io_u.c revision 5a7c56804dafab5770797044a4f1d259fe708dfb
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 "os.h"
10
11/*
12 * Change this define to play with the timeout handling
13 */
14#undef FIO_USE_TIMEOUT
15
16struct io_completion_data {
17	int nr;				/* input */
18
19	int error;			/* output */
20	unsigned long bytes_done[2];	/* output */
21	struct timeval time;		/* output */
22};
23
24/*
25 * The ->file_map[] contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
28static int random_map_free(struct thread_data *td, struct fio_file *f,
29			   unsigned long long block)
30{
31	unsigned int idx = RAND_MAP_IDX(td, f, block);
32	unsigned int bit = RAND_MAP_BIT(td, f, block);
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;
46	unsigned int nr_blocks;
47
48	block = io_u->offset / (unsigned long long) min_bs;
49	blocks = 0;
50	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
51
52	while (blocks < nr_blocks) {
53		unsigned int idx, bit;
54
55		/*
56		 * If we have a mixed random workload, we may
57		 * encounter blocks we already did IO to.
58		 */
59		if (!td->o.ddir_nr && !random_map_free(td, f, block))
60			break;
61
62		idx = RAND_MAP_IDX(td, f, block);
63		bit = RAND_MAP_BIT(td, f, block);
64
65		fio_assert(td, idx < f->num_maps);
66
67		f->file_map[idx] |= (1UL << bit);
68		block++;
69		blocks++;
70	}
71
72	if ((blocks * min_bs) < io_u->buflen)
73		io_u->buflen = blocks * min_bs;
74}
75
76/*
77 * Return the next free block in the map.
78 */
79static int get_next_free_block(struct thread_data *td, struct fio_file *f,
80			       unsigned long long *b)
81{
82	int i;
83
84	i = f->last_free_lookup;
85	*b = (i * BLOCKS_PER_MAP);
86	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
87		if (f->file_map[i] != -1UL) {
88			*b += ffz(f->file_map[i]);
89			f->last_free_lookup = i;
90			return 0;
91		}
92
93		*b += BLOCKS_PER_MAP;
94		i++;
95	}
96
97	return 1;
98}
99
100static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
101				int ddir, unsigned long long *b)
102{
103	unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
104	unsigned long long r, rb;
105	int loops = 5;
106
107	do {
108		r = os_random_long(&td->random_state);
109		if (!max_blocks)
110			*b = 0;
111		else
112			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
113		if (td->o.norandommap)
114			break;
115		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
116		loops--;
117	} while (!random_map_free(td, f, rb) && loops);
118
119	/*
120	 * if we failed to retrieve a truly random offset within
121	 * the loops assigned, see if there are free ones left at all
122	 */
123	if (!loops && get_next_free_block(td, f, b))
124		return 1;
125
126	return 0;
127}
128
129/*
130 * For random io, generate a random new block and see if it's used. Repeat
131 * until we find a free one. For sequential io, just return the end of
132 * the last io issued.
133 */
134static int get_next_offset(struct thread_data *td, struct io_u *io_u)
135{
136	struct fio_file *f = io_u->file;
137	const int ddir = io_u->ddir;
138	unsigned long long b;
139
140	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
141		td->ddir_nr = td->o.ddir_nr;
142
143		if (get_next_rand_offset(td, f, ddir, &b))
144			return 1;
145	} else {
146		if (f->last_pos >= f->real_file_size)
147			return 1;
148
149		b = f->last_pos / td->o.min_bs[ddir];
150	}
151
152	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
153	if (io_u->offset >= f->real_file_size)
154		return 1;
155
156	return 0;
157}
158
159static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
160{
161	struct fio_file *f = io_u->file;
162	const int ddir = io_u->ddir;
163	unsigned int buflen;
164	long r;
165
166	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
167		buflen = td->o.min_bs[ddir];
168	else {
169		r = os_random_long(&td->bsrange_state);
170		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
171		if (!td->o.bs_unaligned)
172			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
173	}
174
175	return buflen;
176}
177
178static void set_rwmix_bytes(struct thread_data *td)
179{
180	unsigned long long rbytes;
181	unsigned int diff;
182
183	/*
184	 * we do time or byte based switch. this is needed because
185	 * buffered writes may issue a lot quicker than they complete,
186	 * whereas reads do not.
187	 */
188	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
189	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
190
191	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
192}
193
194static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
195{
196	unsigned int v;
197	long r;
198
199	r = os_random_long(&td->rwmix_state);
200	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
201	if (v < td->o.rwmix[DDIR_READ])
202		return DDIR_READ;
203
204	return DDIR_WRITE;
205}
206
207/*
208 * Return the data direction for the next io_u. If the job is a
209 * mixed read/write workload, check the rwmix cycle and switch if
210 * necessary.
211 */
212static enum fio_ddir get_rw_ddir(struct thread_data *td)
213{
214	if (td_rw(td)) {
215		struct timeval now;
216		unsigned long elapsed;
217		unsigned int cycle;
218
219		fio_gettime(&now, NULL);
220	 	elapsed = mtime_since_now(&td->rwmix_switch);
221
222		/*
223		 * if this is the first cycle, make it shorter
224		 */
225		cycle = td->o.rwmixcycle;
226		if (!td->rwmix_bytes)
227			cycle /= 10;
228
229		/*
230		 * Check if it's time to seed a new data direction.
231		 */
232		if (elapsed >= cycle ||
233		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
234			unsigned long long max_bytes;
235			enum fio_ddir ddir;
236
237			/*
238			 * Put a top limit on how many bytes we do for
239			 * one data direction, to avoid overflowing the
240			 * ranges too much
241			 */
242			ddir = get_rand_ddir(td);
243			max_bytes = td->this_io_bytes[ddir];
244			if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
245				if (!td->rw_end_set[ddir]) {
246					td->rw_end_set[ddir] = 1;
247					memcpy(&td->rw_end[ddir], &now, sizeof(now));
248				}
249				ddir ^= 1;
250			}
251
252			if (ddir != td->rwmix_ddir)
253				set_rwmix_bytes(td);
254
255			td->rwmix_ddir = ddir;
256			memcpy(&td->rwmix_switch, &now, sizeof(now));
257		}
258		return td->rwmix_ddir;
259	} else if (td_read(td))
260		return DDIR_READ;
261	else
262		return DDIR_WRITE;
263}
264
265void put_io_u(struct thread_data *td, struct io_u *io_u)
266{
267	assert((io_u->flags & IO_U_F_FREE) == 0);
268	io_u->flags |= IO_U_F_FREE;
269
270	io_u->file = NULL;
271	list_del(&io_u->list);
272	list_add(&io_u->list, &td->io_u_freelist);
273	td->cur_depth--;
274}
275
276void requeue_io_u(struct thread_data *td, struct io_u **io_u)
277{
278	struct io_u *__io_u = *io_u;
279
280	__io_u->flags |= IO_U_F_FREE;
281	__io_u->flags &= ~IO_U_F_FLIGHT;
282
283	list_del(&__io_u->list);
284	list_add_tail(&__io_u->list, &td->io_u_requeues);
285	td->cur_depth--;
286	*io_u = NULL;
287}
288
289static int fill_io_u(struct thread_data *td, struct io_u *io_u)
290{
291	/*
292	 * If using an iolog, grab next piece if any available.
293	 */
294	if (td->o.read_iolog)
295		return read_iolog_get(td, io_u);
296
297	/*
298	 * see if it's time to sync
299	 */
300	if (td->o.fsync_blocks &&
301	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
302	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
303		io_u->ddir = DDIR_SYNC;
304		goto out;
305	}
306
307	io_u->ddir = get_rw_ddir(td);
308
309	/*
310	 * No log, let the seq/rand engine retrieve the next buflen and
311	 * position.
312	 */
313	if (get_next_offset(td, io_u))
314		return 1;
315
316	io_u->buflen = get_next_buflen(td, io_u);
317	if (!io_u->buflen)
318		return 1;
319
320	/*
321	 * mark entry before potentially trimming io_u
322	 */
323	if (td_random(td) && !td->o.norandommap)
324		mark_random_map(td, io_u);
325
326	/*
327	 * If using a write iolog, store this entry.
328	 */
329out:
330	if (td->o.write_iolog_file)
331		write_iolog_put(td, io_u);
332
333	return 0;
334}
335
336void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
337{
338	int index = 0;
339
340	if (io_u->ddir == DDIR_SYNC)
341		return;
342
343	switch (td->cur_depth) {
344	default:
345		index++;
346	case 32 ... 63:
347		index++;
348	case 16 ... 31:
349		index++;
350	case 8 ... 15:
351		index++;
352	case 4 ... 7:
353		index++;
354	case 2 ... 3:
355		index++;
356	case 1:
357		break;
358	}
359
360	td->ts.io_u_map[index]++;
361	td->ts.total_io_u[io_u->ddir]++;
362}
363
364static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
365{
366	int index = 0;
367
368	switch (msec) {
369	default:
370		index++;
371	case 1000 ... 1999:
372		index++;
373	case 750 ... 999:
374		index++;
375	case 500 ... 749:
376		index++;
377	case 250 ... 499:
378		index++;
379	case 100 ... 249:
380		index++;
381	case 50 ... 99:
382		index++;
383	case 20 ... 49:
384		index++;
385	case 10 ... 19:
386		index++;
387	case 4 ... 9:
388		index++;
389	case 2 ... 3:
390		index++;
391	case 0 ... 1:
392		break;
393	}
394
395	td->ts.io_u_lat[index]++;
396}
397
398/*
399 * Get next file to service by choosing one at random
400 */
401static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
402					   int badf)
403{
404	struct fio_file *f;
405	int fno;
406
407	do {
408		long r = os_random_long(&td->next_file_state);
409
410		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
411		f = &td->files[fno];
412		if (f->flags & FIO_FILE_DONE)
413			continue;
414
415		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
416			return f;
417	} while (1);
418}
419
420/*
421 * Get next file to service by doing round robin between all available ones
422 */
423static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
424					 int badf)
425{
426	unsigned int old_next_file = td->next_file;
427	struct fio_file *f;
428
429	do {
430		f = &td->files[td->next_file];
431
432		td->next_file++;
433		if (td->next_file >= td->o.nr_files)
434			td->next_file = 0;
435
436		if (f->flags & FIO_FILE_DONE) {
437			f = NULL;
438			continue;
439		}
440
441		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
442			break;
443
444		f = NULL;
445	} while (td->next_file != old_next_file);
446
447	return f;
448}
449
450static struct fio_file *get_next_file(struct thread_data *td)
451{
452	struct fio_file *f;
453
454	assert(td->o.nr_files <= td->files_index);
455
456	if (!td->nr_open_files)
457		return NULL;
458
459	f = td->file_service_file;
460	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
461		return f;
462
463	if (td->o.file_service_type == FIO_FSERVICE_RR)
464		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
465	else
466		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
467
468	td->file_service_file = f;
469	td->file_service_left = td->file_service_nr - 1;
470	return f;
471}
472
473static struct fio_file *find_next_new_file(struct thread_data *td)
474{
475	struct fio_file *f;
476
477	if (td->o.file_service_type == FIO_FSERVICE_RR)
478		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
479	else
480		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
481
482	return f;
483}
484
485struct io_u *__get_io_u(struct thread_data *td)
486{
487	struct io_u *io_u = NULL;
488
489	if (!list_empty(&td->io_u_requeues))
490		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
491	else if (!queue_full(td)) {
492		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
493
494		io_u->buflen = 0;
495		io_u->resid = 0;
496		io_u->file = NULL;
497		io_u->end_io = NULL;
498	}
499
500	if (io_u) {
501		assert(io_u->flags & IO_U_F_FREE);
502		io_u->flags &= ~IO_U_F_FREE;
503
504		io_u->error = 0;
505		list_del(&io_u->list);
506		list_add(&io_u->list, &td->io_u_busylist);
507		td->cur_depth++;
508	}
509
510	return io_u;
511}
512
513/*
514 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
515 * etc. The returned io_u is fully ready to be prepped and submitted.
516 */
517struct io_u *get_io_u(struct thread_data *td)
518{
519	struct fio_file *f;
520	struct io_u *io_u;
521	int ret;
522
523	io_u = __get_io_u(td);
524	if (!io_u)
525		return NULL;
526
527	/*
528	 * from a requeue, io_u already setup
529	 */
530	if (io_u->file)
531		goto out;
532
533	do {
534		f = get_next_file(td);
535		if (!f) {
536			put_io_u(td, io_u);
537			return NULL;
538		}
539
540set_file:
541		io_u->file = f;
542
543		if (!fill_io_u(td, io_u))
544			break;
545
546		/*
547		 * No more to do for this file, close it
548		 */
549		io_u->file = NULL;
550		td_io_close_file(td, f);
551		f->flags |= FIO_FILE_DONE;
552
553		/*
554		 * probably not the right place to do this, but see
555		 * if we need to open a new file
556		 */
557		if (td->nr_open_files < td->o.open_files &&
558		    td->o.open_files != td->o.nr_files) {
559			f = find_next_new_file(td);
560
561			if (!f || (ret = td_io_open_file(td, f))) {
562				put_io_u(td, io_u);
563				return NULL;
564			}
565			goto set_file;
566		}
567	} while (1);
568
569	if (td->zone_bytes >= td->o.zone_size) {
570		td->zone_bytes = 0;
571		f->last_pos += td->o.zone_skip;
572	}
573
574	if (io_u->ddir != DDIR_SYNC) {
575		if (!io_u->buflen) {
576			put_io_u(td, io_u);
577			return NULL;
578		}
579
580		f->last_pos = io_u->offset + io_u->buflen;
581
582		if (td->o.verify != VERIFY_NONE)
583			populate_verify_io_u(td, io_u);
584	}
585
586	/*
587	 * Set io data pointers.
588	 */
589out:
590	io_u->xfer_buf = io_u->buf;
591	io_u->xfer_buflen = io_u->buflen;
592
593	if (td_io_prep(td, io_u)) {
594		put_io_u(td, io_u);
595		return NULL;
596	}
597
598	fio_gettime(&io_u->start_time, NULL);
599	return io_u;
600}
601
602void io_u_log_error(struct thread_data *td, struct io_u *io_u)
603{
604	const char *msg[] = { "read", "write", "sync" };
605
606	log_err("fio: io_u error");
607
608	if (io_u->file)
609		log_err(" on file %s", io_u->file->file_name);
610
611	log_err(": %s\n", strerror(io_u->error));
612
613	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
614
615	if (!td->error)
616		td_verror(td, io_u->error, "io_u error");
617}
618
619static void io_completed(struct thread_data *td, struct io_u *io_u,
620			 struct io_completion_data *icd)
621{
622	unsigned long msec;
623
624	assert(io_u->flags & IO_U_F_FLIGHT);
625	io_u->flags &= ~IO_U_F_FLIGHT;
626
627	put_file(td, io_u->file);
628
629	if (io_u->ddir == DDIR_SYNC) {
630		td->last_was_sync = 1;
631		return;
632	}
633
634	td->last_was_sync = 0;
635
636	if (!io_u->error) {
637		unsigned int bytes = io_u->buflen - io_u->resid;
638		const enum fio_ddir idx = io_u->ddir;
639		int ret;
640
641		td->io_blocks[idx]++;
642		td->io_bytes[idx] += bytes;
643		td->zone_bytes += bytes;
644		td->this_io_bytes[idx] += bytes;
645
646		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
647
648		msec = mtime_since(&io_u->issue_time, &icd->time);
649
650		add_clat_sample(td, idx, msec);
651		add_bw_sample(td, idx, &icd->time);
652		io_u_mark_latency(td, msec);
653
654		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
655		    td->o.verify != VERIFY_NONE)
656			log_io_piece(td, io_u);
657
658		icd->bytes_done[idx] += bytes;
659
660		if (io_u->end_io) {
661			ret = io_u->end_io(td, io_u);
662			if (ret && !icd->error)
663				icd->error = ret;
664		}
665	} else {
666		icd->error = io_u->error;
667		io_u_log_error(td, io_u);
668	}
669}
670
671static void init_icd(struct io_completion_data *icd, int nr)
672{
673	fio_gettime(&icd->time, NULL);
674
675	icd->nr = nr;
676
677	icd->error = 0;
678	icd->bytes_done[0] = icd->bytes_done[1] = 0;
679}
680
681static void ios_completed(struct thread_data *td,
682			  struct io_completion_data *icd)
683{
684	struct io_u *io_u;
685	int i;
686
687	for (i = 0; i < icd->nr; i++) {
688		io_u = td->io_ops->event(td, i);
689
690		io_completed(td, io_u, icd);
691		put_io_u(td, io_u);
692	}
693}
694
695/*
696 * Complete a single io_u for the sync engines.
697 */
698long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
699{
700	struct io_completion_data icd;
701
702	init_icd(&icd, 1);
703	io_completed(td, io_u, &icd);
704	put_io_u(td, io_u);
705
706	if (!icd.error)
707		return icd.bytes_done[0] + icd.bytes_done[1];
708
709	td_verror(td, icd.error, "io_u_sync_complete");
710	return -1;
711}
712
713/*
714 * Called to complete min_events number of io for the async engines.
715 */
716long io_u_queued_complete(struct thread_data *td, int min_events)
717{
718	struct io_completion_data icd;
719	struct timespec *tvp = NULL;
720	int ret;
721	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
722
723	if (!min_events)
724		tvp = &ts;
725
726	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
727	if (ret < 0) {
728		td_verror(td, -ret, "td_io_getevents");
729		return ret;
730	} else if (!ret)
731		return ret;
732
733	init_icd(&icd, ret);
734	ios_completed(td, &icd);
735	if (!icd.error)
736		return icd.bytes_done[0] + icd.bytes_done[1];
737
738	td_verror(td, icd.error, "io_u_queued_complete");
739	return -1;
740}
741
742/*
743 * Call when io_u is really queued, to update the submission latency.
744 */
745void io_u_queued(struct thread_data *td, struct io_u *io_u)
746{
747	unsigned long slat_time;
748
749	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
750	add_slat_sample(td, io_u->ddir, slat_time);
751}
752
753#ifdef FIO_USE_TIMEOUT
754void io_u_set_timeout(struct thread_data *td)
755{
756	assert(td->cur_depth);
757
758	td->timer.it_interval.tv_sec = 0;
759	td->timer.it_interval.tv_usec = 0;
760	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
761	td->timer.it_value.tv_usec = 0;
762	setitimer(ITIMER_REAL, &td->timer, NULL);
763	fio_gettime(&td->timeout_end, NULL);
764}
765
766static void io_u_dump(struct io_u *io_u)
767{
768	unsigned long t_start = mtime_since_now(&io_u->start_time);
769	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
770
771	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
772	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
773	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
774}
775#else
776void io_u_set_timeout(struct thread_data fio_unused *td)
777{
778}
779#endif
780
781#ifdef FIO_USE_TIMEOUT
782static void io_u_timeout_handler(int fio_unused sig)
783{
784	struct thread_data *td, *__td;
785	pid_t pid = getpid();
786	struct list_head *entry;
787	struct io_u *io_u;
788	int i;
789
790	log_err("fio: io_u timeout\n");
791
792	/*
793	 * TLS would be nice...
794	 */
795	td = NULL;
796	for_each_td(__td, i) {
797		if (__td->pid == pid) {
798			td = __td;
799			break;
800		}
801	}
802
803	if (!td) {
804		log_err("fio: io_u timeout, can't find job\n");
805		exit(1);
806	}
807
808	if (!td->cur_depth) {
809		log_err("fio: timeout without pending work?\n");
810		return;
811	}
812
813	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
814
815	list_for_each(entry, &td->io_u_busylist) {
816		io_u = list_entry(entry, struct io_u, list);
817
818		io_u_dump(io_u);
819	}
820
821	td_verror(td, ETIMEDOUT, "io_u timeout");
822	exit(1);
823}
824#endif
825
826void io_u_init_timeout(void)
827{
828#ifdef FIO_USE_TIMEOUT
829	signal(SIGALRM, io_u_timeout_handler);
830#endif
831}
832