io_u.c revision 6b1190fdf64da94a5d6b22439fa31825160f8734
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13
14struct io_completion_data {
15	int nr;				/* input */
16	int account;			/* input */
17
18	int error;			/* output */
19	unsigned long bytes_done[2];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const unsigned long long block)
28{
29	unsigned int idx = RAND_MAP_IDX(f, block);
30	unsigned int bit = RAND_MAP_BIT(f, block);
31
32	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33
34	return (f->file_map[idx] & (1UL << bit)) == 0;
35}
36
37/*
38 * Mark a given offset as used in the map.
39 */
40static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41{
42	unsigned int min_bs = td->o.rw_min_bs;
43	struct fio_file *f = io_u->file;
44	unsigned long long block;
45	unsigned int blocks, nr_blocks;
46	int busy_check;
47
48	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
49	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50	blocks = 0;
51	busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
52
53	while (nr_blocks) {
54		unsigned int idx, bit;
55		unsigned long mask, this_blocks;
56
57		/*
58		 * If we have a mixed random workload, we may
59		 * encounter blocks we already did IO to.
60		 */
61		if (!busy_check) {
62			blocks = nr_blocks;
63			break;
64		}
65		if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
66			break;
67
68		idx = RAND_MAP_IDX(f, block);
69		bit = RAND_MAP_BIT(f, block);
70
71		fio_assert(td, idx < f->num_maps);
72
73		this_blocks = nr_blocks;
74		if (this_blocks + bit > BLOCKS_PER_MAP)
75			this_blocks = BLOCKS_PER_MAP - bit;
76
77		do {
78			if (this_blocks == BLOCKS_PER_MAP)
79				mask = -1UL;
80			else
81				mask = ((1UL << this_blocks) - 1) << bit;
82
83			if (!(f->file_map[idx] & mask))
84				break;
85
86			this_blocks--;
87		} while (this_blocks);
88
89		if (!this_blocks)
90			break;
91
92		f->file_map[idx] |= mask;
93		nr_blocks -= this_blocks;
94		blocks += this_blocks;
95		block += this_blocks;
96	}
97
98	if ((blocks * min_bs) < io_u->buflen)
99		io_u->buflen = blocks * min_bs;
100}
101
102static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
103				     enum fio_ddir ddir)
104{
105	unsigned long long max_blocks;
106	unsigned long long max_size;
107
108	assert(ddir_rw(ddir));
109
110	/*
111	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
112	 */
113	max_size = f->io_size;
114	if (max_size > f->real_file_size)
115		max_size = f->real_file_size;
116
117	if (td->o.zone_range)
118		max_size = td->o.zone_range;
119
120	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
121	if (!max_blocks)
122		return 0;
123
124	return max_blocks;
125}
126
127/*
128 * Return the next free block in the map.
129 */
130static int get_next_free_block(struct thread_data *td, struct fio_file *f,
131			       enum fio_ddir ddir, unsigned long long *b)
132{
133	unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
134	int i;
135
136	lastb = last_block(td, f, ddir);
137	if (!lastb)
138		return 1;
139
140	i = f->last_free_lookup;
141	block = i * BLOCKS_PER_MAP;
142	while (block * min_bs < f->real_file_size &&
143		block * min_bs < f->io_size) {
144		if (f->file_map[i] != -1UL) {
145			block += ffz(f->file_map[i]);
146			if (block > lastb)
147				break;
148			f->last_free_lookup = i;
149			*b = block;
150			return 0;
151		}
152
153		block += BLOCKS_PER_MAP;
154		i++;
155	}
156
157	dprint(FD_IO, "failed finding a free block\n");
158	return 1;
159}
160
161static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
162				enum fio_ddir ddir, unsigned long long *b)
163{
164	unsigned long long rmax, r, lastb;
165	int loops = 5;
166
167	lastb = last_block(td, f, ddir);
168	if (!lastb)
169		return 1;
170
171	if (f->failed_rands >= 200)
172		goto ffz;
173
174	rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
175	do {
176		if (td->o.use_os_rand)
177			r = os_random_long(&td->random_state);
178		else
179			r = __rand(&td->__random_state);
180
181		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
182
183		dprint(FD_RANDOM, "off rand %llu\n", r);
184
185
186		/*
187		 * if we are not maintaining a random map, we are done.
188		 */
189		if (!file_randommap(td, f))
190			goto ret_good;
191
192		/*
193		 * calculate map offset and check if it's free
194		 */
195		if (random_map_free(f, *b))
196			goto ret_good;
197
198		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
199									*b);
200	} while (--loops);
201
202	if (!f->failed_rands++)
203		f->last_free_lookup = 0;
204
205	/*
206	 * we get here, if we didn't suceed in looking up a block. generate
207	 * a random start offset into the filemap, and find the first free
208	 * block from there.
209	 */
210	loops = 10;
211	do {
212		f->last_free_lookup = (f->num_maps - 1) *
213					(r / ((unsigned long long) rmax + 1.0));
214		if (!get_next_free_block(td, f, ddir, b))
215			goto ret;
216
217		if (td->o.use_os_rand)
218			r = os_random_long(&td->random_state);
219		else
220			r = __rand(&td->__random_state);
221	} while (--loops);
222
223	/*
224	 * that didn't work either, try exhaustive search from the start
225	 */
226	f->last_free_lookup = 0;
227ffz:
228	if (!get_next_free_block(td, f, ddir, b))
229		return 0;
230	f->last_free_lookup = 0;
231	return get_next_free_block(td, f, ddir, b);
232ret_good:
233	f->failed_rands = 0;
234ret:
235	return 0;
236}
237
238static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239			       enum fio_ddir ddir, unsigned long long *b)
240{
241	if (get_next_rand_offset(td, f, ddir, b)) {
242		dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
243				f->file_name, f->last_pos, f->real_file_size);
244		return 1;
245	}
246
247	return 0;
248}
249
250static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
251			      enum fio_ddir ddir, unsigned long long *b)
252{
253	assert(ddir_rw(ddir));
254
255	if (f->last_pos < f->real_file_size) {
256		unsigned long long pos;
257
258		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
259			f->last_pos = f->real_file_size;
260
261		pos = f->last_pos - f->file_offset;
262		if (pos)
263			pos += td->o.ddir_seq_add;
264
265		*b = pos / td->o.min_bs[ddir];
266		return 0;
267	}
268
269	return 1;
270}
271
272static int get_next_block(struct thread_data *td, struct io_u *io_u,
273			  enum fio_ddir ddir, int rw_seq, unsigned long long *b)
274{
275	struct fio_file *f = io_u->file;
276	int ret;
277
278	assert(ddir_rw(ddir));
279
280	if (rw_seq) {
281		if (td_random(td))
282			ret = get_next_rand_block(td, f, ddir, b);
283		else
284			ret = get_next_seq_block(td, f, ddir, b);
285	} else {
286		io_u->flags |= IO_U_F_BUSY_OK;
287
288		if (td->o.rw_seq == RW_SEQ_SEQ) {
289			ret = get_next_seq_block(td, f, ddir, b);
290			if (ret)
291				ret = get_next_rand_block(td, f, ddir, b);
292		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
293			if (f->last_start != -1ULL)
294				*b = (f->last_start - f->file_offset)
295					/ td->o.min_bs[ddir];
296			else
297				*b = 0;
298			ret = 0;
299		} else {
300			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
301			ret = 1;
302		}
303	}
304
305	return ret;
306}
307
308/*
309 * For random io, generate a random new block and see if it's used. Repeat
310 * until we find a free one. For sequential io, just return the end of
311 * the last io issued.
312 */
313static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
314{
315	struct fio_file *f = io_u->file;
316	unsigned long long b;
317	enum fio_ddir ddir = io_u->ddir;
318	int rw_seq_hit = 0;
319
320	assert(ddir_rw(ddir));
321
322	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
323		rw_seq_hit = 1;
324		td->ddir_seq_nr = td->o.ddir_seq_nr;
325	}
326
327	if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
328		return 1;
329
330	io_u->offset = b * td->o.ba[ddir];
331	if (io_u->offset >= f->io_size) {
332		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
333					io_u->offset, f->io_size);
334		return 1;
335	}
336
337	io_u->offset += f->file_offset;
338	if (io_u->offset >= f->real_file_size) {
339		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
340					io_u->offset, f->real_file_size);
341		return 1;
342	}
343
344	return 0;
345}
346
347static int get_next_offset(struct thread_data *td, struct io_u *io_u)
348{
349	struct prof_io_ops *ops = &td->prof_io_ops;
350
351	if (ops->fill_io_u_off)
352		return ops->fill_io_u_off(td, io_u);
353
354	return __get_next_offset(td, io_u);
355}
356
357static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
358			    unsigned int buflen)
359{
360	struct fio_file *f = io_u->file;
361
362	return io_u->offset + buflen <= f->io_size + td->o.start_offset;
363}
364
365static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
366{
367	const int ddir = io_u->ddir;
368	unsigned int uninitialized_var(buflen);
369	unsigned int minbs, maxbs;
370	unsigned long r, rand_max;
371
372	assert(ddir_rw(ddir));
373
374	minbs = td->o.min_bs[ddir];
375	maxbs = td->o.max_bs[ddir];
376
377	if (minbs == maxbs)
378		return minbs;
379
380	/*
381	 * If we can't satisfy the min block size from here, then fail
382	 */
383	if (!io_u_fits(td, io_u, minbs))
384		return 0;
385
386	if (td->o.use_os_rand)
387		rand_max = OS_RAND_MAX;
388	else
389		rand_max = FRAND_MAX;
390
391	do {
392		if (td->o.use_os_rand)
393			r = os_random_long(&td->bsrange_state);
394		else
395			r = __rand(&td->__bsrange_state);
396
397		if (!td->o.bssplit_nr[ddir]) {
398			buflen = 1 + (unsigned int) ((double) maxbs *
399					(r / (rand_max + 1.0)));
400			if (buflen < minbs)
401				buflen = minbs;
402		} else {
403			long perc = 0;
404			unsigned int i;
405
406			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
407				struct bssplit *bsp = &td->o.bssplit[ddir][i];
408
409				buflen = bsp->bs;
410				perc += bsp->perc;
411				if ((r <= ((rand_max / 100L) * perc)) &&
412				    io_u_fits(td, io_u, buflen))
413					break;
414			}
415		}
416
417		if (!td->o.bs_unaligned && is_power_of_2(minbs))
418			buflen = (buflen + minbs - 1) & ~(minbs - 1);
419
420	} while (!io_u_fits(td, io_u, buflen));
421
422	return buflen;
423}
424
425static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
426{
427	struct prof_io_ops *ops = &td->prof_io_ops;
428
429	if (ops->fill_io_u_size)
430		return ops->fill_io_u_size(td, io_u);
431
432	return __get_next_buflen(td, io_u);
433}
434
435static void set_rwmix_bytes(struct thread_data *td)
436{
437	unsigned int diff;
438
439	/*
440	 * we do time or byte based switch. this is needed because
441	 * buffered writes may issue a lot quicker than they complete,
442	 * whereas reads do not.
443	 */
444	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
445	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
446}
447
448static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
449{
450	unsigned int v;
451	unsigned long r;
452
453	if (td->o.use_os_rand) {
454		r = os_random_long(&td->rwmix_state);
455		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
456	} else {
457		r = __rand(&td->__rwmix_state);
458		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
459	}
460
461	if (v <= td->o.rwmix[DDIR_READ])
462		return DDIR_READ;
463
464	return DDIR_WRITE;
465}
466
467static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
468{
469	enum fio_ddir odir = ddir ^ 1;
470	struct timeval t;
471	long usec;
472
473	assert(ddir_rw(ddir));
474
475	if (td->rate_pending_usleep[ddir] <= 0)
476		return ddir;
477
478	/*
479	 * We have too much pending sleep in this direction. See if we
480	 * should switch.
481	 */
482	if (td_rw(td)) {
483		/*
484		 * Other direction does not have too much pending, switch
485		 */
486		if (td->rate_pending_usleep[odir] < 100000)
487			return odir;
488
489		/*
490		 * Both directions have pending sleep. Sleep the minimum time
491		 * and deduct from both.
492		 */
493		if (td->rate_pending_usleep[ddir] <=
494			td->rate_pending_usleep[odir]) {
495			usec = td->rate_pending_usleep[ddir];
496		} else {
497			usec = td->rate_pending_usleep[odir];
498			ddir = odir;
499		}
500	} else
501		usec = td->rate_pending_usleep[ddir];
502
503	/*
504	 * We are going to sleep, ensure that we flush anything pending as
505	 * not to skew our latency numbers.
506	 *
507	 * Changed to only monitor 'in flight' requests here instead of the
508	 * td->cur_depth, b/c td->cur_depth does not accurately represent
509	 * io's that have been actually submitted to an async engine,
510	 * and cur_depth is meaningless for sync engines.
511	 */
512	if (td->io_u_in_flight) {
513		int fio_unused ret;
514
515		ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
516	}
517
518	fio_gettime(&t, NULL);
519	usec_sleep(td, usec);
520	usec = utime_since_now(&t);
521
522	td->rate_pending_usleep[ddir] -= usec;
523
524	odir = ddir ^ 1;
525	if (td_rw(td) && __should_check_rate(td, odir))
526		td->rate_pending_usleep[odir] -= usec;
527
528	return ddir;
529}
530
531/*
532 * Return the data direction for the next io_u. If the job is a
533 * mixed read/write workload, check the rwmix cycle and switch if
534 * necessary.
535 */
536static enum fio_ddir get_rw_ddir(struct thread_data *td)
537{
538	enum fio_ddir ddir;
539
540	/*
541	 * see if it's time to fsync
542	 */
543	if (td->o.fsync_blocks &&
544	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
545	     td->io_issues[DDIR_WRITE] && should_fsync(td))
546		return DDIR_SYNC;
547
548	/*
549	 * see if it's time to fdatasync
550	 */
551	if (td->o.fdatasync_blocks &&
552	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
553	     td->io_issues[DDIR_WRITE] && should_fsync(td))
554		return DDIR_DATASYNC;
555
556	/*
557	 * see if it's time to sync_file_range
558	 */
559	if (td->sync_file_range_nr &&
560	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
561	     td->io_issues[DDIR_WRITE] && should_fsync(td))
562		return DDIR_SYNC_FILE_RANGE;
563
564	if (td_rw(td)) {
565		/*
566		 * Check if it's time to seed a new data direction.
567		 */
568		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
569			/*
570			 * Put a top limit on how many bytes we do for
571			 * one data direction, to avoid overflowing the
572			 * ranges too much
573			 */
574			ddir = get_rand_ddir(td);
575
576			if (ddir != td->rwmix_ddir)
577				set_rwmix_bytes(td);
578
579			td->rwmix_ddir = ddir;
580		}
581		ddir = td->rwmix_ddir;
582	} else if (td_read(td))
583		ddir = DDIR_READ;
584	else
585		ddir = DDIR_WRITE;
586
587	td->rwmix_ddir = rate_ddir(td, ddir);
588	return td->rwmix_ddir;
589}
590
591static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
592{
593	io_u->ddir = get_rw_ddir(td);
594
595	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
596	    td->o.barrier_blocks &&
597	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
598	     td->io_issues[DDIR_WRITE])
599		io_u->flags |= IO_U_F_BARRIER;
600}
601
602void put_file_log(struct thread_data *td, struct fio_file *f)
603{
604	int ret = put_file(td, f);
605
606	if (ret)
607		td_verror(td, ret, "file close");
608}
609
610void put_io_u(struct thread_data *td, struct io_u *io_u)
611{
612	td_io_u_lock(td);
613
614	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
615		put_file_log(td, io_u->file);
616	io_u->file = NULL;
617	io_u->flags &= ~IO_U_F_FREE_DEF;
618	io_u->flags |= IO_U_F_FREE;
619
620	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
621		td->cur_depth--;
622	flist_del_init(&io_u->list);
623	flist_add(&io_u->list, &td->io_u_freelist);
624	td_io_u_unlock(td);
625	td_io_u_free_notify(td);
626}
627
628void clear_io_u(struct thread_data *td, struct io_u *io_u)
629{
630	io_u->flags &= ~IO_U_F_FLIGHT;
631	put_io_u(td, io_u);
632}
633
634void requeue_io_u(struct thread_data *td, struct io_u **io_u)
635{
636	struct io_u *__io_u = *io_u;
637
638	dprint(FD_IO, "requeue %p\n", __io_u);
639
640	td_io_u_lock(td);
641
642	__io_u->flags |= IO_U_F_FREE;
643	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
644		td->io_issues[__io_u->ddir]--;
645
646	__io_u->flags &= ~IO_U_F_FLIGHT;
647	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
648		td->cur_depth--;
649	flist_del(&__io_u->list);
650	flist_add_tail(&__io_u->list, &td->io_u_requeues);
651	td_io_u_unlock(td);
652	*io_u = NULL;
653}
654
655static int fill_io_u(struct thread_data *td, struct io_u *io_u)
656{
657	if (td->io_ops->flags & FIO_NOIO)
658		goto out;
659
660	set_rw_ddir(td, io_u);
661
662	/*
663	 * fsync() or fdatasync() or trim etc, we are done
664	 */
665	if (!ddir_rw(io_u->ddir))
666		goto out;
667
668	/*
669	 * See if it's time to switch to a new zone
670	 */
671	if (td->zone_bytes >= td->o.zone_size) {
672		td->zone_bytes = 0;
673		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
674		io_u->file->last_pos = io_u->file->file_offset;
675		td->io_skip_bytes += td->o.zone_skip;
676	}
677
678	/*
679	 * No log, let the seq/rand engine retrieve the next buflen and
680	 * position.
681	 */
682	if (get_next_offset(td, io_u)) {
683		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
684		return 1;
685	}
686
687	io_u->buflen = get_next_buflen(td, io_u);
688	if (!io_u->buflen) {
689		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
690		return 1;
691	}
692
693	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
694		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
695		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
696				io_u->buflen, io_u->file->real_file_size);
697		return 1;
698	}
699
700	/*
701	 * mark entry before potentially trimming io_u
702	 */
703	if (td_random(td) && file_randommap(td, io_u->file))
704		mark_random_map(td, io_u);
705
706	/*
707	 * If using a write iolog, store this entry.
708	 */
709out:
710	dprint_io_u(io_u, "fill_io_u");
711	td->zone_bytes += io_u->buflen;
712	log_io_u(td, io_u);
713	return 0;
714}
715
716static void __io_u_mark_map(unsigned int *map, unsigned int nr)
717{
718	int idx = 0;
719
720	switch (nr) {
721	default:
722		idx = 6;
723		break;
724	case 33 ... 64:
725		idx = 5;
726		break;
727	case 17 ... 32:
728		idx = 4;
729		break;
730	case 9 ... 16:
731		idx = 3;
732		break;
733	case 5 ... 8:
734		idx = 2;
735		break;
736	case 1 ... 4:
737		idx = 1;
738	case 0:
739		break;
740	}
741
742	map[idx]++;
743}
744
745void io_u_mark_submit(struct thread_data *td, unsigned int nr)
746{
747	__io_u_mark_map(td->ts.io_u_submit, nr);
748	td->ts.total_submit++;
749}
750
751void io_u_mark_complete(struct thread_data *td, unsigned int nr)
752{
753	__io_u_mark_map(td->ts.io_u_complete, nr);
754	td->ts.total_complete++;
755}
756
757void io_u_mark_depth(struct thread_data *td, unsigned int nr)
758{
759	int idx = 0;
760
761	switch (td->cur_depth) {
762	default:
763		idx = 6;
764		break;
765	case 32 ... 63:
766		idx = 5;
767		break;
768	case 16 ... 31:
769		idx = 4;
770		break;
771	case 8 ... 15:
772		idx = 3;
773		break;
774	case 4 ... 7:
775		idx = 2;
776		break;
777	case 2 ... 3:
778		idx = 1;
779	case 1:
780		break;
781	}
782
783	td->ts.io_u_map[idx] += nr;
784}
785
786static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
787{
788	int idx = 0;
789
790	assert(usec < 1000);
791
792	switch (usec) {
793	case 750 ... 999:
794		idx = 9;
795		break;
796	case 500 ... 749:
797		idx = 8;
798		break;
799	case 250 ... 499:
800		idx = 7;
801		break;
802	case 100 ... 249:
803		idx = 6;
804		break;
805	case 50 ... 99:
806		idx = 5;
807		break;
808	case 20 ... 49:
809		idx = 4;
810		break;
811	case 10 ... 19:
812		idx = 3;
813		break;
814	case 4 ... 9:
815		idx = 2;
816		break;
817	case 2 ... 3:
818		idx = 1;
819	case 0 ... 1:
820		break;
821	}
822
823	assert(idx < FIO_IO_U_LAT_U_NR);
824	td->ts.io_u_lat_u[idx]++;
825}
826
827static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
828{
829	int idx = 0;
830
831	switch (msec) {
832	default:
833		idx = 11;
834		break;
835	case 1000 ... 1999:
836		idx = 10;
837		break;
838	case 750 ... 999:
839		idx = 9;
840		break;
841	case 500 ... 749:
842		idx = 8;
843		break;
844	case 250 ... 499:
845		idx = 7;
846		break;
847	case 100 ... 249:
848		idx = 6;
849		break;
850	case 50 ... 99:
851		idx = 5;
852		break;
853	case 20 ... 49:
854		idx = 4;
855		break;
856	case 10 ... 19:
857		idx = 3;
858		break;
859	case 4 ... 9:
860		idx = 2;
861		break;
862	case 2 ... 3:
863		idx = 1;
864	case 0 ... 1:
865		break;
866	}
867
868	assert(idx < FIO_IO_U_LAT_M_NR);
869	td->ts.io_u_lat_m[idx]++;
870}
871
872static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
873{
874	if (usec < 1000)
875		io_u_mark_lat_usec(td, usec);
876	else
877		io_u_mark_lat_msec(td, usec / 1000);
878}
879
880/*
881 * Get next file to service by choosing one at random
882 */
883static struct fio_file *get_next_file_rand(struct thread_data *td,
884					   enum fio_file_flags goodf,
885					   enum fio_file_flags badf)
886{
887	struct fio_file *f;
888	int fno;
889
890	do {
891		int opened = 0;
892		unsigned long r;
893
894		if (td->o.use_os_rand) {
895			r = os_random_long(&td->next_file_state);
896			fno = (unsigned int) ((double) td->o.nr_files
897				* (r / (OS_RAND_MAX + 1.0)));
898		} else {
899			r = __rand(&td->__next_file_state);
900			fno = (unsigned int) ((double) td->o.nr_files
901				* (r / (FRAND_MAX + 1.0)));
902		}
903
904		f = td->files[fno];
905		if (fio_file_done(f))
906			continue;
907
908		if (!fio_file_open(f)) {
909			int err;
910
911			err = td_io_open_file(td, f);
912			if (err)
913				continue;
914			opened = 1;
915		}
916
917		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
918			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
919			return f;
920		}
921		if (opened)
922			td_io_close_file(td, f);
923	} while (1);
924}
925
926/*
927 * Get next file to service by doing round robin between all available ones
928 */
929static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
930					 int badf)
931{
932	unsigned int old_next_file = td->next_file;
933	struct fio_file *f;
934
935	do {
936		int opened = 0;
937
938		f = td->files[td->next_file];
939
940		td->next_file++;
941		if (td->next_file >= td->o.nr_files)
942			td->next_file = 0;
943
944		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
945		if (fio_file_done(f)) {
946			f = NULL;
947			continue;
948		}
949
950		if (!fio_file_open(f)) {
951			int err;
952
953			err = td_io_open_file(td, f);
954			if (err) {
955				dprint(FD_FILE, "error %d on open of %s\n",
956					err, f->file_name);
957				f = NULL;
958				continue;
959			}
960			opened = 1;
961		}
962
963		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
964								f->flags);
965		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
966			break;
967
968		if (opened)
969			td_io_close_file(td, f);
970
971		f = NULL;
972	} while (td->next_file != old_next_file);
973
974	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
975	return f;
976}
977
978static struct fio_file *__get_next_file(struct thread_data *td)
979{
980	struct fio_file *f;
981
982	assert(td->o.nr_files <= td->files_index);
983
984	if (td->nr_done_files >= td->o.nr_files) {
985		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
986				" nr_files=%d\n", td->nr_open_files,
987						  td->nr_done_files,
988						  td->o.nr_files);
989		return NULL;
990	}
991
992	f = td->file_service_file;
993	if (f && fio_file_open(f) && !fio_file_closing(f)) {
994		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
995			goto out;
996		if (td->file_service_left--)
997			goto out;
998	}
999
1000	if (td->o.file_service_type == FIO_FSERVICE_RR ||
1001	    td->o.file_service_type == FIO_FSERVICE_SEQ)
1002		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1003	else
1004		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1005
1006	td->file_service_file = f;
1007	td->file_service_left = td->file_service_nr - 1;
1008out:
1009	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1010	return f;
1011}
1012
1013static struct fio_file *get_next_file(struct thread_data *td)
1014{
1015	struct prof_io_ops *ops = &td->prof_io_ops;
1016
1017	if (ops->get_next_file)
1018		return ops->get_next_file(td);
1019
1020	return __get_next_file(td);
1021}
1022
1023static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1024{
1025	struct fio_file *f;
1026
1027	do {
1028		f = get_next_file(td);
1029		if (!f)
1030			return 1;
1031
1032		io_u->file = f;
1033		get_file(f);
1034
1035		if (!fill_io_u(td, io_u))
1036			break;
1037
1038		put_file_log(td, f);
1039		td_io_close_file(td, f);
1040		io_u->file = NULL;
1041		fio_file_set_done(f);
1042		td->nr_done_files++;
1043		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1044					td->nr_done_files, td->o.nr_files);
1045	} while (1);
1046
1047	return 0;
1048}
1049
1050
1051struct io_u *__get_io_u(struct thread_data *td)
1052{
1053	struct io_u *io_u = NULL;
1054
1055	td_io_u_lock(td);
1056
1057again:
1058	if (!flist_empty(&td->io_u_requeues))
1059		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1060	else if (!queue_full(td)) {
1061		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1062
1063		io_u->buflen = 0;
1064		io_u->resid = 0;
1065		io_u->file = NULL;
1066		io_u->end_io = NULL;
1067	}
1068
1069	if (io_u) {
1070		assert(io_u->flags & IO_U_F_FREE);
1071		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1072		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1073
1074		io_u->error = 0;
1075		flist_del(&io_u->list);
1076		flist_add(&io_u->list, &td->io_u_busylist);
1077		td->cur_depth++;
1078		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1079	} else if (td->o.verify_async) {
1080		/*
1081		 * We ran out, wait for async verify threads to finish and
1082		 * return one
1083		 */
1084		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1085		goto again;
1086	}
1087
1088	td_io_u_unlock(td);
1089	return io_u;
1090}
1091
1092static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1093{
1094	if (td->o.trim_backlog && td->trim_entries) {
1095		int get_trim = 0;
1096
1097		if (td->trim_batch) {
1098			td->trim_batch--;
1099			get_trim = 1;
1100		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1101			 td->last_ddir != DDIR_READ) {
1102			td->trim_batch = td->o.trim_batch;
1103			if (!td->trim_batch)
1104				td->trim_batch = td->o.trim_backlog;
1105			get_trim = 1;
1106		}
1107
1108		if (get_trim && !get_next_trim(td, io_u))
1109			return 1;
1110	}
1111
1112	return 0;
1113}
1114
1115static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1116{
1117	if (td->o.verify_backlog && td->io_hist_len) {
1118		int get_verify = 0;
1119
1120		if (td->verify_batch) {
1121			td->verify_batch--;
1122			get_verify = 1;
1123		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
1124			 td->last_ddir != DDIR_READ) {
1125			td->verify_batch = td->o.verify_batch;
1126			if (!td->verify_batch)
1127				td->verify_batch = td->o.verify_backlog;
1128			get_verify = 1;
1129		}
1130
1131		if (get_verify && !get_next_verify(td, io_u))
1132			return 1;
1133	}
1134
1135	return 0;
1136}
1137
1138/*
1139 * Fill offset and start time into the buffer content, to prevent too
1140 * easy compressible data for simple de-dupe attempts. Do this for every
1141 * 512b block in the range, since that should be the smallest block size
1142 * we can expect from a device.
1143 */
1144static void small_content_scramble(struct io_u *io_u)
1145{
1146	unsigned int i, nr_blocks = io_u->buflen / 512;
1147	unsigned long long boffset;
1148	unsigned int offset;
1149	void *p, *end;
1150
1151	if (!nr_blocks)
1152		return;
1153
1154	p = io_u->xfer_buf;
1155	boffset = io_u->offset;
1156	io_u->buf_filled_len = 0;
1157
1158	for (i = 0; i < nr_blocks; i++) {
1159		/*
1160		 * Fill the byte offset into a "random" start offset of
1161		 * the buffer, given by the product of the usec time
1162		 * and the actual offset.
1163		 */
1164		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1165		offset &= ~(sizeof(unsigned long long) - 1);
1166		if (offset >= 512 - sizeof(unsigned long long))
1167			offset -= sizeof(unsigned long long);
1168		memcpy(p + offset, &boffset, sizeof(boffset));
1169
1170		end = p + 512 - sizeof(io_u->start_time);
1171		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1172		p += 512;
1173		boffset += 512;
1174	}
1175}
1176
1177/*
1178 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1179 * etc. The returned io_u is fully ready to be prepped and submitted.
1180 */
1181struct io_u *get_io_u(struct thread_data *td)
1182{
1183	struct fio_file *f;
1184	struct io_u *io_u;
1185	int do_scramble = 0;
1186
1187	io_u = __get_io_u(td);
1188	if (!io_u) {
1189		dprint(FD_IO, "__get_io_u failed\n");
1190		return NULL;
1191	}
1192
1193	if (check_get_verify(td, io_u))
1194		goto out;
1195	if (check_get_trim(td, io_u))
1196		goto out;
1197
1198	/*
1199	 * from a requeue, io_u already setup
1200	 */
1201	if (io_u->file)
1202		goto out;
1203
1204	/*
1205	 * If using an iolog, grab next piece if any available.
1206	 */
1207	if (td->o.read_iolog_file) {
1208		if (read_iolog_get(td, io_u))
1209			goto err_put;
1210	} else if (set_io_u_file(td, io_u)) {
1211		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1212		goto err_put;
1213	}
1214
1215	f = io_u->file;
1216	assert(fio_file_open(f));
1217
1218	if (ddir_rw(io_u->ddir)) {
1219		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1220			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1221			goto err_put;
1222		}
1223
1224		f->last_start = io_u->offset;
1225		f->last_pos = io_u->offset + io_u->buflen;
1226
1227		if (io_u->ddir == DDIR_WRITE) {
1228			if (td->o.verify != VERIFY_NONE)
1229				populate_verify_io_u(td, io_u);
1230			else if (td->o.refill_buffers)
1231				io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1232			else if (td->o.scramble_buffers)
1233				do_scramble = 1;
1234		} else if (io_u->ddir == DDIR_READ) {
1235			/*
1236			 * Reset the buf_filled parameters so next time if the
1237			 * buffer is used for writes it is refilled.
1238			 */
1239			io_u->buf_filled_len = 0;
1240		}
1241	}
1242
1243	/*
1244	 * Set io data pointers.
1245	 */
1246	io_u->xfer_buf = io_u->buf;
1247	io_u->xfer_buflen = io_u->buflen;
1248
1249out:
1250	assert(io_u->file);
1251	if (!td_io_prep(td, io_u)) {
1252		if (!td->o.disable_slat)
1253			fio_gettime(&io_u->start_time, NULL);
1254		if (do_scramble)
1255			small_content_scramble(io_u);
1256		return io_u;
1257	}
1258err_put:
1259	dprint(FD_IO, "get_io_u failed\n");
1260	put_io_u(td, io_u);
1261	return NULL;
1262}
1263
1264void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1265{
1266	const char *msg[] = { "read", "write", "sync", "datasync",
1267				"sync_file_range", "wait", "trim" };
1268
1269
1270
1271	log_err("fio: io_u error");
1272
1273	if (io_u->file)
1274		log_err(" on file %s", io_u->file->file_name);
1275
1276	log_err(": %s\n", strerror(io_u->error));
1277
1278	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1279					io_u->offset, io_u->xfer_buflen);
1280
1281	if (!td->error)
1282		td_verror(td, io_u->error, "io_u error");
1283}
1284
1285static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1286				  struct io_completion_data *icd,
1287				  const enum fio_ddir idx, unsigned int bytes)
1288{
1289	unsigned long uninitialized_var(lusec);
1290
1291	if (!icd->account)
1292		return;
1293
1294	if (!td->o.disable_clat || !td->o.disable_bw)
1295		lusec = utime_since(&io_u->issue_time, &icd->time);
1296
1297	if (!td->o.disable_lat) {
1298		unsigned long tusec;
1299
1300		tusec = utime_since(&io_u->start_time, &icd->time);
1301		add_lat_sample(td, idx, tusec, bytes);
1302	}
1303
1304	if (!td->o.disable_clat) {
1305		add_clat_sample(td, idx, lusec, bytes);
1306		io_u_mark_latency(td, lusec);
1307	}
1308
1309	if (!td->o.disable_bw)
1310		add_bw_sample(td, idx, bytes, &icd->time);
1311
1312	add_iops_sample(td, idx, &icd->time);
1313}
1314
1315static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1316{
1317	unsigned long long secs, remainder, bps, bytes;
1318	bytes = td->this_io_bytes[ddir];
1319	bps = td->rate_bps[ddir];
1320	secs = bytes / bps;
1321	remainder = bytes % bps;
1322	return remainder * 1000000 / bps + secs * 1000000;
1323}
1324
1325static void io_completed(struct thread_data *td, struct io_u *io_u,
1326			 struct io_completion_data *icd)
1327{
1328	/*
1329	 * Older gcc's are too dumb to realize that usec is always used
1330	 * initialized, silence that warning.
1331	 */
1332	unsigned long uninitialized_var(usec);
1333	struct fio_file *f;
1334
1335	dprint_io_u(io_u, "io complete");
1336
1337	td_io_u_lock(td);
1338	assert(io_u->flags & IO_U_F_FLIGHT);
1339	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1340	td_io_u_unlock(td);
1341
1342	if (ddir_sync(io_u->ddir)) {
1343		td->last_was_sync = 1;
1344		f = io_u->file;
1345		if (f) {
1346			f->first_write = -1ULL;
1347			f->last_write = -1ULL;
1348		}
1349		return;
1350	}
1351
1352	td->last_was_sync = 0;
1353	td->last_ddir = io_u->ddir;
1354
1355	if (!io_u->error && ddir_rw(io_u->ddir)) {
1356		unsigned int bytes = io_u->buflen - io_u->resid;
1357		const enum fio_ddir idx = io_u->ddir;
1358		const enum fio_ddir odx = io_u->ddir ^ 1;
1359		int ret;
1360
1361		td->io_blocks[idx]++;
1362		td->this_io_blocks[idx]++;
1363		td->io_bytes[idx] += bytes;
1364		td->this_io_bytes[idx] += bytes;
1365
1366		if (idx == DDIR_WRITE) {
1367			f = io_u->file;
1368			if (f) {
1369				if (f->first_write == -1ULL ||
1370				    io_u->offset < f->first_write)
1371					f->first_write = io_u->offset;
1372				if (f->last_write == -1ULL ||
1373				    ((io_u->offset + bytes) > f->last_write))
1374					f->last_write = io_u->offset + bytes;
1375			}
1376		}
1377
1378		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1379					   td->runstate == TD_VERIFYING)) {
1380			account_io_completion(td, io_u, icd, idx, bytes);
1381
1382			if (__should_check_rate(td, idx)) {
1383				td->rate_pending_usleep[idx] =
1384					(usec_for_io(td, idx) -
1385					 utime_since_now(&td->start));
1386			}
1387			if (__should_check_rate(td, odx))
1388				td->rate_pending_usleep[odx] =
1389					(usec_for_io(td, odx) -
1390					 utime_since_now(&td->start));
1391		}
1392
1393		if (td_write(td) && idx == DDIR_WRITE &&
1394		    td->o.do_verify &&
1395		    td->o.verify != VERIFY_NONE)
1396			log_io_piece(td, io_u);
1397
1398		icd->bytes_done[idx] += bytes;
1399
1400		if (io_u->end_io) {
1401			ret = io_u->end_io(td, io_u);
1402			if (ret && !icd->error)
1403				icd->error = ret;
1404		}
1405	} else if (io_u->error) {
1406		icd->error = io_u->error;
1407		io_u_log_error(td, io_u);
1408	}
1409	if (icd->error && td_non_fatal_error(icd->error) &&
1410           (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1411		/*
1412		 * If there is a non_fatal error, then add to the error count
1413		 * and clear all the errors.
1414		 */
1415		update_error_count(td, icd->error);
1416		td_clear_error(td);
1417		icd->error = 0;
1418		io_u->error = 0;
1419	}
1420}
1421
1422static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1423		     int nr)
1424{
1425	if (!td->o.disable_clat || !td->o.disable_bw)
1426		fio_gettime(&icd->time, NULL);
1427
1428	icd->nr = nr;
1429	icd->account = 1;
1430
1431	icd->error = 0;
1432	icd->bytes_done[0] = icd->bytes_done[1] = 0;
1433}
1434
1435static void ios_completed(struct thread_data *td,
1436			  struct io_completion_data *icd)
1437{
1438	struct io_u *io_u;
1439	int i;
1440
1441	for (i = 0; i < icd->nr; i++) {
1442		io_u = td->io_ops->event(td, i);
1443
1444		io_completed(td, io_u, icd);
1445
1446		if (!(io_u->flags & IO_U_F_FREE_DEF))
1447			put_io_u(td, io_u);
1448
1449		icd->account = 0;
1450	}
1451}
1452
1453/*
1454 * Complete a single io_u for the sync engines.
1455 */
1456int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1457		       unsigned long *bytes)
1458{
1459	struct io_completion_data icd;
1460
1461	init_icd(td, &icd, 1);
1462	io_completed(td, io_u, &icd);
1463
1464	if (!(io_u->flags & IO_U_F_FREE_DEF))
1465		put_io_u(td, io_u);
1466
1467	if (icd.error) {
1468		td_verror(td, icd.error, "io_u_sync_complete");
1469		return -1;
1470	}
1471
1472	if (bytes) {
1473		bytes[0] += icd.bytes_done[0];
1474		bytes[1] += icd.bytes_done[1];
1475	}
1476
1477	return 0;
1478}
1479
1480/*
1481 * Called to complete min_events number of io for the async engines.
1482 */
1483int io_u_queued_complete(struct thread_data *td, int min_evts,
1484			 unsigned long *bytes)
1485{
1486	struct io_completion_data icd;
1487	struct timespec *tvp = NULL;
1488	int ret;
1489	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1490
1491	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1492
1493	if (!min_evts)
1494		tvp = &ts;
1495
1496	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1497	if (ret < 0) {
1498		td_verror(td, -ret, "td_io_getevents");
1499		return ret;
1500	} else if (!ret)
1501		return ret;
1502
1503	init_icd(td, &icd, ret);
1504	ios_completed(td, &icd);
1505	if (icd.error) {
1506		td_verror(td, icd.error, "io_u_queued_complete");
1507		return -1;
1508	}
1509
1510	if (bytes) {
1511		bytes[0] += icd.bytes_done[0];
1512		bytes[1] += icd.bytes_done[1];
1513	}
1514
1515	return 0;
1516}
1517
1518/*
1519 * Call when io_u is really queued, to update the submission latency.
1520 */
1521void io_u_queued(struct thread_data *td, struct io_u *io_u)
1522{
1523	if (!td->o.disable_slat) {
1524		unsigned long slat_time;
1525
1526		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1527		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1528	}
1529}
1530
1531/*
1532 * "randomly" fill the buffer contents
1533 */
1534void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1535		      unsigned int max_bs)
1536{
1537	io_u->buf_filled_len = 0;
1538
1539	if (!td->o.zero_buffers)
1540		fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1541	else
1542		memset(io_u->buf, 0, max_bs);
1543}
1544