io_u.c revision 7477673323a943b99ea203bb9434661d13a0159c
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13#include "lib/axmap.h"
14
15struct io_completion_data {
16	int nr;				/* input */
17
18	int error;			/* output */
19	unsigned long bytes_done[DDIR_RWDIR_CNT];	/* output */
20	struct timeval time;		/* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const unsigned long long block)
28{
29	return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37	unsigned int min_bs = td->o.rw_min_bs;
38	struct fio_file *f = io_u->file;
39	unsigned long long block;
40	unsigned int nr_blocks;
41
42	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
43	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45	if (!(io_u->flags & IO_U_F_BUSY_OK))
46		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47
48	if ((nr_blocks * min_bs) < io_u->buflen)
49		io_u->buflen = nr_blocks * min_bs;
50}
51
52static uint64_t last_block(struct thread_data *td, struct fio_file *f,
53			   enum fio_ddir ddir)
54{
55	uint64_t max_blocks;
56	uint64_t max_size;
57
58	assert(ddir_rw(ddir));
59
60	/*
61	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
62	 */
63	max_size = f->io_size;
64	if (max_size > f->real_file_size)
65		max_size = f->real_file_size;
66
67	if (td->o.zone_range)
68		max_size = td->o.zone_range;
69
70	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
71	if (!max_blocks)
72		return 0;
73
74	return max_blocks;
75}
76
77static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
78				  enum fio_ddir ddir, unsigned long long *b)
79{
80	unsigned long long r, lastb;
81
82	lastb = last_block(td, f, ddir);
83	if (!lastb)
84		return 1;
85
86	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
87		unsigned long long rmax;
88
89		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
90
91		if (td->o.use_os_rand) {
92			rmax = OS_RAND_MAX;
93			r = os_random_long(&td->random_state);
94		} else {
95			rmax = FRAND_MAX;
96			r = __rand(&td->__random_state);
97		}
98
99		dprint(FD_RANDOM, "off rand %llu\n", r);
100
101		*b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
102	} else {
103		uint64_t off = 0;
104
105		if (lfsr_next(&f->lfsr, &off, lastb))
106			return 1;
107
108		*b = off;
109	}
110
111	/*
112	 * if we are not maintaining a random map, we are done.
113	 */
114	if (!file_randommap(td, f))
115		goto ret;
116
117	/*
118	 * calculate map offset and check if it's free
119	 */
120	if (random_map_free(f, *b))
121		goto ret;
122
123	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", *b);
124
125	*b = axmap_next_free(f->io_axmap, *b);
126	if (*b == (uint64_t) -1ULL)
127		return 1;
128ret:
129	return 0;
130}
131
132static int __get_next_rand_offset_zipf(struct thread_data *td,
133				       struct fio_file *f, enum fio_ddir ddir,
134				       unsigned long long *b)
135{
136	*b = zipf_next(&f->zipf);
137	return 0;
138}
139
140static int __get_next_rand_offset_pareto(struct thread_data *td,
141					 struct fio_file *f, enum fio_ddir ddir,
142					 unsigned long long *b)
143{
144	*b = pareto_next(&f->zipf);
145	return 0;
146}
147
148static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
149				enum fio_ddir ddir, unsigned long long *b)
150{
151	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
152		return __get_next_rand_offset(td, f, ddir, b);
153	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
154		return __get_next_rand_offset_zipf(td, f, ddir, b);
155	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
156		return __get_next_rand_offset_pareto(td, f, ddir, b);
157
158	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
159	return 1;
160}
161
162static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
163			       enum fio_ddir ddir, unsigned long long *b)
164{
165	if (!get_next_rand_offset(td, f, ddir, b))
166		return 0;
167
168	if (td->o.time_based) {
169		fio_file_reset(f);
170		if (!get_next_rand_offset(td, f, ddir, b))
171			return 0;
172	}
173
174	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
175			f->file_name, f->last_pos, f->real_file_size);
176	return 1;
177}
178
179static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
180			       enum fio_ddir ddir, unsigned long long *offset)
181{
182	assert(ddir_rw(ddir));
183
184	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
185		f->last_pos = f->last_pos - f->io_size;
186
187	if (f->last_pos < f->real_file_size) {
188		unsigned long long pos;
189
190		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
191			f->last_pos = f->real_file_size;
192
193		pos = f->last_pos - f->file_offset;
194		if (pos)
195			pos += td->o.ddir_seq_add;
196
197		*offset = pos;
198		return 0;
199	}
200
201	return 1;
202}
203
204static int get_next_block(struct thread_data *td, struct io_u *io_u,
205			  enum fio_ddir ddir, int rw_seq)
206{
207	struct fio_file *f = io_u->file;
208	unsigned long long b, offset;
209	int ret;
210
211	assert(ddir_rw(ddir));
212
213	b = offset = -1ULL;
214
215	if (rw_seq) {
216		if (td_random(td))
217			ret = get_next_rand_block(td, f, ddir, &b);
218		else
219			ret = get_next_seq_offset(td, f, ddir, &offset);
220	} else {
221		io_u->flags |= IO_U_F_BUSY_OK;
222
223		if (td->o.rw_seq == RW_SEQ_SEQ) {
224			ret = get_next_seq_offset(td, f, ddir, &offset);
225			if (ret)
226				ret = get_next_rand_block(td, f, ddir, &b);
227		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
228			if (f->last_start != -1ULL)
229				offset = f->last_start - f->file_offset;
230			else
231				offset = 0;
232			ret = 0;
233		} else {
234			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
235			ret = 1;
236		}
237	}
238
239	if (!ret) {
240		if (offset != -1ULL)
241			io_u->offset = offset;
242		else if (b != -1ULL)
243			io_u->offset = b * td->o.ba[ddir];
244		else {
245			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
246								offset, b);
247			ret = 1;
248		}
249	}
250
251	return ret;
252}
253
254/*
255 * For random io, generate a random new block and see if it's used. Repeat
256 * until we find a free one. For sequential io, just return the end of
257 * the last io issued.
258 */
259static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
260{
261	struct fio_file *f = io_u->file;
262	enum fio_ddir ddir = io_u->ddir;
263	int rw_seq_hit = 0;
264
265	assert(ddir_rw(ddir));
266
267	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
268		rw_seq_hit = 1;
269		td->ddir_seq_nr = td->o.ddir_seq_nr;
270	}
271
272	if (get_next_block(td, io_u, ddir, rw_seq_hit))
273		return 1;
274
275	if (io_u->offset >= f->io_size) {
276		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
277					io_u->offset, f->io_size);
278		return 1;
279	}
280
281	io_u->offset += f->file_offset;
282	if (io_u->offset >= f->real_file_size) {
283		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
284					io_u->offset, f->real_file_size);
285		return 1;
286	}
287
288	return 0;
289}
290
291static int get_next_offset(struct thread_data *td, struct io_u *io_u)
292{
293	if (td->flags & TD_F_PROFILE_OPS) {
294		struct prof_io_ops *ops = &td->prof_io_ops;
295
296		if (ops->fill_io_u_off)
297			return ops->fill_io_u_off(td, io_u);
298	}
299
300	return __get_next_offset(td, io_u);
301}
302
303static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
304			    unsigned int buflen)
305{
306	struct fio_file *f = io_u->file;
307
308	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
309}
310
311static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
312{
313	const int ddir = io_u->ddir;
314	unsigned int buflen = 0;
315	unsigned int minbs, maxbs;
316	unsigned long r, rand_max;
317
318	assert(ddir_rw(ddir));
319
320	minbs = td->o.min_bs[ddir];
321	maxbs = td->o.max_bs[ddir];
322
323	if (minbs == maxbs)
324		return minbs;
325
326	/*
327	 * If we can't satisfy the min block size from here, then fail
328	 */
329	if (!io_u_fits(td, io_u, minbs))
330		return 0;
331
332	if (td->o.use_os_rand)
333		rand_max = OS_RAND_MAX;
334	else
335		rand_max = FRAND_MAX;
336
337	do {
338		if (td->o.use_os_rand)
339			r = os_random_long(&td->bsrange_state);
340		else
341			r = __rand(&td->__bsrange_state);
342
343		if (!td->o.bssplit_nr[ddir]) {
344			buflen = 1 + (unsigned int) ((double) maxbs *
345					(r / (rand_max + 1.0)));
346			if (buflen < minbs)
347				buflen = minbs;
348		} else {
349			long perc = 0;
350			unsigned int i;
351
352			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
353				struct bssplit *bsp = &td->o.bssplit[ddir][i];
354
355				buflen = bsp->bs;
356				perc += bsp->perc;
357				if ((r <= ((rand_max / 100L) * perc)) &&
358				    io_u_fits(td, io_u, buflen))
359					break;
360			}
361		}
362
363		if (!td->o.bs_unaligned && is_power_of_2(minbs))
364			buflen = (buflen + minbs - 1) & ~(minbs - 1);
365
366	} while (!io_u_fits(td, io_u, buflen));
367
368	return buflen;
369}
370
371static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
372{
373	if (td->flags & TD_F_PROFILE_OPS) {
374		struct prof_io_ops *ops = &td->prof_io_ops;
375
376		if (ops->fill_io_u_size)
377			return ops->fill_io_u_size(td, io_u);
378	}
379
380	return __get_next_buflen(td, io_u);
381}
382
383static void set_rwmix_bytes(struct thread_data *td)
384{
385	unsigned int diff;
386
387	/*
388	 * we do time or byte based switch. this is needed because
389	 * buffered writes may issue a lot quicker than they complete,
390	 * whereas reads do not.
391	 */
392	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
393	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
394}
395
396static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
397{
398	unsigned int v;
399	unsigned long r;
400
401	if (td->o.use_os_rand) {
402		r = os_random_long(&td->rwmix_state);
403		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
404	} else {
405		r = __rand(&td->__rwmix_state);
406		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
407	}
408
409	if (v <= td->o.rwmix[DDIR_READ])
410		return DDIR_READ;
411
412	return DDIR_WRITE;
413}
414
415static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
416{
417	enum fio_ddir odir = ddir ^ 1;
418	struct timeval t;
419	long usec;
420
421	assert(ddir_rw(ddir));
422
423	if (td->rate_pending_usleep[ddir] <= 0)
424		return ddir;
425
426	/*
427	 * We have too much pending sleep in this direction. See if we
428	 * should switch.
429	 */
430	if (td_rw(td)) {
431		/*
432		 * Other direction does not have too much pending, switch
433		 */
434		if (td->rate_pending_usleep[odir] < 100000)
435			return odir;
436
437		/*
438		 * Both directions have pending sleep. Sleep the minimum time
439		 * and deduct from both.
440		 */
441		if (td->rate_pending_usleep[ddir] <=
442			td->rate_pending_usleep[odir]) {
443			usec = td->rate_pending_usleep[ddir];
444		} else {
445			usec = td->rate_pending_usleep[odir];
446			ddir = odir;
447		}
448	} else
449		usec = td->rate_pending_usleep[ddir];
450
451	/*
452	 * We are going to sleep, ensure that we flush anything pending as
453	 * not to skew our latency numbers.
454	 *
455	 * Changed to only monitor 'in flight' requests here instead of the
456	 * td->cur_depth, b/c td->cur_depth does not accurately represent
457	 * io's that have been actually submitted to an async engine,
458	 * and cur_depth is meaningless for sync engines.
459	 */
460	if (td->io_u_in_flight) {
461		int fio_unused ret;
462
463		ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
464	}
465
466	fio_gettime(&t, NULL);
467	usec_sleep(td, usec);
468	usec = utime_since_now(&t);
469
470	td->rate_pending_usleep[ddir] -= usec;
471
472	odir = ddir ^ 1;
473	if (td_rw(td) && __should_check_rate(td, odir))
474		td->rate_pending_usleep[odir] -= usec;
475
476	if (ddir_trim(ddir))
477		return ddir;
478	return ddir;
479}
480
481/*
482 * Return the data direction for the next io_u. If the job is a
483 * mixed read/write workload, check the rwmix cycle and switch if
484 * necessary.
485 */
486static enum fio_ddir get_rw_ddir(struct thread_data *td)
487{
488	enum fio_ddir ddir;
489
490	/*
491	 * see if it's time to fsync
492	 */
493	if (td->o.fsync_blocks &&
494	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
495	     td->io_issues[DDIR_WRITE] && should_fsync(td))
496		return DDIR_SYNC;
497
498	/*
499	 * see if it's time to fdatasync
500	 */
501	if (td->o.fdatasync_blocks &&
502	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
503	     td->io_issues[DDIR_WRITE] && should_fsync(td))
504		return DDIR_DATASYNC;
505
506	/*
507	 * see if it's time to sync_file_range
508	 */
509	if (td->sync_file_range_nr &&
510	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
511	     td->io_issues[DDIR_WRITE] && should_fsync(td))
512		return DDIR_SYNC_FILE_RANGE;
513
514	if (td_rw(td)) {
515		/*
516		 * Check if it's time to seed a new data direction.
517		 */
518		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
519			/*
520			 * Put a top limit on how many bytes we do for
521			 * one data direction, to avoid overflowing the
522			 * ranges too much
523			 */
524			ddir = get_rand_ddir(td);
525
526			if (ddir != td->rwmix_ddir)
527				set_rwmix_bytes(td);
528
529			td->rwmix_ddir = ddir;
530		}
531		ddir = td->rwmix_ddir;
532	} else if (td_read(td))
533		ddir = DDIR_READ;
534	else if (td_write(td))
535		ddir = DDIR_WRITE;
536	else
537		ddir = DDIR_TRIM;
538
539	td->rwmix_ddir = rate_ddir(td, ddir);
540	return td->rwmix_ddir;
541}
542
543static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
544{
545	io_u->ddir = get_rw_ddir(td);
546
547	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
548	    td->o.barrier_blocks &&
549	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
550	     td->io_issues[DDIR_WRITE])
551		io_u->flags |= IO_U_F_BARRIER;
552}
553
554void put_file_log(struct thread_data *td, struct fio_file *f)
555{
556	int ret = put_file(td, f);
557
558	if (ret)
559		td_verror(td, ret, "file close");
560}
561
562void put_io_u(struct thread_data *td, struct io_u *io_u)
563{
564	td_io_u_lock(td);
565
566	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
567		put_file_log(td, io_u->file);
568	io_u->file = NULL;
569	io_u->flags &= ~IO_U_F_FREE_DEF;
570	io_u->flags |= IO_U_F_FREE;
571
572	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
573		td->cur_depth--;
574	flist_del_init(&io_u->list);
575	flist_add(&io_u->list, &td->io_u_freelist);
576	td_io_u_unlock(td);
577	td_io_u_free_notify(td);
578}
579
580void clear_io_u(struct thread_data *td, struct io_u *io_u)
581{
582	io_u->flags &= ~IO_U_F_FLIGHT;
583	put_io_u(td, io_u);
584}
585
586void requeue_io_u(struct thread_data *td, struct io_u **io_u)
587{
588	struct io_u *__io_u = *io_u;
589
590	dprint(FD_IO, "requeue %p\n", __io_u);
591
592	td_io_u_lock(td);
593
594	__io_u->flags |= IO_U_F_FREE;
595	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
596		td->io_issues[__io_u->ddir]--;
597
598	__io_u->flags &= ~IO_U_F_FLIGHT;
599	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
600		td->cur_depth--;
601	flist_del(&__io_u->list);
602	flist_add_tail(&__io_u->list, &td->io_u_requeues);
603	td_io_u_unlock(td);
604	*io_u = NULL;
605}
606
607static int fill_io_u(struct thread_data *td, struct io_u *io_u)
608{
609	if (td->io_ops->flags & FIO_NOIO)
610		goto out;
611
612	set_rw_ddir(td, io_u);
613
614	/*
615	 * fsync() or fdatasync() or trim etc, we are done
616	 */
617	if (!ddir_rw(io_u->ddir))
618		goto out;
619
620	/*
621	 * See if it's time to switch to a new zone
622	 */
623	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
624		td->zone_bytes = 0;
625		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
626		io_u->file->last_pos = io_u->file->file_offset;
627		td->io_skip_bytes += td->o.zone_skip;
628	}
629
630	/*
631	 * No log, let the seq/rand engine retrieve the next buflen and
632	 * position.
633	 */
634	if (get_next_offset(td, io_u)) {
635		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
636		return 1;
637	}
638
639	io_u->buflen = get_next_buflen(td, io_u);
640	if (!io_u->buflen) {
641		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
642		return 1;
643	}
644
645	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
646		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
647		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
648				io_u->buflen, io_u->file->real_file_size);
649		return 1;
650	}
651
652	/*
653	 * mark entry before potentially trimming io_u
654	 */
655	if (td_random(td) && file_randommap(td, io_u->file))
656		mark_random_map(td, io_u);
657
658	/*
659	 * If using a write iolog, store this entry.
660	 */
661out:
662	dprint_io_u(io_u, "fill_io_u");
663	td->zone_bytes += io_u->buflen;
664	log_io_u(td, io_u);
665	return 0;
666}
667
668static void __io_u_mark_map(unsigned int *map, unsigned int nr)
669{
670	int idx = 0;
671
672	switch (nr) {
673	default:
674		idx = 6;
675		break;
676	case 33 ... 64:
677		idx = 5;
678		break;
679	case 17 ... 32:
680		idx = 4;
681		break;
682	case 9 ... 16:
683		idx = 3;
684		break;
685	case 5 ... 8:
686		idx = 2;
687		break;
688	case 1 ... 4:
689		idx = 1;
690	case 0:
691		break;
692	}
693
694	map[idx]++;
695}
696
697void io_u_mark_submit(struct thread_data *td, unsigned int nr)
698{
699	__io_u_mark_map(td->ts.io_u_submit, nr);
700	td->ts.total_submit++;
701}
702
703void io_u_mark_complete(struct thread_data *td, unsigned int nr)
704{
705	__io_u_mark_map(td->ts.io_u_complete, nr);
706	td->ts.total_complete++;
707}
708
709void io_u_mark_depth(struct thread_data *td, unsigned int nr)
710{
711	int idx = 0;
712
713	switch (td->cur_depth) {
714	default:
715		idx = 6;
716		break;
717	case 32 ... 63:
718		idx = 5;
719		break;
720	case 16 ... 31:
721		idx = 4;
722		break;
723	case 8 ... 15:
724		idx = 3;
725		break;
726	case 4 ... 7:
727		idx = 2;
728		break;
729	case 2 ... 3:
730		idx = 1;
731	case 1:
732		break;
733	}
734
735	td->ts.io_u_map[idx] += nr;
736}
737
738static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
739{
740	int idx = 0;
741
742	assert(usec < 1000);
743
744	switch (usec) {
745	case 750 ... 999:
746		idx = 9;
747		break;
748	case 500 ... 749:
749		idx = 8;
750		break;
751	case 250 ... 499:
752		idx = 7;
753		break;
754	case 100 ... 249:
755		idx = 6;
756		break;
757	case 50 ... 99:
758		idx = 5;
759		break;
760	case 20 ... 49:
761		idx = 4;
762		break;
763	case 10 ... 19:
764		idx = 3;
765		break;
766	case 4 ... 9:
767		idx = 2;
768		break;
769	case 2 ... 3:
770		idx = 1;
771	case 0 ... 1:
772		break;
773	}
774
775	assert(idx < FIO_IO_U_LAT_U_NR);
776	td->ts.io_u_lat_u[idx]++;
777}
778
779static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
780{
781	int idx = 0;
782
783	switch (msec) {
784	default:
785		idx = 11;
786		break;
787	case 1000 ... 1999:
788		idx = 10;
789		break;
790	case 750 ... 999:
791		idx = 9;
792		break;
793	case 500 ... 749:
794		idx = 8;
795		break;
796	case 250 ... 499:
797		idx = 7;
798		break;
799	case 100 ... 249:
800		idx = 6;
801		break;
802	case 50 ... 99:
803		idx = 5;
804		break;
805	case 20 ... 49:
806		idx = 4;
807		break;
808	case 10 ... 19:
809		idx = 3;
810		break;
811	case 4 ... 9:
812		idx = 2;
813		break;
814	case 2 ... 3:
815		idx = 1;
816	case 0 ... 1:
817		break;
818	}
819
820	assert(idx < FIO_IO_U_LAT_M_NR);
821	td->ts.io_u_lat_m[idx]++;
822}
823
824static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
825{
826	if (usec < 1000)
827		io_u_mark_lat_usec(td, usec);
828	else
829		io_u_mark_lat_msec(td, usec / 1000);
830}
831
832/*
833 * Get next file to service by choosing one at random
834 */
835static struct fio_file *get_next_file_rand(struct thread_data *td,
836					   enum fio_file_flags goodf,
837					   enum fio_file_flags badf)
838{
839	struct fio_file *f;
840	int fno;
841
842	do {
843		int opened = 0;
844		unsigned long r;
845
846		if (td->o.use_os_rand) {
847			r = os_random_long(&td->next_file_state);
848			fno = (unsigned int) ((double) td->o.nr_files
849				* (r / (OS_RAND_MAX + 1.0)));
850		} else {
851			r = __rand(&td->__next_file_state);
852			fno = (unsigned int) ((double) td->o.nr_files
853				* (r / (FRAND_MAX + 1.0)));
854		}
855
856		f = td->files[fno];
857		if (fio_file_done(f))
858			continue;
859
860		if (!fio_file_open(f)) {
861			int err;
862
863			err = td_io_open_file(td, f);
864			if (err)
865				continue;
866			opened = 1;
867		}
868
869		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
870			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
871			return f;
872		}
873		if (opened)
874			td_io_close_file(td, f);
875	} while (1);
876}
877
878/*
879 * Get next file to service by doing round robin between all available ones
880 */
881static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
882					 int badf)
883{
884	unsigned int old_next_file = td->next_file;
885	struct fio_file *f;
886
887	do {
888		int opened = 0;
889
890		f = td->files[td->next_file];
891
892		td->next_file++;
893		if (td->next_file >= td->o.nr_files)
894			td->next_file = 0;
895
896		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
897		if (fio_file_done(f)) {
898			f = NULL;
899			continue;
900		}
901
902		if (!fio_file_open(f)) {
903			int err;
904
905			err = td_io_open_file(td, f);
906			if (err) {
907				dprint(FD_FILE, "error %d on open of %s\n",
908					err, f->file_name);
909				f = NULL;
910				continue;
911			}
912			opened = 1;
913		}
914
915		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
916								f->flags);
917		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
918			break;
919
920		if (opened)
921			td_io_close_file(td, f);
922
923		f = NULL;
924	} while (td->next_file != old_next_file);
925
926	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
927	return f;
928}
929
930static struct fio_file *__get_next_file(struct thread_data *td)
931{
932	struct fio_file *f;
933
934	assert(td->o.nr_files <= td->files_index);
935
936	if (td->nr_done_files >= td->o.nr_files) {
937		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
938				" nr_files=%d\n", td->nr_open_files,
939						  td->nr_done_files,
940						  td->o.nr_files);
941		return NULL;
942	}
943
944	f = td->file_service_file;
945	if (f && fio_file_open(f) && !fio_file_closing(f)) {
946		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
947			goto out;
948		if (td->file_service_left--)
949			goto out;
950	}
951
952	if (td->o.file_service_type == FIO_FSERVICE_RR ||
953	    td->o.file_service_type == FIO_FSERVICE_SEQ)
954		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
955	else
956		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
957
958	td->file_service_file = f;
959	td->file_service_left = td->file_service_nr - 1;
960out:
961	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
962	return f;
963}
964
965static struct fio_file *get_next_file(struct thread_data *td)
966{
967	if (!(td->flags & TD_F_PROFILE_OPS)) {
968		struct prof_io_ops *ops = &td->prof_io_ops;
969
970		if (ops->get_next_file)
971			return ops->get_next_file(td);
972	}
973
974	return __get_next_file(td);
975}
976
977static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
978{
979	struct fio_file *f;
980
981	do {
982		f = get_next_file(td);
983		if (!f)
984			return 1;
985
986		io_u->file = f;
987		get_file(f);
988
989		if (!fill_io_u(td, io_u))
990			break;
991
992		put_file_log(td, f);
993		td_io_close_file(td, f);
994		io_u->file = NULL;
995		fio_file_set_done(f);
996		td->nr_done_files++;
997		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
998					td->nr_done_files, td->o.nr_files);
999	} while (1);
1000
1001	return 0;
1002}
1003
1004
1005struct io_u *__get_io_u(struct thread_data *td)
1006{
1007	struct io_u *io_u = NULL;
1008
1009	td_io_u_lock(td);
1010
1011again:
1012	if (!flist_empty(&td->io_u_requeues))
1013		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1014	else if (!queue_full(td)) {
1015		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1016
1017		io_u->buflen = 0;
1018		io_u->resid = 0;
1019		io_u->file = NULL;
1020		io_u->end_io = NULL;
1021	}
1022
1023	if (io_u) {
1024		assert(io_u->flags & IO_U_F_FREE);
1025		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1026		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1027		io_u->flags &= ~IO_U_F_VER_LIST;
1028
1029		io_u->error = 0;
1030		flist_del(&io_u->list);
1031		flist_add_tail(&io_u->list, &td->io_u_busylist);
1032		td->cur_depth++;
1033		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1034	} else if (td->o.verify_async) {
1035		/*
1036		 * We ran out, wait for async verify threads to finish and
1037		 * return one
1038		 */
1039		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1040		goto again;
1041	}
1042
1043	td_io_u_unlock(td);
1044	return io_u;
1045}
1046
1047static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1048{
1049	if (!(td->flags & TD_F_TRIM_BACKLOG))
1050		return 0;
1051
1052	if (td->trim_entries) {
1053		int get_trim = 0;
1054
1055		if (td->trim_batch) {
1056			td->trim_batch--;
1057			get_trim = 1;
1058		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
1059			 td->last_ddir != DDIR_READ) {
1060			td->trim_batch = td->o.trim_batch;
1061			if (!td->trim_batch)
1062				td->trim_batch = td->o.trim_backlog;
1063			get_trim = 1;
1064		}
1065
1066		if (get_trim && !get_next_trim(td, io_u))
1067			return 1;
1068	}
1069
1070	return 0;
1071}
1072
1073static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1074{
1075	if (!(td->flags & TD_F_VER_BACKLOG))
1076		return 0;
1077
1078	if (td->io_hist_len) {
1079		int get_verify = 0;
1080
1081		if (td->verify_batch)
1082			get_verify = 1;
1083		else if (!(td->io_hist_len % td->o.verify_backlog) &&
1084			 td->last_ddir != DDIR_READ) {
1085			td->verify_batch = td->o.verify_batch;
1086			if (!td->verify_batch)
1087				td->verify_batch = td->o.verify_backlog;
1088			get_verify = 1;
1089		}
1090
1091		if (get_verify && !get_next_verify(td, io_u)) {
1092			td->verify_batch--;
1093			return 1;
1094		}
1095	}
1096
1097	return 0;
1098}
1099
1100/*
1101 * Fill offset and start time into the buffer content, to prevent too
1102 * easy compressible data for simple de-dupe attempts. Do this for every
1103 * 512b block in the range, since that should be the smallest block size
1104 * we can expect from a device.
1105 */
1106static void small_content_scramble(struct io_u *io_u)
1107{
1108	unsigned int i, nr_blocks = io_u->buflen / 512;
1109	unsigned long long boffset;
1110	unsigned int offset;
1111	void *p, *end;
1112
1113	if (!nr_blocks)
1114		return;
1115
1116	p = io_u->xfer_buf;
1117	boffset = io_u->offset;
1118	io_u->buf_filled_len = 0;
1119
1120	for (i = 0; i < nr_blocks; i++) {
1121		/*
1122		 * Fill the byte offset into a "random" start offset of
1123		 * the buffer, given by the product of the usec time
1124		 * and the actual offset.
1125		 */
1126		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1127		offset &= ~(sizeof(unsigned long long) - 1);
1128		if (offset >= 512 - sizeof(unsigned long long))
1129			offset -= sizeof(unsigned long long);
1130		memcpy(p + offset, &boffset, sizeof(boffset));
1131
1132		end = p + 512 - sizeof(io_u->start_time);
1133		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1134		p += 512;
1135		boffset += 512;
1136	}
1137}
1138
1139/*
1140 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1141 * etc. The returned io_u is fully ready to be prepped and submitted.
1142 */
1143struct io_u *get_io_u(struct thread_data *td)
1144{
1145	struct fio_file *f;
1146	struct io_u *io_u;
1147	int do_scramble = 0;
1148
1149	io_u = __get_io_u(td);
1150	if (!io_u) {
1151		dprint(FD_IO, "__get_io_u failed\n");
1152		return NULL;
1153	}
1154
1155	if (check_get_verify(td, io_u))
1156		goto out;
1157	if (check_get_trim(td, io_u))
1158		goto out;
1159
1160	/*
1161	 * from a requeue, io_u already setup
1162	 */
1163	if (io_u->file)
1164		goto out;
1165
1166	/*
1167	 * If using an iolog, grab next piece if any available.
1168	 */
1169	if (td->flags & TD_F_READ_IOLOG) {
1170		if (read_iolog_get(td, io_u))
1171			goto err_put;
1172	} else if (set_io_u_file(td, io_u)) {
1173		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1174		goto err_put;
1175	}
1176
1177	f = io_u->file;
1178	assert(fio_file_open(f));
1179
1180	if (ddir_rw(io_u->ddir)) {
1181		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1182			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1183			goto err_put;
1184		}
1185
1186		f->last_start = io_u->offset;
1187		f->last_pos = io_u->offset + io_u->buflen;
1188
1189		if (io_u->ddir == DDIR_WRITE) {
1190			if (td->flags & TD_F_REFILL_BUFFERS) {
1191				io_u_fill_buffer(td, io_u,
1192					io_u->xfer_buflen, io_u->xfer_buflen);
1193			} else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1194				do_scramble = 1;
1195			if (td->flags & TD_F_VER_NONE) {
1196				populate_verify_io_u(td, io_u);
1197				do_scramble = 0;
1198			}
1199		} else if (io_u->ddir == DDIR_READ) {
1200			/*
1201			 * Reset the buf_filled parameters so next time if the
1202			 * buffer is used for writes it is refilled.
1203			 */
1204			io_u->buf_filled_len = 0;
1205		}
1206	}
1207
1208	/*
1209	 * Set io data pointers.
1210	 */
1211	io_u->xfer_buf = io_u->buf;
1212	io_u->xfer_buflen = io_u->buflen;
1213
1214out:
1215	assert(io_u->file);
1216	if (!td_io_prep(td, io_u)) {
1217		if (!td->o.disable_slat)
1218			fio_gettime(&io_u->start_time, NULL);
1219		if (do_scramble)
1220			small_content_scramble(io_u);
1221		return io_u;
1222	}
1223err_put:
1224	dprint(FD_IO, "get_io_u failed\n");
1225	put_io_u(td, io_u);
1226	return NULL;
1227}
1228
1229void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1230{
1231	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1232	const char *msg[] = { "read", "write", "sync", "datasync",
1233				"sync_file_range", "wait", "trim" };
1234
1235	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1236		return;
1237
1238	log_err("fio: io_u error");
1239
1240	if (io_u->file)
1241		log_err(" on file %s", io_u->file->file_name);
1242
1243	log_err(": %s\n", strerror(io_u->error));
1244
1245	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1246					io_u->offset, io_u->xfer_buflen);
1247
1248	if (!td->error)
1249		td_verror(td, io_u->error, "io_u error");
1250}
1251
1252static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1253				  struct io_completion_data *icd,
1254				  const enum fio_ddir idx, unsigned int bytes)
1255{
1256	unsigned long lusec = 0;
1257
1258	if (!td->o.disable_clat || !td->o.disable_bw)
1259		lusec = utime_since(&io_u->issue_time, &icd->time);
1260
1261	if (!td->o.disable_lat) {
1262		unsigned long tusec;
1263
1264		tusec = utime_since(&io_u->start_time, &icd->time);
1265		add_lat_sample(td, idx, tusec, bytes);
1266
1267		if (td->o.max_latency && tusec > td->o.max_latency) {
1268			if (!td->error)
1269				log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1270			td_verror(td, ETIMEDOUT, "max latency exceeded");
1271			icd->error = ETIMEDOUT;
1272		}
1273	}
1274
1275	if (!td->o.disable_clat) {
1276		add_clat_sample(td, idx, lusec, bytes);
1277		io_u_mark_latency(td, lusec);
1278	}
1279
1280	if (!td->o.disable_bw)
1281		add_bw_sample(td, idx, bytes, &icd->time);
1282
1283	add_iops_sample(td, idx, &icd->time);
1284}
1285
1286static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1287{
1288	unsigned long long secs, remainder, bps, bytes;
1289	bytes = td->this_io_bytes[ddir];
1290	bps = td->rate_bps[ddir];
1291	secs = bytes / bps;
1292	remainder = bytes % bps;
1293	return remainder * 1000000 / bps + secs * 1000000;
1294}
1295
1296static void io_completed(struct thread_data *td, struct io_u *io_u,
1297			 struct io_completion_data *icd)
1298{
1299	struct fio_file *f;
1300
1301	dprint_io_u(io_u, "io complete");
1302
1303	td_io_u_lock(td);
1304	assert(io_u->flags & IO_U_F_FLIGHT);
1305	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1306	td_io_u_unlock(td);
1307
1308	if (ddir_sync(io_u->ddir)) {
1309		td->last_was_sync = 1;
1310		f = io_u->file;
1311		if (f) {
1312			f->first_write = -1ULL;
1313			f->last_write = -1ULL;
1314		}
1315		return;
1316	}
1317
1318	td->last_was_sync = 0;
1319	td->last_ddir = io_u->ddir;
1320
1321	if (!io_u->error && ddir_rw(io_u->ddir)) {
1322		unsigned int bytes = io_u->buflen - io_u->resid;
1323		const enum fio_ddir idx = io_u->ddir;
1324		const enum fio_ddir odx = io_u->ddir ^ 1;
1325		int ret;
1326
1327		td->io_blocks[idx]++;
1328		td->this_io_blocks[idx]++;
1329		td->io_bytes[idx] += bytes;
1330
1331		if (!(io_u->flags & IO_U_F_VER_LIST))
1332			td->this_io_bytes[idx] += bytes;
1333
1334		if (idx == DDIR_WRITE) {
1335			f = io_u->file;
1336			if (f) {
1337				if (f->first_write == -1ULL ||
1338				    io_u->offset < f->first_write)
1339					f->first_write = io_u->offset;
1340				if (f->last_write == -1ULL ||
1341				    ((io_u->offset + bytes) > f->last_write))
1342					f->last_write = io_u->offset + bytes;
1343			}
1344		}
1345
1346		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1347					   td->runstate == TD_VERIFYING)) {
1348			account_io_completion(td, io_u, icd, idx, bytes);
1349
1350			if (__should_check_rate(td, idx)) {
1351				td->rate_pending_usleep[idx] =
1352					(usec_for_io(td, idx) -
1353					 utime_since_now(&td->start));
1354			}
1355			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1356				td->rate_pending_usleep[odx] =
1357					(usec_for_io(td, odx) -
1358					 utime_since_now(&td->start));
1359		}
1360
1361		if (td_write(td) && idx == DDIR_WRITE &&
1362		    td->o.do_verify &&
1363		    td->o.verify != VERIFY_NONE)
1364			log_io_piece(td, io_u);
1365
1366		icd->bytes_done[idx] += bytes;
1367
1368		if (io_u->end_io) {
1369			ret = io_u->end_io(td, io_u);
1370			if (ret && !icd->error)
1371				icd->error = ret;
1372		}
1373	} else if (io_u->error) {
1374		icd->error = io_u->error;
1375		io_u_log_error(td, io_u);
1376	}
1377	if (icd->error) {
1378		enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1379		if (!td_non_fatal_error(td, eb, icd->error))
1380			return;
1381		/*
1382		 * If there is a non_fatal error, then add to the error count
1383		 * and clear all the errors.
1384		 */
1385		update_error_count(td, icd->error);
1386		td_clear_error(td);
1387		icd->error = 0;
1388		io_u->error = 0;
1389	}
1390}
1391
1392static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1393		     int nr)
1394{
1395	int ddir;
1396	if (!td->o.disable_clat || !td->o.disable_bw)
1397		fio_gettime(&icd->time, NULL);
1398
1399	icd->nr = nr;
1400
1401	icd->error = 0;
1402	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1403		icd->bytes_done[ddir] = 0;
1404}
1405
1406static void ios_completed(struct thread_data *td,
1407			  struct io_completion_data *icd)
1408{
1409	struct io_u *io_u;
1410	int i;
1411
1412	for (i = 0; i < icd->nr; i++) {
1413		io_u = td->io_ops->event(td, i);
1414
1415		io_completed(td, io_u, icd);
1416
1417		if (!(io_u->flags & IO_U_F_FREE_DEF))
1418			put_io_u(td, io_u);
1419	}
1420}
1421
1422/*
1423 * Complete a single io_u for the sync engines.
1424 */
1425int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1426		       unsigned long *bytes)
1427{
1428	struct io_completion_data icd;
1429
1430	init_icd(td, &icd, 1);
1431	io_completed(td, io_u, &icd);
1432
1433	if (!(io_u->flags & IO_U_F_FREE_DEF))
1434		put_io_u(td, io_u);
1435
1436	if (icd.error) {
1437		td_verror(td, icd.error, "io_u_sync_complete");
1438		return -1;
1439	}
1440
1441	if (bytes) {
1442		int ddir;
1443
1444		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1445			bytes[ddir] += icd.bytes_done[ddir];
1446	}
1447
1448	return 0;
1449}
1450
1451/*
1452 * Called to complete min_events number of io for the async engines.
1453 */
1454int io_u_queued_complete(struct thread_data *td, int min_evts,
1455			 unsigned long *bytes)
1456{
1457	struct io_completion_data icd;
1458	struct timespec *tvp = NULL;
1459	int ret;
1460	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1461
1462	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1463
1464	if (!min_evts)
1465		tvp = &ts;
1466
1467	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1468	if (ret < 0) {
1469		td_verror(td, -ret, "td_io_getevents");
1470		return ret;
1471	} else if (!ret)
1472		return ret;
1473
1474	init_icd(td, &icd, ret);
1475	ios_completed(td, &icd);
1476	if (icd.error) {
1477		td_verror(td, icd.error, "io_u_queued_complete");
1478		return -1;
1479	}
1480
1481	if (bytes) {
1482		int ddir;
1483
1484		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1485			bytes[ddir] += icd.bytes_done[ddir];
1486	}
1487
1488	return 0;
1489}
1490
1491/*
1492 * Call when io_u is really queued, to update the submission latency.
1493 */
1494void io_u_queued(struct thread_data *td, struct io_u *io_u)
1495{
1496	if (!td->o.disable_slat) {
1497		unsigned long slat_time;
1498
1499		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1500		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1501	}
1502}
1503
1504/*
1505 * "randomly" fill the buffer contents
1506 */
1507void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1508		      unsigned int min_write, unsigned int max_bs)
1509{
1510	io_u->buf_filled_len = 0;
1511
1512	if (!td->o.zero_buffers) {
1513		unsigned int perc = td->o.compress_percentage;
1514
1515		if (perc) {
1516			unsigned int seg = min_write;
1517
1518			seg = min(min_write, td->o.compress_chunk);
1519			fill_random_buf_percentage(&td->buf_state, io_u->buf,
1520						perc, seg, max_bs);
1521		} else
1522			fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1523	} else
1524		memset(io_u->buf, 0, max_bs);
1525}
1526