[PATCH] blkparse: Add support for the plug/unplug events
[blktrace.git] / blkparse.c
1 /*
2  * block queue tracing parse application
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  *
6  *  This program is free software; you can redistribute it and/or modify
7  *  it under the terms of the GNU General Public License as published by
8  *  the Free Software Foundation; either version 2 of the License, or
9  *  (at your option) any later version.
10  *
11  *  This program is distributed in the hope that it will be useful,
12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  *  GNU General Public License for more details.
15  *
16  *  You should have received a copy of the GNU General Public License
17  *  along with this program; if not, write to the Free Software
18  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
19  *
20  */
21 #include <sys/types.h>
22 #include <sys/stat.h>
23 #include <unistd.h>
24 #include <stdio.h>
25 #include <fcntl.h>
26 #include <stdlib.h>
27 #include <string.h>
28 #include <getopt.h>
29 #include <errno.h>
30 #include <signal.h>
31 #include <locale.h>
32 #include <limits.h>
33
34 #include "blktrace.h"
35 #include "rbtree.h"
36
37 #define SECONDS(x)              ((unsigned long long)(x) / 1000000000)
38 #define NANO_SECONDS(x)         ((unsigned long long)(x) % 1000000000)
39 #define DOUBLE_TO_NANO_ULL(d)   ((unsigned long long)((d) * 1000000000))
40
41 #define MINORBITS       20
42 #define MINORMASK       ((1U << MINORBITS) - 1)
43 #define MAJOR(dev)      ((unsigned int) ((dev) >> MINORBITS))
44 #define MINOR(dev)      ((unsigned int) ((dev) & MINORMASK))
45
46 #define min(a, b)       ((a) < (b) ? (a) : (b))
47
48 struct io_stats {
49         unsigned long qreads, qwrites, creads, cwrites, mreads, mwrites;
50         unsigned long ireads, iwrites;
51         unsigned long long qread_kb, qwrite_kb, cread_kb, cwrite_kb;
52         unsigned long long iread_kb, iwrite_kb;
53 };
54
55 struct per_cpu_info {
56         int cpu;
57         int nelems;
58
59         int fd;
60         char fname[128];
61
62         struct io_stats io_stats;
63 };
64
65 struct per_dev_info {
66         dev_t id;
67         char *name;
68
69         int backwards;
70         unsigned long long events;
71         unsigned long long last_reported_time;
72         struct io_stats io_stats;
73
74         int ncpus;
75         struct per_cpu_info *cpus;
76 };
77
78 struct per_process_info {
79         char name[16];
80         __u32 pid;
81         struct io_stats io_stats;
82         struct per_process_info *hash_next, *list_next;
83
84         /*
85          * individual io stats
86          */
87         unsigned long long longest_allocation_wait[2];
88         unsigned long long longest_dispatch_wait[2];
89         unsigned long long longest_completion_wait[2];
90 };
91
92 #define PPI_HASH_SHIFT  (8)
93 static struct per_process_info *ppi_hash[1 << PPI_HASH_SHIFT];
94 static struct per_process_info *ppi_list;
95
96 #define S_OPTS  "i:o:b:stqw:"
97 static struct option l_opts[] = {
98         {
99                 .name = "input",
100                 .has_arg = 1,
101                 .flag = NULL,
102                 .val = 'i'
103         },
104         {
105                 .name = "output",
106                 .has_arg = 1,
107                 .flag = NULL,
108                 .val = 'o'
109         },
110         {
111                 .name = "batch",
112                 .has_arg = 1,
113                 .flag = NULL,
114                 .val = 'b'
115         },
116         {
117                 .name = "per program stats",
118                 .has_arg = 0,
119                 .flag = NULL,
120                 .val = 's'
121         },
122         {
123                 .name = "track ios",
124                 .has_arg = 0,
125                 .flag = NULL,
126                 .val = 't'
127         },
128         {
129                 .name = "quiet",
130                 .has_arg = 0,
131                 .flag = NULL,
132                 .val = 'q'
133         },
134         {
135                 .name = "stopwatch",
136                 .has_arg = 1,
137                 .flag = NULL,
138                 .val = 'w'
139         },
140         {
141                 .name = NULL,
142                 .has_arg = 0,
143                 .flag = NULL,
144                 .val = 0
145         }
146 };
147
148 static struct rb_root rb_sort_root;
149 static struct rb_root rb_track_root;
150
151 /*
152  * for sorting the displayed output
153  */
154 struct trace {
155         struct blk_io_trace *bit;
156         struct rb_node rb_node;
157 };
158
159 /*
160  * for tracking individual ios
161  */
162 struct io_track {
163         struct rb_node rb_node;
164
165         dev_t device;
166         __u64 sector;
167         __u32 pid;
168         unsigned long long allocation_time;
169         unsigned long long queue_time;
170         unsigned long long dispatch_time;
171         unsigned long long completion_time;
172 };
173
174 static int ndevices;
175 static struct per_dev_info *devices;
176 static char *get_dev_name(struct per_dev_info *, char *, int);
177
178 static FILE *ofp;
179 static char *output_name;
180
181 static unsigned long long genesis_time;
182 static unsigned long long stopwatch_start;      /* start from zero by default */
183 static unsigned long long stopwatch_end = ULONG_LONG_MAX;       /* "infinity" */
184
185 static int per_process_stats;
186 static int track_ios;
187
188 #define RB_BATCH_DEFAULT        (1024)
189 static int rb_batch = RB_BATCH_DEFAULT;
190
191 static int pipeline;
192
193 #define is_done()       (*(volatile int *)(&done))
194 static volatile int done;
195
196 static inline unsigned long hash_long(unsigned long val)
197 {
198 #if __WORDSIZE == 32
199         val *= 0x9e370001UL;
200 #elif __WORDSIZE == 64
201         val *= 0x9e37fffffffc0001UL;
202 #else
203 #error unknown word size
204 #endif
205
206         return val >> (__WORDSIZE - PPI_HASH_SHIFT);
207 }
208
209 static inline void add_process_to_hash(struct per_process_info *ppi)
210 {
211         const int hash_idx = hash_long(ppi->pid);
212
213         ppi->hash_next = ppi_hash[hash_idx];
214         ppi_hash[hash_idx] = ppi;
215 }
216
217 static inline void add_process_to_list(struct per_process_info *ppi)
218 {
219         ppi->list_next = ppi_list;
220         ppi_list = ppi;
221 }
222
223 static struct per_process_info *find_process_by_pid(__u32 pid)
224 {
225         const int hash_idx = hash_long(pid);
226         struct per_process_info *ppi;
227
228         ppi = ppi_hash[hash_idx];
229         while (ppi) {
230                 if (ppi->pid == pid)
231                         return ppi;
232
233                 ppi = ppi->hash_next;
234         }
235
236         return NULL;
237 }
238
239 static inline int trace_rb_insert(struct trace *t)
240 {
241         struct rb_node **p = &rb_sort_root.rb_node;
242         struct rb_node *parent = NULL;
243         struct trace *__t;
244
245         if (genesis_time == 0 || t->bit->time < genesis_time)
246                 genesis_time = t->bit->time;
247
248         while (*p) {
249                 parent = *p;
250                 __t = rb_entry(parent, struct trace, rb_node);
251
252                 if (t->bit->time < __t->bit->time)
253                         p = &(*p)->rb_left;
254                 else if (t->bit->time > __t->bit->time)
255                         p = &(*p)->rb_right;
256                 else if (t->bit->device < __t->bit->device)
257                         p = &(*p)->rb_left;
258                 else if (t->bit->device > __t->bit->device)
259                         p = &(*p)->rb_right;
260                 else if (t->bit->sequence < __t->bit->sequence)
261                         p = &(*p)->rb_left;
262                 else if (t->bit->sequence > __t->bit->sequence)
263                         p = &(*p)->rb_right;
264                 else if (t->bit->device == __t->bit->device) {
265                         fprintf(stderr,
266                                 "sequence alias (%d) on device %d,%d!\n",
267                                 t->bit->sequence,
268                                 MAJOR(t->bit->device), MINOR(t->bit->device));
269                         return 1;
270                 }
271         }
272
273         rb_link_node(&t->rb_node, parent, p);
274         rb_insert_color(&t->rb_node, &rb_sort_root);
275         return 0;
276 }
277
278 static inline int track_rb_insert(struct io_track *iot)
279 {
280         struct rb_node **p = &rb_track_root.rb_node;
281         struct rb_node *parent = NULL;
282         struct io_track *__iot;
283
284         while (*p) {
285                 parent = *p;
286
287                 __iot = rb_entry(parent, struct io_track, rb_node);
288
289                 if (iot->device < __iot->device)
290                         p = &(*p)->rb_left;
291                 else if (iot->device > __iot->device)
292                         p = &(*p)->rb_right;
293                 else if (iot->sector < __iot->sector)
294                         p = &(*p)->rb_left;
295                 else if (iot->sector > __iot->sector)
296                         p = &(*p)->rb_right;
297                 else {
298                         fprintf(stderr,
299                                 "sector alias (%llu) on device %d,%d!\n",
300                                 iot->sector,
301                                 MAJOR(iot->device), MINOR(iot->device));
302                         return 1;
303                 }
304         }
305
306         rb_link_node(&iot->rb_node, parent, p);
307         rb_insert_color(&iot->rb_node, &rb_track_root);
308         return 0;
309 }
310
311 static struct io_track *__find_track(dev_t device, __u64 sector)
312 {
313         struct rb_node **p = &rb_track_root.rb_node;
314         struct rb_node *parent = NULL;
315         struct io_track *__iot;
316
317         while (*p) {
318                 parent = *p;
319                 
320                 __iot = rb_entry(parent, struct io_track, rb_node);
321
322                 if (device < __iot->device)
323                         p = &(*p)->rb_left;
324                 else if (device > __iot->device)
325                         p = &(*p)->rb_right;
326                 else if (sector < __iot->sector)
327                         p = &(*p)->rb_left;
328                 else if (sector > __iot->sector)
329                         p = &(*p)->rb_right;
330                 else
331                         return __iot;
332         }
333
334         return NULL;
335 }
336
337 static struct io_track *find_track(__u32 pid, dev_t device, __u64 sector)
338 {
339         struct io_track *iot;
340
341         iot = __find_track(device, sector);
342         if (!iot) {
343                 iot = malloc(sizeof(*iot));
344                 iot->pid = pid;
345                 iot->device = device;
346                 iot->sector = sector;
347                 track_rb_insert(iot);
348         }
349
350         return iot;
351 }
352
353 static void log_track_merge(struct blk_io_trace *t)
354 {
355         struct io_track *iot;
356
357         if (!track_ios)
358                 return;
359         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
360                 return;
361
362         iot = __find_track(t->device, t->sector - (t->bytes >> 10));
363         if (!iot) {
364                 fprintf(stderr, "Trying to merge on non-existing request\n");
365                 return;
366         }
367
368         rb_erase(&iot->rb_node, &rb_track_root);
369         iot->sector -= t->bytes >> 10;
370         track_rb_insert(iot);
371 }
372
373 static void log_track_getrq(struct blk_io_trace *t)
374 {
375         struct io_track *iot;
376
377         if (!track_ios)
378                 return;
379
380         iot = find_track(t->pid, t->device, t->sector);
381         iot->allocation_time = t->time;
382 }
383
384
385 /*
386  * return time between rq allocation and queue
387  */
388 static unsigned long long log_track_queue(struct blk_io_trace *t)
389 {
390         unsigned long long elapsed;
391         struct io_track *iot;
392
393         if (!track_ios)
394                 return -1;
395
396         iot = find_track(t->pid, t->device, t->sector);
397         iot->queue_time = t->time;
398         elapsed = iot->queue_time - iot->allocation_time;
399
400         if (per_process_stats) {
401                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
402                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
403
404                 if (ppi && elapsed > ppi->longest_allocation_wait[w])
405                         ppi->longest_allocation_wait[w] = elapsed;
406         }
407
408         return elapsed;
409 }
410
411 /*
412  * return time between queue and issue
413  */
414 static unsigned long long log_track_issue(struct blk_io_trace *t)
415 {
416         unsigned long long elapsed;
417         struct io_track *iot;
418
419         if (!track_ios)
420                 return -1;
421         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
422                 return -1;
423
424         iot = __find_track(t->device, t->sector);
425         if (!iot) {
426                 fprintf(stderr, "Trying to issue on non-existing request\n");
427                 return -1;
428         }
429
430         iot->dispatch_time = t->time;
431         elapsed = iot->dispatch_time - iot->queue_time;
432
433         if (per_process_stats) {
434                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
435                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
436
437                 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
438                         ppi->longest_dispatch_wait[w] = elapsed;
439         }
440
441         return elapsed;
442 }
443
444 /*
445  * return time between dispatch and complete
446  */
447 static unsigned long long log_track_complete(struct blk_io_trace *t)
448 {
449         unsigned long long elapsed;
450         struct io_track *iot;
451
452         if (!track_ios)
453                 return -1;
454         if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
455                 return -1;
456
457         iot = __find_track(t->device, t->sector);
458         if (!iot) {
459                 fprintf(stderr, "Trying to dispatch on non-existing request\n");
460                 return -1;
461         }
462
463         iot->completion_time = t->time;
464         elapsed = iot->completion_time - iot->dispatch_time;
465
466         if (per_process_stats) {
467                 struct per_process_info *ppi = find_process_by_pid(iot->pid);
468                 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
469
470                 if (ppi && elapsed > ppi->longest_completion_wait[w])
471                         ppi->longest_completion_wait[w] = elapsed;
472         }
473
474         /*
475          * kill the trace, we don't need it after completion
476          */
477         rb_erase(&iot->rb_node, &rb_track_root);
478         free(iot);
479
480         return elapsed;
481 }
482
483
484 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
485 {
486         struct per_process_info *ppi = find_process_by_pid(pid);
487
488         if (!ppi) {
489                 ppi = malloc(sizeof(*ppi));
490                 memset(ppi, 0, sizeof(*ppi));
491                 strncpy(ppi->name, name, sizeof(ppi->name));
492                 ppi->pid = pid;
493                 add_process_to_hash(ppi);
494                 add_process_to_list(ppi);
495         }
496
497         return &ppi->io_stats;
498 }
499
500
501 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
502 {
503         struct per_cpu_info *cpus = pdi->cpus;
504         int ncpus = pdi->ncpus;
505         int new_count = cpu + 1;
506         int new_space, size;
507         char *new_start;
508
509         size = new_count * sizeof(struct per_cpu_info);
510         cpus = realloc(cpus, size);
511         if (!cpus) {
512                 char name[20];
513                 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
514                         get_dev_name(pdi, name, sizeof(name)), size);
515                 exit(1);
516         }
517
518         new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
519         new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
520         memset(new_start, 0, new_space);
521
522         pdi->ncpus = new_count;
523         pdi->cpus = cpus;
524 }
525   
526 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
527 {
528         if (cpu >= pdi->ncpus)
529                 resize_cpu_info(pdi, cpu);
530         return &pdi->cpus[cpu];
531 }
532
533
534 static int resize_devices(char *name)
535 {
536         int size = (ndevices + 1) * sizeof(struct per_dev_info);
537
538         devices = realloc(devices, size);
539         if (!devices) {
540                 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
541                 return 1;
542         }
543         memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
544         devices[ndevices].name = name;
545         ndevices++;
546         return 0;
547 }
548
549 static struct per_dev_info *get_dev_info(dev_t id, int create)
550 {
551         int i;
552
553         for (i = 0; i < ndevices; i++)
554                 if (devices[i].id == id)
555                         return &devices[i];
556         if (!create)
557                 return NULL;
558         if (resize_devices(NULL) != 0)
559                 return NULL;
560         return &devices[ndevices-1];
561 }
562
563 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
564 {
565         if (pdi->name)
566                 snprintf(buffer, size, "%s", pdi->name);
567         else
568                 snprintf(buffer, size, "%d,%d", MAJOR(pdi->id), MINOR(pdi->id));
569         return buffer;
570 }
571
572
573 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
574 {
575         unsigned long long this = bit->time;
576         unsigned long long last = pdi->last_reported_time;
577
578         pdi->backwards = (this < last) ? 'B' : ' ';
579         pdi->last_reported_time = this;
580 }
581
582
583 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
584                                int rw)
585 {
586         if (rw) {
587                 ios->mwrites++;
588                 ios->qwrite_kb += t->bytes >> 10;
589         } else {
590                 ios->mreads++;
591                 ios->qread_kb += t->bytes >> 10;
592         }
593 }
594
595 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
596                              int rw)
597 {
598         __account_m(&pci->io_stats, t, rw);
599
600         if (per_process_stats) {
601                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
602
603                 __account_m(ios, t, rw);
604         }
605 }
606
607 static inline void __account_q(struct io_stats *ios, struct blk_io_trace *t,
608                                int rw)
609 {
610         if (rw) {
611                 ios->qwrites++;
612                 ios->qwrite_kb += t->bytes >> 10;
613         } else {
614                 ios->qreads++;
615                 ios->qread_kb += t->bytes >> 10;
616         }
617 }
618
619 static inline void account_q(struct blk_io_trace *t, struct per_cpu_info *pci,
620                              int rw)
621 {
622         __account_q(&pci->io_stats, t, rw);
623
624         if (per_process_stats) {
625                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
626
627                 __account_q(ios, t, rw);
628         }
629 }
630
631 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
632 {
633         if (rw) {
634                 ios->cwrites++;
635                 ios->cwrite_kb += bytes >> 10;
636         } else {
637                 ios->creads++;
638                 ios->cread_kb += bytes >> 10;
639         }
640 }
641
642 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
643                              int rw, int bytes)
644 {
645         __account_c(&pci->io_stats, rw, bytes);
646
647         if (per_process_stats) {
648                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
649
650                 __account_c(ios, rw, bytes);
651         }
652 }
653
654 static inline void __account_i(struct io_stats *ios, int rw, unsigned int bytes)
655 {
656         if (rw) {
657                 ios->iwrites++;
658                 ios->iwrite_kb += bytes >> 10;
659         } else {
660                 ios->ireads++;
661                 ios->iread_kb += bytes >> 10;
662         }
663 }
664
665 static inline void account_i(struct blk_io_trace *t, struct per_cpu_info *pci,
666                              int rw)
667 {
668         __account_i(&pci->io_stats, rw, t->bytes);
669
670         if (per_process_stats) {
671                 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
672
673                 __account_i(ios, rw, t->bytes);
674         }
675 }
676
677 static void output(struct per_cpu_info *pci, char *s)
678 {
679         fprintf(ofp, "%s", s);
680 }
681
682 static char hstring[256];
683 static char tstring[256];
684
685 static inline char *setup_header(struct per_cpu_info *pci,
686                                  struct blk_io_trace *t, char act)
687 {
688         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
689         int b = t->action & BLK_TC_ACT(BLK_TC_BARRIER);
690         int s = t->action & BLK_TC_ACT(BLK_TC_SYNC);
691         char rwbs[4];
692         int i = 0;
693
694         if (w)
695                 rwbs[i++] = 'W';
696         else
697                 rwbs[i++] = 'R';
698         if (b)
699                 rwbs[i++] = 'B';
700         if (s)
701                 rwbs[i++] = 'S';
702
703         rwbs[i] = '\0';
704
705         sprintf(hstring, "%3d,%-3d %2d %8ld %5Lu.%09Lu %5u %c %3s",
706                 MAJOR(t->device), MINOR(t->device), pci->cpu,
707                 (unsigned long)t->sequence, SECONDS(t->time), 
708                 NANO_SECONDS(t->time), t->pid, act, rwbs);
709
710         return hstring;
711 }
712
713 static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
714                          char act)
715 {
716         unsigned long long elapsed = log_track_complete(t);
717
718         if (elapsed != -1ULL) {
719                 unsigned long usec = elapsed / 1000;
720
721                 sprintf(tstring,"%s %Lu + %u (%8lu) [%d]\n",
722                         setup_header(pci, t, act),
723                         (unsigned long long)t->sector, t->bytes >> 9,
724                         usec, t->error);
725         } else {
726                 sprintf(tstring,"%s %Lu + %u [%d]\n", setup_header(pci, t, act),
727                         (unsigned long long)t->sector, t->bytes >> 9, t->error);
728         }
729         
730         output(pci, tstring);
731 }
732
733 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
734                       char act)
735 {
736         unsigned long long elapsed = log_track_queue(t);
737
738         if (elapsed != -1ULL) {
739                 unsigned long usec = elapsed / 1000;
740
741                 sprintf(tstring,"%s %Lu + %u (%8lu) [%s]\n",
742                         setup_header(pci, t, act),
743                         (unsigned long long)t->sector, t->bytes >> 9,
744                         usec, t->comm);
745         } else {
746                 sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
747                         (unsigned long long)t->sector, t->bytes >> 9, t->comm);
748         }
749         output(pci, tstring);
750 }
751
752 static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
753                       char act)
754 {
755         unsigned long long elapsed = log_track_issue(t);
756
757         if (elapsed != -1ULL) {
758                 double usec = (double) elapsed / 1000;
759
760                 sprintf(tstring,"%s %Lu + %u (%8.2f) [%s]\n",
761                         setup_header(pci, t, act),
762                         (unsigned long long)t->sector, t->bytes >> 9,
763                         usec, t->comm);
764         } else {
765                 sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
766                         (unsigned long long)t->sector, t->bytes >> 9, t->comm);
767         }
768
769         output(pci, tstring);
770 }
771
772 static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
773                       char act)
774 {
775         log_track_merge(t);
776
777         sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
778                 (unsigned long long)t->sector, t->bytes >> 9, t->comm);
779         output(pci, tstring);
780 }
781
782 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
783                         char act)
784 {
785         sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
786                 (unsigned long long)t->sector, t->bytes >> 9, t->comm);
787         output(pci, tstring);
788 }
789
790 static int log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char act)
791 {
792         unsigned char *buf;
793         int i;
794
795         sprintf(tstring,"%s ", setup_header(pci, t, act));
796         output(pci, tstring);
797
798         buf = (unsigned char *) t + sizeof(*t);
799         for (i = 0; i < t->pdu_len; i++) {
800                 sprintf(tstring,"%02x ", buf[i]);
801                 output(pci, tstring);
802         }
803
804         if (act == 'C') {
805                 sprintf(tstring,"[%d]\n", t->error);
806                 output(pci, tstring);
807         } else {
808                 sprintf(tstring,"[%s]\n", t->comm);
809                 output(pci, tstring);
810         }
811         return 0;
812 }
813
814 static int dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
815 {
816         int ret = 0;
817
818         switch (t->action & 0xffff) {
819                 case __BLK_TA_QUEUE:
820                         log_generic(pci, t, 'Q');
821                         break;
822                 case __BLK_TA_GETRQ:
823                         log_generic(pci, t, 'G');
824                         break;
825                 case __BLK_TA_SLEEPRQ:
826                         log_generic(pci, t, 'S');
827                         break;
828                 case __BLK_TA_REQUEUE:
829                         log_generic(pci, t, 'R');
830                         break;
831                 case __BLK_TA_ISSUE:
832                         ret = log_pc(pci, t, 'D');
833                         break;
834                 case __BLK_TA_COMPLETE:
835                         log_pc(pci, t, 'C');
836                         break;
837                 default:
838                         fprintf(stderr, "Bad pc action %x\n", t->action);
839                         ret = 1;
840                         break;
841         }
842         
843         return ret;
844 }
845
846 static void dump_trace_fs(struct blk_io_trace *t, struct per_cpu_info *pci)
847 {
848         int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
849         int act = t->action & 0xffff;
850
851         switch (act) {
852                 case __BLK_TA_QUEUE:
853                         account_q(t, pci, w);
854                         log_queue(pci, t, 'Q');
855                         break;
856                 case __BLK_TA_BACKMERGE:
857                         account_m(t, pci, w);
858                         log_merge(pci, t, 'M');
859                         break;
860                 case __BLK_TA_FRONTMERGE:
861                         account_m(t, pci, w);
862                         log_merge(pci, t, 'F');
863                         break;
864                 case __BLK_TA_GETRQ:
865                         log_track_getrq(t);
866                         log_generic(pci, t, 'G');
867                         break;
868                 case __BLK_TA_SLEEPRQ:
869                         log_generic(pci, t, 'S');
870                         break;
871                 case __BLK_TA_REQUEUE:
872                         account_c(t, pci, w, -t->bytes);
873                         log_queue(pci, t, 'R');
874                         break;
875                 case __BLK_TA_ISSUE:
876                         account_i(t, pci, w);
877                         log_issue(pci, t, 'D');
878                         break;
879                 case __BLK_TA_COMPLETE:
880                         account_c(t, pci, w, t->bytes);
881                         log_complete(pci, t, 'C');
882                         break;
883                 case __BLK_TA_PLUG:
884                         log_generic(pci, t, 'P');
885                         break;
886                 case __BLK_TA_UNPLUG:
887                         log_generic(pci, t, 'U');
888                         break;
889                 default:
890                         fprintf(stderr, "Bad fs action %x\n", t->action);
891                         break;
892         }
893 }
894
895 static int dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
896                         struct per_dev_info *pdi)
897 {
898         int ret = 0;
899
900         if (t->action & BLK_TC_ACT(BLK_TC_PC))
901                 ret = dump_trace_pc(t, pci);
902         else
903                 dump_trace_fs(t, pci);
904
905         pdi->events++;
906         return ret;
907 }
908
909 static void dump_io_stats(struct io_stats *ios, char *msg)
910 {
911         fprintf(ofp, "%s\n", msg);
912
913         fprintf(ofp, " Reads Queued:    %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
914         fprintf(ofp, " Writes Queued:    %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
915
916         fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
917         fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
918         fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
919         fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
920         fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
921
922         fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
923 }
924
925 static void dump_wait_stats(struct per_process_info *ppi)
926 {
927         unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
928         unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
929         unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
930         unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
931         unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
932         unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
933
934         fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
935         fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
936         fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
937         fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
938         fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
939         fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
940 }
941
942 static void show_process_stats(void)
943 {
944         struct per_process_info *ppi;
945
946         ppi = ppi_list;
947         while (ppi) {
948                 dump_io_stats(&ppi->io_stats, ppi->name);
949                 dump_wait_stats(ppi);
950                 ppi = ppi->list_next;
951         }
952
953         fprintf(ofp, "\n");
954 }
955
956 static void show_device_and_cpu_stats(void)
957 {
958         struct per_dev_info *pdi;
959         struct per_cpu_info *pci;
960         struct io_stats total, *ios;
961         int i, j, pci_events;
962         char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
963         char name[32];
964
965         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
966
967                 memset(&total, 0, sizeof(total));
968                 pci_events = 0;
969
970                 if (i > 0)
971                         fprintf(ofp, "\n");
972
973                 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
974                         if (!pci->nelems)
975                                 continue;
976
977                         ios = &pci->io_stats;
978                         total.qreads += ios->qreads;
979                         total.qwrites += ios->qwrites;
980                         total.creads += ios->creads;
981                         total.cwrites += ios->cwrites;
982                         total.mreads += ios->mreads;
983                         total.mwrites += ios->mwrites;
984                         total.ireads += ios->ireads;
985                         total.iwrites += ios->iwrites;
986                         total.qread_kb += ios->qread_kb;
987                         total.qwrite_kb += ios->qwrite_kb;
988                         total.cread_kb += ios->cread_kb;
989                         total.cwrite_kb += ios->cwrite_kb;
990                         total.iread_kb += ios->iread_kb;
991                         total.iwrite_kb += ios->iwrite_kb;
992
993                         snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
994                                  j, get_dev_name(pdi, name, sizeof(name)));
995                         dump_io_stats(ios, line);
996                         pci_events++;
997                 }
998
999                 if (pci_events > 1) {
1000                         fprintf(ofp, "\n");
1001                         snprintf(line, sizeof(line) - 1, "Total (%s):",
1002                                  get_dev_name(pdi, name, sizeof(name)));
1003                         dump_io_stats(&total, line);
1004                 }
1005
1006                 fprintf(ofp, "Events (%s): %'Lu\n",
1007                         get_dev_name(pdi, line, sizeof(line)), pdi->events);
1008         }
1009 }
1010
1011 static struct blk_io_trace *find_trace(void *p, unsigned long offset, int nr)
1012 {
1013         unsigned long max_offset = min(offset,nr * sizeof(struct blk_io_trace));
1014         unsigned long off;
1015         struct blk_io_trace *bit;
1016         __u32 magic;
1017
1018         for (off = 0; off < max_offset; off++) {
1019                 bit = p + off;
1020
1021                 magic = be32_to_cpu(bit->magic);
1022                 if ((magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
1023                         return bit;
1024         }
1025
1026         return NULL;
1027 }
1028
1029 static int sort_entries(void *traces, unsigned long offset, int nr,
1030                         struct per_dev_info *fpdi, struct per_cpu_info *fpci)
1031 {
1032         struct per_dev_info *pdi;
1033         struct per_cpu_info *pci;
1034         struct blk_io_trace *bit;
1035         struct trace *t;
1036         void *start = traces;
1037
1038         while (traces - start <= offset - sizeof(*bit)) {
1039                 if (!nr)
1040                         break;
1041
1042                 bit = find_trace(traces, offset - (traces - start), nr);
1043                 if (!bit)
1044                         break;
1045
1046                 t = malloc(sizeof(*t));
1047                 if (!t) {
1048                         fprintf(stderr, "Out of memory, seq %d on dev %d,%d\n",
1049                                 bit->sequence,
1050                                 MAJOR(bit->device), MINOR(bit->device));
1051                         return -1;
1052                 }
1053                 t->bit = bit;
1054                 memset(&t->rb_node, 0, sizeof(t->rb_node));
1055
1056                 trace_to_cpu(bit);
1057
1058                 if (verify_trace(bit)) {
1059                         free(t);
1060                         break;
1061                 }
1062
1063                 pdi = fpdi ? fpdi : get_dev_info(bit->device, 1);
1064                 pdi->id = bit->device;
1065                 pci = fpci ? fpci : get_cpu_info(pdi, bit->cpu);
1066                 pci->cpu = bit->cpu;
1067                 pci->nelems++;
1068
1069                 if (trace_rb_insert(t)) {
1070                         free(t);
1071                         return -1;
1072                 }
1073
1074                 traces += sizeof(*bit) + bit->pdu_len;
1075                 nr--;
1076         }
1077
1078         return 0;
1079 }
1080
1081 static void free_entries_rb(void)
1082 {
1083         struct rb_node *n;
1084
1085         while ((n = rb_first(&rb_sort_root)) != NULL) {
1086                 struct trace *t = rb_entry(n, struct trace, rb_node);
1087
1088                 rb_erase(&t->rb_node, &rb_sort_root);
1089                 free(t);
1090         }
1091 }
1092
1093 static void show_entries_rb(void)
1094 {
1095         struct per_dev_info *pdi;
1096         struct blk_io_trace *bit;
1097         struct rb_node *n;
1098         struct trace *t;
1099         int cpu;
1100
1101         n = rb_first(&rb_sort_root);
1102         if (!n)
1103                 return;
1104
1105         do {
1106                 t = rb_entry(n, struct trace, rb_node);
1107                 bit = t->bit;
1108
1109                 pdi = get_dev_info(bit->device, 0);
1110                 if (!pdi) {
1111                         fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1112                                 MAJOR(bit->device), MINOR(bit->device));
1113                         break;
1114                 }
1115                 cpu = bit->cpu;
1116                 if (cpu > pdi->ncpus) {
1117                         fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
1118                                 cpu, MAJOR(bit->device), MINOR(bit->device));
1119                         break;
1120                 }
1121
1122                 bit->time -= genesis_time;
1123                 if (bit->time < stopwatch_start)
1124                         continue;
1125                 if (bit->time >= stopwatch_end)
1126                         break;
1127
1128                 check_time(pdi, bit);
1129
1130                 if (dump_trace(bit, &pdi->cpus[cpu], pdi))
1131                         break;
1132
1133         } while ((n = rb_next(n)) != NULL);
1134 }
1135
1136 static int read_data(int fd, void *buffer, int bytes, int block)
1137 {
1138         int ret, bytes_left, fl;
1139         void *p;
1140
1141         fl = fcntl(fd, F_GETFL);
1142
1143         if (!block)
1144                 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1145         else
1146                 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1147
1148         bytes_left = bytes;
1149         p = buffer;
1150         while (bytes_left > 0) {
1151                 ret = read(fd, p, bytes_left);
1152                 if (!ret)
1153                         return 1;
1154                 else if (ret < 0) {
1155                         if (errno != EAGAIN)
1156                                 perror("read");
1157                         return -1;
1158                 } else {
1159                         p += ret;
1160                         bytes_left -= ret;
1161                 }
1162         }
1163
1164         return 0;
1165 }
1166
1167 static int do_file(void)
1168 {
1169         struct per_dev_info *pdi;
1170         int i, j, nfiles = 0;
1171
1172         for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1173                 for (j = 0;; j++, nfiles++) {
1174                         struct per_cpu_info *pci;
1175                         struct stat st;
1176                         void *tb;
1177
1178                         pci = get_cpu_info(pdi, j);
1179                         pci->cpu = j;
1180
1181                         snprintf(pci->fname, sizeof(pci->fname)-1,
1182                                  "%s_out.%d", pdi->name, j);
1183                         if (stat(pci->fname, &st) < 0)
1184                                 break;
1185                         if (!st.st_size)
1186                                 continue;
1187
1188                         printf("Processing %s\n", pci->fname);
1189
1190                         tb = malloc(st.st_size);
1191                         if (!tb) {
1192                                 fprintf(stderr, "Out of memory, skip file %s\n",
1193                                         pci->fname);
1194                                 continue;
1195                         }
1196
1197                         pci->fd = open(pci->fname, O_RDONLY);
1198                         if (pci->fd < 0) {
1199                                 perror(pci->fname);
1200                                 free(tb);
1201                                 continue;
1202                         }
1203
1204                         if (read_data(pci->fd, tb, st.st_size, 1)) {
1205                                 close(pci->fd);
1206                                 free(tb);
1207                                 continue;
1208                         }
1209
1210                         if (sort_entries(tb, st.st_size, ~0U, pdi, pci) == -1) {
1211                                 close(pci->fd);
1212                                 free(tb);
1213                                 continue;
1214                         }
1215
1216                         printf("Completed %s (CPU%d %d, entries)\n",
1217                                 pci->fname, j, pci->nelems);
1218                         close(pci->fd);
1219                 }
1220         }
1221
1222         if (!nfiles) {
1223                 fprintf(stderr, "No files found\n");
1224                 return 1;
1225         }
1226
1227         show_entries_rb();
1228         return 0;
1229 }
1230
1231 static void resize_buffer(void **buffer, long *size, long offset)
1232 {
1233         long old_size = *size;
1234
1235         if (*size == 0)
1236                 *size = 64 * sizeof(struct blk_io_trace);
1237
1238         *size *= 2;
1239         *buffer = realloc(*buffer, *size);
1240
1241         if (old_size)
1242                 memset(*buffer + offset, 0, *size - old_size);
1243 }
1244
1245 static int read_sort_events(int fd, void **buffer, long *max_offset)
1246 {
1247         long offset;
1248         int events;
1249
1250         events = offset = 0;
1251         do {
1252                 struct blk_io_trace *t;
1253                 int pdu_len;
1254                 __u32 magic;
1255
1256                 if (*max_offset - offset < sizeof(*t))
1257                         resize_buffer(buffer, max_offset, offset);
1258
1259                 if (read_data(fd, *buffer + offset, sizeof(*t), !events))
1260                         break;
1261
1262                 t = *buffer + offset;
1263                 offset += sizeof(*t);
1264
1265                 magic = be32_to_cpu(t->magic);
1266                 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1267                         fprintf(stderr, "Bad magic %x\n", magic);
1268                         break;
1269                 }
1270
1271                 pdu_len = be16_to_cpu(t->pdu_len);
1272                 if (pdu_len) {
1273                         if (*max_offset - offset <= pdu_len)
1274                                 resize_buffer(buffer, max_offset, offset);
1275
1276                         if (read_data(fd, *buffer + offset, pdu_len, 1))
1277                                 break;
1278
1279                         offset += pdu_len;
1280                 }
1281
1282                 events++;
1283         } while (!is_done() && events < rb_batch);
1284
1285         return events;
1286 }
1287
1288 static int do_stdin(void)
1289 {
1290         int fd;
1291         void *ptr = NULL;
1292         long max_offset;
1293
1294         fd = dup(STDIN_FILENO);
1295         max_offset = 0;
1296         do {
1297                 int events;
1298
1299                 events = read_sort_events(fd, &ptr, &max_offset);
1300                 if (!events)
1301                         break;
1302         
1303                 if (sort_entries(ptr, ~0UL, events, NULL, NULL) == -1)
1304                         break;
1305
1306                 show_entries_rb();
1307                 free_entries_rb();
1308         } while (1);
1309
1310         if (ptr)
1311                 free(ptr);
1312
1313         close(fd);
1314         return 0;
1315 }
1316
1317 static void flush_output(void)
1318 {
1319         fflush(ofp);
1320 }
1321
1322 static void handle_sigint(int sig)
1323 {
1324         done = 1;
1325         flush_output();
1326 }
1327
1328 /*
1329  * Extract start and duration times from a string, allowing
1330  * us to specify a time interval of interest within a trace.
1331  * Format: "duration" (start is zero) or "start:duration".
1332  */
1333 static int find_stopwatch_interval(char *string)
1334 {
1335         double value;
1336         char *sp;
1337
1338         value = strtod(string, &sp);
1339         if (sp == string) {
1340                 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1341                 return 1;
1342         }
1343         if (*sp == ':') {
1344                 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1345                 string = sp + 1;
1346                 value = strtod(string, &sp);
1347                 if (sp == string || *sp != '\0') {
1348                         fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1349                                 string);
1350                         return 1;
1351                 }
1352         } else if (*sp != '\0') {
1353                 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1354                 return 1;
1355         }
1356         stopwatch_end = stopwatch_start + DOUBLE_TO_NANO_ULL(value);
1357         return 0;
1358 }
1359
1360 static void usage(char *prog)
1361 {
1362         fprintf(stderr, "Usage: %s "
1363                 "[-i <name>] [-o <output>] [-s] [-w N[:n]] <name>...\n",
1364                 prog);
1365 }
1366
1367 int main(int argc, char *argv[])
1368 {
1369         char *ofp_buffer;
1370         int c, ret, mode;
1371         int per_device_and_cpu_stats = 1;
1372
1373         while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1374                 switch (c) {
1375                 case 'i':
1376                         if (!strcmp(optarg, "-") && !pipeline)
1377                                 pipeline = 1;
1378                         else if (resize_devices(optarg) != 0)
1379                                 return 1;
1380                         break;
1381                 case 'o':
1382                         output_name = optarg;
1383                         break;
1384                 case 'b':
1385                         rb_batch = atoi(optarg);
1386                         if (rb_batch <= 0)
1387                                 rb_batch = RB_BATCH_DEFAULT;
1388                         break;
1389                 case 's':
1390                         per_process_stats = 1;
1391                         break;
1392                 case 't':
1393                         track_ios = 1;
1394                         break;
1395                 case 'q':
1396                         per_device_and_cpu_stats = 0;
1397                         break;
1398                 case 'w':
1399                         if (find_stopwatch_interval(optarg) != 0)
1400                                 return 1;
1401                         break;
1402                 default:
1403                         usage(argv[0]);
1404                         return 1;
1405                 }
1406         }
1407
1408         while (optind < argc) {
1409                 if (!strcmp(argv[optind], "-") && !pipeline)
1410                         pipeline = 1;
1411                 else if (resize_devices(argv[optind]) != 0)
1412                         return 1;
1413                 optind++;
1414         }
1415
1416         if (!pipeline && !ndevices) {
1417                 usage(argv[0]);
1418                 return 1;
1419         }
1420
1421         memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1422         memset(&rb_track_root, 0, sizeof(rb_track_root));
1423
1424         signal(SIGINT, handle_sigint);
1425         signal(SIGHUP, handle_sigint);
1426         signal(SIGTERM, handle_sigint);
1427
1428         setlocale(LC_NUMERIC, "en_US");
1429
1430         if (!output_name) {
1431                 ofp = fdopen(STDOUT_FILENO, "w");
1432                 mode = _IOLBF;
1433         } else {
1434                 char ofname[128];
1435
1436                 snprintf(ofname, sizeof(ofname) - 1, "%s.log", output_name);
1437                 ofp = fopen(ofname, "w");
1438                 mode = _IOFBF;
1439         }
1440
1441         if (!ofp) {
1442                 perror("fopen");
1443                 return 1;
1444         }
1445
1446         ofp_buffer = malloc(4096);      
1447         if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1448                 perror("setvbuf");
1449                 return 1;
1450         }
1451
1452         if (pipeline)
1453                 ret = do_stdin();
1454         else
1455                 ret = do_file();
1456
1457         if (per_process_stats)
1458                 show_process_stats();
1459
1460         if (per_device_and_cpu_stats)
1461                 show_device_and_cpu_stats();
1462
1463         flush_output();
1464         return ret;
1465 }