io_u.c revision c8eeb9df1f52f28567a5937e141decc6a26ec30b
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 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	do {
172		if (td->o.use_os_rand) {
173			r = os_random_long(&td->random_state);
174			*b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
175		} else {
176			r = __rand(&td->__random_state);
177			*b = (lastb - 1) * (r / ((unsigned long long) FRAND_MAX + 1.0));
178		}
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 / (OS_RAND_MAX + 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	io_u->flags |= IO_U_F_FREE;
601	io_u->flags &= ~IO_U_F_FREE_DEF;
602
603	if (io_u->file)
604		put_file_log(td, io_u->file);
605
606	io_u->file = NULL;
607	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
608		td->cur_depth--;
609	flist_del_init(&io_u->list);
610	flist_add(&io_u->list, &td->io_u_freelist);
611	td_io_u_unlock(td);
612	td_io_u_free_notify(td);
613}
614
615void clear_io_u(struct thread_data *td, struct io_u *io_u)
616{
617	io_u->flags &= ~IO_U_F_FLIGHT;
618	put_io_u(td, io_u);
619}
620
621void requeue_io_u(struct thread_data *td, struct io_u **io_u)
622{
623	struct io_u *__io_u = *io_u;
624
625	dprint(FD_IO, "requeue %p\n", __io_u);
626
627	td_io_u_lock(td);
628
629	__io_u->flags |= IO_U_F_FREE;
630	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
631		td->io_issues[__io_u->ddir]--;
632
633	__io_u->flags &= ~IO_U_F_FLIGHT;
634	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
635		td->cur_depth--;
636	flist_del(&__io_u->list);
637	flist_add_tail(&__io_u->list, &td->io_u_requeues);
638	td_io_u_unlock(td);
639	*io_u = NULL;
640}
641
642static int fill_io_u(struct thread_data *td, struct io_u *io_u)
643{
644	if (td->io_ops->flags & FIO_NOIO)
645		goto out;
646
647	set_rw_ddir(td, io_u);
648
649	/*
650	 * fsync() or fdatasync() or trim etc, we are done
651	 */
652	if (!ddir_rw(io_u->ddir))
653		goto out;
654
655	/*
656	 * See if it's time to switch to a new zone
657	 */
658	if (td->zone_bytes >= td->o.zone_size) {
659		td->zone_bytes = 0;
660		io_u->file->last_pos += td->o.zone_skip;
661		td->io_skip_bytes += td->o.zone_skip;
662	}
663
664	/*
665	 * No log, let the seq/rand engine retrieve the next buflen and
666	 * position.
667	 */
668	if (get_next_offset(td, io_u)) {
669		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
670		return 1;
671	}
672
673	io_u->buflen = get_next_buflen(td, io_u);
674	if (!io_u->buflen) {
675		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
676		return 1;
677	}
678
679	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
680		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
681		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
682				io_u->buflen, io_u->file->real_file_size);
683		return 1;
684	}
685
686	/*
687	 * mark entry before potentially trimming io_u
688	 */
689	if (td_random(td) && file_randommap(td, io_u->file))
690		mark_random_map(td, io_u);
691
692	/*
693	 * If using a write iolog, store this entry.
694	 */
695out:
696	dprint_io_u(io_u, "fill_io_u");
697	td->zone_bytes += io_u->buflen;
698	log_io_u(td, io_u);
699	return 0;
700}
701
702static void __io_u_mark_map(unsigned int *map, unsigned int nr)
703{
704	int idx = 0;
705
706	switch (nr) {
707	default:
708		idx = 6;
709		break;
710	case 33 ... 64:
711		idx = 5;
712		break;
713	case 17 ... 32:
714		idx = 4;
715		break;
716	case 9 ... 16:
717		idx = 3;
718		break;
719	case 5 ... 8:
720		idx = 2;
721		break;
722	case 1 ... 4:
723		idx = 1;
724	case 0:
725		break;
726	}
727
728	map[idx]++;
729}
730
731void io_u_mark_submit(struct thread_data *td, unsigned int nr)
732{
733	__io_u_mark_map(td->ts.io_u_submit, nr);
734	td->ts.total_submit++;
735}
736
737void io_u_mark_complete(struct thread_data *td, unsigned int nr)
738{
739	__io_u_mark_map(td->ts.io_u_complete, nr);
740	td->ts.total_complete++;
741}
742
743void io_u_mark_depth(struct thread_data *td, unsigned int nr)
744{
745	int idx = 0;
746
747	switch (td->cur_depth) {
748	default:
749		idx = 6;
750		break;
751	case 32 ... 63:
752		idx = 5;
753		break;
754	case 16 ... 31:
755		idx = 4;
756		break;
757	case 8 ... 15:
758		idx = 3;
759		break;
760	case 4 ... 7:
761		idx = 2;
762		break;
763	case 2 ... 3:
764		idx = 1;
765	case 1:
766		break;
767	}
768
769	td->ts.io_u_map[idx] += nr;
770}
771
772static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
773{
774	int idx = 0;
775
776	assert(usec < 1000);
777
778	switch (usec) {
779	case 750 ... 999:
780		idx = 9;
781		break;
782	case 500 ... 749:
783		idx = 8;
784		break;
785	case 250 ... 499:
786		idx = 7;
787		break;
788	case 100 ... 249:
789		idx = 6;
790		break;
791	case 50 ... 99:
792		idx = 5;
793		break;
794	case 20 ... 49:
795		idx = 4;
796		break;
797	case 10 ... 19:
798		idx = 3;
799		break;
800	case 4 ... 9:
801		idx = 2;
802		break;
803	case 2 ... 3:
804		idx = 1;
805	case 0 ... 1:
806		break;
807	}
808
809	assert(idx < FIO_IO_U_LAT_U_NR);
810	td->ts.io_u_lat_u[idx]++;
811}
812
813static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
814{
815	int idx = 0;
816
817	switch (msec) {
818	default:
819		idx = 11;
820		break;
821	case 1000 ... 1999:
822		idx = 10;
823		break;
824	case 750 ... 999:
825		idx = 9;
826		break;
827	case 500 ... 749:
828		idx = 8;
829		break;
830	case 250 ... 499:
831		idx = 7;
832		break;
833	case 100 ... 249:
834		idx = 6;
835		break;
836	case 50 ... 99:
837		idx = 5;
838		break;
839	case 20 ... 49:
840		idx = 4;
841		break;
842	case 10 ... 19:
843		idx = 3;
844		break;
845	case 4 ... 9:
846		idx = 2;
847		break;
848	case 2 ... 3:
849		idx = 1;
850	case 0 ... 1:
851		break;
852	}
853
854	assert(idx < FIO_IO_U_LAT_M_NR);
855	td->ts.io_u_lat_m[idx]++;
856}
857
858static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
859{
860	if (usec < 1000)
861		io_u_mark_lat_usec(td, usec);
862	else
863		io_u_mark_lat_msec(td, usec / 1000);
864}
865
866/*
867 * Get next file to service by choosing one at random
868 */
869static struct fio_file *get_next_file_rand(struct thread_data *td,
870					   enum fio_file_flags goodf,
871					   enum fio_file_flags badf)
872{
873	struct fio_file *f;
874	int fno;
875
876	do {
877		int opened = 0;
878		unsigned long r;
879
880		if (td->o.use_os_rand) {
881			r = os_random_long(&td->next_file_state);
882			fno = (unsigned int) ((double) td->o.nr_files
883				* (r / (OS_RAND_MAX + 1.0)));
884		} else {
885			r = __rand(&td->__next_file_state);
886			fno = (unsigned int) ((double) td->o.nr_files
887				* (r / (FRAND_MAX + 1.0)));
888		}
889
890		f = td->files[fno];
891		if (fio_file_done(f))
892			continue;
893
894		if (!fio_file_open(f)) {
895			int err;
896
897			err = td_io_open_file(td, f);
898			if (err)
899				continue;
900			opened = 1;
901		}
902
903		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
904			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
905			return f;
906		}
907		if (opened)
908			td_io_close_file(td, f);
909	} while (1);
910}
911
912/*
913 * Get next file to service by doing round robin between all available ones
914 */
915static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
916					 int badf)
917{
918	unsigned int old_next_file = td->next_file;
919	struct fio_file *f;
920
921	do {
922		int opened = 0;
923
924		f = td->files[td->next_file];
925
926		td->next_file++;
927		if (td->next_file >= td->o.nr_files)
928			td->next_file = 0;
929
930		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
931		if (fio_file_done(f)) {
932			f = NULL;
933			continue;
934		}
935
936		if (!fio_file_open(f)) {
937			int err;
938
939			err = td_io_open_file(td, f);
940			if (err) {
941				dprint(FD_FILE, "error %d on open of %s\n",
942					err, f->file_name);
943				f = NULL;
944				continue;
945			}
946			opened = 1;
947		}
948
949		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
950								f->flags);
951		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
952			break;
953
954		if (opened)
955			td_io_close_file(td, f);
956
957		f = NULL;
958	} while (td->next_file != old_next_file);
959
960	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
961	return f;
962}
963
964static struct fio_file *__get_next_file(struct thread_data *td)
965{
966	struct fio_file *f;
967
968	assert(td->o.nr_files <= td->files_index);
969
970	if (td->nr_done_files >= td->o.nr_files) {
971		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
972				" nr_files=%d\n", td->nr_open_files,
973						  td->nr_done_files,
974						  td->o.nr_files);
975		return NULL;
976	}
977
978	f = td->file_service_file;
979	if (f && fio_file_open(f) && !fio_file_closing(f)) {
980		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
981			goto out;
982		if (td->file_service_left--)
983			goto out;
984	}
985
986	if (td->o.file_service_type == FIO_FSERVICE_RR ||
987	    td->o.file_service_type == FIO_FSERVICE_SEQ)
988		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
989	else
990		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
991
992	td->file_service_file = f;
993	td->file_service_left = td->file_service_nr - 1;
994out:
995	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
996	return f;
997}
998
999static struct fio_file *get_next_file(struct thread_data *td)
1000{
1001	struct prof_io_ops *ops = &td->prof_io_ops;
1002
1003	if (ops->get_next_file)
1004		return ops->get_next_file(td);
1005
1006	return __get_next_file(td);
1007}
1008
1009static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1010{
1011	struct fio_file *f;
1012
1013	do {
1014		f = get_next_file(td);
1015		if (!f)
1016			return 1;
1017
1018		io_u->file = f;
1019		get_file(f);
1020
1021		if (!fill_io_u(td, io_u))
1022			break;
1023
1024		put_file_log(td, f);
1025		td_io_close_file(td, f);
1026		io_u->file = NULL;
1027		fio_file_set_done(f);
1028		td->nr_done_files++;
1029		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1030					td->nr_done_files, td->o.nr_files);
1031	} while (1);
1032
1033	return 0;
1034}
1035
1036
1037struct io_u *__get_io_u(struct thread_data *td)
1038{
1039	struct io_u *io_u = NULL;
1040
1041	td_io_u_lock(td);
1042
1043again:
1044	if (!flist_empty(&td->io_u_requeues))
1045		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1046	else if (!queue_full(td)) {
1047		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1048
1049		io_u->buflen = 0;
1050		io_u->resid = 0;
1051		io_u->file = NULL;
1052		io_u->end_io = NULL;
1053	}
1054
1055	if (io_u) {
1056		assert(io_u->flags & IO_U_F_FREE);
1057		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1058		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1059
1060		io_u->error = 0;
1061		flist_del(&io_u->list);
1062		flist_add(&io_u->list, &td->io_u_busylist);
1063		td->cur_depth++;
1064		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1065	} else if (td->o.verify_async) {
1066		/*
1067		 * We ran out, wait for async verify threads to finish and
1068		 * return one
1069		 */
1070		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1071		goto again;
1072	}
1073
1074	td_io_u_unlock(td);
1075	return io_u;
1076}
1077
1078static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1079{
1080	if (td->o.trim_backlog && td->trim_entries) {
1081		int get_trim = 0;
1082
1083		if (td->trim_batch) {
1084			td->trim_batch--;
1085			get_trim = 1;
1086		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1087			 td->last_ddir != DDIR_READ) {
1088			td->trim_batch = td->o.trim_batch;
1089			if (!td->trim_batch)
1090				td->trim_batch = td->o.trim_backlog;
1091			get_trim = 1;
1092		}
1093
1094		if (get_trim && !get_next_trim(td, io_u))
1095			return 1;
1096	}
1097
1098	return 0;
1099}
1100
1101static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1102{
1103	if (td->o.verify_backlog && td->io_hist_len) {
1104		int get_verify = 0;
1105
1106		if (td->verify_batch) {
1107			td->verify_batch--;
1108			get_verify = 1;
1109		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
1110			 td->last_ddir != DDIR_READ) {
1111			td->verify_batch = td->o.verify_batch;
1112			if (!td->verify_batch)
1113				td->verify_batch = td->o.verify_backlog;
1114			get_verify = 1;
1115		}
1116
1117		if (get_verify && !get_next_verify(td, io_u))
1118			return 1;
1119	}
1120
1121	return 0;
1122}
1123
1124/*
1125 * Fill offset and start time into the buffer content, to prevent too
1126 * easy compressible data for simple de-dupe attempts. Do this for every
1127 * 512b block in the range, since that should be the smallest block size
1128 * we can expect from a device.
1129 */
1130static void small_content_scramble(struct io_u *io_u)
1131{
1132	unsigned int i, nr_blocks = io_u->buflen / 512;
1133	unsigned long long boffset;
1134	unsigned int offset;
1135	void *p, *end;
1136
1137	if (!nr_blocks)
1138		return;
1139
1140	p = io_u->xfer_buf;
1141	boffset = io_u->offset;
1142
1143	for (i = 0; i < nr_blocks; i++) {
1144		/*
1145		 * Fill the byte offset into a "random" start offset of
1146		 * the buffer, given by the product of the usec time
1147		 * and the actual offset.
1148		 */
1149		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1150		offset &= ~(sizeof(unsigned long long) - 1);
1151		if (offset >= 512 - sizeof(unsigned long long))
1152			offset -= sizeof(unsigned long long);
1153		memcpy(p + offset, &boffset, sizeof(boffset));
1154
1155		end = p + 512 - sizeof(io_u->start_time);
1156		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1157		p += 512;
1158		boffset += 512;
1159	}
1160}
1161
1162/*
1163 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1164 * etc. The returned io_u is fully ready to be prepped and submitted.
1165 */
1166struct io_u *get_io_u(struct thread_data *td)
1167{
1168	struct fio_file *f;
1169	struct io_u *io_u;
1170	int do_scramble = 0;
1171
1172	io_u = __get_io_u(td);
1173	if (!io_u) {
1174		dprint(FD_IO, "__get_io_u failed\n");
1175		return NULL;
1176	}
1177
1178	if (check_get_verify(td, io_u))
1179		goto out;
1180	if (check_get_trim(td, io_u))
1181		goto out;
1182
1183	/*
1184	 * from a requeue, io_u already setup
1185	 */
1186	if (io_u->file)
1187		goto out;
1188
1189	/*
1190	 * If using an iolog, grab next piece if any available.
1191	 */
1192	if (td->o.read_iolog_file) {
1193		if (read_iolog_get(td, io_u))
1194			goto err_put;
1195	} else if (set_io_u_file(td, io_u)) {
1196		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1197		goto err_put;
1198	}
1199
1200	f = io_u->file;
1201	assert(fio_file_open(f));
1202
1203	if (ddir_rw(io_u->ddir)) {
1204		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1205			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1206			goto err_put;
1207		}
1208
1209		f->last_start = io_u->offset;
1210		f->last_pos = io_u->offset + io_u->buflen;
1211
1212		if (io_u->ddir == DDIR_WRITE) {
1213			if (td->o.verify != VERIFY_NONE)
1214				populate_verify_io_u(td, io_u);
1215			else if (td->o.refill_buffers)
1216				io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1217			else if (td->o.scramble_buffers)
1218				do_scramble = 1;
1219		} else if (io_u->ddir == DDIR_READ) {
1220			/*
1221			 * Reset the buf_filled parameters so next time if the
1222			 * buffer is used for writes it is refilled.
1223			 */
1224			io_u->buf_filled_len = 0;
1225		}
1226	}
1227
1228	/*
1229	 * Set io data pointers.
1230	 */
1231	io_u->xfer_buf = io_u->buf;
1232	io_u->xfer_buflen = io_u->buflen;
1233
1234out:
1235	assert(io_u->file);
1236	if (!td_io_prep(td, io_u)) {
1237		if (!td->o.disable_slat)
1238			fio_gettime(&io_u->start_time, NULL);
1239		if (do_scramble)
1240			small_content_scramble(io_u);
1241		return io_u;
1242	}
1243err_put:
1244	dprint(FD_IO, "get_io_u failed\n");
1245	put_io_u(td, io_u);
1246	return NULL;
1247}
1248
1249void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1250{
1251	const char *msg[] = { "read", "write", "sync", "datasync",
1252				"sync_file_range", "wait", "trim" };
1253
1254
1255
1256	log_err("fio: io_u error");
1257
1258	if (io_u->file)
1259		log_err(" on file %s", io_u->file->file_name);
1260
1261	log_err(": %s\n", strerror(io_u->error));
1262
1263	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1264					io_u->offset, io_u->xfer_buflen);
1265
1266	if (!td->error)
1267		td_verror(td, io_u->error, "io_u error");
1268}
1269
1270static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1271				  struct io_completion_data *icd,
1272				  const enum fio_ddir idx, unsigned int bytes)
1273{
1274	unsigned long uninitialized_var(lusec);
1275
1276	if (!icd->account)
1277		return;
1278
1279	if (!td->o.disable_clat || !td->o.disable_bw)
1280		lusec = utime_since(&io_u->issue_time, &icd->time);
1281
1282	if (!td->o.disable_lat) {
1283		unsigned long tusec;
1284
1285		tusec = utime_since(&io_u->start_time, &icd->time);
1286		add_lat_sample(td, idx, tusec, bytes);
1287	}
1288
1289	if (!td->o.disable_clat) {
1290		add_clat_sample(td, idx, lusec, bytes);
1291		io_u_mark_latency(td, lusec);
1292	}
1293
1294	if (!td->o.disable_bw)
1295		add_bw_sample(td, idx, bytes, &icd->time);
1296
1297	add_iops_sample(td, idx, &icd->time);
1298}
1299
1300static void io_completed(struct thread_data *td, struct io_u *io_u,
1301			 struct io_completion_data *icd)
1302{
1303	/*
1304	 * Older gcc's are too dumb to realize that usec is always used
1305	 * initialized, silence that warning.
1306	 */
1307	unsigned long uninitialized_var(usec);
1308	struct fio_file *f;
1309
1310	dprint_io_u(io_u, "io complete");
1311
1312	td_io_u_lock(td);
1313	assert(io_u->flags & IO_U_F_FLIGHT);
1314	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1315	td_io_u_unlock(td);
1316
1317	if (ddir_sync(io_u->ddir)) {
1318		td->last_was_sync = 1;
1319		f = io_u->file;
1320		if (f) {
1321			f->first_write = -1ULL;
1322			f->last_write = -1ULL;
1323		}
1324		return;
1325	}
1326
1327	td->last_was_sync = 0;
1328	td->last_ddir = io_u->ddir;
1329
1330	if (!io_u->error && ddir_rw(io_u->ddir)) {
1331		unsigned int bytes = io_u->buflen - io_u->resid;
1332		const enum fio_ddir idx = io_u->ddir;
1333		const enum fio_ddir odx = io_u->ddir ^ 1;
1334		int ret;
1335
1336		td->io_blocks[idx]++;
1337		td->this_io_blocks[idx]++;
1338		td->io_bytes[idx] += bytes;
1339		td->this_io_bytes[idx] += bytes;
1340
1341		if (idx == DDIR_WRITE) {
1342			f = io_u->file;
1343			if (f) {
1344				if (f->first_write == -1ULL ||
1345				    io_u->offset < f->first_write)
1346					f->first_write = io_u->offset;
1347				if (f->last_write == -1ULL ||
1348				    ((io_u->offset + bytes) > f->last_write))
1349					f->last_write = io_u->offset + bytes;
1350			}
1351		}
1352
1353		if (ramp_time_over(td) && td->runstate == TD_RUNNING) {
1354			account_io_completion(td, io_u, icd, idx, bytes);
1355
1356			if (__should_check_rate(td, idx)) {
1357				td->rate_pending_usleep[idx] =
1358					((td->this_io_bytes[idx] *
1359					  td->rate_nsec_cycle[idx]) / 1000 -
1360					 utime_since_now(&td->start));
1361			}
1362			if (__should_check_rate(td, idx ^ 1))
1363				td->rate_pending_usleep[odx] =
1364					((td->this_io_bytes[odx] *
1365					  td->rate_nsec_cycle[odx]) / 1000 -
1366					 utime_since_now(&td->start));
1367		}
1368
1369		if (td_write(td) && idx == DDIR_WRITE &&
1370		    td->o.do_verify &&
1371		    td->o.verify != VERIFY_NONE)
1372			log_io_piece(td, io_u);
1373
1374		icd->bytes_done[idx] += bytes;
1375
1376		if (io_u->end_io) {
1377			ret = io_u->end_io(td, io_u);
1378			if (ret && !icd->error)
1379				icd->error = ret;
1380		}
1381	} else if (io_u->error) {
1382		icd->error = io_u->error;
1383		io_u_log_error(td, io_u);
1384	}
1385	if (td->o.continue_on_error && icd->error &&
1386	    td_non_fatal_error(icd->error)) {
1387		/*
1388		 * If there is a non_fatal error, then add to the error count
1389		 * and clear all the errors.
1390		 */
1391		update_error_count(td, icd->error);
1392		td_clear_error(td);
1393		icd->error = 0;
1394		io_u->error = 0;
1395	}
1396}
1397
1398static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1399		     int nr)
1400{
1401	if (!td->o.disable_clat || !td->o.disable_bw)
1402		fio_gettime(&icd->time, NULL);
1403
1404	icd->nr = nr;
1405	icd->account = 1;
1406
1407	icd->error = 0;
1408	icd->bytes_done[0] = icd->bytes_done[1] = 0;
1409}
1410
1411static void ios_completed(struct thread_data *td,
1412			  struct io_completion_data *icd)
1413{
1414	struct io_u *io_u;
1415	int i;
1416
1417	for (i = 0; i < icd->nr; i++) {
1418		io_u = td->io_ops->event(td, i);
1419
1420		io_completed(td, io_u, icd);
1421
1422		if (!(io_u->flags & IO_U_F_FREE_DEF))
1423			put_io_u(td, io_u);
1424
1425		icd->account = 0;
1426	}
1427}
1428
1429/*
1430 * Complete a single io_u for the sync engines.
1431 */
1432int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1433		       unsigned long *bytes)
1434{
1435	struct io_completion_data icd;
1436
1437	init_icd(td, &icd, 1);
1438	io_completed(td, io_u, &icd);
1439
1440	if (!(io_u->flags & IO_U_F_FREE_DEF))
1441		put_io_u(td, io_u);
1442
1443	if (icd.error) {
1444		td_verror(td, icd.error, "io_u_sync_complete");
1445		return -1;
1446	}
1447
1448	if (bytes) {
1449		bytes[0] += icd.bytes_done[0];
1450		bytes[1] += icd.bytes_done[1];
1451	}
1452
1453	return 0;
1454}
1455
1456/*
1457 * Called to complete min_events number of io for the async engines.
1458 */
1459int io_u_queued_complete(struct thread_data *td, int min_evts,
1460			 unsigned long *bytes)
1461{
1462	struct io_completion_data icd;
1463	struct timespec *tvp = NULL;
1464	int ret;
1465	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1466
1467	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1468
1469	if (!min_evts)
1470		tvp = &ts;
1471
1472	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1473	if (ret < 0) {
1474		td_verror(td, -ret, "td_io_getevents");
1475		return ret;
1476	} else if (!ret)
1477		return ret;
1478
1479	init_icd(td, &icd, ret);
1480	ios_completed(td, &icd);
1481	if (icd.error) {
1482		td_verror(td, icd.error, "io_u_queued_complete");
1483		return -1;
1484	}
1485
1486	if (bytes) {
1487		bytes[0] += icd.bytes_done[0];
1488		bytes[1] += icd.bytes_done[1];
1489	}
1490
1491	return 0;
1492}
1493
1494/*
1495 * Call when io_u is really queued, to update the submission latency.
1496 */
1497void io_u_queued(struct thread_data *td, struct io_u *io_u)
1498{
1499	if (!td->o.disable_slat) {
1500		unsigned long slat_time;
1501
1502		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1503		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1504	}
1505}
1506
1507/*
1508 * "randomly" fill the buffer contents
1509 */
1510void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1511		      unsigned int max_bs)
1512{
1513	io_u->buf_filled_len = 0;
1514
1515	if (!td->o.zero_buffers)
1516		fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1517	else
1518		memset(io_u->buf, 0, max_bs);
1519}
1520