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;
44 extern int plot_io_action;
50 BLK_TC_READ = 1 << 0, /* reads */
51 BLK_TC_WRITE = 1 << 1, /* writes */
52 BLK_TC_FLUSH = 1 << 2, /* flush */
53 BLK_TC_SYNC = 1 << 3, /* sync */
54 BLK_TC_QUEUE = 1 << 4, /* queueing/merging */
55 BLK_TC_REQUEUE = 1 << 5, /* requeueing */
56 BLK_TC_ISSUE = 1 << 6, /* issue */
57 BLK_TC_COMPLETE = 1 << 7, /* completions */
58 BLK_TC_FS = 1 << 8, /* fs requests */
59 BLK_TC_PC = 1 << 9, /* pc requests */
60 BLK_TC_NOTIFY = 1 << 10, /* special message */
61 BLK_TC_AHEAD = 1 << 11, /* readahead */
62 BLK_TC_META = 1 << 12, /* metadata */
63 BLK_TC_DISCARD = 1 << 13, /* discard requests */
64 BLK_TC_DRV_DATA = 1 << 14, /* binary driver data */
65 BLK_TC_FUA = 1 << 15, /* fua requests */
67 BLK_TC_END = 1 << 15, /* we've run out of bits! */
70 #define BLK_TC_SHIFT (16)
71 #define BLK_TC_ACT(act) ((act) << BLK_TC_SHIFT)
72 #define BLK_DATADIR(a) (((a) >> BLK_TC_SHIFT) & (BLK_TC_READ | BLK_TC_WRITE))
78 __BLK_TA_QUEUE = 1, /* queued */
79 __BLK_TA_BACKMERGE, /* back merged to existing rq */
80 __BLK_TA_FRONTMERGE, /* front merge to existing rq */
81 __BLK_TA_GETRQ, /* allocated new request */
82 __BLK_TA_SLEEPRQ, /* sleeping on rq allocation */
83 __BLK_TA_REQUEUE, /* request requeued */
84 __BLK_TA_ISSUE, /* sent to driver */
85 __BLK_TA_COMPLETE, /* completed by driver */
86 __BLK_TA_PLUG, /* queue was plugged */
87 __BLK_TA_UNPLUG_IO, /* queue was unplugged by io */
88 __BLK_TA_UNPLUG_TIMER, /* queue was unplugged by timer */
89 __BLK_TA_INSERT, /* insert request */
90 __BLK_TA_SPLIT, /* bio was split */
91 __BLK_TA_BOUNCE, /* bio was bounced */
92 __BLK_TA_REMAP, /* bio was remapped */
93 __BLK_TA_ABORT, /* request aborted */
94 __BLK_TA_DRV_DATA, /* binary driver data */
97 #define BLK_TA_MASK ((1 << BLK_TC_SHIFT) - 1)
102 enum blktrace_notify {
103 __BLK_TN_PROCESS = 0, /* establish pid/name mapping */
104 __BLK_TN_TIMESTAMP, /* include system clock */
105 __BLK_TN_MESSAGE, /* Character string message */
109 * Trace actions in full. Additionally, read or write is masked
111 #define BLK_TA_QUEUE (__BLK_TA_QUEUE | BLK_TC_ACT(BLK_TC_QUEUE))
112 #define BLK_TA_BACKMERGE (__BLK_TA_BACKMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
113 #define BLK_TA_FRONTMERGE (__BLK_TA_FRONTMERGE | BLK_TC_ACT(BLK_TC_QUEUE))
114 #define BLK_TA_GETRQ (__BLK_TA_GETRQ | BLK_TC_ACT(BLK_TC_QUEUE))
115 #define BLK_TA_SLEEPRQ (__BLK_TA_SLEEPRQ | BLK_TC_ACT(BLK_TC_QUEUE))
116 #define BLK_TA_REQUEUE (__BLK_TA_REQUEUE | BLK_TC_ACT(BLK_TC_REQUEUE))
117 #define BLK_TA_ISSUE (__BLK_TA_ISSUE | BLK_TC_ACT(BLK_TC_ISSUE))
118 #define BLK_TA_COMPLETE (__BLK_TA_COMPLETE| BLK_TC_ACT(BLK_TC_COMPLETE))
119 #define BLK_TA_PLUG (__BLK_TA_PLUG | BLK_TC_ACT(BLK_TC_QUEUE))
120 #define BLK_TA_UNPLUG_IO (__BLK_TA_UNPLUG_IO | BLK_TC_ACT(BLK_TC_QUEUE))
121 #define BLK_TA_UNPLUG_TIMER (__BLK_TA_UNPLUG_TIMER | BLK_TC_ACT(BLK_TC_QUEUE))
122 #define BLK_TA_INSERT (__BLK_TA_INSERT | BLK_TC_ACT(BLK_TC_QUEUE))
123 #define BLK_TA_SPLIT (__BLK_TA_SPLIT)
124 #define BLK_TA_BOUNCE (__BLK_TA_BOUNCE)
125 #define BLK_TA_REMAP (__BLK_TA_REMAP | BLK_TC_ACT(BLK_TC_QUEUE))
126 #define BLK_TA_ABORT (__BLK_TA_ABORT | BLK_TC_ACT(BLK_TC_QUEUE))
127 #define BLK_TA_DRV_DATA (__BLK_TA_DRV_DATA | BLK_TC_ACT(BLK_TC_DRV_DATA))
129 #define BLK_TN_PROCESS (__BLK_TN_PROCESS | BLK_TC_ACT(BLK_TC_NOTIFY))
130 #define BLK_TN_TIMESTAMP (__BLK_TN_TIMESTAMP | BLK_TC_ACT(BLK_TC_NOTIFY))
131 #define BLK_TN_MESSAGE (__BLK_TN_MESSAGE | BLK_TC_ACT(BLK_TC_NOTIFY))
133 #define BLK_IO_TRACE_MAGIC 0x65617400
134 #define BLK_IO_TRACE_VERSION 0x07
138 struct blk_io_trace {
139 __u32 magic; /* MAGIC << 8 | version */
140 __u32 sequence; /* event number */
141 __u64 time; /* in nanoseconds */
142 __u64 sector; /* disk offset */
143 __u32 bytes; /* transfer length */
144 __u32 action; /* what happened */
145 __u32 pid; /* who did it */
146 __u32 device; /* device identifier (dev_t) */
147 __u32 cpu; /* on what cpu did it happen */
148 __u16 error; /* completion error */
149 __u16 pdu_len; /* length of data after this trace */
153 /* sector offset of this IO */
156 /* time this IO was dispatched */
158 /* time this IO was finished */
160 struct list_head hash_list;
164 #define MINORMASK ((1 << MINORBITS) - 1)
165 #define SECONDS(x) ((unsigned long long)(x) / 1000000000)
166 #define NANO_SECONDS(x) ((unsigned long long)(x) % 1000000000)
167 #define DOUBLE_TO_NANO_ULL(d) ((unsigned long long)((d) * 1000000000))
168 #define CHECK_MAGIC(t) (((t)->magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
170 void init_io_hash_table(void)
173 struct list_head *head;
175 for (i = 0; i < IO_HASH_TABLE_SIZE; i++) {
176 head = io_hash_table + i;
177 INIT_LIST_HEAD(head);
181 /* taken from the kernel hash.h */
182 static inline u64 hash_sector(u64 val)
186 /* Sigh, gcc can't optimise this alone like it does for 32 bits. */
201 /* High bits are more random, so use them. */
202 return hash >> (64 - IO_HASH_TABLE_BITS);
205 static int hash_table_insert(struct pending_io *ins_pio)
207 u64 sector = ins_pio->sector;
208 int slot = hash_sector(sector);
209 struct list_head *head;
210 struct pending_io *pio;
212 head = io_hash_table + slot;
213 list_for_each_entry(pio, head, hash_list) {
214 if (pio->sector == sector)
217 list_add_tail(&ins_pio->hash_list, head);
221 static struct pending_io *hash_table_search(u64 sector)
223 int slot = hash_sector(sector);
224 struct list_head *head;
225 struct pending_io *pio;
227 head = io_hash_table + slot;
228 list_for_each_entry(pio, head, hash_list) {
229 if (pio->sector == sector)
235 static int hash_dispatched_io(struct blk_io_trace *io)
237 struct pending_io *pio;
240 pio = calloc(1, sizeof(*pio));
241 pio->sector = io->sector;
242 pio->dispatch_time = io->time;
244 ret = hash_table_insert(pio);
245 if (ret == -EEXIST) {
246 /* crud, the IO isn't here */
252 static struct pending_io *hash_completed_io(struct blk_io_trace *io)
254 struct pending_io *pio;
256 pio = hash_table_search(io->sector);
263 static void handle_notify(struct trace *trace)
265 struct blk_io_trace *io = trace->io;
266 void *payload = (char *)io + sizeof(*io);
270 if (io->action != BLK_TN_TIMESTAMP)
273 if (io->pdu_len != sizeof(two32))
276 memcpy(two32, payload, sizeof(two32));
277 trace->start_timestamp = io->time;
278 trace->abs_start_time.tv_sec = two32[0];
279 trace->abs_start_time.tv_nsec = two32[1];
280 if (trace->abs_start_time.tv_nsec < 0) {
281 trace->abs_start_time.tv_sec--;
282 trace->abs_start_time.tv_nsec += 1000000000;
286 int next_record(struct trace *trace)
288 int skip = trace->io->pdu_len;
291 trace->cur += sizeof(*trace->io) + skip;
292 offset = trace->cur - trace->start;
293 if (offset >= trace->len)
296 trace->io = (struct blk_io_trace *)trace->cur;
300 void first_record(struct trace *trace)
302 trace->cur = trace->start;
303 trace->io = (struct blk_io_trace *)trace->cur;
306 int is_io_event(struct blk_io_trace *test)
309 if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
311 if (test->action == BLK_TN_MESSAGE) {
312 int len = test->pdu_len;
315 message = (char *)(test + 1);
316 if (strncmp(message, "fio ", 4) == 0) {
323 u64 find_last_time(struct trace *trace)
325 char *p = trace->start + trace->len;
326 struct blk_io_trace *test;
330 if (trace->len < sizeof(*trace->io))
332 p -= sizeof(*trace->io);
333 while (p >= trace->start) {
334 test = (struct blk_io_trace *)p;
335 if (CHECK_MAGIC(test) && is_io_event(test)) {
336 u64 offset = p - trace->start;
337 if (offset + sizeof(*test) + test->pdu_len == trace->len) {
343 if (search_len > 8192) {
348 /* searching backwards didn't work out, we'll have to scan the file */
351 if (is_io_event(trace->io))
352 found = trace->io->time;
353 if (next_record(trace))
360 int parse_fio_bank_message(struct trace *trace, u64 *bank_ret, u64 *offset_ret,
366 struct blk_io_trace *test = trace->io;
367 int len = test->pdu_len;
372 if (!(test->action & BLK_TC_ACT(BLK_TC_NOTIFY)))
374 if (test->action != BLK_TN_MESSAGE)
377 /* the message is fio rw bank offset num_banks */
380 message = (char *)(test + 1);
381 if (strncmp(message, "fio r ", 6) != 0)
384 message = strndup(message, len);
385 s = strchr(message, ' ');
393 bank = strtoll(s, &next, 10);
398 offset = strtoll(s, &next, 10);
403 num_banks = strtoll(s, &next, 10);
408 *offset_ret = offset;
409 *num_banks_ret = num_banks;
417 void find_extreme_offsets(struct trace *trace, u64 *min_ret, u64 *max_ret, u64 *max_bank_ret,
421 u64 max = 0, min = ~(u64)0;
423 u64 max_bank_offset = 0;
427 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
428 found = trace->io->sector << 9;
432 found += trace->io->bytes;
438 if (!parse_fio_bank_message(trace, &bank,
439 &offset, &num_banks)) {
442 if (offset > max_bank_offset)
443 max_bank_offset = offset;
446 if (next_record(trace))
452 *max_bank_ret = max_bank;
453 *max_offset_ret = max_bank_offset;
456 int filter_outliers(struct trace *trace, u64 min_offset, u64 max_offset,
457 u64 *yzoom_min, u64 *yzoom_max)
460 u64 max_per_bucket[11];
461 u64 min_per_bucket[11];
462 u64 bytes_per_bucket = (max_offset - min_offset + 1) / 10;
466 memset(hits, 0, sizeof(int) * 11);
467 memset(max_per_bucket, 0, sizeof(u64) * 11);
468 memset(min_per_bucket, 0xff, sizeof(u64) * 11);
471 if (!(trace->io->action & BLK_TC_ACT(BLK_TC_NOTIFY)) &&
472 (trace->io->action & BLK_TA_MASK) == __BLK_TA_QUEUE) {
473 u64 off = (trace->io->sector << 9) - min_offset;
475 slot = (int)(off / bytes_per_bucket);
477 if (off < min_per_bucket[slot])
478 min_per_bucket[slot] = off;
480 off += trace->io->bytes;
481 slot = (int)(off / bytes_per_bucket);
483 if (off > max_per_bucket[slot])
484 max_per_bucket[slot] = off;
486 if (next_record(trace))
490 for (slot = 0; slot < 11; slot++) {
491 if (hits[slot] > fat_count) {
492 fat_count = hits[slot];
496 *yzoom_max = max_offset;
497 for (slot = 10; slot >= 0; slot--) {
498 double d = hits[slot];
500 if (d >= (double)fat_count * .05) {
501 *yzoom_max = max_per_bucket[slot] + min_offset;
506 *yzoom_min = min_offset;
507 for (slot = 0; slot < 10; slot++) {
508 double d = hits[slot];
510 if (d >= (double)fat_count * .05) {
511 *yzoom_min = min_per_bucket[slot] + min_offset;
518 static char *find_trace_file(char *filename)
526 ret = stat(filename, &st);
528 return strdup(filename);
530 snprintf(line, 1024, "%s.%s", filename, "dump");
531 ret = stat(line, &st);
535 try = strdup(filename);
536 dot = strrchr(try, '.');
537 if (!dot || strcmp(".dump", dot) != 0) {
540 snprintf(line, 1024, "%s%s", try, ".blktrace.0");
541 ret = stat(line, &st);
543 blktrace_to_dump(try);
544 snprintf(line, 1024, "%s.%s", try, "dump");
545 ret = stat(line, &st);
555 struct trace *open_trace(char *filename)
562 char *found_filename;
564 trace = calloc(1, sizeof(*trace));
566 fprintf(stderr, "unable to allocate memory for trace\n");
570 found_filename = find_trace_file(filename);
571 if (!found_filename) {
572 fprintf(stderr, "Unable to find trace file %s\n", filename);
575 filename = found_filename;
577 fd = open(filename, O_RDONLY);
579 fprintf(stderr, "Unable to open trace file %s err %s\n", filename, strerror(errno));
582 ret = fstat(fd, &st);
584 fprintf(stderr, "stat failed on %s err %s\n", filename, strerror(errno));
587 p = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
588 if (p == MAP_FAILED) {
589 fprintf(stderr, "Unable to mmap trace file %s, err %s\n", filename, strerror(errno));
593 trace->len = st.st_size;
596 trace->io = (struct blk_io_trace *)p;
605 static inline int tput_event(struct trace *trace)
607 if (trace->found_completion)
608 return __BLK_TA_COMPLETE;
609 if (trace->found_issue)
610 return __BLK_TA_ISSUE;
611 if (trace->found_queue)
612 return __BLK_TA_QUEUE;
614 return __BLK_TA_COMPLETE;
617 int action_char_to_num(char action)
621 return __BLK_TA_QUEUE;
623 return __BLK_TA_ISSUE;
625 return __BLK_TA_COMPLETE;
630 static inline int io_event(struct trace *trace)
633 return plot_io_action;
634 if (trace->found_queue)
635 return __BLK_TA_QUEUE;
636 if (trace->found_issue)
637 return __BLK_TA_ISSUE;
638 if (trace->found_completion)
639 return __BLK_TA_COMPLETE;
641 return __BLK_TA_COMPLETE;
644 void add_tput(struct trace *trace, struct graph_line_data *gld)
646 struct blk_io_trace *io = trace->io;
647 int action = io->action & BLK_TA_MASK;
650 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
653 if (action != tput_event(trace))
656 seconds = SECONDS(io->time);
657 if (seconds > gld->max_seconds)
660 gld->data[seconds].sum += io->bytes;
661 gld->data[seconds].count = 1;
662 if (gld->data[seconds].sum > gld->max)
663 gld->max = gld->data[seconds].sum;
666 void add_io(struct trace *trace, struct graph_dot_data *gdd_writes,
667 struct graph_dot_data *gdd_reads)
669 struct blk_io_trace *io = trace->io;
670 int action = io->action & BLK_TA_MASK;
673 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
676 if (action != io_event(trace))
679 offset = io->sector << 9;
681 if (BLK_DATADIR(io->action) & BLK_TC_READ)
682 set_gdd_bit(gdd_reads, offset, io->bytes, io->time);
683 else if (BLK_DATADIR(io->action) & BLK_TC_WRITE)
684 set_gdd_bit(gdd_writes, offset, io->bytes, io->time);
687 void add_pending_io(struct trace *trace, struct graph_line_data *gld)
691 struct blk_io_trace *io = trace->io;
692 int action = io->action & BLK_TA_MASK;
695 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
698 if (action != __BLK_TA_ISSUE)
701 seconds = SECONDS(io->time);
702 if (seconds > gld->max_seconds)
705 ret = hash_dispatched_io(trace->io);
711 gld->data[seconds].sum += ios_in_flight;
712 gld->data[seconds].count++;
714 avg = (double)gld->data[seconds].sum / gld->data[seconds].count;
715 if (gld->max < (u64)avg) {
720 void add_completed_io(struct trace *trace,
721 struct graph_line_data *latency_gld)
723 struct blk_io_trace *io = trace->io;
725 int action = io->action & BLK_TA_MASK;
726 struct pending_io *pio;
730 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
733 if (action != __BLK_TA_COMPLETE)
736 seconds = SECONDS(io->time);
738 pio = hash_completed_io(trace->io);
742 if (ios_in_flight > 0)
744 if (io->time >= pio->dispatch_time) {
745 latency = io->time - pio->dispatch_time;
746 latency_gld->data[seconds].sum += latency;
747 latency_gld->data[seconds].count++;
750 list_del(&pio->hash_list);
753 avg = (double)latency_gld->data[seconds].sum /
754 latency_gld->data[seconds].count;
755 if (latency_gld->max < (u64)avg) {
756 latency_gld->max = avg;
760 void add_iop(struct trace *trace, struct graph_line_data *gld)
762 struct blk_io_trace *io = trace->io;
763 int action = io->action & BLK_TA_MASK;
766 if (io->action & BLK_TC_ACT(BLK_TC_NOTIFY))
769 /* iops and tput use the same events */
770 if (action != tput_event(trace))
773 seconds = SECONDS(io->time);
774 if (seconds > gld->max_seconds)
777 gld->data[seconds].sum += 1;
778 gld->data[seconds].count = 1;
779 if (gld->data[seconds].sum > gld->max)
780 gld->max = gld->data[seconds].sum;
783 void check_record(struct trace *trace)
785 struct blk_io_trace *io = trace->io;
786 int action = io->action & BLK_TA_MASK;
788 if (!(io->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
790 case __BLK_TA_COMPLETE:
791 trace->found_completion = 1;
794 trace->found_issue = 1;
797 trace->found_queue = 1;
801 handle_notify(trace);