b)
249{
250	assert(ddir_rw(ddir));
251
252	if (f->last_pos < f->real_file_size) {
253		unsigned long long pos;
254
255		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
256			f->last_pos = f->real_file_size;
257
258		pos = f->last_pos - f->file_offset;
259		if (pos)
260			pos += td->o.ddir_seq_add;
261
262		*b = pos / td->o.min_bs[ddir];
263		return 0;
264	}
265
266	return 1;
267}
268
269static int get_next_block(struct thread_data *td, struct io_u *io_u,
270			  enum fio_ddir ddir, int rw_seq, unsigned long long *b)
271{
272	struct fio_file *f = io_u->file;
273	int ret;
274
275	assert(ddir_rw(ddir));
276
277	if (rw_seq) {
278		if (td_random(td))
279			ret = get_next_rand_block(td, f, ddir, b);
280		else
281			ret = get_next_seq_block(td, f, ddir, b);
282	} else {
283		io_u->flags |= IO_U_F_BUSY_OK;
284
285		if (td->o.rw_seq == RW_SEQ_SEQ) {
286			ret = get_next_seq_block(td, f, ddir, b);
287			if (ret)
288				ret = get_next_rand_block(td, f, ddir, b);
289		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
290			if (f->last_start != -1ULL)
291				*b = (f->last_start - f->file_offset)
292					/ td->o.min_bs[ddir];
293			else
294				*b = 0;
295			ret = 0;
296		} else {
297			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
298			ret = 1;
299		}
300	}
301
302	return ret;
303}
304
305/*
306 * For random io, generate a random new block and see if it's used. Repeat
307 * until we find a free one. For sequential io, just return the end of
308 * the last io issued.
309 */
310static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
311{
312	struct fio_file *f = io_u->file;
313	unsigned long long b;
314	enum fio_ddir ddir = io_u->ddir;
315	int rw_seq_hit = 0;
316
317	assert(ddir_rw(ddir));
318
319	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
320		rw_seq_hit = 1;
321		td->ddir_seq_nr = td->o.ddir_seq_nr;
322	}
323
324	if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
325		return 1;
326
327	io_u->offset = b * td->o.ba[ddir];
328	if (io_u->offset >= f->io_size) {
329		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
330					io_u->offset, f->io_size);
331		return 1;
332	}
333
334	io_u->offset += f->file_offset;
335	if (io_u->offset >= f->real_file_size) {
336		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
337					io_u->offset, f->real_file_size);
338		return 1;
339	}
340
341	return 0;
342}
343
344static int get_next_offset(struct thread_data *td, struct io_u *io_u)
345{
346	struct prof_io_ops *ops = &td->prof_io_ops;
347
348	if (ops->fill_io_u_off)
349		return ops->fill_io_u_off(td, io_u);
350
351	return __get_next_offset(td, io_u);
352}
353
354static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
355			    unsigned int buflen)
356{
357	struct fio_file *f = io_u->file;
358
359	return io_u->offset + buflen <= f->io_size + td->o.start_offset;
360}
361
362static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
363{
364	const int ddir = io_u->ddir;
365	unsigned int uninitialized_var(buflen);
366	unsigned int minbs, maxbs;
367	unsigned long r, rand_max;
368
369	assert(ddir_rw(ddir));
370
371	minbs = td->o.min_bs[ddir];
372	maxbs = td->o.max_bs[ddir];
373
374	if (minbs == maxbs)
375		return minbs;
376
377	if (td->o.use_os_rand)
378		rand_max = OS_RAND_MAX;
379	else
380		rand_max = FRAND_MAX;
381
382	do {
383		if (td->o.use_os_rand)
384			r = os_random_long(&td->bsrange_state);
385		else
386			r = __rand(&td->__bsrange_state);
387
388		if (!td->o.bssplit_nr[ddir]) {
389			buflen = 1 + (unsigned int) ((double) maxbs *
390					(r / (rand_max + 1.0)));
391			if (buflen < minbs)
392				buflen = minbs;
393		} else {
394			long perc = 0;
395			unsigned int i;
396
397			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
398				struct bssplit *bsp = &td->o.bssplit[ddir][i];
399
400				buflen = bsp->bs;
401				perc += bsp->perc;
402				if ((r <= ((rand_max / 100L) * perc)) &&
403				    io_u_fits(td, io_u, buflen))
404					break;
405			}
406		}
407
408		if (!td->o.bs_unaligned && is_power_of_2(minbs))
409			buflen = (buflen + minbs - 1) & ~(minbs - 1);
410
411	} while (!io_u_fits(td, io_u, buflen));
412
413	return buflen;
414}
415
416static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
417{
418	struct prof_io_ops *ops = &td->prof_io_ops;
419
420	if (ops->fill_io_u_size)
421		return ops->fill_io_u_size(td, io_u);
422
423	return __get_next_buflen(td, io_u);
424}
425
426static void set_rwmix_bytes(struct thread_data *td)
427{
428	unsigned int diff;
429
430	/*
431	 * we do time or byte based switch. this is needed because
432	 * buffered writes may issue a lot quicker than they complete,
433	 * whereas reads do not.
434	 */
435	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
436	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
437}
438
439static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
440{
441	unsigned int v;
442	unsigned long r;
443
444	if (td->o.use_os_rand) {
445		r = os_random_long(&td->rwmix_state);
446		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
447	} else {
448		r = __rand(&td->__rwmix_state);
449		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
450	}
451
452	if (v <= td->o.rwmix[DDIR_READ])
453		return DDIR_READ;
454
455	return DDIR_WRITE;
456}
457
458static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
459{
460	enum fio_ddir odir = ddir ^ 1;
461	struct timeval t;
462	long usec;
463
464	assert(ddir_rw(ddir));
465
466	if (td->rate_pending_usleep[ddir] <= 0)
467		return ddir;
468
469	/*
470	 * We have too much pending sleep in this direction. See if we
471	 * should switch.
472	 */
473	if (td_rw(td)) {
474		/*
475		 * Other direction does not have too much pending, switch
476		 */
477		if (td->rate_pending_usleep[odir] < 100000)
478			return odir;
479
480		/*
481		 * Both directions have pending sleep. Sleep the minimum time
482		 * and deduct from both.
483		 */
484		if (td->rate_pending_usleep[ddir] <=
485			td->rate_pending_usleep[odir]) {
486			usec = td->rate_pending_usleep[ddir];
487		} else {
488			usec = td->rate_pending_usleep[odir];
489			ddir = odir;
490		}
491	} else
492		usec = td->rate_pending_usleep[ddir];
493
494	/*
495	 * We are going to sleep, ensure that we flush anything pending as
496	 * not to skew our latency numbers
497	 */
498	if (td->cur_depth) {
499		int fio_unused ret;
500
501		ret = io_u_queued_complete(td, td->cur_depth, NULL);
502	}
503
504	fio_gettime(&t, NULL);
505	usec_sleep(td, usec);
506	usec = utime_since_now(&t);
507
508	td->rate_pending_usleep[ddir] -= usec;
509
510	odir = ddir ^ 1;
511	if (td_rw(td) && __should_check_rate(td, odir))
512		td->rate_pending_usleep[odir] -= usec;
513
514	return ddir;
515}
516
517/*
518 * Return the data direction for the next io_u. If the job is a
519 * mixed read/write workload, check the rwmix cycle and switch if
520 * necessary.
521 */
522static enum fio_ddir get_rw_ddir(struct thread_data *td)
523{
524	enum fio_ddir ddir;
525
526	/*
527	 * see if it's time to fsync
528	 */
529	if (td->o.fsync_blocks &&
530	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
531	     td->io_issues[DDIR_WRITE] && should_fsync(td))
532		return DDIR_SYNC;
533
534	/*
535	 * see if it's time to fdatasync
536	 */
537	if (td->o.fdatasync_blocks &&
538	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
539	     td->io_issues[DDIR_WRITE] && should_fsync(td))
540		return DDIR_DATASYNC;
541
542	/*
543	 * see if it's time to sync_file_range
544	 */
545	if (td->sync_file_range_nr &&
546	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
547	     td->io_issues[DDIR_WRITE] && should_fsync(td))
548		return DDIR_SYNC_FILE_RANGE;
549
550	if (td_rw(td)) {
551		/*
552		 * Check if it's time to seed a new data direction.
553		 */
554		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
555			/*
556			 * Put a top limit on how many bytes we do for
557			 * one data direction, to avoid overflowing the
558			 * ranges too much
559			 */
560			ddir = get_rand_ddir(td);
561
562			if (ddir != td->rwmix_ddir)
563				set_rwmix_bytes(td);
564
565			td->rwmix_ddir = ddir;
566		}
567		ddir = td->rwmix_ddir;
568	} else if (td_read(td))
569		ddir = DDIR_READ;
570	else
571		ddir = DDIR_WRITE;
572
573	td->rwmix_ddir = rate_ddir(td, ddir);
574	return td->rwmix_ddir;
575}
576
577static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
578{
579	io_u->ddir = get_rw_ddir(td);
580
581	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
582	    td->o.barrier_blocks &&
583	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
584	     td->io_issues[DDIR_WRITE])
585		io_u->flags |= IO_U_F_BARRIER;
586}
587
588void put_file_log(struct thread_data *td, struct fio_file *f)
589{
590	int ret = put_file(td, f);
591
592	if (ret)
593		td_verror(td, ret, "file close");
594}
595
596void put_io_u(struct thread_data *td, struct io_u *io_u)
597{
598	td_io_u_lock(td);
599
600	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
601		put_file_log(td, io_u->file);
602	io_u->file = NULL;
603	io_u->flags &= ~IO_U_F_FREE_DEF;
604	io_u->flags |= IO_U_F_FREE;
605
606	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
607		td->cur_depth--;
608	flist_del_init(&io_u->list);
609	flist_add(&io_u->list, &td->io_u_freelist);
610	td_io_u_unlock(td);
611	td_io_u_free_notify(td);
612}
613
614void clear_io_u(struct thread_data *td, struct io_u *io_u)
615{
616	io_u->flags &= ~IO_U_F_FLIGHT;
617	put_io_u(td, io_u);
618}
619
620void requeue_io_u(struct thread_data *td, struct io_u **io_u)
621{
622	struct io_u *__io_u = *io_u;
623
624	dprint(FD_IO, "requeue %p\n", __io_u);
625
626	td_io_u_lock(td);
627
628	__io_u->flags |= IO_U_F_FREE;
629	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
630		td->io_issues[__io_u->ddir]--;
631
632	__io_u->flags &= ~IO_U_F_FLIGHT;
633	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
634		td->cur_depth--;
635	flist_del(&__io_u->list);
636	flist_add_tail(&__io_u->list, &td->io_u_requeues);
637	td_io_u_unlock(td);
638	*io_u = NULL;
639}
640
641static int fill_io_u(struct thread_data *td, struct io_u *io_u)
642{
643	if (td->io_ops->flags & FIO_NOIO)
644		goto out;
645
646	set_rw_ddir(td, io_u);
647
648	/*
649	 * fsync() or fdatasync() or trim etc, we are done
650	 */
651	if (!ddir_rw(io_u->ddir))
652		goto out;
653
654	/*
655	 * See if it's time to switch to a new zone
656	 */
657	if (td->zone_bytes >= td->o.zone_size) {
658		td->zone_bytes = 0;
659		io_u->file->last_pos += td->o.zone_skip;
660		td->io_skip_bytes += td->o.zone_skip;
661	}
662
663	/*
664	 * No log, let the seq/rand engine retrieve the next buflen and
665	 * position.
666	 */
667	if (get_next_offset(td, io_u)) {
668		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
669		return 1;
670	}
671
672	io_u->buflen = get_next_buflen(td, io_u);
673	if (!io_u->buflen) {
674		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
675		return 1;
676	}
677
678	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
679		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
680		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
681				io_u->buflen, io_u->file->real_file_size);
682		return 1;
683	}
684
685	/*
686	 * mark entry before potentially trimming io_u
687	 */
688	if (td_random(td) && file_randommap(td, io_u->file))
689		mark_random_map(td, io_u);
690
691	/*
692	 * If using a write iolog, store this entry.
693	 */
694out:
695	dprint_io_u(io_u, "fill_io_u");
696	td->zone_bytes += io_u->buflen;
697	log_io_u(td, io_u);
698	return 0;
699}
700
701static void __io_u_mark_map(unsigned int *map, unsigned int nr)
702{
703	int idx = 0;
704
705	switch (nr) {
706	default:
707		idx = 6;
708		break;
709	case 33 ... 64:
710		idx = 5;
711		break;
712	case 17 ... 32:
713		idx = 4;
714		break;
715	case 9 ... 16:
716		idx = 3;
717		break;
718	case 5 ... 8:
719		idx = 2;
720		break;
721	case 1 ... 4:
722		idx = 1;
723	case 0:
724		break;
725	}
726
727	map[idx]++;
728}
729
730void io_u_mark_submit(struct thread_data *td, unsigned int nr)
731{
732	__io_u_mark_map(td->ts.io_u_submit, nr);
733	td->ts.total_submit++;
734}
735
736void io_u_mark_complete(struct thread_data *td, unsigned int nr)
737{
738	__io_u_mark_map(td->ts.io_u_complete, nr);
739	td->ts.total_complete++;
740}
741
742void io_u_mark_depth(struct thread_data *td, unsigned int nr)
743{
744	int idx = 0;
745
746	switch (td->cur_depth) {
747	default:
748		idx = 6;
749		break;
750	case 32 ... 63:
751		idx = 5;
752		break;
753	case 16 ... 31:
754		idx = 4;
755		break;
756	case 8 ... 15:
757		idx = 3;
758		break;
759	case 4 ... 7:
760		idx = 2;
761		break;
762	case 2 ... 3:
763		idx = 1;
764	case 1:
765		break;
766	}
767
768	td->ts.io_u_map[idx] += nr;
769}
770
771static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
772{
773	int idx = 0;
774
775	assert(usec < 1000);
776
777	switch (usec) {
778	case 750 ... 999:
779		idx = 9;
780		break;
781	case 500 ... 749:
782		idx = 8;
783		break;
784	case 250 ... 499:
785		idx = 7;
786		break;
787	case 100 ... 249:
788		idx = 6;
789		break;
790	case 50 ... 99:
791		idx = 5;
792		break;
793	case 20 ... 49:
794		idx = 4;
795		break;
796	case 10 ... 19:
797		idx = 3;
798		break;
799	case 4 ... 9:
800		idx = 2;
801		break;
802	case 2 ... 3:
803		idx = 1;
804	case 0 ... 1:
805		break;
806	}
807
808	assert(idx < FIO_IO_U_LAT_U_NR);
809	td->ts.io_u_lat_u[idx]++;
810}
811
812static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
813{
814	int idx = 0;
815
816	switch (msec) {
817	default:
818		idx = 11;
819		break;
820	case 1000 ... 1999:
821		idx = 10;
822		break;
823	case 750 ... 999:
824		idx = 9;
825		break;
826	case 500 ... 749:
827		idx = 8;
828		break;
829	case 250 ... 499:
830		idx = 7;
831		break;
832	case 100 ... 249:
833		idx = 6;
834		break;
835	case 50 ... 99:
836		idx = 5;
837		break;
838	case 20 ... 49:
839		idx = 4;
840		break;
841	case 10 ... 19:
842		idx = 3;
843		break;
844	case 4 ... 9:
845		idx = 2;
846		break;
847	case 2 ... 3:
848		idx = 1;
849	case 0 ... 1:
850		break;
851	}
852
853	assert(idx < FIO_IO_U_LAT_M_NR);
854	td->ts.io_u_lat_m[idx]++;
855}
856
857static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
858{
859	if (usec < 1000)
860		io_u_mark_lat_usec(td, usec);
861	else
862		io_u_mark_lat_msec(td, usec / 1000);
863}
864
865/*
866 * Get next file to service by choosing one at random
867 */
868static struct fio_file *get_next_file_rand(struct thread_data *td,
869					   enum fio_file_flags goodf,
870					   enum fio_file_flags badf)
871{
872	struct fio_file *f;
873	int fno;
874
875	do {
876		int opened = 0;
877		unsigned long r;
878
879		if (td->o.use_os_rand) {
880			r = os_random_long(&td->next_file_state);
881			fno = (unsigned int) ((double) td->o.nr_files
882				* (r / (OS_RAND_MAX + 1.0)));
883		} else {
884			r = __rand(&td->__next_file_state);
885			fno = (unsigned int) ((double) td->o.nr_files
886				* (r / (FRAND_MAX + 1.0)));
887		}
888
889		f = td->files[fno];
890		if (fio_file_done(f))
891			continue;
892
893		if (!fio_file_open(f)) {
894			int err;
895
896			err = td_io_open_file(td, f);
897			if (err)
898				continue;
899			opened = 1;
900		}
901
902		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
903			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
904			return f;
905		}
906		if (opened)
907			td_io_close_file(td, f);
908	} while (1);
909}
910
911/*
912 * Get next file to service by doing round robin between all available ones
913 */
914static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
915					 int badf)
916{
917	unsigned int old_next_file = td->next_file;
918	struct fio_file *f;
919
920	do {
921		int opened = 0;
922
923		f = td->files[td->next_file];
924
925		td->next_file++;
926		if (td->next_file >= td->o.nr_files)
927			td->next_file = 0;
928
929		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
930		if (fio_file_done(f)) {
931			f = NULL;
932			continue;
933		}
934
935		if (!fio_file_open(f)) {
936			int err;
937
938			err = td_io_open_file(td, f);
939			if (err) {
940				dprint(FD_FILE, "error %d on open of %s\n",
941					err, f->file_name);
942				f = NULL;
943				continue;
944			}
945			opened = 1;
946		}
947
948		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
949								f->flags);
950		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
951			break;
952
953		if (opened)
954			td_io_close_file(td, f);
955
956		f = NULL;
957	} while (td->next_file != old_next_file);
958
959	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
960	return f;
961}
962
963static struct fio_file *__get_next_file(struct thread_data *td)
964{
965	struct fio_file *f;
966
967	assert(td->o.nr_files <= td->files_index);
968
969	if (td->nr_done_files >= td->o.nr_files) {
970		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
971				" nr_files=%d\n", td->nr_open_files,
972						  td->nr_done_files,
973						  td->o.nr_files);
974		return NULL;
975	}
976
977	f = td->file_service_file;
978	if (f && fio_file_open(f) && !fio_file_closing(f)) {
979		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
980			goto out;
981		if (td->file_service_left--)
982			goto out;
983	}
984
985	if (td->o.file_service_type == FIO_FSERVICE_RR ||
986	    td->o.file_service_type == FIO_FSERVICE_SEQ)
987		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
988	else
989		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
990
991	td->file_service_file = f;
992	td->file_service_left = td->file_service_nr - 1;
993out:
994	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
995	return f;
996}
997
998static struct fio_file *get_next_file(struct thread_data *td)
999{
1000	struct prof_io_ops *ops = &td->prof_io_ops;
1001
1002	if (ops->get_next_file)
1003		return ops->get_next_file(td);
1004
1005	return __get_next_file(td);
1006}
1007
1008static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1009{
1010	struct fio_file *f;
1011
1012	do {
1013		f = get_next_file(td);
1014		if (!f)
1015			return 1;
1016
1017		io_u->file = f;
1018		get_file(f);
1019
1020		if (!fill_io_u(td, io_u))
1021			break;
1022
1023		put_file_log(td, f);
1024		td_io_close_file(td, f);
1025		io_u->file = NULL;
1026		fio_file_set_done(f);
1027		td->nr_done_files++;
1028		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1029					td->nr_done_files, td->o.nr_files);
1030	} while (1);
1031
1032	return 0;
1033}
1034
1035
1036struct io_u *__get_io_u(struct thread_data *td)
1037{
1038	struct io_u *io_u = NULL;
1039
1040	td_io_u_lock(td);
1041
1042again:
1043	if (!flist_empty(&td->io_u_requeues))
1044		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1045	else if (!queue_full(td)) {
1046		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1047
1048		io_u->buflen = 0;
1049		io_u->resid = 0;
1050		io_u->file = NULL;
1051		io_u->end_io = NULL;
1052	}
1053
1054	if (io_u) {
1055		assert(io_u->flags & IO_U_F_FREE);
1056		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1057		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1058
1059		io_u->error = 0;
1060		flist_del(&io_u->list);
1061		flist_add(&io_u->list, &td->io_u_busylist);
1062		td->cur_depth++;
1063		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1064	} else if (td->o.verify_async) {
1065		/*
1066		 * We ran out, wait for async verify threads to finish and
1067		 * return one
1068		 */
1069		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1070		goto again;
1071	}
1072
1073	td_io_u_unlock(td);
1074	return io_u;
1075}
1076
1077static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1078{
1079	if (td->o.trim_backlog && td->trim_entries) {
1080		int get_trim = 0;
1081
1082		if (td->trim_batch) {
1083			td->trim_batch--;
1084			get_trim = 1;
1085		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1086			 td->last_ddir != DDIR_READ) {
1087			td->trim_batch = td->o.trim_batch;
1088			if (!td->trim_batch)
1089				td->trim_batch = td->o.trim_backlog;
1090			get_trim = 1;
1091		}
1092
1093		if (get_trim && !get_next_trim(td, io_u))
1094			return 1;
1095	}
1096
1097	return 0;
1098}
1099
1100static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1101{
1102	if (td->o.verify_backlog && td->io_hist_len) {
1103		int get_verify = 0;
1104
1105		if (td->verify_batch) {
1106			td->verify_batch--;
1107			get_verify = 1;
1108		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
1109			 td->last_ddir != DDIR_READ) {
1110			td->verify_batch = td->o.verify_batch;
1111			if (!td->verify_batch)
1112				td->verify_batch = td->o.verify_backlog;
1113			get_verify = 1;
1114		}
1115
1116		if (get_verify && !get_next_verify(td, io_u))
1117			return 1;
1118	}
1119
1120	return 0;
1121}
1122
1123/*
1124 * Fill offset and start time into the buffer content, to prevent too
1125 * easy compressible data for simple de-dupe attempts. Do this for every
1126 * 512b block in the range, since that should be the smallest block size
1127 * we can expect from a device.
1128 */
1129static void small_content_scramble(struct io_u *io_u)
1130{
1131	unsigned int i, nr_blocks = io_u->buflen / 512;
1132	unsigned long long boffset;
1133	unsigned int offset;
1134	void *p, *end;
1135
1136	if (!nr_blocks)
1137		return;
1138
1139	p = io_u->xfer_buf;
1140	boffset = io_u->offset;
1141
1142	for (i = 0; i < nr_blocks; i++) {
1143		/*
1144		 * Fill the byte offset into a "random" start offset of
1145		 * the buffer, given by the product of the usec time
1146		 * and the actual offset.
1147		 */
1148		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1149		offset &= ~(sizeof(unsigned long long) - 1);
1150		if (offset >= 512 - sizeof(unsigned long long))
1151			offset -= sizeof(unsigned long long);
1152		memcpy(p + offset, &boffset, sizeof(boffset));
1153
1154		end = p + 512 - sizeof(io_u->start_time);
1155		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1156		p += 512;
1157		boffset += 512;
1158	}
1159}
1160
1161/*
1162 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1163 * etc. The returned io_u is fully ready to be prepped and submitted.
1164 */
1165struct io_u *get_io_u(struct thread_data *td)
1166{
1167	struct fio_file *f;
1168	struct io_u *io_u;
1169	int do_scramble = 0;
1170
1171	io_u = __get_io_u(td);
1172	if (!io_u) {
1173		dprint(FD_IO, "__get_io_u failed\n");
1174		return NULL;
1175	}
1176
1177	if (check_get_verify(td, io_u))
1178		goto out;
1179	if (check_get_trim(td, io_u))
1180		goto out;
1181
1182	/*
1183	 * from a requeue, io_u already setup
1184	 */
1185	if (io_u->file)
1186		goto out;
1187
1188	/*
1189	 * If using an iolog, grab next piece if any available.
1190	 */
1191	if (td->o.read_iolog_file) {
1192		if (read_iolog_get(td, io_u))
1193			goto err_put;
1194	} else if (set_io_u_file(td, io_u)) {
1195		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1196		goto err_put;
1197	}
1198
1199	f = io_u->file;
1200	assert(fio_file_open(f));
1201
1202	if (ddir_rw(io_u->ddir)) {
1203		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1204			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1205			goto err_put;
1206		}
1207
1208		f->last_start = io_u->offset;
1209		f->last_pos = io_u->offset + io_u->buflen;
1210
1211		if (io_u->ddir == DDIR_WRITE) {
1212			if (td->o.verify != VERIFY_NONE)
1213				populate_verify_io_u(td, io_u);
1214			else if (td->o.refill_buffers)
1215				io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1216			else if (td->o.scramble_buffers)
1217				do_scramble = 1;
1218		} else if (io_u->ddir == DDIR_READ) {
1219			/*
1220			 * Reset the buf_filled parameters so next time if the
1221			 * buffer is used for writes it is refilled.
1222			 */
1223			io_u->buf_filled_len = 0;
1224		}
1225	}
1226
1227	/*
1228	 * Set io data pointers.
1229	 */
1230	io_u->xfer_buf = io_u->buf;
1231	io_u->xfer_buflen = io_u->buflen;
1232
1233out:
1234	assert(io_u->file);
1235	if (!td_io_prep(td, io_u)) {
1236		if (!td->o.disable_slat)
1237			fio_gettime(&io_u->start_time, NULL);
1238		if (do_scramble)
1239			small_content_scramble(io_u);
1240		return io_u;
1241	}
1242err_put:
1243	dprint(FD_IO, "get_io_u failed\n");
1244	put_io_u(td, io_u);
1245	return NULL;
1246}
1247
1248void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1249{
1250	const char *msg[] = { "read", "write", "sync", "datasync",
1251				"sync_file_range", "wait", "trim" };
1252
1253
1254
1255	log_err("fio: io_u error");
1256
1257	if (io_u->file)
1258		log_err(" on file %s", io_u->file->file_name);
1259
1260	log_err(": %s\n", strerror(io_u->error));
1261
1262	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1263					io_u->offset, io_u->xfer_buflen);
1264
1265	if (!td->error)
1266		td_verror(td, io_u->error, "io_u error");
1267}
1268
1269static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1270				  struct io_completion_data *icd,
1271				  const enum fio_ddir idx, unsigned int bytes)
1272{
1273	unsigned long uninitialized_var(lusec);
1274
1275	if (!icd->account)
1276		return;
1277
1278	if (!td->o.disable_clat || !td->o.disable_bw)
1279		lusec = utime_since(&io_u->issue_time, &icd->time);
1280
1281	if (!td->o.disable_lat) {
1282		unsigned long tusec;
1283
1284		tusec = utime_since(&io_u->start_time, &icd->time);
1285		add_lat_sample(td, idx, tusec, bytes);
1286	}
1287
1288	if (!td->o.disable_clat) {
1289		add_clat_sample(td, idx, lusec, bytes);
1290		io_u_mark_latency(td, lusec);
1291	}
1292
1293	if (!td->o.disable_bw)
1294		add_bw_sample(td, idx, bytes, &icd->time);
1295
1296	add_iops_sample(td, idx, &icd->time);
1297}
1298
1299static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1300{
1301	unsigned long long secs, remainder, bps, bytes;
1302	bytes = td->this_io_bytes[ddir];
1303	bps = td->rate_bps[ddir];
1304	secs = bytes / bps;
1305	remainder = bytes % bps;
1306	return remainder * 1000000 / bps + secs * 1000000;
1307}
1308
1309static void io_completed(struct thread_data *td, struct io_u *io_u,
1310			 struct io_completion_data *icd)
1311{
1312	/*
1313	 * Older gcc's are too dumb to realize that usec is always used
1314	 * initialized, silence that warning.
1315	 */
1316	unsigned long uninitialized_var(usec);
1317	struct fio_file *f;
1318
1319	dprint_io_u(io_u, "io complete");
1320
1321	td_io_u_lock(td);
1322	assert(io_u->flags & IO_U_F_FLIGHT);
1323	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1324	td_io_u_unlock(td);
1325
1326	if (ddir_sync(io_u->ddir)) {
1327		td->last_was_sync = 1;
1328		f = io_u->file;
1329		if (f) {
1330			f->first_write = -1ULL;
1331			f->last_write = -1ULL;
1332		}
1333		return;
1334	}
1335
1336	td->last_was_sync = 0;
1337	td->last_ddir = io_u->ddir;
1338
1339	if (!io_u->error && ddir_rw(io_u->ddir)) {
1340		unsigned int bytes = io_u->buflen - io_u->resid;
1341		const enum fio_ddir idx = io_u->ddir;
1342		const enum fio_ddir odx = io_u->ddir ^ 1;
1343		int ret;
1344
1345		td->io_blocks[idx]++;
1346		td->this_io_blocks[idx]++;
1347		td->io_bytes[idx] += bytes;
1348		td->this_io_bytes[idx] += bytes;
1349
1350		if (idx == DDIR_WRITE) {
1351			f = io_u->file;
1352			if (f) {
1353				if (f->first_write == -1ULL ||
1354				    io_u->offset < f->first_write)
1355					f->first_write = io_u->offset;
1356				if (f->last_write == -1ULL ||
1357				    ((io_u->offset + bytes) > f->last_write))
1358					f->last_write = io_u->offset + bytes;
1359			}
1360		}
1361
1362		if (ramp_time_over(td) && td->runstate == TD_RUNNING) {
1363			account_io_completion(td, io_u, icd, idx, bytes);
1364
1365			if (__should_check_rate(td, idx)) {
1366				td->rate_pending_usleep[idx] =
1367					(usec_for_io(td, idx) -
1368					 utime_since_now(&td->start));
1369			}
1370			if (__should_check_rate(td, odx))
1371				td->rate_pending_usleep[odx] =
1372					(usec_for_io(td, odx) -
1373					 utime_since_now(&td->start));
1374		}
1375
1376		if (td_write(td) && idx == DDIR_WRITE &&
1377		    td->o.do_verify &&
1378		    td->o.verify != VERIFY_NONE)
1379			log_io_piece(td, io_u);
1380
1381		icd->bytes_done[idx] += bytes;
1382
1383		if (io_u->end_io) {
1384			ret = io_u->end_io(td, io_u);
1385			if (ret && !icd->error)
1386				icd->error = ret;
1387		}
1388	} else if (io_u->error) {
1389		icd->error = io_u->error;
1390		io_u_log_error(td, io_u);
1391	}
1392	if (icd->error && td_non_fatal_error(icd->error) &&
1393           (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1394		/*
1395		 * If there is a non_fatal error, then add to the error count
1396		 * and clear all the errors.
1397		 */
1398		update_error_count(td, icd->error);
1399		td_clear_error(td);
1400		icd->error = 0;
1401		io_u->error = 0;
1402	}
1403}
1404
1405static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1406		     int nr)
1407{
1408	if (!td->o.disable_clat || !td->o.disable_bw)
1409		fio_gettime(&icd->time, NULL);
1410
1411	icd->nr = nr;
1412	icd->account = 1;
1413
1414	icd->error = 0;
1415	icd->bytes_done[0] = icd->bytes_done[1] = 0;
1416}
1417
1418static void ios_completed(struct thread_data *td,
1419			  struct io_completion_data *icd)
1420{
1421	struct io_u *io_u;
1422	int i;
1423
1424	for (i = 0; i < icd->nr; i++) {
1425		io_u = td->io_ops->event(td, i);
1426
1427		io_completed(td, io_u, icd);
1428
1429		if (!(io_u->flags & IO_U_F_FREE_DEF))
1430			put_io_u(td, io_u);
1431
1432		icd->account = 0;
1433	}
1434}
1435
1436/*
1437 * Complete a single io_u for the sync engines.
1438 */
1439int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1440		       unsigned long *bytes)
1441{
1442	struct io_completion_data icd;
1443
1444	init_icd(td, &icd, 1);
1445	io_completed(td, io_u, &icd);
1446
1447	if (!(io_u->flags & IO_U_F_FREE_DEF))
1448		put_io_u(td, io_u);
1449
1450	if (icd.error) {
1451		td_verror(td, icd.error, "io_u_sync_complete");
1452		return -1;
1453	}
1454
1455	if (bytes) {
1456		bytes[0] += icd.bytes_done[0];
1457		bytes[1] += icd.bytes_done[1];
1458	}
1459
1460	return 0;
1461}
1462
1463/*
1464 * Called to complete min_events number of io for the async engines.
1465 */
1466int io_u_queued_complete(struct thread_data *td, int min_evts,
1467			 unsigned long *bytes)
1468{
1469	struct io_completion_data icd;
1470	struct timespec *tvp = NULL;
1471	int ret;
1472	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1473
1474	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1475
1476	if (!min_evts)
1477		tvp = &ts;
1478
1479	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1480	if (ret < 0) {
1481		td_verror(td, -ret, "td_io_getevents");
1482		return ret;
1483	} else if (!ret)
1484		return ret;
1485
1486	init_icd(td, &icd, ret);
1487	ios_completed(td, &icd);
1488	if (icd.error) {
1489		td_verror(td, icd.error, "io_u_queued_complete");
1490		return -1;
1491	}
1492
1493	if (bytes) {
1494		bytes[0] += icd.bytes_done[0];
1495		bytes[1] += icd.bytes_done[1];
1496	}
1497
1498	return 0;
1499}
1500
1501/*
1502 * Call when io_u is really queued, to update the submission latency.
1503 */
1504void io_u_queued(struct thread_data *td, struct io_u *io_u)
1505{
1506	if (!td->o.disable_slat) {
1507		unsigned long slat_time;
1508
1509		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1510		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1511	}
1512}
1513
1514/*
1515 * "randomly" fill the buffer contents
1516 */
1517void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1518		      unsigned int max_bs)
1519{
1520	io_u->buf_filled_len = 0;
1521
1522	if (!td->o.zero_buffers)
1523		fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1524	else
1525		memset(io_u->buf, 0, max_bs);
1526}
1527