io_u.c revision 60f2c658b923afdd491d556e15a655584b9db306
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 fio_file *f, const unsigned long long block)
28{
29	unsigned int idx = RAND_MAP_IDX(f, block);
30	unsigned int bit = RAND_MAP_BIT(f, block);
31
32	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
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 - f->file_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 == 1) && !random_map_free(f, block))
60			break;
61
62		idx = RAND_MAP_IDX(f, block);
63		bit = RAND_MAP_BIT(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
76static inline unsigned long long last_block(struct thread_data *td,
77					    struct fio_file *f,
78					    enum fio_ddir ddir)
79{
80	unsigned long long max_blocks;
81
82	max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
83	if (!max_blocks)
84		return 0;
85
86	return max_blocks;
87}
88
89/*
90 * Return the next free block in the map.
91 */
92static int get_next_free_block(struct thread_data *td, struct fio_file *f,
93			       enum fio_ddir ddir, unsigned long long *b)
94{
95	unsigned long long min_bs = td->o.rw_min_bs;
96	int i;
97
98	i = f->last_free_lookup;
99	*b = (i * BLOCKS_PER_MAP);
100	while ((*b) * min_bs < f->real_file_size) {
101		if (f->file_map[i] != -1UL) {
102			*b += fio_ffz(f->file_map[i]);
103			if (*b > last_block(td, f, ddir))
104				break;
105			f->last_free_lookup = i;
106			return 0;
107		}
108
109		*b += BLOCKS_PER_MAP;
110		i++;
111	}
112
113	dprint(FD_IO, "failed finding a free block\n");
114	return 1;
115}
116
117static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
118				enum fio_ddir ddir, unsigned long long *b)
119{
120	unsigned long long r;
121	int loops = 5;
122
123	do {
124		r = os_random_long(&td->random_state);
125		dprint(FD_RANDOM, "off rand %llu\n", r);
126		*b = (last_block(td, f, ddir) - 1)
127			* (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 (!file_randommap(td, f))
133			return 0;
134
135		/*
136		 * calculate map offset and check if it's free
137		 */
138		if (random_map_free(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];
192	if (io_u->offset >= f->io_size) {
193		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
194					io_u->offset, f->io_size);
195		return 1;
196	}
197
198	io_u->offset += f->file_offset;
199	if (io_u->offset >= f->real_file_size) {
200		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
201					io_u->offset, f->real_file_size);
202		return 1;
203	}
204
205	return 0;
206}
207
208static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
209{
210	const int ddir = io_u->ddir;
211	unsigned int buflen;
212	long r;
213
214	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
215		buflen = td->o.min_bs[ddir];
216	else {
217		r = os_random_long(&td->bsrange_state);
218		if (!td->o.bssplit_nr) {
219			buflen = (unsigned int)
220					(1 + (double) (td->o.max_bs[ddir] - 1)
221					* r / (RAND_MAX + 1.0));
222		} else {
223			long perc = 0;
224			unsigned int i;
225
226			for (i = 0; i < td->o.bssplit_nr; i++) {
227				struct bssplit *bsp = &td->o.bssplit[i];
228
229				buflen = bsp->bs;
230				perc += bsp->perc;
231				if (r <= ((LONG_MAX / 100L) * perc))
232					break;
233			}
234		}
235		if (!td->o.bs_unaligned) {
236			buflen = (buflen + td->o.min_bs[ddir] - 1)
237					& ~(td->o.min_bs[ddir] - 1);
238		}
239	}
240
241	if (io_u->offset + buflen > io_u->file->real_file_size) {
242		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
243						td->o.min_bs[ddir], ddir);
244		buflen = td->o.min_bs[ddir];
245	}
246
247	return buflen;
248}
249
250static void set_rwmix_bytes(struct thread_data *td)
251{
252	unsigned long issues;
253	unsigned int diff;
254
255	/*
256	 * we do time or byte based switch. this is needed because
257	 * buffered writes may issue a lot quicker than they complete,
258	 * whereas reads do not.
259	 */
260	issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
261	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
262
263	td->rwmix_issues = td->io_issues[td->rwmix_ddir]
264				+ (issues * ((100 - diff)) / diff);
265}
266
267static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
268{
269	unsigned int v;
270	long r;
271
272	r = os_random_long(&td->rwmix_state);
273	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
274	if (v < td->o.rwmix[DDIR_READ])
275		return DDIR_READ;
276
277	return DDIR_WRITE;
278}
279
280/*
281 * Return the data direction for the next io_u. If the job is a
282 * mixed read/write workload, check the rwmix cycle and switch if
283 * necessary.
284 */
285static enum fio_ddir get_rw_ddir(struct thread_data *td)
286{
287	if (td_rw(td)) {
288		/*
289		 * Check if it's time to seed a new data direction.
290		 */
291		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
292			unsigned long long max_bytes;
293			enum fio_ddir ddir;
294
295			/*
296			 * Put a top limit on how many bytes we do for
297			 * one data direction, to avoid overflowing the
298			 * ranges too much
299			 */
300			ddir = get_rand_ddir(td);
301			max_bytes = td->this_io_bytes[ddir];
302			if (max_bytes >=
303			    (td->o.size * td->o.rwmix[ddir] / 100)) {
304				if (!td->rw_end_set[ddir])
305					td->rw_end_set[ddir] = 1;
306
307				ddir ^= 1;
308			}
309
310			if (ddir != td->rwmix_ddir)
311				set_rwmix_bytes(td);
312
313			td->rwmix_ddir = ddir;
314		}
315		return td->rwmix_ddir;
316	} else if (td_read(td))
317		return DDIR_READ;
318	else
319		return DDIR_WRITE;
320}
321
322static void put_file_log(struct thread_data *td, struct fio_file *f)
323{
324	int ret = put_file(td, f);
325
326	if (ret)
327		td_verror(td, ret, "file close");
328}
329
330void put_io_u(struct thread_data *td, struct io_u *io_u)
331{
332	assert((io_u->flags & IO_U_F_FREE) == 0);
333	io_u->flags |= IO_U_F_FREE;
334
335	if (io_u->file)
336		put_file_log(td, io_u->file);
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) && file_randommap(td, io_u->file))
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, unsigned int nr)
425{
426	int index = 0;
427
428	switch (td->cur_depth) {
429	default:
430		index = 6;
431		break;
432	case 32 ... 63:
433		index = 5;
434		break;
435	case 16 ... 31:
436		index = 4;
437		break;
438	case 8 ... 15:
439		index = 3;
440		break;
441	case 4 ... 7:
442		index = 2;
443		break;
444	case 2 ... 3:
445		index = 1;
446	case 1:
447		break;
448	}
449
450	td->ts.io_u_map[index] += nr;
451}
452
453static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
454{
455	int index = 0;
456
457	assert(usec < 1000);
458
459	switch (usec) {
460	case 750 ... 999:
461		index = 9;
462		break;
463	case 500 ... 749:
464		index = 8;
465		break;
466	case 250 ... 499:
467		index = 7;
468		break;
469	case 100 ... 249:
470		index = 6;
471		break;
472	case 50 ... 99:
473		index = 5;
474		break;
475	case 20 ... 49:
476		index = 4;
477		break;
478	case 10 ... 19:
479		index = 3;
480		break;
481	case 4 ... 9:
482		index = 2;
483		break;
484	case 2 ... 3:
485		index = 1;
486	case 0 ... 1:
487		break;
488	}
489
490	assert(index < FIO_IO_U_LAT_U_NR);
491	td->ts.io_u_lat_u[index]++;
492}
493
494static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
495{
496	int index = 0;
497
498	switch (msec) {
499	default:
500		index = 11;
501		break;
502	case 1000 ... 1999:
503		index = 10;
504		break;
505	case 750 ... 999:
506		index = 9;
507		break;
508	case 500 ... 749:
509		index = 8;
510		break;
511	case 250 ... 499:
512		index = 7;
513		break;
514	case 100 ... 249:
515		index = 6;
516		break;
517	case 50 ... 99:
518		index = 5;
519		break;
520	case 20 ... 49:
521		index = 4;
522		break;
523	case 10 ... 19:
524		index = 3;
525		break;
526	case 4 ... 9:
527		index = 2;
528		break;
529	case 2 ... 3:
530		index = 1;
531	case 0 ... 1:
532		break;
533	}
534
535	assert(index < FIO_IO_U_LAT_M_NR);
536	td->ts.io_u_lat_m[index]++;
537}
538
539static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
540{
541	if (usec < 1000)
542		io_u_mark_lat_usec(td, usec);
543	else
544		io_u_mark_lat_msec(td, usec / 1000);
545}
546
547/*
548 * Get next file to service by choosing one at random
549 */
550static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
551					   int badf)
552{
553	struct fio_file *f;
554	int fno;
555
556	do {
557		long r = os_random_long(&td->next_file_state);
558
559		fno = (unsigned int) ((double) td->o.nr_files
560			* (r / (RAND_MAX + 1.0)));
561		f = td->files[fno];
562		if (f->flags & FIO_FILE_DONE)
563			continue;
564
565		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
566			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
567			return f;
568		}
569	} while (1);
570}
571
572/*
573 * Get next file to service by doing round robin between all available ones
574 */
575static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
576					 int badf)
577{
578	unsigned int old_next_file = td->next_file;
579	struct fio_file *f;
580
581	do {
582		f = td->files[td->next_file];
583
584		td->next_file++;
585		if (td->next_file >= td->o.nr_files)
586			td->next_file = 0;
587
588		if (f->flags & FIO_FILE_DONE) {
589			f = NULL;
590			continue;
591		}
592
593		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
594			break;
595
596		f = NULL;
597	} while (td->next_file != old_next_file);
598
599	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
600	return f;
601}
602
603static struct fio_file *get_next_file(struct thread_data *td)
604{
605	struct fio_file *f;
606
607	assert(td->o.nr_files <= td->files_index);
608
609	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
610		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
611				" nr_files=%d\n", td->nr_open_files,
612						  td->nr_done_files,
613						  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		 * optimization to prevent close/open of the same file. This
666		 * way we preserve queueing etc.
667		 */
668		if (td->o.nr_files == 1 && td->o.time_based) {
669			put_file_log(td, f);
670			fio_file_reset(f);
671			goto set_file;
672		}
673
674		/*
675		 * td_io_close() does a put_file() as well, so no need to
676		 * do that here.
677		 */
678		io_u->file = NULL;
679		td_io_close_file(td, f);
680		f->flags |= FIO_FILE_DONE;
681		td->nr_done_files++;
682
683		/*
684		 * probably not the right place to do this, but see
685		 * if we need to open a new file
686		 */
687		if (td->nr_open_files < td->o.open_files &&
688		    td->o.open_files != td->o.nr_files) {
689			f = find_next_new_file(td);
690
691			if (!f || td_io_open_file(td, f))
692				return 1;
693
694			goto set_file;
695		}
696	} while (1);
697
698	return 0;
699}
700
701
702struct io_u *__get_io_u(struct thread_data *td)
703{
704	struct io_u *io_u = NULL;
705
706	if (!list_empty(&td->io_u_requeues))
707		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
708	else if (!queue_full(td)) {
709		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
710
711		io_u->buflen = 0;
712		io_u->resid = 0;
713		io_u->file = NULL;
714		io_u->end_io = NULL;
715	}
716
717	if (io_u) {
718		assert(io_u->flags & IO_U_F_FREE);
719		io_u->flags &= ~IO_U_F_FREE;
720
721		io_u->error = 0;
722		list_del(&io_u->list);
723		list_add(&io_u->list, &td->io_u_busylist);
724		td->cur_depth++;
725	}
726
727	return io_u;
728}
729
730/*
731 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
732 * etc. The returned io_u is fully ready to be prepped and submitted.
733 */
734struct io_u *get_io_u(struct thread_data *td)
735{
736	struct fio_file *f;
737	struct io_u *io_u;
738
739	io_u = __get_io_u(td);
740	if (!io_u) {
741		dprint(FD_IO, "__get_io_u failed\n");
742		return NULL;
743	}
744
745	/*
746	 * from a requeue, io_u already setup
747	 */
748	if (io_u->file)
749		goto out;
750
751	/*
752	 * If using an iolog, grab next piece if any available.
753	 */
754	if (td->o.read_iolog_file) {
755		if (read_iolog_get(td, io_u))
756			goto err_put;
757	} else if (set_io_u_file(td, io_u)) {
758		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
759		goto err_put;
760	}
761
762	f = io_u->file;
763	assert(f->flags & FIO_FILE_OPEN);
764
765	if (io_u->ddir != DDIR_SYNC) {
766		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
767			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
768			goto err_put;
769		}
770
771		f->last_pos = io_u->offset + io_u->buflen;
772
773		if (td->o.verify != VERIFY_NONE)
774			populate_verify_io_u(td, io_u);
775	}
776
777	/*
778	 * Set io data pointers.
779	 */
780	io_u->endpos = io_u->offset + io_u->buflen;
781	io_u->xfer_buf = io_u->buf;
782	io_u->xfer_buflen = io_u->buflen;
783out:
784	if (!td_io_prep(td, io_u)) {
785		fio_gettime(&io_u->start_time, NULL);
786		return io_u;
787	}
788err_put:
789	dprint(FD_IO, "get_io_u failed\n");
790	put_io_u(td, io_u);
791	return NULL;
792}
793
794void io_u_log_error(struct thread_data *td, struct io_u *io_u)
795{
796	const char *msg[] = { "read", "write", "sync" };
797
798	log_err("fio: io_u error");
799
800	if (io_u->file)
801		log_err(" on file %s", io_u->file->file_name);
802
803	log_err(": %s\n", strerror(io_u->error));
804
805	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
806					io_u->offset, io_u->xfer_buflen);
807
808	if (!td->error)
809		td_verror(td, io_u->error, "io_u error");
810}
811
812static void io_completed(struct thread_data *td, struct io_u *io_u,
813			 struct io_completion_data *icd)
814{
815	unsigned long usec;
816
817	dprint_io_u(io_u, "io complete");
818
819	assert(io_u->flags & IO_U_F_FLIGHT);
820	io_u->flags &= ~IO_U_F_FLIGHT;
821
822	if (io_u->ddir == DDIR_SYNC) {
823		td->last_was_sync = 1;
824		return;
825	}
826
827	td->last_was_sync = 0;
828
829	if (!io_u->error) {
830		unsigned int bytes = io_u->buflen - io_u->resid;
831		const enum fio_ddir idx = io_u->ddir;
832		int ret;
833
834		td->io_blocks[idx]++;
835		td->io_bytes[idx] += bytes;
836		td->this_io_bytes[idx] += bytes;
837
838		usec = utime_since(&io_u->issue_time, &icd->time);
839
840		add_clat_sample(td, idx, usec);
841		add_bw_sample(td, idx, &icd->time);
842		io_u_mark_latency(td, usec);
843
844		if (td_write(td) && idx == DDIR_WRITE &&
845		    td->o.do_verify &&
846		    td->o.verify != VERIFY_NONE)
847			log_io_piece(td, io_u);
848
849		icd->bytes_done[idx] += bytes;
850
851		if (io_u->end_io) {
852			ret = io_u->end_io(td, io_u);
853			if (ret && !icd->error)
854				icd->error = ret;
855		}
856	} else {
857		icd->error = io_u->error;
858		io_u_log_error(td, io_u);
859	}
860}
861
862static void init_icd(struct io_completion_data *icd, int nr)
863{
864	fio_gettime(&icd->time, NULL);
865
866	icd->nr = nr;
867
868	icd->error = 0;
869	icd->bytes_done[0] = icd->bytes_done[1] = 0;
870}
871
872static void ios_completed(struct thread_data *td,
873			  struct io_completion_data *icd)
874{
875	struct io_u *io_u;
876	int i;
877
878	for (i = 0; i < icd->nr; i++) {
879		io_u = td->io_ops->event(td, i);
880
881		io_completed(td, io_u, icd);
882		put_io_u(td, io_u);
883	}
884}
885
886/*
887 * Complete a single io_u for the sync engines.
888 */
889long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
890{
891	struct io_completion_data icd;
892
893	init_icd(&icd, 1);
894	io_completed(td, io_u, &icd);
895	put_io_u(td, io_u);
896
897	if (!icd.error)
898		return icd.bytes_done[0] + icd.bytes_done[1];
899
900	td_verror(td, icd.error, "io_u_sync_complete");
901	return -1;
902}
903
904/*
905 * Called to complete min_events number of io for the async engines.
906 */
907long io_u_queued_complete(struct thread_data *td, int min_events)
908{
909	struct io_completion_data icd;
910	struct timespec *tvp = NULL;
911	int ret;
912	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
913
914	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
915
916	if (!min_events)
917		tvp = &ts;
918
919	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
920	if (ret < 0) {
921		td_verror(td, -ret, "td_io_getevents");
922		return ret;
923	} else if (!ret)
924		return ret;
925
926	init_icd(&icd, ret);
927	ios_completed(td, &icd);
928	if (!icd.error)
929		return icd.bytes_done[0] + icd.bytes_done[1];
930
931	td_verror(td, icd.error, "io_u_queued_complete");
932	return -1;
933}
934
935/*
936 * Call when io_u is really queued, to update the submission latency.
937 */
938void io_u_queued(struct thread_data *td, struct io_u *io_u)
939{
940	unsigned long slat_time;
941
942	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
943	add_slat_sample(td, io_u->ddir, slat_time);
944}
945
946#ifdef FIO_USE_TIMEOUT
947void io_u_set_timeout(struct thread_data *td)
948{
949	assert(td->cur_depth);
950
951	td->timer.it_interval.tv_sec = 0;
952	td->timer.it_interval.tv_usec = 0;
953	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
954	td->timer.it_value.tv_usec = 0;
955	setitimer(ITIMER_REAL, &td->timer, NULL);
956	fio_gettime(&td->timeout_end, NULL);
957}
958
959static void io_u_dump(struct io_u *io_u)
960{
961	unsigned long t_start = mtime_since_now(&io_u->start_time);
962	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
963
964	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
965	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
966						io_u->xfer_buf, io_u->buflen,
967						io_u->xfer_buflen,
968						io_u->offset);
969	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
970}
971#else
972void io_u_set_timeout(struct thread_data fio_unused *td)
973{
974}
975#endif
976
977#ifdef FIO_USE_TIMEOUT
978static void io_u_timeout_handler(int fio_unused sig)
979{
980	struct thread_data *td, *__td;
981	pid_t pid = getpid();
982	struct list_head *entry;
983	struct io_u *io_u;
984	int i;
985
986	log_err("fio: io_u timeout\n");
987
988	/*
989	 * TLS would be nice...
990	 */
991	td = NULL;
992	for_each_td(__td, i) {
993		if (__td->pid == pid) {
994			td = __td;
995			break;
996		}
997	}
998
999	if (!td) {
1000		log_err("fio: io_u timeout, can't find job\n");
1001		exit(1);
1002	}
1003
1004	if (!td->cur_depth) {
1005		log_err("fio: timeout without pending work?\n");
1006		return;
1007	}
1008
1009	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1010
1011	list_for_each(entry, &td->io_u_busylist) {
1012		io_u = list_entry(entry, struct io_u, list);
1013
1014		io_u_dump(io_u);
1015	}
1016
1017	td_verror(td, ETIMEDOUT, "io_u timeout");
1018	exit(1);
1019}
1020#endif
1021
1022void io_u_init_timeout(void)
1023{
1024#ifdef FIO_USE_TIMEOUT
1025	signal(SIGALRM, io_u_timeout_handler);
1026#endif
1027}
1028