1#include <stdio.h>
2#include <stdio.h>
3#include <unistd.h>
4#include <inttypes.h>
5#include <math.h>
6#include <assert.h>
7
8#include "../io_ddir.h"
9#include "../flist.h"
10#include "../hash.h"
11#include "../fifo.h"
12#include "../blktrace_api.h"
13#include "../os/os.h"
14#include "../log.h"
15#include "../oslib/linux-dev-lookup.h"
16
17#define TRACE_FIFO_SIZE	8192
18
19static unsigned int rt_threshold = 1000000;
20static unsigned int ios_threshold = 10;
21static unsigned int rate_threshold;
22static unsigned int set_rate;
23static unsigned int max_depth = 256;
24static int output_ascii = 1;
25static char *filename;
26
27static char **add_opts;
28static int n_add_opts;
29
30/*
31 * Collapse defaults
32 */
33static unsigned int collapse_entries = 0;
34static unsigned int depth_diff = 1;
35static unsigned int random_diff = 5;
36
37struct bs {
38	unsigned int bs;
39	unsigned int nr;
40	int merges;
41};
42
43struct trace_file {
44	char *name;
45	int major, minor;
46};
47
48struct btrace_out {
49	unsigned long ios[DDIR_RWDIR_CNT];
50	unsigned long merges[DDIR_RWDIR_CNT];
51
52	uint64_t last_end[DDIR_RWDIR_CNT];
53	uint64_t seq[DDIR_RWDIR_CNT];
54
55	struct bs *bs[DDIR_RWDIR_CNT];
56	unsigned int nr_bs[DDIR_RWDIR_CNT];
57
58	int inflight;
59	unsigned int depth;
60	int depth_disabled;
61	int complete_seen;
62
63	uint64_t first_ttime[DDIR_RWDIR_CNT];
64	uint64_t last_ttime[DDIR_RWDIR_CNT];
65	uint64_t kib[DDIR_RWDIR_CNT];
66
67	uint64_t start_delay;
68};
69
70struct btrace_pid {
71	struct flist_head hash_list;
72	struct flist_head pid_list;
73	pid_t pid;
74
75	pid_t *merge_pids;
76	unsigned int nr_merge_pids;
77
78	struct trace_file *files;
79	int nr_files;
80	unsigned int last_major, last_minor;
81	int numjobs;
82	int ignore;
83
84	struct btrace_out o;
85};
86
87struct inflight {
88	struct flist_head list;
89	struct btrace_pid *p;
90	uint64_t end_sector;
91};
92
93#define PID_HASH_BITS	10
94#define PID_HASH_SIZE	(1U << PID_HASH_BITS)
95
96static struct flist_head pid_hash[PID_HASH_SIZE];
97static FLIST_HEAD(pid_list);
98
99#define INFLIGHT_HASH_BITS	8
100#define INFLIGHT_HASH_SIZE	(1U << INFLIGHT_HASH_BITS)
101static struct flist_head inflight_hash[INFLIGHT_HASH_SIZE];
102
103static uint64_t first_ttime = -1ULL;
104
105static struct inflight *inflight_find(uint64_t sector)
106{
107	struct flist_head *inflight_list;
108	struct flist_head *e;
109
110	inflight_list = &inflight_hash[hash_long(sector, INFLIGHT_HASH_BITS)];
111
112	flist_for_each(e, inflight_list) {
113		struct inflight *i = flist_entry(e, struct inflight, list);
114
115		if (i->end_sector == sector)
116			return i;
117	}
118
119	return NULL;
120}
121
122static void inflight_remove(struct inflight *i)
123{
124	struct btrace_out *o = &i->p->o;
125
126	o->inflight--;
127	assert(o->inflight >= 0);
128	flist_del(&i->list);
129	free(i);
130}
131
132static void __inflight_add(struct inflight *i)
133{
134	struct flist_head *list;
135
136	list = &inflight_hash[hash_long(i->end_sector, INFLIGHT_HASH_BITS)];
137	flist_add_tail(&i->list, list);
138}
139
140static void inflight_add(struct btrace_pid *p, uint64_t sector, uint32_t len)
141{
142	struct btrace_out *o = &p->o;
143	struct inflight *i;
144
145	i = calloc(1, sizeof(*i));
146	i->p = p;
147	o->inflight++;
148	if (!o->depth_disabled) {
149		o->depth = max((int) o->depth, o->inflight);
150		if (o->depth >= max_depth && !o->complete_seen) {
151			o->depth_disabled = 1;
152			o->depth = max_depth;
153		}
154	}
155	i->end_sector = sector + (len >> 9);
156	__inflight_add(i);
157}
158
159static void inflight_merge(struct inflight *i, int rw, unsigned int size)
160{
161	i->p->o.merges[rw]++;
162	if (size) {
163		i->end_sector += (size >> 9);
164		flist_del(&i->list);
165		__inflight_add(i);
166	}
167}
168
169/*
170 * fifo refill frontend, to avoid reading data in trace sized bites
171 */
172static int refill_fifo(struct fifo *fifo, int fd)
173{
174	char buf[TRACE_FIFO_SIZE];
175	unsigned int total;
176	int ret;
177
178	total = sizeof(buf);
179	if (total > fifo_room(fifo))
180		total = fifo_room(fifo);
181
182	ret = read(fd, buf, total);
183	if (ret < 0) {
184		perror("read refill");
185		return -1;
186	}
187
188	if (ret > 0)
189		ret = fifo_put(fifo, buf, ret);
190
191	return ret;
192}
193
194/*
195 * Retrieve 'len' bytes from the fifo, refilling if necessary.
196 */
197static int trace_fifo_get(struct fifo *fifo, int fd, void *buf,
198			  unsigned int len)
199{
200	if (fifo_len(fifo) < len) {
201		int ret = refill_fifo(fifo, fd);
202
203		if (ret < 0)
204			return ret;
205	}
206
207	return fifo_get(fifo, buf, len);
208}
209
210/*
211 * Just discard the pdu by seeking past it.
212 */
213static int discard_pdu(struct fifo *fifo, int fd, struct blk_io_trace *t)
214{
215	if (t->pdu_len == 0)
216		return 0;
217
218	return trace_fifo_get(fifo, fd, NULL, t->pdu_len);
219}
220
221static int handle_trace_notify(struct blk_io_trace *t)
222{
223	switch (t->action) {
224	case BLK_TN_PROCESS:
225		//printf("got process notify: %x, %d\n", t->action, t->pid);
226		break;
227	case BLK_TN_TIMESTAMP:
228		//printf("got timestamp notify: %x, %d\n", t->action, t->pid);
229		break;
230	case BLK_TN_MESSAGE:
231		break;
232	default:
233		log_err("unknown trace act %x\n", t->action);
234		return 1;
235	}
236
237	return 0;
238}
239
240static void __add_bs(struct btrace_out *o, unsigned int len, int rw)
241{
242	o->bs[rw] = realloc(o->bs[rw], (o->nr_bs[rw] + 1) * sizeof(struct bs));
243	o->bs[rw][o->nr_bs[rw]].bs = len;
244	o->bs[rw][o->nr_bs[rw]].nr = 1;
245	o->nr_bs[rw]++;
246}
247
248static void add_bs(struct btrace_out *o, unsigned int len, int rw)
249{
250	struct bs *bs = o->bs[rw];
251	int i;
252
253	if (!o->nr_bs[rw]) {
254		__add_bs(o, len, rw);
255		return;
256	}
257
258	for (i = 0; i < o->nr_bs[rw]; i++) {
259		if (bs[i].bs == len) {
260			bs[i].nr++;
261			return;
262		}
263	}
264
265	__add_bs(o, len, rw);
266}
267
268#define FMINORBITS	20
269#define FMINORMASK	((1U << FMINORBITS) - 1)
270#define FMAJOR(dev)	((unsigned int) ((dev) >> FMINORBITS))
271#define FMINOR(dev)	((unsigned int) ((dev) & FMINORMASK))
272
273static int btrace_add_file(struct btrace_pid *p, uint32_t devno)
274{
275	unsigned int maj = FMAJOR(devno);
276	unsigned int min = FMINOR(devno);
277	struct trace_file *f;
278	unsigned int i;
279	char dev[256];
280
281	if (filename)
282		return 0;
283	if (p->last_major == maj && p->last_minor == min)
284		return 0;
285
286	p->last_major = maj;
287	p->last_minor = min;
288
289	/*
290	 * check for this file in our list
291	 */
292	for (i = 0; i < p->nr_files; i++) {
293		f = &p->files[i];
294
295		if (f->major == maj && f->minor == min)
296			return 0;
297	}
298
299	strcpy(dev, "/dev");
300	if (!blktrace_lookup_device(NULL, dev, maj, min)) {
301		log_err("fio: failed to find device %u/%u\n", maj, min);
302		if (!output_ascii) {
303			log_err("fio: use -d to specify device\n");
304			return 1;
305		}
306		return 0;
307	}
308
309	p->files = realloc(p->files, (p->nr_files + 1) * sizeof(*f));
310	f = &p->files[p->nr_files];
311	f->name = strdup(dev);
312	f->major = maj;
313	f->minor = min;
314	p->nr_files++;
315	return 0;
316}
317
318static int t_to_rwdir(struct blk_io_trace *t)
319{
320	if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
321		return DDIR_TRIM;
322
323	return (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
324}
325
326static int handle_trace_discard(struct blk_io_trace *t, struct btrace_pid *p)
327{
328	struct btrace_out *o = &p->o;
329
330	if (btrace_add_file(p, t->device))
331		return 1;
332
333	if (o->first_ttime[2] == -1ULL)
334		o->first_ttime[2] = t->time;
335
336	o->ios[DDIR_TRIM]++;
337	add_bs(o, t->bytes, DDIR_TRIM);
338	return 0;
339}
340
341static int handle_trace_fs(struct blk_io_trace *t, struct btrace_pid *p)
342{
343	struct btrace_out *o = &p->o;
344	int rw;
345
346	if (btrace_add_file(p, t->device))
347		return 1;
348
349	first_ttime = min(first_ttime, (uint64_t) t->time);
350
351	rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
352
353	if (o->first_ttime[rw] == -1ULL)
354		o->first_ttime[rw] = t->time;
355
356	add_bs(o, t->bytes, rw);
357	o->ios[rw]++;
358
359	if (t->sector == o->last_end[rw] || o->last_end[rw] == -1ULL)
360		o->seq[rw]++;
361
362	o->last_end[rw] = t->sector + (t->bytes >> 9);
363	return 0;
364}
365
366static int handle_queue_trace(struct blk_io_trace *t, struct btrace_pid *p)
367{
368	if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
369		return handle_trace_notify(t);
370	else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
371		return handle_trace_discard(t, p);
372	else
373		return handle_trace_fs(t, p);
374}
375
376static int handle_trace(struct blk_io_trace *t, struct btrace_pid *p)
377{
378	unsigned int act = t->action & 0xffff;
379	int ret = 0;
380
381	if (act == __BLK_TA_QUEUE) {
382		inflight_add(p, t->sector, t->bytes);
383		ret = handle_queue_trace(t, p);
384	} else if (act == __BLK_TA_BACKMERGE) {
385		struct inflight *i;
386
387		i = inflight_find(t->sector + (t->bytes >> 9));
388		if (i)
389			inflight_remove(i);
390
391		i = inflight_find(t->sector);
392		if (i)
393			inflight_merge(i, t_to_rwdir(t), t->bytes);
394	} else if (act == __BLK_TA_FRONTMERGE) {
395		struct inflight *i;
396
397		i = inflight_find(t->sector + (t->bytes >> 9));
398		if (i)
399			inflight_remove(i);
400
401		i = inflight_find(t->sector);
402		if (i)
403			inflight_merge(i, t_to_rwdir(t), 0);
404	} else if (act == __BLK_TA_COMPLETE) {
405		struct inflight *i;
406
407		i = inflight_find(t->sector + (t->bytes >> 9));
408		if (i) {
409			i->p->o.kib[t_to_rwdir(t)] += (t->bytes >> 10);
410			i->p->o.complete_seen = 1;
411			inflight_remove(i);
412		}
413	}
414
415	return ret;
416}
417
418static void byteswap_trace(struct blk_io_trace *t)
419{
420	t->magic = fio_swap32(t->magic);
421	t->sequence = fio_swap32(t->sequence);
422	t->time = fio_swap64(t->time);
423	t->sector = fio_swap64(t->sector);
424	t->bytes = fio_swap32(t->bytes);
425	t->action = fio_swap32(t->action);
426	t->pid = fio_swap32(t->pid);
427	t->device = fio_swap32(t->device);
428	t->cpu = fio_swap32(t->cpu);
429	t->error = fio_swap16(t->error);
430	t->pdu_len = fio_swap16(t->pdu_len);
431}
432
433static struct btrace_pid *pid_hash_find(pid_t pid, struct flist_head *list)
434{
435	struct flist_head *e;
436	struct btrace_pid *p;
437
438	flist_for_each(e, list) {
439		p = flist_entry(e, struct btrace_pid, hash_list);
440		if (p->pid == pid)
441			return p;
442	}
443
444	return NULL;
445}
446
447static struct btrace_pid *pid_hash_get(pid_t pid)
448{
449	struct flist_head *hash_list;
450	struct btrace_pid *p;
451
452	hash_list = &pid_hash[hash_long(pid, PID_HASH_BITS)];
453
454	p = pid_hash_find(pid, hash_list);
455	if (!p) {
456		int i;
457
458		p = calloc(1, sizeof(*p));
459
460		for (i = 0; i < DDIR_RWDIR_CNT; i++) {
461			p->o.first_ttime[i] = -1ULL;
462			p->o.last_ttime[i] = -1ULL;
463			p->o.last_end[i] = -1ULL;
464		}
465
466		p->pid = pid;
467		p->numjobs = 1;
468		flist_add_tail(&p->hash_list, hash_list);
469		flist_add_tail(&p->pid_list, &pid_list);
470	}
471
472	return p;
473}
474
475/*
476 * Load a blktrace file by reading all the blk_io_trace entries, and storing
477 * them as io_pieces like the fio text version would do.
478 */
479static int load_blktrace(const char *fname, int need_swap)
480{
481	struct btrace_pid *p;
482	unsigned long traces;
483	struct blk_io_trace t;
484	struct fifo *fifo;
485	int fd, ret = 0;
486
487	fd = open(fname, O_RDONLY);
488	if (fd < 0) {
489		perror("open trace file\n");
490		return 1;
491	}
492
493	fifo = fifo_alloc(TRACE_FIFO_SIZE);
494
495	traces = 0;
496	do {
497		ret = trace_fifo_get(fifo, fd, &t, sizeof(t));
498		if (ret < 0)
499			goto err;
500		else if (!ret)
501			break;
502		else if (ret < (int) sizeof(t)) {
503			log_err("fio: short fifo get\n");
504			break;
505		}
506
507		if (need_swap)
508			byteswap_trace(&t);
509
510		if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
511			log_err("fio: bad magic in blktrace data: %x\n", t.magic);
512			goto err;
513		}
514		if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
515			log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
516			goto err;
517		}
518		ret = discard_pdu(fifo, fd, &t);
519		if (ret < 0) {
520			log_err("blktrace lseek\n");
521			goto err;
522		} else if (t.pdu_len != ret) {
523			log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
524			goto err;
525		}
526
527		p = pid_hash_get(t.pid);
528		ret = handle_trace(&t, p);
529		if (ret)
530			break;
531		p->o.last_ttime[t_to_rwdir(&t)] = t.time;
532		traces++;
533	} while (1);
534
535	fifo_free(fifo);
536	close(fd);
537
538	if (ret)
539		return ret;
540
541	if (output_ascii)
542		printf("Traces loaded: %lu\n", traces);
543
544	return 0;
545err:
546	close(fd);
547	fifo_free(fifo);
548	return 1;
549}
550
551static int bs_cmp(const void *ba, const void *bb)
552{
553	const struct bs *bsa = ba;
554	const struct bs *bsb = bb;
555
556	return bsb->nr - bsa->nr;
557}
558
559static unsigned long o_to_kib_rate(struct btrace_out *o, int rw)
560{
561	uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL;
562	uint64_t val;
563
564	if (!usec)
565		return 0;
566
567	usec /= 1000;
568	if (!usec)
569		return 0;
570
571	val = o->kib[rw] * 1000ULL;
572	return val / usec;
573}
574
575static uint64_t o_first_ttime(struct btrace_out *o)
576{
577	uint64_t first;
578
579	first = min(o->first_ttime[0], o->first_ttime[1]);
580	return min(first, o->first_ttime[2]);
581}
582
583static uint64_t o_longest_ttime(struct btrace_out *o)
584{
585	uint64_t ret = 0;
586	int i;
587
588	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
589		uint64_t diff;
590
591		diff = o->last_ttime[i] - o->first_ttime[i];
592		ret = max(diff, ret);
593	}
594
595	return ret;
596}
597
598static void __output_p_ascii(struct btrace_pid *p, unsigned long *ios)
599{
600	const char *msg[] = { "reads", "writes", "trims" };
601	struct btrace_out *o = &p->o;
602	unsigned long total, usec;
603	int i, j;
604
605	printf("[pid:\t%u", p->pid);
606	if (p->nr_merge_pids)
607		for (i = 0; i < p->nr_merge_pids; i++)
608			printf(", %u", p->merge_pids[i]);
609	printf("]\n");
610
611	total = ddir_rw_sum(o->ios);
612	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
613		float perc;
614
615		if (!o->ios[i])
616			continue;
617
618		ios[i] += o->ios[i] + o->merges[i];
619		printf("%s\n", msg[i]);
620		perc = ((float) o->ios[i] * 100.0) / (float) total;
621		printf("\tios:    %lu (perc=%3.2f%%)\n", o->ios[i], perc);
622		perc = ((float) o->merges[i] * 100.0) / (float) total;
623		printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc);
624		perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
625		printf("\tseq:    %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc);
626		printf("\trate:   %lu KiB/sec\n", o_to_kib_rate(o, i));
627
628		for (j = 0; j < o->nr_bs[i]; j++) {
629			struct bs *bs = &o->bs[i][j];
630
631			perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
632			printf("\tbs=%u, perc=%3.2f%%\n", bs->bs, perc);
633		}
634	}
635
636	printf("depth:\t%u\n", o->depth);
637	usec = o_longest_ttime(o) / 1000ULL;
638	printf("usec:\t%lu (delay=%llu)\n", usec, (unsigned long long) o->start_delay);
639
640	printf("files:\t");
641	for (i = 0; i < p->nr_files; i++)
642		printf("%s,", p->files[i].name);
643	printf("\n");
644
645	printf("\n");
646}
647
648static int __output_p_fio(struct btrace_pid *p, unsigned long *ios)
649{
650	struct btrace_out *o = &p->o;
651	unsigned long total;
652	unsigned long long time;
653	float perc;
654	int i, j;
655
656	if ((o->ios[0] + o->ios[1]) && o->ios[2]) {
657		log_err("fio: trace has both read/write and trim\n");
658		return 1;
659	}
660	if (!p->nr_files) {
661		log_err("fio: no devices found\n");
662		return 1;
663	}
664
665	printf("[pid%u", p->pid);
666	if (p->nr_merge_pids)
667		for (i = 0; i < p->nr_merge_pids; i++)
668			printf(",pid%u", p->merge_pids[i]);
669	printf("]\n");
670
671	printf("numjobs=%u\n", p->numjobs);
672	printf("direct=1\n");
673	if (o->depth == 1)
674		printf("ioengine=sync\n");
675	else
676		printf("ioengine=libaio\niodepth=%u\n", o->depth);
677
678	if (o->ios[0] && !o->ios[1])
679		printf("rw=randread\n");
680	else if (!o->ios[0] && o->ios[1])
681		printf("rw=randwrite\n");
682	else if (o->ios[2])
683		printf("rw=randtrim\n");
684	else {
685		printf("rw=randrw\n");
686		total = ddir_rw_sum(o->ios);
687		perc = ((float) o->ios[0] * 100.0) / (float) total;
688		printf("rwmixread=%u\n", (int) floor(perc + 0.50));
689	}
690
691	printf("percentage_random=");
692	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
693		if (o->seq[i] && o->ios[i]) {
694			perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
695			if (perc >= 99.0)
696				perc = 100.0;
697		} else
698			perc = 100.0;
699
700		if (i)
701			printf(",");
702		perc = 100.0 - perc;
703		printf("%u", (int) floor(perc + 0.5));
704	}
705	printf("\n");
706
707	printf("filename=");
708	for (i = 0; i < p->nr_files; i++) {
709		if (i)
710			printf(":");
711		printf("%s", p->files[i].name);
712	}
713	printf("\n");
714
715	if (o->start_delay / 1000000ULL)
716		printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
717
718	time = o_longest_ttime(o);
719	time = (time + 1000000000ULL - 1) / 1000000000ULL;
720	printf("runtime=%llus\n", time);
721
722	printf("bssplit=");
723	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
724
725		if (i && o->nr_bs[i - 1] && o->nr_bs[i])
726			printf(",");
727
728		for (j = 0; j < o->nr_bs[i]; j++) {
729			struct bs *bs = &o->bs[i][j];
730
731			perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
732			if (perc < 1.00)
733				continue;
734			if (j)
735				printf(":");
736			if (j + 1 == o->nr_bs[i])
737				printf("%u/", bs->bs);
738			else
739				printf("%u/%u", bs->bs, (int) floor(perc + 0.5));
740		}
741	}
742	printf("\n");
743
744	if (set_rate) {
745		printf("rate=");
746		for (i = 0; i < DDIR_RWDIR_CNT; i++) {
747			unsigned long rate;
748
749			rate = o_to_kib_rate(o, i);
750			if (i)
751				printf(",");
752			if (rate)
753				printf("%luk", rate);
754		}
755		printf("\n");
756	}
757
758	if (n_add_opts)
759		for (i = 0; i < n_add_opts; i++)
760			printf("%s\n", add_opts[i]);
761
762	printf("\n");
763	return 0;
764}
765
766static int __output_p(struct btrace_pid *p, unsigned long *ios)
767{
768	struct btrace_out *o = &p->o;
769	int i, ret = 0;
770
771	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
772		if (o->nr_bs[i] <= 1)
773			continue;
774		qsort(o->bs[i], o->nr_bs[i], sizeof(struct bs), bs_cmp);
775	}
776
777	if (filename) {
778		p->files = malloc(sizeof(struct trace_file));
779		p->nr_files++;
780		p->files[0].name = filename;
781	}
782
783	if (output_ascii)
784		__output_p_ascii(p, ios);
785	else
786		ret = __output_p_fio(p, ios);
787
788	return ret;
789}
790
791static void remove_ddir(struct btrace_out *o, int rw)
792{
793	o->ios[rw] = 0;
794}
795
796static int prune_entry(struct btrace_out *o)
797{
798	unsigned long rate;
799	uint64_t time;
800	int i;
801
802	if (ddir_rw_sum(o->ios) < ios_threshold)
803		return 1;
804
805	time = o_longest_ttime(o) / 1000ULL;
806	if (time < rt_threshold)
807		return 1;
808
809	rate = 0;
810	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
811		unsigned long this_rate;
812
813		this_rate = o_to_kib_rate(o, i);
814		if (this_rate < rate_threshold) {
815			remove_ddir(o, i);
816			this_rate = 0;
817		}
818		rate += this_rate;
819	}
820
821	if (rate < rate_threshold)
822		return 1;
823
824	return 0;
825}
826
827static int entry_cmp(void *priv, struct flist_head *a, struct flist_head *b)
828{
829	struct btrace_pid *pa = flist_entry(a, struct btrace_pid, pid_list);
830	struct btrace_pid *pb = flist_entry(b, struct btrace_pid, pid_list);
831
832	return ddir_rw_sum(pb->o.ios) - ddir_rw_sum(pa->o.ios);
833}
834
835static void free_p(struct btrace_pid *p)
836{
837	struct btrace_out *o = &p->o;
838	int i;
839
840	for (i = 0; i < p->nr_files; i++) {
841		if (p->files[i].name && p->files[i].name != filename)
842			free(p->files[i].name);
843	}
844
845	for (i = 0; i < DDIR_RWDIR_CNT; i++)
846		free(o->bs[i]);
847
848	free(p->files);
849	flist_del(&p->pid_list);
850	flist_del(&p->hash_list);
851	free(p);
852}
853
854static int entries_close(struct btrace_pid *pida, struct btrace_pid *pidb)
855{
856	float perca, percb, fdiff;
857	int i, idiff;
858
859	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
860		if ((pida->o.ios[i] && !pidb->o.ios[i]) ||
861		    (pidb->o.ios[i] && !pida->o.ios[i]))
862			return 0;
863		if (pida->o.ios[i] && pidb->o.ios[i]) {
864			perca = ((float) pida->o.seq[i] * 100.0) / (float) pida->o.ios[i];
865			percb = ((float) pidb->o.seq[i] * 100.0) / (float) pidb->o.ios[i];
866			fdiff = perca - percb;
867			if (fabs(fdiff) > random_diff)
868				return 0;
869		}
870
871		idiff = pida->o.depth - pidb->o.depth;
872		if (abs(idiff) > depth_diff)
873			return 0;
874	}
875
876	return 1;
877}
878
879static void merge_bs(struct bs **bsap, unsigned int *nr_bsap,
880		     struct bs *bsb, unsigned int nr_bsb)
881{
882	struct bs *bsa = *bsap;
883	unsigned int nr_bsa = *nr_bsap;
884	int a, b;
885
886	for (b = 0; b < nr_bsb; b++) {
887		int next, found = 0;
888
889		for (a = 0; a < nr_bsa; a++) {
890			if (bsb[b].bs != bsa[a].bs)
891				continue;
892
893			bsa[a].nr += bsb[b].nr;
894			bsa[a].merges += bsb[b].merges;
895			found = 1;
896			break;
897		}
898
899		if (found)
900			continue;
901
902		next = *nr_bsap;
903		bsa = realloc(bsa, (next + 1) * sizeof(struct bs));
904		bsa[next].bs = bsb[b].bs;
905		bsa[next].nr = bsb[b].nr;
906		(*nr_bsap)++;
907		*bsap = bsa;
908	}
909}
910
911static int merge_entries(struct btrace_pid *pida, struct btrace_pid *pidb)
912{
913	int i;
914
915	if (!entries_close(pida, pidb))
916		return 0;
917
918	pida->nr_merge_pids++;
919	pida->merge_pids = realloc(pida->merge_pids, pida->nr_merge_pids * sizeof(pid_t));
920	pida->merge_pids[pida->nr_merge_pids - 1] = pidb->pid;
921
922	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
923		struct btrace_out *oa = &pida->o;
924		struct btrace_out *ob = &pidb->o;
925
926		oa->ios[i] += ob->ios[i];
927		oa->merges[i] += ob->merges[i];
928		oa->seq[i] += ob->seq[i];
929		oa->kib[i] += ob->kib[i];
930		oa->first_ttime[i] = min(oa->first_ttime[i], ob->first_ttime[i]);
931		oa->last_ttime[i] = max(oa->last_ttime[i], ob->last_ttime[i]);
932		merge_bs(&oa->bs[i], &oa->nr_bs[i], ob->bs[i], ob->nr_bs[i]);
933	}
934
935	pida->o.start_delay = min(pida->o.start_delay, pidb->o.start_delay);
936	pida->o.depth = (pida->o.depth + pidb->o.depth) / 2;
937	return 1;
938}
939
940static void check_merges(struct btrace_pid *p, struct flist_head *pidlist)
941{
942	struct flist_head *e, *tmp;
943
944	if (p->ignore)
945		return;
946
947	flist_for_each_safe(e, tmp, pidlist) {
948		struct btrace_pid *pidb;
949
950		pidb = flist_entry(e, struct btrace_pid, pid_list);
951		if (pidb == p)
952			continue;
953
954		if (merge_entries(p, pidb)) {
955			pidb->ignore = 1;
956			p->numjobs++;
957		}
958	}
959}
960
961static int output_p(void)
962{
963	unsigned long ios[DDIR_RWDIR_CNT];
964	struct flist_head *e, *tmp;
965	int depth_disabled = 0;
966	int ret = 0;
967
968	flist_for_each_safe(e, tmp, &pid_list) {
969		struct btrace_pid *p;
970
971		p = flist_entry(e, struct btrace_pid, pid_list);
972		if (prune_entry(&p->o)) {
973			free_p(p);
974			continue;
975		}
976		p->o.start_delay = (o_first_ttime(&p->o) / 1000ULL) - first_ttime;
977		depth_disabled += p->o.depth_disabled;
978	}
979
980	if (collapse_entries) {
981		struct btrace_pid *p;
982
983		flist_for_each_safe(e, tmp, &pid_list) {
984			p = flist_entry(e, struct btrace_pid, pid_list);
985			check_merges(p, &pid_list);
986		}
987
988		flist_for_each_safe(e, tmp, &pid_list) {
989			p = flist_entry(e, struct btrace_pid, pid_list);
990			if (p->ignore)
991				free_p(p);
992		}
993	}
994
995	if (depth_disabled)
996		log_err("fio: missing completion traces, depths capped at %u\n", max_depth);
997
998	memset(ios, 0, sizeof(ios));
999
1000	flist_sort(NULL, &pid_list, entry_cmp);
1001
1002	flist_for_each(e, &pid_list) {
1003		struct btrace_pid *p;
1004
1005		p = flist_entry(e, struct btrace_pid, pid_list);
1006		ret |= __output_p(p, ios);
1007		if (ret && !output_ascii)
1008			break;
1009	}
1010
1011	if (output_ascii)
1012		printf("Total: reads=%lu, writes=%lu\n", ios[0], ios[1]);
1013
1014	return ret;
1015}
1016
1017static int usage(char *argv[])
1018{
1019	log_err("%s: [options] <blktrace bin file>\n", argv[0]);
1020	log_err("\t-t\tUsec threshold to ignore task\n");
1021	log_err("\t-n\tNumber IOS threshold to ignore task\n");
1022	log_err("\t-f\tFio job file output\n");
1023	log_err("\t-d\tUse this file/device for replay\n");
1024	log_err("\t-r\tIgnore jobs with less than this KiB/sec rate\n");
1025	log_err("\t-R\tSet rate in fio job (def=%u)\n", set_rate);
1026	log_err("\t-D\tCap queue depth at this value (def=%u)\n", max_depth);
1027	log_err("\t-c\tCollapse \"identical\" jobs (def=%u)\n", collapse_entries);
1028	log_err("\t-u\tDepth difference for collapse (def=%u)\n", depth_diff);
1029	log_err("\t-x\tRandom difference for collapse (def=%u)\n", random_diff);
1030	log_err("\t-a\tAdditional fio option to add to job file\n");
1031	return 1;
1032}
1033
1034static int trace_needs_swap(const char *trace_file, int *swap)
1035{
1036	struct blk_io_trace t;
1037	int fd, ret;
1038
1039	*swap = -1;
1040
1041	fd = open(trace_file, O_RDONLY);
1042	if (fd < 0) {
1043		perror("open");
1044		return 1;
1045	}
1046
1047	ret = read(fd, &t, sizeof(t));
1048	if (ret < 0) {
1049		close(fd);
1050		perror("read");
1051		return 1;
1052	} else if (ret != sizeof(t)) {
1053		close(fd);
1054		log_err("fio: short read on trace file\n");
1055		return 1;
1056	}
1057
1058	close(fd);
1059
1060	if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1061		*swap = 0;
1062	else {
1063		/*
1064		 * Maybe it needs to be endian swapped...
1065		 */
1066		t.magic = fio_swap32(t.magic);
1067		if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1068			*swap = 1;
1069	}
1070
1071	if (*swap == -1) {
1072		log_err("fio: blktrace appears corrupt\n");
1073		return 1;
1074	}
1075
1076	return 0;
1077}
1078
1079int main(int argc, char *argv[])
1080{
1081	int need_swap, i, c;
1082
1083	if (argc < 2)
1084		return usage(argv);
1085
1086	while ((c = getopt(argc, argv, "t:n:fd:r:RD:c:u:x:a:")) != -1) {
1087		switch (c) {
1088		case 'R':
1089			set_rate = 1;
1090			break;
1091		case 'r':
1092			rate_threshold = atoi(optarg);
1093			break;
1094		case 't':
1095			rt_threshold = atoi(optarg);
1096			break;
1097		case 'n':
1098			ios_threshold = atoi(optarg);
1099			break;
1100		case 'f':
1101			output_ascii = 0;
1102			break;
1103		case 'd':
1104			filename = strdup(optarg);
1105			break;
1106		case 'D':
1107			max_depth = atoi(optarg);
1108			break;
1109		case 'c':
1110			collapse_entries = atoi(optarg);
1111			break;
1112		case 'u':
1113			depth_diff = atoi(optarg);
1114			break;
1115		case 'x':
1116			random_diff = atoi(optarg);
1117			break;
1118		case 'a':
1119			add_opts = realloc(add_opts, (n_add_opts + 1) * sizeof(char *));
1120			add_opts[n_add_opts] = strdup(optarg);
1121			n_add_opts++;
1122			break;
1123		case '?':
1124		default:
1125			return usage(argv);
1126		}
1127	}
1128
1129	if (argc == optind)
1130		return usage(argv);
1131
1132	if (trace_needs_swap(argv[optind], &need_swap))
1133		return 1;
1134
1135	for (i = 0; i < PID_HASH_SIZE; i++)
1136		INIT_FLIST_HEAD(&pid_hash[i]);
1137	for (i = 0; i < INFLIGHT_HASH_SIZE; i++)
1138		INIT_FLIST_HEAD(&inflight_hash[i]);
1139
1140	load_blktrace(argv[optind], need_swap);
1141	first_ttime /= 1000ULL;
1142
1143	return output_p();
1144}
1145