io_u.c revision c311cd2a9e3ca77f4d2cb2f6dd8bd08185e8c0f6
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#include "hash.h"
10
11struct io_completion_data {
12	int nr;				/* input */
13
14	int error;			/* output */
15	unsigned long bytes_done[2];	/* output */
16	struct timeval time;		/* output */
17};
18
19/*
20 * The ->file_map[] contains a map of blocks we have or have not done io
21 * to yet. Used to make sure we cover the entire range in a fair fashion.
22 */
23static int random_map_free(struct fio_file *f, const unsigned long long block)
24{
25	unsigned int idx = RAND_MAP_IDX(f, block);
26	unsigned int bit = RAND_MAP_BIT(f, block);
27
28	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
29
30	return (f->file_map[idx] & (1 << bit)) == 0;
31}
32
33/*
34 * Mark a given offset as used in the map.
35 */
36static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37{
38	unsigned int min_bs = td->o.rw_min_bs;
39	struct fio_file *f = io_u->file;
40	unsigned long long block;
41	unsigned int blocks, nr_blocks;
42
43	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
44	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45	blocks = 0;
46
47	while (nr_blocks) {
48		unsigned int this_blocks, mask;
49		unsigned int idx, bit;
50
51		/*
52		 * If we have a mixed random workload, we may
53		 * encounter blocks we already did IO to.
54		 */
55		if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
56			if (!blocks)
57				blocks = 1;
58			break;
59		}
60
61		idx = RAND_MAP_IDX(f, block);
62		bit = RAND_MAP_BIT(f, block);
63
64		fio_assert(td, idx < f->num_maps);
65
66		this_blocks = nr_blocks;
67		if (this_blocks + bit > BLOCKS_PER_MAP)
68			this_blocks = BLOCKS_PER_MAP - bit;
69
70		if (this_blocks == BLOCKS_PER_MAP)
71			mask = -1U;
72		else
73			mask = ((1U << this_blocks) - 1) << bit;
74
75		f->file_map[idx] |= mask;
76		nr_blocks -= this_blocks;
77		blocks += this_blocks;
78		block += this_blocks;
79	}
80
81	if ((blocks * min_bs) < io_u->buflen)
82		io_u->buflen = blocks * min_bs;
83}
84
85static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
86				     enum fio_ddir ddir)
87{
88	unsigned long long max_blocks;
89	unsigned long long max_size;
90
91	/*
92	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
93	 */
94	max_size = f->io_size;
95	if (max_size > f->real_file_size)
96		max_size = f->real_file_size;
97
98	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
99	if (!max_blocks)
100		return 0;
101
102	return max_blocks;
103}
104
105/*
106 * Return the next free block in the map.
107 */
108static int get_next_free_block(struct thread_data *td, struct fio_file *f,
109			       enum fio_ddir ddir, unsigned long long *b)
110{
111	unsigned long long min_bs = td->o.rw_min_bs;
112	int i;
113
114	i = f->last_free_lookup;
115	*b = (i * BLOCKS_PER_MAP);
116	while ((*b) * min_bs < f->real_file_size &&
117		(*b) * min_bs < f->io_size) {
118		if (f->file_map[i] != (unsigned int) -1) {
119			*b += ffz(f->file_map[i]);
120			if (*b > last_block(td, f, ddir))
121				break;
122			f->last_free_lookup = i;
123			return 0;
124		}
125
126		*b += BLOCKS_PER_MAP;
127		i++;
128	}
129
130	dprint(FD_IO, "failed finding a free block\n");
131	return 1;
132}
133
134static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
135				enum fio_ddir ddir, unsigned long long *b)
136{
137	unsigned long long r;
138	int loops = 5;
139
140	do {
141		r = os_random_long(&td->random_state);
142		dprint(FD_RANDOM, "off rand %llu\n", r);
143		*b = (last_block(td, f, ddir) - 1)
144			* (r / ((unsigned long long) OS_RAND_MAX + 1.0));
145
146		/*
147		 * if we are not maintaining a random map, we are done.
148		 */
149		if (!file_randommap(td, f))
150			return 0;
151
152		/*
153		 * calculate map offset and check if it's free
154		 */
155		if (random_map_free(f, *b))
156			return 0;
157
158		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
159									*b);
160	} while (--loops);
161
162	/*
163	 * we get here, if we didn't suceed in looking up a block. generate
164	 * a random start offset into the filemap, and find the first free
165	 * block from there.
166	 */
167	loops = 10;
168	do {
169		f->last_free_lookup = (f->num_maps - 1) *
170					(r / (OS_RAND_MAX + 1.0));
171		if (!get_next_free_block(td, f, ddir, b))
172			return 0;
173
174		r = os_random_long(&td->random_state);
175	} while (--loops);
176
177	/*
178	 * that didn't work either, try exhaustive search from the start
179	 */
180	f->last_free_lookup = 0;
181	return get_next_free_block(td, f, ddir, b);
182}
183
184/*
185 * For random io, generate a random new block and see if it's used. Repeat
186 * until we find a free one. For sequential io, just return the end of
187 * the last io issued.
188 */
189static int get_next_offset(struct thread_data *td, struct io_u *io_u)
190{
191	struct fio_file *f = io_u->file;
192	unsigned long long b;
193	enum fio_ddir ddir = io_u->ddir;
194
195	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
196		td->ddir_nr = td->o.ddir_nr;
197
198		if (get_next_rand_offset(td, f, ddir, &b)) {
199			dprint(FD_IO, "%s: getting rand offset failed\n",
200				f->file_name);
201			return 1;
202		}
203	} else {
204		if (f->last_pos >= f->real_file_size) {
205			if (!td_random(td) ||
206			     get_next_rand_offset(td, f, ddir, &b)) {
207				dprint(FD_IO, "%s: pos %llu > size %llu\n",
208						f->file_name, f->last_pos,
209						f->real_file_size);
210				return 1;
211			}
212		} else
213			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
214	}
215
216	io_u->offset = b * td->o.ba[ddir];
217	if (io_u->offset >= f->io_size) {
218		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
219					io_u->offset, f->io_size);
220		return 1;
221	}
222
223	io_u->offset += f->file_offset;
224	if (io_u->offset >= f->real_file_size) {
225		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
226					io_u->offset, f->real_file_size);
227		return 1;
228	}
229
230	return 0;
231}
232
233static inline int is_power_of_2(unsigned int val)
234{
235	return (val != 0 && ((val & (val - 1)) == 0));
236}
237
238static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
239{
240	const int ddir = io_u->ddir;
241	unsigned int uninitialized_var(buflen);
242	unsigned int minbs, maxbs;
243	long r;
244
245	minbs = td->o.min_bs[ddir];
246	maxbs = td->o.max_bs[ddir];
247
248	if (minbs == maxbs)
249		buflen = minbs;
250	else {
251		r = os_random_long(&td->bsrange_state);
252		if (!td->o.bssplit_nr[ddir]) {
253			buflen = 1 + (unsigned int) ((double) maxbs *
254					(r / (OS_RAND_MAX + 1.0)));
255			if (buflen < minbs)
256				buflen = minbs;
257		} else {
258			long perc = 0;
259			unsigned int i;
260
261			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
262				struct bssplit *bsp = &td->o.bssplit[ddir][i];
263
264				buflen = bsp->bs;
265				perc += bsp->perc;
266				if (r <= ((OS_RAND_MAX / 100L) * perc))
267					break;
268			}
269		}
270		if (!td->o.bs_unaligned && is_power_of_2(minbs))
271			buflen = (buflen + minbs - 1) & ~(minbs - 1);
272	}
273
274	if (io_u->offset + buflen > io_u->file->real_file_size) {
275		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
276						minbs, ddir);
277		buflen = minbs;
278	}
279
280	return buflen;
281}
282
283static void set_rwmix_bytes(struct thread_data *td)
284{
285	unsigned int diff;
286
287	/*
288	 * we do time or byte based switch. this is needed because
289	 * buffered writes may issue a lot quicker than they complete,
290	 * whereas reads do not.
291	 */
292	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
293	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
294}
295
296static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
297{
298	unsigned int v;
299	long r;
300
301	r = os_random_long(&td->rwmix_state);
302	v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
303	if (v <= td->o.rwmix[DDIR_READ])
304		return DDIR_READ;
305
306	return DDIR_WRITE;
307}
308
309/*
310 * Return the data direction for the next io_u. If the job is a
311 * mixed read/write workload, check the rwmix cycle and switch if
312 * necessary.
313 */
314static enum fio_ddir get_rw_ddir(struct thread_data *td)
315{
316	if (td_rw(td)) {
317		/*
318		 * Check if it's time to seed a new data direction.
319		 */
320		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
321			unsigned long long max_bytes;
322			enum fio_ddir ddir;
323
324			/*
325			 * Put a top limit on how many bytes we do for
326			 * one data direction, to avoid overflowing the
327			 * ranges too much
328			 */
329			ddir = get_rand_ddir(td);
330			max_bytes = td->this_io_bytes[ddir];
331			if (max_bytes >=
332			    (td->o.size * td->o.rwmix[ddir] / 100)) {
333				if (!td->rw_end_set[ddir]) {
334					td->rw_end_set[ddir] = 1;
335					fio_gettime(&td->rw_end[ddir], NULL);
336				}
337
338				ddir ^= 1;
339			}
340
341			if (ddir != td->rwmix_ddir)
342				set_rwmix_bytes(td);
343
344			td->rwmix_ddir = ddir;
345		}
346		return td->rwmix_ddir;
347	} else if (td_read(td))
348		return DDIR_READ;
349	else
350		return DDIR_WRITE;
351}
352
353static void put_file_log(struct thread_data *td, struct fio_file *f)
354{
355	int ret = put_file(td, f);
356
357	if (ret)
358		td_verror(td, ret, "file close");
359}
360
361void put_io_u(struct thread_data *td, struct io_u *io_u)
362{
363	assert((io_u->flags & IO_U_F_FREE) == 0);
364	io_u->flags |= IO_U_F_FREE;
365
366	if (io_u->file)
367		put_file_log(td, io_u->file);
368
369	io_u->file = NULL;
370	flist_del(&io_u->list);
371	flist_add(&io_u->list, &td->io_u_freelist);
372	td->cur_depth--;
373}
374
375void requeue_io_u(struct thread_data *td, struct io_u **io_u)
376{
377	struct io_u *__io_u = *io_u;
378
379	dprint(FD_IO, "requeue %p\n", __io_u);
380
381	__io_u->flags |= IO_U_F_FREE;
382	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
383		td->io_issues[__io_u->ddir]--;
384
385	__io_u->flags &= ~IO_U_F_FLIGHT;
386
387	flist_del(&__io_u->list);
388	flist_add_tail(&__io_u->list, &td->io_u_requeues);
389	td->cur_depth--;
390	*io_u = NULL;
391}
392
393static int fill_io_u(struct thread_data *td, struct io_u *io_u)
394{
395	if (td->io_ops->flags & FIO_NOIO)
396		goto out;
397
398	/*
399	 * see if it's time to sync
400	 */
401	if (td->o.fsync_blocks &&
402	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
403	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
404		io_u->ddir = DDIR_SYNC;
405		goto out;
406	}
407
408	io_u->ddir = get_rw_ddir(td);
409
410	/*
411	 * See if it's time to switch to a new zone
412	 */
413	if (td->zone_bytes >= td->o.zone_size) {
414		td->zone_bytes = 0;
415		io_u->file->last_pos += td->o.zone_skip;
416		td->io_skip_bytes += td->o.zone_skip;
417	}
418
419	/*
420	 * No log, let the seq/rand engine retrieve the next buflen and
421	 * position.
422	 */
423	if (get_next_offset(td, io_u)) {
424		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
425		return 1;
426	}
427
428	io_u->buflen = get_next_buflen(td, io_u);
429	if (!io_u->buflen) {
430		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
431		return 1;
432	}
433
434	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
435		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
436		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
437				io_u->buflen, io_u->file->real_file_size);
438		return 1;
439	}
440
441	/*
442	 * mark entry before potentially trimming io_u
443	 */
444	if (td_random(td) && file_randommap(td, io_u->file))
445		mark_random_map(td, io_u);
446
447	/*
448	 * If using a write iolog, store this entry.
449	 */
450out:
451	dprint_io_u(io_u, "fill_io_u");
452	td->zone_bytes += io_u->buflen;
453	log_io_u(td, io_u);
454	return 0;
455}
456
457static void __io_u_mark_map(unsigned int *map, unsigned int nr)
458{
459	int index = 0;
460
461	switch (nr) {
462	default:
463		index = 6;
464		break;
465	case 33 ... 64:
466		index = 5;
467		break;
468	case 17 ... 32:
469		index = 4;
470		break;
471	case 9 ... 16:
472		index = 3;
473		break;
474	case 5 ... 8:
475		index = 2;
476		break;
477	case 1 ... 4:
478		index = 1;
479	case 0:
480		break;
481	}
482
483	map[index]++;
484}
485
486void io_u_mark_submit(struct thread_data *td, unsigned int nr)
487{
488	__io_u_mark_map(td->ts.io_u_submit, nr);
489	td->ts.total_submit++;
490}
491
492void io_u_mark_complete(struct thread_data *td, unsigned int nr)
493{
494	__io_u_mark_map(td->ts.io_u_complete, nr);
495	td->ts.total_complete++;
496}
497
498void io_u_mark_depth(struct thread_data *td, unsigned int nr)
499{
500	int index = 0;
501
502	switch (td->cur_depth) {
503	default:
504		index = 6;
505		break;
506	case 32 ... 63:
507		index = 5;
508		break;
509	case 16 ... 31:
510		index = 4;
511		break;
512	case 8 ... 15:
513		index = 3;
514		break;
515	case 4 ... 7:
516		index = 2;
517		break;
518	case 2 ... 3:
519		index = 1;
520	case 1:
521		break;
522	}
523
524	td->ts.io_u_map[index] += nr;
525}
526
527static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
528{
529	int index = 0;
530
531	assert(usec < 1000);
532
533	switch (usec) {
534	case 750 ... 999:
535		index = 9;
536		break;
537	case 500 ... 749:
538		index = 8;
539		break;
540	case 250 ... 499:
541		index = 7;
542		break;
543	case 100 ... 249:
544		index = 6;
545		break;
546	case 50 ... 99:
547		index = 5;
548		break;
549	case 20 ... 49:
550		index = 4;
551		break;
552	case 10 ... 19:
553		index = 3;
554		break;
555	case 4 ... 9:
556		index = 2;
557		break;
558	case 2 ... 3:
559		index = 1;
560	case 0 ... 1:
561		break;
562	}
563
564	assert(index < FIO_IO_U_LAT_U_NR);
565	td->ts.io_u_lat_u[index]++;
566}
567
568static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
569{
570	int index = 0;
571
572	switch (msec) {
573	default:
574		index = 11;
575		break;
576	case 1000 ... 1999:
577		index = 10;
578		break;
579	case 750 ... 999:
580		index = 9;
581		break;
582	case 500 ... 749:
583		index = 8;
584		break;
585	case 250 ... 499:
586		index = 7;
587		break;
588	case 100 ... 249:
589		index = 6;
590		break;
591	case 50 ... 99:
592		index = 5;
593		break;
594	case 20 ... 49:
595		index = 4;
596		break;
597	case 10 ... 19:
598		index = 3;
599		break;
600	case 4 ... 9:
601		index = 2;
602		break;
603	case 2 ... 3:
604		index = 1;
605	case 0 ... 1:
606		break;
607	}
608
609	assert(index < FIO_IO_U_LAT_M_NR);
610	td->ts.io_u_lat_m[index]++;
611}
612
613static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
614{
615	if (usec < 1000)
616		io_u_mark_lat_usec(td, usec);
617	else
618		io_u_mark_lat_msec(td, usec / 1000);
619}
620
621/*
622 * Get next file to service by choosing one at random
623 */
624static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
625					   int badf)
626{
627	struct fio_file *f;
628	int fno;
629
630	do {
631		long r = os_random_long(&td->next_file_state);
632		int opened = 0;
633
634		fno = (unsigned int) ((double) td->o.nr_files
635			* (r / (OS_RAND_MAX + 1.0)));
636		f = td->files[fno];
637		if (f->flags & FIO_FILE_DONE)
638			continue;
639
640		if (!(f->flags & FIO_FILE_OPEN)) {
641			int err;
642
643			err = td_io_open_file(td, f);
644			if (err)
645				continue;
646			opened = 1;
647		}
648
649		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
650			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
651			return f;
652		}
653		if (opened)
654			td_io_close_file(td, f);
655	} while (1);
656}
657
658/*
659 * Get next file to service by doing round robin between all available ones
660 */
661static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
662					 int badf)
663{
664	unsigned int old_next_file = td->next_file;
665	struct fio_file *f;
666
667	do {
668		int opened = 0;
669
670		f = td->files[td->next_file];
671
672		td->next_file++;
673		if (td->next_file >= td->o.nr_files)
674			td->next_file = 0;
675
676		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
677		if (f->flags & FIO_FILE_DONE) {
678			f = NULL;
679			continue;
680		}
681
682		if (!(f->flags & FIO_FILE_OPEN)) {
683			int err;
684
685			err = td_io_open_file(td, f);
686			if (err) {
687				dprint(FD_FILE, "error %d on open of %s\n",
688					err, f->file_name);
689				continue;
690			}
691			opened = 1;
692		}
693
694		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
695		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
696			break;
697
698		if (opened)
699			td_io_close_file(td, f);
700
701		f = NULL;
702	} while (td->next_file != old_next_file);
703
704	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
705	return f;
706}
707
708static struct fio_file *get_next_file(struct thread_data *td)
709{
710	struct fio_file *f;
711
712	assert(td->o.nr_files <= td->files_index);
713
714	if (td->nr_done_files >= td->o.nr_files) {
715		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
716				" nr_files=%d\n", td->nr_open_files,
717						  td->nr_done_files,
718						  td->o.nr_files);
719		return NULL;
720	}
721
722	f = td->file_service_file;
723	if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) {
724		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
725			goto out;
726		if (td->file_service_left--)
727			goto out;
728	}
729
730	if (td->o.file_service_type == FIO_FSERVICE_RR ||
731	    td->o.file_service_type == FIO_FSERVICE_SEQ)
732		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
733	else
734		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
735
736	td->file_service_file = f;
737	td->file_service_left = td->file_service_nr - 1;
738out:
739	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
740	return f;
741}
742
743static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
744{
745	struct fio_file *f;
746
747	do {
748		f = get_next_file(td);
749		if (!f)
750			return 1;
751
752		io_u->file = f;
753		get_file(f);
754
755		if (!fill_io_u(td, io_u))
756			break;
757
758		put_file_log(td, f);
759		td_io_close_file(td, f);
760		io_u->file = NULL;
761		f->flags |= FIO_FILE_DONE;
762		td->nr_done_files++;
763		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
764	} while (1);
765
766	return 0;
767}
768
769
770struct io_u *__get_io_u(struct thread_data *td)
771{
772	struct io_u *io_u = NULL;
773
774	if (!flist_empty(&td->io_u_requeues))
775		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
776	else if (!queue_full(td)) {
777		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
778
779		io_u->buflen = 0;
780		io_u->resid = 0;
781		io_u->file = NULL;
782		io_u->end_io = NULL;
783	}
784
785	if (io_u) {
786		assert(io_u->flags & IO_U_F_FREE);
787		io_u->flags &= ~IO_U_F_FREE;
788
789		io_u->error = 0;
790		flist_del(&io_u->list);
791		flist_add(&io_u->list, &td->io_u_busylist);
792		td->cur_depth++;
793	}
794
795	return io_u;
796}
797
798/*
799 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
800 * etc. The returned io_u is fully ready to be prepped and submitted.
801 */
802struct io_u *get_io_u(struct thread_data *td)
803{
804	struct fio_file *f;
805	struct io_u *io_u;
806
807	io_u = __get_io_u(td);
808	if (!io_u) {
809		dprint(FD_IO, "__get_io_u failed\n");
810		return NULL;
811	}
812
813	/*
814	 * from a requeue, io_u already setup
815	 */
816	if (io_u->file)
817		goto out;
818
819	/*
820	 * If using an iolog, grab next piece if any available.
821	 */
822	if (td->o.read_iolog_file) {
823		if (read_iolog_get(td, io_u))
824			goto err_put;
825	} else if (set_io_u_file(td, io_u)) {
826		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
827		goto err_put;
828	}
829
830	f = io_u->file;
831	assert(f->flags & FIO_FILE_OPEN);
832
833	if (io_u->ddir != DDIR_SYNC) {
834		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
835			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
836			goto err_put;
837		}
838
839		f->last_pos = io_u->offset + io_u->buflen;
840
841		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
842			populate_verify_io_u(td, io_u);
843		else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
844			io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
845	}
846
847	/*
848	 * Set io data pointers.
849	 */
850	io_u->endpos = io_u->offset + io_u->buflen;
851	io_u->xfer_buf = io_u->buf;
852	io_u->xfer_buflen = io_u->buflen;
853
854out:
855	if (!td_io_prep(td, io_u)) {
856		if (!td->o.disable_slat)
857			fio_gettime(&io_u->start_time, NULL);
858		return io_u;
859	}
860err_put:
861	dprint(FD_IO, "get_io_u failed\n");
862	put_io_u(td, io_u);
863	return NULL;
864}
865
866void io_u_log_error(struct thread_data *td, struct io_u *io_u)
867{
868	const char *msg[] = { "read", "write", "sync" };
869
870	log_err("fio: io_u error");
871
872	if (io_u->file)
873		log_err(" on file %s", io_u->file->file_name);
874
875	log_err(": %s\n", strerror(io_u->error));
876
877	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
878					io_u->offset, io_u->xfer_buflen);
879
880	if (!td->error)
881		td_verror(td, io_u->error, "io_u error");
882}
883
884static void io_completed(struct thread_data *td, struct io_u *io_u,
885			 struct io_completion_data *icd)
886{
887	/*
888	 * Older gcc's are too dumb to realize that usec is always used
889	 * initialized, silence that warning.
890	 */
891	unsigned long uninitialized_var(usec);
892
893	dprint_io_u(io_u, "io complete");
894
895	assert(io_u->flags & IO_U_F_FLIGHT);
896	io_u->flags &= ~IO_U_F_FLIGHT;
897
898	if (io_u->ddir == DDIR_SYNC) {
899		td->last_was_sync = 1;
900		return;
901	}
902
903	td->last_was_sync = 0;
904
905	if (!io_u->error) {
906		unsigned int bytes = io_u->buflen - io_u->resid;
907		const enum fio_ddir idx = io_u->ddir;
908		int ret;
909
910		td->io_blocks[idx]++;
911		td->io_bytes[idx] += bytes;
912		td->this_io_bytes[idx] += bytes;
913
914		if (ramp_time_over(td)) {
915			if (!td->o.disable_clat || !td->o.disable_bw)
916				usec = utime_since(&io_u->issue_time,
917							&icd->time);
918
919			if (!td->o.disable_clat) {
920				add_clat_sample(td, idx, usec);
921				io_u_mark_latency(td, usec);
922			}
923			if (!td->o.disable_bw)
924				add_bw_sample(td, idx, &icd->time);
925		}
926
927		if (td_write(td) && idx == DDIR_WRITE &&
928		    td->o.do_verify &&
929		    td->o.verify != VERIFY_NONE)
930			log_io_piece(td, io_u);
931
932		icd->bytes_done[idx] += bytes;
933
934		if (io_u->end_io) {
935			ret = io_u->end_io(td, io_u);
936			if (ret && !icd->error)
937				icd->error = ret;
938		}
939	} else {
940		icd->error = io_u->error;
941		io_u_log_error(td, io_u);
942	}
943}
944
945static void init_icd(struct thread_data *td, struct io_completion_data *icd,
946		     int nr)
947{
948	if (!td->o.disable_clat || !td->o.disable_bw)
949		fio_gettime(&icd->time, NULL);
950
951	icd->nr = nr;
952
953	icd->error = 0;
954	icd->bytes_done[0] = icd->bytes_done[1] = 0;
955}
956
957static void ios_completed(struct thread_data *td,
958			  struct io_completion_data *icd)
959{
960	struct io_u *io_u;
961	int i;
962
963	for (i = 0; i < icd->nr; i++) {
964		io_u = td->io_ops->event(td, i);
965
966		io_completed(td, io_u, icd);
967		put_io_u(td, io_u);
968	}
969}
970
971/*
972 * Complete a single io_u for the sync engines.
973 */
974long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
975{
976	struct io_completion_data icd;
977
978	init_icd(td, &icd, 1);
979	io_completed(td, io_u, &icd);
980	put_io_u(td, io_u);
981
982	if (!icd.error)
983		return icd.bytes_done[0] + icd.bytes_done[1];
984
985	td_verror(td, icd.error, "io_u_sync_complete");
986	return -1;
987}
988
989/*
990 * Called to complete min_events number of io for the async engines.
991 */
992long io_u_queued_complete(struct thread_data *td, int min_evts)
993{
994	struct io_completion_data icd;
995	struct timespec *tvp = NULL;
996	int ret;
997	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
998
999	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1000
1001	if (!min_evts)
1002		tvp = &ts;
1003
1004	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1005	if (ret < 0) {
1006		td_verror(td, -ret, "td_io_getevents");
1007		return ret;
1008	} else if (!ret)
1009		return ret;
1010
1011	init_icd(td, &icd, ret);
1012	ios_completed(td, &icd);
1013	if (!icd.error)
1014		return icd.bytes_done[0] + icd.bytes_done[1];
1015
1016	td_verror(td, icd.error, "io_u_queued_complete");
1017	return -1;
1018}
1019
1020/*
1021 * Call when io_u is really queued, to update the submission latency.
1022 */
1023void io_u_queued(struct thread_data *td, struct io_u *io_u)
1024{
1025	if (!td->o.disable_slat) {
1026		unsigned long slat_time;
1027
1028		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1029		add_slat_sample(td, io_u->ddir, slat_time);
1030	}
1031}
1032
1033/*
1034 * "randomly" fill the buffer contents
1035 */
1036void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1037		      unsigned int max_bs)
1038{
1039	long *ptr = io_u->buf;
1040
1041	if (!td->o.zero_buffers) {
1042		while ((void *) ptr - io_u->buf < max_bs) {
1043			*ptr = rand() * GOLDEN_RATIO_PRIME;
1044			ptr++;
1045		}
1046	} else
1047		memset(ptr, 0, max_bs);
1048}
1049