io_u.c revision d8005759746a2cb5c8269201911b1997aa714e80
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]) + 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)
213					(1 + (double) (td->o.max_bs[ddir] - 1)
214					* r / (RAND_MAX + 1.0));
215		} else {
216			long perc = 0;
217			unsigned int i;
218
219			for (i = 0; i < td->o.bssplit_nr; i++) {
220				struct bssplit *bsp = &td->o.bssplit[i];
221
222				buflen = bsp->bs;
223				perc += bsp->perc;
224				if (r <= ((LONG_MAX / 100L) * perc))
225					break;
226			}
227		}
228		if (!td->o.bs_unaligned) {
229			buflen = (buflen + td->o.min_bs[ddir] - 1)
230					& ~(td->o.min_bs[ddir] - 1);
231		}
232	}
233
234	if (io_u->offset + buflen > io_u->file->real_file_size) {
235		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
236						td->o.min_bs[ddir], ddir);
237		buflen = td->o.min_bs[ddir];
238	}
239
240	return buflen;
241}
242
243static void set_rwmix_bytes(struct thread_data *td)
244{
245	unsigned long issues;
246	unsigned int diff;
247
248	/*
249	 * we do time or byte based switch. this is needed because
250	 * buffered writes may issue a lot quicker than they complete,
251	 * whereas reads do not.
252	 */
253	issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
254	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
255
256	td->rwmix_issues = td->io_issues[td->rwmix_ddir]
257				+ (issues * ((100 - diff)) / diff);
258}
259
260static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
261{
262	unsigned int v;
263	long r;
264
265	r = os_random_long(&td->rwmix_state);
266	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
267	if (v < td->o.rwmix[DDIR_READ])
268		return DDIR_READ;
269
270	return DDIR_WRITE;
271}
272
273/*
274 * Return the data direction for the next io_u. If the job is a
275 * mixed read/write workload, check the rwmix cycle and switch if
276 * necessary.
277 */
278static enum fio_ddir get_rw_ddir(struct thread_data *td)
279{
280	if (td_rw(td)) {
281		/*
282		 * Check if it's time to seed a new data direction.
283		 */
284		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
285			unsigned long long max_bytes;
286			enum fio_ddir ddir;
287
288			/*
289			 * Put a top limit on how many bytes we do for
290			 * one data direction, to avoid overflowing the
291			 * ranges too much
292			 */
293			ddir = get_rand_ddir(td);
294			max_bytes = td->this_io_bytes[ddir];
295			if (max_bytes >=
296			    (td->o.size * td->o.rwmix[ddir] / 100)) {
297				if (!td->rw_end_set[ddir])
298					td->rw_end_set[ddir] = 1;
299
300				ddir ^= 1;
301			}
302
303			if (ddir != td->rwmix_ddir)
304				set_rwmix_bytes(td);
305
306			td->rwmix_ddir = ddir;
307		}
308		return td->rwmix_ddir;
309	} else if (td_read(td))
310		return DDIR_READ;
311	else
312		return DDIR_WRITE;
313}
314
315void put_io_u(struct thread_data *td, struct io_u *io_u)
316{
317	assert((io_u->flags & IO_U_F_FREE) == 0);
318	io_u->flags |= IO_U_F_FREE;
319
320	if (io_u->file) {
321		int ret = put_file(td, io_u->file);
322
323		if (ret)
324			td_verror(td, ret, "file close");
325	}
326
327	io_u->file = NULL;
328	list_del(&io_u->list);
329	list_add(&io_u->list, &td->io_u_freelist);
330	td->cur_depth--;
331}
332
333void requeue_io_u(struct thread_data *td, struct io_u **io_u)
334{
335	struct io_u *__io_u = *io_u;
336
337	__io_u->flags |= IO_U_F_FREE;
338	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
339		td->io_issues[__io_u->ddir]--;
340
341	__io_u->flags &= ~IO_U_F_FLIGHT;
342
343	list_del(&__io_u->list);
344	list_add_tail(&__io_u->list, &td->io_u_requeues);
345	td->cur_depth--;
346	*io_u = NULL;
347}
348
349static int fill_io_u(struct thread_data *td, struct io_u *io_u)
350{
351	if (td->io_ops->flags & FIO_NOIO)
352		goto out;
353
354	/*
355	 * see if it's time to sync
356	 */
357	if (td->o.fsync_blocks &&
358	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
359	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
360		io_u->ddir = DDIR_SYNC;
361		goto out;
362	}
363
364	io_u->ddir = get_rw_ddir(td);
365
366	/*
367	 * See if it's time to switch to a new zone
368	 */
369	if (td->zone_bytes >= td->o.zone_size) {
370		td->zone_bytes = 0;
371		io_u->file->last_pos += td->o.zone_skip;
372		td->io_skip_bytes += td->o.zone_skip;
373	}
374
375	/*
376	 * No log, let the seq/rand engine retrieve the next buflen and
377	 * position.
378	 */
379	if (get_next_offset(td, io_u)) {
380		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
381		return 1;
382	}
383
384	io_u->buflen = get_next_buflen(td, io_u);
385	if (!io_u->buflen) {
386		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
387		return 1;
388	}
389
390	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
391		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
392		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
393				io_u->buflen, io_u->file->real_file_size);
394		return 1;
395	}
396
397	/*
398	 * mark entry before potentially trimming io_u
399	 */
400	if (td_random(td) && file_randommap(td, io_u->file))
401		mark_random_map(td, io_u);
402
403	/*
404	 * If using a write iolog, store this entry.
405	 */
406out:
407	dprint_io_u(io_u, "fill_io_u");
408	td->zone_bytes += io_u->buflen;
409	log_io_u(td, io_u);
410	return 0;
411}
412
413void io_u_mark_depth(struct thread_data *td, unsigned int nr)
414{
415	int index = 0;
416
417	switch (td->cur_depth) {
418	default:
419		index = 6;
420		break;
421	case 32 ... 63:
422		index = 5;
423		break;
424	case 16 ... 31:
425		index = 4;
426		break;
427	case 8 ... 15:
428		index = 3;
429		break;
430	case 4 ... 7:
431		index = 2;
432		break;
433	case 2 ... 3:
434		index = 1;
435	case 1:
436		break;
437	}
438
439	td->ts.io_u_map[index] += nr;
440}
441
442static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
443{
444	int index = 0;
445
446	assert(usec < 1000);
447
448	switch (usec) {
449	case 750 ... 999:
450		index = 9;
451		break;
452	case 500 ... 749:
453		index = 8;
454		break;
455	case 250 ... 499:
456		index = 7;
457		break;
458	case 100 ... 249:
459		index = 6;
460		break;
461	case 50 ... 99:
462		index = 5;
463		break;
464	case 20 ... 49:
465		index = 4;
466		break;
467	case 10 ... 19:
468		index = 3;
469		break;
470	case 4 ... 9:
471		index = 2;
472		break;
473	case 2 ... 3:
474		index = 1;
475	case 0 ... 1:
476		break;
477	}
478
479	assert(index < FIO_IO_U_LAT_U_NR);
480	td->ts.io_u_lat_u[index]++;
481}
482
483static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
484{
485	int index = 0;
486
487	switch (msec) {
488	default:
489		index = 11;
490		break;
491	case 1000 ... 1999:
492		index = 10;
493		break;
494	case 750 ... 999:
495		index = 9;
496		break;
497	case 500 ... 749:
498		index = 8;
499		break;
500	case 250 ... 499:
501		index = 7;
502		break;
503	case 100 ... 249:
504		index = 6;
505		break;
506	case 50 ... 99:
507		index = 5;
508		break;
509	case 20 ... 49:
510		index = 4;
511		break;
512	case 10 ... 19:
513		index = 3;
514		break;
515	case 4 ... 9:
516		index = 2;
517		break;
518	case 2 ... 3:
519		index = 1;
520	case 0 ... 1:
521		break;
522	}
523
524	assert(index < FIO_IO_U_LAT_M_NR);
525	td->ts.io_u_lat_m[index]++;
526}
527
528static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
529{
530	if (usec < 1000)
531		io_u_mark_lat_usec(td, usec);
532	else
533		io_u_mark_lat_msec(td, usec / 1000);
534}
535
536/*
537 * Get next file to service by choosing one at random
538 */
539static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
540					   int badf)
541{
542	struct fio_file *f;
543	int fno;
544
545	do {
546		long r = os_random_long(&td->next_file_state);
547
548		fno = (unsigned int) ((double) td->o.nr_files
549			* (r / (RAND_MAX + 1.0)));
550		f = td->files[fno];
551		if (f->flags & FIO_FILE_DONE)
552			continue;
553
554		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
555			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
556			return f;
557		}
558	} while (1);
559}
560
561/*
562 * Get next file to service by doing round robin between all available ones
563 */
564static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
565					 int badf)
566{
567	unsigned int old_next_file = td->next_file;
568	struct fio_file *f;
569
570	do {
571		f = td->files[td->next_file];
572
573		td->next_file++;
574		if (td->next_file >= td->o.nr_files)
575			td->next_file = 0;
576
577		if (f->flags & FIO_FILE_DONE) {
578			f = NULL;
579			continue;
580		}
581
582		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
583			break;
584
585		f = NULL;
586	} while (td->next_file != old_next_file);
587
588	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
589	return f;
590}
591
592static struct fio_file *get_next_file(struct thread_data *td)
593{
594	struct fio_file *f;
595
596	assert(td->o.nr_files <= td->files_index);
597
598	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
599		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
600				" nr_files=%d\n", td->nr_open_files,
601						  td->nr_done_files,
602						  td->o.nr_files);
603		return NULL;
604	}
605
606	f = td->file_service_file;
607	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
608		goto out;
609
610	if (td->o.file_service_type == FIO_FSERVICE_RR)
611		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
612	else
613		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
614
615	td->file_service_file = f;
616	td->file_service_left = td->file_service_nr - 1;
617out:
618	dprint(FD_FILE, "get_next_file: %p\n", f);
619	return f;
620}
621
622static struct fio_file *find_next_new_file(struct thread_data *td)
623{
624	struct fio_file *f;
625
626	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
627		return NULL;
628
629	if (td->o.file_service_type == FIO_FSERVICE_RR)
630		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
631	else
632		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
633
634	return f;
635}
636
637static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
638{
639	struct fio_file *f;
640
641	do {
642		f = get_next_file(td);
643		if (!f)
644			return 1;
645
646set_file:
647		io_u->file = f;
648		get_file(f);
649
650		if (!fill_io_u(td, io_u))
651			break;
652
653		/*
654		 * td_io_close() does a put_file() as well, so no need to
655		 * do that here.
656		 */
657		io_u->file = NULL;
658		td_io_close_file(td, f);
659		f->flags |= FIO_FILE_DONE;
660		td->nr_done_files++;
661
662		/*
663		 * probably not the right place to do this, but see
664		 * if we need to open a new file
665		 */
666		if (td->nr_open_files < td->o.open_files &&
667		    td->o.open_files != td->o.nr_files) {
668			f = find_next_new_file(td);
669
670			if (!f || td_io_open_file(td, f))
671				return 1;
672
673			goto set_file;
674		}
675	} while (1);
676
677	return 0;
678}
679
680
681struct io_u *__get_io_u(struct thread_data *td)
682{
683	struct io_u *io_u = NULL;
684
685	if (!list_empty(&td->io_u_requeues))
686		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
687	else if (!queue_full(td)) {
688		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
689
690		io_u->buflen = 0;
691		io_u->resid = 0;
692		io_u->file = NULL;
693		io_u->end_io = NULL;
694	}
695
696	if (io_u) {
697		assert(io_u->flags & IO_U_F_FREE);
698		io_u->flags &= ~IO_U_F_FREE;
699
700		io_u->error = 0;
701		list_del(&io_u->list);
702		list_add(&io_u->list, &td->io_u_busylist);
703		td->cur_depth++;
704	}
705
706	return io_u;
707}
708
709/*
710 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
711 * etc. The returned io_u is fully ready to be prepped and submitted.
712 */
713struct io_u *get_io_u(struct thread_data *td)
714{
715	struct fio_file *f;
716	struct io_u *io_u;
717
718	io_u = __get_io_u(td);
719	if (!io_u) {
720		dprint(FD_IO, "__get_io_u failed\n");
721		return NULL;
722	}
723
724	/*
725	 * from a requeue, io_u already setup
726	 */
727	if (io_u->file)
728		goto out;
729
730	/*
731	 * If using an iolog, grab next piece if any available.
732	 */
733	if (td->o.read_iolog_file) {
734		if (read_iolog_get(td, io_u))
735			goto err_put;
736	} else if (set_io_u_file(td, io_u)) {
737		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
738		goto err_put;
739	}
740
741	f = io_u->file;
742	assert(f->flags & FIO_FILE_OPEN);
743
744	if (io_u->ddir != DDIR_SYNC) {
745		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
746			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
747			goto err_put;
748		}
749
750		f->last_pos = io_u->offset + io_u->buflen;
751
752		if (td->o.verify != VERIFY_NONE)
753			populate_verify_io_u(td, io_u);
754	}
755
756	/*
757	 * Set io data pointers.
758	 */
759	io_u->endpos = io_u->offset + io_u->buflen;
760	io_u->xfer_buf = io_u->buf;
761	io_u->xfer_buflen = io_u->buflen;
762out:
763	if (!td_io_prep(td, io_u)) {
764		fio_gettime(&io_u->start_time, NULL);
765		return io_u;
766	}
767err_put:
768	dprint(FD_IO, "get_io_u failed\n");
769	put_io_u(td, io_u);
770	return NULL;
771}
772
773void io_u_log_error(struct thread_data *td, struct io_u *io_u)
774{
775	const char *msg[] = { "read", "write", "sync" };
776
777	log_err("fio: io_u error");
778
779	if (io_u->file)
780		log_err(" on file %s", io_u->file->file_name);
781
782	log_err(": %s\n", strerror(io_u->error));
783
784	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
785					io_u->offset, io_u->xfer_buflen);
786
787	if (!td->error)
788		td_verror(td, io_u->error, "io_u error");
789}
790
791static void io_completed(struct thread_data *td, struct io_u *io_u,
792			 struct io_completion_data *icd)
793{
794	unsigned long usec;
795
796	dprint_io_u(io_u, "io complete");
797
798	assert(io_u->flags & IO_U_F_FLIGHT);
799	io_u->flags &= ~IO_U_F_FLIGHT;
800
801	if (io_u->ddir == DDIR_SYNC) {
802		td->last_was_sync = 1;
803		return;
804	}
805
806	td->last_was_sync = 0;
807
808	if (!io_u->error) {
809		unsigned int bytes = io_u->buflen - io_u->resid;
810		const enum fio_ddir idx = io_u->ddir;
811		int ret;
812
813		td->io_blocks[idx]++;
814		td->io_bytes[idx] += bytes;
815		td->this_io_bytes[idx] += bytes;
816
817		usec = utime_since(&io_u->issue_time, &icd->time);
818
819		add_clat_sample(td, idx, usec);
820		add_bw_sample(td, idx, &icd->time);
821		io_u_mark_latency(td, usec);
822
823		if (td_write(td) && idx == DDIR_WRITE &&
824		    td->o.do_verify &&
825		    td->o.verify != VERIFY_NONE)
826			log_io_piece(td, io_u);
827
828		icd->bytes_done[idx] += bytes;
829
830		if (io_u->end_io) {
831			ret = io_u->end_io(td, io_u);
832			if (ret && !icd->error)
833				icd->error = ret;
834		}
835	} else {
836		icd->error = io_u->error;
837		io_u_log_error(td, io_u);
838	}
839}
840
841static void init_icd(struct io_completion_data *icd, int nr)
842{
843	fio_gettime(&icd->time, NULL);
844
845	icd->nr = nr;
846
847	icd->error = 0;
848	icd->bytes_done[0] = icd->bytes_done[1] = 0;
849}
850
851static void ios_completed(struct thread_data *td,
852			  struct io_completion_data *icd)
853{
854	struct io_u *io_u;
855	int i;
856
857	for (i = 0; i < icd->nr; i++) {
858		io_u = td->io_ops->event(td, i);
859
860		io_completed(td, io_u, icd);
861		put_io_u(td, io_u);
862	}
863}
864
865/*
866 * Complete a single io_u for the sync engines.
867 */
868long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
869{
870	struct io_completion_data icd;
871
872	init_icd(&icd, 1);
873	io_completed(td, io_u, &icd);
874	put_io_u(td, io_u);
875
876	if (!icd.error)
877		return icd.bytes_done[0] + icd.bytes_done[1];
878
879	td_verror(td, icd.error, "io_u_sync_complete");
880	return -1;
881}
882
883/*
884 * Called to complete min_events number of io for the async engines.
885 */
886long io_u_queued_complete(struct thread_data *td, int min_events)
887{
888	struct io_completion_data icd;
889	struct timespec *tvp = NULL;
890	int ret;
891	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
892
893	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
894
895	if (!min_events)
896		tvp = &ts;
897
898	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
899	if (ret < 0) {
900		td_verror(td, -ret, "td_io_getevents");
901		return ret;
902	} else if (!ret)
903		return ret;
904
905	init_icd(&icd, ret);
906	ios_completed(td, &icd);
907	if (!icd.error)
908		return icd.bytes_done[0] + icd.bytes_done[1];
909
910	td_verror(td, icd.error, "io_u_queued_complete");
911	return -1;
912}
913
914/*
915 * Call when io_u is really queued, to update the submission latency.
916 */
917void io_u_queued(struct thread_data *td, struct io_u *io_u)
918{
919	unsigned long slat_time;
920
921	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
922	add_slat_sample(td, io_u->ddir, slat_time);
923}
924
925#ifdef FIO_USE_TIMEOUT
926void io_u_set_timeout(struct thread_data *td)
927{
928	assert(td->cur_depth);
929
930	td->timer.it_interval.tv_sec = 0;
931	td->timer.it_interval.tv_usec = 0;
932	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
933	td->timer.it_value.tv_usec = 0;
934	setitimer(ITIMER_REAL, &td->timer, NULL);
935	fio_gettime(&td->timeout_end, NULL);
936}
937
938static void io_u_dump(struct io_u *io_u)
939{
940	unsigned long t_start = mtime_since_now(&io_u->start_time);
941	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
942
943	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
944	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
945						io_u->xfer_buf, io_u->buflen,
946						io_u->xfer_buflen,
947						io_u->offset);
948	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
949}
950#else
951void io_u_set_timeout(struct thread_data fio_unused *td)
952{
953}
954#endif
955
956#ifdef FIO_USE_TIMEOUT
957static void io_u_timeout_handler(int fio_unused sig)
958{
959	struct thread_data *td, *__td;
960	pid_t pid = getpid();
961	struct list_head *entry;
962	struct io_u *io_u;
963	int i;
964
965	log_err("fio: io_u timeout\n");
966
967	/*
968	 * TLS would be nice...
969	 */
970	td = NULL;
971	for_each_td(__td, i) {
972		if (__td->pid == pid) {
973			td = __td;
974			break;
975		}
976	}
977
978	if (!td) {
979		log_err("fio: io_u timeout, can't find job\n");
980		exit(1);
981	}
982
983	if (!td->cur_depth) {
984		log_err("fio: timeout without pending work?\n");
985		return;
986	}
987
988	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
989
990	list_for_each(entry, &td->io_u_busylist) {
991		io_u = list_entry(entry, struct io_u, list);
992
993		io_u_dump(io_u);
994	}
995
996	td_verror(td, ETIMEDOUT, "io_u timeout");
997	exit(1);
998}
999#endif
1000
1001void io_u_init_timeout(void)
1002{
1003#ifdef FIO_USE_TIMEOUT
1004	signal(SIGALRM, io_u_timeout_handler);
1005#endif
1006}
1007