Fix for a race when fio prints I/O statistics periodically
[fio.git] / t / btrace2fio.c
1 #include <stdio.h>
2 #include <stdio.h>
3 #include <unistd.h>
4 #include <inttypes.h>
5 #include <assert.h>
6
7 #include "../io_ddir.h"
8 #include "../flist.h"
9 #include "../hash.h"
10 #include "../fifo.h"
11 #include "../blktrace_api.h"
12 #include "../os/os.h"
13 #include "../log.h"
14 #include "../lib/linux-dev-lookup.h"
15
16 #define TRACE_FIFO_SIZE 8192
17
18 static unsigned int rt_threshold = 1000000;
19 static unsigned int ios_threshold = 10;
20 static unsigned int rate_threshold;
21 static unsigned int set_rate;
22 static int output_ascii = 1;
23 static char *filename;
24
25 struct bs {
26         unsigned int bs;
27         unsigned int nr;
28         int merges;
29 };
30
31 struct trace_file {
32         char *name;
33         int major, minor;
34 };
35
36 struct btrace_out {
37         unsigned long ios[DDIR_RWDIR_CNT];
38         unsigned long merges[DDIR_RWDIR_CNT];
39
40         uint64_t last_end[DDIR_RWDIR_CNT];
41         uint64_t seq[DDIR_RWDIR_CNT];
42
43         struct bs *bs[DDIR_RWDIR_CNT];
44         unsigned int nr_bs[DDIR_RWDIR_CNT];
45
46         int inflight;
47         unsigned int depth;
48         uint64_t first_ttime[DDIR_RWDIR_CNT];
49         uint64_t last_ttime[DDIR_RWDIR_CNT];
50         uint64_t kb[DDIR_RWDIR_CNT];
51
52         uint64_t start_delay;
53 };
54
55 struct btrace_pid {
56         struct flist_head hash_list;
57         struct flist_head pid_list;
58         pid_t pid;
59
60         struct trace_file *files;
61         int nr_files;
62         unsigned int last_major, last_minor;
63
64         struct btrace_out o;
65 };
66
67 struct inflight {
68         struct flist_head list;
69         struct btrace_pid *p;
70         uint64_t end_sector;
71 };
72
73 #define PID_HASH_BITS   10
74 #define PID_HASH_SIZE   (1U << PID_HASH_BITS)
75
76 static struct flist_head pid_hash[PID_HASH_SIZE];
77 static FLIST_HEAD(pid_list);
78
79 #define INFLIGHT_HASH_BITS      8
80 #define INFLIGHT_HASH_SIZE      (1U << INFLIGHT_HASH_BITS)
81 static struct flist_head inflight_hash[INFLIGHT_HASH_SIZE];
82
83 static uint64_t first_ttime = -1ULL;
84
85 static struct inflight *inflight_find(uint64_t sector)
86 {
87         struct flist_head *inflight_list;
88         struct flist_head *e;
89
90         inflight_list = &inflight_hash[hash_long(sector, INFLIGHT_HASH_BITS)];
91
92         flist_for_each(e, inflight_list) {
93                 struct inflight *i = flist_entry(e, struct inflight, list);
94
95                 if (i->end_sector == sector)
96                         return i;
97         }
98
99         return NULL;
100 }
101
102 static void inflight_remove(struct inflight *i)
103 {
104         struct btrace_out *o = &i->p->o;
105
106         o->inflight--;
107         assert(o->inflight >= 0);
108         flist_del(&i->list);
109         free(i);
110 }
111
112 static void __inflight_add(struct inflight *i)
113 {
114         struct flist_head *list;
115
116         list = &inflight_hash[hash_long(i->end_sector, INFLIGHT_HASH_BITS)];
117         flist_add_tail(&i->list, list);
118 }
119
120 static void inflight_add(struct btrace_pid *p, uint64_t sector, uint32_t len)
121 {
122         struct btrace_out *o = &p->o;
123         struct inflight *i;
124
125         i = calloc(1, sizeof(*i));
126         i->p = p;
127         o->inflight++;
128         o->depth = max((int) o->depth, o->inflight);
129         i->end_sector = sector + (len >> 9);
130         __inflight_add(i);
131 }
132
133 static void inflight_merge(struct inflight *i, int rw, unsigned int size)
134 {
135         i->p->o.merges[rw]++;
136         if (size) {
137                 i->end_sector += (size >> 9);
138                 flist_del(&i->list);
139                 __inflight_add(i);
140         }
141 }
142
143 /*
144  * fifo refill frontend, to avoid reading data in trace sized bites
145  */
146 static int refill_fifo(struct fifo *fifo, int fd)
147 {
148         char buf[TRACE_FIFO_SIZE];
149         unsigned int total;
150         int ret;
151
152         total = sizeof(buf);
153         if (total > fifo_room(fifo))
154                 total = fifo_room(fifo);
155
156         ret = read(fd, buf, total);
157         if (ret < 0) {
158                 perror("read refill");
159                 return -1;
160         }
161
162         if (ret > 0)
163                 ret = fifo_put(fifo, buf, ret);
164
165         return ret;
166 }
167
168 /*
169  * Retrieve 'len' bytes from the fifo, refilling if necessary.
170  */
171 static int trace_fifo_get(struct fifo *fifo, int fd, void *buf,
172                           unsigned int len)
173 {
174         if (fifo_len(fifo) < len) {
175                 int ret = refill_fifo(fifo, fd);
176
177                 if (ret < 0)
178                         return ret;
179         }
180
181         return fifo_get(fifo, buf, len);
182 }
183
184 /*
185  * Just discard the pdu by seeking past it.
186  */
187 static int discard_pdu(struct fifo *fifo, int fd, struct blk_io_trace *t)
188 {
189         if (t->pdu_len == 0)
190                 return 0;
191
192         return trace_fifo_get(fifo, fd, NULL, t->pdu_len);
193 }
194
195 static int handle_trace_notify(struct blk_io_trace *t)
196 {
197         switch (t->action) {
198         case BLK_TN_PROCESS:
199                 //printf("got process notify: %x, %d\n", t->action, t->pid);
200                 break;
201         case BLK_TN_TIMESTAMP:
202                 //printf("got timestamp notify: %x, %d\n", t->action, t->pid);
203                 break;
204         case BLK_TN_MESSAGE:
205                 break;
206         default:
207                 log_err("unknown trace act %x\n", t->action);
208                 return 1;
209         }
210
211         return 0;
212 }
213
214 static void __add_bs(struct btrace_out *o, unsigned int len, int rw)
215 {
216         o->bs[rw] = realloc(o->bs[rw], (o->nr_bs[rw] + 1) * sizeof(struct bs));
217         o->bs[rw][o->nr_bs[rw]].bs = len;
218         o->bs[rw][o->nr_bs[rw]].nr = 1;
219         o->nr_bs[rw]++;
220 }
221
222 static void add_bs(struct btrace_out *o, unsigned int len, int rw)
223 {
224         struct bs *bs = o->bs[rw];
225         int i;
226
227         if (!o->nr_bs[rw]) {
228                 __add_bs(o, len, rw);
229                 return;
230         }
231
232         for (i = 0; i < o->nr_bs[rw]; i++) {
233                 if (bs[i].bs == len) {
234                         bs[i].nr++;
235                         return;
236                 }
237         }
238
239         __add_bs(o, len, rw);
240 }
241
242 #define FMINORBITS      20
243 #define FMINORMASK      ((1U << FMINORBITS) - 1)
244 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
245 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
246
247 static int btrace_add_file(struct btrace_pid *p, uint32_t devno)
248 {
249         unsigned int maj = FMAJOR(devno);
250         unsigned int min = FMINOR(devno);
251         struct trace_file *f;
252         unsigned int i;
253         char dev[256];
254
255         if (filename)
256                 return 0;
257         if (p->last_major == maj && p->last_minor == min)
258                 return 0;
259
260         p->last_major = maj;
261         p->last_minor = min;
262
263         /*
264          * check for this file in our list
265          */
266         for (i = 0; i < p->nr_files; i++) {
267                 f = &p->files[i];
268
269                 if (f->major == maj && f->minor == min)
270                         return 0;
271         }
272
273         strcpy(dev, "/dev");
274         if (!blktrace_lookup_device(NULL, dev, maj, min)) {
275                 log_err("fio: failed to find device %u/%u\n", maj, min);
276                 if (!output_ascii) {
277                         log_err("fio: use -d to specify device\n");
278                         return 1;
279                 }
280                 return 0;
281         }
282
283         p->files = realloc(p->files, (p->nr_files + 1) * sizeof(*f));
284         f = &p->files[p->nr_files];
285         f->name = strdup(dev);
286         f->major = maj;
287         f->minor = min;
288         p->nr_files++;
289         return 0;
290 }
291
292 static int t_to_rwdir(struct blk_io_trace *t)
293 {
294         if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
295                 return DDIR_TRIM;
296
297         return (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
298 }
299
300 static int handle_trace_discard(struct blk_io_trace *t, struct btrace_pid *p)
301 {
302         struct btrace_out *o = &p->o;
303
304         if (btrace_add_file(p, t->device))
305                 return 1;
306
307         if (o->first_ttime[2] == -1ULL)
308                 o->first_ttime[2] = t->time;
309
310         o->ios[DDIR_TRIM]++;
311         add_bs(o, t->bytes, DDIR_TRIM);
312         return 0;
313 }
314
315 static int handle_trace_fs(struct blk_io_trace *t, struct btrace_pid *p)
316 {
317         struct btrace_out *o = &p->o;
318         int rw;
319
320         if (btrace_add_file(p, t->device))
321                 return 1;
322
323         first_ttime = min(first_ttime, (uint64_t) t->time);
324
325         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
326
327         if (o->first_ttime[rw] == -1ULL)
328                 o->first_ttime[rw] = t->time;
329
330         add_bs(o, t->bytes, rw);
331         o->ios[rw]++;
332
333         if (t->sector == o->last_end[rw] || o->last_end[rw] == -1ULL)
334                 o->seq[rw]++;
335
336         o->last_end[rw] = t->sector + (t->bytes >> 9);
337         return 0;
338 }
339
340 static int handle_queue_trace(struct blk_io_trace *t, struct btrace_pid *p)
341 {
342         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
343                 return handle_trace_notify(t);
344         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
345                 return handle_trace_discard(t, p);
346         else
347                 return handle_trace_fs(t, p);
348 }
349
350 static int handle_trace(struct blk_io_trace *t, struct btrace_pid *p)
351 {
352         unsigned int act = t->action & 0xffff;
353         int ret = 0;
354
355         if (act == __BLK_TA_QUEUE) {
356                 inflight_add(p, t->sector, t->bytes);
357                 ret = handle_queue_trace(t, p);
358         } else if (act == __BLK_TA_REQUEUE) {
359                 p->o.inflight--;
360         } else if (act == __BLK_TA_BACKMERGE) {
361                 struct inflight *i;
362
363                 i = inflight_find(t->sector + (t->bytes >> 9));
364                 if (i)
365                         inflight_remove(i);
366
367                 i = inflight_find(t->sector);
368                 if (i)
369                         inflight_merge(i, t_to_rwdir(t), t->bytes);
370         } else if (act == __BLK_TA_FRONTMERGE) {
371                 struct inflight *i;
372
373                 i = inflight_find(t->sector + (t->bytes >> 9));
374                 if (i)
375                         inflight_remove(i);
376
377                 i = inflight_find(t->sector);
378                 if (i)
379                         inflight_merge(i, t_to_rwdir(t), 0);
380         } else if (act == __BLK_TA_COMPLETE) {
381                 struct inflight *i;
382
383                 i = inflight_find(t->sector + (t->bytes >> 9));
384                 if (i) {
385                         i->p->o.kb[t_to_rwdir(t)] += (t->bytes >> 10);
386                         inflight_remove(i);
387                 }
388         }
389
390         return ret;
391 }
392
393 static void byteswap_trace(struct blk_io_trace *t)
394 {
395         t->magic = fio_swap32(t->magic);
396         t->sequence = fio_swap32(t->sequence);
397         t->time = fio_swap64(t->time);
398         t->sector = fio_swap64(t->sector);
399         t->bytes = fio_swap32(t->bytes);
400         t->action = fio_swap32(t->action);
401         t->pid = fio_swap32(t->pid);
402         t->device = fio_swap32(t->device);
403         t->cpu = fio_swap32(t->cpu);
404         t->error = fio_swap16(t->error);
405         t->pdu_len = fio_swap16(t->pdu_len);
406 }
407
408 static struct btrace_pid *pid_hash_find(pid_t pid, struct flist_head *list)
409 {
410         struct flist_head *e;
411         struct btrace_pid *p;
412
413         flist_for_each(e, list) {
414                 p = flist_entry(e, struct btrace_pid, hash_list);
415                 if (p->pid == pid)
416                         return p;
417         }
418
419         return NULL;
420 }
421
422 static struct btrace_pid *pid_hash_get(pid_t pid)
423 {
424         struct flist_head *hash_list;
425         struct btrace_pid *p;
426
427         hash_list = &pid_hash[hash_long(pid, PID_HASH_BITS)];
428
429         p = pid_hash_find(pid, hash_list);
430         if (!p) {
431                 int i;
432
433                 p = calloc(1, sizeof(*p));
434
435                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
436                         p->o.first_ttime[i] = -1ULL;
437                         p->o.last_ttime[i] = -1ULL;
438                         p->o.last_end[i] = -1ULL;
439                 }
440
441                 p->pid = pid;
442                 flist_add_tail(&p->hash_list, hash_list);
443                 flist_add_tail(&p->pid_list, &pid_list);
444         }
445
446         return p;
447 }
448
449 /*
450  * Load a blktrace file by reading all the blk_io_trace entries, and storing
451  * them as io_pieces like the fio text version would do.
452  */
453 static int load_blktrace(const char *fname, int need_swap)
454 {
455         struct btrace_pid *p;
456         unsigned long traces;
457         struct blk_io_trace t;
458         struct fifo *fifo;
459         int fd, ret = 0;
460
461         fd = open(fname, O_RDONLY);
462         if (fd < 0) {
463                 perror("open trace file\n");
464                 return 1;
465         }
466
467         fifo = fifo_alloc(TRACE_FIFO_SIZE);
468
469         traces = 0;
470         do {
471                 ret = trace_fifo_get(fifo, fd, &t, sizeof(t));
472                 if (ret < 0)
473                         goto err;
474                 else if (!ret)
475                         break;
476                 else if (ret < (int) sizeof(t)) {
477                         log_err("fio: short fifo get\n");
478                         break;
479                 }
480
481                 if (need_swap)
482                         byteswap_trace(&t);
483
484                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
485                         log_err("fio: bad magic in blktrace data: %x\n", t.magic);
486                         goto err;
487                 }
488                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
489                         log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
490                         goto err;
491                 }
492                 ret = discard_pdu(fifo, fd, &t);
493                 if (ret < 0) {
494                         log_err("blktrace lseek\n");
495                         goto err;
496                 } else if (t.pdu_len != ret) {
497                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
498                         goto err;
499                 }
500
501                 p = pid_hash_get(t.pid);
502                 ret = handle_trace(&t, p);
503                 if (ret)
504                         break;
505                 p->o.last_ttime[t_to_rwdir(&t)] = t.time;
506                 traces++;
507         } while (1);
508
509         fifo_free(fifo);
510         close(fd);
511
512         if (ret)
513                 return ret;
514
515         if (output_ascii)
516                 printf("Traces loaded: %lu\n", traces);
517
518         return 0;
519 err:
520         close(fd);
521         fifo_free(fifo);
522         return 1;
523 }
524
525 static int bs_cmp(const void *ba, const void *bb)
526 {
527         const struct bs *bsa = ba;
528         const struct bs *bsb = bb;
529
530         return bsb->nr - bsa->nr;
531 }
532
533 static unsigned long o_to_kb_rate(struct btrace_out *o, int rw)
534 {
535         uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL;
536         uint64_t val;
537
538         if (!usec)
539                 return 0;
540
541         val = o->kb[rw] * 1000ULL;
542         return val / (usec / 1000ULL);
543 }
544
545 static uint64_t o_first_ttime(struct btrace_out *o)
546 {
547         uint64_t first;
548
549         first = min(o->first_ttime[0], o->first_ttime[1]);
550         return min(first, o->first_ttime[2]);
551 }
552
553 static uint64_t o_longest_ttime(struct btrace_out *o)
554 {
555         uint64_t ret = 0;
556         int i;
557
558         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
559                 uint64_t diff;
560
561                 diff = o->last_ttime[i] - o->first_ttime[i];
562                 ret = max(diff, ret);
563         }
564
565         return ret;
566 }
567
568 static void __output_p_ascii(struct btrace_pid *p, unsigned long *ios)
569 {
570         const char *msg[] = { "reads", "writes", "trims" };
571         struct btrace_out *o = &p->o;
572         unsigned long total, usec;
573         int i, j;
574
575         printf("[pid:\t%u]\n", p->pid);
576
577         total = ddir_rw_sum(o->ios);
578         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
579                 float perc;
580
581                 if (!o->ios[i])
582                         continue;
583
584                 ios[i] += o->ios[i] + o->merges[i];
585                 printf("%s\n", msg[i]);
586                 perc = ((float) o->ios[i] * 100.0) / (float) total;
587                 printf("\tios:    %lu (perc=%3.2f%%)\n", o->ios[i], perc);
588                 perc = ((float) o->merges[i] * 100.0) / (float) total;
589                 printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc);
590                 perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
591                 printf("\tseq:    %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc);
592                 printf("\trate:   %lu KB/sec\n", o_to_kb_rate(o, i));
593
594                 for (j = 0; j < o->nr_bs[i]; j++) {
595                         struct bs *bs = &o->bs[i][j];
596
597                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
598                         printf("\tbs=%u, perc=%3.2f%%\n", bs->bs, perc);
599                 }
600         }
601
602         printf("depth:\t%u\n", o->depth);
603         usec = o_longest_ttime(o) / 1000ULL;
604         printf("usec:\t%lu (delay=%llu)\n", usec, (unsigned long long) o->start_delay);
605
606         printf("files:\t");
607         for (i = 0; i < p->nr_files; i++)
608                 printf("%s,", p->files[i].name);
609         printf("\n");
610
611         printf("\n");
612 }
613
614 static int __output_p_fio(struct btrace_pid *p, unsigned long *ios)
615 {
616         struct btrace_out *o = &p->o;
617         unsigned long total;
618         unsigned long long time;
619         float perc;
620         int i, j;
621
622         if ((o->ios[0] + o->ios[1]) && o->ios[2]) {
623                 log_err("fio: trace has both read/write and trim\n");
624                 return 1;
625         }
626         if (!p->nr_files) {
627                 log_err("fio: no devices found\n");
628                 return 1;
629         }
630
631         printf("[pid%u]\n", p->pid);
632         printf("direct=1\n");
633         if (o->depth == 1)
634                 printf("ioengine=sync\n");
635         else
636                 printf("ioengine=libaio\niodepth=%u\n", o->depth);
637
638         if (o->ios[0] && !o->ios[1])
639                 printf("rw=randread\n");
640         else if (!o->ios[0] && o->ios[1])
641                 printf("rw=randwrite\n");
642         else if (o->ios[2])
643                 printf("rw=randtrim\n");
644         else {
645                 printf("rw=randrw\n");
646                 total = ddir_rw_sum(o->ios);
647                 perc = ((float) o->ios[0] * 100.0) / (float) total;
648                 printf("rwmixread=%u\n", (int) (perc + 0.99));
649         }
650
651         printf("percentage_random=");
652         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
653                 if (o->seq[i] && o->ios[i]) {
654                         perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
655                         if (perc >= 99.0)
656                                 perc = 100.0;
657                 } else
658                         perc = 100.0;
659
660                 if (i)
661                         printf(",");
662                 perc = 100.0 - perc;
663                 printf("%u", (int) perc);
664         }
665         printf("\n");
666
667         printf("filename=");
668         for (i = 0; i < p->nr_files; i++) {
669                 if (i)
670                         printf(":");
671                 printf("%s", p->files[i].name);
672         }
673         printf("\n");
674
675         printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
676
677         time = o_longest_ttime(o);
678         time = (time + 1000000000ULL - 1) / 1000000000ULL;
679         printf("runtime=%llus\n", time);
680
681         printf("bssplit=");
682         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
683
684                 if (i && o->nr_bs[i - 1] && o->nr_bs[i])
685                         printf(",");
686
687                 for (j = 0; j < o->nr_bs[i]; j++) {
688                         struct bs *bs = &o->bs[i][j];
689
690                         perc = (((float) bs->nr * 100.0) / (float) o->ios[i]);
691                         if (perc < 1.00)
692                                 continue;
693                         if (j)
694                                 printf(":");
695                         if (j + 1 == o->nr_bs[i])
696                                 printf("%u/", bs->bs);
697                         else
698                                 printf("%u/%u", bs->bs, (int) perc);
699                 }
700         }
701         printf("\n");
702
703         if (set_rate) {
704                 printf("rate=");
705                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
706                         unsigned long rate;
707
708                         rate = o_to_kb_rate(o, i);
709                         if (i)
710                                 printf(",");
711                         if (rate)
712                                 printf("%luk", rate);
713                 }
714                 printf("\n");
715         }
716
717         printf("\n");
718         return 0;
719 }
720
721 static int __output_p(struct btrace_pid *p, unsigned long *ios)
722 {
723         struct btrace_out *o = &p->o;
724         int i, ret = 0;
725
726         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
727                 if (o->nr_bs[i] <= 1)
728                         continue;
729                 qsort(o->bs[i], o->nr_bs[i], sizeof(struct bs), bs_cmp);
730         }
731
732         if (filename) {
733                 p->files = malloc(sizeof(struct trace_file));
734                 p->nr_files++;
735                 p->files[0].name = filename;
736         }
737
738         if (output_ascii)
739                 __output_p_ascii(p, ios);
740         else
741                 ret = __output_p_fio(p, ios);
742
743         return ret;
744 }
745
746 static void remove_ddir(struct btrace_out *o, int rw)
747 {
748         o->ios[rw] = 0;
749 }
750
751 static int prune_entry(struct btrace_out *o)
752 {
753         unsigned long rate;
754         uint64_t time;
755         int i;
756
757         if (ddir_rw_sum(o->ios) < ios_threshold)
758                 return 1;
759
760         time = o_longest_ttime(o) / 1000ULL;
761         if (time < rt_threshold)
762                 return 1;
763
764         rate = 0;
765         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
766                 unsigned long this_rate;
767
768                 this_rate = o_to_kb_rate(o, i);
769                 if (this_rate < rate_threshold) {
770                         remove_ddir(o, i);
771                         this_rate = 0;
772                 }
773                 rate += this_rate;
774         }
775
776         if (rate < rate_threshold)
777                 return 1;
778
779         return 0;
780 }
781
782 static int entry_cmp(void *priv, struct flist_head *a, struct flist_head *b)
783 {
784         struct btrace_pid *pa = flist_entry(a, struct btrace_pid, pid_list);
785         struct btrace_pid *pb = flist_entry(b, struct btrace_pid, pid_list);
786
787         return ddir_rw_sum(pb->o.ios) - ddir_rw_sum(pa->o.ios);
788 }
789
790 static void free_p(struct btrace_pid *p)
791 {
792         struct btrace_out *o = &p->o;
793         int i;
794
795         for (i = 0; i < p->nr_files; i++) {
796                 if (p->files[i].name && p->files[i].name != filename)
797                         free(p->files[i].name);
798         }
799
800         for (i = 0; i < DDIR_RWDIR_CNT; i++)
801                 free(o->bs[i]);
802
803         free(p->files);
804         flist_del(&p->pid_list);
805         flist_del(&p->hash_list);
806         free(p);
807 }
808
809 static int output_p(void)
810 {
811         unsigned long ios[DDIR_RWDIR_CNT];
812         struct flist_head *e, *tmp;
813         int ret = 0;
814
815         flist_for_each_safe(e, tmp, &pid_list) {
816                 struct btrace_pid *p;
817
818                 p = flist_entry(e, struct btrace_pid, pid_list);
819                 if (prune_entry(&p->o)) {
820                         free_p(p);
821                         continue;
822                 }
823                 p->o.start_delay = (o_first_ttime(&p->o) / 1000ULL) - first_ttime;
824         }
825
826         memset(ios, 0, sizeof(ios));
827
828         flist_sort(NULL, &pid_list, entry_cmp);
829
830         flist_for_each(e, &pid_list) {
831                 struct btrace_pid *p;
832
833                 p = flist_entry(e, struct btrace_pid, pid_list);
834                 ret |= __output_p(p, ios);
835                 if (ret && !output_ascii)
836                         break;
837         }
838
839         if (output_ascii)
840                 printf("Total: reads=%lu, writes=%lu\n", ios[0], ios[1]);
841
842         return ret;
843 }
844
845 static int usage(char *argv[])
846 {
847         log_err("%s: <blktrace bin file>\n", argv[0]);
848         log_err("\t-t\tUsec threshold to ignore task\n");
849         log_err("\t-n\tNumber IOS threshold to ignore task\n");
850         log_err("\t-f\tFio job file output\n");
851         log_err("\t-d\tUse this file/device for replay\n");
852         log_err("\t-r\tIgnore jobs with less than this KB/sec rate\n");
853         log_err("\t-R\tSet rate in fio job\n");
854         return 1;
855 }
856
857 static int trace_needs_swap(const char *trace_file, int *swap)
858 {
859         struct blk_io_trace t;
860         int fd, ret;
861
862         *swap = -1;
863         
864         fd = open(trace_file, O_RDONLY);
865         if (fd < 0) {
866                 perror("open");
867                 return 1;
868         }
869
870         ret = read(fd, &t, sizeof(t));
871         if (ret < 0) {
872                 close(fd);
873                 perror("read");
874                 return 1;
875         } else if (ret != sizeof(t)) {
876                 close(fd);
877                 log_err("fio: short read on trace file\n");
878                 return 1;
879         }
880
881         close(fd);
882
883         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
884                 *swap = 0;
885         else {
886                 /*
887                  * Maybe it needs to be endian swapped...
888                  */
889                 t.magic = fio_swap32(t.magic);
890                 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
891                         *swap = 1;
892         }
893
894         if (*swap == -1) {
895                 log_err("fio: blktrace appears corrupt\n");
896                 return 1;
897         }
898
899         return 0;
900 }
901
902 int main(int argc, char *argv[])
903 {
904         int need_swap, i, c;
905
906         if (argc < 2)
907                 return usage(argv);
908
909         while ((c = getopt(argc, argv, "t:n:fd:r:R")) != -1) {
910                 switch (c) {
911                 case 'R':
912                         set_rate = 1;
913                         break;
914                 case 'r':
915                         rate_threshold = atoi(optarg);
916                         break;
917                 case 't':
918                         rt_threshold = atoi(optarg);
919                         break;
920                 case 'n':
921                         ios_threshold = atoi(optarg);
922                         break;
923                 case 'f':
924                         output_ascii = 0;
925                         break;
926                 case 'd':
927                         filename = strdup(optarg);
928                         break;
929                 case '?':
930                 default:
931                         return usage(argv);
932                 }
933         }
934
935         if (argc == optind)
936                 return usage(argv);
937
938         if (trace_needs_swap(argv[optind], &need_swap))
939                 return 1;
940
941         for (i = 0; i < PID_HASH_SIZE; i++)
942                 INIT_FLIST_HEAD(&pid_hash[i]);
943         for (i = 0; i < INFLIGHT_HASH_SIZE; i++)
944                 INIT_FLIST_HEAD(&inflight_hash[i]);
945
946         load_blktrace(argv[optind], need_swap);
947         first_ttime /= 1000ULL;
948
949         return output_p();
950 }