io_u.c revision 126d65c6fc97d6acdc568aa5a969c012018daf15
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) * (r / ((unsigned long long) RAND_MAX + 1.0));
128
129		/*
130		 * if we are not maintaining a random map, we are done.
131		 */
132		if (td->o.norandommap)
133			return 0;
134
135		/*
136		 * calculate map offset and check if it's free
137		 */
138		if (random_map_free(td, f, *b))
139			return 0;
140
141		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
142									*b);
143	} while (--loops);
144
145	/*
146	 * we get here, if we didn't suceed in looking up a block. generate
147	 * a random start offset into the filemap, and find the first free
148	 * block from there.
149	 */
150	loops = 10;
151	do {
152		f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
153		if (!get_next_free_block(td, f, ddir, b))
154			return 0;
155
156		r = os_random_long(&td->random_state);
157	} while (--loops);
158
159	/*
160	 * that didn't work either, try exhaustive search from the start
161	 */
162	f->last_free_lookup = 0;
163	return get_next_free_block(td, f, ddir, b);
164}
165
166/*
167 * For random io, generate a random new block and see if it's used. Repeat
168 * until we find a free one. For sequential io, just return the end of
169 * the last io issued.
170 */
171static int get_next_offset(struct thread_data *td, struct io_u *io_u)
172{
173	struct fio_file *f = io_u->file;
174	unsigned long long b;
175	enum fio_ddir ddir = io_u->ddir;
176
177	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
178		td->ddir_nr = td->o.ddir_nr;
179
180		if (get_next_rand_offset(td, f, ddir, &b))
181			return 1;
182	} else {
183		if (f->last_pos >= f->real_file_size) {
184			if (!td_random(td) ||
185			     get_next_rand_offset(td, f, ddir, &b))
186				return 1;
187		} else
188			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
189	}
190
191	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
192	if (io_u->offset >= f->real_file_size) {
193		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
194					io_u->offset, f->real_file_size);
195		return 1;
196	}
197
198	return 0;
199}
200
201static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
202{
203	const int ddir = io_u->ddir;
204	unsigned int buflen;
205	long r;
206
207	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
208		buflen = td->o.min_bs[ddir];
209	else {
210		r = os_random_long(&td->bsrange_state);
211		if (!td->o.bssplit_nr)
212			buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
213		else {
214			long perc = 0;
215			unsigned int i;
216
217			for (i = 0; i < td->o.bssplit_nr; i++) {
218				struct bssplit *bsp = &td->o.bssplit[i];
219
220				buflen = bsp->bs;
221				perc += bsp->perc;
222				if (r <= ((LONG_MAX / 100L) * perc))
223					break;
224			}
225		}
226		if (!td->o.bs_unaligned)
227			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
228	}
229
230	if (io_u->offset + buflen > io_u->file->real_file_size) {
231		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
232						td->o.min_bs[ddir], ddir);
233		buflen = td->o.min_bs[ddir];
234	}
235
236	return buflen;
237}
238
239static void set_rwmix_bytes(struct thread_data *td)
240{
241	unsigned long long rbytes;
242	unsigned int diff;
243
244	/*
245	 * we do time or byte based switch. this is needed because
246	 * buffered writes may issue a lot quicker than they complete,
247	 * whereas reads do not.
248	 */
249	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
250	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
251
252	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
253}
254
255static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
256{
257	unsigned int v;
258	long r;
259
260	r = os_random_long(&td->rwmix_state);
261	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
262	if (v < td->o.rwmix[DDIR_READ])
263		return DDIR_READ;
264
265	return DDIR_WRITE;
266}
267
268/*
269 * Return the data direction for the next io_u. If the job is a
270 * mixed read/write workload, check the rwmix cycle and switch if
271 * necessary.
272 */
273static enum fio_ddir get_rw_ddir(struct thread_data *td)
274{
275	if (td_rw(td)) {
276		struct timeval now;
277		unsigned long elapsed;
278		unsigned int cycle;
279
280		fio_gettime(&now, NULL);
281	 	elapsed = mtime_since_now(&td->rwmix_switch);
282
283		/*
284		 * if this is the first cycle, make it shorter
285		 */
286		cycle = td->o.rwmixcycle;
287		if (!td->rwmix_bytes)
288			cycle /= 10;
289
290		/*
291		 * Check if it's time to seed a new data direction.
292		 */
293		if (elapsed >= cycle ||
294		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
295			unsigned long long max_bytes;
296			enum fio_ddir ddir;
297
298			/*
299			 * Put a top limit on how many bytes we do for
300			 * one data direction, to avoid overflowing the
301			 * ranges too much
302			 */
303			ddir = get_rand_ddir(td);
304			max_bytes = td->this_io_bytes[ddir];
305			if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
306				if (!td->rw_end_set[ddir]) {
307					td->rw_end_set[ddir] = 1;
308					memcpy(&td->rw_end[ddir], &now, sizeof(now));
309				}
310				ddir ^= 1;
311			}
312
313			if (ddir != td->rwmix_ddir)
314				set_rwmix_bytes(td);
315
316			td->rwmix_ddir = ddir;
317			memcpy(&td->rwmix_switch, &now, sizeof(now));
318		}
319		return td->rwmix_ddir;
320	} else if (td_read(td))
321		return DDIR_READ;
322	else
323		return DDIR_WRITE;
324}
325
326void put_io_u(struct thread_data *td, struct io_u *io_u)
327{
328	assert((io_u->flags & IO_U_F_FREE) == 0);
329	io_u->flags |= IO_U_F_FREE;
330
331	if (io_u->file) {
332		int ret = put_file(td, io_u->file);
333
334		if (ret)
335			td_verror(td, ret, "file close");
336	}
337
338	io_u->file = NULL;
339	list_del(&io_u->list);
340	list_add(&io_u->list, &td->io_u_freelist);
341	td->cur_depth--;
342}
343
344void requeue_io_u(struct thread_data *td, struct io_u **io_u)
345{
346	struct io_u *__io_u = *io_u;
347
348	__io_u->flags |= IO_U_F_FREE;
349	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
350		td->io_issues[__io_u->ddir]--;
351
352	__io_u->flags &= ~IO_U_F_FLIGHT;
353
354	list_del(&__io_u->list);
355	list_add_tail(&__io_u->list, &td->io_u_requeues);
356	td->cur_depth--;
357	*io_u = NULL;
358}
359
360static int fill_io_u(struct thread_data *td, struct io_u *io_u)
361{
362	if (td->io_ops->flags & FIO_NOIO)
363		goto out;
364
365	/*
366	 * see if it's time to sync
367	 */
368	if (td->o.fsync_blocks &&
369	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
370	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
371		io_u->ddir = DDIR_SYNC;
372		goto out;
373	}
374
375	io_u->ddir = get_rw_ddir(td);
376
377	/*
378	 * See if it's time to switch to a new zone
379	 */
380	if (td->zone_bytes >= td->o.zone_size) {
381		td->zone_bytes = 0;
382		io_u->file->last_pos += td->o.zone_skip;
383		td->io_skip_bytes += td->o.zone_skip;
384	}
385
386	/*
387	 * No log, let the seq/rand engine retrieve the next buflen and
388	 * position.
389	 */
390	if (get_next_offset(td, io_u)) {
391		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
392		return 1;
393	}
394
395	io_u->buflen = get_next_buflen(td, io_u);
396	if (!io_u->buflen) {
397		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
398		return 1;
399	}
400
401	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
402		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
403		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
404				io_u->buflen, io_u->file->real_file_size);
405		return 1;
406	}
407
408	/*
409	 * mark entry before potentially trimming io_u
410	 */
411	if (td_random(td) && !td->o.norandommap)
412		mark_random_map(td, io_u);
413
414	/*
415	 * If using a write iolog, store this entry.
416	 */
417out:
418	dprint_io_u(io_u, "fill_io_u");
419	td->zone_bytes += io_u->buflen;
420	log_io_u(td, io_u);
421	return 0;
422}
423
424void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
425{
426	int index = 0;
427
428	if (io_u->ddir == DDIR_SYNC)
429		return;
430
431	switch (td->cur_depth) {
432	default:
433		index = 6;
434		break;
435	case 32 ... 63:
436		index = 5;
437		break;
438	case 16 ... 31:
439		index = 4;
440		break;
441	case 8 ... 15:
442		index = 3;
443		break;
444	case 4 ... 7:
445		index = 2;
446		break;
447	case 2 ... 3:
448		index = 1;
449	case 1:
450		break;
451	}
452
453	td->ts.io_u_map[index]++;
454	td->ts.total_io_u[io_u->ddir]++;
455}
456
457static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
458{
459	int index = 0;
460
461	assert(usec < 1000);
462
463	switch (usec) {
464	case 750 ... 999:
465		index = 9;
466		break;
467	case 500 ... 749:
468		index = 8;
469		break;
470	case 250 ... 499:
471		index = 7;
472		break;
473	case 100 ... 249:
474		index = 6;
475		break;
476	case 50 ... 99:
477		index = 5;
478		break;
479	case 20 ... 49:
480		index = 4;
481		break;
482	case 10 ... 19:
483		index = 3;
484		break;
485	case 4 ... 9:
486		index = 2;
487		break;
488	case 2 ... 3:
489		index = 1;
490	case 0 ... 1:
491		break;
492	}
493
494	assert(index < FIO_IO_U_LAT_U_NR);
495	td->ts.io_u_lat_u[index]++;
496}
497
498static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
499{
500	int index = 0;
501
502	switch (msec) {
503	default:
504		index = 11;
505		break;
506	case 1000 ... 1999:
507		index = 10;
508		break;
509	case 750 ... 999:
510		index = 9;
511		break;
512	case 500 ... 749:
513		index = 8;
514		break;
515	case 250 ... 499:
516		index = 7;
517		break;
518	case 100 ... 249:
519		index = 6;
520		break;
521	case 50 ... 99:
522		index = 5;
523		break;
524	case 20 ... 49:
525		index = 4;
526		break;
527	case 10 ... 19:
528		index = 3;
529		break;
530	case 4 ... 9:
531		index = 2;
532		break;
533	case 2 ... 3:
534		index = 1;
535	case 0 ... 1:
536		break;
537	}
538
539	assert(index < FIO_IO_U_LAT_M_NR);
540	td->ts.io_u_lat_m[index]++;
541}
542
543static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
544{
545	if (usec < 1000)
546		io_u_mark_lat_usec(td, usec);
547	else
548		io_u_mark_lat_msec(td, usec / 1000);
549}
550
551/*
552 * Get next file to service by choosing one at random
553 */
554static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
555					   int badf)
556{
557	struct fio_file *f;
558	int fno;
559
560	do {
561		long r = os_random_long(&td->next_file_state);
562
563		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
564		f = td->files[fno];
565		if (f->flags & FIO_FILE_DONE)
566			continue;
567
568		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
569			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
570			return f;
571		}
572	} while (1);
573}
574
575/*
576 * Get next file to service by doing round robin between all available ones
577 */
578static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
579					 int badf)
580{
581	unsigned int old_next_file = td->next_file;
582	struct fio_file *f;
583
584	do {
585		f = td->files[td->next_file];
586
587		td->next_file++;
588		if (td->next_file >= td->o.nr_files)
589			td->next_file = 0;
590
591		if (f->flags & FIO_FILE_DONE) {
592			f = NULL;
593			continue;
594		}
595
596		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
597			break;
598
599		f = NULL;
600	} while (td->next_file != old_next_file);
601
602	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
603	return f;
604}
605
606static struct fio_file *get_next_file(struct thread_data *td)
607{
608	struct fio_file *f;
609
610	assert(td->o.nr_files <= td->files_index);
611
612	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
613		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files);
614		return NULL;
615	}
616
617	f = td->file_service_file;
618	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
619		goto out;
620
621	if (td->o.file_service_type == FIO_FSERVICE_RR)
622		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
623	else
624		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
625
626	td->file_service_file = f;
627	td->file_service_left = td->file_service_nr - 1;
628out:
629	dprint(FD_FILE, "get_next_file: %p\n", f);
630	return f;
631}
632
633static struct fio_file *find_next_new_file(struct thread_data *td)
634{
635	struct fio_file *f;
636
637	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
638		return NULL;
639
640	if (td->o.file_service_type == FIO_FSERVICE_RR)
641		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
642	else
643		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
644
645	return f;
646}
647
648static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
649{
650	struct fio_file *f;
651
652	do {
653		f = get_next_file(td);
654		if (!f)
655			return 1;
656
657set_file:
658		io_u->file = f;
659		get_file(f);
660
661		if (!fill_io_u(td, io_u))
662			break;
663
664		/*
665		 * td_io_close() does a put_file() as well, so no need to
666		 * do that here.
667		 */
668		io_u->file = NULL;
669		td_io_close_file(td, f);
670		f->flags |= FIO_FILE_DONE;
671		td->nr_done_files++;
672
673		/*
674		 * probably not the right place to do this, but see
675		 * if we need to open a new file
676		 */
677		if (td->nr_open_files < td->o.open_files &&
678		    td->o.open_files != td->o.nr_files) {
679			f = find_next_new_file(td);
680
681			if (!f || td_io_open_file(td, f))
682				return 1;
683
684			goto set_file;
685		}
686	} while (1);
687
688	return 0;
689}
690
691
692struct io_u *__get_io_u(struct thread_data *td)
693{
694	struct io_u *io_u = NULL;
695
696	if (!list_empty(&td->io_u_requeues))
697		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
698	else if (!queue_full(td)) {
699		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
700
701		io_u->buflen = 0;
702		io_u->resid = 0;
703		io_u->file = NULL;
704		io_u->end_io = NULL;
705	}
706
707	if (io_u) {
708		assert(io_u->flags & IO_U_F_FREE);
709		io_u->flags &= ~IO_U_F_FREE;
710
711		io_u->error = 0;
712		list_del(&io_u->list);
713		list_add(&io_u->list, &td->io_u_busylist);
714		td->cur_depth++;
715	}
716
717	return io_u;
718}
719
720/*
721 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
722 * etc. The returned io_u is fully ready to be prepped and submitted.
723 */
724struct io_u *get_io_u(struct thread_data *td)
725{
726	struct fio_file *f;
727	struct io_u *io_u;
728
729	io_u = __get_io_u(td);
730	if (!io_u) {
731		dprint(FD_IO, "__get_io_u failed\n");
732		return NULL;
733	}
734
735	/*
736	 * from a requeue, io_u already setup
737	 */
738	if (io_u->file)
739		goto out;
740
741	/*
742	 * If using an iolog, grab next piece if any available.
743	 */
744	if (td->o.read_iolog_file) {
745		if (read_iolog_get(td, io_u))
746			goto err_put;
747	} else if (set_io_u_file(td, io_u)) {
748		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
749		goto err_put;
750	}
751
752	f = io_u->file;
753	assert(f->flags & FIO_FILE_OPEN);
754
755	if (io_u->ddir != DDIR_SYNC) {
756		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
757			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
758			goto err_put;
759		}
760
761		f->last_pos = io_u->offset + io_u->buflen;
762
763		if (td->o.verify != VERIFY_NONE)
764			populate_verify_io_u(td, io_u);
765	}
766
767	/*
768	 * Set io data pointers.
769	 */
770	io_u->endpos = io_u->offset + io_u->buflen;
771	io_u->xfer_buf = io_u->buf;
772	io_u->xfer_buflen = io_u->buflen;
773out:
774	if (!td_io_prep(td, io_u)) {
775		fio_gettime(&io_u->start_time, NULL);
776		return io_u;
777	}
778err_put:
779	dprint(FD_IO, "get_io_u failed\n");
780	put_io_u(td, io_u);
781	return NULL;
782}
783
784void io_u_log_error(struct thread_data *td, struct io_u *io_u)
785{
786	const char *msg[] = { "read", "write", "sync" };
787
788	log_err("fio: io_u error");
789
790	if (io_u->file)
791		log_err(" on file %s", io_u->file->file_name);
792
793	log_err(": %s\n", strerror(io_u->error));
794
795	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
796
797	if (!td->error)
798		td_verror(td, io_u->error, "io_u error");
799}
800
801static void io_completed(struct thread_data *td, struct io_u *io_u,
802			 struct io_completion_data *icd)
803{
804	unsigned long usec;
805
806	dprint_io_u(io_u, "io complete");
807
808	assert(io_u->flags & IO_U_F_FLIGHT);
809	io_u->flags &= ~IO_U_F_FLIGHT;
810
811	if (io_u->ddir == DDIR_SYNC) {
812		td->last_was_sync = 1;
813		return;
814	}
815
816	td->last_was_sync = 0;
817
818	if (!io_u->error) {
819		unsigned int bytes = io_u->buflen - io_u->resid;
820		const enum fio_ddir idx = io_u->ddir;
821		int ret;
822
823		td->io_blocks[idx]++;
824		td->io_bytes[idx] += bytes;
825		td->this_io_bytes[idx] += bytes;
826
827		io_u->file->last_completed_pos = io_u->endpos;
828
829		usec = utime_since(&io_u->issue_time, &icd->time);
830
831		add_clat_sample(td, idx, usec);
832		add_bw_sample(td, idx, &icd->time);
833		io_u_mark_latency(td, usec);
834
835		if (td_write(td) && idx == DDIR_WRITE &&
836		    td->o.do_verify &&
837		    td->o.verify != VERIFY_NONE)
838			log_io_piece(td, io_u);
839
840		icd->bytes_done[idx] += bytes;
841
842		if (io_u->end_io) {
843			ret = io_u->end_io(td, io_u);
844			if (ret && !icd->error)
845				icd->error = ret;
846		}
847	} else {
848		icd->error = io_u->error;
849		io_u_log_error(td, io_u);
850	}
851}
852
853static void init_icd(struct io_completion_data *icd, int nr)
854{
855	fio_gettime(&icd->time, NULL);
856
857	icd->nr = nr;
858
859	icd->error = 0;
860	icd->bytes_done[0] = icd->bytes_done[1] = 0;
861}
862
863static void ios_completed(struct thread_data *td,
864			  struct io_completion_data *icd)
865{
866	struct io_u *io_u;
867	int i;
868
869	for (i = 0; i < icd->nr; i++) {
870		io_u = td->io_ops->event(td, i);
871
872		io_completed(td, io_u, icd);
873		put_io_u(td, io_u);
874	}
875}
876
877/*
878 * Complete a single io_u for the sync engines.
879 */
880long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
881{
882	struct io_completion_data icd;
883
884	init_icd(&icd, 1);
885	io_completed(td, io_u, &icd);
886	put_io_u(td, io_u);
887
888	if (!icd.error)
889		return icd.bytes_done[0] + icd.bytes_done[1];
890
891	td_verror(td, icd.error, "io_u_sync_complete");
892	return -1;
893}
894
895/*
896 * Called to complete min_events number of io for the async engines.
897 */
898long io_u_queued_complete(struct thread_data *td, int min_events)
899{
900	struct io_completion_data icd;
901	struct timespec *tvp = NULL;
902	int ret;
903	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
904
905	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
906
907	if (!min_events)
908		tvp = &ts;
909
910	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
911	if (ret < 0) {
912		td_verror(td, -ret, "td_io_getevents");
913		return ret;
914	} else if (!ret)
915		return ret;
916
917	init_icd(&icd, ret);
918	ios_completed(td, &icd);
919	if (!icd.error)
920		return icd.bytes_done[0] + icd.bytes_done[1];
921
922	td_verror(td, icd.error, "io_u_queued_complete");
923	return -1;
924}
925
926/*
927 * Call when io_u is really queued, to update the submission latency.
928 */
929void io_u_queued(struct thread_data *td, struct io_u *io_u)
930{
931	unsigned long slat_time;
932
933	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
934	add_slat_sample(td, io_u->ddir, slat_time);
935}
936
937#ifdef FIO_USE_TIMEOUT
938void io_u_set_timeout(struct thread_data *td)
939{
940	assert(td->cur_depth);
941
942	td->timer.it_interval.tv_sec = 0;
943	td->timer.it_interval.tv_usec = 0;
944	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
945	td->timer.it_value.tv_usec = 0;
946	setitimer(ITIMER_REAL, &td->timer, NULL);
947	fio_gettime(&td->timeout_end, NULL);
948}
949
950static void io_u_dump(struct io_u *io_u)
951{
952	unsigned long t_start = mtime_since_now(&io_u->start_time);
953	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
954
955	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
956	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);
957	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
958}
959#else
960void io_u_set_timeout(struct thread_data fio_unused *td)
961{
962}
963#endif
964
965#ifdef FIO_USE_TIMEOUT
966static void io_u_timeout_handler(int fio_unused sig)
967{
968	struct thread_data *td, *__td;
969	pid_t pid = getpid();
970	struct list_head *entry;
971	struct io_u *io_u;
972	int i;
973
974	log_err("fio: io_u timeout\n");
975
976	/*
977	 * TLS would be nice...
978	 */
979	td = NULL;
980	for_each_td(__td, i) {
981		if (__td->pid == pid) {
982			td = __td;
983			break;
984		}
985	}
986
987	if (!td) {
988		log_err("fio: io_u timeout, can't find job\n");
989		exit(1);
990	}
991
992	if (!td->cur_depth) {
993		log_err("fio: timeout without pending work?\n");
994		return;
995	}
996
997	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
998
999	list_for_each(entry, &td->io_u_busylist) {
1000		io_u = list_entry(entry, struct io_u, list);
1001
1002		io_u_dump(io_u);
1003	}
1004
1005	td_verror(td, ETIMEDOUT, "io_u timeout");
1006	exit(1);
1007}
1008#endif
1009
1010void io_u_init_timeout(void)
1011{
1012#ifdef FIO_USE_TIMEOUT
1013	signal(SIGALRM, io_u_timeout_handler);
1014#endif
1015}
1016