io_u.c revision 303032ae4e13c7ac10b8198c40df98d95a4524a6
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			   const 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	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
34
35	return (f->file_map[idx] & (1UL << bit)) == 0;
36}
37
38/*
39 * Mark a given offset as used in the map.
40 */
41static void mark_random_map(struct thread_data *td, struct io_u *io_u)
42{
43	unsigned int min_bs = td->o.rw_min_bs;
44	struct fio_file *f = io_u->file;
45	unsigned long long block;
46	unsigned int blocks;
47	unsigned int nr_blocks;
48
49	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
50	blocks = 0;
51	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
52
53	while (blocks < nr_blocks) {
54		unsigned int idx, bit;
55
56		/*
57		 * If we have a mixed random workload, we may
58		 * encounter blocks we already did IO to.
59		 */
60		if ((td->o.ddir_nr == 1) && !random_map_free(td, f, block))
61			break;
62
63		idx = RAND_MAP_IDX(td, f, block);
64		bit = RAND_MAP_BIT(td, f, block);
65
66		fio_assert(td, idx < f->num_maps);
67
68		f->file_map[idx] |= (1UL << bit);
69		block++;
70		blocks++;
71	}
72
73	if ((blocks * min_bs) < io_u->buflen)
74		io_u->buflen = blocks * min_bs;
75}
76
77static inline unsigned long long last_block(struct thread_data *td,
78					    struct fio_file *f,
79					    enum fio_ddir ddir)
80{
81	unsigned long long max_blocks;
82
83	max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
84	if (!max_blocks)
85		return 0;
86
87	return max_blocks - 1;
88}
89
90/*
91 * Return the next free block in the map.
92 */
93static int get_next_free_block(struct thread_data *td, struct fio_file *f,
94			       enum fio_ddir ddir, unsigned long long *b)
95{
96	unsigned long long min_bs = td->o.rw_min_bs;
97	int i;
98
99	i = f->last_free_lookup;
100	*b = (i * BLOCKS_PER_MAP);
101	while ((*b) * min_bs < f->real_file_size) {
102		if (f->file_map[i] != -1UL) {
103			*b += fio_ffz(f->file_map[i]);
104			if (*b > last_block(td, f, ddir))
105				break;
106			f->last_free_lookup = i;
107			return 0;
108		}
109
110		*b += BLOCKS_PER_MAP;
111		i++;
112	}
113
114	dprint(FD_IO, "failed finding a free block\n");
115	return 1;
116}
117
118static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
119				enum fio_ddir ddir, unsigned long long *b)
120{
121	unsigned long long r;
122	int loops = 5;
123
124	do {
125		r = os_random_long(&td->random_state);
126		dprint(FD_RANDOM, "off rand %llu\n", r);
127		*b = (last_block(td, f, ddir) - 1)
128			* (r / ((unsigned long long) RAND_MAX + 1.0));
129
130		/*
131		 * if we are not maintaining a random map, we are done.
132		 */
133		if (!file_randommap(td, f))
134			return 0;
135
136		/*
137		 * calculate map offset and check if it's free
138		 */
139		if (random_map_free(td, f, *b))
140			return 0;
141
142		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
143									*b);
144	} while (--loops);
145
146	/*
147	 * we get here, if we didn't suceed in looking up a block. generate
148	 * a random start offset into the filemap, and find the first free
149	 * block from there.
150	 */
151	loops = 10;
152	do {
153		f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
154		if (!get_next_free_block(td, f, ddir, b))
155			return 0;
156
157		r = os_random_long(&td->random_state);
158	} while (--loops);
159
160	/*
161	 * that didn't work either, try exhaustive search from the start
162	 */
163	f->last_free_lookup = 0;
164	return get_next_free_block(td, f, ddir, b);
165}
166
167/*
168 * For random io, generate a random new block and see if it's used. Repeat
169 * until we find a free one. For sequential io, just return the end of
170 * the last io issued.
171 */
172static int get_next_offset(struct thread_data *td, struct io_u *io_u)
173{
174	struct fio_file *f = io_u->file;
175	unsigned long long b;
176	enum fio_ddir ddir = io_u->ddir;
177
178	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
179		td->ddir_nr = td->o.ddir_nr;
180
181		if (get_next_rand_offset(td, f, ddir, &b))
182			return 1;
183	} else {
184		if (f->last_pos >= f->real_file_size) {
185			if (!td_random(td) ||
186			     get_next_rand_offset(td, f, ddir, &b))
187				return 1;
188		} else
189			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
190	}
191
192	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
193	if (io_u->offset >= f->real_file_size) {
194		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
195					io_u->offset, f->real_file_size);
196		return 1;
197	}
198
199	return 0;
200}
201
202static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
203{
204	const int ddir = io_u->ddir;
205	unsigned int buflen;
206	long r;
207
208	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
209		buflen = td->o.min_bs[ddir];
210	else {
211		r = os_random_long(&td->bsrange_state);
212		if (!td->o.bssplit_nr) {
213			buflen = (unsigned int)
214					(1 + (double) (td->o.max_bs[ddir] - 1)
215					* r / (RAND_MAX + 1.0));
216		} else {
217			long perc = 0;
218			unsigned int i;
219
220			for (i = 0; i < td->o.bssplit_nr; i++) {
221				struct bssplit *bsp = &td->o.bssplit[i];
222
223				buflen = bsp->bs;
224				perc += bsp->perc;
225				if (r <= ((LONG_MAX / 100L) * perc))
226					break;
227			}
228		}
229		if (!td->o.bs_unaligned) {
230			buflen = (buflen + td->o.min_bs[ddir] - 1)
231					& ~(td->o.min_bs[ddir] - 1);
232		}
233	}
234
235	if (io_u->offset + buflen > io_u->file->real_file_size) {
236		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
237						td->o.min_bs[ddir], ddir);
238		buflen = td->o.min_bs[ddir];
239	}
240
241	return buflen;
242}
243
244static void set_rwmix_bytes(struct thread_data *td)
245{
246	unsigned long long rbytes;
247	unsigned int diff;
248
249	/*
250	 * we do time or byte based switch. this is needed because
251	 * buffered writes may issue a lot quicker than they complete,
252	 * whereas reads do not.
253	 */
254	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
255	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
256
257	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir]
258				+ (rbytes * ((100 - diff)) / diff);
259}
260
261static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
262{
263	unsigned int v;
264	long r;
265
266	r = os_random_long(&td->rwmix_state);
267	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
268	if (v < td->o.rwmix[DDIR_READ])
269		return DDIR_READ;
270
271	return DDIR_WRITE;
272}
273
274/*
275 * Return the data direction for the next io_u. If the job is a
276 * mixed read/write workload, check the rwmix cycle and switch if
277 * necessary.
278 */
279static enum fio_ddir get_rw_ddir(struct thread_data *td)
280{
281	if (td_rw(td)) {
282		struct timeval now;
283		unsigned long elapsed;
284		unsigned int cycle;
285
286		fio_gettime(&now, NULL);
287		elapsed = mtime_since_now(&td->rwmix_switch);
288
289		/*
290		 * if this is the first cycle, make it shorter
291		 */
292		cycle = td->o.rwmixcycle;
293		if (!td->rwmix_bytes)
294			cycle /= 10;
295
296		/*
297		 * Check if it's time to seed a new data direction.
298		 */
299		if (elapsed >= cycle ||
300		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
301			unsigned long long max_bytes;
302			enum fio_ddir ddir;
303
304			/*
305			 * Put a top limit on how many bytes we do for
306			 * one data direction, to avoid overflowing the
307			 * ranges too much
308			 */
309			ddir = get_rand_ddir(td);
310			max_bytes = td->this_io_bytes[ddir];
311			if (max_bytes >=
312			    (td->o.size * td->o.rwmix[ddir] / 100)) {
313				if (!td->rw_end_set[ddir]) {
314					td->rw_end_set[ddir] = 1;
315					memcpy(&td->rw_end[ddir], &now,
316						sizeof(now));
317				}
318				ddir ^= 1;
319			}
320
321			if (ddir != td->rwmix_ddir)
322				set_rwmix_bytes(td);
323
324			td->rwmix_ddir = ddir;
325			memcpy(&td->rwmix_switch, &now, sizeof(now));
326		}
327		return td->rwmix_ddir;
328	} else if (td_read(td))
329		return DDIR_READ;
330	else
331		return DDIR_WRITE;
332}
333
334void put_io_u(struct thread_data *td, struct io_u *io_u)
335{
336	assert((io_u->flags & IO_U_F_FREE) == 0);
337	io_u->flags |= IO_U_F_FREE;
338
339	if (io_u->file) {
340		int ret = put_file(td, io_u->file);
341
342		if (ret)
343			td_verror(td, ret, "file close");
344	}
345
346	io_u->file = NULL;
347	list_del(&io_u->list);
348	list_add(&io_u->list, &td->io_u_freelist);
349	td->cur_depth--;
350}
351
352void requeue_io_u(struct thread_data *td, struct io_u **io_u)
353{
354	struct io_u *__io_u = *io_u;
355
356	__io_u->flags |= IO_U_F_FREE;
357	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
358		td->io_issues[__io_u->ddir]--;
359
360	__io_u->flags &= ~IO_U_F_FLIGHT;
361
362	list_del(&__io_u->list);
363	list_add_tail(&__io_u->list, &td->io_u_requeues);
364	td->cur_depth--;
365	*io_u = NULL;
366}
367
368static int fill_io_u(struct thread_data *td, struct io_u *io_u)
369{
370	if (td->io_ops->flags & FIO_NOIO)
371		goto out;
372
373	/*
374	 * see if it's time to sync
375	 */
376	if (td->o.fsync_blocks &&
377	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
378	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
379		io_u->ddir = DDIR_SYNC;
380		goto out;
381	}
382
383	io_u->ddir = get_rw_ddir(td);
384
385	/*
386	 * See if it's time to switch to a new zone
387	 */
388	if (td->zone_bytes >= td->o.zone_size) {
389		td->zone_bytes = 0;
390		io_u->file->last_pos += td->o.zone_skip;
391		td->io_skip_bytes += td->o.zone_skip;
392	}
393
394	/*
395	 * No log, let the seq/rand engine retrieve the next buflen and
396	 * position.
397	 */
398	if (get_next_offset(td, io_u)) {
399		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
400		return 1;
401	}
402
403	io_u->buflen = get_next_buflen(td, io_u);
404	if (!io_u->buflen) {
405		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
406		return 1;
407	}
408
409	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
410		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
411		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
412				io_u->buflen, io_u->file->real_file_size);
413		return 1;
414	}
415
416	/*
417	 * mark entry before potentially trimming io_u
418	 */
419	if (td_random(td) && file_randommap(td, io_u->file))
420		mark_random_map(td, io_u);
421
422	/*
423	 * If using a write iolog, store this entry.
424	 */
425out:
426	dprint_io_u(io_u, "fill_io_u");
427	td->zone_bytes += io_u->buflen;
428	log_io_u(td, io_u);
429	return 0;
430}
431
432void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
433{
434	int index = 0;
435
436	if (io_u->ddir == DDIR_SYNC)
437		return;
438
439	switch (td->cur_depth) {
440	default:
441		index = 6;
442		break;
443	case 32 ... 63:
444		index = 5;
445		break;
446	case 16 ... 31:
447		index = 4;
448		break;
449	case 8 ... 15:
450		index = 3;
451		break;
452	case 4 ... 7:
453		index = 2;
454		break;
455	case 2 ... 3:
456		index = 1;
457	case 1:
458		break;
459	}
460
461	td->ts.io_u_map[index]++;
462	td->ts.total_io_u[io_u->ddir]++;
463}
464
465static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
466{
467	int index = 0;
468
469	assert(usec < 1000);
470
471	switch (usec) {
472	case 750 ... 999:
473		index = 9;
474		break;
475	case 500 ... 749:
476		index = 8;
477		break;
478	case 250 ... 499:
479		index = 7;
480		break;
481	case 100 ... 249:
482		index = 6;
483		break;
484	case 50 ... 99:
485		index = 5;
486		break;
487	case 20 ... 49:
488		index = 4;
489		break;
490	case 10 ... 19:
491		index = 3;
492		break;
493	case 4 ... 9:
494		index = 2;
495		break;
496	case 2 ... 3:
497		index = 1;
498	case 0 ... 1:
499		break;
500	}
501
502	assert(index < FIO_IO_U_LAT_U_NR);
503	td->ts.io_u_lat_u[index]++;
504}
505
506static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
507{
508	int index = 0;
509
510	switch (msec) {
511	default:
512		index = 11;
513		break;
514	case 1000 ... 1999:
515		index = 10;
516		break;
517	case 750 ... 999:
518		index = 9;
519		break;
520	case 500 ... 749:
521		index = 8;
522		break;
523	case 250 ... 499:
524		index = 7;
525		break;
526	case 100 ... 249:
527		index = 6;
528		break;
529	case 50 ... 99:
530		index = 5;
531		break;
532	case 20 ... 49:
533		index = 4;
534		break;
535	case 10 ... 19:
536		index = 3;
537		break;
538	case 4 ... 9:
539		index = 2;
540		break;
541	case 2 ... 3:
542		index = 1;
543	case 0 ... 1:
544		break;
545	}
546
547	assert(index < FIO_IO_U_LAT_M_NR);
548	td->ts.io_u_lat_m[index]++;
549}
550
551static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
552{
553	if (usec < 1000)
554		io_u_mark_lat_usec(td, usec);
555	else
556		io_u_mark_lat_msec(td, usec / 1000);
557}
558
559/*
560 * Get next file to service by choosing one at random
561 */
562static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
563					   int badf)
564{
565	struct fio_file *f;
566	int fno;
567
568	do {
569		long r = os_random_long(&td->next_file_state);
570
571		fno = (unsigned int) ((double) td->o.nr_files
572			* (r / (RAND_MAX + 1.0)));
573		f = td->files[fno];
574		if (f->flags & FIO_FILE_DONE)
575			continue;
576
577		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
578			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
579			return f;
580		}
581	} while (1);
582}
583
584/*
585 * Get next file to service by doing round robin between all available ones
586 */
587static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
588					 int badf)
589{
590	unsigned int old_next_file = td->next_file;
591	struct fio_file *f;
592
593	do {
594		f = td->files[td->next_file];
595
596		td->next_file++;
597		if (td->next_file >= td->o.nr_files)
598			td->next_file = 0;
599
600		if (f->flags & FIO_FILE_DONE) {
601			f = NULL;
602			continue;
603		}
604
605		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
606			break;
607
608		f = NULL;
609	} while (td->next_file != old_next_file);
610
611	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
612	return f;
613}
614
615static struct fio_file *get_next_file(struct thread_data *td)
616{
617	struct fio_file *f;
618
619	assert(td->o.nr_files <= td->files_index);
620
621	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
622		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
623				" nr_files=%d\n", td->nr_open_files,
624						  td->nr_done_files,
625						  td->o.nr_files);
626		return NULL;
627	}
628
629	f = td->file_service_file;
630	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
631		goto out;
632
633	if (td->o.file_service_type == FIO_FSERVICE_RR)
634		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
635	else
636		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
637
638	td->file_service_file = f;
639	td->file_service_left = td->file_service_nr - 1;
640out:
641	dprint(FD_FILE, "get_next_file: %p\n", f);
642	return f;
643}
644
645static struct fio_file *find_next_new_file(struct thread_data *td)
646{
647	struct fio_file *f;
648
649	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
650		return NULL;
651
652	if (td->o.file_service_type == FIO_FSERVICE_RR)
653		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
654	else
655		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
656
657	return f;
658}
659
660static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
661{
662	struct fio_file *f;
663
664	do {
665		f = get_next_file(td);
666		if (!f)
667			return 1;
668
669set_file:
670		io_u->file = f;
671		get_file(f);
672
673		if (!fill_io_u(td, io_u))
674			break;
675
676		/*
677		 * td_io_close() does a put_file() as well, so no need to
678		 * do that here.
679		 */
680		io_u->file = NULL;
681		td_io_close_file(td, f);
682		f->flags |= FIO_FILE_DONE;
683		td->nr_done_files++;
684
685		/*
686		 * probably not the right place to do this, but see
687		 * if we need to open a new file
688		 */
689		if (td->nr_open_files < td->o.open_files &&
690		    td->o.open_files != td->o.nr_files) {
691			f = find_next_new_file(td);
692
693			if (!f || td_io_open_file(td, f))
694				return 1;
695
696			goto set_file;
697		}
698	} while (1);
699
700	return 0;
701}
702
703
704struct io_u *__get_io_u(struct thread_data *td)
705{
706	struct io_u *io_u = NULL;
707
708	if (!list_empty(&td->io_u_requeues))
709		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
710	else if (!queue_full(td)) {
711		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
712
713		io_u->buflen = 0;
714		io_u->resid = 0;
715		io_u->file = NULL;
716		io_u->end_io = NULL;
717	}
718
719	if (io_u) {
720		assert(io_u->flags & IO_U_F_FREE);
721		io_u->flags &= ~IO_U_F_FREE;
722
723		io_u->error = 0;
724		list_del(&io_u->list);
725		list_add(&io_u->list, &td->io_u_busylist);
726		td->cur_depth++;
727	}
728
729	return io_u;
730}
731
732/*
733 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
734 * etc. The returned io_u is fully ready to be prepped and submitted.
735 */
736struct io_u *get_io_u(struct thread_data *td)
737{
738	struct fio_file *f;
739	struct io_u *io_u;
740
741	io_u = __get_io_u(td);
742	if (!io_u) {
743		dprint(FD_IO, "__get_io_u failed\n");
744		return NULL;
745	}
746
747	/*
748	 * from a requeue, io_u already setup
749	 */
750	if (io_u->file)
751		goto out;
752
753	/*
754	 * If using an iolog, grab next piece if any available.
755	 */
756	if (td->o.read_iolog_file) {
757		if (read_iolog_get(td, io_u))
758			goto err_put;
759	} else if (set_io_u_file(td, io_u)) {
760		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
761		goto err_put;
762	}
763
764	f = io_u->file;
765	assert(f->flags & FIO_FILE_OPEN);
766
767	if (io_u->ddir != DDIR_SYNC) {
768		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
769			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
770			goto err_put;
771		}
772
773		f->last_pos = io_u->offset + io_u->buflen;
774
775		if (td->o.verify != VERIFY_NONE)
776			populate_verify_io_u(td, io_u);
777	}
778
779	/*
780	 * Set io data pointers.
781	 */
782	io_u->endpos = io_u->offset + io_u->buflen;
783	io_u->xfer_buf = io_u->buf;
784	io_u->xfer_buflen = io_u->buflen;
785out:
786	if (!td_io_prep(td, io_u)) {
787		fio_gettime(&io_u->start_time, NULL);
788		return io_u;
789	}
790err_put:
791	dprint(FD_IO, "get_io_u failed\n");
792	put_io_u(td, io_u);
793	return NULL;
794}
795
796void io_u_log_error(struct thread_data *td, struct io_u *io_u)
797{
798	const char *msg[] = { "read", "write", "sync" };
799
800	log_err("fio: io_u error");
801
802	if (io_u->file)
803		log_err(" on file %s", io_u->file->file_name);
804
805	log_err(": %s\n", strerror(io_u->error));
806
807	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
808					io_u->offset, io_u->xfer_buflen);
809
810	if (!td->error)
811		td_verror(td, io_u->error, "io_u error");
812}
813
814static void io_completed(struct thread_data *td, struct io_u *io_u,
815			 struct io_completion_data *icd)
816{
817	unsigned long usec;
818
819	dprint_io_u(io_u, "io complete");
820
821	assert(io_u->flags & IO_U_F_FLIGHT);
822	io_u->flags &= ~IO_U_F_FLIGHT;
823
824	if (io_u->ddir == DDIR_SYNC) {
825		td->last_was_sync = 1;
826		return;
827	}
828
829	td->last_was_sync = 0;
830
831	if (!io_u->error) {
832		unsigned int bytes = io_u->buflen - io_u->resid;
833		const enum fio_ddir idx = io_u->ddir;
834		int ret;
835
836		td->io_blocks[idx]++;
837		td->io_bytes[idx] += bytes;
838		td->this_io_bytes[idx] += bytes;
839
840		usec = utime_since(&io_u->issue_time, &icd->time);
841
842		add_clat_sample(td, idx, usec);
843		add_bw_sample(td, idx, &icd->time);
844		io_u_mark_latency(td, usec);
845
846		if (td_write(td) && idx == DDIR_WRITE &&
847		    td->o.do_verify &&
848		    td->o.verify != VERIFY_NONE)
849			log_io_piece(td, io_u);
850
851		icd->bytes_done[idx] += bytes;
852
853		if (io_u->end_io) {
854			ret = io_u->end_io(td, io_u);
855			if (ret && !icd->error)
856				icd->error = ret;
857		}
858	} else {
859		icd->error = io_u->error;
860		io_u_log_error(td, io_u);
861	}
862}
863
864static void init_icd(struct io_completion_data *icd, int nr)
865{
866	fio_gettime(&icd->time, NULL);
867
868	icd->nr = nr;
869
870	icd->error = 0;
871	icd->bytes_done[0] = icd->bytes_done[1] = 0;
872}
873
874static void ios_completed(struct thread_data *td,
875			  struct io_completion_data *icd)
876{
877	struct io_u *io_u;
878	int i;
879
880	for (i = 0; i < icd->nr; i++) {
881		io_u = td->io_ops->event(td, i);
882
883		io_completed(td, io_u, icd);
884		put_io_u(td, io_u);
885	}
886}
887
888/*
889 * Complete a single io_u for the sync engines.
890 */
891long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
892{
893	struct io_completion_data icd;
894
895	init_icd(&icd, 1);
896	io_completed(td, io_u, &icd);
897	put_io_u(td, io_u);
898
899	if (!icd.error)
900		return icd.bytes_done[0] + icd.bytes_done[1];
901
902	td_verror(td, icd.error, "io_u_sync_complete");
903	return -1;
904}
905
906/*
907 * Called to complete min_events number of io for the async engines.
908 */
909long io_u_queued_complete(struct thread_data *td, int min_events)
910{
911	struct io_completion_data icd;
912	struct timespec *tvp = NULL;
913	int ret;
914	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
915
916	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
917
918	if (!min_events)
919		tvp = &ts;
920
921	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
922	if (ret < 0) {
923		td_verror(td, -ret, "td_io_getevents");
924		return ret;
925	} else if (!ret)
926		return ret;
927
928	init_icd(&icd, ret);
929	ios_completed(td, &icd);
930	if (!icd.error)
931		return icd.bytes_done[0] + icd.bytes_done[1];
932
933	td_verror(td, icd.error, "io_u_queued_complete");
934	return -1;
935}
936
937/*
938 * Call when io_u is really queued, to update the submission latency.
939 */
940void io_u_queued(struct thread_data *td, struct io_u *io_u)
941{
942	unsigned long slat_time;
943
944	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
945	add_slat_sample(td, io_u->ddir, slat_time);
946}
947
948#ifdef FIO_USE_TIMEOUT
949void io_u_set_timeout(struct thread_data *td)
950{
951	assert(td->cur_depth);
952
953	td->timer.it_interval.tv_sec = 0;
954	td->timer.it_interval.tv_usec = 0;
955	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
956	td->timer.it_value.tv_usec = 0;
957	setitimer(ITIMER_REAL, &td->timer, NULL);
958	fio_gettime(&td->timeout_end, NULL);
959}
960
961static void io_u_dump(struct io_u *io_u)
962{
963	unsigned long t_start = mtime_since_now(&io_u->start_time);
964	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
965
966	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
967	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
968						io_u->xfer_buf, io_u->buflen,
969						io_u->xfer_buflen,
970						io_u->offset);
971	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
972}
973#else
974void io_u_set_timeout(struct thread_data fio_unused *td)
975{
976}
977#endif
978
979#ifdef FIO_USE_TIMEOUT
980static void io_u_timeout_handler(int fio_unused sig)
981{
982	struct thread_data *td, *__td;
983	pid_t pid = getpid();
984	struct list_head *entry;
985	struct io_u *io_u;
986	int i;
987
988	log_err("fio: io_u timeout\n");
989
990	/*
991	 * TLS would be nice...
992	 */
993	td = NULL;
994	for_each_td(__td, i) {
995		if (__td->pid == pid) {
996			td = __td;
997			break;
998		}
999	}
1000
1001	if (!td) {
1002		log_err("fio: io_u timeout, can't find job\n");
1003		exit(1);
1004	}
1005
1006	if (!td->cur_depth) {
1007		log_err("fio: timeout without pending work?\n");
1008		return;
1009	}
1010
1011	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1012
1013	list_for_each(entry, &td->io_u_busylist) {
1014		io_u = list_entry(entry, struct io_u, list);
1015
1016		io_u_dump(io_u);
1017	}
1018
1019	td_verror(td, ETIMEDOUT, "io_u timeout");
1020	exit(1);
1021}
1022#endif
1023
1024void io_u_init_timeout(void)
1025{
1026#ifdef FIO_USE_TIMEOUT
1027	signal(SIGALRM, io_u_timeout_handler);
1028#endif
1029}
1030