2 * Copyright (C) 2012 Fusion-io
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
13 * You should have received a copy of the GNU General Public License
14 * along with this program; if not, write to the Free Software
15 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 * Parts of this file were imported from Jens Axboe's blktrace sources (also GPL)
19 #include <sys/types.h>
28 #include <asm/types.h>
39 #define IO_HASH_TABLE_BITS 11
40 #define IO_HASH_TABLE_SIZE (1 << IO_HASH_TABLE_BITS)
41 static struct list_head io_hash_table[IO_HASH_TABLE_SIZE];
42 static u64 ios_in_flight = 0;
49 BLK_TC_READ = 1 << 0, /* reads */
50 BLK_TC_WRITE = 1 << 1, /* writes */
51 BLK_TC_FLUSH = 1 << 2, /* flush */
52 BLK_TC_SYNC = 1 << 3, /* sync */
53 BLK_TC_QUEUE = 1 << 4, /* queueing/merging */
54 BLK_TC_REQUEUE = 1 << 5, /* requeueing */
55 BLK_TC_ISSUE = 1 << 6, /* issue */
56 BLK_TC_COMPLETE = 1 << 7, /* completions */
57 BLK_TC_FS = 1 << 8, /* fs requests */
58 BLK_TC_PC = 1 << 9, /* pc requests */
59 BLK_TC_NOTIFY = 1 << 10, /* special message */
60 BLK_TC_AHEAD = 1 << 11, /* readahead */
61 BLK_TC_META = 1 << 12, /* metadata */
62 BLK_TC_DISCARD = 1 << 13, /* discard requests */
63 BLK_TC_DRV_DATA = 1 << 14, /* binary driver data */
64 BLK_TC_FUA = 1 << 15, /* fua requests */
66 BLK_TC_END = 1 << 15, /* we've run out of bits! */
69 #define BLK_TC_SHIFT (16)
70 #define BLK_TC_ACT(act) ((act) << BLK_TC_SHIFT)
71 #define BLK_DATADIR(a) (((a) >> BLK_TC_SHIFT) & (BLK_TC_READ | BLK_TC_WRITE))
77 __BLK_TA_QUEUE = 1, /* queued */
78 __BLK_TA_BACKMERGE, /* back merged to existing rq */
79 __BLK_TA_FRONTMERGE, /* front merge to existing rq */
80 __BLK_TA_GETRQ, /* allocated new request */
81 __BLK_TA_SLEEPRQ, /* sleeping on rq allocation */
82 __BLK_TA_REQUEUE, /* request requeued */
83 __BLK_TA_ISSUE, /* sent to driver */
84 __BLK_TA_COMPLETE, /* completed by driver */
85 __BLK_TA_PLUG, /* queue was plugged */
86 __BLK_TA_UNPLUG_IO, /* queue was unplugged by io */
87 __BLK_TA_UNPLUG_TIMER, /* queue was unplugged by timer */
88 __BLK_TA_INSERT, /* insert request */
89 __BLK_TA_SPLIT, /* bio was split */
90 __BLK_TA_BOUNCE, /* bio was bounced */
91 __BLK_TA_REMAP, /* bio was remapped */
92 __BLK_TA_ABORT, /* request aborted */
93 __BLK_TA_DRV_DATA, /* binary driver data */
99 enum blktrace_notify {
100 __BLK_TN_PROCESS = 0, /* establish pid/name mapping */
101 __BLK_TN_TIMESTAMP, /* include system clock */
102 __BLK_TN_MESSAGE, /* Character string message */
106 * Trace actions in full. Additionally, read or write is masked
108 #define BLK_TA_QUEUE (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
109 #define BLK_TA_BACKMERGE (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
110 #define BLK_TA_FRONTMERGE (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
111 #define BLK_TA_GETRQ (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
112 #define BLK_TA_SLEEPRQ (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
113 #define BLK_TA_REQUEUE (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
114 #define BLK_TA_ISSUE (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
115 #define BLK_TA_COMPLETE (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
116 #define BLK_TA_PLUG (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
117 #define BLK_TA_UNPLUG_IO (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
118 #define BLK_TA_UNPLUG_TIMER (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
119 #define BLK_TA_INSERT (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
120 #define BLK_TA_SPLIT (__BLK_TA_SPLIT)
121 #define BLK_TA_BOUNCE (__BLK_TA_BOUNCE)
122 #define BLK_TA_REMAP (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
123 #define BLK_TA_ABORT (__BLK_TA_ABORT | BLK_TC_ACT(BLK_TC_QUEUE))
124 #define BLK_TA_DRV_DATA (__BLK_TA_DRV_DATA | BLK_TC_ACT(BLK_TC_DRV_DATA))
126 #define BLK_TN_PROCESS (__BLK_TN_PROCESS | BLK_TC_ACT(BLK_TC_NOTIFY))
127 #define BLK_TN_TIMESTAMP (__BLK_TN_TIMESTAMP | BLK_TC_ACT(BLK_TC_NOTIFY))
128 #define BLK_TN_MESSAGE (__BLK_TN_MESSAGE | BLK_TC_ACT(BLK_TC_NOTIFY))
130 #define BLK_IO_TRACE_MAGIC 0x65617400
131 #define BLK_IO_TRACE_VERSION 0x07
135 struct blk_io_trace {
136 __u32 magic; /* MAGIC << 8 | version */
137 __u32 sequence; /* event number */
138 __u64 time; /* in nanoseconds */
139 __u64 sector; /* disk offset */
140 __u32 bytes; /* transfer length */
141 __u32 action; /* what happened */
142 __u32 pid; /* who did it */
143 __u32 device; /* device identifier (dev_t) */
144 __u32 cpu; /* on what cpu did it happen */
145 __u16 error; /* completion error */
146 __u16 pdu_len; /* length of data after this trace */
150 /* sector offset of this IO */
153 /* time this IO was dispatched */
155 /* time this IO was finished */
157 struct list_head hash_list;
161 #define MINORMASK ((1 << MINORBITS) - 1)
162 #define SECONDS(x) ((unsigned long long)(x) / 1000000000)
163 #define NANO_SECONDS(x) ((unsigned long long)(x) % 1000000000)
164 #define DOUBLE_TO_NANO_ULL(d) ((unsigned long long)((d) * 1000000000))
165 #define CHECK_MAGIC(t) (((t)->magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
167 void init_io_hash_table(void)
170 struct list_head *head;
172 for (i = 0; i < IO_HASH_TABLE_SIZE; i++) {
173 head = io_hash_table + i;
174 INIT_LIST_HEAD(head);
178 /* taken from the kernel hash.h */
179 static inline u64 hash_sector(u64 val)
183 /* Sigh, gcc can't optimise this alone like it does for 32 bits. */
198 /* High bits are more random, so use them. */
199 return hash >> (64 - IO_HASH_TABLE_BITS);
202 static int hash_table_insert(struct pending_io *ins_pio)
204 u64 sector = ins_pio->sector;
205 int slot = hash_sector(sector);
206 struct list_head *head;
207 struct pending_io *pio;
209 head = io_hash_table + slot;
210 list_for_each_entry(pio, head, hash_list) {
211 if (pio->sector == sector)
214 list_add_tail(&ins_pio->hash_list, head);
218 static struct pending_io *hash_table_search(u64 sector)
220 int slot = hash_sector(sector);
221 struct list_head *head;
222 struct pending_io *pio;
224 head = io_hash_table + slot;
225 list_for_each_entry(pio, head, hash_list) {
226 if (pio->sector == sector)
232 static int hash_dispatched_io(struct blk_io_trace *io)
234 struct pending_io *pio;
237 pio = calloc(1, sizeof(*pio));
238 pio->sector = io->sector;
239 pio->dispatch_time = io->time;
241 ret = hash_table_insert(pio);
242 if (ret == -EEXIST) {
243 /* crud, the IO isn't here */
249 static struct pending_io *hash_completed_io(struct blk_io_trace *io)
251 struct pending_io *pio;
253 pio = hash_table_search(io->sector);
260 static void handle_notify(struct trace *trace)
262 struct blk_io_trace *io = trace->io;
263 void *payload = (char *)io + sizeof(*io);
267 if (io->action != BLK_TN_TIMESTAMP)
270 if (io->pdu_len != sizeof(two32))
273 memcpy(two32, payload, sizeof(two32));
274 trace->start_timestamp = io->time;
275 trace->abs_start_time.tv_sec = two32[0];
276 trace->abs_start_time.tv_nsec = two32[1];
277 if (trace->abs_start_time.tv_nsec < 0) {
278 trace->abs_start_time.tv_sec--;
279 trace->abs_start_time.tv_nsec += 1000000000;
283 int next_record(struct trace *trace)
285 int skip = trace->io->pdu_len;
288 trace->cur += sizeof(*trace->io) + skip;
289 offset = trace->cur - trace->start;
290 if (offset >= trace->len)
293 trace->io = (struct blk_io_trace *)trace->cur;
297 void first_record(struct trace *trace)
299 trace->cur = trace->start;
300 trace->io = (struct blk_io_trace *)trace->cur;
303 int is_io_event(struct blk_io_trace *test)
306 if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
308 if (test->action == BLK_TN_MESSAGE) {
309 int len = test->pdu_len;
312 message = (char *)(test + 1);
313 if (strncmp(message, "fio ", 4) == 0) {
320 u64 find_last_time(struct trace *trace)
322 char *p = trace->start + trace->len;
323 struct blk_io_trace *test;
327 if (trace->len < sizeof(*trace->io))
329 p -= sizeof(*trace->io);
330 while (p >= trace->start) {
331 test = (struct blk_io_trace *)p;
332 if (CHECK_MAGIC(test) && is_io_event(test)) {
333 u64 offset = p - trace->start;
334 if (offset + sizeof(*test) + test->pdu_len == trace->len) {
340 if (search_len > 8192) {
345 /* searching backwards didn't work out, we'll have to scan the file */
348 if (is_io_event(trace->io))
349 found = trace->io->time;
350 if (next_record(trace))
357 int parse_fio_bank_message(struct trace *trace, u64 *bank_ret, u64 *offset_ret,
363 struct blk_io_trace *test = trace->io;
364 int len = test->pdu_len;
369 if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
371 if (test->action != BLK_TN_MESSAGE)
374 /* the message is fio rw bank offset num_banks */
377 message = (char *)(test + 1);
378 if (strncmp(message, "fio r ", 6) != 0)
381 message = strndup(message, len);
382 s = strchr(message, ' ');
390 bank = strtoll(s, &next, 10);
395 offset = strtoll(s, &next, 10);
400 num_banks = strtoll(s, &next, 10);
405 *offset_ret = offset;
406 *num_banks_ret = num_banks;
414 void find_highest_offset(struct trace *trace, u64 *max_ret, u64 *max_bank_ret,
420 u64 max_bank_offset = 0;
424 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
425 found = trace->io->sector << 9;
426 found += trace->io->bytes;
434 if (!parse_fio_bank_message(trace, &bank,
435 &offset, &num_banks)) {
438 if (offset > max_bank_offset)
439 max_bank_offset = offset;
442 if (next_record(trace))
447 *max_bank_ret = max_bank;
448 *max_offset_ret = max_bank_offset;
451 int filter_outliers(struct trace *trace, u64 max_offset,
452 u64 *yzoom_min, u64 *yzoom_max)
455 u64 max_per_bucket[11];
456 u64 bytes_per_bucket = max_offset / 10;
460 memset(hits, 0, sizeof(int) * 11);
461 memset(max_per_bucket, 0, sizeof(u64) * 11);
464 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
465 u64 top = (trace->io->sector << 9) + trace->io->bytes;
466 slot = (int)(top / bytes_per_bucket);
468 if (top > max_per_bucket[slot])
469 max_per_bucket[slot] = top;
471 if (next_record(trace))
475 for (slot = 0; slot < 11; slot++) {
476 if (hits[slot] > fat_count) {
477 fat_count = hits[slot];
481 *yzoom_max = max_offset;
482 for (slot = 10; slot >= 0; slot--) {
483 double d = hits[slot];
485 if (d >= (double)fat_count * .05) {
486 *yzoom_max = max_per_bucket[slot];
492 for (slot = 0; slot < 10; slot++) {
493 double d = hits[slot];
495 if (d >= (double)fat_count * .05) {
496 *yzoom_min = slot * bytes_per_bucket;
503 static char *find_trace_file(char *filename)
511 ret = stat(filename, &st);
513 return strdup(filename);
515 snprintf(line, 1024, "%s.%s", filename, "dump");
516 ret = stat(line, &st);
520 try = strdup(filename);
521 dot = strrchr(try, '.');
522 if (!dot || strcmp(".dump", dot) != 0) {
525 snprintf(line, 1024, "%s%s", try, ".blktrace.0");
526 ret = stat(line, &st);
528 blktrace_to_dump(try);
529 snprintf(line, 1024, "%s.%s", try, "dump");
530 ret = stat(line, &st);
540 struct trace *open_trace(char *filename)
547 char *found_filename;
549 trace = calloc(1, sizeof(*trace));
551 fprintf(stderr, "unable to allocate memory for trace\n");
555 found_filename = find_trace_file(filename);
556 if (!found_filename) {
557 fprintf(stderr, "Unable to find trace file %s\n", filename);
560 filename = found_filename;
562 fd = open(filename, O_RDONLY);
564 fprintf(stderr, "Unable to open trace file %s err %s\n", filename, strerror(errno));
567 ret = fstat(fd, &st);
569 fprintf(stderr, "stat failed on %s err %s\n", filename, strerror(errno));
572 p = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
573 if (p == MAP_FAILED) {
574 fprintf(stderr, "Unable to mmap trace file %s, err %s\n", filename, strerror(errno));
578 trace->len = st.st_size;
581 trace->io = (struct blk_io_trace *)p;
590 static inline int tput_event(struct trace *trace)
592 if (trace->found_completion)
593 return __BLK_TA_COMPLETE;
594 if (trace->found_issue)
595 return __BLK_TA_ISSUE;
596 if (trace->found_queue)
597 return __BLK_TA_QUEUE;
599 return __BLK_TA_COMPLETE;
602 static inline int io_event(struct trace *trace)
604 if (trace->found_queue)
605 return __BLK_TA_QUEUE;
606 if (trace->found_issue)
607 return __BLK_TA_ISSUE;
608 if (trace->found_completion)
609 return __BLK_TA_COMPLETE;
611 return __BLK_TA_COMPLETE;
614 void add_tput(struct trace *trace, struct graph_line_data *gld)
616 struct blk_io_trace *io = trace->io;
617 int action = io->action & 0xffff;
620 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
623 if (action != tput_event(trace))
626 seconds = SECONDS(io->time);
627 if (seconds > gld->seconds) {
628 fprintf(stderr, "Bad record %d %d %d\n", seconds, gld->seconds, action);
632 gld->data[seconds].sum += io->bytes;
633 gld->data[seconds].count = 1;
634 if (gld->data[seconds].sum > gld->max)
635 gld->max = gld->data[seconds].sum;
638 void add_io(struct trace *trace, struct graph_dot_data *gdd_writes,
639 struct graph_dot_data *gdd_reads)
641 struct blk_io_trace *io = trace->io;
642 int action = io->action & 0xffff;
645 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
648 if (action != io_event(trace))
651 offset = io->sector << 9;
653 if (BLK_DATADIR(io->action) & BLK_TC_READ)
654 set_gdd_bit(gdd_reads, offset, io->bytes, io->time);
655 else if (BLK_DATADIR(io->action) & BLK_TC_WRITE)
656 set_gdd_bit(gdd_writes, offset, io->bytes, io->time);
659 void add_pending_io(struct trace *trace, struct graph_line_data *gld)
663 struct blk_io_trace *io = trace->io;
664 int action = io->action & 0xffff;
667 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
670 if (action != __BLK_TA_ISSUE)
673 seconds = SECONDS(io->time);
674 if (seconds > gld->seconds) {
675 fprintf(stderr, "Bad record %d %d\n", seconds, gld->seconds);
679 ret = hash_dispatched_io(trace->io);
685 gld->data[seconds].sum += ios_in_flight;
686 gld->data[seconds].count++;
688 avg = (double)gld->data[seconds].sum / gld->data[seconds].count;
689 if (gld->max < (u64)avg) {
694 void add_completed_io(struct trace *trace,
695 struct graph_line_data *latency_gld)
697 struct blk_io_trace *io = trace->io;
699 int action = io->action & 0xffff;
700 struct pending_io *pio;
704 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
707 if (action != __BLK_TA_COMPLETE)
710 seconds = SECONDS(io->time);
712 pio = hash_completed_io(trace->io);
716 if (ios_in_flight > 0)
718 if (io->time >= pio->dispatch_time) {
719 latency = io->time - pio->dispatch_time;
720 latency_gld->data[seconds].sum += latency;
721 latency_gld->data[seconds].count++;
724 list_del(&pio->hash_list);
727 avg = (double)latency_gld->data[seconds].sum /
728 latency_gld->data[seconds].count;
729 if (latency_gld->max < (u64)avg) {
730 latency_gld->max = avg;
734 void add_iop(struct trace *trace, struct graph_line_data *gld)
736 struct blk_io_trace *io = trace->io;
737 int action = io->action & 0xffff;
740 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
743 /* iops and tput use the same events */
744 if (action != tput_event(trace))
747 seconds = SECONDS(io->time);
748 if (seconds > gld->seconds) {
749 fprintf(stderr, "Bad record %d %d\n", seconds, gld->seconds);
753 gld->data[seconds].sum += 1;
754 gld->data[seconds].count = 1;
755 if (gld->data[seconds].sum > gld->max)
756 gld->max = gld->data[seconds].sum;
759 void check_record(struct trace *trace)
761 struct blk_io_trace *io = trace->io;
762 int action = io->action & 0xffff;
764 if (!(io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
766 case __BLK_TA_COMPLETE:
767 trace->found_completion = 1;
770 trace->found_issue = 1;
773 trace->found_queue = 1;
777 handle_notify(trace);