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 "i:o:b:stqw:f:F:vnm"
85 static struct option l_opts[] = {
88 .has_arg = required_argument,
94 .has_arg = required_argument,
100 .has_arg = required_argument,
105 .name = "per program stats",
106 .has_arg = no_argument,
112 .has_arg = no_argument,
118 .has_arg = no_argument,
124 .has_arg = required_argument,
130 .has_arg = required_argument,
135 .name = "format-spec",
136 .has_arg = required_argument,
141 .name = "hash by name",
142 .has_arg = no_argument,
148 .has_arg = no_argument,
154 .has_arg = no_argument,
161 * for sorting the displayed output
164 struct blk_io_trace *bit;
165 struct rb_node rb_node;
169 static struct rb_root rb_sort_root;
170 static unsigned long rb_sort_entries;
172 static struct trace *trace_list;
177 static struct blk_io_trace *bit_alloc_list;
178 static struct trace *t_alloc_list;
181 * for tracking individual ios
184 struct rb_node rb_node;
189 unsigned long long allocation_time;
190 unsigned long long queue_time;
191 unsigned long long dispatch_time;
192 unsigned long long completion_time;
196 static struct per_dev_info *devices;
197 static char *get_dev_name(struct per_dev_info *, char *, int);
200 static char *output_name;
202 static unsigned long long genesis_time;
203 static unsigned long long last_allowed_time;
204 static unsigned int smallest_seq_read;
205 static unsigned long long stopwatch_start; /* start from zero by default */
206 static unsigned long long stopwatch_end = ULONG_LONG_MAX; /* "infinity" */
208 static int per_process_stats;
209 static int track_ios;
210 static int ppi_hash_by_pid = 1;
211 static int print_missing;
213 static unsigned int t_alloc_cache;
214 static unsigned int bit_alloc_cache;
216 #define RB_BATCH_DEFAULT (512)
217 static unsigned int rb_batch = RB_BATCH_DEFAULT;
221 #define is_done() (*(volatile int *)(&done))
222 static volatile int done;
224 #define JHASH_RANDOM (0x3af5f2ee)
226 static inline int ppi_hash_pid(__u32 pid)
228 return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
231 static inline int ppi_hash_name(const char *name)
233 return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
236 static inline int ppi_hash(struct per_process_info *ppi)
239 return ppi_hash_pid(ppi->pid);
241 return ppi_hash_name(ppi->name);
244 static inline void add_process_to_hash(struct per_process_info *ppi)
246 const int hash_idx = ppi_hash(ppi);
248 ppi->hash_next = ppi_hash_table[hash_idx];
249 ppi_hash_table[hash_idx] = ppi;
252 static inline void add_process_to_list(struct per_process_info *ppi)
254 ppi->list_next = ppi_list;
259 static struct per_process_info *find_process_by_name(char *name)
261 const int hash_idx = ppi_hash_name(name);
262 struct per_process_info *ppi;
264 ppi = ppi_hash_table[hash_idx];
266 if (!strcmp(ppi->name, name))
269 ppi = ppi->hash_next;
275 static struct per_process_info *find_process_by_pid(__u32 pid)
277 const int hash_idx = ppi_hash_pid(pid);
278 struct per_process_info *ppi;
280 ppi = ppi_hash_table[hash_idx];
285 ppi = ppi->hash_next;
291 static struct per_process_info *find_process(__u32 pid, char *name)
293 struct per_process_info *ppi;
296 return find_process_by_pid(pid);
298 ppi = find_process_by_name(name);
299 if (ppi && ppi->pid != pid)
300 ppi->more_than_one = 1;
305 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
308 struct rb_node **p = &root->rb_node;
309 struct rb_node *parent = NULL;
315 __t = rb_entry(parent, struct trace, rb_node);
318 if (t->bit->time < __t->bit->time) {
321 } else if (t->bit->time > __t->bit->time) {
326 if (t->bit->device < __t->bit->device)
328 else if (t->bit->device > __t->bit->device)
330 else if (t->bit->sequence < __t->bit->sequence)
332 else /* >= sequence */
336 rb_link_node(&t->rb_node, parent, p);
337 rb_insert_color(&t->rb_node, root);
341 static inline int trace_rb_insert_sort(struct trace *t)
343 if (!trace_rb_insert(t, &rb_sort_root, 1)) {
351 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
353 if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
354 pdi->rb_last_entries++;
361 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
362 struct rb_root *root, int order)
364 struct rb_node *n = root->rb_node;
365 struct rb_node *prev = NULL;
369 __t = rb_entry(n, struct trace, rb_node);
372 if (device < __t->bit->device)
374 else if (device > __t->bit->device)
376 else if (sequence < __t->bit->sequence)
378 else if (sequence > __t->bit->sequence)
385 * hack - the list may not be sequence ordered because some
386 * events don't have sequence and time matched. so we end up
387 * being a little off in the rb lookup here, because we don't
388 * know the time we are looking for. compensate by browsing
389 * a little ahead from the last entry to find the match
394 while (((n = rb_next(prev)) != NULL) && max--) {
395 __t = rb_entry(n, struct trace, rb_node);
397 if (__t->bit->device == device &&
398 __t->bit->sequence == sequence)
408 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
410 return trace_rb_find(dev, seq, &rb_sort_root, 1);
413 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
416 return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
419 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
421 struct rb_node **p = &pdi->rb_track.rb_node;
422 struct rb_node *parent = NULL;
423 struct io_track *__iot;
427 __iot = rb_entry(parent, struct io_track, rb_node);
429 if (iot->sector < __iot->sector)
431 else if (iot->sector > __iot->sector)
435 "sector alias (%Lu) on device %d,%d!\n",
436 (unsigned long long) iot->sector,
437 MAJOR(pdi->dev), MINOR(pdi->dev));
442 rb_link_node(&iot->rb_node, parent, p);
443 rb_insert_color(&iot->rb_node, &pdi->rb_track);
447 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
449 struct rb_node *n = pdi->rb_track.rb_node;
450 struct io_track *__iot;
453 __iot = rb_entry(n, struct io_track, rb_node);
455 if (sector < __iot->sector)
457 else if (sector > __iot->sector)
466 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
467 char *comm, __u64 sector)
469 struct io_track *iot;
471 iot = __find_track(pdi, sector);
473 iot = malloc(sizeof(*iot));
475 memcpy(iot->comm, comm, sizeof(iot->comm));
476 iot->sector = sector;
477 track_rb_insert(pdi, iot);
483 static void log_track_frontmerge(struct per_dev_info *pdi,
484 struct blk_io_trace *t)
486 struct io_track *iot;
491 iot = __find_track(pdi, t->sector + (t->bytes >> 9));
493 fprintf(stderr, "merge not found for (%d,%d): %llu\n",
494 MAJOR(pdi->dev), MINOR(pdi->dev),
495 (unsigned long long) t->sector + (t->bytes >> 9));
499 rb_erase(&iot->rb_node, &pdi->rb_track);
500 iot->sector -= t->bytes >> 9;
501 track_rb_insert(pdi, iot);
504 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
506 struct io_track *iot;
511 iot = find_track(pdi, t->pid, t->comm, t->sector);
512 iot->allocation_time = t->time;
516 * return time between rq allocation and insertion
518 static unsigned long long log_track_insert(struct per_dev_info *pdi,
519 struct blk_io_trace *t)
521 unsigned long long elapsed;
522 struct io_track *iot;
527 iot = find_track(pdi, t->pid, t->comm, t->sector);
528 iot->queue_time = t->time;
530 if (!iot->allocation_time)
533 elapsed = iot->queue_time - iot->allocation_time;
535 if (per_process_stats) {
536 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
537 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
539 if (ppi && elapsed > ppi->longest_allocation_wait[w])
540 ppi->longest_allocation_wait[w] = elapsed;
547 * return time between queue and issue
549 static unsigned long long log_track_issue(struct per_dev_info *pdi,
550 struct blk_io_trace *t)
552 unsigned long long elapsed;
553 struct io_track *iot;
557 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
560 iot = __find_track(pdi, t->sector);
562 fprintf(stderr, "issue not found for (%d,%d): %llu\n",
563 MAJOR(pdi->dev), MINOR(pdi->dev),
564 (unsigned long long) t->sector);
568 iot->dispatch_time = t->time;
569 elapsed = iot->dispatch_time - iot->queue_time;
571 if (per_process_stats) {
572 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
573 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
575 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
576 ppi->longest_dispatch_wait[w] = elapsed;
583 * return time between dispatch and complete
585 static unsigned long long log_track_complete(struct per_dev_info *pdi,
586 struct blk_io_trace *t)
588 unsigned long long elapsed;
589 struct io_track *iot;
593 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
596 iot = __find_track(pdi, t->sector);
598 fprintf(stderr, "complete not found for (%d,%d): %llu\n",
599 MAJOR(pdi->dev), MINOR(pdi->dev),
600 (unsigned long long) t->sector);
604 iot->completion_time = t->time;
605 elapsed = iot->completion_time - iot->dispatch_time;
607 if (per_process_stats) {
608 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
609 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
611 if (ppi && elapsed > ppi->longest_completion_wait[w])
612 ppi->longest_completion_wait[w] = elapsed;
616 * kill the trace, we don't need it after completion
618 rb_erase(&iot->rb_node, &pdi->rb_track);
625 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
627 struct per_process_info *ppi = find_process(pid, name);
630 ppi = malloc(sizeof(*ppi));
631 memset(ppi, 0, sizeof(*ppi));
632 memcpy(ppi->name, name, 16);
634 add_process_to_hash(ppi);
635 add_process_to_list(ppi);
638 return &ppi->io_stats;
641 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
643 struct per_cpu_info *cpus = pdi->cpus;
644 int ncpus = pdi->ncpus;
645 int new_count = cpu + 1;
649 size = new_count * sizeof(struct per_cpu_info);
650 cpus = realloc(cpus, size);
653 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
654 get_dev_name(pdi, name, sizeof(name)), size);
658 new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
659 new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
660 memset(new_start, 0, new_space);
662 pdi->ncpus = new_count;
666 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
668 struct per_cpu_info *pci;
670 if (cpu >= pdi->ncpus)
671 resize_cpu_info(pdi, cpu);
673 pci = &pdi->cpus[cpu];
679 static int resize_devices(char *name)
681 int size = (ndevices + 1) * sizeof(struct per_dev_info);
683 devices = realloc(devices, size);
685 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
688 memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
689 devices[ndevices].name = name;
694 static struct per_dev_info *get_dev_info(dev_t dev)
696 struct per_dev_info *pdi;
699 for (i = 0; i < ndevices; i++) {
701 devices[i].dev = dev;
702 if (devices[i].dev == dev)
706 if (resize_devices(NULL))
709 pdi = &devices[ndevices - 1];
711 pdi->last_sequence = -1;
712 pdi->last_read_time = 0;
713 memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
714 pdi->rb_last_entries = 0;
718 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
721 snprintf(buffer, size, "%s", pdi->name);
723 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
727 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
729 unsigned long long this = bit->time;
730 unsigned long long last = pdi->last_reported_time;
732 pdi->backwards = (this < last) ? 'B' : ' ';
733 pdi->last_reported_time = this;
736 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
741 ios->qwrite_kb += t->bytes >> 10;
744 ios->qread_kb += t->bytes >> 10;
748 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
751 __account_m(&pci->io_stats, t, rw);
753 if (per_process_stats) {
754 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
756 __account_m(ios, t, rw);
760 static inline void __account_queue(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_queue(struct blk_io_trace *t,
773 struct per_cpu_info *pci, int rw)
775 __account_queue(&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_queue(ios, t, rw);
784 static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
788 ios->cwrite_kb += bytes >> 10;
791 ios->cread_kb += bytes >> 10;
795 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
798 __account_c(&pci->io_stats, rw, bytes);
800 if (per_process_stats) {
801 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
803 __account_c(ios, rw, bytes);
807 static inline void __account_issue(struct io_stats *ios, int rw,
812 ios->iwrite_kb += bytes >> 10;
815 ios->iread_kb += bytes >> 10;
819 static inline void account_issue(struct blk_io_trace *t,
820 struct per_cpu_info *pci, int rw)
822 __account_issue(&pci->io_stats, rw, t->bytes);
824 if (per_process_stats) {
825 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
827 __account_issue(ios, rw, t->bytes);
831 static inline void __account_unplug(struct io_stats *ios, int timer)
834 ios->timer_unplugs++;
839 static inline void account_unplug(struct blk_io_trace *t,
840 struct per_cpu_info *pci, int timer)
842 __account_unplug(&pci->io_stats, timer);
844 if (per_process_stats) {
845 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
847 __account_unplug(ios, timer);
851 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
852 struct blk_io_trace *t, char *act)
854 process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
857 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
858 struct blk_io_trace *t, char *act)
860 process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
863 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
866 process_fmt(act, pci, t, -1, 0, NULL);
869 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
870 struct blk_io_trace *t, char *act)
872 process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
875 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
876 struct blk_io_trace *t, char *act)
879 log_track_frontmerge(pdi, t);
881 process_fmt(act, pci, t, -1ULL, 0, NULL);
884 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
887 process_fmt(act, pci, t, -1ULL, 0, NULL);
890 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
893 process_fmt(act, pci, t, -1ULL, 0, NULL);
896 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
899 process_fmt(act, pci, t, -1ULL, 0, NULL);
902 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
905 process_fmt(act, pci, t, -1ULL, 0, NULL);
908 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
910 unsigned char *buf = (unsigned char *) t + sizeof(*t);
912 process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
915 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
917 int act = t->action & 0xffff;
921 log_generic(pci, t, "Q");
924 log_generic(pci, t, "G");
926 case __BLK_TA_SLEEPRQ:
927 log_generic(pci, t, "S");
929 case __BLK_TA_REQUEUE:
930 log_generic(pci, t, "R");
935 case __BLK_TA_COMPLETE:
938 case __BLK_TA_INSERT:
942 fprintf(stderr, "Bad pc action %x\n", act);
947 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
948 struct per_cpu_info *pci)
950 int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
951 int act = t->action & 0xffff;
955 account_queue(t, pci, w);
956 log_queue(pci, t, "Q");
958 case __BLK_TA_INSERT:
959 log_insert(pdi, pci, t, "I");
961 case __BLK_TA_BACKMERGE:
962 account_m(t, pci, w);
963 log_merge(pdi, pci, t, "M");
965 case __BLK_TA_FRONTMERGE:
966 account_m(t, pci, w);
967 log_merge(pdi, pci, t, "F");
970 log_track_getrq(pdi, t);
971 log_generic(pci, t, "G");
973 case __BLK_TA_SLEEPRQ:
974 log_generic(pci, t, "S");
976 case __BLK_TA_REQUEUE:
977 account_c(t, pci, w, -t->bytes);
978 log_queue(pci, t, "R");
981 account_issue(t, pci, w);
982 log_issue(pdi, pci, t, "D");
984 case __BLK_TA_COMPLETE:
985 account_c(t, pci, w, t->bytes);
986 log_complete(pdi, pci, t, "C");
989 log_action(pci, t, "P");
991 case __BLK_TA_UNPLUG_IO:
992 account_unplug(t, pci, 0);
993 log_unplug(pci, t, "U");
995 case __BLK_TA_UNPLUG_TIMER:
996 account_unplug(t, pci, 1);
997 log_unplug(pci, t, "UT");
1000 log_split(pci, t, "X");
1002 case __BLK_TA_BOUNCE:
1003 log_generic(pci, t, "B");
1006 fprintf(stderr, "Bad fs action %x\n", t->action);
1011 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1012 struct per_dev_info *pdi)
1014 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1015 dump_trace_pc(t, pci);
1017 dump_trace_fs(t, pdi, pci);
1022 static void dump_io_stats(struct io_stats *ios, char *msg)
1024 fprintf(ofp, "%s\n", msg);
1026 fprintf(ofp, " Reads Queued: %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
1027 fprintf(ofp, " Writes Queued: %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
1029 fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
1030 fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
1031 fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
1032 fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
1033 fprintf(ofp, " Read Merges: %'8lu%8c\t", ios->mreads, ' ');
1034 fprintf(ofp, " Write Merges: %'8lu\n", ios->mwrites);
1035 fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' ');
1036 fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs);
1039 static void dump_wait_stats(struct per_process_info *ppi)
1041 unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1042 unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1043 unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1044 unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1045 unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1046 unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1048 fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1049 fprintf(ofp, " Allocation wait: %'8lu\n", wawait);
1050 fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' ');
1051 fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait);
1052 fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1053 fprintf(ofp, " Completion wait: %'8lu\n", wcwait);
1056 static int ppi_name_compare(const void *p1, const void *p2)
1058 struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1059 struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1062 res = strverscmp(ppi1->name, ppi2->name);
1064 res = ppi1->pid > ppi2->pid;
1069 static void sort_process_list(void)
1071 struct per_process_info **ppis;
1072 struct per_process_info *ppi;
1075 ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1080 ppi = ppi->list_next;
1083 qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1085 i = ppi_list_entries - 1;
1090 ppi->list_next = ppi_list;
1098 static void show_process_stats(void)
1100 struct per_process_info *ppi;
1102 sort_process_list();
1108 if (ppi->more_than_one)
1109 sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1111 sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1113 dump_io_stats(&ppi->io_stats, name);
1114 dump_wait_stats(ppi);
1115 ppi = ppi->list_next;
1121 static void show_device_and_cpu_stats(void)
1123 struct per_dev_info *pdi;
1124 struct per_cpu_info *pci;
1125 struct io_stats total, *ios;
1126 int i, j, pci_events;
1127 char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1130 for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1132 memset(&total, 0, sizeof(total));
1138 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1142 ios = &pci->io_stats;
1143 total.qreads += ios->qreads;
1144 total.qwrites += ios->qwrites;
1145 total.creads += ios->creads;
1146 total.cwrites += ios->cwrites;
1147 total.mreads += ios->mreads;
1148 total.mwrites += ios->mwrites;
1149 total.ireads += ios->ireads;
1150 total.iwrites += ios->iwrites;
1151 total.qread_kb += ios->qread_kb;
1152 total.qwrite_kb += ios->qwrite_kb;
1153 total.cread_kb += ios->cread_kb;
1154 total.cwrite_kb += ios->cwrite_kb;
1155 total.iread_kb += ios->iread_kb;
1156 total.iwrite_kb += ios->iwrite_kb;
1157 total.timer_unplugs += ios->timer_unplugs;
1158 total.io_unplugs += ios->io_unplugs;
1160 snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1161 j, get_dev_name(pdi, name, sizeof(name)));
1162 dump_io_stats(ios, line);
1166 if (pci_events > 1) {
1168 snprintf(line, sizeof(line) - 1, "Total (%s):",
1169 get_dev_name(pdi, name, sizeof(name)));
1170 dump_io_stats(&total, line);
1173 fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
1174 get_dev_name(pdi, line, sizeof(line)), pdi->events,
1180 * struct trace and blktrace allocation cache, we do potentially
1181 * millions of mallocs for these structures while only using at most
1182 * a few thousand at the time
1184 static inline void t_free(struct trace *t)
1186 if (t_alloc_cache < 1024) {
1187 t->next = t_alloc_list;
1194 static inline struct trace *t_alloc(void)
1196 struct trace *t = t_alloc_list;
1199 t_alloc_list = t->next;
1204 return malloc(sizeof(*t));
1207 static inline void bit_free(struct blk_io_trace *bit)
1209 if (bit_alloc_cache < 1024) {
1211 * abuse a 64-bit field for a next pointer for the free item
1213 bit->time = (__u64) (unsigned long) bit_alloc_list;
1214 bit_alloc_list = (struct blk_io_trace *) bit;
1220 static inline struct blk_io_trace *bit_alloc(void)
1222 struct blk_io_trace *bit = bit_alloc_list;
1225 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1231 return malloc(sizeof(*bit));
1234 static void find_genesis(void)
1236 struct trace *t = trace_list;
1238 genesis_time = -1ULL;
1240 if (t->bit->time < genesis_time)
1241 genesis_time = t->bit->time;
1247 static inline int check_stopwatch(struct blk_io_trace *bit)
1249 if (bit->time < stopwatch_end &&
1250 bit->time >= stopwatch_start)
1257 * return youngest entry read
1259 static int sort_entries(unsigned long long *youngest)
1267 while ((t = trace_list) != NULL) {
1268 struct blk_io_trace *bit = t->bit;
1270 trace_list = t->next;
1272 bit->time -= genesis_time;
1274 if (bit->time < *youngest || !*youngest)
1275 *youngest = bit->time;
1277 if (check_stopwatch(bit)) {
1283 if (trace_rb_insert_sort(t))
1286 if (bit->sequence < smallest_seq_read)
1287 smallest_seq_read = bit->sequence;
1293 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1295 rb_erase(&t->rb_node, &pdi->rb_last);
1296 pdi->rb_last_entries--;
1302 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1304 rb_erase(&t->rb_node, &rb_sort_root);
1307 trace_rb_insert_last(pdi, t);
1309 if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1310 struct rb_node *n = rb_first(&pdi->rb_last);
1312 t = rb_entry(n, struct trace, rb_node);
1313 __put_trace_last(pdi, t);
1317 static int check_sequence(struct per_dev_info *pdi, struct blk_io_trace *bit)
1319 unsigned long expected_sequence = pdi->last_sequence + 1;
1323 * first entry, always ok
1325 if (!expected_sequence)
1328 if (bit->sequence == expected_sequence)
1332 * we may not have seen that sequence yet. if we are not doing
1333 * the final run, break and wait for more entries.
1335 if (expected_sequence < smallest_seq_read) {
1336 t = trace_rb_find_last(pdi, expected_sequence);
1340 __put_trace_last(pdi, t);
1344 if (print_missing) {
1345 fprintf(stderr, "(%d,%d): skipping %lu -> %u\n",
1346 MAJOR(pdi->dev), MINOR(pdi->dev),
1347 pdi->last_sequence, bit->sequence);
1354 static void show_entries_rb(int force)
1356 struct per_dev_info *pdi = NULL;
1357 struct per_cpu_info *pci = NULL;
1358 struct blk_io_trace *bit;
1362 while ((n = rb_first(&rb_sort_root)) != NULL) {
1363 if (is_done() && !force)
1366 t = rb_entry(n, struct trace, rb_node);
1369 if (!pdi || pdi->dev != bit->device)
1370 pdi = get_dev_info(bit->device);
1373 fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1374 MAJOR(bit->device), MINOR(bit->device));
1379 if (check_sequence(pdi, bit))
1382 if (bit->time > last_allowed_time)
1386 pdi->last_sequence = bit->sequence;
1388 check_time(pdi, bit);
1390 if (!pci || pci->cpu != bit->cpu)
1391 pci = get_cpu_info(pdi, bit->cpu);
1393 dump_trace(bit, pci, pdi);
1399 static int read_data(int fd, void *buffer, int bytes, int block)
1401 int ret, bytes_left, fl;
1404 fl = fcntl(fd, F_GETFL);
1407 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1409 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1413 while (bytes_left > 0) {
1414 ret = read(fd, p, bytes_left);
1418 if (errno != EAGAIN)
1431 static int read_events(int fd, int always_block)
1433 struct per_dev_info *pdi = NULL;
1434 unsigned int events = 0;
1436 while (!is_done() && events < rb_batch) {
1437 struct blk_io_trace *bit;
1444 if (read_data(fd, bit, sizeof(*bit), !events || always_block))
1447 magic = be32_to_cpu(bit->magic);
1448 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1449 fprintf(stderr, "Bad magic %x\n", magic);
1453 pdu_len = be16_to_cpu(bit->pdu_len);
1455 void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1457 if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
1465 if (verify_trace(bit)) {
1471 memset(t, 0, sizeof(*t));
1474 t->next = trace_list;
1477 if (!pdi || pdi->dev != bit->device)
1478 pdi = get_dev_info(bit->device);
1480 if (bit->time > pdi->last_read_time)
1481 pdi->last_read_time = bit->time;
1489 static int do_file(void)
1491 struct per_cpu_info *pci;
1492 struct per_dev_info *pdi;
1493 int i, j, events, events_added;
1496 * first prepare all files for reading
1498 for (i = 0; i < ndevices; i++) {
1501 pdi->last_sequence = -1;
1506 pci = get_cpu_info(pdi, j);
1510 snprintf(pci->fname, sizeof(pci->fname)-1,
1511 "%s.blktrace.%d", pdi->name, pci->cpu);
1512 if (stat(pci->fname, &st) < 0)
1515 pci->fd = open(pci->fname, O_RDONLY);
1522 printf("Input file %s added\n", pci->fname);
1528 * now loop over the files reading in the data
1531 unsigned long long youngest;
1534 last_allowed_time = -1ULL;
1535 smallest_seq_read = -1U;
1537 for (i = 0; i < ndevices; i++) {
1540 for (j = 0; j < pdi->nfiles; j++) {
1542 pci = get_cpu_info(pdi, j);
1547 events = read_events(pci->fd, 1);
1554 if (pdi->last_read_time < last_allowed_time)
1555 last_allowed_time = pdi->last_read_time;
1557 events_added += events;
1561 if (sort_entries(&youngest))
1564 if (youngest > stopwatch_end)
1569 } while (events_added);
1571 if (rb_sort_entries)
1577 static int do_stdin(void)
1579 unsigned long long youngest;
1580 int fd, events, loops;
1582 last_allowed_time = -1ULL;
1583 fd = dup(STDIN_FILENO);
1585 perror("dup stdin");
1590 while ((events = read_events(fd, 0)) != 0) {
1592 smallest_seq_read = -1U;
1594 if (sort_entries(&youngest))
1597 if (youngest > stopwatch_end)
1604 if (rb_sort_entries)
1611 static void flush_output(void)
1616 static void handle_sigint(__attribute__((__unused__)) int sig)
1623 * Extract start and duration times from a string, allowing
1624 * us to specify a time interval of interest within a trace.
1625 * Format: "duration" (start is zero) or "start:duration".
1627 static int find_stopwatch_interval(char *string)
1632 value = strtod(string, &sp);
1634 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1638 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1640 value = strtod(string, &sp);
1641 if (sp == string || *sp != '\0') {
1642 fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1646 } else if (*sp != '\0') {
1647 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
1650 stopwatch_end = DOUBLE_TO_NANO_ULL(value);
1651 if (stopwatch_end <= stopwatch_start) {
1652 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
1653 stopwatch_start, stopwatch_end);
1660 static char usage_str[] = \
1661 "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
1662 "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
1663 "\t-i Input file containing trace data, or '-' for stdin\n" \
1664 "\t-o Output file. If not given, output is stdout\n" \
1665 "\t-b stdin read batching\n" \
1666 "\t-s Show per-program io statistics\n" \
1667 "\t-n Hash processes by name, not pid\n" \
1668 "\t-t Track individual ios. Will tell you the time a request took\n" \
1669 "\t to get queued, to get dispatched, and to get completed\n" \
1670 "\t-q Quiet. Don't display any stats at the end of the trace\n" \
1671 "\t-w Only parse data between the given time interval in seconds.\n" \
1672 "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \
1673 "\t -f Output format. Customize the output format. The format field\n" \
1674 "\t identifies can be found in the documentation\n" \
1675 "\t-F Format specification. Can be found in the documentation\n" \
1676 "\t-m Print missing entries\n" \
1677 "\t-v Print program version info\n\n";
1679 static void usage(char *prog)
1681 fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
1684 int main(int argc, char *argv[])
1688 int per_device_and_cpu_stats = 1;
1690 while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
1693 if (!strcmp(optarg, "-") && !pipeline)
1695 else if (resize_devices(optarg) != 0)
1699 output_name = optarg;
1702 rb_batch = atoi(optarg);
1704 rb_batch = RB_BATCH_DEFAULT;
1707 per_process_stats = 1;
1713 per_device_and_cpu_stats = 0;
1716 if (find_stopwatch_interval(optarg) != 0)
1720 set_all_format_specs(optarg);
1723 if (add_format_spec(optarg) != 0)
1727 ppi_hash_by_pid = 0;
1733 printf("%s version %s\n", argv[0], blkparse_version);
1741 while (optind < argc) {
1742 if (!strcmp(argv[optind], "-") && !pipeline)
1744 else if (resize_devices(argv[optind]) != 0)
1749 if (!pipeline && !ndevices) {
1754 memset(&rb_sort_root, 0, sizeof(rb_sort_root));
1756 signal(SIGINT, handle_sigint);
1757 signal(SIGHUP, handle_sigint);
1758 signal(SIGTERM, handle_sigint);
1760 setlocale(LC_NUMERIC, "en_US");
1763 ofp = fdopen(STDOUT_FILENO, "w");
1768 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
1769 ofp = fopen(ofname, "w");
1778 ofp_buffer = malloc(4096);
1779 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
1789 if (per_process_stats)
1790 show_process_stats();
1792 if (per_device_and_cpu_stats)
1793 show_device_and_cpu_stats();