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[] = "0.99";
41 unsigned long start, end;
42 struct skip_info *prev, *next;
50 unsigned long long events;
51 unsigned long long first_reported_time;
52 unsigned long long last_reported_time;
53 unsigned long long last_read_time;
54 struct io_stats io_stats;
55 unsigned long last_sequence;
56 unsigned long skips, nskips;
57 unsigned long long seq_skips, seq_nskips;
58 unsigned int max_depth[2];
59 unsigned int cur_depth[2];
61 struct rb_root rb_last;
62 unsigned long rb_last_entries;
64 struct rb_root rb_track;
69 unsigned long *cpu_map;
70 unsigned int cpu_map_max;
72 struct per_cpu_info *cpus;
73 struct skip_info *skips_head;
74 struct skip_info *skips_tail;
77 struct per_process_info {
80 struct io_stats io_stats;
81 struct per_process_info *hash_next, *list_next;
87 unsigned long long longest_allocation_wait[2];
88 unsigned long long longest_dispatch_wait[2];
89 unsigned long long longest_completion_wait[2];
92 #define PPI_HASH_SHIFT (8)
93 #define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT)
94 #define PPI_HASH_MASK (PPI_HASH_SIZE - 1)
95 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
96 static struct per_process_info *ppi_list;
97 static int ppi_list_entries;
99 #define S_OPTS "a:A:i:o:b:stqw:f:F:vVhD:"
100 static struct option l_opts[] = {
103 .has_arg = required_argument,
109 .has_arg = required_argument,
115 .has_arg = required_argument,
121 .has_arg = required_argument,
127 .has_arg = required_argument,
132 .name = "per-program-stats",
133 .has_arg = no_argument,
139 .has_arg = no_argument,
145 .has_arg = no_argument,
151 .has_arg = required_argument,
157 .has_arg = required_argument,
162 .name = "format-spec",
163 .has_arg = required_argument,
168 .name = "hash-by-name",
169 .has_arg = no_argument,
175 .has_arg = no_argument,
181 .has_arg = no_argument,
186 .name = "input-directory",
187 .has_arg = required_argument,
197 * for sorting the displayed output
200 struct blk_io_trace *bit;
201 struct rb_node rb_node;
205 static struct rb_root rb_sort_root;
206 static unsigned long rb_sort_entries;
208 static struct trace *trace_list;
213 static struct blk_io_trace *bit_alloc_list;
214 static struct trace *t_alloc_list;
217 * for tracking individual ios
220 struct rb_node rb_node;
225 unsigned long long allocation_time;
226 unsigned long long queue_time;
227 unsigned long long dispatch_time;
228 unsigned long long completion_time;
232 static struct per_dev_info *devices;
233 static char *get_dev_name(struct per_dev_info *, char *, int);
236 static char *output_name;
237 static char *input_dir;
239 static unsigned long long genesis_time;
240 static unsigned long long last_allowed_time;
241 static unsigned int smallest_seq_read;
242 static unsigned long long stopwatch_start; /* start from zero by default */
243 static unsigned long long stopwatch_end = -1ULL; /* "infinity" */
245 static int per_process_stats;
246 static int per_device_and_cpu_stats = 1;
247 static int track_ios;
248 static int ppi_hash_by_pid = 1;
250 static unsigned int act_mask = -1U;
251 static int stats_printed;
253 static unsigned int t_alloc_cache;
254 static unsigned int bit_alloc_cache;
256 #define RB_BATCH_DEFAULT (512)
257 static unsigned int rb_batch = RB_BATCH_DEFAULT;
261 #define is_done() (*(volatile int *)(&done))
262 static volatile int done;
264 #define JHASH_RANDOM (0x3af5f2ee)
266 #define CPUS_PER_LONG (8 * sizeof(unsigned long))
267 #define CPU_IDX(cpu) ((cpu) / CPUS_PER_LONG)
268 #define CPU_BIT(cpu) ((cpu) & (CPUS_PER_LONG - 1))
270 static void insert_skip(struct per_dev_info *pdi, unsigned long start,
273 struct skip_info *sip;
275 for (sip = pdi->skips_tail; sip != NULL; sip = sip->prev) {
276 if (end == (sip->start - 1)) {
279 } else if (start == (sip->end + 1)) {
285 sip = malloc(sizeof(struct skip_info));
288 sip->prev = sip->next = NULL;
289 if (pdi->skips_tail == NULL)
290 pdi->skips_head = pdi->skips_tail = sip;
292 sip->prev = pdi->skips_tail;
293 pdi->skips_tail->next = sip;
294 pdi->skips_tail = sip;
298 static void remove_sip(struct per_dev_info *pdi, struct skip_info *sip)
300 if (sip->prev == NULL) {
301 if (sip->next == NULL)
302 pdi->skips_head = pdi->skips_tail = NULL;
304 pdi->skips_head = sip->next;
305 sip->next->prev = NULL;
307 } else if (sip->next == NULL) {
308 pdi->skips_tail = sip->prev;
309 sip->prev->next = NULL;
311 sip->prev->next = sip->next;
312 sip->next->prev = sip->prev;
315 sip->prev = sip->next = NULL;
319 #define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
320 static int check_current_skips(struct per_dev_info *pdi, unsigned long seq)
322 struct skip_info *sip;
324 for (sip = pdi->skips_tail; sip != NULL; sip = sip->prev) {
325 if (IN_SKIP(sip,seq)) {
326 if (sip->start == seq) {
331 } else if (sip->end == seq)
335 insert_skip(pdi,seq+1,sip->end);
343 static void collect_pdi_skips(struct per_dev_info *pdi)
345 struct skip_info *sip;
349 for (sip = pdi->skips_head; sip != NULL; sip = sip->next) {
351 pdi->seq_skips += (sip->end - sip->start + 1);
353 fprintf(stderr, "(%d,%d): skipping %lu -> %lu\n",
354 MAJOR(pdi->dev), MINOR(pdi->dev),
355 sip->start, sip->end);
359 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
361 if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
362 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
363 unsigned long *map = malloc(new_max / sizeof(long));
365 memset(map, 0, new_max / sizeof(long));
368 memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
373 pdi->cpu_map_max = new_max;
376 pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
379 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
381 pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
384 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
386 return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
389 static inline int ppi_hash_pid(__u32 pid)
391 return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
394 static inline int ppi_hash_name(const char *name)
396 return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
399 static inline int ppi_hash(struct per_process_info *ppi)
402 return ppi_hash_pid(ppi->pid);
404 return ppi_hash_name(ppi->name);
407 static inline void add_process_to_hash(struct per_process_info *ppi)
409 const int hash_idx = ppi_hash(ppi);
411 ppi->hash_next = ppi_hash_table[hash_idx];
412 ppi_hash_table[hash_idx] = ppi;
415 static inline void add_process_to_list(struct per_process_info *ppi)
417 ppi->list_next = ppi_list;
422 static struct per_process_info *find_process_by_name(char *name)
424 const int hash_idx = ppi_hash_name(name);
425 struct per_process_info *ppi;
427 ppi = ppi_hash_table[hash_idx];
429 if (!strcmp(ppi->name, name))
432 ppi = ppi->hash_next;
438 static struct per_process_info *find_process_by_pid(__u32 pid)
440 const int hash_idx = ppi_hash_pid(pid);
441 struct per_process_info *ppi;
443 ppi = ppi_hash_table[hash_idx];
448 ppi = ppi->hash_next;
454 static struct per_process_info *find_process(__u32 pid, char *name)
456 struct per_process_info *ppi;
459 return find_process_by_pid(pid);
461 ppi = find_process_by_name(name);
462 if (ppi && ppi->pid != pid)
463 ppi->more_than_one = 1;
468 static inline int trace_rb_insert(struct trace *t, struct rb_root *root,
471 struct rb_node **p = &root->rb_node;
472 struct rb_node *parent = NULL;
478 __t = rb_entry(parent, struct trace, rb_node);
481 if (t->bit->time < __t->bit->time) {
484 } else if (t->bit->time > __t->bit->time) {
489 if (t->bit->device < __t->bit->device)
491 else if (t->bit->device > __t->bit->device)
493 else if (t->bit->sequence < __t->bit->sequence)
495 else /* >= sequence */
499 rb_link_node(&t->rb_node, parent, p);
500 rb_insert_color(&t->rb_node, root);
504 static inline int trace_rb_insert_sort(struct trace *t)
506 if (!trace_rb_insert(t, &rb_sort_root, 1)) {
514 static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t)
516 if (!trace_rb_insert(t, &pdi->rb_last, 1)) {
517 pdi->rb_last_entries++;
524 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
525 struct rb_root *root, int order)
527 struct rb_node *n = root->rb_node;
528 struct rb_node *prev = NULL;
532 __t = rb_entry(n, struct trace, rb_node);
535 if (device < __t->bit->device)
537 else if (device > __t->bit->device)
539 else if (sequence < __t->bit->sequence)
541 else if (sequence > __t->bit->sequence)
548 * hack - the list may not be sequence ordered because some
549 * events don't have sequence and time matched. so we end up
550 * being a little off in the rb lookup here, because we don't
551 * know the time we are looking for. compensate by browsing
552 * a little ahead from the last entry to find the match
557 while (((n = rb_next(prev)) != NULL) && max--) {
558 __t = rb_entry(n, struct trace, rb_node);
560 if (__t->bit->device == device &&
561 __t->bit->sequence == sequence)
571 static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq)
573 return trace_rb_find(dev, seq, &rb_sort_root, 1);
576 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
579 return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0);
582 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
584 struct rb_node **p = &pdi->rb_track.rb_node;
585 struct rb_node *parent = NULL;
586 struct io_track *__iot;
590 __iot = rb_entry(parent, struct io_track, rb_node);
592 if (iot->sector < __iot->sector)
594 else if (iot->sector > __iot->sector)
598 "sector alias (%Lu) on device %d,%d!\n",
599 (unsigned long long) iot->sector,
600 MAJOR(pdi->dev), MINOR(pdi->dev));
605 rb_link_node(&iot->rb_node, parent, p);
606 rb_insert_color(&iot->rb_node, &pdi->rb_track);
610 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
612 struct rb_node *n = pdi->rb_track.rb_node;
613 struct io_track *__iot;
616 __iot = rb_entry(n, struct io_track, rb_node);
618 if (sector < __iot->sector)
620 else if (sector > __iot->sector)
629 static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid,
630 char *comm, __u64 sector)
632 struct io_track *iot;
634 iot = __find_track(pdi, sector);
636 iot = malloc(sizeof(*iot));
638 memcpy(iot->comm, comm, sizeof(iot->comm));
639 iot->sector = sector;
640 track_rb_insert(pdi, iot);
646 static void log_track_frontmerge(struct per_dev_info *pdi,
647 struct blk_io_trace *t)
649 struct io_track *iot;
654 iot = __find_track(pdi, t->sector + t_sec(t));
657 fprintf(stderr, "merge not found for (%d,%d): %llu\n",
658 MAJOR(pdi->dev), MINOR(pdi->dev),
659 (unsigned long long) t->sector + t_sec(t));
663 rb_erase(&iot->rb_node, &pdi->rb_track);
664 iot->sector -= t_sec(t);
665 track_rb_insert(pdi, iot);
668 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
670 struct io_track *iot;
675 iot = find_track(pdi, t->pid, t->comm, t->sector);
676 iot->allocation_time = t->time;
680 * return time between rq allocation and insertion
682 static unsigned long long log_track_insert(struct per_dev_info *pdi,
683 struct blk_io_trace *t)
685 unsigned long long elapsed;
686 struct io_track *iot;
691 iot = find_track(pdi, t->pid, t->comm, t->sector);
692 iot->queue_time = t->time;
694 if (!iot->allocation_time)
697 elapsed = iot->queue_time - iot->allocation_time;
699 if (per_process_stats) {
700 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
701 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
703 if (ppi && elapsed > ppi->longest_allocation_wait[w])
704 ppi->longest_allocation_wait[w] = elapsed;
711 * return time between queue and issue
713 static unsigned long long log_track_issue(struct per_dev_info *pdi,
714 struct blk_io_trace *t)
716 unsigned long long elapsed;
717 struct io_track *iot;
721 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
724 iot = __find_track(pdi, t->sector);
727 fprintf(stderr, "issue not found for (%d,%d): %llu\n",
728 MAJOR(pdi->dev), MINOR(pdi->dev),
729 (unsigned long long) t->sector);
733 iot->dispatch_time = t->time;
734 elapsed = iot->dispatch_time - iot->queue_time;
736 if (per_process_stats) {
737 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
738 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
740 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
741 ppi->longest_dispatch_wait[w] = elapsed;
748 * return time between dispatch and complete
750 static unsigned long long log_track_complete(struct per_dev_info *pdi,
751 struct blk_io_trace *t)
753 unsigned long long elapsed;
754 struct io_track *iot;
758 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
761 iot = __find_track(pdi, t->sector);
764 fprintf(stderr,"complete not found for (%d,%d): %llu\n",
765 MAJOR(pdi->dev), MINOR(pdi->dev),
766 (unsigned long long) t->sector);
770 iot->completion_time = t->time;
771 elapsed = iot->completion_time - iot->dispatch_time;
773 if (per_process_stats) {
774 struct per_process_info *ppi = find_process(iot->pid,iot->comm);
775 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
777 if (ppi && elapsed > ppi->longest_completion_wait[w])
778 ppi->longest_completion_wait[w] = elapsed;
782 * kill the trace, we don't need it after completion
784 rb_erase(&iot->rb_node, &pdi->rb_track);
791 static struct io_stats *find_process_io_stats(__u32 pid, char *name)
793 struct per_process_info *ppi = find_process(pid, name);
796 ppi = malloc(sizeof(*ppi));
797 memset(ppi, 0, sizeof(*ppi));
798 memcpy(ppi->name, name, 16);
800 add_process_to_hash(ppi);
801 add_process_to_list(ppi);
804 return &ppi->io_stats;
807 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
809 struct per_cpu_info *cpus = pdi->cpus;
810 int ncpus = pdi->ncpus;
811 int new_count = cpu + 1;
815 size = new_count * sizeof(struct per_cpu_info);
816 cpus = realloc(cpus, size);
819 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
820 get_dev_name(pdi, name, sizeof(name)), size);
824 new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
825 new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
826 memset(new_start, 0, new_space);
828 pdi->ncpus = new_count;
831 for (new_count = 0; new_count < pdi->ncpus; new_count++)
832 if (!pdi->cpus[new_count].fd)
833 pdi->cpus[new_count].fd = -1;
836 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
838 struct per_cpu_info *pci;
840 if (cpu >= pdi->ncpus)
841 resize_cpu_info(pdi, cpu);
843 pci = &pdi->cpus[cpu];
849 static int resize_devices(char *name)
851 int size = (ndevices + 1) * sizeof(struct per_dev_info);
853 devices = realloc(devices, size);
855 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
858 memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
859 devices[ndevices].name = name;
864 static struct per_dev_info *get_dev_info(dev_t dev)
866 struct per_dev_info *pdi;
869 for (i = 0; i < ndevices; i++) {
871 devices[i].dev = dev;
872 if (devices[i].dev == dev)
876 if (resize_devices(NULL))
879 pdi = &devices[ndevices - 1];
881 pdi->first_reported_time = 0;
882 pdi->last_sequence = -1;
883 pdi->last_read_time = 0;
884 memset(&pdi->rb_last, 0, sizeof(pdi->rb_last));
885 pdi->rb_last_entries = 0;
887 pdi->skips_head = pdi->skips_tail = NULL;
892 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
895 snprintf(buffer, size, "%s", pdi->name);
897 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
901 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
903 unsigned long long this = bit->time;
904 unsigned long long last = pdi->last_reported_time;
906 pdi->backwards = (this < last) ? 'B' : ' ';
907 pdi->last_reported_time = this;
910 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
915 ios->qwrite_kb += t_kb(t);
918 ios->qread_kb += t_kb(t);
922 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
925 __account_m(&pci->io_stats, t, rw);
927 if (per_process_stats) {
928 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
930 __account_m(ios, t, rw);
934 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
939 ios->qwrite_kb += t_kb(t);
942 ios->qread_kb += t_kb(t);
946 static inline void account_queue(struct blk_io_trace *t,
947 struct per_cpu_info *pci, int rw)
949 __account_queue(&pci->io_stats, t, rw);
951 if (per_process_stats) {
952 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
954 __account_queue(ios, t, rw);
958 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
962 ios->cwrite_kb += bytes >> 10;
965 ios->cread_kb += bytes >> 10;
969 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
972 __account_c(&pci->io_stats, rw, bytes);
974 if (per_process_stats) {
975 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
977 __account_c(ios, rw, bytes);
981 static inline void __account_issue(struct io_stats *ios, int rw,
986 ios->iwrite_kb += bytes >> 10;
989 ios->iread_kb += bytes >> 10;
993 static inline void account_issue(struct blk_io_trace *t,
994 struct per_cpu_info *pci, int rw)
996 __account_issue(&pci->io_stats, rw, t->bytes);
998 if (per_process_stats) {
999 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
1001 __account_issue(ios, rw, t->bytes);
1005 static inline void __account_unplug(struct io_stats *ios, int timer)
1008 ios->timer_unplugs++;
1013 static inline void account_unplug(struct blk_io_trace *t,
1014 struct per_cpu_info *pci, int timer)
1016 __account_unplug(&pci->io_stats, timer);
1018 if (per_process_stats) {
1019 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
1021 __account_unplug(ios, timer);
1025 static inline void __account_requeue(struct io_stats *ios,
1026 struct blk_io_trace *t, int rw)
1030 ios->iwrite_kb -= t_kb(t);
1033 ios->iread_kb -= t_kb(t);
1037 static inline void account_requeue(struct blk_io_trace *t,
1038 struct per_cpu_info *pci, int rw)
1040 __account_requeue(&pci->io_stats, t, rw);
1042 if (per_process_stats) {
1043 struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
1045 __account_requeue(ios, t, rw);
1049 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1050 struct blk_io_trace *t, char *act)
1052 process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1055 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1056 struct blk_io_trace *t, char *act)
1058 process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1061 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1064 process_fmt(act, pci, t, -1, 0, NULL);
1067 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1068 struct blk_io_trace *t, char *act)
1070 process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1073 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1074 struct blk_io_trace *t, char *act)
1077 log_track_frontmerge(pdi, t);
1079 process_fmt(act, pci, t, -1ULL, 0, NULL);
1082 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1085 process_fmt(act, pci, t, -1ULL, 0, NULL);
1088 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1091 process_fmt(act, pci, t, -1ULL, 0, NULL);
1094 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1097 process_fmt(act, pci, t, -1ULL, 0, NULL);
1100 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1103 process_fmt(act, pci, t, -1ULL, 0, NULL);
1106 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1108 unsigned char *buf = (unsigned char *) t + sizeof(*t);
1110 process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1113 static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
1115 int act = t->action & 0xffff;
1118 case __BLK_TA_QUEUE:
1119 log_generic(pci, t, "Q");
1121 case __BLK_TA_GETRQ:
1122 log_generic(pci, t, "G");
1124 case __BLK_TA_SLEEPRQ:
1125 log_generic(pci, t, "S");
1127 case __BLK_TA_REQUEUE:
1128 log_generic(pci, t, "R");
1130 case __BLK_TA_ISSUE:
1131 log_pc(pci, t, "D");
1133 case __BLK_TA_COMPLETE:
1134 log_pc(pci, t, "C");
1136 case __BLK_TA_INSERT:
1137 log_pc(pci, t, "I");
1140 fprintf(stderr, "Bad pc action %x\n", act);
1145 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1146 struct per_cpu_info *pci)
1148 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1149 int act = t->action & 0xffff;
1152 case __BLK_TA_QUEUE:
1153 account_queue(t, pci, w);
1154 log_queue(pci, t, "Q");
1156 case __BLK_TA_INSERT:
1157 log_insert(pdi, pci, t, "I");
1159 case __BLK_TA_BACKMERGE:
1160 account_m(t, pci, w);
1161 log_merge(pdi, pci, t, "M");
1163 case __BLK_TA_FRONTMERGE:
1164 account_m(t, pci, w);
1165 log_merge(pdi, pci, t, "F");
1167 case __BLK_TA_GETRQ:
1168 log_track_getrq(pdi, t);
1169 log_generic(pci, t, "G");
1171 case __BLK_TA_SLEEPRQ:
1172 log_generic(pci, t, "S");
1174 case __BLK_TA_REQUEUE:
1175 pdi->cur_depth[w]--;
1176 account_requeue(t, pci, w);
1177 log_queue(pci, t, "R");
1179 case __BLK_TA_ISSUE:
1180 account_issue(t, pci, w);
1181 pdi->cur_depth[w]++;
1182 if (pdi->cur_depth[w] > pdi->max_depth[w])
1183 pdi->max_depth[w] = pdi->cur_depth[w];
1184 log_issue(pdi, pci, t, "D");
1186 case __BLK_TA_COMPLETE:
1187 pdi->cur_depth[w]--;
1188 account_c(t, pci, w, t->bytes);
1189 log_complete(pdi, pci, t, "C");
1192 log_action(pci, t, "P");
1194 case __BLK_TA_UNPLUG_IO:
1195 account_unplug(t, pci, 0);
1196 log_unplug(pci, t, "U");
1198 case __BLK_TA_UNPLUG_TIMER:
1199 account_unplug(t, pci, 1);
1200 log_unplug(pci, t, "UT");
1202 case __BLK_TA_SPLIT:
1203 log_split(pci, t, "X");
1205 case __BLK_TA_BOUNCE:
1206 log_generic(pci, t, "B");
1208 case __BLK_TA_REMAP:
1209 log_generic(pci, t, "A");
1212 fprintf(stderr, "Bad fs action %x\n", t->action);
1217 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1218 struct per_dev_info *pdi)
1220 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1221 dump_trace_pc(t, pci);
1223 dump_trace_fs(t, pdi, pci);
1226 pdi->first_reported_time = t->time;
1232 * print in a proper way, not too small and not too big. if more than
1233 * 1000,000K, turn into M and so on
1235 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1237 char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1243 while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1248 sprintf(dst, "%'8Lu%c", num, suff[i]);
1252 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1255 static char x[256], y[256];
1257 fprintf(ofp, "%s\n", msg);
1259 fprintf(ofp, " Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1260 fprintf(ofp, " Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1262 fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1263 fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1264 fprintf(ofp, " Reads Requeued: %s\t\t", size_cnv(x, ios->rrqueue, 0));
1265 fprintf(ofp, " Writes Requeued: %s\n", size_cnv(x, ios->wrqueue, 0));
1266 fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1267 fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1268 fprintf(ofp, " Read Merges: %'8lu%8c\t", ios->mreads, ' ');
1269 fprintf(ofp, " Write Merges: %'8lu\n", ios->mwrites);
1271 fprintf(ofp, " Read depth: %'8u%8c\t", pdi->max_depth[0], ' ');
1272 fprintf(ofp, " Write depth: %'8u\n", pdi->max_depth[1]);
1274 fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' ');
1275 fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs);
1278 static void dump_wait_stats(struct per_process_info *ppi)
1280 unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1281 unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1282 unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1283 unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1284 unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1285 unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1287 fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1288 fprintf(ofp, " Allocation wait: %'8lu\n", wawait);
1289 fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' ');
1290 fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait);
1291 fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1292 fprintf(ofp, " Completion wait: %'8lu\n", wcwait);
1295 static int ppi_name_compare(const void *p1, const void *p2)
1297 struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1298 struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1301 res = strverscmp(ppi1->name, ppi2->name);
1303 res = ppi1->pid > ppi2->pid;
1308 static void sort_process_list(void)
1310 struct per_process_info **ppis;
1311 struct per_process_info *ppi;
1314 ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1319 ppi = ppi->list_next;
1322 qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1324 i = ppi_list_entries - 1;
1329 ppi->list_next = ppi_list;
1337 static void show_process_stats(void)
1339 struct per_process_info *ppi;
1341 sort_process_list();
1347 if (ppi->more_than_one)
1348 sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
1350 sprintf(name, "%s (%u)", ppi->name, ppi->pid);
1352 dump_io_stats(NULL, &ppi->io_stats, name);
1353 dump_wait_stats(ppi);
1354 ppi = ppi->list_next;
1360 static void show_device_and_cpu_stats(void)
1362 struct per_dev_info *pdi;
1363 struct per_cpu_info *pci;
1364 struct io_stats total, *ios;
1365 unsigned long long rrate, wrate, msec;
1366 int i, j, pci_events;
1367 char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1370 for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1372 memset(&total, 0, sizeof(total));
1378 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1382 ios = &pci->io_stats;
1383 total.qreads += ios->qreads;
1384 total.qwrites += ios->qwrites;
1385 total.creads += ios->creads;
1386 total.cwrites += ios->cwrites;
1387 total.mreads += ios->mreads;
1388 total.mwrites += ios->mwrites;
1389 total.ireads += ios->ireads;
1390 total.iwrites += ios->iwrites;
1391 total.rrqueue += ios->rrqueue;
1392 total.wrqueue += ios->wrqueue;
1393 total.qread_kb += ios->qread_kb;
1394 total.qwrite_kb += ios->qwrite_kb;
1395 total.cread_kb += ios->cread_kb;
1396 total.cwrite_kb += ios->cwrite_kb;
1397 total.iread_kb += ios->iread_kb;
1398 total.iwrite_kb += ios->iwrite_kb;
1399 total.timer_unplugs += ios->timer_unplugs;
1400 total.io_unplugs += ios->io_unplugs;
1402 snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1403 j, get_dev_name(pdi, name, sizeof(name)));
1404 dump_io_stats(pdi, ios, line);
1408 if (pci_events > 1) {
1410 snprintf(line, sizeof(line) - 1, "Total (%s):",
1411 get_dev_name(pdi, name, sizeof(name)));
1412 dump_io_stats(NULL, &total, line);
1416 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1418 rrate = 1000 * total.cread_kb / msec;
1419 wrate = 1000 * total.cwrite_kb / msec;
1422 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1424 fprintf(ofp, "Events (%s): %'Lu entries\n",
1425 get_dev_name(pdi, line, sizeof(line)), pdi->events);
1427 collect_pdi_skips(pdi);
1428 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1429 pdi->skips,pdi->seq_skips,
1430 100.0 * ((double)pdi->seq_skips /
1431 (double)(pdi->events + pdi->seq_skips)));
1436 * struct trace and blktrace allocation cache, we do potentially
1437 * millions of mallocs for these structures while only using at most
1438 * a few thousand at the time
1440 static inline void t_free(struct trace *t)
1442 if (t_alloc_cache < 1024) {
1443 t->next = t_alloc_list;
1450 static inline struct trace *t_alloc(void)
1452 struct trace *t = t_alloc_list;
1455 t_alloc_list = t->next;
1460 return malloc(sizeof(*t));
1463 static inline void bit_free(struct blk_io_trace *bit)
1465 if (bit_alloc_cache < 1024 && !bit->pdu_len) {
1467 * abuse a 64-bit field for a next pointer for the free item
1469 bit->time = (__u64) (unsigned long) bit_alloc_list;
1470 bit_alloc_list = (struct blk_io_trace *) bit;
1476 static inline struct blk_io_trace *bit_alloc(void)
1478 struct blk_io_trace *bit = bit_alloc_list;
1481 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
1487 return malloc(sizeof(*bit));
1490 static void find_genesis(void)
1492 struct trace *t = trace_list;
1494 genesis_time = -1ULL;
1496 if (t->bit->time < genesis_time)
1497 genesis_time = t->bit->time;
1503 static inline int check_stopwatch(struct blk_io_trace *bit)
1505 if (bit->time < stopwatch_end &&
1506 bit->time >= stopwatch_start)
1513 * return youngest entry read
1515 static int sort_entries(unsigned long long *youngest)
1523 while ((t = trace_list) != NULL) {
1524 struct blk_io_trace *bit = t->bit;
1526 trace_list = t->next;
1528 bit->time -= genesis_time;
1530 if (bit->time < *youngest || !*youngest)
1531 *youngest = bit->time;
1533 if (bit->sequence < smallest_seq_read)
1534 smallest_seq_read = bit->sequence;
1536 if (check_stopwatch(bit)) {
1542 if (trace_rb_insert_sort(t))
1549 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
1551 rb_erase(&t->rb_node, &pdi->rb_last);
1552 pdi->rb_last_entries--;
1558 static void put_trace(struct per_dev_info *pdi, struct trace *t)
1560 rb_erase(&t->rb_node, &rb_sort_root);
1563 trace_rb_insert_last(pdi, t);
1565 if (pdi->rb_last_entries > rb_batch * pdi->nfiles) {
1566 struct rb_node *n = rb_first(&pdi->rb_last);
1568 t = rb_entry(n, struct trace, rb_node);
1569 __put_trace_last(pdi, t);
1574 * to continue, we must have traces from all online cpus in the tree
1576 static int check_cpu_map(struct per_dev_info *pdi)
1578 unsigned long *cpu_map;
1585 * create a map of the cpus we have traces for
1587 cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1588 n = rb_first(&rb_sort_root);
1590 __t = rb_entry(n, struct trace, rb_node);
1591 cpu = __t->bit->cpu;
1593 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1598 * we can't continue if pdi->cpu_map has entries set that we don't
1599 * have in the sort rbtree. the opposite is not a problem, though
1602 for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1603 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1613 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1615 unsigned long expected_sequence = pdi->last_sequence + 1;
1616 struct blk_io_trace *bit = t->bit;
1619 if (!expected_sequence) {
1621 * 1 should be the first entry, just allow it
1623 if (bit->sequence == 1)
1625 if (bit->sequence == smallest_seq_read)
1628 return check_cpu_map(pdi);
1631 if (bit->sequence == expected_sequence)
1635 * we may not have seen that sequence yet. if we are not doing
1636 * the final run, break and wait for more entries.
1638 if (expected_sequence < smallest_seq_read) {
1639 __t = trace_rb_find_last(pdi, expected_sequence);
1643 __put_trace_last(pdi, __t);
1645 } else if (!force) {
1649 if (check_current_skips(pdi,bit->sequence))
1652 if (expected_sequence < bit->sequence)
1653 insert_skip(pdi, expected_sequence, bit->sequence - 1);
1658 static void show_entries_rb(int force)
1660 struct per_dev_info *pdi = NULL;
1661 struct per_cpu_info *pci = NULL;
1662 struct blk_io_trace *bit;
1666 while ((n = rb_first(&rb_sort_root)) != NULL) {
1667 if (is_done() && !force && !pipeline)
1670 t = rb_entry(n, struct trace, rb_node);
1673 if (!pdi || pdi->dev != bit->device)
1674 pdi = get_dev_info(bit->device);
1677 fprintf(stderr, "Unknown device ID? (%d,%d)\n",
1678 MAJOR(bit->device), MINOR(bit->device));
1682 if (check_sequence(pdi, t, force))
1685 if (!force && bit->time > last_allowed_time)
1688 pdi->last_sequence = bit->sequence;
1690 check_time(pdi, bit);
1692 if (!pci || pci->cpu != bit->cpu)
1693 pci = get_cpu_info(pdi, bit->cpu);
1697 if (bit->action & (act_mask << BLK_TC_SHIFT))
1698 dump_trace(bit, pci, pdi);
1704 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
1706 int ret, bytes_left, fl;
1709 if (block != *fdblock) {
1710 fl = fcntl(fd, F_GETFL);
1714 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
1717 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
1723 while (bytes_left > 0) {
1724 ret = read(fd, p, bytes_left);
1728 if (errno != EAGAIN)
1732 * never do partial reads. we can return if we
1733 * didn't read anything and we should not block,
1734 * otherwise wait for data
1736 if ((bytes_left == bytes) && !block)
1750 static int read_events(int fd, int always_block, int *fdblock)
1752 struct per_dev_info *pdi = NULL;
1753 unsigned int events = 0;
1755 while (!is_done() && events < rb_batch) {
1756 struct blk_io_trace *bit;
1758 int pdu_len, should_block;
1763 should_block = !events || always_block;
1765 if (read_data(fd, bit, sizeof(*bit), should_block, fdblock)) {
1770 magic = be32_to_cpu(bit->magic);
1771 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
1772 fprintf(stderr, "Bad magic %x\n", magic);
1776 pdu_len = be16_to_cpu(bit->pdu_len);
1778 void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
1780 if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
1790 if (verify_trace(bit)) {
1796 memset(t, 0, sizeof(*t));
1799 t->next = trace_list;
1802 if (!pdi || pdi->dev != bit->device)
1803 pdi = get_dev_info(bit->device);
1805 if (bit->time > pdi->last_read_time)
1806 pdi->last_read_time = bit->time;
1814 static int do_file(void)
1816 struct per_cpu_info *pci;
1817 struct per_dev_info *pdi;
1818 int i, j, events, events_added;
1821 * first prepare all files for reading
1823 for (i = 0; i < ndevices; i++) {
1826 pdi->last_sequence = -1;
1833 pci = get_cpu_info(pdi, j);
1838 p = strdup(pdi->name);
1840 if (strcmp(dname, ".")) {
1842 p = strdup(pdi->name);
1843 strcpy(pdi->name, basename(p));
1848 len = sprintf(pci->fname, "%s/", input_dir);
1850 snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
1851 "%s.blktrace.%d", pdi->name, pci->cpu);
1852 if (stat(pci->fname, &st) < 0)
1855 pci->fd = open(pci->fname, O_RDONLY);
1862 printf("Input file %s added\n", pci->fname);
1864 cpu_mark_online(pdi, pci->cpu);
1869 * now loop over the files reading in the data
1872 unsigned long long youngest;
1875 last_allowed_time = -1ULL;
1876 smallest_seq_read = -1U;
1878 for (i = 0; i < ndevices; i++) {
1881 for (j = 0; j < pdi->nfiles; j++) {
1883 pci = get_cpu_info(pdi, j);
1888 events = read_events(pci->fd, 1, &pci->fdblock);
1890 cpu_mark_offline(pdi, pci->cpu);
1896 if (pdi->last_read_time < last_allowed_time)
1897 last_allowed_time = pdi->last_read_time;
1899 events_added += events;
1903 if (sort_entries(&youngest))
1906 if (youngest > stopwatch_end)
1911 } while (events_added);
1913 if (rb_sort_entries)
1919 static int do_stdin(void)
1921 unsigned long long youngest;
1922 int fd, events, fdblock;
1924 last_allowed_time = -1ULL;
1925 fd = dup(STDIN_FILENO);
1927 perror("dup stdin");
1932 while ((events = read_events(fd, 0, &fdblock)) != 0) {
1934 smallest_seq_read = -1U;
1936 if (sort_entries(&youngest))
1939 if (youngest > stopwatch_end)
1945 if (rb_sort_entries)
1952 static void show_stats(void)
1961 if (per_process_stats)
1962 show_process_stats();
1964 if (per_device_and_cpu_stats)
1965 show_device_and_cpu_stats();
1970 static void handle_sigint(__attribute__((__unused__)) int sig)
1976 * Extract start and duration times from a string, allowing
1977 * us to specify a time interval of interest within a trace.
1978 * Format: "duration" (start is zero) or "start:duration".
1980 static int find_stopwatch_interval(char *string)
1985 value = strtod(string, &sp);
1987 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
1991 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
1993 value = strtod(string, &sp);
1994 if (sp == string || *sp != '\0') {
1995 fprintf(stderr,"Invalid stopwatch duration time: %s\n",
1999 } else if (*sp != '\0') {
2000 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2003 stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2004 if (stopwatch_end <= stopwatch_start) {
2005 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2006 stopwatch_start, stopwatch_end);
2013 static char usage_str[] = \
2014 "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
2015 "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
2016 "\t-i Input file containing trace data, or '-' for stdin\n" \
2017 "\t-D Directory to prepend to input file names\n" \
2018 "\t-o Output file. If not given, output is stdout\n" \
2019 "\t-b stdin read batching\n" \
2020 "\t-s Show per-program io statistics\n" \
2021 "\t-h Hash processes by name, not pid\n" \
2022 "\t-t Track individual ios. Will tell you the time a request took\n" \
2023 "\t to get queued, to get dispatched, and to get completed\n" \
2024 "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2025 "\t-w Only parse data between the given time interval in seconds.\n" \
2026 "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \
2027 "\t-f Output format. Customize the output format. The format field\n" \
2028 "\t identifies can be found in the documentation\n" \
2029 "\t-F Format specification. Can be found in the documentation\n" \
2030 "\t-v More verbose for marginal errors\n" \
2031 "\t-V Print program version info\n\n";
2033 static void usage(char *prog)
2035 fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2038 int main(int argc, char *argv[])
2041 int i, c, ret, mode;
2042 int act_mask_tmp = 0;
2044 while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2047 i = find_mask_map(optarg);
2049 fprintf(stderr,"Invalid action mask %s\n",
2057 if ((sscanf(optarg, "%x", &i) != 1) ||
2058 !valid_act_opt(i)) {
2060 "Invalid set action mask %s/0x%x\n",
2067 if (!strcmp(optarg, "-") && !pipeline)
2069 else if (resize_devices(optarg) != 0)
2076 output_name = optarg;
2079 rb_batch = atoi(optarg);
2081 rb_batch = RB_BATCH_DEFAULT;
2084 per_process_stats = 1;
2090 per_device_and_cpu_stats = 0;
2093 if (find_stopwatch_interval(optarg) != 0)
2097 set_all_format_specs(optarg);
2100 if (add_format_spec(optarg) != 0)
2104 ppi_hash_by_pid = 0;
2110 printf("%s version %s\n", argv[0], blkparse_version);
2118 while (optind < argc) {
2119 if (!strcmp(argv[optind], "-") && !pipeline)
2121 else if (resize_devices(argv[optind]) != 0)
2126 if (!pipeline && !ndevices) {
2131 if (act_mask_tmp != 0)
2132 act_mask = act_mask_tmp;
2134 memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2136 signal(SIGINT, handle_sigint);
2137 signal(SIGHUP, handle_sigint);
2138 signal(SIGTERM, handle_sigint);
2140 setlocale(LC_NUMERIC, "en_US");
2143 ofp = fdopen(STDOUT_FILENO, "w");
2148 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2149 ofp = fopen(ofname, "w");
2158 ofp_buffer = malloc(4096);
2159 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {