io_u.c revision 317b95d07d4921d2594a1be6e014c9c2d062fe75
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
10/*
11 * Change this define to play with the timeout handling
12 */
13#undef FIO_USE_TIMEOUT
14
15struct io_completion_data {
16	int nr;				/* 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 thread_data *td, struct fio_file *f,
28			   unsigned long long block)
29{
30	unsigned int idx = RAND_MAP_IDX(td, f, block);
31	unsigned int bit = RAND_MAP_BIT(td, f, block);
32
33	return (f->file_map[idx] & (1UL << bit)) == 0;
34}
35
36/*
37 * Mark a given offset as used in the map.
38 */
39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40{
41	unsigned int min_bs = td->o.rw_min_bs;
42	struct fio_file *f = io_u->file;
43	unsigned long long block;
44	unsigned int blocks;
45	unsigned int nr_blocks;
46
47	block = io_u->offset / (unsigned long long) min_bs;
48	blocks = 0;
49	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50
51	while (blocks < nr_blocks) {
52		unsigned int idx, bit;
53
54		/*
55		 * If we have a mixed random workload, we may
56		 * encounter blocks we already did IO to.
57		 */
58		if (!td->o.ddir_nr && !random_map_free(td, f, block))
59			break;
60
61		idx = RAND_MAP_IDX(td, f, block);
62		bit = RAND_MAP_BIT(td, f, block);
63
64		fio_assert(td, idx < f->num_maps);
65
66		f->file_map[idx] |= (1UL << bit);
67		block++;
68		blocks++;
69	}
70
71	if ((blocks * min_bs) < io_u->buflen)
72		io_u->buflen = blocks * min_bs;
73}
74
75/*
76 * Return the next free block in the map.
77 */
78static int get_next_free_block(struct thread_data *td, struct fio_file *f,
79			       unsigned long long *b)
80{
81	int i;
82
83	i = f->last_free_lookup;
84	*b = (i * BLOCKS_PER_MAP);
85	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
86		if (f->file_map[i] != -1UL) {
87			*b += ffz(f->file_map[i]);
88			f->last_free_lookup = i;
89			return 0;
90		}
91
92		*b += BLOCKS_PER_MAP;
93		i++;
94	}
95
96	return 1;
97}
98
99static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
100				int ddir, unsigned long long *b)
101{
102	unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
103	unsigned long long r, rb;
104	int loops = 5;
105
106	do {
107		r = os_random_long(&td->random_state);
108		if (!max_blocks)
109			*b = 0;
110		else
111			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
112		if (td->o.norandommap)
113			break;
114		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
115		loops--;
116	} while (!random_map_free(td, f, rb) && loops);
117
118	/*
119	 * if we failed to retrieve a truly random offset within
120	 * the loops assigned, see if there are free ones left at all
121	 */
122	if (!loops && get_next_free_block(td, f, b))
123		return 1;
124
125	return 0;
126}
127
128/*
129 * For random io, generate a random new block and see if it's used. Repeat
130 * until we find a free one. For sequential io, just return the end of
131 * the last io issued.
132 */
133static int get_next_offset(struct thread_data *td, struct io_u *io_u)
134{
135	struct fio_file *f = io_u->file;
136	const int ddir = io_u->ddir;
137	unsigned long long b;
138
139	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
140		td->ddir_nr = td->o.ddir_nr;
141
142		if (get_next_rand_offset(td, f, ddir, &b))
143			return 1;
144	} else {
145		if (f->last_pos >= f->real_file_size)
146			return 1;
147
148		b = f->last_pos / td->o.min_bs[ddir];
149	}
150
151	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
152	if (io_u->offset >= f->real_file_size)
153		return 1;
154
155	return 0;
156}
157
158static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
159{
160	const int ddir = io_u->ddir;
161	unsigned int buflen;
162	long r;
163
164	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
165		buflen = td->o.min_bs[ddir];
166	else {
167		r = os_random_long(&td->bsrange_state);
168		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
169		if (!td->o.bs_unaligned)
170			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
171	}
172
173	return buflen;
174}
175
176static void set_rwmix_bytes(struct thread_data *td)
177{
178	unsigned long long rbytes;
179	unsigned int diff;
180
181	/*
182	 * we do time or byte based switch. this is needed because
183	 * buffered writes may issue a lot quicker than they complete,
184	 * whereas reads do not.
185	 */
186	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
187	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
188
189	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
190}
191
192static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
193{
194	unsigned int v;
195	long r;
196
197	r = os_random_long(&td->rwmix_state);
198	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
199	if (v < td->o.rwmix[DDIR_READ])
200		return DDIR_READ;
201
202	return DDIR_WRITE;
203}
204
205/*
206 * Return the data direction for the next io_u. If the job is a
207 * mixed read/write workload, check the rwmix cycle and switch if
208 * necessary.
209 */
210static enum fio_ddir get_rw_ddir(struct thread_data *td)
211{
212	if (td_rw(td)) {
213		struct timeval now;
214		unsigned long elapsed;
215		unsigned int cycle;
216
217		fio_gettime(&now, NULL);
218	 	elapsed = mtime_since_now(&td->rwmix_switch);
219
220		/*
221		 * if this is the first cycle, make it shorter
222		 */
223		cycle = td->o.rwmixcycle;
224		if (!td->rwmix_bytes)
225			cycle /= 10;
226
227		/*
228		 * Check if it's time to seed a new data direction.
229		 */
230		if (elapsed >= cycle ||
231		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
232			unsigned long long max_bytes;
233			enum fio_ddir ddir;
234
235			/*
236			 * Put a top limit on how many bytes we do for
237			 * one data direction, to avoid overflowing the
238			 * ranges too much
239			 */
240			ddir = get_rand_ddir(td);
241			max_bytes = td->this_io_bytes[ddir];
242			if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
243				if (!td->rw_end_set[ddir]) {
244					td->rw_end_set[ddir] = 1;
245					memcpy(&td->rw_end[ddir], &now, sizeof(now));
246				}
247				ddir ^= 1;
248			}
249
250			if (ddir != td->rwmix_ddir)
251				set_rwmix_bytes(td);
252
253			td->rwmix_ddir = ddir;
254			memcpy(&td->rwmix_switch, &now, sizeof(now));
255		}
256		return td->rwmix_ddir;
257	} else if (td_read(td))
258		return DDIR_READ;
259	else
260		return DDIR_WRITE;
261}
262
263void put_io_u(struct thread_data *td, struct io_u *io_u)
264{
265	assert((io_u->flags & IO_U_F_FREE) == 0);
266	io_u->flags |= IO_U_F_FREE;
267
268	io_u->file = NULL;
269	list_del(&io_u->list);
270	list_add(&io_u->list, &td->io_u_freelist);
271	td->cur_depth--;
272}
273
274void requeue_io_u(struct thread_data *td, struct io_u **io_u)
275{
276	struct io_u *__io_u = *io_u;
277
278	__io_u->flags |= IO_U_F_FREE;
279	__io_u->flags &= ~IO_U_F_FLIGHT;
280
281	list_del(&__io_u->list);
282	list_add_tail(&__io_u->list, &td->io_u_requeues);
283	td->cur_depth--;
284	*io_u = NULL;
285}
286
287static int fill_io_u(struct thread_data *td, struct io_u *io_u)
288{
289	/*
290	 * If using an iolog, grab next piece if any available.
291	 */
292	if (td->o.read_iolog)
293		return read_iolog_get(td, io_u);
294
295	/*
296	 * see if it's time to sync
297	 */
298	if (td->o.fsync_blocks &&
299	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
300	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
301		io_u->ddir = DDIR_SYNC;
302		goto out;
303	}
304
305	io_u->ddir = get_rw_ddir(td);
306
307	/*
308	 * No log, let the seq/rand engine retrieve the next buflen and
309	 * position.
310	 */
311	if (get_next_offset(td, io_u))
312		return 1;
313
314	io_u->buflen = get_next_buflen(td, io_u);
315	if (!io_u->buflen)
316		return 1;
317
318	/*
319	 * mark entry before potentially trimming io_u
320	 */
321	if (td_random(td) && !td->o.norandommap)
322		mark_random_map(td, io_u);
323
324	/*
325	 * If using a write iolog, store this entry.
326	 */
327out:
328	if (td->o.write_iolog_file)
329		write_iolog_put(td, io_u);
330
331	return 0;
332}
333
334void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
335{
336	int index = 0;
337
338	if (io_u->ddir == DDIR_SYNC)
339		return;
340
341	switch (td->cur_depth) {
342	default:
343		index++;
344	case 32 ... 63:
345		index++;
346	case 16 ... 31:
347		index++;
348	case 8 ... 15:
349		index++;
350	case 4 ... 7:
351		index++;
352	case 2 ... 3:
353		index++;
354	case 1:
355		break;
356	}
357
358	td->ts.io_u_map[index]++;
359	td->ts.total_io_u[io_u->ddir]++;
360}
361
362static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
363{
364	int index = 0;
365
366	switch (msec) {
367	default:
368		index++;
369	case 1000 ... 1999:
370		index++;
371	case 750 ... 999:
372		index++;
373	case 500 ... 749:
374		index++;
375	case 250 ... 499:
376		index++;
377	case 100 ... 249:
378		index++;
379	case 50 ... 99:
380		index++;
381	case 20 ... 49:
382		index++;
383	case 10 ... 19:
384		index++;
385	case 4 ... 9:
386		index++;
387	case 2 ... 3:
388		index++;
389	case 0 ... 1:
390		break;
391	}
392
393	td->ts.io_u_lat[index]++;
394}
395
396/*
397 * Get next file to service by choosing one at random
398 */
399static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
400					   int badf)
401{
402	struct fio_file *f;
403	int fno;
404
405	do {
406		long r = os_random_long(&td->next_file_state);
407
408		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
409		f = &td->files[fno];
410		if (f->flags & FIO_FILE_DONE)
411			continue;
412
413		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
414			return f;
415	} while (1);
416}
417
418/*
419 * Get next file to service by doing round robin between all available ones
420 */
421static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
422					 int badf)
423{
424	unsigned int old_next_file = td->next_file;
425	struct fio_file *f;
426
427	do {
428		f = &td->files[td->next_file];
429
430		td->next_file++;
431		if (td->next_file >= td->o.nr_files)
432			td->next_file = 0;
433
434		if (f->flags & FIO_FILE_DONE) {
435			f = NULL;
436			continue;
437		}
438
439		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
440			break;
441
442		f = NULL;
443	} while (td->next_file != old_next_file);
444
445	return f;
446}
447
448static struct fio_file *get_next_file(struct thread_data *td)
449{
450	struct fio_file *f;
451
452	assert(td->o.nr_files <= td->files_index);
453
454	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
455		return NULL;
456
457	f = td->file_service_file;
458	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
459		return f;
460
461	if (td->o.file_service_type == FIO_FSERVICE_RR)
462		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
463	else
464		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
465
466	td->file_service_file = f;
467	td->file_service_left = td->file_service_nr - 1;
468	return f;
469}
470
471static struct fio_file *find_next_new_file(struct thread_data *td)
472{
473	struct fio_file *f;
474
475	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
476		return NULL;
477
478	if (td->o.file_service_type == FIO_FSERVICE_RR)
479		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
480	else
481		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
482
483	return f;
484}
485
486struct io_u *__get_io_u(struct thread_data *td)
487{
488	struct io_u *io_u = NULL;
489
490	if (!list_empty(&td->io_u_requeues))
491		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
492	else if (!queue_full(td)) {
493		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
494
495		io_u->buflen = 0;
496		io_u->resid = 0;
497		io_u->file = NULL;
498		io_u->end_io = NULL;
499	}
500
501	if (io_u) {
502		assert(io_u->flags & IO_U_F_FREE);
503		io_u->flags &= ~IO_U_F_FREE;
504
505		io_u->error = 0;
506		list_del(&io_u->list);
507		list_add(&io_u->list, &td->io_u_busylist);
508		td->cur_depth++;
509	}
510
511	return io_u;
512}
513
514/*
515 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
516 * etc. The returned io_u is fully ready to be prepped and submitted.
517 */
518struct io_u *get_io_u(struct thread_data *td)
519{
520	struct fio_file *f;
521	struct io_u *io_u;
522	int ret;
523
524	io_u = __get_io_u(td);
525	if (!io_u)
526		return NULL;
527
528	/*
529	 * from a requeue, io_u already setup
530	 */
531	if (io_u->file)
532		goto out;
533
534	do {
535		f = get_next_file(td);
536		if (!f) {
537			put_io_u(td, io_u);
538			return NULL;
539		}
540
541set_file:
542		io_u->file = f;
543
544		if (!fill_io_u(td, io_u))
545			break;
546
547		/*
548		 * No more to do for this file, close it
549		 */
550		io_u->file = NULL;
551		td_io_close_file(td, f);
552		f->flags |= FIO_FILE_DONE;
553		td->nr_done_files++;
554
555		/*
556		 * probably not the right place to do this, but see
557		 * if we need to open a new file
558		 */
559		if (td->nr_open_files < td->o.open_files &&
560		    td->o.open_files != td->o.nr_files) {
561			f = find_next_new_file(td);
562
563			if (!f || (ret = td_io_open_file(td, f))) {
564				put_io_u(td, io_u);
565				return NULL;
566			}
567			goto set_file;
568		}
569	} while (1);
570
571	if (td->zone_bytes >= td->o.zone_size) {
572		td->zone_bytes = 0;
573		f->last_pos += td->o.zone_skip;
574	}
575
576	if (io_u->ddir != DDIR_SYNC) {
577		if (!io_u->buflen) {
578			put_io_u(td, io_u);
579			return NULL;
580		}
581
582		f->last_pos = io_u->offset + io_u->buflen;
583
584		if (td->o.verify != VERIFY_NONE)
585			populate_verify_io_u(td, io_u);
586	}
587
588	/*
589	 * Set io data pointers.
590	 */
591out:
592	io_u->xfer_buf = io_u->buf;
593	io_u->xfer_buflen = io_u->buflen;
594
595	if (td_io_prep(td, io_u)) {
596		put_io_u(td, io_u);
597		return NULL;
598	}
599
600	fio_gettime(&io_u->start_time, NULL);
601	return io_u;
602}
603
604void io_u_log_error(struct thread_data *td, struct io_u *io_u)
605{
606	const char *msg[] = { "read", "write", "sync" };
607
608	log_err("fio: io_u error");
609
610	if (io_u->file)
611		log_err(" on file %s", io_u->file->file_name);
612
613	log_err(": %s\n", strerror(io_u->error));
614
615	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
616
617	if (!td->error)
618		td_verror(td, io_u->error, "io_u error");
619}
620
621static void io_completed(struct thread_data *td, struct io_u *io_u,
622			 struct io_completion_data *icd)
623{
624	unsigned long msec;
625
626	assert(io_u->flags & IO_U_F_FLIGHT);
627	io_u->flags &= ~IO_U_F_FLIGHT;
628
629	put_file(td, io_u->file);
630
631	if (io_u->ddir == DDIR_SYNC) {
632		td->last_was_sync = 1;
633		return;
634	}
635
636	td->last_was_sync = 0;
637
638	if (!io_u->error) {
639		unsigned int bytes = io_u->buflen - io_u->resid;
640		const enum fio_ddir idx = io_u->ddir;
641		int ret;
642
643		td->io_blocks[idx]++;
644		td->io_bytes[idx] += bytes;
645		td->zone_bytes += bytes;
646		td->this_io_bytes[idx] += bytes;
647
648		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
649
650		msec = mtime_since(&io_u->issue_time, &icd->time);
651
652		add_clat_sample(td, idx, msec);
653		add_bw_sample(td, idx, &icd->time);
654		io_u_mark_latency(td, msec);
655
656		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
657		    td->o.verify != VERIFY_NONE)
658			log_io_piece(td, io_u);
659
660		icd->bytes_done[idx] += bytes;
661
662		if (io_u->end_io) {
663			ret = io_u->end_io(td, io_u);
664			if (ret && !icd->error)
665				icd->error = ret;
666		}
667	} else {
668		icd->error = io_u->error;
669		io_u_log_error(td, io_u);
670	}
671}
672
673static void init_icd(struct io_completion_data *icd, int nr)
674{
675	fio_gettime(&icd->time, NULL);
676
677	icd->nr = nr;
678
679	icd->error = 0;
680	icd->bytes_done[0] = icd->bytes_done[1] = 0;
681}
682
683static void ios_completed(struct thread_data *td,
684			  struct io_completion_data *icd)
685{
686	struct io_u *io_u;
687	int i;
688
689	for (i = 0; i < icd->nr; i++) {
690		io_u = td->io_ops->event(td, i);
691
692		io_completed(td, io_u, icd);
693		put_io_u(td, io_u);
694	}
695}
696
697/*
698 * Complete a single io_u for the sync engines.
699 */
700long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
701{
702	struct io_completion_data icd;
703
704	init_icd(&icd, 1);
705	io_completed(td, io_u, &icd);
706	put_io_u(td, io_u);
707
708	if (!icd.error)
709		return icd.bytes_done[0] + icd.bytes_done[1];
710
711	td_verror(td, icd.error, "io_u_sync_complete");
712	return -1;
713}
714
715/*
716 * Called to complete min_events number of io for the async engines.
717 */
718long io_u_queued_complete(struct thread_data *td, int min_events)
719{
720	struct io_completion_data icd;
721	struct timespec *tvp = NULL;
722	int ret;
723	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
724
725	if (!min_events)
726		tvp = &ts;
727
728	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
729	if (ret < 0) {
730		td_verror(td, -ret, "td_io_getevents");
731		return ret;
732	} else if (!ret)
733		return ret;
734
735	init_icd(&icd, ret);
736	ios_completed(td, &icd);
737	if (!icd.error)
738		return icd.bytes_done[0] + icd.bytes_done[1];
739
740	td_verror(td, icd.error, "io_u_queued_complete");
741	return -1;
742}
743
744/*
745 * Call when io_u is really queued, to update the submission latency.
746 */
747void io_u_queued(struct thread_data *td, struct io_u *io_u)
748{
749	unsigned long slat_time;
750
751	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
752	add_slat_sample(td, io_u->ddir, slat_time);
753}
754
755#ifdef FIO_USE_TIMEOUT
756void io_u_set_timeout(struct thread_data *td)
757{
758	assert(td->cur_depth);
759
760	td->timer.it_interval.tv_sec = 0;
761	td->timer.it_interval.tv_usec = 0;
762	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
763	td->timer.it_value.tv_usec = 0;
764	setitimer(ITIMER_REAL, &td->timer, NULL);
765	fio_gettime(&td->timeout_end, NULL);
766}
767
768static void io_u_dump(struct io_u *io_u)
769{
770	unsigned long t_start = mtime_since_now(&io_u->start_time);
771	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
772
773	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
774	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);
775	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
776}
777#else
778void io_u_set_timeout(struct thread_data fio_unused *td)
779{
780}
781#endif
782
783#ifdef FIO_USE_TIMEOUT
784static void io_u_timeout_handler(int fio_unused sig)
785{
786	struct thread_data *td, *__td;
787	pid_t pid = getpid();
788	struct list_head *entry;
789	struct io_u *io_u;
790	int i;
791
792	log_err("fio: io_u timeout\n");
793
794	/*
795	 * TLS would be nice...
796	 */
797	td = NULL;
798	for_each_td(__td, i) {
799		if (__td->pid == pid) {
800			td = __td;
801			break;
802		}
803	}
804
805	if (!td) {
806		log_err("fio: io_u timeout, can't find job\n");
807		exit(1);
808	}
809
810	if (!td->cur_depth) {
811		log_err("fio: timeout without pending work?\n");
812		return;
813	}
814
815	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
816
817	list_for_each(entry, &td->io_u_busylist) {
818		io_u = list_entry(entry, struct io_u, list);
819
820		io_u_dump(io_u);
821	}
822
823	td_verror(td, ETIMEDOUT, "io_u timeout");
824	exit(1);
825}
826#endif
827
828void io_u_init_timeout(void)
829{
830#ifdef FIO_USE_TIMEOUT
831	signal(SIGALRM, io_u_timeout_handler);
832#endif
833}
834