2 * block queue tracing parse application
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
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.
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.
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
21 #include <sys/types.h>
38 static char blkparse_version[] = "0.90";
45 unsigned long long events;
46 unsigned long long last_reported_time;
47 unsigned long long last_read_time;
48 struct io_stats io_stats;
49 unsigned long last_sequence;
52 struct rb_root rb_last;
53 unsigned long rb_last_entries;
55 struct rb_root rb_track;
59 struct per_cpu_info *cpus;
62 struct per_process_info {
65 struct io_stats io_stats;
66 struct per_process_info *hash_next, *list_next;
72 unsigned long long longest_allocation_wait[2];
73 unsigned long long longest_dispatch_wait[2];
74 unsigned long long longest_completion_wait[2];
77 #define PPI_HASH_SHIFT (8)
78 #define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT)
79 #define PPI_HASH_MASK (PPI_HASH_SIZE - 1)
80 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
81 static struct per_process_info *ppi_list;
82 static int ppi_list_entries;
84 #define S_OPTS "a:A:i:o:b:stqw:f:F:vnmD:"
85 static struct option l_opts[] = {
88 .has_arg = required_argument,
94 .has_arg = required_argument,
100 .has_arg = required_argument,
106 .has_arg = required_argument,
112 .has_arg = required_argument,
117 .name = "per-program-stats",
118 .has_arg = no_argument,
124 .has_arg = no_argument,
130 .has_arg = no_argument,
136 .has_arg = required_argument,
142 .has_arg = required_argument,
147 .name = "format-spec",
148 .has_arg = required_argument,
153 .name = "hash-by-name",
154 .has_arg = no_argument,
160 .has_arg = no_argument,
166 .has_arg = no_argument,
171 .name = "input-directory",
172 .has_arg = required_argument,
182 * for sorting the displayed output
185 struct blk_io_trace *bit;
186 struct rb_node rb_node;
191 static struct rb_root rb_sort_root;
192 static unsigned long rb_sort_entries;
194 static struct trace *trace_list;
199 static struct blk_io_trace *bit_alloc_list;
200 static struct trace *t_alloc_list;
203 * for tracking individual ios
206 struct rb_node rb_node;
211 unsigned long long allocation_time;
212 unsigned long long queue_time;
213 unsigned long long dispatch_time;
214 unsigned long long completion_time;
218 static struct per_dev_info *devices;
219 static char *get_dev_name(struct per_dev_info *, char *, int);
222 static char *output_name;
223 static char *input_dir;
225 static unsigned long long genesis_time;
226 static unsigned long long last_allowed_time;
227 static unsigned int smallest_seq_read;
228 static unsigned long long stopwatch_start; /* start from zero by default */
229 static unsigned long long stopwatch_end = ULONG_LONG_MAX; /* "infinity" */
231 static int per_process_stats;
232 static int track_ios;
233 static int ppi_hash_by_pid = 1;
234 static int print_missing;
235 static unsigned int act_mask = -1U;
237 static unsigned int t_alloc_cache;
238 static unsigned int bit_alloc_cache;
240 #define RB_BATCH_DEFAULT (512)
241 static unsigned int rb_batch = RB_BATCH_DEFAULT;
245 #define is_done() (*(volatile int *)(&done))
246 static volatile int done;
248 #define JHASH_RANDOM (0x3af5f2ee)
250 static inline int ppi_hash_pid(__u32 pid)
252 return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
255 static inline int ppi_hash_name(const char *name)
257 return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
260 static inline int ppi_hash(struct per_process_info *ppi)
263 return ppi_hash_pid(ppi->pid);
265 return ppi_hash_name(ppi->name);
268 static inline void add_process_to_hash(struct per_process_info *ppi)
270 const int hash_idx = ppi_hash(ppi);
272 ppi->hash_next = ppi_hash_table[hash_idx];
273 ppi_hash_table[hash_idx] = ppi;
276 static inline void add_process_to_list(struct per_process_info *ppi)
278 ppi->list_next = ppi_list;
283 static struct per_process_info *find_process_by_name(char *name)
285 const int hash_idx = ppi_hash_name(name);
286 struct per_process_info *ppi;
288 ppi = ppi_hash_table[hash_idx];
290 if (!strcmp(ppi->name, name))
293 ppi = ppi->hash_next;
299 static struct per_process_info *find_process_by_pid(__u32 pid)
301 const int hash_idx = ppi_hash_pid(pid);
302 struct per_process_info *ppi;
304 ppi = ppi_hash_table[hash_idx];
309 ppi = ppi->hash_next;
315 static struct per_process_info *find_process(__u32 pid, char *name)
317 struct per_process_info *ppi;
320 return find_process_by_pid(pid);
322 ppi = find_process_by_name(name);
323 if (ppi && ppi->pid != pid)
324 ppi->more_than_one = 1;
329 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
332 struct rb_node **p = &root->rb_node;
333 struct rb_node *parent = NULL;
339 __t = rb_entry(parent, struct trace, rb_node);
342 if (t->bit->time < __t->bit->time) {
345 } else if (t->bit->time > __t->bit->time) {
350 if (t->bit->device < __t->bit->device)
352 else if (t->bit->device > __t->bit->device)
354 else if (t->bit->sequence < __t->bit->sequence)
356 else /* >= sequence */
360 rb_link_node(&t->rb_node, parent, p);
361 rb_insert_color(&t->rb_node, root);
365 static inline int trace_rb_insert_sort(struct trace *t)
367 if (!trace_rb_insert(t, &rb_sort_root, 1)) {
375 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
377 if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
378 pdi->rb_last_entries++;
385 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
386 struct rb_root *root, int order)
388 struct rb_node *n = root->rb_node;
389 struct rb_node *prev = NULL;
393 __t = rb_entry(n, struct trace, rb_node);
396 if (device < __t->bit->device)
398 else if (device > __t->bit->device)
400 else if (sequence < __t->bit->sequence)
402 else if (sequence > __t->bit->sequence)
409 * hack - the list may not be sequence ordered because some
410 * events don't have sequence and time matched. so we end up
411 * being a little off in the rb lookup here, because we don't
412 * know the time we are looking for. compensate by browsing
413 * a little ahead from the last entry to find the match
418 while (((n = rb_next(prev)) != NULL) && max--) {
419 __t = rb_entry(n, struct trace, rb_node);
421 if (__t->bit->device == device &&
422 __t->bit->sequence == sequence)
432 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
434 return trace_rb_find(dev, seq, &rb_sort_root, 1);
437 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
440 return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
443 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
445 struct rb_node **p = &pdi->rb_track.rb_node;
446 struct rb_node *parent = NULL;
447 struct io_track *__iot;
451 __iot = rb_entry(parent, struct io_track, rb_node);
453 if (iot->sector < __iot->sector)
455 else if (iot->sector > __iot->sector)
459 "sector alias (%Lu) on device %d,%d!\n",
460 (unsigned long long) iot->sector,
461 MAJOR(pdi->dev), MINOR(pdi->dev));
466 rb_link_node(&iot->rb_node, parent, p);
467 rb_insert_color(&iot->rb_node, &pdi->rb_track);
471 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
473 struct rb_node *n = pdi->rb_track.rb_node;
474 struct io_track *__iot;
477 __iot = rb_entry(n, struct io_track, rb_node);
479 if (sector < __iot->sector)
481 else if (sector > __iot->sector)
490 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
491 char *comm, __u64 sector)
493 struct io_track *iot;
495 iot = __find_track(pdi, sector);
497 iot = malloc(sizeof(*iot));
499 memcpy(iot->comm, comm, sizeof(iot->comm));
500 iot->sector = sector;
501 track_rb_insert(pdi, iot);
507 static void log_track_frontmerge(struct per_dev_info *pdi,
508 struct blk_io_trace *t)
510 struct io_track *iot;
515 iot = __find_track(pdi, t->sector + (t->bytes >> 9));
517 fprintf(stderr, "merge not found for (%d,%d): %llu\n",
518 MAJOR(pdi->dev), MINOR(pdi->dev),
519 (unsigned long long) t->sector + (t->bytes >> 9));
523 rb_erase(&iot->rb_node, &pdi->rb_track);
524 iot->sector -= t->bytes >> 9;
525 track_rb_insert(pdi, iot);
528 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
530 struct io_track *iot;
535 iot = find_track(pdi, t->pid, t->comm, t->sector);
536 iot->allocation_time = t->time;
540 * return time between rq allocation and insertion
542 static unsigned long long log_track_insert(struct per_dev_info *pdi,
543 struct blk_io_trace *t)
545 unsigned long long elapsed;
546 struct io_track *iot;
551 iot = find_track(pdi, t->pid, t->comm, t->sector);
552 iot->queue_time = t->time;
554 if (!iot->allocation_time)
557 elapsed = iot->queue_time - iot->allocation_time;
559 if (per_process_stats) {
560 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
561 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
563 if (ppi && elapsed > ppi->longest_allocation_wait[w])
564 ppi->longest_allocation_wait[w] = elapsed;
571 * return time between queue and issue
573 static unsigned long long log_track_issue(struct per_dev_info *pdi,
574 struct blk_io_trace *t)
576 unsigned long long elapsed;
577 struct io_track *iot;
581 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
584 iot = __find_track(pdi, t->sector);
586 fprintf(stderr, "issue not found for (%d,%d): %llu\n",
587 MAJOR(pdi->dev), MINOR(pdi->dev),
588 (unsigned long long) t->sector);
592 iot->dispatch_time = t->time;
593 elapsed = iot->dispatch_time - iot->queue_time;
595 if (per_process_stats) {
596 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
597 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
599 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
600 ppi->longest_dispatch_wait[w] = elapsed;
607 * return time between dispatch and complete
609 static unsigned long long log_track_complete(struct per_dev_info *pdi,
610 struct blk_io_trace *t)
612 unsigned long long elapsed;
613 struct io_track *iot;
617 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
620 iot = __find_track(pdi, t->sector);
622 fprintf(stderr, "complete not found for (%d,%d): %llu\n",
623 MAJOR(pdi->dev), MINOR(pdi->dev),
624 (unsigned long long) t->sector);
628 iot->completion_time = t->time;
629 elapsed = iot->completion_time - iot->dispatch_time;
631 if (per_process_stats) {
632 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
633 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
635 if (ppi && elapsed > ppi->longest_completion_wait[w])
636 ppi->longest_completion_wait[w] = elapsed;
640 * kill the trace, we don't need it after completion
642 rb_erase(&iot->rb_node, &pdi->rb_track);
649 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
651 struct per_process_info *ppi = find_process(pid, name);
654 ppi = malloc(sizeof(*ppi));
655 memset(ppi, 0, sizeof(*ppi));
656 memcpy(ppi->name, name, 16);
658 add_process_to_hash(ppi);
659 add_process_to_list(ppi);
662 return &ppi->io_stats;
665 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
667 struct per_cpu_info *cpus = pdi->cpus;
668 int ncpus = pdi->ncpus;
669 int new_count = cpu + 1;
673 size = new_count * sizeof(struct per_cpu_info);
674 cpus = realloc(cpus, size);
677 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
678 get_dev_name(pdi, name, sizeof(name)), size);
682 new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
683 new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
684 memset(new_start, 0, new_space);
686 pdi->ncpus = new_count;
690 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
692 struct per_cpu_info *pci;
694 if (cpu >= pdi->ncpus)
695 resize_cpu_info(pdi, cpu);
697 pci = &pdi->cpus[cpu];
703 static int resize_devices(char *name)
705 int size = (ndevices + 1) * sizeof(struct per_dev_info);
707 devices = realloc(devices, size);
709 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
712 memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
713 devices[ndevices].name = name;
718 static struct per_dev_info *get_dev_info(dev_t dev)
720 struct per_dev_info *pdi;
723 for (i = 0; i < ndevices; i++) {
725 devices[i].dev = dev;
726 if (devices[i].dev == dev)
730 if (resize_devices(NULL))
733 pdi = &devices[ndevices - 1];
735 pdi->last_sequence = -1;
736 pdi->last_read_time = 0;
737 memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
738 pdi->rb_last_entries = 0;
742 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
745 snprintf(buffer, size, "%s", pdi->name);
747 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
751 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
753 unsigned long long this = bit->time;
754 unsigned long long last = pdi->last_reported_time;
756 pdi->backwards = (this < last) ? 'B' : ' ';
757 pdi->last_reported_time = this;
760 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
765 ios->qwrite_kb += t->bytes >> 10;
768 ios->qread_kb += t->bytes >> 10;
772 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
775 __account_m(&pci->io_stats, t, rw);
777 if (per_process_stats) {
778 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
780 __account_m(ios, t, rw);
784 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
789 ios->qwrite_kb += t->bytes >> 10;
792 ios->qread_kb += t->bytes >> 10;
796 static inline void account_queue(struct blk_io_trace *t,
797 struct per_cpu_info *pci, int rw)
799 __account_queue(&pci->io_stats, t, rw);
801 if (per_process_stats) {
802 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
804 __account_queue(ios, t, rw);
808 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
812 ios->cwrite_kb += bytes >> 10;
815 ios->cread_kb += bytes >> 10;
819 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
822 __account_c(&pci->io_stats, rw, bytes);
824 if (per_process_stats) {
825 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
827 __account_c(ios, rw, bytes);
831 static inline void __account_issue(struct io_stats *ios, int rw,
836 ios->iwrite_kb += bytes >> 10;
839 ios->iread_kb += bytes >> 10;
843 static inline void account_issue(struct blk_io_trace *t,
844 struct per_cpu_info *pci, int rw)
846 __account_issue(&pci->io_stats, rw, t->bytes);
848 if (per_process_stats) {
849 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
851 __account_issue(ios, rw, t->bytes);
855 static inline void __account_unplug(struct io_stats *ios, int timer)
858 ios->timer_unplugs++;
863 static inline void account_unplug(struct blk_io_trace *t,
864 struct per_cpu_info *pci, int timer)
866 __account_unplug(&pci->io_stats, timer);
868 if (per_process_stats) {
869 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
871 __account_unplug(ios, timer);
875 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
876 struct blk_io_trace *t, char *act)
878 process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
881 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
882 struct blk_io_trace *t, char *act)
884 process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
887 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
890 process_fmt(act, pci, t, -1, 0, NULL);
893 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
894 struct blk_io_trace *t, char *act)
896 process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
899 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
900 struct blk_io_trace *t, char *act)
903 log_track_frontmerge(pdi, t);
905 process_fmt(act, pci, t, -1ULL, 0, NULL);
908 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
911 process_fmt(act, pci, t, -1ULL, 0, NULL);
914 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
917 process_fmt(act, pci, t, -1ULL, 0, NULL);
920 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
923 process_fmt(act, pci, t, -1ULL, 0, NULL);
926 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
929 process_fmt(act, pci, t, -1ULL, 0, NULL);
932 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
934 unsigned char *buf = (unsigned char *) t + sizeof(*t);
936 process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
939 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
941 int act = t->action & 0xffff;
945 log_generic(pci, t, "Q");
948 log_generic(pci, t, "G");
950 case __BLK_TA_SLEEPRQ:
951 log_generic(pci, t, "S");
953 case __BLK_TA_REQUEUE:
954 log_generic(pci, t, "R");
959 case __BLK_TA_COMPLETE:
962 case __BLK_TA_INSERT:
966 fprintf(stderr, "Bad pc action %x\n", act);
971 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
972 struct per_cpu_info *pci)
974 int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
975 int act = t->action & 0xffff;
979 account_queue(t, pci, w);
980 log_queue(pci, t, "Q");
982 case __BLK_TA_INSERT:
983 log_insert(pdi, pci, t, "I");
985 case __BLK_TA_BACKMERGE:
986 account_m(t, pci, w);
987 log_merge(pdi, pci, t, "M");
989 case __BLK_TA_FRONTMERGE:
990 account_m(t, pci, w);
991 log_merge(pdi, pci, t, "F");
994 log_track_getrq(pdi, t);
995 log_generic(pci, t, "G");
997 case __BLK_TA_SLEEPRQ:
998 log_generic(pci, t, "S");
1000 case __BLK_TA_REQUEUE:
1001 account_c(t, pci, w, -t->bytes);
1002 log_queue(pci, t, "R");
1004 case __BLK_TA_ISSUE:
1005 account_issue(t, pci, w);
1006 log_issue(pdi, pci, t, "D");
1008 case __BLK_TA_COMPLETE:
1009 account_c(t, pci, w, t->bytes);
1010 log_complete(pdi, pci, t, "C");
1013 log_action(pci, t, "P");
1015 case __BLK_TA_UNPLUG_IO:
1016 account_unplug(t, pci, 0);
1017 log_unplug(pci, t, "U");
1019 case __BLK_TA_UNPLUG_TIMER:
1020 account_unplug(t, pci, 1);
1021 log_unplug(pci, t, "UT");
1023 case __BLK_TA_SPLIT:
1024 log_split(pci, t, "X");
1026 case __BLK_TA_BOUNCE:
1027 log_generic(pci, t, "B");
1030 fprintf(stderr, "Bad fs action %x\n", t->action);
1035 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1036 struct per_dev_info *pdi)
1038 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1039 dump_trace_pc(t, pci);
1041 dump_trace_fs(t, pdi, pci);
1046 static void dump_io_stats(struct io_stats *ios, char *msg)
1048 fprintf(ofp, "%s\n", msg);
1050 fprintf(ofp, " Reads Queued: %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
1051 fprintf(ofp, " Writes Queued: %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
1053 fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
1054 fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
1055 fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
1056 fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
1057 fprintf(ofp, " Read Merges: %'8lu%8c\t", ios->mreads, ' ');
1058 fprintf(ofp, " Write Merges: %'8lu\n", ios->mwrites);
1059 fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' ');
1060 fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs);
1063 static void dump_wait_stats(struct per_process_info *ppi)
1065 unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1066 unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1067 unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1068 unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1069 unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1070 unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1072 fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1073 fprintf(ofp, " Allocation wait: %'8lu\n", wawait);
1074 fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' ');
1075 fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait);
1076 fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1077 fprintf(ofp, " Completion wait: %'8lu\n", wcwait);
1080 static int ppi_name_compare(const void *p1, const void *p2)
1082 struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1083 struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1086 res = strverscmp(ppi1->name, ppi2->name);
1088 res = ppi1->pid > ppi2->pid;
1093 static void sort_process_list(void)
1095 struct per_process_info **ppis;
1096 struct per_process_info *ppi;
1099 ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1104 ppi = ppi->list_next;
1107 qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1109 i = ppi_list_entries - 1;
1114 ppi->list_next = ppi_list;
1122 static void show_process_stats(void)
1124 struct per_process_info *ppi;
1126 sort_process_list();
1132 if (ppi->more_than_one)
1133 sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1135 sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1137 dump_io_stats(&ppi->io_stats, name);
1138 dump_wait_stats(ppi);
1139 ppi = ppi->list_next;
1145 static void show_device_and_cpu_stats(void)
1147 struct per_dev_info *pdi;
1148 struct per_cpu_info *pci;
1149 struct io_stats total, *ios;
1150 int i, j, pci_events;
1151 char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1154 for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1156 memset(&total, 0, sizeof(total));
1162 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1166 ios = &pci->io_stats;
1167 total.qreads += ios->qreads;
1168 total.qwrites += ios->qwrites;
1169 total.creads += ios->creads;
1170 total.cwrites += ios->cwrites;
1171 total.mreads += ios->mreads;
1172 total.mwrites += ios->mwrites;
1173 total.ireads += ios->ireads;
1174 total.iwrites += ios->iwrites;
1175 total.qread_kb += ios->qread_kb;
1176 total.qwrite_kb += ios->qwrite_kb;
1177 total.cread_kb += ios->cread_kb;
1178 total.cwrite_kb += ios->cwrite_kb;
1179 total.iread_kb += ios->iread_kb;
1180 total.iwrite_kb += ios->iwrite_kb;
1181 total.timer_unplugs += ios->timer_unplugs;
1182 total.io_unplugs += ios->io_unplugs;
1184 snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1185 j, get_dev_name(pdi, name, sizeof(name)));
1186 dump_io_stats(ios, line);
1190 if (pci_events > 1) {
1192 snprintf(line, sizeof(line) - 1, "Total (%s):",
1193 get_dev_name(pdi, name, sizeof(name)));
1194 dump_io_stats(&total, line);
1197 fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
1198 get_dev_name(pdi, line, sizeof(line)), pdi->events,
1204 * struct trace and blktrace allocation cache, we do potentially
1205 * millions of mallocs for these structures while only using at most
1206 * a few thousand at the time
1208 static inline void t_free(struct trace *t)
1210 if (t_alloc_cache < 1024) {
1211 t->next = t_alloc_list;
1218 static inline struct trace *t_alloc(void)
1220 struct trace *t = t_alloc_list;
1223 t_alloc_list = t->next;
1228 return malloc(sizeof(*t));
1231 static inline void bit_free(struct blk_io_trace *bit)
1233 if (bit_alloc_cache < 1024) {
1235 * abuse a 64-bit field for a next pointer for the free item
1237 bit->time = (__u64) (unsigned long) bit_alloc_list;
1238 bit_alloc_list = (struct blk_io_trace *) bit;
1244 static inline struct blk_io_trace *bit_alloc(void)
1246 struct blk_io_trace *bit = bit_alloc_list;
1249 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1255 return malloc(sizeof(*bit));
1258 static void find_genesis(void)
1260 struct trace *t = trace_list;
1262 genesis_time = -1ULL;
1264 if (t->bit->time < genesis_time)
1265 genesis_time = t->bit->time;
1271 static inline int check_stopwatch(struct blk_io_trace *bit)
1273 if (bit->time < stopwatch_end &&
1274 bit->time >= stopwatch_start)
1281 * return youngest entry read
1283 static int sort_entries(unsigned long long *youngest)
1291 while ((t = trace_list) != NULL) {
1292 struct blk_io_trace *bit = t->bit;
1294 trace_list = t->next;
1296 bit->time -= genesis_time;
1298 if (bit->time < *youngest || !*youngest)
1299 *youngest = bit->time;
1301 if (check_stopwatch(bit)) {
1307 if (trace_rb_insert_sort(t))
1310 if (bit->sequence < smallest_seq_read)
1311 smallest_seq_read = bit->sequence;
1317 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1319 rb_erase(&t->rb_node, &pdi->rb_last);
1320 pdi->rb_last_entries--;
1326 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1328 rb_erase(&t->rb_node, &rb_sort_root);
1331 trace_rb_insert_last(pdi, t);
1333 if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1334 struct rb_node *n = rb_first(&pdi->rb_last);
1336 t = rb_entry(n, struct trace, rb_node);
1337 __put_trace_last(pdi, t);
1341 static int check_sequence(struct per_dev_info *pdi, struct trace *t)
1343 unsigned long expected_sequence = pdi->last_sequence + 1;
1344 struct blk_io_trace *bit = t->bit;
1348 * first entry, always ok
1350 if (!expected_sequence)
1353 if (bit->sequence == expected_sequence)
1357 * we may not have seen that sequence yet. if we are not doing
1358 * the final run, break and wait for more entries.
1360 if (expected_sequence < smallest_seq_read) {
1361 __t = trace_rb_find_last(pdi, expected_sequence);
1365 __put_trace_last(pdi, __t);
1374 if (print_missing) {
1375 fprintf(stderr, "(%d,%d): skipping %lu -> %u\n",
1376 MAJOR(pdi->dev), MINOR(pdi->dev),
1377 pdi->last_sequence, bit->sequence);
1384 static void show_entries_rb(int force)
1386 struct per_dev_info *pdi = NULL;
1387 struct per_cpu_info *pci = NULL;
1388 struct blk_io_trace *bit;
1392 while ((n = rb_first(&rb_sort_root)) != NULL) {
1393 if (is_done() && !force && !pipeline)
1396 t = rb_entry(n, struct trace, rb_node);
1399 if (!pdi || pdi->dev != bit->device)
1400 pdi = get_dev_info(bit->device);
1403 fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1404 MAJOR(bit->device), MINOR(bit->device));
1409 if (check_sequence(pdi, t))
1412 if (bit->time > last_allowed_time)
1416 pdi->last_sequence = bit->sequence;
1418 check_time(pdi, bit);
1420 if (!pci || pci->cpu != bit->cpu)
1421 pci = get_cpu_info(pdi, bit->cpu);
1423 if (bit->action & (act_mask << BLK_TC_SHIFT))
1424 dump_trace(bit, pci, pdi);
1430 static int read_data(int fd, void *buffer, int bytes, int block)
1432 int ret, bytes_left, fl;
1435 fl = fcntl(fd, F_GETFL);
1438 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1440 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1444 while (bytes_left > 0) {
1445 ret = read(fd, p, bytes_left);
1449 if (errno != EAGAIN)
1462 static int read_events(int fd, int always_block)
1464 struct per_dev_info *pdi = NULL;
1465 unsigned int events = 0;
1467 while (!is_done() && events < rb_batch) {
1468 struct blk_io_trace *bit;
1475 if (read_data(fd, bit, sizeof(*bit), !events || always_block))
1478 magic = be32_to_cpu(bit->magic);
1479 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1480 fprintf(stderr, "Bad magic %x\n", magic);
1484 pdu_len = be16_to_cpu(bit->pdu_len);
1486 void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1488 if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
1496 if (verify_trace(bit)) {
1502 memset(t, 0, sizeof(*t));
1505 t->next = trace_list;
1508 if (!pdi || pdi->dev != bit->device)
1509 pdi = get_dev_info(bit->device);
1511 if (bit->time > pdi->last_read_time)
1512 pdi->last_read_time = bit->time;
1520 static int do_file(void)
1522 struct per_cpu_info *pci;
1523 struct per_dev_info *pdi;
1524 int i, j, events, events_added;
1527 * first prepare all files for reading
1529 for (i = 0; i < ndevices; i++) {
1532 pdi->last_sequence = -1;
1538 pci = get_cpu_info(pdi, j);
1543 len = sprintf(pci->fname, "%s/", input_dir);
1545 snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
1546 "%s.blktrace.%d", pdi->name, pci->cpu);
1547 if (stat(pci->fname, &st) < 0)
1550 pci->fd = open(pci->fname, O_RDONLY);
1557 printf("Input file %s added\n", pci->fname);
1563 * now loop over the files reading in the data
1566 unsigned long long youngest;
1569 last_allowed_time = -1ULL;
1570 smallest_seq_read = -1U;
1572 for (i = 0; i < ndevices; i++) {
1575 for (j = 0; j < pdi->nfiles; j++) {
1577 pci = get_cpu_info(pdi, j);
1582 events = read_events(pci->fd, 1);
1589 if (pdi->last_read_time < last_allowed_time)
1590 last_allowed_time = pdi->last_read_time;
1592 events_added += events;
1596 if (sort_entries(&youngest))
1599 if (youngest > stopwatch_end)
1604 } while (events_added);
1606 if (rb_sort_entries)
1612 static int do_stdin(void)
1614 unsigned long long youngest;
1615 int fd, events, loops;
1617 last_allowed_time = -1ULL;
1618 fd = dup(STDIN_FILENO);
1620 perror("dup stdin");
1625 while ((events = read_events(fd, 0)) != 0) {
1627 smallest_seq_read = -1U;
1629 if (sort_entries(&youngest))
1632 if (youngest > stopwatch_end)
1639 if (rb_sort_entries)
1646 static void flush_output(void)
1651 static void handle_sigint(__attribute__((__unused__)) int sig)
1658 * Extract start and duration times from a string, allowing
1659 * us to specify a time interval of interest within a trace.
1660 * Format: "duration" (start is zero) or "start:duration".
1662 static int find_stopwatch_interval(char *string)
1667 value = strtod(string, &sp);
1669 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1673 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1675 value = strtod(string, &sp);
1676 if (sp == string || *sp != '\0') {
1677 fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1681 } else if (*sp != '\0') {
1682 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1685 stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1686 if (stopwatch_end <= stopwatch_start) {
1687 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1688 stopwatch_start, stopwatch_end);
1695 static char usage_str[] = \
1696 "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1697 "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1698 "\t-i Input file containing trace data, or '-' for stdin\n" \
1699 "\t-D Directory to prepend to input file names\n" \
1700 "\t-o Output file. If not given, output is stdout\n" \
1701 "\t-b stdin read batching\n" \
1702 "\t-s Show per-program io statistics\n" \
1703 "\t-n Hash processes by name, not pid\n" \
1704 "\t-t Track individual ios. Will tell you the time a request took\n" \
1705 "\t to get queued, to get dispatched, and to get completed\n" \
1706 "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1707 "\t-w Only parse data between the given time interval in seconds.\n" \
1708 "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \
1709 "\t -f Output format. Customize the output format. The format field\n" \
1710 "\t identifies can be found in the documentation\n" \
1711 "\t-F Format specification. Can be found in the documentation\n" \
1712 "\t-m Print missing entries\n" \
1713 "\t-v Print program version info\n\n";
1715 static void usage(char *prog)
1717 fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1720 int main(int argc, char *argv[])
1723 int i, c, ret, mode;
1724 int per_device_and_cpu_stats = 1;
1725 int act_mask_tmp = 0;
1727 while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1730 i = find_mask_map(optarg);
1732 fprintf(stderr,"Invalid action mask %s\n",
1740 if ((sscanf(optarg, "%x", &i) != 1) ||
1741 !valid_act_opt(i)) {
1743 "Invalid set action mask %s/0x%x\n",
1750 if (!strcmp(optarg, "-") && !pipeline)
1752 else if (resize_devices(optarg) != 0)
1759 output_name = optarg;
1762 rb_batch = atoi(optarg);
1764 rb_batch = RB_BATCH_DEFAULT;
1767 per_process_stats = 1;
1773 per_device_and_cpu_stats = 0;
1776 if (find_stopwatch_interval(optarg) != 0)
1780 set_all_format_specs(optarg);
1783 if (add_format_spec(optarg) != 0)
1787 ppi_hash_by_pid = 0;
1793 printf("%s version %s\n", argv[0], blkparse_version);
1801 while (optind < argc) {
1802 if (!strcmp(argv[optind], "-") && !pipeline)
1804 else if (resize_devices(argv[optind]) != 0)
1809 if (!pipeline && !ndevices) {
1814 if (act_mask_tmp != 0)
1815 act_mask = act_mask_tmp;
1817 memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1819 signal(SIGINT, handle_sigint);
1820 signal(SIGHUP, handle_sigint);
1821 signal(SIGTERM, handle_sigint);
1823 setlocale(LC_NUMERIC, "en_US");
1826 ofp = fdopen(STDOUT_FILENO, "w");
1831 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
1832 ofp = fopen(ofname, "w");
1841 ofp_buffer = malloc(4096);
1842 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1852 if (per_process_stats)
1853 show_process_stats();
1855 if (per_device_and_cpu_stats)
1856 show_device_and_cpu_stats();