2 * block queue tracing parse application
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 #include <sys/types.h>
39 static char blkparse_version[] = "0.99.3";
42 unsigned long start, end;
43 struct skip_info *prev, *next;
51 unsigned long long events;
52 unsigned long long first_reported_time;
53 unsigned long long last_reported_time;
54 unsigned long long last_read_time;
55 struct io_stats io_stats;
57 unsigned long long seq_skips;
58 unsigned int max_depth[2];
59 unsigned int cur_depth[2];
61 struct rb_root rb_track;
66 unsigned long *cpu_map;
67 unsigned int cpu_map_max;
69 struct per_cpu_info *cpus;
73 * some duplicated effort here, we can unify this hash and the ppi hash later
75 struct process_pid_map {
78 struct process_pid_map *hash_next, *list_next;
81 #define PPM_HASH_SHIFT (8)
82 #define PPM_HASH_SIZE (1 << PPM_HASH_SHIFT)
83 #define PPM_HASH_MASK (PPM_HASH_SIZE - 1)
84 static struct process_pid_map *ppm_hash_table[PPM_HASH_SIZE];
86 struct per_process_info {
87 struct process_pid_map *ppm;
88 struct io_stats io_stats;
89 struct per_process_info *hash_next, *list_next;
95 unsigned long long longest_allocation_wait[2];
96 unsigned long long longest_dispatch_wait[2];
97 unsigned long long longest_completion_wait[2];
100 #define PPI_HASH_SHIFT (8)
101 #define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT)
102 #define PPI_HASH_MASK (PPI_HASH_SIZE - 1)
103 static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
104 static struct per_process_info *ppi_list;
105 static int ppi_list_entries;
107 #define S_OPTS "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
108 static struct option l_opts[] = {
111 .has_arg = required_argument,
117 .has_arg = required_argument,
123 .has_arg = required_argument,
128 .name = "input-directory",
129 .has_arg = required_argument,
134 .name = "dump-binary",
135 .has_arg = required_argument,
141 .has_arg = required_argument,
146 .name = "format-spec",
147 .has_arg = required_argument,
152 .name = "hash-by-name",
153 .has_arg = no_argument,
159 .has_arg = required_argument,
165 .has_arg = required_argument,
170 .name = "no-text-output",
171 .has_arg = no_argument,
177 .has_arg = no_argument,
182 .name = "per-program-stats",
183 .has_arg = no_argument,
189 .has_arg = no_argument,
195 .has_arg = required_argument,
201 .has_arg = no_argument,
207 .has_arg = no_argument,
217 * for sorting the displayed output
220 struct blk_io_trace *bit;
221 struct rb_node rb_node;
223 unsigned long read_sequence;
226 static struct rb_root rb_sort_root;
227 static unsigned long rb_sort_entries;
229 static struct trace *trace_list;
234 static struct blk_io_trace *bit_alloc_list;
235 static struct trace *t_alloc_list;
238 * for tracking individual ios
241 struct rb_node rb_node;
243 struct process_pid_map *ppm;
245 unsigned long long allocation_time;
246 unsigned long long queue_time;
247 unsigned long long dispatch_time;
248 unsigned long long completion_time;
252 static struct per_dev_info *devices;
253 static char *get_dev_name(struct per_dev_info *, char *, int);
254 static int trace_rb_insert_last(struct per_dev_info *, struct trace *);
257 static char *output_name;
258 static char *input_dir;
260 static unsigned long long genesis_time;
261 static unsigned long long last_allowed_time;
262 static unsigned long long stopwatch_start; /* start from zero by default */
263 static unsigned long long stopwatch_end = -1ULL; /* "infinity" */
264 static unsigned long read_sequence;
266 static int per_process_stats;
267 static int per_device_and_cpu_stats = 1;
268 static int track_ios;
269 static int ppi_hash_by_pid = 1;
271 static unsigned int act_mask = -1U;
272 static int stats_printed;
273 int data_is_native = -1;
275 static FILE *dump_fp;
276 static char *dump_binary;
278 static unsigned int t_alloc_cache;
279 static unsigned int bit_alloc_cache;
281 #define RB_BATCH_DEFAULT (512)
282 static unsigned int rb_batch = RB_BATCH_DEFAULT;
285 static char *pipename;
287 static int text_output = 1;
289 #define is_done() (*(volatile int *)(&done))
290 static volatile int done;
292 struct timespec abs_start_time;
293 static unsigned long long start_timestamp;
295 #define JHASH_RANDOM (0x3af5f2ee)
297 #define CPUS_PER_LONG (8 * sizeof(unsigned long))
298 #define CPU_IDX(cpu) ((cpu) / CPUS_PER_LONG)
299 #define CPU_BIT(cpu) ((cpu) & (CPUS_PER_LONG - 1))
301 static void output_binary(void *buf, int len)
304 size_t n = fwrite(buf, len, 1, dump_fp);
313 static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
315 struct per_cpu_info *cpus = pdi->cpus;
316 int ncpus = pdi->ncpus;
317 int new_count = cpu + 1;
321 size = new_count * sizeof(struct per_cpu_info);
322 cpus = realloc(cpus, size);
325 fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
326 get_dev_name(pdi, name, sizeof(name)), size);
330 new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
331 new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
332 memset(new_start, 0, new_space);
334 pdi->ncpus = new_count;
337 for (new_count = 0; new_count < pdi->ncpus; new_count++) {
338 struct per_cpu_info *pci = &pdi->cpus[new_count];
342 memset(&pci->rb_last, 0, sizeof(pci->rb_last));
343 pci->rb_last_entries = 0;
344 pci->last_sequence = -1;
349 static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
351 struct per_cpu_info *pci;
353 if (cpu >= pdi->ncpus)
354 resize_cpu_info(pdi, cpu);
356 pci = &pdi->cpus[cpu];
362 static int resize_devices(char *name)
364 int size = (ndevices + 1) * sizeof(struct per_dev_info);
366 devices = realloc(devices, size);
368 fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
371 memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
372 devices[ndevices].name = name;
377 static struct per_dev_info *get_dev_info(dev_t dev)
379 struct per_dev_info *pdi;
382 for (i = 0; i < ndevices; i++) {
384 devices[i].dev = dev;
385 if (devices[i].dev == dev)
389 if (resize_devices(NULL))
392 pdi = &devices[ndevices - 1];
394 pdi->first_reported_time = 0;
395 pdi->last_read_time = 0;
400 static void insert_skip(struct per_cpu_info *pci, unsigned long start,
403 struct skip_info *sip;
405 for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
406 if (end == (sip->start - 1)) {
409 } else if (start == (sip->end + 1)) {
415 sip = malloc(sizeof(struct skip_info));
418 sip->prev = sip->next = NULL;
419 if (pci->skips_tail == NULL)
420 pci->skips_head = pci->skips_tail = sip;
422 sip->prev = pci->skips_tail;
423 pci->skips_tail->next = sip;
424 pci->skips_tail = sip;
428 static void remove_sip(struct per_cpu_info *pci, struct skip_info *sip)
430 if (sip->prev == NULL) {
431 if (sip->next == NULL)
432 pci->skips_head = pci->skips_tail = NULL;
434 pci->skips_head = sip->next;
435 sip->next->prev = NULL;
437 } else if (sip->next == NULL) {
438 pci->skips_tail = sip->prev;
439 sip->prev->next = NULL;
441 sip->prev->next = sip->next;
442 sip->next->prev = sip->prev;
445 sip->prev = sip->next = NULL;
449 #define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
450 static int check_current_skips(struct per_cpu_info *pci, unsigned long seq)
452 struct skip_info *sip;
454 for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
455 if (IN_SKIP(sip, seq)) {
456 if (sip->start == seq) {
458 remove_sip(pci, sip);
461 } else if (sip->end == seq)
465 insert_skip(pci, seq + 1, sip->end);
474 static void collect_pdi_skips(struct per_dev_info *pdi)
476 struct skip_info *sip;
482 for (cpu = 0; cpu < pdi->ncpus; cpu++) {
483 struct per_cpu_info *pci = &pdi->cpus[cpu];
485 for (sip = pci->skips_head; sip != NULL; sip = sip->next) {
487 pdi->seq_skips += (sip->end - sip->start + 1);
489 fprintf(stderr,"(%d,%d): skipping %lu -> %lu\n",
490 MAJOR(pdi->dev), MINOR(pdi->dev),
491 sip->start, sip->end);
496 static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
498 if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
499 int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
500 unsigned long *map = malloc(new_max / sizeof(long));
502 memset(map, 0, new_max / sizeof(long));
505 memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
510 pdi->cpu_map_max = new_max;
513 pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
516 static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
518 pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
521 static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
523 return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
526 static inline int ppm_hash_pid(pid_t pid)
528 return jhash_1word(pid, JHASH_RANDOM) & PPM_HASH_MASK;
531 static struct process_pid_map *find_ppm(pid_t pid)
533 const int hash_idx = ppm_hash_pid(pid);
534 struct process_pid_map *ppm;
536 ppm = ppm_hash_table[hash_idx];
541 ppm = ppm->hash_next;
547 static struct process_pid_map *add_ppm_hash(pid_t pid, const char *name)
549 const int hash_idx = ppm_hash_pid(pid);
550 struct process_pid_map *ppm;
554 ppm = malloc(sizeof(*ppm));
555 memset(ppm, 0, sizeof(*ppm));
557 strcpy(ppm->comm, name);
558 ppm->hash_next = ppm_hash_table[hash_idx];
559 ppm_hash_table[hash_idx] = ppm;
565 static void handle_notify(struct blk_io_trace *bit)
567 void *payload = (caddr_t) bit + sizeof(*bit);
570 switch (bit->action) {
572 add_ppm_hash(bit->pid, payload);
575 case BLK_TN_TIMESTAMP:
576 if (bit->pdu_len != sizeof(two32))
578 memcpy(two32, payload, sizeof(two32));
579 if (!data_is_native) {
580 two32[0] = be32_to_cpu(two32[0]);
581 two32[1] = be32_to_cpu(two32[1]);
583 start_timestamp = bit->time;
584 abs_start_time.tv_sec = two32[0];
585 abs_start_time.tv_nsec = two32[1];
586 if (abs_start_time.tv_nsec < 0) {
587 abs_start_time.tv_sec--;
588 abs_start_time.tv_nsec += 1000000000;
594 /* Ignore unknown notify events */
599 char *find_process_name(pid_t pid)
601 struct process_pid_map *ppm = find_ppm(pid);
609 static inline int ppi_hash_pid(pid_t pid)
611 return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
614 static inline int ppi_hash_name(const char *name)
616 return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
619 static inline int ppi_hash(struct per_process_info *ppi)
621 struct process_pid_map *ppm = ppi->ppm;
624 return ppi_hash_pid(ppm->pid);
626 return ppi_hash_name(ppm->comm);
629 static inline void add_ppi_to_hash(struct per_process_info *ppi)
631 const int hash_idx = ppi_hash(ppi);
633 ppi->hash_next = ppi_hash_table[hash_idx];
634 ppi_hash_table[hash_idx] = ppi;
637 static inline void add_ppi_to_list(struct per_process_info *ppi)
639 ppi->list_next = ppi_list;
644 static struct per_process_info *find_ppi_by_name(char *name)
646 const int hash_idx = ppi_hash_name(name);
647 struct per_process_info *ppi;
649 ppi = ppi_hash_table[hash_idx];
651 struct process_pid_map *ppm = ppi->ppm;
653 if (!strcmp(ppm->comm, name))
656 ppi = ppi->hash_next;
662 static struct per_process_info *find_ppi_by_pid(pid_t pid)
664 const int hash_idx = ppi_hash_pid(pid);
665 struct per_process_info *ppi;
667 ppi = ppi_hash_table[hash_idx];
669 struct process_pid_map *ppm = ppi->ppm;
674 ppi = ppi->hash_next;
680 static struct per_process_info *find_ppi(pid_t pid)
682 struct per_process_info *ppi;
686 return find_ppi_by_pid(pid);
688 name = find_process_name(pid);
692 ppi = find_ppi_by_name(name);
693 if (ppi && ppi->ppm->pid != pid)
694 ppi->more_than_one = 1;
700 * struct trace and blktrace allocation cache, we do potentially
701 * millions of mallocs for these structures while only using at most
702 * a few thousand at the time
704 static inline void t_free(struct trace *t)
706 if (t_alloc_cache < 1024) {
707 t->next = t_alloc_list;
714 static inline struct trace *t_alloc(void)
716 struct trace *t = t_alloc_list;
719 t_alloc_list = t->next;
724 return malloc(sizeof(*t));
727 static inline void bit_free(struct blk_io_trace *bit)
729 if (bit_alloc_cache < 1024 && !bit->pdu_len) {
731 * abuse a 64-bit field for a next pointer for the free item
733 bit->time = (__u64) (unsigned long) bit_alloc_list;
734 bit_alloc_list = (struct blk_io_trace *) bit;
740 static inline struct blk_io_trace *bit_alloc(void)
742 struct blk_io_trace *bit = bit_alloc_list;
745 bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
751 return malloc(sizeof(*bit));
754 static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
756 struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
758 rb_erase(&t->rb_node, &pci->rb_last);
759 pci->rb_last_entries--;
765 static void put_trace(struct per_dev_info *pdi, struct trace *t)
767 rb_erase(&t->rb_node, &rb_sort_root);
770 trace_rb_insert_last(pdi, t);
773 static inline int trace_rb_insert(struct trace *t, struct rb_root *root)
775 struct rb_node **p = &root->rb_node;
776 struct rb_node *parent = NULL;
782 __t = rb_entry(parent, struct trace, rb_node);
784 if (t->bit->time < __t->bit->time)
786 else if (t->bit->time > __t->bit->time)
788 else if (t->bit->device < __t->bit->device)
790 else if (t->bit->device > __t->bit->device)
792 else if (t->bit->sequence < __t->bit->sequence)
794 else /* >= sequence */
798 rb_link_node(&t->rb_node, parent, p);
799 rb_insert_color(&t->rb_node, root);
803 static inline int trace_rb_insert_sort(struct trace *t)
805 if (!trace_rb_insert(t, &rb_sort_root)) {
813 static int trace_rb_insert_last(struct per_dev_info *pdi, struct trace *t)
815 struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
817 if (trace_rb_insert(t, &pci->rb_last))
820 pci->rb_last_entries++;
822 if (pci->rb_last_entries > rb_batch * pdi->nfiles) {
823 struct rb_node *n = rb_first(&pci->rb_last);
825 t = rb_entry(n, struct trace, rb_node);
826 __put_trace_last(pdi, t);
832 static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
833 struct rb_root *root, int order)
835 struct rb_node *n = root->rb_node;
836 struct rb_node *prev = NULL;
840 __t = rb_entry(n, struct trace, rb_node);
843 if (device < __t->bit->device)
845 else if (device > __t->bit->device)
847 else if (sequence < __t->bit->sequence)
849 else if (sequence > __t->bit->sequence)
856 * hack - the list may not be sequence ordered because some
857 * events don't have sequence and time matched. so we end up
858 * being a little off in the rb lookup here, because we don't
859 * know the time we are looking for. compensate by browsing
860 * a little ahead from the last entry to find the match
865 while (((n = rb_next(prev)) != NULL) && max--) {
866 __t = rb_entry(n, struct trace, rb_node);
868 if (__t->bit->device == device &&
869 __t->bit->sequence == sequence)
879 static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
880 struct per_cpu_info *pci,
883 return trace_rb_find(pdi->dev, seq, &pci->rb_last, 0);
886 static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
888 struct rb_node **p = &pdi->rb_track.rb_node;
889 struct rb_node *parent = NULL;
890 struct io_track *__iot;
894 __iot = rb_entry(parent, struct io_track, rb_node);
896 if (iot->sector < __iot->sector)
898 else if (iot->sector > __iot->sector)
902 "sector alias (%Lu) on device %d,%d!\n",
903 (unsigned long long) iot->sector,
904 MAJOR(pdi->dev), MINOR(pdi->dev));
909 rb_link_node(&iot->rb_node, parent, p);
910 rb_insert_color(&iot->rb_node, &pdi->rb_track);
914 static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
916 struct rb_node *n = pdi->rb_track.rb_node;
917 struct io_track *__iot;
920 __iot = rb_entry(n, struct io_track, rb_node);
922 if (sector < __iot->sector)
924 else if (sector > __iot->sector)
933 static struct io_track *find_track(struct per_dev_info *pdi, pid_t pid,
936 struct io_track *iot;
938 iot = __find_track(pdi, sector);
940 iot = malloc(sizeof(*iot));
941 iot->ppm = find_ppm(pid);
943 iot->ppm = add_ppm_hash(pid, "unknown");
944 iot->sector = sector;
945 track_rb_insert(pdi, iot);
951 static void log_track_frontmerge(struct per_dev_info *pdi,
952 struct blk_io_trace *t)
954 struct io_track *iot;
959 iot = __find_track(pdi, t->sector + t_sec(t));
962 fprintf(stderr, "merge not found for (%d,%d): %llu\n",
963 MAJOR(pdi->dev), MINOR(pdi->dev),
964 (unsigned long long) t->sector + t_sec(t));
968 rb_erase(&iot->rb_node, &pdi->rb_track);
969 iot->sector -= t_sec(t);
970 track_rb_insert(pdi, iot);
973 static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
975 struct io_track *iot;
980 iot = find_track(pdi, t->pid, t->sector);
981 iot->allocation_time = t->time;
984 static inline int is_remapper(struct per_dev_info *pdi)
986 int major = MAJOR(pdi->dev);
988 return (major == 253 || major == 9);
992 * for md/dm setups, the interesting cycle is Q -> C. So track queueing
993 * time here, as dispatch time
995 static void log_track_queue(struct per_dev_info *pdi, struct blk_io_trace *t)
997 struct io_track *iot;
1001 if (!is_remapper(pdi))
1004 iot = find_track(pdi, t->pid, t->sector);
1005 iot->dispatch_time = t->time;
1009 * return time between rq allocation and insertion
1011 static unsigned long long log_track_insert(struct per_dev_info *pdi,
1012 struct blk_io_trace *t)
1014 unsigned long long elapsed;
1015 struct io_track *iot;
1020 iot = find_track(pdi, t->pid, t->sector);
1021 iot->queue_time = t->time;
1023 if (!iot->allocation_time)
1026 elapsed = iot->queue_time - iot->allocation_time;
1028 if (per_process_stats) {
1029 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1030 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1032 if (ppi && elapsed > ppi->longest_allocation_wait[w])
1033 ppi->longest_allocation_wait[w] = elapsed;
1040 * return time between queue and issue
1042 static unsigned long long log_track_issue(struct per_dev_info *pdi,
1043 struct blk_io_trace *t)
1045 unsigned long long elapsed;
1046 struct io_track *iot;
1050 if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
1053 iot = __find_track(pdi, t->sector);
1056 fprintf(stderr, "issue not found for (%d,%d): %llu\n",
1057 MAJOR(pdi->dev), MINOR(pdi->dev),
1058 (unsigned long long) t->sector);
1062 iot->dispatch_time = t->time;
1063 elapsed = iot->dispatch_time - iot->queue_time;
1065 if (per_process_stats) {
1066 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1067 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1069 if (ppi && elapsed > ppi->longest_dispatch_wait[w])
1070 ppi->longest_dispatch_wait[w] = elapsed;
1077 * return time between dispatch and complete
1079 static unsigned long long log_track_complete(struct per_dev_info *pdi,
1080 struct blk_io_trace *t)
1082 unsigned long long elapsed;
1083 struct io_track *iot;
1088 iot = __find_track(pdi, t->sector);
1091 fprintf(stderr,"complete not found for (%d,%d): %llu\n",
1092 MAJOR(pdi->dev), MINOR(pdi->dev),
1093 (unsigned long long) t->sector);
1097 iot->completion_time = t->time;
1098 elapsed = iot->completion_time - iot->dispatch_time;
1100 if (per_process_stats) {
1101 struct per_process_info *ppi = find_ppi(iot->ppm->pid);
1102 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1104 if (ppi && elapsed > ppi->longest_completion_wait[w])
1105 ppi->longest_completion_wait[w] = elapsed;
1109 * kill the trace, we don't need it after completion
1111 rb_erase(&iot->rb_node, &pdi->rb_track);
1118 static struct io_stats *find_process_io_stats(pid_t pid)
1120 struct per_process_info *ppi = find_ppi(pid);
1123 ppi = malloc(sizeof(*ppi));
1124 memset(ppi, 0, sizeof(*ppi));
1125 ppi->ppm = find_ppm(pid);
1127 ppi->ppm = add_ppm_hash(pid, "unknown");
1128 add_ppi_to_hash(ppi);
1129 add_ppi_to_list(ppi);
1132 return &ppi->io_stats;
1135 static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
1138 snprintf(buffer, size, "%s", pdi->name);
1140 snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
1144 static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
1146 unsigned long long this = bit->time;
1147 unsigned long long last = pdi->last_reported_time;
1149 pdi->backwards = (this < last) ? 'B' : ' ';
1150 pdi->last_reported_time = this;
1153 static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
1158 ios->mwrite_kb += t_kb(t);
1161 ios->mread_kb += t_kb(t);
1165 static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
1168 __account_m(&pci->io_stats, t, rw);
1170 if (per_process_stats) {
1171 struct io_stats *ios = find_process_io_stats(t->pid);
1173 __account_m(ios, t, rw);
1177 static inline void __account_pc_queue(struct io_stats *ios,
1178 struct blk_io_trace *t, int rw)
1182 ios->qwrite_kb_pc += t_kb(t);
1185 ios->qread_kb += t_kb(t);
1189 static inline void account_pc_queue(struct blk_io_trace *t,
1190 struct per_cpu_info *pci, int rw)
1192 __account_pc_queue(&pci->io_stats, t, rw);
1194 if (per_process_stats) {
1195 struct io_stats *ios = find_process_io_stats(t->pid);
1197 __account_pc_queue(ios, t, rw);
1201 static inline void __account_pc_issue(struct io_stats *ios, int rw,
1206 ios->iwrite_kb_pc += bytes >> 10;
1209 ios->iread_kb_pc += bytes >> 10;
1213 static inline void account_pc_issue(struct blk_io_trace *t,
1214 struct per_cpu_info *pci, int rw)
1216 __account_pc_issue(&pci->io_stats, rw, t->bytes);
1218 if (per_process_stats) {
1219 struct io_stats *ios = find_process_io_stats(t->pid);
1221 __account_pc_issue(ios, rw, t->bytes);
1225 static inline void __account_pc_requeue(struct io_stats *ios,
1226 struct blk_io_trace *t, int rw)
1230 ios->iwrite_kb_pc -= t_kb(t);
1233 ios->iread_kb_pc -= t_kb(t);
1237 static inline void account_pc_requeue(struct blk_io_trace *t,
1238 struct per_cpu_info *pci, int rw)
1240 __account_pc_requeue(&pci->io_stats, t, rw);
1242 if (per_process_stats) {
1243 struct io_stats *ios = find_process_io_stats(t->pid);
1245 __account_pc_requeue(ios, t, rw);
1249 static inline void __account_pc_c(struct io_stats *ios, int rw)
1257 static inline void account_pc_c(struct blk_io_trace *t,
1258 struct per_cpu_info *pci, int rw)
1260 __account_pc_c(&pci->io_stats, rw);
1262 if (per_process_stats) {
1263 struct io_stats *ios = find_process_io_stats(t->pid);
1265 __account_pc_c(ios, rw);
1269 static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
1274 ios->qwrite_kb += t_kb(t);
1277 ios->qread_kb += t_kb(t);
1281 static inline void account_queue(struct blk_io_trace *t,
1282 struct per_cpu_info *pci, int rw)
1284 __account_queue(&pci->io_stats, t, rw);
1286 if (per_process_stats) {
1287 struct io_stats *ios = find_process_io_stats(t->pid);
1289 __account_queue(ios, t, rw);
1293 static inline void __account_c(struct io_stats *ios, int rw, int bytes)
1297 ios->cwrite_kb += bytes >> 10;
1300 ios->cread_kb += bytes >> 10;
1304 static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
1307 __account_c(&pci->io_stats, rw, bytes);
1309 if (per_process_stats) {
1310 struct io_stats *ios = find_process_io_stats(t->pid);
1312 __account_c(ios, rw, bytes);
1316 static inline void __account_issue(struct io_stats *ios, int rw,
1321 ios->iwrite_kb += bytes >> 10;
1324 ios->iread_kb += bytes >> 10;
1328 static inline void account_issue(struct blk_io_trace *t,
1329 struct per_cpu_info *pci, int rw)
1331 __account_issue(&pci->io_stats, rw, t->bytes);
1333 if (per_process_stats) {
1334 struct io_stats *ios = find_process_io_stats(t->pid);
1336 __account_issue(ios, rw, t->bytes);
1340 static inline void __account_unplug(struct io_stats *ios, int timer)
1343 ios->timer_unplugs++;
1348 static inline void account_unplug(struct blk_io_trace *t,
1349 struct per_cpu_info *pci, int timer)
1351 __account_unplug(&pci->io_stats, timer);
1353 if (per_process_stats) {
1354 struct io_stats *ios = find_process_io_stats(t->pid);
1356 __account_unplug(ios, timer);
1360 static inline void __account_requeue(struct io_stats *ios,
1361 struct blk_io_trace *t, int rw)
1365 ios->iwrite_kb -= t_kb(t);
1368 ios->iread_kb -= t_kb(t);
1372 static inline void account_requeue(struct blk_io_trace *t,
1373 struct per_cpu_info *pci, int rw)
1375 __account_requeue(&pci->io_stats, t, rw);
1377 if (per_process_stats) {
1378 struct io_stats *ios = find_process_io_stats(t->pid);
1380 __account_requeue(ios, t, rw);
1384 static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
1385 struct blk_io_trace *t, char *act)
1387 process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
1390 static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
1391 struct blk_io_trace *t, char *act)
1393 process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
1396 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
1399 process_fmt(act, pci, t, -1, 0, NULL);
1402 static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
1403 struct blk_io_trace *t, char *act)
1405 process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
1408 static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
1409 struct blk_io_trace *t, char *act)
1412 log_track_frontmerge(pdi, t);
1414 process_fmt(act, pci, t, -1ULL, 0, NULL);
1417 static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
1420 process_fmt(act, pci, t, -1ULL, 0, NULL);
1423 static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
1426 process_fmt(act, pci, t, -1ULL, 0, NULL);
1429 static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
1432 process_fmt(act, pci, t, -1ULL, 0, NULL);
1435 static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
1438 process_fmt(act, pci, t, -1ULL, 0, NULL);
1441 static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
1443 unsigned char *buf = (unsigned char *) t + sizeof(*t);
1445 process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
1448 static void dump_trace_pc(struct blk_io_trace *t, struct per_dev_info *pdi,
1449 struct per_cpu_info *pci)
1451 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1452 int act = t->action & 0xffff;
1455 case __BLK_TA_QUEUE:
1456 log_generic(pci, t, "Q");
1457 account_pc_queue(t, pci, w);
1459 case __BLK_TA_GETRQ:
1460 log_generic(pci, t, "G");
1462 case __BLK_TA_SLEEPRQ:
1463 log_generic(pci, t, "S");
1465 case __BLK_TA_REQUEUE:
1467 * can happen if we miss traces, don't let it go
1470 if (pdi->cur_depth[w])
1471 pdi->cur_depth[w]--;
1472 account_pc_requeue(t, pci, w);
1473 log_generic(pci, t, "R");
1475 case __BLK_TA_ISSUE:
1476 account_pc_issue(t, pci, w);
1477 pdi->cur_depth[w]++;
1478 if (pdi->cur_depth[w] > pdi->max_depth[w])
1479 pdi->max_depth[w] = pdi->cur_depth[w];
1480 log_pc(pci, t, "D");
1482 case __BLK_TA_COMPLETE:
1483 if (pdi->cur_depth[w])
1484 pdi->cur_depth[w]--;
1485 log_pc(pci, t, "C");
1486 account_pc_c(t, pci, w);
1488 case __BLK_TA_INSERT:
1489 log_pc(pci, t, "I");
1492 fprintf(stderr, "Bad pc action %x\n", act);
1497 static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
1498 struct per_cpu_info *pci)
1500 int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
1501 int act = t->action & 0xffff;
1504 case __BLK_TA_QUEUE:
1505 log_track_queue(pdi, t);
1506 account_queue(t, pci, w);
1507 log_queue(pci, t, "Q");
1509 case __BLK_TA_INSERT:
1510 log_insert(pdi, pci, t, "I");
1512 case __BLK_TA_BACKMERGE:
1513 account_m(t, pci, w);
1514 log_merge(pdi, pci, t, "M");
1516 case __BLK_TA_FRONTMERGE:
1517 account_m(t, pci, w);
1518 log_merge(pdi, pci, t, "F");
1520 case __BLK_TA_GETRQ:
1521 log_track_getrq(pdi, t);
1522 log_generic(pci, t, "G");
1524 case __BLK_TA_SLEEPRQ:
1525 log_generic(pci, t, "S");
1527 case __BLK_TA_REQUEUE:
1529 * can happen if we miss traces, don't let it go
1532 if (pdi->cur_depth[w])
1533 pdi->cur_depth[w]--;
1534 account_requeue(t, pci, w);
1535 log_queue(pci, t, "R");
1537 case __BLK_TA_ISSUE:
1538 account_issue(t, pci, w);
1539 pdi->cur_depth[w]++;
1540 if (pdi->cur_depth[w] > pdi->max_depth[w])
1541 pdi->max_depth[w] = pdi->cur_depth[w];
1542 log_issue(pdi, pci, t, "D");
1544 case __BLK_TA_COMPLETE:
1545 if (pdi->cur_depth[w])
1546 pdi->cur_depth[w]--;
1547 account_c(t, pci, w, t->bytes);
1548 log_complete(pdi, pci, t, "C");
1551 log_action(pci, t, "P");
1553 case __BLK_TA_UNPLUG_IO:
1554 account_unplug(t, pci, 0);
1555 log_unplug(pci, t, "U");
1557 case __BLK_TA_UNPLUG_TIMER:
1558 account_unplug(t, pci, 1);
1559 log_unplug(pci, t, "UT");
1561 case __BLK_TA_SPLIT:
1562 log_split(pci, t, "X");
1564 case __BLK_TA_BOUNCE:
1565 log_generic(pci, t, "B");
1567 case __BLK_TA_REMAP:
1568 log_generic(pci, t, "A");
1571 fprintf(stderr, "Bad fs action %x\n", t->action);
1576 static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
1577 struct per_dev_info *pdi)
1580 if (t->action & BLK_TC_ACT(BLK_TC_PC))
1581 dump_trace_pc(t, pdi, pci);
1583 dump_trace_fs(t, pdi, pci);
1587 pdi->first_reported_time = t->time;
1591 output_binary(t, sizeof(*t) + t->pdu_len);
1595 * print in a proper way, not too small and not too big. if more than
1596 * 1000,000K, turn into M and so on
1598 static char *size_cnv(char *dst, unsigned long long num, int in_kb)
1600 char suff[] = { '\0', 'K', 'M', 'G', 'P' };
1606 while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
1611 sprintf(dst, "%'8Lu%c", num, suff[i]);
1615 static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
1618 static char x[256], y[256];
1620 fprintf(ofp, "%s\n", msg);
1622 fprintf(ofp, " Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1));
1623 fprintf(ofp, " Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1));
1624 fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1));
1625 fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1));
1626 fprintf(ofp, " Reads Requeued: %s\t\t", size_cnv(x, ios->rrqueue, 0));
1627 fprintf(ofp, " Writes Requeued: %s\n", size_cnv(x, ios->wrqueue, 0));
1628 fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1));
1629 fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1));
1630 fprintf(ofp, " Read Merges: %s, %siB\t", size_cnv(x, ios->mreads, 0), size_cnv(y, ios->mread_kb, 1));
1631 fprintf(ofp, " Write Merges: %s, %siB\n", size_cnv(x, ios->mwrites, 0), size_cnv(y, ios->mwrite_kb, 1));
1633 fprintf(ofp, " Read depth: %'8u%8c\t", pdi->max_depth[0], ' ');
1634 fprintf(ofp, " Write depth: %'8u\n", pdi->max_depth[1]);
1636 if (ios->qreads_pc || ios->qwrites_pc || ios->ireads_pc || ios->iwrites_pc ||
1637 ios->rrqueue_pc || ios->wrqueue_pc || ios->creads_pc || ios->cwrites_pc) {
1638 fprintf(ofp, " PC Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads_pc, 0), size_cnv(y, ios->qread_kb_pc, 1));
1639 fprintf(ofp, " PC Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites_pc, 0), size_cnv(y, ios->qwrite_kb_pc, 1));
1640 fprintf(ofp, " PC Read Disp.: %s, %siB\t", size_cnv(x, ios->ireads_pc, 0), size_cnv(y, ios->iread_kb_pc, 1));
1641 fprintf(ofp, " PC Write Disp.: %s, %siB\n", size_cnv(x, ios->iwrites_pc, 0), size_cnv(y, ios->iwrite_kb_pc, 1));
1642 fprintf(ofp, " PC Reads Req.: %s\t\t", size_cnv(x, ios->rrqueue_pc, 0));
1643 fprintf(ofp, " PC Writes Req.: %s\n", size_cnv(x, ios->wrqueue_pc, 0));
1644 fprintf(ofp, " PC Reads Compl.: %s\t\t", size_cnv(x, ios->creads_pc, 0));
1645 fprintf(ofp, " PC Writes Compl.: %s\n", size_cnv(x, ios->cwrites, 0));
1647 fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' ');
1648 fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs);
1651 static void dump_wait_stats(struct per_process_info *ppi)
1653 unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
1654 unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
1655 unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
1656 unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
1657 unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
1658 unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
1660 fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
1661 fprintf(ofp, " Allocation wait: %'8lu\n", wawait);
1662 fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' ');
1663 fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait);
1664 fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
1665 fprintf(ofp, " Completion wait: %'8lu\n", wcwait);
1668 static int ppi_name_compare(const void *p1, const void *p2)
1670 struct per_process_info *ppi1 = *((struct per_process_info **) p1);
1671 struct per_process_info *ppi2 = *((struct per_process_info **) p2);
1674 res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
1676 res = ppi1->ppm->pid > ppi2->ppm->pid;
1681 static void sort_process_list(void)
1683 struct per_process_info **ppis;
1684 struct per_process_info *ppi;
1687 ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
1692 ppi = ppi->list_next;
1695 qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
1697 i = ppi_list_entries - 1;
1702 ppi->list_next = ppi_list;
1710 static void show_process_stats(void)
1712 struct per_process_info *ppi;
1714 sort_process_list();
1718 struct process_pid_map *ppm = ppi->ppm;
1721 if (ppi->more_than_one)
1722 sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
1724 sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
1726 dump_io_stats(NULL, &ppi->io_stats, name);
1727 dump_wait_stats(ppi);
1728 ppi = ppi->list_next;
1734 static void show_device_and_cpu_stats(void)
1736 struct per_dev_info *pdi;
1737 struct per_cpu_info *pci;
1738 struct io_stats total, *ios;
1739 unsigned long long rrate, wrate, msec;
1740 int i, j, pci_events;
1741 char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
1745 for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
1747 memset(&total, 0, sizeof(total));
1753 for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
1757 ios = &pci->io_stats;
1758 total.qreads += ios->qreads;
1759 total.qwrites += ios->qwrites;
1760 total.creads += ios->creads;
1761 total.cwrites += ios->cwrites;
1762 total.mreads += ios->mreads;
1763 total.mwrites += ios->mwrites;
1764 total.ireads += ios->ireads;
1765 total.iwrites += ios->iwrites;
1766 total.rrqueue += ios->rrqueue;
1767 total.wrqueue += ios->wrqueue;
1768 total.qread_kb += ios->qread_kb;
1769 total.qwrite_kb += ios->qwrite_kb;
1770 total.cread_kb += ios->cread_kb;
1771 total.cwrite_kb += ios->cwrite_kb;
1772 total.iread_kb += ios->iread_kb;
1773 total.iwrite_kb += ios->iwrite_kb;
1774 total.mread_kb += ios->mread_kb;
1775 total.mwrite_kb += ios->mwrite_kb;
1777 total.qreads_pc += ios->qreads_pc;
1778 total.qwrites_pc += ios->qwrites_pc;
1779 total.creads_pc += ios->creads_pc;
1780 total.cwrites_pc += ios->cwrites_pc;
1781 total.ireads_pc += ios->ireads_pc;
1782 total.iwrites_pc += ios->iwrites_pc;
1783 total.rrqueue_pc += ios->rrqueue_pc;
1784 total.wrqueue_pc += ios->wrqueue_pc;
1785 total.qread_kb_pc += ios->qread_kb_pc;
1786 total.qwrite_kb_pc += ios->qwrite_kb_pc;
1787 total.iread_kb_pc += ios->iread_kb_pc;
1788 total.iwrite_kb_pc += ios->iwrite_kb_pc;
1790 total.timer_unplugs += ios->timer_unplugs;
1791 total.io_unplugs += ios->io_unplugs;
1793 snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
1794 j, get_dev_name(pdi, name, sizeof(name)));
1795 dump_io_stats(pdi, ios, line);
1799 if (pci_events > 1) {
1801 snprintf(line, sizeof(line) - 1, "Total (%s):",
1802 get_dev_name(pdi, name, sizeof(name)));
1803 dump_io_stats(NULL, &total, line);
1807 msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
1809 rrate = 1000 * total.cread_kb / msec;
1810 wrate = 1000 * total.cwrite_kb / msec;
1813 fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
1815 fprintf(ofp, "Events (%s): %'Lu entries\n",
1816 get_dev_name(pdi, line, sizeof(line)), pdi->events);
1818 collect_pdi_skips(pdi);
1819 if (!pdi->skips && !pdi->events)
1822 ratio = 100.0 * ((double)pdi->seq_skips /
1823 (double)(pdi->events + pdi->seq_skips));
1824 fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
1825 pdi->skips, pdi->seq_skips, ratio);
1829 static void find_genesis(void)
1831 struct trace *t = trace_list;
1833 genesis_time = -1ULL;
1835 if (t->bit->time < genesis_time)
1836 genesis_time = t->bit->time;
1841 /* The time stamp record will usually be the first
1842 * record in the trace, but not always.
1845 && start_timestamp != genesis_time) {
1846 long delta = genesis_time - start_timestamp;
1848 abs_start_time.tv_sec += SECONDS(delta);
1849 abs_start_time.tv_nsec += NANO_SECONDS(delta);
1850 if (abs_start_time.tv_nsec < 0) {
1851 abs_start_time.tv_nsec += 1000000000;
1852 abs_start_time.tv_sec -= 1;
1854 if (abs_start_time.tv_nsec > 1000000000) {
1855 abs_start_time.tv_nsec -= 1000000000;
1856 abs_start_time.tv_sec += 1;
1861 static inline int check_stopwatch(struct blk_io_trace *bit)
1863 if (bit->time < stopwatch_end &&
1864 bit->time >= stopwatch_start)
1871 * return youngest entry read
1873 static int sort_entries(unsigned long long *youngest)
1875 struct per_dev_info *pdi = NULL;
1876 struct per_cpu_info *pci = NULL;
1883 while ((t = trace_list) != NULL) {
1884 struct blk_io_trace *bit = t->bit;
1886 trace_list = t->next;
1888 bit->time -= genesis_time;
1890 if (bit->time < *youngest || !*youngest)
1891 *youngest = bit->time;
1893 if (!pdi || pdi->dev != bit->device) {
1894 pdi = get_dev_info(bit->device);
1898 if (!pci || pci->cpu != bit->cpu)
1899 pci = get_cpu_info(pdi, bit->cpu);
1901 if (bit->sequence < pci->smallest_seq_read)
1902 pci->smallest_seq_read = bit->sequence;
1904 if (check_stopwatch(bit)) {
1910 if (trace_rb_insert_sort(t))
1918 * to continue, we must have traces from all online cpus in the tree
1920 static int check_cpu_map(struct per_dev_info *pdi)
1922 unsigned long *cpu_map;
1929 * create a map of the cpus we have traces for
1931 cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
1932 n = rb_first(&rb_sort_root);
1934 __t = rb_entry(n, struct trace, rb_node);
1935 cpu = __t->bit->cpu;
1937 cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
1942 * we can't continue if pdi->cpu_map has entries set that we don't
1943 * have in the sort rbtree. the opposite is not a problem, though
1946 for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
1947 if (pdi->cpu_map[i] & ~(cpu_map[i])) {
1957 static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
1959 struct blk_io_trace *bit = t->bit;
1960 unsigned long expected_sequence;
1961 struct per_cpu_info *pci;
1964 pci = get_cpu_info(pdi, bit->cpu);
1965 expected_sequence = pci->last_sequence + 1;
1967 if (!expected_sequence) {
1969 * 1 should be the first entry, just allow it
1971 if (bit->sequence == 1)
1973 if (bit->sequence == pci->smallest_seq_read)
1976 return check_cpu_map(pdi);
1979 if (bit->sequence == expected_sequence)
1983 * we may not have seen that sequence yet. if we are not doing
1984 * the final run, break and wait for more entries.
1986 if (expected_sequence < pci->smallest_seq_read) {
1987 __t = trace_rb_find_last(pdi, pci, expected_sequence);
1991 __put_trace_last(pdi, __t);
1993 } else if (!force) {
1997 if (check_current_skips(pci, bit->sequence))
2000 if (expected_sequence < bit->sequence)
2001 insert_skip(pci, expected_sequence, bit->sequence - 1);
2006 static void show_entries_rb(int force)
2008 struct per_dev_info *pdi = NULL;
2009 struct per_cpu_info *pci = NULL;
2010 struct blk_io_trace *bit;
2014 while ((n = rb_first(&rb_sort_root)) != NULL) {
2015 if (is_done() && !force && !pipeline)
2018 t = rb_entry(n, struct trace, rb_node);
2021 if (read_sequence - t->read_sequence < 1 && !force)
2024 if (!pdi || pdi->dev != bit->device) {
2025 pdi = get_dev_info(bit->device);
2030 fprintf(stderr, "Unknown device ID? (%d,%d)\n",
2031 MAJOR(bit->device), MINOR(bit->device));
2035 if (check_sequence(pdi, t, force))
2038 if (!force && bit->time > last_allowed_time)
2041 check_time(pdi, bit);
2043 if (!pci || pci->cpu != bit->cpu)
2044 pci = get_cpu_info(pdi, bit->cpu);
2046 pci->last_sequence = bit->sequence;
2050 if (bit->action & (act_mask << BLK_TC_SHIFT))
2051 dump_trace(bit, pci, pdi);
2057 static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
2059 int ret, bytes_left, fl;
2062 if (block != *fdblock) {
2063 fl = fcntl(fd, F_GETFL);
2067 fcntl(fd, F_SETFL, fl | O_NONBLOCK);
2070 fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
2076 while (bytes_left > 0) {
2077 ret = read(fd, p, bytes_left);
2081 if (errno != EAGAIN) {
2087 * never do partial reads. we can return if we
2088 * didn't read anything and we should not block,
2089 * otherwise wait for data
2091 if ((bytes_left == bytes) && !block)
2105 static inline __u16 get_pdulen(struct blk_io_trace *bit)
2108 return bit->pdu_len;
2110 return __bswap_16(bit->pdu_len);
2113 static inline __u32 get_magic(struct blk_io_trace *bit)
2118 return __bswap_32(bit->magic);
2121 static int read_events(int fd, int always_block, int *fdblock)
2123 struct per_dev_info *pdi = NULL;
2124 unsigned int events = 0;
2126 while (!is_done() && events < rb_batch) {
2127 struct blk_io_trace *bit;
2129 int pdu_len, should_block, ret;
2134 should_block = !events || always_block;
2136 ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
2139 if (!events && ret < 0)
2145 * look at first trace to check whether we need to convert
2146 * data in the future
2148 if (data_is_native == -1 && check_data_endianness(bit->magic))
2151 magic = get_magic(bit);
2152 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2153 fprintf(stderr, "Bad magic %x\n", magic);
2157 pdu_len = get_pdulen(bit);
2159 void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2161 if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
2171 if (verify_trace(bit)) {
2177 * not a real trace, so grab and handle it here
2179 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2181 output_binary(bit, sizeof(*bit) + bit->pdu_len);
2186 memset(t, 0, sizeof(*t));
2188 t->read_sequence = read_sequence;
2190 t->next = trace_list;
2193 if (!pdi || pdi->dev != bit->device)
2194 pdi = get_dev_info(bit->device);
2196 if (bit->time > pdi->last_read_time)
2197 pdi->last_read_time = bit->time;
2206 * Managing input streams
2210 struct ms_stream *next;
2211 struct trace *first, *last;
2212 struct per_dev_info *pdi;
2216 #define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
2218 struct ms_stream *ms_head;
2219 struct ms_stream *ms_hash[256];
2221 static void ms_sort(struct ms_stream *msp);
2222 static int ms_prime(struct ms_stream *msp);
2224 static inline struct trace *ms_peek(struct ms_stream *msp)
2226 return (msp == NULL) ? NULL : msp->first;
2229 static inline __u64 ms_peek_time(struct ms_stream *msp)
2231 return ms_peek(msp)->bit->time;
2234 static inline void ms_resort(struct ms_stream *msp)
2236 if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
2237 ms_head = msp->next;
2243 static inline void ms_deq(struct ms_stream *msp)
2245 msp->first = msp->first->next;
2248 if (!ms_prime(msp)) {
2249 ms_head = msp->next;
2258 static void ms_sort(struct ms_stream *msp)
2260 __u64 msp_t = ms_peek_time(msp);
2261 struct ms_stream *this_msp = ms_head;
2263 if (this_msp == NULL)
2265 else if (msp_t < ms_peek_time(this_msp)) {
2266 msp->next = this_msp;
2270 while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
2271 this_msp = this_msp->next;
2273 msp->next = this_msp->next;
2274 this_msp->next = msp;
2278 static int ms_prime(struct ms_stream *msp)
2283 struct per_dev_info *pdi = msp->pdi;
2284 struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
2285 struct blk_io_trace *bit = NULL;
2286 int ret, pdu_len, ndone = 0;
2288 for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
2290 ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
2294 if (data_is_native == -1 && check_data_endianness(bit->magic))
2297 magic = get_magic(bit);
2298 if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
2299 fprintf(stderr, "Bad magic %x\n", magic);
2304 pdu_len = get_pdulen(bit);
2306 void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
2307 ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
2319 if (verify_trace(bit))
2322 if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY)) {
2324 output_binary(bit, sizeof(*bit) + bit->pdu_len);
2331 if (bit->time > pdi->last_read_time)
2332 pdi->last_read_time = bit->time;
2335 memset(t, 0, sizeof(*t));
2338 if (msp->first == NULL)
2339 msp->first = msp->last = t;
2341 msp->last->next = t;
2351 if (bit) bit_free(bit);
2353 cpu_mark_offline(pdi, pci->cpu);
2360 static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
2362 struct ms_stream *msp = malloc(sizeof(*msp));
2365 msp->first = msp->last = NULL;
2375 static int setup_file(struct per_dev_info *pdi, int cpu)
2380 struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
2385 p = strdup(pdi->name);
2387 if (strcmp(dname, ".")) {
2389 p = strdup(pdi->name);
2390 strcpy(pdi->name, basename(p));
2395 len = sprintf(pci->fname, "%s/", input_dir);
2397 snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
2398 "%s.blktrace.%d", pdi->name, pci->cpu);
2399 if (stat(pci->fname, &st) < 0)
2404 pci->fd = open(pci->fname, O_RDONLY);
2410 printf("Input file %s added\n", pci->fname);
2411 cpu_mark_online(pdi, pci->cpu);
2414 ms_alloc(pdi, pci->cpu);
2419 static int handle(struct ms_stream *msp)
2422 struct per_dev_info *pdi;
2423 struct per_cpu_info *pci;
2424 struct blk_io_trace *bit;
2430 pci = get_cpu_info(pdi, msp->cpu);
2432 bit->time -= genesis_time;
2434 if (t->bit->time > stopwatch_end)
2437 pdi->last_reported_time = bit->time;
2438 if ((bit->action & (act_mask << BLK_TC_SHIFT))&&
2439 t->bit->time >= stopwatch_start)
2440 dump_trace(bit, pci, pdi);
2445 trace_rb_insert_last(pdi, t);
2455 * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
2456 * is given, then strip back down to 'foo' to avoid missing files.
2458 static int name_fixup(char *name)
2465 b = strstr(name, ".blktrace.");
2472 static int do_file(void)
2475 struct per_dev_info *pdi;
2478 * first prepare all files for reading
2480 for (i = 0; i < ndevices; i++) {
2482 ret = name_fixup(pdi->name);
2486 for (cpu = 0; setup_file(pdi, cpu); cpu++)
2491 * Get the initial time stamp
2494 genesis_time = ms_peek_time(ms_head);
2497 * Keep processing traces while any are left
2499 while (!is_done() && ms_head && handle(ms_head))
2505 static void do_pipe(int fd)
2507 unsigned long long youngest;
2508 int events, fdblock;
2510 last_allowed_time = -1ULL;
2512 while ((events = read_events(fd, 0, &fdblock)) > 0) {
2516 smallest_seq_read = -1U;
2519 if (sort_entries(&youngest))
2522 if (youngest > stopwatch_end)
2528 if (rb_sort_entries)
2532 static int do_fifo(void)
2536 if (!strcmp(pipename, "-"))
2537 fd = dup(STDIN_FILENO);
2539 fd = open(pipename, O_RDONLY);
2542 perror("dup stdin");
2551 static void show_stats(void)
2560 if (per_process_stats)
2561 show_process_stats();
2563 if (per_device_and_cpu_stats)
2564 show_device_and_cpu_stats();
2569 static void handle_sigint(__attribute__((__unused__)) int sig)
2575 * Extract start and duration times from a string, allowing
2576 * us to specify a time interval of interest within a trace.
2577 * Format: "duration" (start is zero) or "start:duration".
2579 static int find_stopwatch_interval(char *string)
2584 value = strtod(string, &sp);
2586 fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
2590 stopwatch_start = DOUBLE_TO_NANO_ULL(value);
2592 value = strtod(string, &sp);
2593 if (sp == string || *sp != '\0') {
2594 fprintf(stderr,"Invalid stopwatch duration time: %s\n",
2598 } else if (*sp != '\0') {
2599 fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
2602 stopwatch_end = DOUBLE_TO_NANO_ULL(value);
2603 if (stopwatch_end <= stopwatch_start) {
2604 fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
2605 stopwatch_start, stopwatch_end);
2612 static int is_pipe(const char *str)
2616 if (!strcmp(str, "-"))
2618 if (!stat(str, &st) && S_ISFIFO(st.st_mode))
2624 #define S_OPTS "a:A:b:D:d:f:F:hi:o:Oqstw:vV"
2625 static char usage_str[] = "\n\n" \
2626 "-i <file> | --input=<file>\n" \
2627 "[ -a <action field> | --act-mask=<action field> ]\n" \
2628 "[ -A <action mask> | --set-mask=<action mask> ]\n" \
2629 "[ -b <traces> | --batch=<traces> ]\n" \
2630 "[ -d <file> | --dump-binary=<file> ]\n" \
2631 "[ -D <dir> | --input-directory=<dir> ]\n" \
2632 "[ -f <format> | --format=<format> ]\n" \
2633 "[ -F <spec> | --format-spec=<spec> ]\n" \
2634 "[ -h | --hash-by-name ]\n" \
2635 "[ -o <file> | --output=<file> ]\n" \
2636 "[ -O | --no-text-output ]\n" \
2637 "[ -q | --quiet ]\n" \
2638 "[ -s | --per-program-stats ]\n" \
2639 "[ -t | --track-ios ]\n" \
2640 "[ -w <time> | --stopwatch=<time> ]\n" \
2641 "[ -v | --verbose ]\n" \
2642 "[ -V | --version ]\n\n" \
2643 "\t-b stdin read batching\n" \
2644 "\t-d Output file. If specified, binary data is written to file\n" \
2645 "\t-D Directory to prepend to input file names\n" \
2646 "\t-f Output format. Customize the output format. The format field\n" \
2647 "\t identifies can be found in the documentation\n" \
2648 "\t-F Format specification. Can be found in the documentation\n" \
2649 "\t-h Hash processes by name, not pid\n" \
2650 "\t-i Input file containing trace data, or '-' for stdin\n" \
2651 "\t-o Output file. If not given, output is stdout\n" \
2652 "\t-O Do NOT output text data\n" \
2653 "\t-q Quiet. Don't display any stats at the end of the trace\n" \
2654 "\t-s Show per-program io statistics\n" \
2655 "\t-t Track individual ios. Will tell you the time a request took\n" \
2656 "\t to get queued, to get dispatched, and to get completed\n" \
2657 "\t-w Only parse data between the given time interval in seconds.\n" \
2658 "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \
2659 "\t-v More verbose for marginal errors\n" \
2660 "\t-V Print program version info\n\n";
2662 static void usage(char *prog)
2664 fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
2667 int main(int argc, char *argv[])
2669 int i, c, ret, mode;
2670 int act_mask_tmp = 0;
2671 char *ofp_buffer = NULL;
2672 char *bin_ofp_buffer = NULL;
2674 while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
2677 i = find_mask_map(optarg);
2679 fprintf(stderr,"Invalid action mask %s\n",
2687 if ((sscanf(optarg, "%x", &i) != 1) ||
2688 !valid_act_opt(i)) {
2690 "Invalid set action mask %s/0x%x\n",
2697 if (is_pipe(optarg) && !pipeline) {
2699 pipename = strdup(optarg);
2700 } else if (resize_devices(optarg) != 0)
2707 output_name = optarg;
2713 rb_batch = atoi(optarg);
2715 rb_batch = RB_BATCH_DEFAULT;
2718 per_process_stats = 1;
2724 per_device_and_cpu_stats = 0;
2727 if (find_stopwatch_interval(optarg) != 0)
2731 set_all_format_specs(optarg);
2734 if (add_format_spec(optarg) != 0)
2738 ppi_hash_by_pid = 0;
2744 printf("%s version %s\n", argv[0], blkparse_version);
2747 dump_binary = optarg;
2755 while (optind < argc) {
2756 if (is_pipe(argv[optind]) && !pipeline) {
2758 pipename = strdup(argv[optind]);
2759 } else if (resize_devices(argv[optind]) != 0)
2764 if (!pipeline && !ndevices) {
2769 if (act_mask_tmp != 0)
2770 act_mask = act_mask_tmp;
2772 memset(&rb_sort_root, 0, sizeof(rb_sort_root));
2774 signal(SIGINT, handle_sigint);
2775 signal(SIGHUP, handle_sigint);
2776 signal(SIGTERM, handle_sigint);
2778 setlocale(LC_NUMERIC, "en_US");
2782 ofp = fdopen(STDOUT_FILENO, "w");
2787 snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
2788 ofp = fopen(ofname, "w");
2797 ofp_buffer = malloc(4096);
2798 if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
2805 dump_fp = fopen(dump_binary, "w");
2807 perror(dump_binary);
2811 bin_ofp_buffer = malloc(128 * 1024);
2812 if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
2813 perror("setvbuf binary");
2830 if (bin_ofp_buffer) {
2832 free(bin_ofp_buffer);