Commit | Line | Data |
---|---|---|
d956a2cd JA |
1 | /* |
2 | * block queue tracing parse application | |
3 | * | |
4 | * Copyright (C) 2005 Jens Axboe <axboe@suse.de> | |
5 | * | |
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. | |
10 | * | |
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. | |
15 | * | |
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 | |
19 | * | |
20 | */ | |
d0ca268b JA |
21 | #include <sys/types.h> |
22 | #include <sys/stat.h> | |
23 | #include <unistd.h> | |
24 | #include <stdio.h> | |
25 | #include <fcntl.h> | |
26 | #include <stdlib.h> | |
8fc0abbc | 27 | #include <string.h> |
d5396421 | 28 | #include <getopt.h> |
412819ce JA |
29 | #include <errno.h> |
30 | #include <signal.h> | |
d69db225 | 31 | #include <locale.h> |
6e0073ed | 32 | #include <libgen.h> |
d0ca268b | 33 | |
8fc0abbc JA |
34 | #include "blktrace.h" |
35 | #include "rbtree.h" | |
bf0720af | 36 | #include "jhash.h" |
d0ca268b | 37 | |
13d928f0 | 38 | static char blkparse_version[] = "0.99"; |
52724a0e | 39 | |
e7c9f3ff | 40 | struct per_dev_info { |
f7bd1a9b | 41 | dev_t dev; |
e7c9f3ff NS |
42 | char *name; |
43 | ||
44 | int backwards; | |
45 | unsigned long long events; | |
20ed6177 | 46 | unsigned long long first_reported_time; |
e7c9f3ff | 47 | unsigned long long last_reported_time; |
287fa3d6 | 48 | unsigned long long last_read_time; |
e7c9f3ff | 49 | struct io_stats io_stats; |
cb2a1a62 | 50 | unsigned long last_sequence; |
dce0f678 AB |
51 | unsigned long skips, nskips; |
52 | unsigned long long seq_skips, seq_nskips; | |
e7c9f3ff | 53 | |
2a1b3424 JA |
54 | struct rb_root rb_last; |
55 | unsigned long rb_last_entries; | |
56 | ||
f7bd1a9b JA |
57 | struct rb_root rb_track; |
58 | ||
73877e12 | 59 | int nfiles; |
e7c9f3ff | 60 | int ncpus; |
824c2b39 JA |
61 | |
62 | unsigned long *cpu_map; | |
63 | unsigned int cpu_map_max; | |
64 | ||
e7c9f3ff NS |
65 | struct per_cpu_info *cpus; |
66 | }; | |
67 | ||
152f6476 JA |
68 | struct per_process_info { |
69 | char name[16]; | |
70 | __u32 pid; | |
71 | struct io_stats io_stats; | |
72 | struct per_process_info *hash_next, *list_next; | |
715d8021 | 73 | int more_than_one; |
50adc0ba JA |
74 | |
75 | /* | |
76 | * individual io stats | |
77 | */ | |
b9d40d6f JA |
78 | unsigned long long longest_allocation_wait[2]; |
79 | unsigned long long longest_dispatch_wait[2]; | |
80 | unsigned long long longest_completion_wait[2]; | |
d0ca268b JA |
81 | }; |
82 | ||
152f6476 | 83 | #define PPI_HASH_SHIFT (8) |
bf0720af JA |
84 | #define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT) |
85 | #define PPI_HASH_MASK (PPI_HASH_SIZE - 1) | |
86 | static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE]; | |
152f6476 | 87 | static struct per_process_info *ppi_list; |
886ecf0e | 88 | static int ppi_list_entries; |
152f6476 | 89 | |
d915dee6 | 90 | #define S_OPTS "a:A:i:o:b:stqw:f:F:vVhD:" |
d5396421 | 91 | static struct option l_opts[] = { |
98f8386b AB |
92 | { |
93 | .name = "act-mask", | |
94 | .has_arg = required_argument, | |
95 | .flag = NULL, | |
96 | .val = 'a' | |
97 | }, | |
98 | { | |
99 | .name = "set-mask", | |
100 | .has_arg = required_argument, | |
101 | .flag = NULL, | |
102 | .val = 'A' | |
103 | }, | |
d5396421 JA |
104 | { |
105 | .name = "input", | |
428683db | 106 | .has_arg = required_argument, |
d5396421 JA |
107 | .flag = NULL, |
108 | .val = 'i' | |
109 | }, | |
110 | { | |
111 | .name = "output", | |
428683db | 112 | .has_arg = required_argument, |
d5396421 JA |
113 | .flag = NULL, |
114 | .val = 'o' | |
115 | }, | |
79f19470 JA |
116 | { |
117 | .name = "batch", | |
428683db | 118 | .has_arg = required_argument, |
79f19470 JA |
119 | .flag = NULL, |
120 | .val = 'b' | |
121 | }, | |
152f6476 | 122 | { |
3f65c585 | 123 | .name = "per-program-stats", |
428683db | 124 | .has_arg = no_argument, |
152f6476 JA |
125 | .flag = NULL, |
126 | .val = 's' | |
127 | }, | |
7997c5b0 | 128 | { |
3f65c585 | 129 | .name = "track-ios", |
428683db | 130 | .has_arg = no_argument, |
7997c5b0 JA |
131 | .flag = NULL, |
132 | .val = 't' | |
133 | }, | |
1e1c60f1 NS |
134 | { |
135 | .name = "quiet", | |
428683db | 136 | .has_arg = no_argument, |
1e1c60f1 NS |
137 | .flag = NULL, |
138 | .val = 'q' | |
139 | }, | |
46e6968b NS |
140 | { |
141 | .name = "stopwatch", | |
428683db | 142 | .has_arg = required_argument, |
46e6968b NS |
143 | .flag = NULL, |
144 | .val = 'w' | |
145 | }, | |
ab197ca7 AB |
146 | { |
147 | .name = "format", | |
428683db | 148 | .has_arg = required_argument, |
ab197ca7 AB |
149 | .flag = NULL, |
150 | .val = 'f' | |
151 | }, | |
152 | { | |
153 | .name = "format-spec", | |
428683db | 154 | .has_arg = required_argument, |
ab197ca7 AB |
155 | .flag = NULL, |
156 | .val = 'F' | |
157 | }, | |
bf0720af | 158 | { |
3f65c585 | 159 | .name = "hash-by-name", |
bf0720af JA |
160 | .has_arg = no_argument, |
161 | .flag = NULL, | |
d915dee6 | 162 | .val = 'h' |
bf0720af | 163 | }, |
7d1c0411 | 164 | { |
57ea8602 | 165 | .name = "verbose", |
7d1c0411 JA |
166 | .has_arg = no_argument, |
167 | .flag = NULL, | |
57ea8602 | 168 | .val = 'v' |
7d1c0411 | 169 | }, |
52724a0e JA |
170 | { |
171 | .name = "version", | |
172 | .has_arg = no_argument, | |
173 | .flag = NULL, | |
57ea8602 | 174 | .val = 'V' |
52724a0e | 175 | }, |
d1d7f15f | 176 | { |
3f65c585 | 177 | .name = "input-directory", |
d1d7f15f JA |
178 | .has_arg = required_argument, |
179 | .flag = NULL, | |
180 | .val = 'D' | |
181 | }, | |
71ef8b7c JA |
182 | { |
183 | .name = NULL, | |
184 | } | |
d5396421 JA |
185 | }; |
186 | ||
7997c5b0 JA |
187 | /* |
188 | * for sorting the displayed output | |
189 | */ | |
8fc0abbc JA |
190 | struct trace { |
191 | struct blk_io_trace *bit; | |
192 | struct rb_node rb_node; | |
cb2a1a62 | 193 | struct trace *next; |
8fc0abbc JA |
194 | }; |
195 | ||
cb2a1a62 | 196 | static struct rb_root rb_sort_root; |
a649216c JA |
197 | static unsigned long rb_sort_entries; |
198 | ||
cb2a1a62 JA |
199 | static struct trace *trace_list; |
200 | ||
d36421e4 JA |
201 | /* |
202 | * allocation cache | |
203 | */ | |
204 | static struct blk_io_trace *bit_alloc_list; | |
205 | static struct trace *t_alloc_list; | |
206 | ||
7997c5b0 JA |
207 | /* |
208 | * for tracking individual ios | |
209 | */ | |
210 | struct io_track { | |
211 | struct rb_node rb_node; | |
212 | ||
213 | __u64 sector; | |
214 | __u32 pid; | |
bf0720af | 215 | char comm[16]; |
95c15013 | 216 | unsigned long long allocation_time; |
7997c5b0 JA |
217 | unsigned long long queue_time; |
218 | unsigned long long dispatch_time; | |
219 | unsigned long long completion_time; | |
220 | }; | |
221 | ||
e7c9f3ff NS |
222 | static int ndevices; |
223 | static struct per_dev_info *devices; | |
224 | static char *get_dev_name(struct per_dev_info *, char *, int); | |
d0ca268b | 225 | |
71d5d4c9 | 226 | FILE *ofp = NULL; |
e7c9f3ff | 227 | static char *output_name; |
d1d7f15f | 228 | static char *input_dir; |
e7c9f3ff NS |
229 | |
230 | static unsigned long long genesis_time; | |
287fa3d6 | 231 | static unsigned long long last_allowed_time; |
1c7c54aa | 232 | static unsigned int smallest_seq_read; |
46e6968b | 233 | static unsigned long long stopwatch_start; /* start from zero by default */ |
bc171579 | 234 | static unsigned long long stopwatch_end = -1ULL; /* "infinity" */ |
152f6476 JA |
235 | |
236 | static int per_process_stats; | |
cbc927b6 | 237 | static int per_device_and_cpu_stats = 1; |
7997c5b0 | 238 | static int track_ios; |
bf0720af | 239 | static int ppi_hash_by_pid = 1; |
57ea8602 | 240 | static int verbose; |
98f8386b | 241 | static unsigned int act_mask = -1U; |
cbc927b6 | 242 | static int stats_printed; |
d0ca268b | 243 | |
1d24fc14 JA |
244 | static unsigned int t_alloc_cache; |
245 | static unsigned int bit_alloc_cache; | |
246 | ||
7d747d22 | 247 | #define RB_BATCH_DEFAULT (512) |
e820abd7 | 248 | static unsigned int rb_batch = RB_BATCH_DEFAULT; |
79f19470 | 249 | |
e7c9f3ff NS |
250 | static int pipeline; |
251 | ||
412819ce JA |
252 | #define is_done() (*(volatile int *)(&done)) |
253 | static volatile int done; | |
254 | ||
bf0720af JA |
255 | #define JHASH_RANDOM (0x3af5f2ee) |
256 | ||
824c2b39 JA |
257 | #define CPUS_PER_LONG (8 * sizeof(unsigned long)) |
258 | #define CPU_IDX(cpu) ((cpu) / CPUS_PER_LONG) | |
259 | #define CPU_BIT(cpu) ((cpu) & (CPUS_PER_LONG - 1)) | |
260 | ||
261 | static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu) | |
262 | { | |
263 | if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) { | |
264 | int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1); | |
265 | unsigned long *map = malloc(new_max / sizeof(long)); | |
266 | ||
267 | memset(map, 0, new_max / sizeof(long)); | |
268 | ||
269 | if (pdi->cpu_map) { | |
270 | memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long)); | |
271 | free(pdi->cpu_map); | |
272 | } | |
273 | ||
274 | pdi->cpu_map = map; | |
275 | pdi->cpu_map_max = new_max; | |
276 | } | |
277 | ||
278 | pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu)); | |
279 | } | |
280 | ||
281 | static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu) | |
282 | { | |
283 | pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu)); | |
284 | } | |
285 | ||
286 | static inline int cpu_is_online(struct per_dev_info *pdi, int cpu) | |
287 | { | |
288 | return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0; | |
289 | } | |
290 | ||
bf0720af JA |
291 | static inline int ppi_hash_pid(__u32 pid) |
292 | { | |
293 | return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK; | |
294 | } | |
295 | ||
296 | static inline int ppi_hash_name(const char *name) | |
152f6476 | 297 | { |
bf0720af JA |
298 | return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK; |
299 | } | |
300 | ||
301 | static inline int ppi_hash(struct per_process_info *ppi) | |
302 | { | |
303 | if (ppi_hash_by_pid) | |
304 | return ppi_hash_pid(ppi->pid); | |
305 | ||
bf0720af | 306 | return ppi_hash_name(ppi->name); |
152f6476 JA |
307 | } |
308 | ||
309 | static inline void add_process_to_hash(struct per_process_info *ppi) | |
310 | { | |
bf0720af | 311 | const int hash_idx = ppi_hash(ppi); |
152f6476 | 312 | |
bf0720af JA |
313 | ppi->hash_next = ppi_hash_table[hash_idx]; |
314 | ppi_hash_table[hash_idx] = ppi; | |
152f6476 JA |
315 | } |
316 | ||
317 | static inline void add_process_to_list(struct per_process_info *ppi) | |
318 | { | |
319 | ppi->list_next = ppi_list; | |
320 | ppi_list = ppi; | |
886ecf0e | 321 | ppi_list_entries++; |
152f6476 JA |
322 | } |
323 | ||
bf0720af JA |
324 | static struct per_process_info *find_process_by_name(char *name) |
325 | { | |
326 | const int hash_idx = ppi_hash_name(name); | |
327 | struct per_process_info *ppi; | |
328 | ||
329 | ppi = ppi_hash_table[hash_idx]; | |
330 | while (ppi) { | |
331 | if (!strcmp(ppi->name, name)) | |
332 | return ppi; | |
333 | ||
334 | ppi = ppi->hash_next; | |
335 | } | |
336 | ||
337 | return NULL; | |
338 | } | |
339 | ||
152f6476 JA |
340 | static struct per_process_info *find_process_by_pid(__u32 pid) |
341 | { | |
bf0720af | 342 | const int hash_idx = ppi_hash_pid(pid); |
152f6476 JA |
343 | struct per_process_info *ppi; |
344 | ||
bf0720af | 345 | ppi = ppi_hash_table[hash_idx]; |
152f6476 JA |
346 | while (ppi) { |
347 | if (ppi->pid == pid) | |
348 | return ppi; | |
349 | ||
350 | ppi = ppi->hash_next; | |
351 | } | |
352 | ||
353 | return NULL; | |
354 | } | |
355 | ||
bf0720af JA |
356 | static struct per_process_info *find_process(__u32 pid, char *name) |
357 | { | |
715d8021 JA |
358 | struct per_process_info *ppi; |
359 | ||
bf0720af JA |
360 | if (ppi_hash_by_pid) |
361 | return find_process_by_pid(pid); | |
362 | ||
715d8021 JA |
363 | ppi = find_process_by_name(name); |
364 | if (ppi && ppi->pid != pid) | |
365 | ppi->more_than_one = 1; | |
366 | ||
367 | return ppi; | |
bf0720af JA |
368 | } |
369 | ||
23211fdf JA |
370 | static inline int trace_rb_insert(struct trace *t, struct rb_root *root, |
371 | int check_time) | |
7997c5b0 | 372 | { |
2a1b3424 | 373 | struct rb_node **p = &root->rb_node; |
7997c5b0 JA |
374 | struct rb_node *parent = NULL; |
375 | struct trace *__t; | |
376 | ||
377 | while (*p) { | |
378 | parent = *p; | |
2a1b3424 | 379 | |
7997c5b0 JA |
380 | __t = rb_entry(parent, struct trace, rb_node); |
381 | ||
23211fdf JA |
382 | if (check_time) { |
383 | if (t->bit->time < __t->bit->time) { | |
384 | p = &(*p)->rb_left; | |
385 | continue; | |
386 | } else if (t->bit->time > __t->bit->time) { | |
387 | p = &(*p)->rb_right; | |
388 | continue; | |
389 | } | |
390 | } | |
391 | if (t->bit->device < __t->bit->device) | |
e7c9f3ff NS |
392 | p = &(*p)->rb_left; |
393 | else if (t->bit->device > __t->bit->device) | |
394 | p = &(*p)->rb_right; | |
dcf0f7ed JA |
395 | else if (t->bit->sequence < __t->bit->sequence) |
396 | p = &(*p)->rb_left; | |
0b07f23e | 397 | else /* >= sequence */ |
dcf0f7ed | 398 | p = &(*p)->rb_right; |
7997c5b0 JA |
399 | } |
400 | ||
401 | rb_link_node(&t->rb_node, parent, p); | |
2a1b3424 | 402 | rb_insert_color(&t->rb_node, root); |
7997c5b0 JA |
403 | return 0; |
404 | } | |
405 | ||
2a1b3424 | 406 | static inline int trace_rb_insert_sort(struct trace *t) |
e3556946 | 407 | { |
23211fdf | 408 | if (!trace_rb_insert(t, &rb_sort_root, 1)) { |
2a1b3424 JA |
409 | rb_sort_entries++; |
410 | return 0; | |
411 | } | |
412 | ||
413 | return 1; | |
414 | } | |
415 | ||
416 | static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t) | |
417 | { | |
23211fdf | 418 | if (!trace_rb_insert(t, &pdi->rb_last, 1)) { |
2a1b3424 JA |
419 | pdi->rb_last_entries++; |
420 | return 0; | |
421 | } | |
422 | ||
423 | return 1; | |
424 | } | |
425 | ||
426 | static struct trace *trace_rb_find(dev_t device, unsigned long sequence, | |
427 | struct rb_root *root, int order) | |
428 | { | |
429 | struct rb_node *n = root->rb_node; | |
430 | struct rb_node *prev = NULL; | |
e3556946 JA |
431 | struct trace *__t; |
432 | ||
2a1b3424 JA |
433 | while (n) { |
434 | __t = rb_entry(n, struct trace, rb_node); | |
435 | prev = n; | |
e3556946 | 436 | |
0583b6a2 | 437 | if (device < __t->bit->device) |
2a1b3424 | 438 | n = n->rb_left; |
0583b6a2 | 439 | else if (device > __t->bit->device) |
2a1b3424 | 440 | n = n->rb_right; |
0583b6a2 | 441 | else if (sequence < __t->bit->sequence) |
2a1b3424 | 442 | n = n->rb_left; |
e3556946 | 443 | else if (sequence > __t->bit->sequence) |
2a1b3424 | 444 | n = n->rb_right; |
e3556946 JA |
445 | else |
446 | return __t; | |
447 | } | |
448 | ||
2a1b3424 JA |
449 | /* |
450 | * hack - the list may not be sequence ordered because some | |
451 | * events don't have sequence and time matched. so we end up | |
452 | * being a little off in the rb lookup here, because we don't | |
453 | * know the time we are looking for. compensate by browsing | |
454 | * a little ahead from the last entry to find the match | |
455 | */ | |
456 | if (order && prev) { | |
457 | int max = 5; | |
458 | ||
459 | while (((n = rb_next(prev)) != NULL) && max--) { | |
460 | __t = rb_entry(n, struct trace, rb_node); | |
461 | ||
462 | if (__t->bit->device == device && | |
463 | __t->bit->sequence == sequence) | |
464 | return __t; | |
465 | ||
466 | prev = n; | |
467 | } | |
468 | } | |
469 | ||
e3556946 JA |
470 | return NULL; |
471 | } | |
472 | ||
2a1b3424 JA |
473 | static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq) |
474 | { | |
475 | return trace_rb_find(dev, seq, &rb_sort_root, 1); | |
476 | } | |
477 | ||
478 | static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi, | |
479 | unsigned long seq) | |
480 | { | |
f7bd1a9b | 481 | return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0); |
2a1b3424 JA |
482 | } |
483 | ||
f7bd1a9b | 484 | static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot) |
7997c5b0 | 485 | { |
f7bd1a9b | 486 | struct rb_node **p = &pdi->rb_track.rb_node; |
7997c5b0 JA |
487 | struct rb_node *parent = NULL; |
488 | struct io_track *__iot; | |
489 | ||
490 | while (*p) { | |
491 | parent = *p; | |
7997c5b0 JA |
492 | __iot = rb_entry(parent, struct io_track, rb_node); |
493 | ||
f7bd1a9b | 494 | if (iot->sector < __iot->sector) |
7997c5b0 JA |
495 | p = &(*p)->rb_left; |
496 | else if (iot->sector > __iot->sector) | |
497 | p = &(*p)->rb_right; | |
498 | else { | |
e7c9f3ff | 499 | fprintf(stderr, |
ab197ca7 AB |
500 | "sector alias (%Lu) on device %d,%d!\n", |
501 | (unsigned long long) iot->sector, | |
f7bd1a9b | 502 | MAJOR(pdi->dev), MINOR(pdi->dev)); |
7997c5b0 JA |
503 | return 1; |
504 | } | |
505 | } | |
506 | ||
507 | rb_link_node(&iot->rb_node, parent, p); | |
f7bd1a9b | 508 | rb_insert_color(&iot->rb_node, &pdi->rb_track); |
7997c5b0 JA |
509 | return 0; |
510 | } | |
511 | ||
f7bd1a9b | 512 | static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector) |
7997c5b0 | 513 | { |
f7bd1a9b | 514 | struct rb_node *n = pdi->rb_track.rb_node; |
7997c5b0 JA |
515 | struct io_track *__iot; |
516 | ||
2a1b3424 JA |
517 | while (n) { |
518 | __iot = rb_entry(n, struct io_track, rb_node); | |
7997c5b0 | 519 | |
f7bd1a9b | 520 | if (sector < __iot->sector) |
2a1b3424 | 521 | n = n->rb_left; |
7997c5b0 | 522 | else if (sector > __iot->sector) |
2a1b3424 | 523 | n = n->rb_right; |
7997c5b0 JA |
524 | else |
525 | return __iot; | |
526 | } | |
527 | ||
528 | return NULL; | |
529 | } | |
530 | ||
f7bd1a9b JA |
531 | static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid, |
532 | char *comm, __u64 sector) | |
7997c5b0 | 533 | { |
916b5501 | 534 | struct io_track *iot; |
7997c5b0 | 535 | |
f7bd1a9b | 536 | iot = __find_track(pdi, sector); |
7997c5b0 JA |
537 | if (!iot) { |
538 | iot = malloc(sizeof(*iot)); | |
50adc0ba | 539 | iot->pid = pid; |
bf0720af | 540 | memcpy(iot->comm, comm, sizeof(iot->comm)); |
7997c5b0 | 541 | iot->sector = sector; |
f7bd1a9b | 542 | track_rb_insert(pdi, iot); |
7997c5b0 JA |
543 | } |
544 | ||
545 | return iot; | |
546 | } | |
547 | ||
f7bd1a9b JA |
548 | static void log_track_frontmerge(struct per_dev_info *pdi, |
549 | struct blk_io_trace *t) | |
2e3e8ded JA |
550 | { |
551 | struct io_track *iot; | |
552 | ||
553 | if (!track_ios) | |
554 | return; | |
2e3e8ded | 555 | |
ae957cbc | 556 | iot = __find_track(pdi, t->sector + t_sec(t)); |
cb2a1a62 | 557 | if (!iot) { |
57ea8602 JA |
558 | if (verbose) |
559 | fprintf(stderr, "merge not found for (%d,%d): %llu\n", | |
560 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
561 | (unsigned long long) t->sector + t_sec(t)); | |
cb2a1a62 | 562 | return; |
2e3e8ded | 563 | } |
cb2a1a62 | 564 | |
f7bd1a9b | 565 | rb_erase(&iot->rb_node, &pdi->rb_track); |
ae957cbc | 566 | iot->sector -= t_sec(t); |
f7bd1a9b | 567 | track_rb_insert(pdi, iot); |
2e3e8ded JA |
568 | } |
569 | ||
f7bd1a9b | 570 | static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t) |
2e3e8ded JA |
571 | { |
572 | struct io_track *iot; | |
573 | ||
574 | if (!track_ios) | |
575 | return; | |
576 | ||
f7bd1a9b | 577 | iot = find_track(pdi, t->pid, t->comm, t->sector); |
95c15013 JA |
578 | iot->allocation_time = t->time; |
579 | } | |
580 | ||
95c15013 | 581 | /* |
b6076a9b | 582 | * return time between rq allocation and insertion |
95c15013 | 583 | */ |
f7bd1a9b JA |
584 | static unsigned long long log_track_insert(struct per_dev_info *pdi, |
585 | struct blk_io_trace *t) | |
95c15013 | 586 | { |
50adc0ba | 587 | unsigned long long elapsed; |
95c15013 JA |
588 | struct io_track *iot; |
589 | ||
590 | if (!track_ios) | |
591 | return -1; | |
592 | ||
f7bd1a9b | 593 | iot = find_track(pdi, t->pid, t->comm, t->sector); |
2e3e8ded | 594 | iot->queue_time = t->time; |
acd70d21 JA |
595 | |
596 | if (!iot->allocation_time) | |
597 | return -1; | |
598 | ||
50adc0ba JA |
599 | elapsed = iot->queue_time - iot->allocation_time; |
600 | ||
601 | if (per_process_stats) { | |
bf0720af | 602 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 603 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 604 | |
b9d40d6f JA |
605 | if (ppi && elapsed > ppi->longest_allocation_wait[w]) |
606 | ppi->longest_allocation_wait[w] = elapsed; | |
50adc0ba JA |
607 | } |
608 | ||
609 | return elapsed; | |
2e3e8ded JA |
610 | } |
611 | ||
612 | /* | |
613 | * return time between queue and issue | |
614 | */ | |
f7bd1a9b JA |
615 | static unsigned long long log_track_issue(struct per_dev_info *pdi, |
616 | struct blk_io_trace *t) | |
2e3e8ded | 617 | { |
50adc0ba | 618 | unsigned long long elapsed; |
2e3e8ded JA |
619 | struct io_track *iot; |
620 | ||
621 | if (!track_ios) | |
622 | return -1; | |
623 | if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0) | |
624 | return -1; | |
625 | ||
f7bd1a9b | 626 | iot = __find_track(pdi, t->sector); |
cb2a1a62 | 627 | if (!iot) { |
57ea8602 JA |
628 | if (verbose) |
629 | fprintf(stderr, "issue not found for (%d,%d): %llu\n", | |
630 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
631 | (unsigned long long) t->sector); | |
2e3e8ded | 632 | return -1; |
cb2a1a62 | 633 | } |
2e3e8ded JA |
634 | |
635 | iot->dispatch_time = t->time; | |
50adc0ba JA |
636 | elapsed = iot->dispatch_time - iot->queue_time; |
637 | ||
638 | if (per_process_stats) { | |
bf0720af | 639 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 640 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 641 | |
b9d40d6f JA |
642 | if (ppi && elapsed > ppi->longest_dispatch_wait[w]) |
643 | ppi->longest_dispatch_wait[w] = elapsed; | |
50adc0ba JA |
644 | } |
645 | ||
646 | return elapsed; | |
2e3e8ded JA |
647 | } |
648 | ||
649 | /* | |
650 | * return time between dispatch and complete | |
651 | */ | |
f7bd1a9b JA |
652 | static unsigned long long log_track_complete(struct per_dev_info *pdi, |
653 | struct blk_io_trace *t) | |
2e3e8ded JA |
654 | { |
655 | unsigned long long elapsed; | |
656 | struct io_track *iot; | |
657 | ||
658 | if (!track_ios) | |
659 | return -1; | |
660 | if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0) | |
661 | return -1; | |
662 | ||
f7bd1a9b | 663 | iot = __find_track(pdi, t->sector); |
cb2a1a62 | 664 | if (!iot) { |
57ea8602 JA |
665 | if (verbose) |
666 | fprintf(stderr,"complete not found for (%d,%d): %llu\n", | |
667 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
668 | (unsigned long long) t->sector); | |
2e3e8ded | 669 | return -1; |
cb2a1a62 | 670 | } |
2e3e8ded JA |
671 | |
672 | iot->completion_time = t->time; | |
673 | elapsed = iot->completion_time - iot->dispatch_time; | |
674 | ||
50adc0ba | 675 | if (per_process_stats) { |
bf0720af | 676 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 677 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 678 | |
b9d40d6f JA |
679 | if (ppi && elapsed > ppi->longest_completion_wait[w]) |
680 | ppi->longest_completion_wait[w] = elapsed; | |
50adc0ba JA |
681 | } |
682 | ||
2e3e8ded JA |
683 | /* |
684 | * kill the trace, we don't need it after completion | |
685 | */ | |
f7bd1a9b | 686 | rb_erase(&iot->rb_node, &pdi->rb_track); |
2e3e8ded JA |
687 | free(iot); |
688 | ||
689 | return elapsed; | |
690 | } | |
691 | ||
692 | ||
152f6476 JA |
693 | static struct io_stats *find_process_io_stats(__u32 pid, char *name) |
694 | { | |
bf0720af | 695 | struct per_process_info *ppi = find_process(pid, name); |
152f6476 JA |
696 | |
697 | if (!ppi) { | |
698 | ppi = malloc(sizeof(*ppi)); | |
699 | memset(ppi, 0, sizeof(*ppi)); | |
bf0720af | 700 | memcpy(ppi->name, name, 16); |
152f6476 JA |
701 | ppi->pid = pid; |
702 | add_process_to_hash(ppi); | |
703 | add_process_to_list(ppi); | |
704 | } | |
705 | ||
706 | return &ppi->io_stats; | |
707 | } | |
708 | ||
e7c9f3ff | 709 | static void resize_cpu_info(struct per_dev_info *pdi, int cpu) |
a718bd37 | 710 | { |
e7c9f3ff NS |
711 | struct per_cpu_info *cpus = pdi->cpus; |
712 | int ncpus = pdi->ncpus; | |
713 | int new_count = cpu + 1; | |
714 | int new_space, size; | |
a718bd37 NS |
715 | char *new_start; |
716 | ||
e7c9f3ff NS |
717 | size = new_count * sizeof(struct per_cpu_info); |
718 | cpus = realloc(cpus, size); | |
719 | if (!cpus) { | |
720 | char name[20]; | |
721 | fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n", | |
722 | get_dev_name(pdi, name, sizeof(name)), size); | |
a718bd37 NS |
723 | exit(1); |
724 | } | |
725 | ||
e7c9f3ff NS |
726 | new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info)); |
727 | new_space = (new_count - ncpus) * sizeof(struct per_cpu_info); | |
a718bd37 | 728 | memset(new_start, 0, new_space); |
e7c9f3ff NS |
729 | |
730 | pdi->ncpus = new_count; | |
731 | pdi->cpus = cpus; | |
732 | } | |
cb2a1a62 | 733 | |
e7c9f3ff NS |
734 | static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu) |
735 | { | |
cb2a1a62 JA |
736 | struct per_cpu_info *pci; |
737 | ||
e7c9f3ff NS |
738 | if (cpu >= pdi->ncpus) |
739 | resize_cpu_info(pdi, cpu); | |
cb2a1a62 JA |
740 | |
741 | pci = &pdi->cpus[cpu]; | |
742 | pci->cpu = cpu; | |
743 | return pci; | |
a718bd37 NS |
744 | } |
745 | ||
e7c9f3ff NS |
746 | |
747 | static int resize_devices(char *name) | |
a718bd37 | 748 | { |
e7c9f3ff | 749 | int size = (ndevices + 1) * sizeof(struct per_dev_info); |
c499bf38 | 750 | |
e7c9f3ff NS |
751 | devices = realloc(devices, size); |
752 | if (!devices) { | |
753 | fprintf(stderr, "Out of memory, device %s (%d)\n", name, size); | |
754 | return 1; | |
755 | } | |
756 | memset(&devices[ndevices], 0, sizeof(struct per_dev_info)); | |
757 | devices[ndevices].name = name; | |
758 | ndevices++; | |
759 | return 0; | |
760 | } | |
a718bd37 | 761 | |
f7bd1a9b | 762 | static struct per_dev_info *get_dev_info(dev_t dev) |
e7c9f3ff | 763 | { |
cb2a1a62 | 764 | struct per_dev_info *pdi; |
e7c9f3ff | 765 | int i; |
c499bf38 | 766 | |
7d747d22 | 767 | for (i = 0; i < ndevices; i++) { |
f7bd1a9b JA |
768 | if (!devices[i].dev) |
769 | devices[i].dev = dev; | |
770 | if (devices[i].dev == dev) | |
e7c9f3ff | 771 | return &devices[i]; |
7d747d22 | 772 | } |
cb2a1a62 | 773 | |
2a1b3424 | 774 | if (resize_devices(NULL)) |
e7c9f3ff | 775 | return NULL; |
cb2a1a62 JA |
776 | |
777 | pdi = &devices[ndevices - 1]; | |
f7bd1a9b | 778 | pdi->dev = dev; |
20ed6177 | 779 | pdi->first_reported_time = 0; |
d6222db8 | 780 | pdi->last_sequence = -1; |
287fa3d6 | 781 | pdi->last_read_time = 0; |
2a1b3424 JA |
782 | memset(&pdi->rb_last, 0, sizeof(pdi->rb_last)); |
783 | pdi->rb_last_entries = 0; | |
cb2a1a62 | 784 | return pdi; |
a718bd37 NS |
785 | } |
786 | ||
e7c9f3ff NS |
787 | static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size) |
788 | { | |
789 | if (pdi->name) | |
790 | snprintf(buffer, size, "%s", pdi->name); | |
791 | else | |
f7bd1a9b | 792 | snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev)); |
e7c9f3ff NS |
793 | return buffer; |
794 | } | |
795 | ||
e7c9f3ff | 796 | static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit) |
cfab07eb AB |
797 | { |
798 | unsigned long long this = bit->time; | |
e7c9f3ff | 799 | unsigned long long last = pdi->last_reported_time; |
cfab07eb | 800 | |
e7c9f3ff NS |
801 | pdi->backwards = (this < last) ? 'B' : ' '; |
802 | pdi->last_reported_time = this; | |
cfab07eb AB |
803 | } |
804 | ||
152f6476 JA |
805 | static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t, |
806 | int rw) | |
d0ca268b JA |
807 | { |
808 | if (rw) { | |
152f6476 | 809 | ios->mwrites++; |
ae957cbc | 810 | ios->qwrite_kb += t_kb(t); |
d0ca268b | 811 | } else { |
152f6476 | 812 | ios->mreads++; |
ae957cbc | 813 | ios->qread_kb += t_kb(t); |
152f6476 JA |
814 | } |
815 | } | |
816 | ||
817 | static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci, | |
818 | int rw) | |
819 | { | |
820 | __account_m(&pci->io_stats, t, rw); | |
821 | ||
822 | if (per_process_stats) { | |
823 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
824 | ||
825 | __account_m(ios, t, rw); | |
d0ca268b JA |
826 | } |
827 | } | |
828 | ||
b6076a9b JA |
829 | static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t, |
830 | int rw) | |
d0ca268b JA |
831 | { |
832 | if (rw) { | |
152f6476 | 833 | ios->qwrites++; |
ae957cbc | 834 | ios->qwrite_kb += t_kb(t); |
d0ca268b | 835 | } else { |
152f6476 | 836 | ios->qreads++; |
ae957cbc | 837 | ios->qread_kb += t_kb(t); |
152f6476 JA |
838 | } |
839 | } | |
840 | ||
b6076a9b JA |
841 | static inline void account_queue(struct blk_io_trace *t, |
842 | struct per_cpu_info *pci, int rw) | |
152f6476 | 843 | { |
b6076a9b | 844 | __account_queue(&pci->io_stats, t, rw); |
152f6476 JA |
845 | |
846 | if (per_process_stats) { | |
847 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
848 | ||
b6076a9b | 849 | __account_queue(ios, t, rw); |
d0ca268b JA |
850 | } |
851 | } | |
852 | ||
e21dc4dd | 853 | static inline void __account_c(struct io_stats *ios, int rw, int bytes) |
d0ca268b JA |
854 | { |
855 | if (rw) { | |
152f6476 JA |
856 | ios->cwrites++; |
857 | ios->cwrite_kb += bytes >> 10; | |
d0ca268b | 858 | } else { |
152f6476 JA |
859 | ios->creads++; |
860 | ios->cread_kb += bytes >> 10; | |
861 | } | |
862 | } | |
863 | ||
864 | static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci, | |
865 | int rw, int bytes) | |
866 | { | |
867 | __account_c(&pci->io_stats, rw, bytes); | |
868 | ||
869 | if (per_process_stats) { | |
870 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
871 | ||
872 | __account_c(ios, rw, bytes); | |
d0ca268b JA |
873 | } |
874 | } | |
875 | ||
b6076a9b JA |
876 | static inline void __account_issue(struct io_stats *ios, int rw, |
877 | unsigned int bytes) | |
afd2d7ad | 878 | { |
879 | if (rw) { | |
152f6476 JA |
880 | ios->iwrites++; |
881 | ios->iwrite_kb += bytes >> 10; | |
afd2d7ad | 882 | } else { |
152f6476 JA |
883 | ios->ireads++; |
884 | ios->iread_kb += bytes >> 10; | |
afd2d7ad | 885 | } |
886 | } | |
887 | ||
b6076a9b JA |
888 | static inline void account_issue(struct blk_io_trace *t, |
889 | struct per_cpu_info *pci, int rw) | |
d0ca268b | 890 | { |
b6076a9b | 891 | __account_issue(&pci->io_stats, rw, t->bytes); |
152f6476 JA |
892 | |
893 | if (per_process_stats) { | |
894 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
d5396421 | 895 | |
b6076a9b | 896 | __account_issue(ios, rw, t->bytes); |
152f6476 JA |
897 | } |
898 | } | |
899 | ||
06639b27 JA |
900 | static inline void __account_unplug(struct io_stats *ios, int timer) |
901 | { | |
902 | if (timer) | |
903 | ios->timer_unplugs++; | |
904 | else | |
905 | ios->io_unplugs++; | |
906 | } | |
907 | ||
908 | static inline void account_unplug(struct blk_io_trace *t, | |
909 | struct per_cpu_info *pci, int timer) | |
910 | { | |
911 | __account_unplug(&pci->io_stats, timer); | |
912 | ||
913 | if (per_process_stats) { | |
914 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
915 | ||
916 | __account_unplug(ios, timer); | |
917 | } | |
918 | } | |
919 | ||
4054070a JA |
920 | static inline void __account_requeue(struct io_stats *ios, |
921 | struct blk_io_trace *t, int rw) | |
922 | { | |
923 | if (rw) { | |
924 | ios->wrqueue++; | |
925 | ios->iwrite_kb -= t_kb(t); | |
926 | } else { | |
927 | ios->rrqueue++; | |
928 | ios->iread_kb -= t_kb(t); | |
929 | } | |
930 | } | |
931 | ||
932 | static inline void account_requeue(struct blk_io_trace *t, | |
933 | struct per_cpu_info *pci, int rw) | |
934 | { | |
935 | __account_requeue(&pci->io_stats, t, rw); | |
936 | ||
937 | if (per_process_stats) { | |
938 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
939 | ||
940 | __account_requeue(ios, t, rw); | |
941 | } | |
942 | } | |
943 | ||
f7bd1a9b JA |
944 | static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci, |
945 | struct blk_io_trace *t, char *act) | |
ab197ca7 | 946 | { |
f7bd1a9b | 947 | process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL); |
ab197ca7 AB |
948 | } |
949 | ||
f7bd1a9b JA |
950 | static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci, |
951 | struct blk_io_trace *t, char *act) | |
b6076a9b | 952 | { |
f7bd1a9b | 953 | process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL); |
b6076a9b JA |
954 | } |
955 | ||
ab197ca7 AB |
956 | static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t, |
957 | char *act) | |
958 | { | |
b6076a9b | 959 | process_fmt(act, pci, t, -1, 0, NULL); |
ab197ca7 | 960 | } |
2e3e8ded | 961 | |
f7bd1a9b JA |
962 | static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci, |
963 | struct blk_io_trace *t, char *act) | |
ab197ca7 | 964 | { |
f7bd1a9b | 965 | process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL); |
d0ca268b JA |
966 | } |
967 | ||
f7bd1a9b JA |
968 | static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci, |
969 | struct blk_io_trace *t, char *act) | |
d0ca268b | 970 | { |
a01516de | 971 | if (act[0] == 'F') |
f7bd1a9b | 972 | log_track_frontmerge(pdi, t); |
2e3e8ded | 973 | |
ab197ca7 | 974 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
d0ca268b JA |
975 | } |
976 | ||
dfe34da1 | 977 | static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 978 | char *act) |
dfe34da1 | 979 | { |
ab197ca7 | 980 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
dfe34da1 JA |
981 | } |
982 | ||
d5396421 | 983 | static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 984 | char *act) |
d0ca268b | 985 | { |
ab197ca7 | 986 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
d0ca268b JA |
987 | } |
988 | ||
ab197ca7 | 989 | static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 990 | char *act) |
67e14fdc | 991 | { |
ab197ca7 | 992 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
67e14fdc JA |
993 | } |
994 | ||
93f1c611 JA |
995 | static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t, |
996 | char *act) | |
997 | { | |
998 | process_fmt(act, pci, t, -1ULL, 0, NULL); | |
999 | } | |
1000 | ||
ab197ca7 | 1001 | static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act) |
d0ca268b | 1002 | { |
ab197ca7 | 1003 | unsigned char *buf = (unsigned char *) t + sizeof(*t); |
d0ca268b | 1004 | |
ab197ca7 | 1005 | process_fmt(act, pci, t, -1ULL, t->pdu_len, buf); |
d0ca268b JA |
1006 | } |
1007 | ||
ff3a732c | 1008 | static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci) |
d0ca268b | 1009 | { |
56f2af81 JA |
1010 | int act = t->action & 0xffff; |
1011 | ||
1012 | switch (act) { | |
d0ca268b | 1013 | case __BLK_TA_QUEUE: |
3639a11e | 1014 | log_generic(pci, t, "Q"); |
d0ca268b JA |
1015 | break; |
1016 | case __BLK_TA_GETRQ: | |
3639a11e | 1017 | log_generic(pci, t, "G"); |
d0ca268b JA |
1018 | break; |
1019 | case __BLK_TA_SLEEPRQ: | |
3639a11e | 1020 | log_generic(pci, t, "S"); |
d0ca268b JA |
1021 | break; |
1022 | case __BLK_TA_REQUEUE: | |
3639a11e | 1023 | log_generic(pci, t, "R"); |
d0ca268b JA |
1024 | break; |
1025 | case __BLK_TA_ISSUE: | |
ab197ca7 | 1026 | log_pc(pci, t, "D"); |
d0ca268b JA |
1027 | break; |
1028 | case __BLK_TA_COMPLETE: | |
3639a11e | 1029 | log_pc(pci, t, "C"); |
d0ca268b | 1030 | break; |
56f2af81 JA |
1031 | case __BLK_TA_INSERT: |
1032 | log_pc(pci, t, "I"); | |
1033 | break; | |
d0ca268b | 1034 | default: |
56f2af81 | 1035 | fprintf(stderr, "Bad pc action %x\n", act); |
87b72777 | 1036 | break; |
d0ca268b | 1037 | } |
d0ca268b JA |
1038 | } |
1039 | ||
f7bd1a9b JA |
1040 | static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi, |
1041 | struct per_cpu_info *pci) | |
d0ca268b JA |
1042 | { |
1043 | int w = t->action & BLK_TC_ACT(BLK_TC_WRITE); | |
7997c5b0 | 1044 | int act = t->action & 0xffff; |
d0ca268b | 1045 | |
7997c5b0 | 1046 | switch (act) { |
d0ca268b | 1047 | case __BLK_TA_QUEUE: |
b6076a9b | 1048 | account_queue(t, pci, w); |
3639a11e | 1049 | log_queue(pci, t, "Q"); |
d0ca268b | 1050 | break; |
b6076a9b | 1051 | case __BLK_TA_INSERT: |
f7bd1a9b | 1052 | log_insert(pdi, pci, t, "I"); |
b6076a9b | 1053 | break; |
d0ca268b | 1054 | case __BLK_TA_BACKMERGE: |
152f6476 | 1055 | account_m(t, pci, w); |
f7bd1a9b | 1056 | log_merge(pdi, pci, t, "M"); |
d0ca268b JA |
1057 | break; |
1058 | case __BLK_TA_FRONTMERGE: | |
152f6476 | 1059 | account_m(t, pci, w); |
f7bd1a9b | 1060 | log_merge(pdi, pci, t, "F"); |
d0ca268b JA |
1061 | break; |
1062 | case __BLK_TA_GETRQ: | |
f7bd1a9b | 1063 | log_track_getrq(pdi, t); |
3639a11e | 1064 | log_generic(pci, t, "G"); |
d0ca268b JA |
1065 | break; |
1066 | case __BLK_TA_SLEEPRQ: | |
3639a11e | 1067 | log_generic(pci, t, "S"); |
d0ca268b JA |
1068 | break; |
1069 | case __BLK_TA_REQUEUE: | |
4054070a | 1070 | account_requeue(t, pci, w); |
3639a11e | 1071 | log_queue(pci, t, "R"); |
d0ca268b JA |
1072 | break; |
1073 | case __BLK_TA_ISSUE: | |
b6076a9b | 1074 | account_issue(t, pci, w); |
f7bd1a9b | 1075 | log_issue(pdi, pci, t, "D"); |
d0ca268b JA |
1076 | break; |
1077 | case __BLK_TA_COMPLETE: | |
152f6476 | 1078 | account_c(t, pci, w, t->bytes); |
f7bd1a9b | 1079 | log_complete(pdi, pci, t, "C"); |
d0ca268b | 1080 | break; |
88b1a526 | 1081 | case __BLK_TA_PLUG: |
3639a11e | 1082 | log_action(pci, t, "P"); |
88b1a526 | 1083 | break; |
3639a11e | 1084 | case __BLK_TA_UNPLUG_IO: |
06639b27 | 1085 | account_unplug(t, pci, 0); |
3639a11e JA |
1086 | log_unplug(pci, t, "U"); |
1087 | break; | |
1088 | case __BLK_TA_UNPLUG_TIMER: | |
06639b27 | 1089 | account_unplug(t, pci, 1); |
3639a11e | 1090 | log_unplug(pci, t, "UT"); |
88b1a526 | 1091 | break; |
93f1c611 JA |
1092 | case __BLK_TA_SPLIT: |
1093 | log_split(pci, t, "X"); | |
1094 | break; | |
1095 | case __BLK_TA_BOUNCE: | |
1096 | log_generic(pci, t, "B"); | |
1097 | break; | |
a8f30e64 JA |
1098 | case __BLK_TA_REMAP: |
1099 | log_generic(pci, t, "A"); | |
1100 | break; | |
d0ca268b JA |
1101 | default: |
1102 | fprintf(stderr, "Bad fs action %x\n", t->action); | |
1f79c4a0 | 1103 | break; |
d0ca268b | 1104 | } |
d0ca268b JA |
1105 | } |
1106 | ||
ff3a732c JA |
1107 | static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci, |
1108 | struct per_dev_info *pdi) | |
d0ca268b JA |
1109 | { |
1110 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) | |
ff3a732c | 1111 | dump_trace_pc(t, pci); |
d0ca268b | 1112 | else |
f7bd1a9b | 1113 | dump_trace_fs(t, pdi, pci); |
87b72777 | 1114 | |
20ed6177 JA |
1115 | if (!pdi->events) |
1116 | pdi->first_reported_time = t->time; | |
1117 | ||
e7c9f3ff | 1118 | pdi->events++; |
d0ca268b JA |
1119 | } |
1120 | ||
4c523165 JA |
1121 | /* |
1122 | * print in a proper way, not too small and not too big. if more than | |
1123 | * 1000,000K, turn into M and so on | |
1124 | */ | |
1125 | static char *size_cnv(char *dst, unsigned long long num, int in_kb) | |
1126 | { | |
da19e768 | 1127 | char suff[] = { '\0', 'K', 'M', 'G', 'P' }; |
0dc3602c | 1128 | unsigned int i = 0; |
4c523165 JA |
1129 | |
1130 | if (in_kb) | |
1131 | i++; | |
1132 | ||
0dc3602c | 1133 | while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) { |
4c523165 JA |
1134 | i++; |
1135 | num /= 1000; | |
1136 | } | |
1137 | ||
1138 | sprintf(dst, "%'8Lu%c", num, suff[i]); | |
1139 | return dst; | |
1140 | } | |
1141 | ||
152f6476 | 1142 | static void dump_io_stats(struct io_stats *ios, char *msg) |
5c017e4b | 1143 | { |
4c523165 JA |
1144 | static char x[256], y[256]; |
1145 | ||
152f6476 JA |
1146 | fprintf(ofp, "%s\n", msg); |
1147 | ||
4c523165 JA |
1148 | fprintf(ofp, " Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1)); |
1149 | fprintf(ofp, " Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1)); | |
0a6b8fc4 | 1150 | |
4c523165 JA |
1151 | fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1)); |
1152 | fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1)); | |
4054070a JA |
1153 | fprintf(ofp, " Reads Requeued: %s\t\t", size_cnv(x, ios->rrqueue, 0)); |
1154 | fprintf(ofp, " Writes Requeued: %s\n", size_cnv(x, ios->wrqueue, 0)); | |
4c523165 JA |
1155 | fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1)); |
1156 | fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1)); | |
152f6476 | 1157 | fprintf(ofp, " Read Merges: %'8lu%8c\t", ios->mreads, ' '); |
152f6476 | 1158 | fprintf(ofp, " Write Merges: %'8lu\n", ios->mwrites); |
06639b27 JA |
1159 | fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' '); |
1160 | fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs); | |
5c017e4b JA |
1161 | } |
1162 | ||
50adc0ba JA |
1163 | static void dump_wait_stats(struct per_process_info *ppi) |
1164 | { | |
b9d40d6f JA |
1165 | unsigned long rawait = ppi->longest_allocation_wait[0] / 1000; |
1166 | unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000; | |
1167 | unsigned long rcwait = ppi->longest_completion_wait[0] / 1000; | |
1168 | unsigned long wawait = ppi->longest_allocation_wait[1] / 1000; | |
1169 | unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000; | |
1170 | unsigned long wcwait = ppi->longest_completion_wait[1] / 1000; | |
1171 | ||
1172 | fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' '); | |
1173 | fprintf(ofp, " Allocation wait: %'8lu\n", wawait); | |
1174 | fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' '); | |
1175 | fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait); | |
1176 | fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' '); | |
1177 | fprintf(ofp, " Completion wait: %'8lu\n", wcwait); | |
50adc0ba JA |
1178 | } |
1179 | ||
886ecf0e JA |
1180 | static int ppi_name_compare(const void *p1, const void *p2) |
1181 | { | |
1182 | struct per_process_info *ppi1 = *((struct per_process_info **) p1); | |
1183 | struct per_process_info *ppi2 = *((struct per_process_info **) p2); | |
1184 | int res; | |
1185 | ||
1186 | res = strverscmp(ppi1->name, ppi2->name); | |
1187 | if (!res) | |
06e6f286 | 1188 | res = ppi1->pid > ppi2->pid; |
886ecf0e JA |
1189 | |
1190 | return res; | |
1191 | } | |
1192 | ||
1193 | static void sort_process_list(void) | |
1194 | { | |
1195 | struct per_process_info **ppis; | |
1196 | struct per_process_info *ppi; | |
1197 | int i = 0; | |
1198 | ||
1199 | ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *)); | |
1200 | ||
1201 | ppi = ppi_list; | |
1202 | while (ppi) { | |
06e6f286 | 1203 | ppis[i++] = ppi; |
886ecf0e JA |
1204 | ppi = ppi->list_next; |
1205 | } | |
1206 | ||
06e6f286 | 1207 | qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare); |
886ecf0e JA |
1208 | |
1209 | i = ppi_list_entries - 1; | |
1210 | ppi_list = NULL; | |
1211 | while (i >= 0) { | |
1212 | ppi = ppis[i]; | |
1213 | ||
1214 | ppi->list_next = ppi_list; | |
1215 | ppi_list = ppi; | |
1216 | i--; | |
1217 | } | |
50c38702 JA |
1218 | |
1219 | free(ppis); | |
886ecf0e JA |
1220 | } |
1221 | ||
152f6476 JA |
1222 | static void show_process_stats(void) |
1223 | { | |
1224 | struct per_process_info *ppi; | |
1225 | ||
886ecf0e JA |
1226 | sort_process_list(); |
1227 | ||
152f6476 JA |
1228 | ppi = ppi_list; |
1229 | while (ppi) { | |
ce8b6b4f JA |
1230 | char name[64]; |
1231 | ||
715d8021 | 1232 | if (ppi->more_than_one) |
bf0720af | 1233 | sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid); |
715d8021 JA |
1234 | else |
1235 | sprintf(name, "%s (%u)", ppi->name, ppi->pid); | |
bf0720af | 1236 | |
ce8b6b4f | 1237 | dump_io_stats(&ppi->io_stats, name); |
50adc0ba | 1238 | dump_wait_stats(ppi); |
152f6476 JA |
1239 | ppi = ppi->list_next; |
1240 | } | |
1241 | ||
1242 | fprintf(ofp, "\n"); | |
1243 | } | |
1244 | ||
e7c9f3ff | 1245 | static void show_device_and_cpu_stats(void) |
d0ca268b | 1246 | { |
e7c9f3ff NS |
1247 | struct per_dev_info *pdi; |
1248 | struct per_cpu_info *pci; | |
1249 | struct io_stats total, *ios; | |
20ed6177 | 1250 | unsigned long long rrate, wrate, msec; |
e7c9f3ff NS |
1251 | int i, j, pci_events; |
1252 | char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3]; | |
1253 | char name[32]; | |
1254 | ||
1255 | for (pdi = devices, i = 0; i < ndevices; i++, pdi++) { | |
1256 | ||
1257 | memset(&total, 0, sizeof(total)); | |
1258 | pci_events = 0; | |
1259 | ||
1260 | if (i > 0) | |
1261 | fprintf(ofp, "\n"); | |
1262 | ||
1263 | for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) { | |
1264 | if (!pci->nelems) | |
1265 | continue; | |
1266 | ||
1267 | ios = &pci->io_stats; | |
1268 | total.qreads += ios->qreads; | |
1269 | total.qwrites += ios->qwrites; | |
1270 | total.creads += ios->creads; | |
1271 | total.cwrites += ios->cwrites; | |
1272 | total.mreads += ios->mreads; | |
1273 | total.mwrites += ios->mwrites; | |
1274 | total.ireads += ios->ireads; | |
1275 | total.iwrites += ios->iwrites; | |
4054070a JA |
1276 | total.rrqueue += ios->rrqueue; |
1277 | total.wrqueue += ios->wrqueue; | |
e7c9f3ff NS |
1278 | total.qread_kb += ios->qread_kb; |
1279 | total.qwrite_kb += ios->qwrite_kb; | |
1280 | total.cread_kb += ios->cread_kb; | |
1281 | total.cwrite_kb += ios->cwrite_kb; | |
1282 | total.iread_kb += ios->iread_kb; | |
1283 | total.iwrite_kb += ios->iwrite_kb; | |
06639b27 JA |
1284 | total.timer_unplugs += ios->timer_unplugs; |
1285 | total.io_unplugs += ios->io_unplugs; | |
e7c9f3ff NS |
1286 | |
1287 | snprintf(line, sizeof(line) - 1, "CPU%d (%s):", | |
1288 | j, get_dev_name(pdi, name, sizeof(name))); | |
1289 | dump_io_stats(ios, line); | |
1290 | pci_events++; | |
1291 | } | |
5c017e4b | 1292 | |
e7c9f3ff NS |
1293 | if (pci_events > 1) { |
1294 | fprintf(ofp, "\n"); | |
1295 | snprintf(line, sizeof(line) - 1, "Total (%s):", | |
1296 | get_dev_name(pdi, name, sizeof(name))); | |
1297 | dump_io_stats(&total, line); | |
1298 | } | |
d0ca268b | 1299 | |
20ed6177 | 1300 | wrate = rrate = 0; |
20ed6177 JA |
1301 | msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000; |
1302 | if (msec) { | |
1303 | rrate = 1000 * total.cread_kb / msec; | |
1304 | wrate = 1000 * total.cwrite_kb / msec; | |
1305 | } | |
1306 | ||
dce0f678 AB |
1307 | fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n", |
1308 | rrate, wrate); | |
1309 | fprintf(ofp, "Events (%s): %'Lu entries\n", | |
1310 | get_dev_name(pdi, line, sizeof(line)), pdi->events); | |
1311 | fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%) %'lu backward (%'llu - %5.1lf%%)\n", | |
1312 | pdi->skips,pdi->seq_skips, | |
1313 | 100.0 * ((double)pdi->seq_skips / | |
1314 | (double)(pdi->events + pdi->seq_skips)), | |
1315 | pdi->nskips,pdi->seq_nskips, | |
1316 | 100.0 * ((double)pdi->seq_nskips / | |
1317 | (double)(pdi->events + pdi->seq_skips))); | |
e7c9f3ff | 1318 | } |
d0ca268b JA |
1319 | } |
1320 | ||
d36421e4 JA |
1321 | /* |
1322 | * struct trace and blktrace allocation cache, we do potentially | |
1323 | * millions of mallocs for these structures while only using at most | |
1324 | * a few thousand at the time | |
1325 | */ | |
1326 | static inline void t_free(struct trace *t) | |
1327 | { | |
3c8413cb | 1328 | if (t_alloc_cache < 1024) { |
1d24fc14 JA |
1329 | t->next = t_alloc_list; |
1330 | t_alloc_list = t; | |
3c8413cb JA |
1331 | t_alloc_cache++; |
1332 | } else | |
1333 | free(t); | |
d36421e4 JA |
1334 | } |
1335 | ||
1336 | static inline struct trace *t_alloc(void) | |
1337 | { | |
1338 | struct trace *t = t_alloc_list; | |
1339 | ||
1340 | if (t) { | |
1341 | t_alloc_list = t->next; | |
3c8413cb | 1342 | t_alloc_cache--; |
d36421e4 JA |
1343 | return t; |
1344 | } | |
1345 | ||
1346 | return malloc(sizeof(*t)); | |
1347 | } | |
1348 | ||
1349 | static inline void bit_free(struct blk_io_trace *bit) | |
1350 | { | |
eb9bd4e9 | 1351 | if (bit_alloc_cache < 1024 && !bit->pdu_len) { |
1d24fc14 JA |
1352 | /* |
1353 | * abuse a 64-bit field for a next pointer for the free item | |
1354 | */ | |
1355 | bit->time = (__u64) (unsigned long) bit_alloc_list; | |
1356 | bit_alloc_list = (struct blk_io_trace *) bit; | |
3c8413cb JA |
1357 | bit_alloc_cache++; |
1358 | } else | |
1359 | free(bit); | |
d36421e4 JA |
1360 | } |
1361 | ||
1362 | static inline struct blk_io_trace *bit_alloc(void) | |
1363 | { | |
1364 | struct blk_io_trace *bit = bit_alloc_list; | |
1365 | ||
1366 | if (bit) { | |
0c39425c NS |
1367 | bit_alloc_list = (struct blk_io_trace *) (unsigned long) \ |
1368 | bit->time; | |
3c8413cb | 1369 | bit_alloc_cache--; |
d36421e4 JA |
1370 | return bit; |
1371 | } | |
1372 | ||
1373 | return malloc(sizeof(*bit)); | |
1374 | } | |
1375 | ||
4f0ae44f JA |
1376 | static void find_genesis(void) |
1377 | { | |
1378 | struct trace *t = trace_list; | |
1379 | ||
1380 | genesis_time = -1ULL; | |
1381 | while (t != NULL) { | |
1382 | if (t->bit->time < genesis_time) | |
1383 | genesis_time = t->bit->time; | |
1384 | ||
1385 | t = t->next; | |
1386 | } | |
1387 | } | |
1388 | ||
7f4d89e6 | 1389 | static inline int check_stopwatch(struct blk_io_trace *bit) |
4f0ae44f | 1390 | { |
7f4d89e6 JA |
1391 | if (bit->time < stopwatch_end && |
1392 | bit->time >= stopwatch_start) | |
4f0ae44f JA |
1393 | return 0; |
1394 | ||
1395 | return 1; | |
1396 | } | |
1397 | ||
53c68c88 JA |
1398 | /* |
1399 | * return youngest entry read | |
1400 | */ | |
1401 | static int sort_entries(unsigned long long *youngest) | |
4f0ae44f JA |
1402 | { |
1403 | struct trace *t; | |
4f0ae44f JA |
1404 | |
1405 | if (!genesis_time) | |
1406 | find_genesis(); | |
1407 | ||
d6222db8 | 1408 | *youngest = 0; |
4f0ae44f JA |
1409 | while ((t = trace_list) != NULL) { |
1410 | struct blk_io_trace *bit = t->bit; | |
1411 | ||
1412 | trace_list = t->next; | |
1413 | ||
7f4d89e6 | 1414 | bit->time -= genesis_time; |
4f0ae44f | 1415 | |
d6222db8 JA |
1416 | if (bit->time < *youngest || !*youngest) |
1417 | *youngest = bit->time; | |
1418 | ||
774a1a10 JA |
1419 | if (bit->sequence < smallest_seq_read) |
1420 | smallest_seq_read = bit->sequence; | |
1421 | ||
7f4d89e6 | 1422 | if (check_stopwatch(bit)) { |
4f0ae44f JA |
1423 | bit_free(bit); |
1424 | t_free(t); | |
1425 | continue; | |
1426 | } | |
1427 | ||
2a1b3424 | 1428 | if (trace_rb_insert_sort(t)) |
53c68c88 | 1429 | return -1; |
4f0ae44f JA |
1430 | } |
1431 | ||
53c68c88 | 1432 | return 0; |
4f0ae44f JA |
1433 | } |
1434 | ||
1c24add6 JA |
1435 | static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t) |
1436 | { | |
1437 | rb_erase(&t->rb_node, &pdi->rb_last); | |
1438 | pdi->rb_last_entries--; | |
1439 | ||
1440 | bit_free(t->bit); | |
1441 | t_free(t); | |
1442 | } | |
1443 | ||
1444 | static void put_trace(struct per_dev_info *pdi, struct trace *t) | |
2a1b3424 JA |
1445 | { |
1446 | rb_erase(&t->rb_node, &rb_sort_root); | |
1447 | rb_sort_entries--; | |
1448 | ||
1449 | trace_rb_insert_last(pdi, t); | |
1450 | ||
1c24add6 | 1451 | if (pdi->rb_last_entries > rb_batch * pdi->nfiles) { |
2a1b3424 JA |
1452 | struct rb_node *n = rb_first(&pdi->rb_last); |
1453 | ||
1454 | t = rb_entry(n, struct trace, rb_node); | |
1c24add6 | 1455 | __put_trace_last(pdi, t); |
2a1b3424 JA |
1456 | } |
1457 | } | |
1458 | ||
824c2b39 JA |
1459 | /* |
1460 | * to continue, we must have traces from all online cpus in the tree | |
1461 | */ | |
1462 | static int check_cpu_map(struct per_dev_info *pdi) | |
1463 | { | |
1464 | unsigned long *cpu_map; | |
1465 | struct rb_node *n; | |
1466 | struct trace *__t; | |
1467 | unsigned int i; | |
1468 | int ret, cpu; | |
1469 | ||
1470 | /* | |
1471 | * create a map of the cpus we have traces for | |
1472 | */ | |
1473 | cpu_map = malloc(pdi->cpu_map_max / sizeof(long)); | |
1474 | n = rb_first(&rb_sort_root); | |
1475 | while (n) { | |
1476 | __t = rb_entry(n, struct trace, rb_node); | |
1477 | cpu = __t->bit->cpu; | |
1478 | ||
1479 | cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu)); | |
1480 | n = rb_next(n); | |
1481 | } | |
1482 | ||
1483 | /* | |
b1c8e614 JA |
1484 | * we can't continue if pdi->cpu_map has entries set that we don't |
1485 | * have in the sort rbtree. the opposite is not a problem, though | |
824c2b39 JA |
1486 | */ |
1487 | ret = 0; | |
1488 | for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) { | |
1489 | if (pdi->cpu_map[i] & ~(cpu_map[i])) { | |
1490 | ret = 1; | |
1491 | break; | |
1492 | } | |
1493 | } | |
1494 | ||
1495 | free(cpu_map); | |
1496 | return ret; | |
1497 | } | |
1498 | ||
a141a7cd | 1499 | static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force) |
2a1b3424 JA |
1500 | { |
1501 | unsigned long expected_sequence = pdi->last_sequence + 1; | |
1ca323a5 JA |
1502 | struct blk_io_trace *bit = t->bit; |
1503 | struct trace *__t; | |
d6222db8 | 1504 | |
774a1a10 | 1505 | if (!expected_sequence) { |
774a1a10 JA |
1506 | /* |
1507 | * 1 should be the first entry, just allow it | |
1508 | */ | |
1509 | if (bit->sequence == 1) | |
1510 | return 0; | |
1511 | ||
824c2b39 | 1512 | return check_cpu_map(pdi); |
774a1a10 | 1513 | } |
2a1b3424 JA |
1514 | |
1515 | if (bit->sequence == expected_sequence) | |
1516 | return 0; | |
1517 | ||
2a1b3424 | 1518 | /* |
1c7c54aa JA |
1519 | * we may not have seen that sequence yet. if we are not doing |
1520 | * the final run, break and wait for more entries. | |
1c24add6 | 1521 | */ |
1c7c54aa | 1522 | if (expected_sequence < smallest_seq_read) { |
1ca323a5 JA |
1523 | __t = trace_rb_find_last(pdi, expected_sequence); |
1524 | if (!__t) | |
1c7c54aa | 1525 | goto skip; |
2a1b3424 | 1526 | |
1ca323a5 | 1527 | __put_trace_last(pdi, __t); |
2a1b3424 | 1528 | return 0; |
a141a7cd JA |
1529 | } else if (!force) { |
1530 | return 1; | |
0b07f23e | 1531 | } else { |
1c7c54aa | 1532 | skip: |
57ea8602 | 1533 | if (verbose) { |
7d1c0411 JA |
1534 | fprintf(stderr, "(%d,%d): skipping %lu -> %u\n", |
1535 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
1536 | pdi->last_sequence, bit->sequence); | |
1537 | } | |
dce0f678 AB |
1538 | if (bit->sequence > pdi->last_sequence) { |
1539 | pdi->skips++; | |
1540 | pdi->seq_skips += (bit->sequence - pdi->last_sequence); | |
1541 | } else { | |
1542 | pdi->nskips++; | |
1543 | pdi->seq_nskips += (pdi->last_sequence - bit->sequence); | |
1544 | } | |
1c7c54aa JA |
1545 | return 0; |
1546 | } | |
2a1b3424 JA |
1547 | } |
1548 | ||
a649216c | 1549 | static void show_entries_rb(int force) |
8fc0abbc | 1550 | { |
1f7afa72 JA |
1551 | struct per_dev_info *pdi = NULL; |
1552 | struct per_cpu_info *pci = NULL; | |
8fc0abbc | 1553 | struct blk_io_trace *bit; |
3aabcd89 | 1554 | struct rb_node *n; |
8fc0abbc | 1555 | struct trace *t; |
1f7afa72 | 1556 | |
7d747d22 | 1557 | while ((n = rb_first(&rb_sort_root)) != NULL) { |
dd90748f | 1558 | if (is_done() && !force && !pipeline) |
1f7afa72 | 1559 | break; |
8fc0abbc JA |
1560 | |
1561 | t = rb_entry(n, struct trace, rb_node); | |
1562 | bit = t->bit; | |
1563 | ||
f7bd1a9b | 1564 | if (!pdi || pdi->dev != bit->device) |
287fa3d6 | 1565 | pdi = get_dev_info(bit->device); |
1f7afa72 | 1566 | |
e7c9f3ff NS |
1567 | if (!pdi) { |
1568 | fprintf(stderr, "Unknown device ID? (%d,%d)\n", | |
1569 | MAJOR(bit->device), MINOR(bit->device)); | |
1570 | break; | |
1571 | } | |
1f7afa72 | 1572 | |
a141a7cd JA |
1573 | if (check_sequence(pdi, t, force)) |
1574 | break; | |
cb2a1a62 | 1575 | |
a141a7cd JA |
1576 | if (!force && bit->time > last_allowed_time) |
1577 | break; | |
8fc0abbc | 1578 | |
be925321 JA |
1579 | pdi->last_sequence = bit->sequence; |
1580 | ||
4f0ae44f | 1581 | check_time(pdi, bit); |
8fc0abbc | 1582 | |
4f0ae44f JA |
1583 | if (!pci || pci->cpu != bit->cpu) |
1584 | pci = get_cpu_info(pdi, bit->cpu); | |
287fa3d6 | 1585 | |
cbc927b6 JA |
1586 | pci->nelems++; |
1587 | ||
98f8386b AB |
1588 | if (bit->action & (act_mask << BLK_TC_SHIFT)) |
1589 | dump_trace(bit, pci, pdi); | |
87b72777 | 1590 | |
2a1b3424 | 1591 | put_trace(pdi, t); |
cb2a1a62 | 1592 | } |
8fc0abbc JA |
1593 | } |
1594 | ||
1f79c4a0 JA |
1595 | static int read_data(int fd, void *buffer, int bytes, int block) |
1596 | { | |
1597 | int ret, bytes_left, fl; | |
1598 | void *p; | |
1599 | ||
1600 | fl = fcntl(fd, F_GETFL); | |
1601 | ||
1602 | if (!block) | |
1603 | fcntl(fd, F_SETFL, fl | O_NONBLOCK); | |
1604 | else | |
1605 | fcntl(fd, F_SETFL, fl & ~O_NONBLOCK); | |
1606 | ||
1607 | bytes_left = bytes; | |
1608 | p = buffer; | |
1609 | while (bytes_left > 0) { | |
1610 | ret = read(fd, p, bytes_left); | |
1611 | if (!ret) | |
1612 | return 1; | |
1613 | else if (ret < 0) { | |
1614 | if (errno != EAGAIN) | |
1615 | perror("read"); | |
a649216c | 1616 | |
1f79c4a0 JA |
1617 | return -1; |
1618 | } else { | |
1619 | p += ret; | |
1620 | bytes_left -= ret; | |
1621 | } | |
1622 | } | |
1623 | ||
1624 | return 0; | |
1625 | } | |
1626 | ||
a649216c | 1627 | static int read_events(int fd, int always_block) |
cb2a1a62 | 1628 | { |
287fa3d6 | 1629 | struct per_dev_info *pdi = NULL; |
e820abd7 | 1630 | unsigned int events = 0; |
7d747d22 JA |
1631 | |
1632 | while (!is_done() && events < rb_batch) { | |
1633 | struct blk_io_trace *bit; | |
1634 | struct trace *t; | |
1635 | int pdu_len; | |
1636 | __u32 magic; | |
1637 | ||
d36421e4 | 1638 | bit = bit_alloc(); |
cb2a1a62 | 1639 | |
eb9bd4e9 JA |
1640 | if (read_data(fd, bit, sizeof(*bit), !events || always_block)) { |
1641 | bit_free(bit); | |
cb2a1a62 | 1642 | break; |
eb9bd4e9 | 1643 | } |
cb2a1a62 | 1644 | |
7d747d22 JA |
1645 | magic = be32_to_cpu(bit->magic); |
1646 | if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { | |
1647 | fprintf(stderr, "Bad magic %x\n", magic); | |
1648 | break; | |
1649 | } | |
1650 | ||
1651 | pdu_len = be16_to_cpu(bit->pdu_len); | |
1652 | if (pdu_len) { | |
1653 | void *ptr = realloc(bit, sizeof(*bit) + pdu_len); | |
1654 | ||
eb9bd4e9 JA |
1655 | if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1)) { |
1656 | bit_free(ptr); | |
7d747d22 | 1657 | break; |
eb9bd4e9 | 1658 | } |
7d747d22 JA |
1659 | |
1660 | bit = ptr; | |
1661 | } | |
1662 | ||
d6222db8 JA |
1663 | trace_to_cpu(bit); |
1664 | ||
1665 | if (verify_trace(bit)) { | |
1666 | bit_free(bit); | |
1667 | continue; | |
1668 | } | |
1669 | ||
d36421e4 | 1670 | t = t_alloc(); |
cb2a1a62 JA |
1671 | memset(t, 0, sizeof(*t)); |
1672 | t->bit = bit; | |
1673 | ||
7d747d22 JA |
1674 | t->next = trace_list; |
1675 | trace_list = t; | |
1f7afa72 | 1676 | |
f7bd1a9b | 1677 | if (!pdi || pdi->dev != bit->device) |
287fa3d6 JA |
1678 | pdi = get_dev_info(bit->device); |
1679 | ||
1680 | if (bit->time > pdi->last_read_time) | |
1681 | pdi->last_read_time = bit->time; | |
1682 | ||
7d747d22 | 1683 | events++; |
cb2a1a62 JA |
1684 | } |
1685 | ||
7d747d22 | 1686 | return events; |
cb2a1a62 JA |
1687 | } |
1688 | ||
d5396421 | 1689 | static int do_file(void) |
d0ca268b | 1690 | { |
7d747d22 | 1691 | struct per_cpu_info *pci; |
73877e12 NS |
1692 | struct per_dev_info *pdi; |
1693 | int i, j, events, events_added; | |
d0ca268b | 1694 | |
7d747d22 JA |
1695 | /* |
1696 | * first prepare all files for reading | |
1697 | */ | |
e8741a4a | 1698 | for (i = 0; i < ndevices; i++) { |
73877e12 NS |
1699 | pdi = &devices[i]; |
1700 | pdi->nfiles = 0; | |
d6222db8 | 1701 | pdi->last_sequence = -1; |
73877e12 | 1702 | |
74feaeb5 | 1703 | for (j = 0;; j++) { |
e7c9f3ff | 1704 | struct stat st; |
d1d7f15f | 1705 | int len = 0; |
6e0073ed | 1706 | char *p, *dname; |
87b72777 | 1707 | |
e7c9f3ff NS |
1708 | pci = get_cpu_info(pdi, j); |
1709 | pci->cpu = j; | |
7d747d22 | 1710 | pci->fd = -1; |
6e0073ed JA |
1711 | |
1712 | p = strdup(pdi->name); | |
1713 | dname = dirname(p); | |
1714 | if (strcmp(dname, ".")) { | |
1715 | input_dir = dname; | |
1716 | p = strdup(pdi->name); | |
1717 | strcpy(pdi->name, basename(p)); | |
1718 | } | |
1719 | free(p); | |
d0ca268b | 1720 | |
d1d7f15f JA |
1721 | if (input_dir) |
1722 | len = sprintf(pci->fname, "%s/", input_dir); | |
1723 | ||
1724 | snprintf(pci->fname + len, sizeof(pci->fname)-1-len, | |
7d747d22 | 1725 | "%s.blktrace.%d", pdi->name, pci->cpu); |
e7c9f3ff NS |
1726 | if (stat(pci->fname, &st) < 0) |
1727 | break; | |
74feaeb5 AB |
1728 | if (st.st_size) { |
1729 | pci->fd = open(pci->fname, O_RDONLY); | |
1730 | if (pci->fd < 0) { | |
1731 | perror(pci->fname); | |
1732 | continue; | |
1733 | } | |
e7c9f3ff NS |
1734 | } |
1735 | ||
7d747d22 | 1736 | printf("Input file %s added\n", pci->fname); |
73877e12 | 1737 | pdi->nfiles++; |
824c2b39 | 1738 | cpu_mark_online(pdi, pci->cpu); |
d0ca268b | 1739 | } |
d5396421 JA |
1740 | } |
1741 | ||
7d747d22 JA |
1742 | /* |
1743 | * now loop over the files reading in the data | |
1744 | */ | |
412819ce | 1745 | do { |
53c68c88 JA |
1746 | unsigned long long youngest; |
1747 | ||
7d747d22 | 1748 | events_added = 0; |
287fa3d6 | 1749 | last_allowed_time = -1ULL; |
1c7c54aa | 1750 | smallest_seq_read = -1U; |
d5396421 | 1751 | |
7d747d22 | 1752 | for (i = 0; i < ndevices; i++) { |
73877e12 NS |
1753 | pdi = &devices[i]; |
1754 | ||
1755 | for (j = 0; j < pdi->nfiles; j++) { | |
d5396421 | 1756 | |
73877e12 | 1757 | pci = get_cpu_info(pdi, j); |
d5396421 | 1758 | |
7d747d22 JA |
1759 | if (pci->fd == -1) |
1760 | continue; | |
51128a28 | 1761 | |
a649216c | 1762 | events = read_events(pci->fd, 1); |
7d747d22 | 1763 | if (!events) { |
824c2b39 | 1764 | cpu_mark_offline(pdi, pci->cpu); |
7d747d22 JA |
1765 | close(pci->fd); |
1766 | pci->fd = -1; | |
1767 | continue; | |
1768 | } | |
d5396421 | 1769 | |
287fa3d6 JA |
1770 | if (pdi->last_read_time < last_allowed_time) |
1771 | last_allowed_time = pdi->last_read_time; | |
d5396421 | 1772 | |
7d747d22 JA |
1773 | events_added += events; |
1774 | } | |
2ff323b0 | 1775 | } |
d5396421 | 1776 | |
53c68c88 JA |
1777 | if (sort_entries(&youngest)) |
1778 | break; | |
1779 | ||
1780 | if (youngest > stopwatch_end) | |
287fa3d6 JA |
1781 | break; |
1782 | ||
a649216c | 1783 | show_entries_rb(0); |
cb2a1a62 | 1784 | |
7d747d22 | 1785 | } while (events_added); |
d5396421 | 1786 | |
a649216c JA |
1787 | if (rb_sort_entries) |
1788 | show_entries_rb(1); | |
1789 | ||
7d747d22 | 1790 | return 0; |
412819ce | 1791 | } |
d5396421 | 1792 | |
412819ce JA |
1793 | static int do_stdin(void) |
1794 | { | |
53c68c88 | 1795 | unsigned long long youngest; |
763d936e | 1796 | int fd, events; |
d5396421 | 1797 | |
be925321 | 1798 | last_allowed_time = -1ULL; |
1f79c4a0 | 1799 | fd = dup(STDIN_FILENO); |
0b07f23e JA |
1800 | if (fd == -1) { |
1801 | perror("dup stdin"); | |
1802 | return -1; | |
1803 | } | |
d5396421 | 1804 | |
0b07f23e | 1805 | while ((events = read_events(fd, 0)) != 0) { |
412819ce | 1806 | |
0b07f23e JA |
1807 | smallest_seq_read = -1U; |
1808 | ||
53c68c88 JA |
1809 | if (sort_entries(&youngest)) |
1810 | break; | |
1811 | ||
1812 | if (youngest > stopwatch_end) | |
2ff323b0 JA |
1813 | break; |
1814 | ||
763d936e | 1815 | show_entries_rb(0); |
0b07f23e | 1816 | } |
d5396421 | 1817 | |
a649216c JA |
1818 | if (rb_sort_entries) |
1819 | show_entries_rb(1); | |
1820 | ||
d5396421 | 1821 | close(fd); |
d5396421 JA |
1822 | return 0; |
1823 | } | |
d0ca268b | 1824 | |
cbc927b6 | 1825 | static void show_stats(void) |
412819ce | 1826 | { |
cbc927b6 JA |
1827 | if (!ofp) |
1828 | return; | |
1829 | if (stats_printed) | |
1830 | return; | |
1831 | ||
1832 | stats_printed = 1; | |
1833 | ||
1834 | if (per_process_stats) | |
1835 | show_process_stats(); | |
1836 | ||
1837 | if (per_device_and_cpu_stats) | |
1838 | show_device_and_cpu_stats(); | |
1839 | ||
152f6476 | 1840 | fflush(ofp); |
412819ce JA |
1841 | } |
1842 | ||
e820abd7 | 1843 | static void handle_sigint(__attribute__((__unused__)) int sig) |
412819ce JA |
1844 | { |
1845 | done = 1; | |
cbc927b6 | 1846 | show_stats(); |
412819ce JA |
1847 | } |
1848 | ||
46e6968b NS |
1849 | /* |
1850 | * Extract start and duration times from a string, allowing | |
1851 | * us to specify a time interval of interest within a trace. | |
1852 | * Format: "duration" (start is zero) or "start:duration". | |
1853 | */ | |
1854 | static int find_stopwatch_interval(char *string) | |
1855 | { | |
1856 | double value; | |
1857 | char *sp; | |
1858 | ||
1859 | value = strtod(string, &sp); | |
1860 | if (sp == string) { | |
1861 | fprintf(stderr,"Invalid stopwatch timer: %s\n", string); | |
1862 | return 1; | |
1863 | } | |
1864 | if (*sp == ':') { | |
1865 | stopwatch_start = DOUBLE_TO_NANO_ULL(value); | |
1866 | string = sp + 1; | |
1867 | value = strtod(string, &sp); | |
1868 | if (sp == string || *sp != '\0') { | |
1869 | fprintf(stderr,"Invalid stopwatch duration time: %s\n", | |
1870 | string); | |
1871 | return 1; | |
1872 | } | |
1873 | } else if (*sp != '\0') { | |
1874 | fprintf(stderr,"Invalid stopwatch start timer: %s\n", string); | |
1875 | return 1; | |
1876 | } | |
1b928247 JA |
1877 | stopwatch_end = DOUBLE_TO_NANO_ULL(value); |
1878 | if (stopwatch_end <= stopwatch_start) { | |
1879 | fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n", | |
1880 | stopwatch_start, stopwatch_end); | |
1881 | return 1; | |
1882 | } | |
1883 | ||
46e6968b NS |
1884 | return 0; |
1885 | } | |
1886 | ||
52724a0e JA |
1887 | static char usage_str[] = \ |
1888 | "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \ | |
1889 | "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \ | |
1890 | "\t-i Input file containing trace data, or '-' for stdin\n" \ | |
d1d7f15f | 1891 | "\t-D Directory to prepend to input file names\n" \ |
52724a0e JA |
1892 | "\t-o Output file. If not given, output is stdout\n" \ |
1893 | "\t-b stdin read batching\n" \ | |
1894 | "\t-s Show per-program io statistics\n" \ | |
bf0720af | 1895 | "\t-n Hash processes by name, not pid\n" \ |
52724a0e JA |
1896 | "\t-t Track individual ios. Will tell you the time a request took\n" \ |
1897 | "\t to get queued, to get dispatched, and to get completed\n" \ | |
1898 | "\t-q Quiet. Don't display any stats at the end of the trace\n" \ | |
1899 | "\t-w Only parse data between the given time interval in seconds.\n" \ | |
1900 | "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \ | |
d915dee6 JA |
1901 | "\t-f Output format. Customize the output format. The format field\n" \ |
1902 | "\t identifies can be found in the documentation\n" \ | |
52724a0e | 1903 | "\t-F Format specification. Can be found in the documentation\n" \ |
57ea8602 JA |
1904 | "\t-v More verbose for marginal errors\n" \ |
1905 | "\t-V Print program version info\n\n"; | |
52724a0e | 1906 | |
1f79c4a0 JA |
1907 | static void usage(char *prog) |
1908 | { | |
52724a0e | 1909 | fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str); |
1f79c4a0 JA |
1910 | } |
1911 | ||
d5396421 JA |
1912 | int main(int argc, char *argv[]) |
1913 | { | |
152f6476 | 1914 | char *ofp_buffer; |
98f8386b | 1915 | int i, c, ret, mode; |
98f8386b | 1916 | int act_mask_tmp = 0; |
d5396421 JA |
1917 | |
1918 | while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) { | |
1919 | switch (c) { | |
98f8386b AB |
1920 | case 'a': |
1921 | i = find_mask_map(optarg); | |
1922 | if (i < 0) { | |
1923 | fprintf(stderr,"Invalid action mask %s\n", | |
1924 | optarg); | |
1925 | return 1; | |
1926 | } | |
1927 | act_mask_tmp |= i; | |
1928 | break; | |
1929 | ||
1930 | case 'A': | |
1931 | if ((sscanf(optarg, "%x", &i) != 1) || | |
1932 | !valid_act_opt(i)) { | |
1933 | fprintf(stderr, | |
1934 | "Invalid set action mask %s/0x%x\n", | |
1935 | optarg, i); | |
1936 | return 1; | |
1937 | } | |
1938 | act_mask_tmp = i; | |
1939 | break; | |
d5396421 | 1940 | case 'i': |
e7c9f3ff NS |
1941 | if (!strcmp(optarg, "-") && !pipeline) |
1942 | pipeline = 1; | |
1943 | else if (resize_devices(optarg) != 0) | |
1944 | return 1; | |
d5396421 | 1945 | break; |
d1d7f15f JA |
1946 | case 'D': |
1947 | input_dir = optarg; | |
1948 | break; | |
d5396421 | 1949 | case 'o': |
66efebf8 | 1950 | output_name = optarg; |
d5396421 | 1951 | break; |
79f19470 JA |
1952 | case 'b': |
1953 | rb_batch = atoi(optarg); | |
1954 | if (rb_batch <= 0) | |
1955 | rb_batch = RB_BATCH_DEFAULT; | |
1956 | break; | |
152f6476 JA |
1957 | case 's': |
1958 | per_process_stats = 1; | |
1959 | break; | |
7997c5b0 JA |
1960 | case 't': |
1961 | track_ios = 1; | |
1962 | break; | |
1e1c60f1 NS |
1963 | case 'q': |
1964 | per_device_and_cpu_stats = 0; | |
1965 | break; | |
46e6968b NS |
1966 | case 'w': |
1967 | if (find_stopwatch_interval(optarg) != 0) | |
1968 | return 1; | |
1969 | break; | |
ab197ca7 AB |
1970 | case 'f': |
1971 | set_all_format_specs(optarg); | |
1972 | break; | |
1973 | case 'F': | |
1974 | if (add_format_spec(optarg) != 0) | |
1975 | return 1; | |
1976 | break; | |
d915dee6 | 1977 | case 'h': |
715d8021 | 1978 | ppi_hash_by_pid = 0; |
bf0720af | 1979 | break; |
52724a0e | 1980 | case 'v': |
57ea8602 JA |
1981 | verbose++; |
1982 | break; | |
1983 | case 'V': | |
52724a0e JA |
1984 | printf("%s version %s\n", argv[0], blkparse_version); |
1985 | return 0; | |
d5396421 | 1986 | default: |
1f79c4a0 | 1987 | usage(argv[0]); |
d5396421 JA |
1988 | return 1; |
1989 | } | |
d0ca268b JA |
1990 | } |
1991 | ||
e7c9f3ff NS |
1992 | while (optind < argc) { |
1993 | if (!strcmp(argv[optind], "-") && !pipeline) | |
1994 | pipeline = 1; | |
1995 | else if (resize_devices(argv[optind]) != 0) | |
1996 | return 1; | |
1997 | optind++; | |
1998 | } | |
1999 | ||
2000 | if (!pipeline && !ndevices) { | |
1f79c4a0 | 2001 | usage(argv[0]); |
d5396421 JA |
2002 | return 1; |
2003 | } | |
2004 | ||
98f8386b AB |
2005 | if (act_mask_tmp != 0) |
2006 | act_mask = act_mask_tmp; | |
2007 | ||
7997c5b0 | 2008 | memset(&rb_sort_root, 0, sizeof(rb_sort_root)); |
412819ce JA |
2009 | |
2010 | signal(SIGINT, handle_sigint); | |
2011 | signal(SIGHUP, handle_sigint); | |
2012 | signal(SIGTERM, handle_sigint); | |
d5396421 | 2013 | |
d69db225 JA |
2014 | setlocale(LC_NUMERIC, "en_US"); |
2015 | ||
a66877e6 | 2016 | if (!output_name) { |
152f6476 | 2017 | ofp = fdopen(STDOUT_FILENO, "w"); |
a66877e6 JA |
2018 | mode = _IOLBF; |
2019 | } else { | |
152f6476 JA |
2020 | char ofname[128]; |
2021 | ||
7a9690c0 | 2022 | snprintf(ofname, sizeof(ofname) - 1, "%s", output_name); |
152f6476 | 2023 | ofp = fopen(ofname, "w"); |
a66877e6 | 2024 | mode = _IOFBF; |
152f6476 JA |
2025 | } |
2026 | ||
2027 | if (!ofp) { | |
2028 | perror("fopen"); | |
2029 | return 1; | |
2030 | } | |
2031 | ||
2032 | ofp_buffer = malloc(4096); | |
a66877e6 | 2033 | if (setvbuf(ofp, ofp_buffer, mode, 4096)) { |
152f6476 JA |
2034 | perror("setvbuf"); |
2035 | return 1; | |
2036 | } | |
2037 | ||
e7c9f3ff | 2038 | if (pipeline) |
d5396421 JA |
2039 | ret = do_stdin(); |
2040 | else | |
2041 | ret = do_file(); | |
2042 | ||
cbc927b6 | 2043 | show_stats(); |
eb9bd4e9 | 2044 | free(ofp_buffer); |
d5396421 | 2045 | return ret; |
d0ca268b | 2046 | } |