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 | |
52724a0e JA |
38 | static char blkparse_version[] = "0.90"; |
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; |
dcf0f7ed | 51 | unsigned long skips; |
e7c9f3ff | 52 | |
2a1b3424 JA |
53 | struct rb_root rb_last; |
54 | unsigned long rb_last_entries; | |
55 | ||
f7bd1a9b JA |
56 | struct rb_root rb_track; |
57 | ||
73877e12 | 58 | int nfiles; |
e7c9f3ff | 59 | int ncpus; |
824c2b39 JA |
60 | |
61 | unsigned long *cpu_map; | |
62 | unsigned int cpu_map_max; | |
63 | ||
e7c9f3ff NS |
64 | struct per_cpu_info *cpus; |
65 | }; | |
66 | ||
152f6476 JA |
67 | struct per_process_info { |
68 | char name[16]; | |
69 | __u32 pid; | |
70 | struct io_stats io_stats; | |
71 | struct per_process_info *hash_next, *list_next; | |
715d8021 | 72 | int more_than_one; |
50adc0ba JA |
73 | |
74 | /* | |
75 | * individual io stats | |
76 | */ | |
b9d40d6f JA |
77 | unsigned long long longest_allocation_wait[2]; |
78 | unsigned long long longest_dispatch_wait[2]; | |
79 | unsigned long long longest_completion_wait[2]; | |
d0ca268b JA |
80 | }; |
81 | ||
152f6476 | 82 | #define PPI_HASH_SHIFT (8) |
bf0720af JA |
83 | #define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT) |
84 | #define PPI_HASH_MASK (PPI_HASH_SIZE - 1) | |
85 | static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE]; | |
152f6476 | 86 | static struct per_process_info *ppi_list; |
886ecf0e | 87 | static int ppi_list_entries; |
152f6476 | 88 | |
d915dee6 | 89 | #define S_OPTS "a:A:i:o:b:stqw:f:F:vVhD:" |
d5396421 | 90 | static struct option l_opts[] = { |
98f8386b AB |
91 | { |
92 | .name = "act-mask", | |
93 | .has_arg = required_argument, | |
94 | .flag = NULL, | |
95 | .val = 'a' | |
96 | }, | |
97 | { | |
98 | .name = "set-mask", | |
99 | .has_arg = required_argument, | |
100 | .flag = NULL, | |
101 | .val = 'A' | |
102 | }, | |
d5396421 JA |
103 | { |
104 | .name = "input", | |
428683db | 105 | .has_arg = required_argument, |
d5396421 JA |
106 | .flag = NULL, |
107 | .val = 'i' | |
108 | }, | |
109 | { | |
110 | .name = "output", | |
428683db | 111 | .has_arg = required_argument, |
d5396421 JA |
112 | .flag = NULL, |
113 | .val = 'o' | |
114 | }, | |
79f19470 JA |
115 | { |
116 | .name = "batch", | |
428683db | 117 | .has_arg = required_argument, |
79f19470 JA |
118 | .flag = NULL, |
119 | .val = 'b' | |
120 | }, | |
152f6476 | 121 | { |
3f65c585 | 122 | .name = "per-program-stats", |
428683db | 123 | .has_arg = no_argument, |
152f6476 JA |
124 | .flag = NULL, |
125 | .val = 's' | |
126 | }, | |
7997c5b0 | 127 | { |
3f65c585 | 128 | .name = "track-ios", |
428683db | 129 | .has_arg = no_argument, |
7997c5b0 JA |
130 | .flag = NULL, |
131 | .val = 't' | |
132 | }, | |
1e1c60f1 NS |
133 | { |
134 | .name = "quiet", | |
428683db | 135 | .has_arg = no_argument, |
1e1c60f1 NS |
136 | .flag = NULL, |
137 | .val = 'q' | |
138 | }, | |
46e6968b NS |
139 | { |
140 | .name = "stopwatch", | |
428683db | 141 | .has_arg = required_argument, |
46e6968b NS |
142 | .flag = NULL, |
143 | .val = 'w' | |
144 | }, | |
ab197ca7 AB |
145 | { |
146 | .name = "format", | |
428683db | 147 | .has_arg = required_argument, |
ab197ca7 AB |
148 | .flag = NULL, |
149 | .val = 'f' | |
150 | }, | |
151 | { | |
152 | .name = "format-spec", | |
428683db | 153 | .has_arg = required_argument, |
ab197ca7 AB |
154 | .flag = NULL, |
155 | .val = 'F' | |
156 | }, | |
bf0720af | 157 | { |
3f65c585 | 158 | .name = "hash-by-name", |
bf0720af JA |
159 | .has_arg = no_argument, |
160 | .flag = NULL, | |
d915dee6 | 161 | .val = 'h' |
bf0720af | 162 | }, |
7d1c0411 | 163 | { |
57ea8602 | 164 | .name = "verbose", |
7d1c0411 JA |
165 | .has_arg = no_argument, |
166 | .flag = NULL, | |
57ea8602 | 167 | .val = 'v' |
7d1c0411 | 168 | }, |
52724a0e JA |
169 | { |
170 | .name = "version", | |
171 | .has_arg = no_argument, | |
172 | .flag = NULL, | |
57ea8602 | 173 | .val = 'V' |
52724a0e | 174 | }, |
d1d7f15f | 175 | { |
3f65c585 | 176 | .name = "input-directory", |
d1d7f15f JA |
177 | .has_arg = required_argument, |
178 | .flag = NULL, | |
179 | .val = 'D' | |
180 | }, | |
71ef8b7c JA |
181 | { |
182 | .name = NULL, | |
183 | } | |
d5396421 JA |
184 | }; |
185 | ||
7997c5b0 JA |
186 | /* |
187 | * for sorting the displayed output | |
188 | */ | |
8fc0abbc JA |
189 | struct trace { |
190 | struct blk_io_trace *bit; | |
191 | struct rb_node rb_node; | |
cb2a1a62 | 192 | struct trace *next; |
8fc0abbc JA |
193 | }; |
194 | ||
cb2a1a62 | 195 | static struct rb_root rb_sort_root; |
a649216c JA |
196 | static unsigned long rb_sort_entries; |
197 | ||
cb2a1a62 JA |
198 | static struct trace *trace_list; |
199 | ||
d36421e4 JA |
200 | /* |
201 | * allocation cache | |
202 | */ | |
203 | static struct blk_io_trace *bit_alloc_list; | |
204 | static struct trace *t_alloc_list; | |
205 | ||
7997c5b0 JA |
206 | /* |
207 | * for tracking individual ios | |
208 | */ | |
209 | struct io_track { | |
210 | struct rb_node rb_node; | |
211 | ||
212 | __u64 sector; | |
213 | __u32 pid; | |
bf0720af | 214 | char comm[16]; |
95c15013 | 215 | unsigned long long allocation_time; |
7997c5b0 JA |
216 | unsigned long long queue_time; |
217 | unsigned long long dispatch_time; | |
218 | unsigned long long completion_time; | |
219 | }; | |
220 | ||
e7c9f3ff NS |
221 | static int ndevices; |
222 | static struct per_dev_info *devices; | |
223 | static char *get_dev_name(struct per_dev_info *, char *, int); | |
d0ca268b | 224 | |
71d5d4c9 | 225 | FILE *ofp = NULL; |
e7c9f3ff | 226 | static char *output_name; |
d1d7f15f | 227 | static char *input_dir; |
e7c9f3ff NS |
228 | |
229 | static unsigned long long genesis_time; | |
287fa3d6 | 230 | static unsigned long long last_allowed_time; |
1c7c54aa | 231 | static unsigned int smallest_seq_read; |
46e6968b | 232 | static unsigned long long stopwatch_start; /* start from zero by default */ |
bc171579 | 233 | static unsigned long long stopwatch_end = -1ULL; /* "infinity" */ |
152f6476 JA |
234 | |
235 | static int per_process_stats; | |
cbc927b6 | 236 | static int per_device_and_cpu_stats = 1; |
7997c5b0 | 237 | static int track_ios; |
bf0720af | 238 | static int ppi_hash_by_pid = 1; |
57ea8602 | 239 | static int verbose; |
98f8386b | 240 | static unsigned int act_mask = -1U; |
cbc927b6 | 241 | static int stats_printed; |
d0ca268b | 242 | |
1d24fc14 JA |
243 | static unsigned int t_alloc_cache; |
244 | static unsigned int bit_alloc_cache; | |
245 | ||
7d747d22 | 246 | #define RB_BATCH_DEFAULT (512) |
e820abd7 | 247 | static unsigned int rb_batch = RB_BATCH_DEFAULT; |
79f19470 | 248 | |
e7c9f3ff NS |
249 | static int pipeline; |
250 | ||
412819ce JA |
251 | #define is_done() (*(volatile int *)(&done)) |
252 | static volatile int done; | |
253 | ||
bf0720af JA |
254 | #define JHASH_RANDOM (0x3af5f2ee) |
255 | ||
824c2b39 JA |
256 | #define CPUS_PER_LONG (8 * sizeof(unsigned long)) |
257 | #define CPU_IDX(cpu) ((cpu) / CPUS_PER_LONG) | |
258 | #define CPU_BIT(cpu) ((cpu) & (CPUS_PER_LONG - 1)) | |
259 | ||
260 | static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu) | |
261 | { | |
262 | if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) { | |
263 | int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1); | |
264 | unsigned long *map = malloc(new_max / sizeof(long)); | |
265 | ||
266 | memset(map, 0, new_max / sizeof(long)); | |
267 | ||
268 | if (pdi->cpu_map) { | |
269 | memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long)); | |
270 | free(pdi->cpu_map); | |
271 | } | |
272 | ||
273 | pdi->cpu_map = map; | |
274 | pdi->cpu_map_max = new_max; | |
275 | } | |
276 | ||
277 | pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu)); | |
278 | } | |
279 | ||
280 | static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu) | |
281 | { | |
282 | pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu)); | |
283 | } | |
284 | ||
285 | static inline int cpu_is_online(struct per_dev_info *pdi, int cpu) | |
286 | { | |
287 | return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0; | |
288 | } | |
289 | ||
bf0720af JA |
290 | static inline int ppi_hash_pid(__u32 pid) |
291 | { | |
292 | return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK; | |
293 | } | |
294 | ||
295 | static inline int ppi_hash_name(const char *name) | |
152f6476 | 296 | { |
bf0720af JA |
297 | return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK; |
298 | } | |
299 | ||
300 | static inline int ppi_hash(struct per_process_info *ppi) | |
301 | { | |
302 | if (ppi_hash_by_pid) | |
303 | return ppi_hash_pid(ppi->pid); | |
304 | ||
bf0720af | 305 | return ppi_hash_name(ppi->name); |
152f6476 JA |
306 | } |
307 | ||
308 | static inline void add_process_to_hash(struct per_process_info *ppi) | |
309 | { | |
bf0720af | 310 | const int hash_idx = ppi_hash(ppi); |
152f6476 | 311 | |
bf0720af JA |
312 | ppi->hash_next = ppi_hash_table[hash_idx]; |
313 | ppi_hash_table[hash_idx] = ppi; | |
152f6476 JA |
314 | } |
315 | ||
316 | static inline void add_process_to_list(struct per_process_info *ppi) | |
317 | { | |
318 | ppi->list_next = ppi_list; | |
319 | ppi_list = ppi; | |
886ecf0e | 320 | ppi_list_entries++; |
152f6476 JA |
321 | } |
322 | ||
bf0720af JA |
323 | static struct per_process_info *find_process_by_name(char *name) |
324 | { | |
325 | const int hash_idx = ppi_hash_name(name); | |
326 | struct per_process_info *ppi; | |
327 | ||
328 | ppi = ppi_hash_table[hash_idx]; | |
329 | while (ppi) { | |
330 | if (!strcmp(ppi->name, name)) | |
331 | return ppi; | |
332 | ||
333 | ppi = ppi->hash_next; | |
334 | } | |
335 | ||
336 | return NULL; | |
337 | } | |
338 | ||
152f6476 JA |
339 | static struct per_process_info *find_process_by_pid(__u32 pid) |
340 | { | |
bf0720af | 341 | const int hash_idx = ppi_hash_pid(pid); |
152f6476 JA |
342 | struct per_process_info *ppi; |
343 | ||
bf0720af | 344 | ppi = ppi_hash_table[hash_idx]; |
152f6476 JA |
345 | while (ppi) { |
346 | if (ppi->pid == pid) | |
347 | return ppi; | |
348 | ||
349 | ppi = ppi->hash_next; | |
350 | } | |
351 | ||
352 | return NULL; | |
353 | } | |
354 | ||
bf0720af JA |
355 | static struct per_process_info *find_process(__u32 pid, char *name) |
356 | { | |
715d8021 JA |
357 | struct per_process_info *ppi; |
358 | ||
bf0720af JA |
359 | if (ppi_hash_by_pid) |
360 | return find_process_by_pid(pid); | |
361 | ||
715d8021 JA |
362 | ppi = find_process_by_name(name); |
363 | if (ppi && ppi->pid != pid) | |
364 | ppi->more_than_one = 1; | |
365 | ||
366 | return ppi; | |
bf0720af JA |
367 | } |
368 | ||
23211fdf JA |
369 | static inline int trace_rb_insert(struct trace *t, struct rb_root *root, |
370 | int check_time) | |
7997c5b0 | 371 | { |
2a1b3424 | 372 | struct rb_node **p = &root->rb_node; |
7997c5b0 JA |
373 | struct rb_node *parent = NULL; |
374 | struct trace *__t; | |
375 | ||
376 | while (*p) { | |
377 | parent = *p; | |
2a1b3424 | 378 | |
7997c5b0 JA |
379 | __t = rb_entry(parent, struct trace, rb_node); |
380 | ||
23211fdf JA |
381 | if (check_time) { |
382 | if (t->bit->time < __t->bit->time) { | |
383 | p = &(*p)->rb_left; | |
384 | continue; | |
385 | } else if (t->bit->time > __t->bit->time) { | |
386 | p = &(*p)->rb_right; | |
387 | continue; | |
388 | } | |
389 | } | |
390 | if (t->bit->device < __t->bit->device) | |
e7c9f3ff NS |
391 | p = &(*p)->rb_left; |
392 | else if (t->bit->device > __t->bit->device) | |
393 | p = &(*p)->rb_right; | |
dcf0f7ed JA |
394 | else if (t->bit->sequence < __t->bit->sequence) |
395 | p = &(*p)->rb_left; | |
0b07f23e | 396 | else /* >= sequence */ |
dcf0f7ed | 397 | p = &(*p)->rb_right; |
7997c5b0 JA |
398 | } |
399 | ||
400 | rb_link_node(&t->rb_node, parent, p); | |
2a1b3424 | 401 | rb_insert_color(&t->rb_node, root); |
7997c5b0 JA |
402 | return 0; |
403 | } | |
404 | ||
2a1b3424 | 405 | static inline int trace_rb_insert_sort(struct trace *t) |
e3556946 | 406 | { |
23211fdf | 407 | if (!trace_rb_insert(t, &rb_sort_root, 1)) { |
2a1b3424 JA |
408 | rb_sort_entries++; |
409 | return 0; | |
410 | } | |
411 | ||
412 | return 1; | |
413 | } | |
414 | ||
415 | static inline int trace_rb_insert_last(struct per_dev_info *pdi,struct trace *t) | |
416 | { | |
23211fdf | 417 | if (!trace_rb_insert(t, &pdi->rb_last, 1)) { |
2a1b3424 JA |
418 | pdi->rb_last_entries++; |
419 | return 0; | |
420 | } | |
421 | ||
422 | return 1; | |
423 | } | |
424 | ||
425 | static struct trace *trace_rb_find(dev_t device, unsigned long sequence, | |
426 | struct rb_root *root, int order) | |
427 | { | |
428 | struct rb_node *n = root->rb_node; | |
429 | struct rb_node *prev = NULL; | |
e3556946 JA |
430 | struct trace *__t; |
431 | ||
2a1b3424 JA |
432 | while (n) { |
433 | __t = rb_entry(n, struct trace, rb_node); | |
434 | prev = n; | |
e3556946 | 435 | |
0583b6a2 | 436 | if (device < __t->bit->device) |
2a1b3424 | 437 | n = n->rb_left; |
0583b6a2 | 438 | else if (device > __t->bit->device) |
2a1b3424 | 439 | n = n->rb_right; |
0583b6a2 | 440 | else if (sequence < __t->bit->sequence) |
2a1b3424 | 441 | n = n->rb_left; |
e3556946 | 442 | else if (sequence > __t->bit->sequence) |
2a1b3424 | 443 | n = n->rb_right; |
e3556946 JA |
444 | else |
445 | return __t; | |
446 | } | |
447 | ||
2a1b3424 JA |
448 | /* |
449 | * hack - the list may not be sequence ordered because some | |
450 | * events don't have sequence and time matched. so we end up | |
451 | * being a little off in the rb lookup here, because we don't | |
452 | * know the time we are looking for. compensate by browsing | |
453 | * a little ahead from the last entry to find the match | |
454 | */ | |
455 | if (order && prev) { | |
456 | int max = 5; | |
457 | ||
458 | while (((n = rb_next(prev)) != NULL) && max--) { | |
459 | __t = rb_entry(n, struct trace, rb_node); | |
460 | ||
461 | if (__t->bit->device == device && | |
462 | __t->bit->sequence == sequence) | |
463 | return __t; | |
464 | ||
465 | prev = n; | |
466 | } | |
467 | } | |
468 | ||
e3556946 JA |
469 | return NULL; |
470 | } | |
471 | ||
2a1b3424 JA |
472 | static inline struct trace *trace_rb_find_sort(dev_t dev, unsigned long seq) |
473 | { | |
474 | return trace_rb_find(dev, seq, &rb_sort_root, 1); | |
475 | } | |
476 | ||
477 | static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi, | |
478 | unsigned long seq) | |
479 | { | |
f7bd1a9b | 480 | return trace_rb_find(pdi->dev, seq, &pdi->rb_last, 0); |
2a1b3424 JA |
481 | } |
482 | ||
f7bd1a9b | 483 | static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot) |
7997c5b0 | 484 | { |
f7bd1a9b | 485 | struct rb_node **p = &pdi->rb_track.rb_node; |
7997c5b0 JA |
486 | struct rb_node *parent = NULL; |
487 | struct io_track *__iot; | |
488 | ||
489 | while (*p) { | |
490 | parent = *p; | |
7997c5b0 JA |
491 | __iot = rb_entry(parent, struct io_track, rb_node); |
492 | ||
f7bd1a9b | 493 | if (iot->sector < __iot->sector) |
7997c5b0 JA |
494 | p = &(*p)->rb_left; |
495 | else if (iot->sector > __iot->sector) | |
496 | p = &(*p)->rb_right; | |
497 | else { | |
e7c9f3ff | 498 | fprintf(stderr, |
ab197ca7 AB |
499 | "sector alias (%Lu) on device %d,%d!\n", |
500 | (unsigned long long) iot->sector, | |
f7bd1a9b | 501 | MAJOR(pdi->dev), MINOR(pdi->dev)); |
7997c5b0 JA |
502 | return 1; |
503 | } | |
504 | } | |
505 | ||
506 | rb_link_node(&iot->rb_node, parent, p); | |
f7bd1a9b | 507 | rb_insert_color(&iot->rb_node, &pdi->rb_track); |
7997c5b0 JA |
508 | return 0; |
509 | } | |
510 | ||
f7bd1a9b | 511 | static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector) |
7997c5b0 | 512 | { |
f7bd1a9b | 513 | struct rb_node *n = pdi->rb_track.rb_node; |
7997c5b0 JA |
514 | struct io_track *__iot; |
515 | ||
2a1b3424 JA |
516 | while (n) { |
517 | __iot = rb_entry(n, struct io_track, rb_node); | |
7997c5b0 | 518 | |
f7bd1a9b | 519 | if (sector < __iot->sector) |
2a1b3424 | 520 | n = n->rb_left; |
7997c5b0 | 521 | else if (sector > __iot->sector) |
2a1b3424 | 522 | n = n->rb_right; |
7997c5b0 JA |
523 | else |
524 | return __iot; | |
525 | } | |
526 | ||
527 | return NULL; | |
528 | } | |
529 | ||
f7bd1a9b JA |
530 | static struct io_track *find_track(struct per_dev_info *pdi, __u32 pid, |
531 | char *comm, __u64 sector) | |
7997c5b0 | 532 | { |
916b5501 | 533 | struct io_track *iot; |
7997c5b0 | 534 | |
f7bd1a9b | 535 | iot = __find_track(pdi, sector); |
7997c5b0 JA |
536 | if (!iot) { |
537 | iot = malloc(sizeof(*iot)); | |
50adc0ba | 538 | iot->pid = pid; |
bf0720af | 539 | memcpy(iot->comm, comm, sizeof(iot->comm)); |
7997c5b0 | 540 | iot->sector = sector; |
f7bd1a9b | 541 | track_rb_insert(pdi, iot); |
7997c5b0 JA |
542 | } |
543 | ||
544 | return iot; | |
545 | } | |
546 | ||
f7bd1a9b JA |
547 | static void log_track_frontmerge(struct per_dev_info *pdi, |
548 | struct blk_io_trace *t) | |
2e3e8ded JA |
549 | { |
550 | struct io_track *iot; | |
551 | ||
552 | if (!track_ios) | |
553 | return; | |
2e3e8ded | 554 | |
ae957cbc | 555 | iot = __find_track(pdi, t->sector + t_sec(t)); |
cb2a1a62 | 556 | if (!iot) { |
57ea8602 JA |
557 | if (verbose) |
558 | fprintf(stderr, "merge not found for (%d,%d): %llu\n", | |
559 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
560 | (unsigned long long) t->sector + t_sec(t)); | |
cb2a1a62 | 561 | return; |
2e3e8ded | 562 | } |
cb2a1a62 | 563 | |
f7bd1a9b | 564 | rb_erase(&iot->rb_node, &pdi->rb_track); |
ae957cbc | 565 | iot->sector -= t_sec(t); |
f7bd1a9b | 566 | track_rb_insert(pdi, iot); |
2e3e8ded JA |
567 | } |
568 | ||
f7bd1a9b | 569 | static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t) |
2e3e8ded JA |
570 | { |
571 | struct io_track *iot; | |
572 | ||
573 | if (!track_ios) | |
574 | return; | |
575 | ||
f7bd1a9b | 576 | iot = find_track(pdi, t->pid, t->comm, t->sector); |
95c15013 JA |
577 | iot->allocation_time = t->time; |
578 | } | |
579 | ||
95c15013 | 580 | /* |
b6076a9b | 581 | * return time between rq allocation and insertion |
95c15013 | 582 | */ |
f7bd1a9b JA |
583 | static unsigned long long log_track_insert(struct per_dev_info *pdi, |
584 | struct blk_io_trace *t) | |
95c15013 | 585 | { |
50adc0ba | 586 | unsigned long long elapsed; |
95c15013 JA |
587 | struct io_track *iot; |
588 | ||
589 | if (!track_ios) | |
590 | return -1; | |
591 | ||
f7bd1a9b | 592 | iot = find_track(pdi, t->pid, t->comm, t->sector); |
2e3e8ded | 593 | iot->queue_time = t->time; |
acd70d21 JA |
594 | |
595 | if (!iot->allocation_time) | |
596 | return -1; | |
597 | ||
50adc0ba JA |
598 | elapsed = iot->queue_time - iot->allocation_time; |
599 | ||
600 | if (per_process_stats) { | |
bf0720af | 601 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 602 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 603 | |
b9d40d6f JA |
604 | if (ppi && elapsed > ppi->longest_allocation_wait[w]) |
605 | ppi->longest_allocation_wait[w] = elapsed; | |
50adc0ba JA |
606 | } |
607 | ||
608 | return elapsed; | |
2e3e8ded JA |
609 | } |
610 | ||
611 | /* | |
612 | * return time between queue and issue | |
613 | */ | |
f7bd1a9b JA |
614 | static unsigned long long log_track_issue(struct per_dev_info *pdi, |
615 | struct blk_io_trace *t) | |
2e3e8ded | 616 | { |
50adc0ba | 617 | unsigned long long elapsed; |
2e3e8ded JA |
618 | struct io_track *iot; |
619 | ||
620 | if (!track_ios) | |
621 | return -1; | |
622 | if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0) | |
623 | return -1; | |
624 | ||
f7bd1a9b | 625 | iot = __find_track(pdi, t->sector); |
cb2a1a62 | 626 | if (!iot) { |
57ea8602 JA |
627 | if (verbose) |
628 | fprintf(stderr, "issue not found for (%d,%d): %llu\n", | |
629 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
630 | (unsigned long long) t->sector); | |
2e3e8ded | 631 | return -1; |
cb2a1a62 | 632 | } |
2e3e8ded JA |
633 | |
634 | iot->dispatch_time = t->time; | |
50adc0ba JA |
635 | elapsed = iot->dispatch_time - iot->queue_time; |
636 | ||
637 | if (per_process_stats) { | |
bf0720af | 638 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 639 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 640 | |
b9d40d6f JA |
641 | if (ppi && elapsed > ppi->longest_dispatch_wait[w]) |
642 | ppi->longest_dispatch_wait[w] = elapsed; | |
50adc0ba JA |
643 | } |
644 | ||
645 | return elapsed; | |
2e3e8ded JA |
646 | } |
647 | ||
648 | /* | |
649 | * return time between dispatch and complete | |
650 | */ | |
f7bd1a9b JA |
651 | static unsigned long long log_track_complete(struct per_dev_info *pdi, |
652 | struct blk_io_trace *t) | |
2e3e8ded JA |
653 | { |
654 | unsigned long long elapsed; | |
655 | struct io_track *iot; | |
656 | ||
657 | if (!track_ios) | |
658 | return -1; | |
659 | if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0) | |
660 | return -1; | |
661 | ||
f7bd1a9b | 662 | iot = __find_track(pdi, t->sector); |
cb2a1a62 | 663 | if (!iot) { |
57ea8602 JA |
664 | if (verbose) |
665 | fprintf(stderr,"complete not found for (%d,%d): %llu\n", | |
666 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
667 | (unsigned long long) t->sector); | |
2e3e8ded | 668 | return -1; |
cb2a1a62 | 669 | } |
2e3e8ded JA |
670 | |
671 | iot->completion_time = t->time; | |
672 | elapsed = iot->completion_time - iot->dispatch_time; | |
673 | ||
50adc0ba | 674 | if (per_process_stats) { |
bf0720af | 675 | struct per_process_info *ppi = find_process(iot->pid,iot->comm); |
b9d40d6f | 676 | int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
50adc0ba | 677 | |
b9d40d6f JA |
678 | if (ppi && elapsed > ppi->longest_completion_wait[w]) |
679 | ppi->longest_completion_wait[w] = elapsed; | |
50adc0ba JA |
680 | } |
681 | ||
2e3e8ded JA |
682 | /* |
683 | * kill the trace, we don't need it after completion | |
684 | */ | |
f7bd1a9b | 685 | rb_erase(&iot->rb_node, &pdi->rb_track); |
2e3e8ded JA |
686 | free(iot); |
687 | ||
688 | return elapsed; | |
689 | } | |
690 | ||
691 | ||
152f6476 JA |
692 | static struct io_stats *find_process_io_stats(__u32 pid, char *name) |
693 | { | |
bf0720af | 694 | struct per_process_info *ppi = find_process(pid, name); |
152f6476 JA |
695 | |
696 | if (!ppi) { | |
697 | ppi = malloc(sizeof(*ppi)); | |
698 | memset(ppi, 0, sizeof(*ppi)); | |
bf0720af | 699 | memcpy(ppi->name, name, 16); |
152f6476 JA |
700 | ppi->pid = pid; |
701 | add_process_to_hash(ppi); | |
702 | add_process_to_list(ppi); | |
703 | } | |
704 | ||
705 | return &ppi->io_stats; | |
706 | } | |
707 | ||
e7c9f3ff | 708 | static void resize_cpu_info(struct per_dev_info *pdi, int cpu) |
a718bd37 | 709 | { |
e7c9f3ff NS |
710 | struct per_cpu_info *cpus = pdi->cpus; |
711 | int ncpus = pdi->ncpus; | |
712 | int new_count = cpu + 1; | |
713 | int new_space, size; | |
a718bd37 NS |
714 | char *new_start; |
715 | ||
e7c9f3ff NS |
716 | size = new_count * sizeof(struct per_cpu_info); |
717 | cpus = realloc(cpus, size); | |
718 | if (!cpus) { | |
719 | char name[20]; | |
720 | fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n", | |
721 | get_dev_name(pdi, name, sizeof(name)), size); | |
a718bd37 NS |
722 | exit(1); |
723 | } | |
724 | ||
e7c9f3ff NS |
725 | new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info)); |
726 | new_space = (new_count - ncpus) * sizeof(struct per_cpu_info); | |
a718bd37 | 727 | memset(new_start, 0, new_space); |
e7c9f3ff NS |
728 | |
729 | pdi->ncpus = new_count; | |
730 | pdi->cpus = cpus; | |
731 | } | |
cb2a1a62 | 732 | |
e7c9f3ff NS |
733 | static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu) |
734 | { | |
cb2a1a62 JA |
735 | struct per_cpu_info *pci; |
736 | ||
e7c9f3ff NS |
737 | if (cpu >= pdi->ncpus) |
738 | resize_cpu_info(pdi, cpu); | |
cb2a1a62 JA |
739 | |
740 | pci = &pdi->cpus[cpu]; | |
741 | pci->cpu = cpu; | |
742 | return pci; | |
a718bd37 NS |
743 | } |
744 | ||
e7c9f3ff NS |
745 | |
746 | static int resize_devices(char *name) | |
a718bd37 | 747 | { |
e7c9f3ff | 748 | int size = (ndevices + 1) * sizeof(struct per_dev_info); |
c499bf38 | 749 | |
e7c9f3ff NS |
750 | devices = realloc(devices, size); |
751 | if (!devices) { | |
752 | fprintf(stderr, "Out of memory, device %s (%d)\n", name, size); | |
753 | return 1; | |
754 | } | |
755 | memset(&devices[ndevices], 0, sizeof(struct per_dev_info)); | |
756 | devices[ndevices].name = name; | |
757 | ndevices++; | |
758 | return 0; | |
759 | } | |
a718bd37 | 760 | |
f7bd1a9b | 761 | static struct per_dev_info *get_dev_info(dev_t dev) |
e7c9f3ff | 762 | { |
cb2a1a62 | 763 | struct per_dev_info *pdi; |
e7c9f3ff | 764 | int i; |
c499bf38 | 765 | |
7d747d22 | 766 | for (i = 0; i < ndevices; i++) { |
f7bd1a9b JA |
767 | if (!devices[i].dev) |
768 | devices[i].dev = dev; | |
769 | if (devices[i].dev == dev) | |
e7c9f3ff | 770 | return &devices[i]; |
7d747d22 | 771 | } |
cb2a1a62 | 772 | |
2a1b3424 | 773 | if (resize_devices(NULL)) |
e7c9f3ff | 774 | return NULL; |
cb2a1a62 JA |
775 | |
776 | pdi = &devices[ndevices - 1]; | |
f7bd1a9b | 777 | pdi->dev = dev; |
20ed6177 | 778 | pdi->first_reported_time = 0; |
d6222db8 | 779 | pdi->last_sequence = -1; |
287fa3d6 | 780 | pdi->last_read_time = 0; |
2a1b3424 JA |
781 | memset(&pdi->rb_last, 0, sizeof(pdi->rb_last)); |
782 | pdi->rb_last_entries = 0; | |
cb2a1a62 | 783 | return pdi; |
a718bd37 NS |
784 | } |
785 | ||
e7c9f3ff NS |
786 | static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size) |
787 | { | |
788 | if (pdi->name) | |
789 | snprintf(buffer, size, "%s", pdi->name); | |
790 | else | |
f7bd1a9b | 791 | snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev)); |
e7c9f3ff NS |
792 | return buffer; |
793 | } | |
794 | ||
e7c9f3ff | 795 | static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit) |
cfab07eb AB |
796 | { |
797 | unsigned long long this = bit->time; | |
e7c9f3ff | 798 | unsigned long long last = pdi->last_reported_time; |
cfab07eb | 799 | |
e7c9f3ff NS |
800 | pdi->backwards = (this < last) ? 'B' : ' '; |
801 | pdi->last_reported_time = this; | |
cfab07eb AB |
802 | } |
803 | ||
152f6476 JA |
804 | static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t, |
805 | int rw) | |
d0ca268b JA |
806 | { |
807 | if (rw) { | |
152f6476 | 808 | ios->mwrites++; |
ae957cbc | 809 | ios->qwrite_kb += t_kb(t); |
d0ca268b | 810 | } else { |
152f6476 | 811 | ios->mreads++; |
ae957cbc | 812 | ios->qread_kb += t_kb(t); |
152f6476 JA |
813 | } |
814 | } | |
815 | ||
816 | static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci, | |
817 | int rw) | |
818 | { | |
819 | __account_m(&pci->io_stats, t, rw); | |
820 | ||
821 | if (per_process_stats) { | |
822 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
823 | ||
824 | __account_m(ios, t, rw); | |
d0ca268b JA |
825 | } |
826 | } | |
827 | ||
b6076a9b JA |
828 | static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t, |
829 | int rw) | |
d0ca268b JA |
830 | { |
831 | if (rw) { | |
152f6476 | 832 | ios->qwrites++; |
ae957cbc | 833 | ios->qwrite_kb += t_kb(t); |
d0ca268b | 834 | } else { |
152f6476 | 835 | ios->qreads++; |
ae957cbc | 836 | ios->qread_kb += t_kb(t); |
152f6476 JA |
837 | } |
838 | } | |
839 | ||
b6076a9b JA |
840 | static inline void account_queue(struct blk_io_trace *t, |
841 | struct per_cpu_info *pci, int rw) | |
152f6476 | 842 | { |
b6076a9b | 843 | __account_queue(&pci->io_stats, t, rw); |
152f6476 JA |
844 | |
845 | if (per_process_stats) { | |
846 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
847 | ||
b6076a9b | 848 | __account_queue(ios, t, rw); |
d0ca268b JA |
849 | } |
850 | } | |
851 | ||
152f6476 | 852 | static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes) |
d0ca268b JA |
853 | { |
854 | if (rw) { | |
152f6476 JA |
855 | ios->cwrites++; |
856 | ios->cwrite_kb += bytes >> 10; | |
d0ca268b | 857 | } else { |
152f6476 JA |
858 | ios->creads++; |
859 | ios->cread_kb += bytes >> 10; | |
860 | } | |
861 | } | |
862 | ||
863 | static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci, | |
864 | int rw, int bytes) | |
865 | { | |
866 | __account_c(&pci->io_stats, rw, bytes); | |
867 | ||
868 | if (per_process_stats) { | |
869 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
870 | ||
871 | __account_c(ios, rw, bytes); | |
d0ca268b JA |
872 | } |
873 | } | |
874 | ||
b6076a9b JA |
875 | static inline void __account_issue(struct io_stats *ios, int rw, |
876 | unsigned int bytes) | |
afd2d7ad | 877 | { |
878 | if (rw) { | |
152f6476 JA |
879 | ios->iwrites++; |
880 | ios->iwrite_kb += bytes >> 10; | |
afd2d7ad | 881 | } else { |
152f6476 JA |
882 | ios->ireads++; |
883 | ios->iread_kb += bytes >> 10; | |
afd2d7ad | 884 | } |
885 | } | |
886 | ||
b6076a9b JA |
887 | static inline void account_issue(struct blk_io_trace *t, |
888 | struct per_cpu_info *pci, int rw) | |
d0ca268b | 889 | { |
b6076a9b | 890 | __account_issue(&pci->io_stats, rw, t->bytes); |
152f6476 JA |
891 | |
892 | if (per_process_stats) { | |
893 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
d5396421 | 894 | |
b6076a9b | 895 | __account_issue(ios, rw, t->bytes); |
152f6476 JA |
896 | } |
897 | } | |
898 | ||
06639b27 JA |
899 | static inline void __account_unplug(struct io_stats *ios, int timer) |
900 | { | |
901 | if (timer) | |
902 | ios->timer_unplugs++; | |
903 | else | |
904 | ios->io_unplugs++; | |
905 | } | |
906 | ||
907 | static inline void account_unplug(struct blk_io_trace *t, | |
908 | struct per_cpu_info *pci, int timer) | |
909 | { | |
910 | __account_unplug(&pci->io_stats, timer); | |
911 | ||
912 | if (per_process_stats) { | |
913 | struct io_stats *ios = find_process_io_stats(t->pid, t->comm); | |
914 | ||
915 | __account_unplug(ios, timer); | |
916 | } | |
917 | } | |
918 | ||
f7bd1a9b JA |
919 | static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci, |
920 | struct blk_io_trace *t, char *act) | |
ab197ca7 | 921 | { |
f7bd1a9b | 922 | process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL); |
ab197ca7 AB |
923 | } |
924 | ||
f7bd1a9b JA |
925 | static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci, |
926 | struct blk_io_trace *t, char *act) | |
b6076a9b | 927 | { |
f7bd1a9b | 928 | process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL); |
b6076a9b JA |
929 | } |
930 | ||
ab197ca7 AB |
931 | static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t, |
932 | char *act) | |
933 | { | |
b6076a9b | 934 | process_fmt(act, pci, t, -1, 0, NULL); |
ab197ca7 | 935 | } |
2e3e8ded | 936 | |
f7bd1a9b JA |
937 | static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci, |
938 | struct blk_io_trace *t, char *act) | |
ab197ca7 | 939 | { |
f7bd1a9b | 940 | process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL); |
d0ca268b JA |
941 | } |
942 | ||
f7bd1a9b JA |
943 | static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci, |
944 | struct blk_io_trace *t, char *act) | |
d0ca268b | 945 | { |
a01516de | 946 | if (act[0] == 'F') |
f7bd1a9b | 947 | log_track_frontmerge(pdi, t); |
2e3e8ded | 948 | |
ab197ca7 | 949 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
d0ca268b JA |
950 | } |
951 | ||
dfe34da1 | 952 | static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 953 | char *act) |
dfe34da1 | 954 | { |
ab197ca7 | 955 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
dfe34da1 JA |
956 | } |
957 | ||
d5396421 | 958 | static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 959 | char *act) |
d0ca268b | 960 | { |
ab197ca7 | 961 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
d0ca268b JA |
962 | } |
963 | ||
ab197ca7 | 964 | static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t, |
3639a11e | 965 | char *act) |
67e14fdc | 966 | { |
ab197ca7 | 967 | process_fmt(act, pci, t, -1ULL, 0, NULL); |
67e14fdc JA |
968 | } |
969 | ||
93f1c611 JA |
970 | static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t, |
971 | char *act) | |
972 | { | |
973 | process_fmt(act, pci, t, -1ULL, 0, NULL); | |
974 | } | |
975 | ||
ab197ca7 | 976 | static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act) |
d0ca268b | 977 | { |
ab197ca7 | 978 | unsigned char *buf = (unsigned char *) t + sizeof(*t); |
d0ca268b | 979 | |
ab197ca7 | 980 | process_fmt(act, pci, t, -1ULL, t->pdu_len, buf); |
d0ca268b JA |
981 | } |
982 | ||
ff3a732c | 983 | static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci) |
d0ca268b | 984 | { |
56f2af81 JA |
985 | int act = t->action & 0xffff; |
986 | ||
987 | switch (act) { | |
d0ca268b | 988 | case __BLK_TA_QUEUE: |
3639a11e | 989 | log_generic(pci, t, "Q"); |
d0ca268b JA |
990 | break; |
991 | case __BLK_TA_GETRQ: | |
3639a11e | 992 | log_generic(pci, t, "G"); |
d0ca268b JA |
993 | break; |
994 | case __BLK_TA_SLEEPRQ: | |
3639a11e | 995 | log_generic(pci, t, "S"); |
d0ca268b JA |
996 | break; |
997 | case __BLK_TA_REQUEUE: | |
3639a11e | 998 | log_generic(pci, t, "R"); |
d0ca268b JA |
999 | break; |
1000 | case __BLK_TA_ISSUE: | |
ab197ca7 | 1001 | log_pc(pci, t, "D"); |
d0ca268b JA |
1002 | break; |
1003 | case __BLK_TA_COMPLETE: | |
3639a11e | 1004 | log_pc(pci, t, "C"); |
d0ca268b | 1005 | break; |
56f2af81 JA |
1006 | case __BLK_TA_INSERT: |
1007 | log_pc(pci, t, "I"); | |
1008 | break; | |
d0ca268b | 1009 | default: |
56f2af81 | 1010 | fprintf(stderr, "Bad pc action %x\n", act); |
87b72777 | 1011 | break; |
d0ca268b | 1012 | } |
d0ca268b JA |
1013 | } |
1014 | ||
f7bd1a9b JA |
1015 | static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi, |
1016 | struct per_cpu_info *pci) | |
d0ca268b JA |
1017 | { |
1018 | int w = t->action & BLK_TC_ACT(BLK_TC_WRITE); | |
7997c5b0 | 1019 | int act = t->action & 0xffff; |
d0ca268b | 1020 | |
7997c5b0 | 1021 | switch (act) { |
d0ca268b | 1022 | case __BLK_TA_QUEUE: |
b6076a9b | 1023 | account_queue(t, pci, w); |
3639a11e | 1024 | log_queue(pci, t, "Q"); |
d0ca268b | 1025 | break; |
b6076a9b | 1026 | case __BLK_TA_INSERT: |
f7bd1a9b | 1027 | log_insert(pdi, pci, t, "I"); |
b6076a9b | 1028 | break; |
d0ca268b | 1029 | case __BLK_TA_BACKMERGE: |
152f6476 | 1030 | account_m(t, pci, w); |
f7bd1a9b | 1031 | log_merge(pdi, pci, t, "M"); |
d0ca268b JA |
1032 | break; |
1033 | case __BLK_TA_FRONTMERGE: | |
152f6476 | 1034 | account_m(t, pci, w); |
f7bd1a9b | 1035 | log_merge(pdi, pci, t, "F"); |
d0ca268b JA |
1036 | break; |
1037 | case __BLK_TA_GETRQ: | |
f7bd1a9b | 1038 | log_track_getrq(pdi, t); |
3639a11e | 1039 | log_generic(pci, t, "G"); |
d0ca268b JA |
1040 | break; |
1041 | case __BLK_TA_SLEEPRQ: | |
3639a11e | 1042 | log_generic(pci, t, "S"); |
d0ca268b JA |
1043 | break; |
1044 | case __BLK_TA_REQUEUE: | |
152f6476 | 1045 | account_c(t, pci, w, -t->bytes); |
3639a11e | 1046 | log_queue(pci, t, "R"); |
d0ca268b JA |
1047 | break; |
1048 | case __BLK_TA_ISSUE: | |
b6076a9b | 1049 | account_issue(t, pci, w); |
f7bd1a9b | 1050 | log_issue(pdi, pci, t, "D"); |
d0ca268b JA |
1051 | break; |
1052 | case __BLK_TA_COMPLETE: | |
152f6476 | 1053 | account_c(t, pci, w, t->bytes); |
f7bd1a9b | 1054 | log_complete(pdi, pci, t, "C"); |
d0ca268b | 1055 | break; |
88b1a526 | 1056 | case __BLK_TA_PLUG: |
3639a11e | 1057 | log_action(pci, t, "P"); |
88b1a526 | 1058 | break; |
3639a11e | 1059 | case __BLK_TA_UNPLUG_IO: |
06639b27 | 1060 | account_unplug(t, pci, 0); |
3639a11e JA |
1061 | log_unplug(pci, t, "U"); |
1062 | break; | |
1063 | case __BLK_TA_UNPLUG_TIMER: | |
06639b27 | 1064 | account_unplug(t, pci, 1); |
3639a11e | 1065 | log_unplug(pci, t, "UT"); |
88b1a526 | 1066 | break; |
93f1c611 JA |
1067 | case __BLK_TA_SPLIT: |
1068 | log_split(pci, t, "X"); | |
1069 | break; | |
1070 | case __BLK_TA_BOUNCE: | |
1071 | log_generic(pci, t, "B"); | |
1072 | break; | |
a8f30e64 JA |
1073 | case __BLK_TA_REMAP: |
1074 | log_generic(pci, t, "A"); | |
1075 | break; | |
d0ca268b JA |
1076 | default: |
1077 | fprintf(stderr, "Bad fs action %x\n", t->action); | |
1f79c4a0 | 1078 | break; |
d0ca268b | 1079 | } |
d0ca268b JA |
1080 | } |
1081 | ||
ff3a732c JA |
1082 | static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci, |
1083 | struct per_dev_info *pdi) | |
d0ca268b JA |
1084 | { |
1085 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) | |
ff3a732c | 1086 | dump_trace_pc(t, pci); |
d0ca268b | 1087 | else |
f7bd1a9b | 1088 | dump_trace_fs(t, pdi, pci); |
87b72777 | 1089 | |
20ed6177 JA |
1090 | if (!pdi->events) |
1091 | pdi->first_reported_time = t->time; | |
1092 | ||
e7c9f3ff | 1093 | pdi->events++; |
d0ca268b JA |
1094 | } |
1095 | ||
4c523165 JA |
1096 | /* |
1097 | * print in a proper way, not too small and not too big. if more than | |
1098 | * 1000,000K, turn into M and so on | |
1099 | */ | |
1100 | static char *size_cnv(char *dst, unsigned long long num, int in_kb) | |
1101 | { | |
da19e768 | 1102 | char suff[] = { '\0', 'K', 'M', 'G', 'P' }; |
0dc3602c | 1103 | unsigned int i = 0; |
4c523165 JA |
1104 | |
1105 | if (in_kb) | |
1106 | i++; | |
1107 | ||
0dc3602c | 1108 | while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) { |
4c523165 JA |
1109 | i++; |
1110 | num /= 1000; | |
1111 | } | |
1112 | ||
1113 | sprintf(dst, "%'8Lu%c", num, suff[i]); | |
1114 | return dst; | |
1115 | } | |
1116 | ||
152f6476 | 1117 | static void dump_io_stats(struct io_stats *ios, char *msg) |
5c017e4b | 1118 | { |
4c523165 JA |
1119 | static char x[256], y[256]; |
1120 | ||
152f6476 JA |
1121 | fprintf(ofp, "%s\n", msg); |
1122 | ||
4c523165 JA |
1123 | fprintf(ofp, " Reads Queued: %s, %siB\t", size_cnv(x, ios->qreads, 0), size_cnv(y, ios->qread_kb, 1)); |
1124 | fprintf(ofp, " Writes Queued: %s, %siB\n", size_cnv(x, ios->qwrites, 0), size_cnv(y, ios->qwrite_kb, 1)); | |
0a6b8fc4 | 1125 | |
4c523165 JA |
1126 | fprintf(ofp, " Read Dispatches: %s, %siB\t", size_cnv(x, ios->ireads, 0), size_cnv(y, ios->iread_kb, 1)); |
1127 | fprintf(ofp, " Write Dispatches: %s, %siB\n", size_cnv(x, ios->iwrites, 0), size_cnv(y, ios->iwrite_kb, 1)); | |
1128 | fprintf(ofp, " Reads Completed: %s, %siB\t", size_cnv(x, ios->creads, 0), size_cnv(y, ios->cread_kb, 1)); | |
1129 | fprintf(ofp, " Writes Completed: %s, %siB\n", size_cnv(x, ios->cwrites, 0), size_cnv(y, ios->cwrite_kb, 1)); | |
152f6476 | 1130 | fprintf(ofp, " Read Merges: %'8lu%8c\t", ios->mreads, ' '); |
152f6476 | 1131 | fprintf(ofp, " Write Merges: %'8lu\n", ios->mwrites); |
06639b27 JA |
1132 | fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' '); |
1133 | fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs); | |
5c017e4b JA |
1134 | } |
1135 | ||
50adc0ba JA |
1136 | static void dump_wait_stats(struct per_process_info *ppi) |
1137 | { | |
b9d40d6f JA |
1138 | unsigned long rawait = ppi->longest_allocation_wait[0] / 1000; |
1139 | unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000; | |
1140 | unsigned long rcwait = ppi->longest_completion_wait[0] / 1000; | |
1141 | unsigned long wawait = ppi->longest_allocation_wait[1] / 1000; | |
1142 | unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000; | |
1143 | unsigned long wcwait = ppi->longest_completion_wait[1] / 1000; | |
1144 | ||
1145 | fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' '); | |
1146 | fprintf(ofp, " Allocation wait: %'8lu\n", wawait); | |
1147 | fprintf(ofp, " Dispatch wait: %'8lu%8c\t", rdwait, ' '); | |
1148 | fprintf(ofp, " Dispatch wait: %'8lu\n", wdwait); | |
1149 | fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' '); | |
1150 | fprintf(ofp, " Completion wait: %'8lu\n", wcwait); | |
50adc0ba JA |
1151 | } |
1152 | ||
886ecf0e JA |
1153 | static int ppi_name_compare(const void *p1, const void *p2) |
1154 | { | |
1155 | struct per_process_info *ppi1 = *((struct per_process_info **) p1); | |
1156 | struct per_process_info *ppi2 = *((struct per_process_info **) p2); | |
1157 | int res; | |
1158 | ||
1159 | res = strverscmp(ppi1->name, ppi2->name); | |
1160 | if (!res) | |
06e6f286 | 1161 | res = ppi1->pid > ppi2->pid; |
886ecf0e JA |
1162 | |
1163 | return res; | |
1164 | } | |
1165 | ||
1166 | static void sort_process_list(void) | |
1167 | { | |
1168 | struct per_process_info **ppis; | |
1169 | struct per_process_info *ppi; | |
1170 | int i = 0; | |
1171 | ||
1172 | ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *)); | |
1173 | ||
1174 | ppi = ppi_list; | |
1175 | while (ppi) { | |
06e6f286 | 1176 | ppis[i++] = ppi; |
886ecf0e JA |
1177 | ppi = ppi->list_next; |
1178 | } | |
1179 | ||
06e6f286 | 1180 | qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare); |
886ecf0e JA |
1181 | |
1182 | i = ppi_list_entries - 1; | |
1183 | ppi_list = NULL; | |
1184 | while (i >= 0) { | |
1185 | ppi = ppis[i]; | |
1186 | ||
1187 | ppi->list_next = ppi_list; | |
1188 | ppi_list = ppi; | |
1189 | i--; | |
1190 | } | |
50c38702 JA |
1191 | |
1192 | free(ppis); | |
886ecf0e JA |
1193 | } |
1194 | ||
152f6476 JA |
1195 | static void show_process_stats(void) |
1196 | { | |
1197 | struct per_process_info *ppi; | |
1198 | ||
886ecf0e JA |
1199 | sort_process_list(); |
1200 | ||
152f6476 JA |
1201 | ppi = ppi_list; |
1202 | while (ppi) { | |
ce8b6b4f JA |
1203 | char name[64]; |
1204 | ||
715d8021 | 1205 | if (ppi->more_than_one) |
bf0720af | 1206 | sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid); |
715d8021 JA |
1207 | else |
1208 | sprintf(name, "%s (%u)", ppi->name, ppi->pid); | |
bf0720af | 1209 | |
ce8b6b4f | 1210 | dump_io_stats(&ppi->io_stats, name); |
50adc0ba | 1211 | dump_wait_stats(ppi); |
152f6476 JA |
1212 | ppi = ppi->list_next; |
1213 | } | |
1214 | ||
1215 | fprintf(ofp, "\n"); | |
1216 | } | |
1217 | ||
e7c9f3ff | 1218 | static void show_device_and_cpu_stats(void) |
d0ca268b | 1219 | { |
e7c9f3ff NS |
1220 | struct per_dev_info *pdi; |
1221 | struct per_cpu_info *pci; | |
1222 | struct io_stats total, *ios; | |
20ed6177 | 1223 | unsigned long long rrate, wrate, msec; |
e7c9f3ff NS |
1224 | int i, j, pci_events; |
1225 | char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3]; | |
1226 | char name[32]; | |
1227 | ||
1228 | for (pdi = devices, i = 0; i < ndevices; i++, pdi++) { | |
1229 | ||
1230 | memset(&total, 0, sizeof(total)); | |
1231 | pci_events = 0; | |
1232 | ||
1233 | if (i > 0) | |
1234 | fprintf(ofp, "\n"); | |
1235 | ||
1236 | for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) { | |
1237 | if (!pci->nelems) | |
1238 | continue; | |
1239 | ||
1240 | ios = &pci->io_stats; | |
1241 | total.qreads += ios->qreads; | |
1242 | total.qwrites += ios->qwrites; | |
1243 | total.creads += ios->creads; | |
1244 | total.cwrites += ios->cwrites; | |
1245 | total.mreads += ios->mreads; | |
1246 | total.mwrites += ios->mwrites; | |
1247 | total.ireads += ios->ireads; | |
1248 | total.iwrites += ios->iwrites; | |
1249 | total.qread_kb += ios->qread_kb; | |
1250 | total.qwrite_kb += ios->qwrite_kb; | |
1251 | total.cread_kb += ios->cread_kb; | |
1252 | total.cwrite_kb += ios->cwrite_kb; | |
1253 | total.iread_kb += ios->iread_kb; | |
1254 | total.iwrite_kb += ios->iwrite_kb; | |
06639b27 JA |
1255 | total.timer_unplugs += ios->timer_unplugs; |
1256 | total.io_unplugs += ios->io_unplugs; | |
e7c9f3ff NS |
1257 | |
1258 | snprintf(line, sizeof(line) - 1, "CPU%d (%s):", | |
1259 | j, get_dev_name(pdi, name, sizeof(name))); | |
1260 | dump_io_stats(ios, line); | |
1261 | pci_events++; | |
1262 | } | |
5c017e4b | 1263 | |
e7c9f3ff NS |
1264 | if (pci_events > 1) { |
1265 | fprintf(ofp, "\n"); | |
1266 | snprintf(line, sizeof(line) - 1, "Total (%s):", | |
1267 | get_dev_name(pdi, name, sizeof(name))); | |
1268 | dump_io_stats(&total, line); | |
1269 | } | |
d0ca268b | 1270 | |
20ed6177 | 1271 | wrate = rrate = 0; |
20ed6177 JA |
1272 | msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000; |
1273 | if (msec) { | |
1274 | rrate = 1000 * total.cread_kb / msec; | |
1275 | wrate = 1000 * total.cwrite_kb / msec; | |
1276 | } | |
1277 | ||
1278 | fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n", rrate, wrate); | |
1279 | fprintf(ofp, "Events (%s): %'Lu entries, %'lu skips\n", | |
0aa8a62d | 1280 | get_dev_name(pdi, line, sizeof(line)), pdi->events, |
dcf0f7ed | 1281 | pdi->skips); |
e7c9f3ff | 1282 | } |
d0ca268b JA |
1283 | } |
1284 | ||
d36421e4 JA |
1285 | /* |
1286 | * struct trace and blktrace allocation cache, we do potentially | |
1287 | * millions of mallocs for these structures while only using at most | |
1288 | * a few thousand at the time | |
1289 | */ | |
1290 | static inline void t_free(struct trace *t) | |
1291 | { | |
3c8413cb | 1292 | if (t_alloc_cache < 1024) { |
1d24fc14 JA |
1293 | t->next = t_alloc_list; |
1294 | t_alloc_list = t; | |
3c8413cb JA |
1295 | t_alloc_cache++; |
1296 | } else | |
1297 | free(t); | |
d36421e4 JA |
1298 | } |
1299 | ||
1300 | static inline struct trace *t_alloc(void) | |
1301 | { | |
1302 | struct trace *t = t_alloc_list; | |
1303 | ||
1304 | if (t) { | |
1305 | t_alloc_list = t->next; | |
3c8413cb | 1306 | t_alloc_cache--; |
d36421e4 JA |
1307 | return t; |
1308 | } | |
1309 | ||
1310 | return malloc(sizeof(*t)); | |
1311 | } | |
1312 | ||
1313 | static inline void bit_free(struct blk_io_trace *bit) | |
1314 | { | |
eb9bd4e9 | 1315 | if (bit_alloc_cache < 1024 && !bit->pdu_len) { |
1d24fc14 JA |
1316 | /* |
1317 | * abuse a 64-bit field for a next pointer for the free item | |
1318 | */ | |
1319 | bit->time = (__u64) (unsigned long) bit_alloc_list; | |
1320 | bit_alloc_list = (struct blk_io_trace *) bit; | |
3c8413cb JA |
1321 | bit_alloc_cache++; |
1322 | } else | |
1323 | free(bit); | |
d36421e4 JA |
1324 | } |
1325 | ||
1326 | static inline struct blk_io_trace *bit_alloc(void) | |
1327 | { | |
1328 | struct blk_io_trace *bit = bit_alloc_list; | |
1329 | ||
1330 | if (bit) { | |
0c39425c NS |
1331 | bit_alloc_list = (struct blk_io_trace *) (unsigned long) \ |
1332 | bit->time; | |
3c8413cb | 1333 | bit_alloc_cache--; |
d36421e4 JA |
1334 | return bit; |
1335 | } | |
1336 | ||
1337 | return malloc(sizeof(*bit)); | |
1338 | } | |
1339 | ||
4f0ae44f JA |
1340 | static void find_genesis(void) |
1341 | { | |
1342 | struct trace *t = trace_list; | |
1343 | ||
1344 | genesis_time = -1ULL; | |
1345 | while (t != NULL) { | |
1346 | if (t->bit->time < genesis_time) | |
1347 | genesis_time = t->bit->time; | |
1348 | ||
1349 | t = t->next; | |
1350 | } | |
1351 | } | |
1352 | ||
7f4d89e6 | 1353 | static inline int check_stopwatch(struct blk_io_trace *bit) |
4f0ae44f | 1354 | { |
7f4d89e6 JA |
1355 | if (bit->time < stopwatch_end && |
1356 | bit->time >= stopwatch_start) | |
4f0ae44f JA |
1357 | return 0; |
1358 | ||
1359 | return 1; | |
1360 | } | |
1361 | ||
53c68c88 JA |
1362 | /* |
1363 | * return youngest entry read | |
1364 | */ | |
1365 | static int sort_entries(unsigned long long *youngest) | |
4f0ae44f JA |
1366 | { |
1367 | struct trace *t; | |
4f0ae44f JA |
1368 | |
1369 | if (!genesis_time) | |
1370 | find_genesis(); | |
1371 | ||
d6222db8 | 1372 | *youngest = 0; |
4f0ae44f JA |
1373 | while ((t = trace_list) != NULL) { |
1374 | struct blk_io_trace *bit = t->bit; | |
1375 | ||
1376 | trace_list = t->next; | |
1377 | ||
7f4d89e6 | 1378 | bit->time -= genesis_time; |
4f0ae44f | 1379 | |
d6222db8 JA |
1380 | if (bit->time < *youngest || !*youngest) |
1381 | *youngest = bit->time; | |
1382 | ||
774a1a10 JA |
1383 | if (bit->sequence < smallest_seq_read) |
1384 | smallest_seq_read = bit->sequence; | |
1385 | ||
7f4d89e6 | 1386 | if (check_stopwatch(bit)) { |
4f0ae44f JA |
1387 | bit_free(bit); |
1388 | t_free(t); | |
1389 | continue; | |
1390 | } | |
1391 | ||
2a1b3424 | 1392 | if (trace_rb_insert_sort(t)) |
53c68c88 | 1393 | return -1; |
4f0ae44f JA |
1394 | } |
1395 | ||
53c68c88 | 1396 | return 0; |
4f0ae44f JA |
1397 | } |
1398 | ||
1c24add6 JA |
1399 | static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t) |
1400 | { | |
1401 | rb_erase(&t->rb_node, &pdi->rb_last); | |
1402 | pdi->rb_last_entries--; | |
1403 | ||
1404 | bit_free(t->bit); | |
1405 | t_free(t); | |
1406 | } | |
1407 | ||
1408 | static void put_trace(struct per_dev_info *pdi, struct trace *t) | |
2a1b3424 JA |
1409 | { |
1410 | rb_erase(&t->rb_node, &rb_sort_root); | |
1411 | rb_sort_entries--; | |
1412 | ||
1413 | trace_rb_insert_last(pdi, t); | |
1414 | ||
1c24add6 | 1415 | if (pdi->rb_last_entries > rb_batch * pdi->nfiles) { |
2a1b3424 JA |
1416 | struct rb_node *n = rb_first(&pdi->rb_last); |
1417 | ||
1418 | t = rb_entry(n, struct trace, rb_node); | |
1c24add6 | 1419 | __put_trace_last(pdi, t); |
2a1b3424 JA |
1420 | } |
1421 | } | |
1422 | ||
824c2b39 JA |
1423 | /* |
1424 | * to continue, we must have traces from all online cpus in the tree | |
1425 | */ | |
1426 | static int check_cpu_map(struct per_dev_info *pdi) | |
1427 | { | |
1428 | unsigned long *cpu_map; | |
1429 | struct rb_node *n; | |
1430 | struct trace *__t; | |
1431 | unsigned int i; | |
1432 | int ret, cpu; | |
1433 | ||
1434 | /* | |
1435 | * create a map of the cpus we have traces for | |
1436 | */ | |
1437 | cpu_map = malloc(pdi->cpu_map_max / sizeof(long)); | |
1438 | n = rb_first(&rb_sort_root); | |
1439 | while (n) { | |
1440 | __t = rb_entry(n, struct trace, rb_node); | |
1441 | cpu = __t->bit->cpu; | |
1442 | ||
1443 | cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu)); | |
1444 | n = rb_next(n); | |
1445 | } | |
1446 | ||
1447 | /* | |
b1c8e614 JA |
1448 | * we can't continue if pdi->cpu_map has entries set that we don't |
1449 | * have in the sort rbtree. the opposite is not a problem, though | |
824c2b39 JA |
1450 | */ |
1451 | ret = 0; | |
1452 | for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) { | |
1453 | if (pdi->cpu_map[i] & ~(cpu_map[i])) { | |
1454 | ret = 1; | |
1455 | break; | |
1456 | } | |
1457 | } | |
1458 | ||
1459 | free(cpu_map); | |
1460 | return ret; | |
1461 | } | |
1462 | ||
a141a7cd | 1463 | static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force) |
2a1b3424 JA |
1464 | { |
1465 | unsigned long expected_sequence = pdi->last_sequence + 1; | |
1ca323a5 JA |
1466 | struct blk_io_trace *bit = t->bit; |
1467 | struct trace *__t; | |
d6222db8 | 1468 | |
774a1a10 | 1469 | if (!expected_sequence) { |
774a1a10 JA |
1470 | /* |
1471 | * 1 should be the first entry, just allow it | |
1472 | */ | |
1473 | if (bit->sequence == 1) | |
1474 | return 0; | |
1475 | ||
824c2b39 | 1476 | return check_cpu_map(pdi); |
774a1a10 | 1477 | } |
2a1b3424 JA |
1478 | |
1479 | if (bit->sequence == expected_sequence) | |
1480 | return 0; | |
1481 | ||
2a1b3424 | 1482 | /* |
1c7c54aa JA |
1483 | * we may not have seen that sequence yet. if we are not doing |
1484 | * the final run, break and wait for more entries. | |
1c24add6 | 1485 | */ |
1c7c54aa | 1486 | if (expected_sequence < smallest_seq_read) { |
1ca323a5 JA |
1487 | __t = trace_rb_find_last(pdi, expected_sequence); |
1488 | if (!__t) | |
1c7c54aa | 1489 | goto skip; |
2a1b3424 | 1490 | |
1ca323a5 | 1491 | __put_trace_last(pdi, __t); |
2a1b3424 | 1492 | return 0; |
a141a7cd JA |
1493 | } else if (!force) { |
1494 | return 1; | |
0b07f23e | 1495 | } else { |
1c7c54aa | 1496 | skip: |
57ea8602 | 1497 | if (verbose) { |
7d1c0411 JA |
1498 | fprintf(stderr, "(%d,%d): skipping %lu -> %u\n", |
1499 | MAJOR(pdi->dev), MINOR(pdi->dev), | |
1500 | pdi->last_sequence, bit->sequence); | |
1501 | } | |
1c7c54aa JA |
1502 | pdi->skips++; |
1503 | return 0; | |
1504 | } | |
2a1b3424 JA |
1505 | } |
1506 | ||
a649216c | 1507 | static void show_entries_rb(int force) |
8fc0abbc | 1508 | { |
1f7afa72 JA |
1509 | struct per_dev_info *pdi = NULL; |
1510 | struct per_cpu_info *pci = NULL; | |
8fc0abbc | 1511 | struct blk_io_trace *bit; |
3aabcd89 | 1512 | struct rb_node *n; |
8fc0abbc | 1513 | struct trace *t; |
1f7afa72 | 1514 | |
7d747d22 | 1515 | while ((n = rb_first(&rb_sort_root)) != NULL) { |
dd90748f | 1516 | if (is_done() && !force && !pipeline) |
1f7afa72 | 1517 | break; |
8fc0abbc JA |
1518 | |
1519 | t = rb_entry(n, struct trace, rb_node); | |
1520 | bit = t->bit; | |
1521 | ||
f7bd1a9b | 1522 | if (!pdi || pdi->dev != bit->device) |
287fa3d6 | 1523 | pdi = get_dev_info(bit->device); |
1f7afa72 | 1524 | |
e7c9f3ff NS |
1525 | if (!pdi) { |
1526 | fprintf(stderr, "Unknown device ID? (%d,%d)\n", | |
1527 | MAJOR(bit->device), MINOR(bit->device)); | |
1528 | break; | |
1529 | } | |
1f7afa72 | 1530 | |
a141a7cd JA |
1531 | if (check_sequence(pdi, t, force)) |
1532 | break; | |
cb2a1a62 | 1533 | |
a141a7cd JA |
1534 | if (!force && bit->time > last_allowed_time) |
1535 | break; | |
8fc0abbc | 1536 | |
be925321 JA |
1537 | pdi->last_sequence = bit->sequence; |
1538 | ||
4f0ae44f | 1539 | check_time(pdi, bit); |
8fc0abbc | 1540 | |
4f0ae44f JA |
1541 | if (!pci || pci->cpu != bit->cpu) |
1542 | pci = get_cpu_info(pdi, bit->cpu); | |
287fa3d6 | 1543 | |
cbc927b6 JA |
1544 | pci->nelems++; |
1545 | ||
98f8386b AB |
1546 | if (bit->action & (act_mask << BLK_TC_SHIFT)) |
1547 | dump_trace(bit, pci, pdi); | |
87b72777 | 1548 | |
2a1b3424 | 1549 | put_trace(pdi, t); |
cb2a1a62 | 1550 | } |
8fc0abbc JA |
1551 | } |
1552 | ||
1f79c4a0 JA |
1553 | static int read_data(int fd, void *buffer, int bytes, int block) |
1554 | { | |
1555 | int ret, bytes_left, fl; | |
1556 | void *p; | |
1557 | ||
1558 | fl = fcntl(fd, F_GETFL); | |
1559 | ||
1560 | if (!block) | |
1561 | fcntl(fd, F_SETFL, fl | O_NONBLOCK); | |
1562 | else | |
1563 | fcntl(fd, F_SETFL, fl & ~O_NONBLOCK); | |
1564 | ||
1565 | bytes_left = bytes; | |
1566 | p = buffer; | |
1567 | while (bytes_left > 0) { | |
1568 | ret = read(fd, p, bytes_left); | |
1569 | if (!ret) | |
1570 | return 1; | |
1571 | else if (ret < 0) { | |
1572 | if (errno != EAGAIN) | |
1573 | perror("read"); | |
a649216c | 1574 | |
1f79c4a0 JA |
1575 | return -1; |
1576 | } else { | |
1577 | p += ret; | |
1578 | bytes_left -= ret; | |
1579 | } | |
1580 | } | |
1581 | ||
1582 | return 0; | |
1583 | } | |
1584 | ||
a649216c | 1585 | static int read_events(int fd, int always_block) |
cb2a1a62 | 1586 | { |
287fa3d6 | 1587 | struct per_dev_info *pdi = NULL; |
e820abd7 | 1588 | unsigned int events = 0; |
7d747d22 JA |
1589 | |
1590 | while (!is_done() && events < rb_batch) { | |
1591 | struct blk_io_trace *bit; | |
1592 | struct trace *t; | |
1593 | int pdu_len; | |
1594 | __u32 magic; | |
1595 | ||
d36421e4 | 1596 | bit = bit_alloc(); |
cb2a1a62 | 1597 | |
eb9bd4e9 JA |
1598 | if (read_data(fd, bit, sizeof(*bit), !events || always_block)) { |
1599 | bit_free(bit); | |
cb2a1a62 | 1600 | break; |
eb9bd4e9 | 1601 | } |
cb2a1a62 | 1602 | |
7d747d22 JA |
1603 | magic = be32_to_cpu(bit->magic); |
1604 | if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { | |
1605 | fprintf(stderr, "Bad magic %x\n", magic); | |
1606 | break; | |
1607 | } | |
1608 | ||
1609 | pdu_len = be16_to_cpu(bit->pdu_len); | |
1610 | if (pdu_len) { | |
1611 | void *ptr = realloc(bit, sizeof(*bit) + pdu_len); | |
1612 | ||
eb9bd4e9 JA |
1613 | if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1)) { |
1614 | bit_free(ptr); | |
7d747d22 | 1615 | break; |
eb9bd4e9 | 1616 | } |
7d747d22 JA |
1617 | |
1618 | bit = ptr; | |
1619 | } | |
1620 | ||
d6222db8 JA |
1621 | trace_to_cpu(bit); |
1622 | ||
1623 | if (verify_trace(bit)) { | |
1624 | bit_free(bit); | |
1625 | continue; | |
1626 | } | |
1627 | ||
d36421e4 | 1628 | t = t_alloc(); |
cb2a1a62 JA |
1629 | memset(t, 0, sizeof(*t)); |
1630 | t->bit = bit; | |
1631 | ||
7d747d22 JA |
1632 | t->next = trace_list; |
1633 | trace_list = t; | |
1f7afa72 | 1634 | |
f7bd1a9b | 1635 | if (!pdi || pdi->dev != bit->device) |
287fa3d6 JA |
1636 | pdi = get_dev_info(bit->device); |
1637 | ||
1638 | if (bit->time > pdi->last_read_time) | |
1639 | pdi->last_read_time = bit->time; | |
1640 | ||
7d747d22 | 1641 | events++; |
cb2a1a62 JA |
1642 | } |
1643 | ||
7d747d22 | 1644 | return events; |
cb2a1a62 JA |
1645 | } |
1646 | ||
d5396421 | 1647 | static int do_file(void) |
d0ca268b | 1648 | { |
7d747d22 | 1649 | struct per_cpu_info *pci; |
73877e12 NS |
1650 | struct per_dev_info *pdi; |
1651 | int i, j, events, events_added; | |
d0ca268b | 1652 | |
7d747d22 JA |
1653 | /* |
1654 | * first prepare all files for reading | |
1655 | */ | |
e8741a4a | 1656 | for (i = 0; i < ndevices; i++) { |
73877e12 NS |
1657 | pdi = &devices[i]; |
1658 | pdi->nfiles = 0; | |
d6222db8 | 1659 | pdi->last_sequence = -1; |
73877e12 | 1660 | |
74feaeb5 | 1661 | for (j = 0;; j++) { |
e7c9f3ff | 1662 | struct stat st; |
d1d7f15f | 1663 | int len = 0; |
6e0073ed | 1664 | char *p, *dname; |
87b72777 | 1665 | |
e7c9f3ff NS |
1666 | pci = get_cpu_info(pdi, j); |
1667 | pci->cpu = j; | |
7d747d22 | 1668 | pci->fd = -1; |
6e0073ed JA |
1669 | |
1670 | p = strdup(pdi->name); | |
1671 | dname = dirname(p); | |
1672 | if (strcmp(dname, ".")) { | |
1673 | input_dir = dname; | |
1674 | p = strdup(pdi->name); | |
1675 | strcpy(pdi->name, basename(p)); | |
1676 | } | |
1677 | free(p); | |
d0ca268b | 1678 | |
d1d7f15f JA |
1679 | if (input_dir) |
1680 | len = sprintf(pci->fname, "%s/", input_dir); | |
1681 | ||
1682 | snprintf(pci->fname + len, sizeof(pci->fname)-1-len, | |
7d747d22 | 1683 | "%s.blktrace.%d", pdi->name, pci->cpu); |
e7c9f3ff NS |
1684 | if (stat(pci->fname, &st) < 0) |
1685 | break; | |
74feaeb5 AB |
1686 | if (st.st_size) { |
1687 | pci->fd = open(pci->fname, O_RDONLY); | |
1688 | if (pci->fd < 0) { | |
1689 | perror(pci->fname); | |
1690 | continue; | |
1691 | } | |
e7c9f3ff NS |
1692 | } |
1693 | ||
7d747d22 | 1694 | printf("Input file %s added\n", pci->fname); |
73877e12 | 1695 | pdi->nfiles++; |
824c2b39 | 1696 | cpu_mark_online(pdi, pci->cpu); |
d0ca268b | 1697 | } |
d5396421 JA |
1698 | } |
1699 | ||
7d747d22 JA |
1700 | /* |
1701 | * now loop over the files reading in the data | |
1702 | */ | |
412819ce | 1703 | do { |
53c68c88 JA |
1704 | unsigned long long youngest; |
1705 | ||
7d747d22 | 1706 | events_added = 0; |
287fa3d6 | 1707 | last_allowed_time = -1ULL; |
1c7c54aa | 1708 | smallest_seq_read = -1U; |
d5396421 | 1709 | |
7d747d22 | 1710 | for (i = 0; i < ndevices; i++) { |
73877e12 NS |
1711 | pdi = &devices[i]; |
1712 | ||
1713 | for (j = 0; j < pdi->nfiles; j++) { | |
d5396421 | 1714 | |
73877e12 | 1715 | pci = get_cpu_info(pdi, j); |
d5396421 | 1716 | |
7d747d22 JA |
1717 | if (pci->fd == -1) |
1718 | continue; | |
51128a28 | 1719 | |
a649216c | 1720 | events = read_events(pci->fd, 1); |
7d747d22 | 1721 | if (!events) { |
824c2b39 | 1722 | cpu_mark_offline(pdi, pci->cpu); |
7d747d22 JA |
1723 | close(pci->fd); |
1724 | pci->fd = -1; | |
1725 | continue; | |
1726 | } | |
d5396421 | 1727 | |
287fa3d6 JA |
1728 | if (pdi->last_read_time < last_allowed_time) |
1729 | last_allowed_time = pdi->last_read_time; | |
d5396421 | 1730 | |
7d747d22 JA |
1731 | events_added += events; |
1732 | } | |
2ff323b0 | 1733 | } |
d5396421 | 1734 | |
53c68c88 JA |
1735 | if (sort_entries(&youngest)) |
1736 | break; | |
1737 | ||
1738 | if (youngest > stopwatch_end) | |
287fa3d6 JA |
1739 | break; |
1740 | ||
a649216c | 1741 | show_entries_rb(0); |
cb2a1a62 | 1742 | |
7d747d22 | 1743 | } while (events_added); |
d5396421 | 1744 | |
a649216c JA |
1745 | if (rb_sort_entries) |
1746 | show_entries_rb(1); | |
1747 | ||
7d747d22 | 1748 | return 0; |
412819ce | 1749 | } |
d5396421 | 1750 | |
412819ce JA |
1751 | static int do_stdin(void) |
1752 | { | |
53c68c88 | 1753 | unsigned long long youngest; |
763d936e | 1754 | int fd, events; |
d5396421 | 1755 | |
be925321 | 1756 | last_allowed_time = -1ULL; |
1f79c4a0 | 1757 | fd = dup(STDIN_FILENO); |
0b07f23e JA |
1758 | if (fd == -1) { |
1759 | perror("dup stdin"); | |
1760 | return -1; | |
1761 | } | |
d5396421 | 1762 | |
0b07f23e | 1763 | while ((events = read_events(fd, 0)) != 0) { |
412819ce | 1764 | |
0b07f23e JA |
1765 | smallest_seq_read = -1U; |
1766 | ||
53c68c88 JA |
1767 | if (sort_entries(&youngest)) |
1768 | break; | |
1769 | ||
1770 | if (youngest > stopwatch_end) | |
2ff323b0 JA |
1771 | break; |
1772 | ||
763d936e | 1773 | show_entries_rb(0); |
0b07f23e | 1774 | } |
d5396421 | 1775 | |
a649216c JA |
1776 | if (rb_sort_entries) |
1777 | show_entries_rb(1); | |
1778 | ||
d5396421 | 1779 | close(fd); |
d5396421 JA |
1780 | return 0; |
1781 | } | |
d0ca268b | 1782 | |
cbc927b6 | 1783 | static void show_stats(void) |
412819ce | 1784 | { |
cbc927b6 JA |
1785 | if (!ofp) |
1786 | return; | |
1787 | if (stats_printed) | |
1788 | return; | |
1789 | ||
1790 | stats_printed = 1; | |
1791 | ||
1792 | if (per_process_stats) | |
1793 | show_process_stats(); | |
1794 | ||
1795 | if (per_device_and_cpu_stats) | |
1796 | show_device_and_cpu_stats(); | |
1797 | ||
152f6476 | 1798 | fflush(ofp); |
412819ce JA |
1799 | } |
1800 | ||
e820abd7 | 1801 | static void handle_sigint(__attribute__((__unused__)) int sig) |
412819ce JA |
1802 | { |
1803 | done = 1; | |
cbc927b6 | 1804 | show_stats(); |
412819ce JA |
1805 | } |
1806 | ||
46e6968b NS |
1807 | /* |
1808 | * Extract start and duration times from a string, allowing | |
1809 | * us to specify a time interval of interest within a trace. | |
1810 | * Format: "duration" (start is zero) or "start:duration". | |
1811 | */ | |
1812 | static int find_stopwatch_interval(char *string) | |
1813 | { | |
1814 | double value; | |
1815 | char *sp; | |
1816 | ||
1817 | value = strtod(string, &sp); | |
1818 | if (sp == string) { | |
1819 | fprintf(stderr,"Invalid stopwatch timer: %s\n", string); | |
1820 | return 1; | |
1821 | } | |
1822 | if (*sp == ':') { | |
1823 | stopwatch_start = DOUBLE_TO_NANO_ULL(value); | |
1824 | string = sp + 1; | |
1825 | value = strtod(string, &sp); | |
1826 | if (sp == string || *sp != '\0') { | |
1827 | fprintf(stderr,"Invalid stopwatch duration time: %s\n", | |
1828 | string); | |
1829 | return 1; | |
1830 | } | |
1831 | } else if (*sp != '\0') { | |
1832 | fprintf(stderr,"Invalid stopwatch start timer: %s\n", string); | |
1833 | return 1; | |
1834 | } | |
1b928247 JA |
1835 | stopwatch_end = DOUBLE_TO_NANO_ULL(value); |
1836 | if (stopwatch_end <= stopwatch_start) { | |
1837 | fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n", | |
1838 | stopwatch_start, stopwatch_end); | |
1839 | return 1; | |
1840 | } | |
1841 | ||
46e6968b NS |
1842 | return 0; |
1843 | } | |
1844 | ||
52724a0e JA |
1845 | static char usage_str[] = \ |
1846 | "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \ | |
1847 | "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \ | |
1848 | "\t-i Input file containing trace data, or '-' for stdin\n" \ | |
d1d7f15f | 1849 | "\t-D Directory to prepend to input file names\n" \ |
52724a0e JA |
1850 | "\t-o Output file. If not given, output is stdout\n" \ |
1851 | "\t-b stdin read batching\n" \ | |
1852 | "\t-s Show per-program io statistics\n" \ | |
bf0720af | 1853 | "\t-n Hash processes by name, not pid\n" \ |
52724a0e JA |
1854 | "\t-t Track individual ios. Will tell you the time a request took\n" \ |
1855 | "\t to get queued, to get dispatched, and to get completed\n" \ | |
1856 | "\t-q Quiet. Don't display any stats at the end of the trace\n" \ | |
1857 | "\t-w Only parse data between the given time interval in seconds.\n" \ | |
1858 | "\t If 'start' isn't given, blkparse defaults the start time to 0\n" \ | |
d915dee6 JA |
1859 | "\t-f Output format. Customize the output format. The format field\n" \ |
1860 | "\t identifies can be found in the documentation\n" \ | |
52724a0e | 1861 | "\t-F Format specification. Can be found in the documentation\n" \ |
57ea8602 JA |
1862 | "\t-v More verbose for marginal errors\n" \ |
1863 | "\t-V Print program version info\n\n"; | |
52724a0e | 1864 | |
1f79c4a0 JA |
1865 | static void usage(char *prog) |
1866 | { | |
52724a0e | 1867 | fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str); |
1f79c4a0 JA |
1868 | } |
1869 | ||
d5396421 JA |
1870 | int main(int argc, char *argv[]) |
1871 | { | |
152f6476 | 1872 | char *ofp_buffer; |
98f8386b | 1873 | int i, c, ret, mode; |
98f8386b | 1874 | int act_mask_tmp = 0; |
d5396421 JA |
1875 | |
1876 | while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) { | |
1877 | switch (c) { | |
98f8386b AB |
1878 | case 'a': |
1879 | i = find_mask_map(optarg); | |
1880 | if (i < 0) { | |
1881 | fprintf(stderr,"Invalid action mask %s\n", | |
1882 | optarg); | |
1883 | return 1; | |
1884 | } | |
1885 | act_mask_tmp |= i; | |
1886 | break; | |
1887 | ||
1888 | case 'A': | |
1889 | if ((sscanf(optarg, "%x", &i) != 1) || | |
1890 | !valid_act_opt(i)) { | |
1891 | fprintf(stderr, | |
1892 | "Invalid set action mask %s/0x%x\n", | |
1893 | optarg, i); | |
1894 | return 1; | |
1895 | } | |
1896 | act_mask_tmp = i; | |
1897 | break; | |
d5396421 | 1898 | case 'i': |
e7c9f3ff NS |
1899 | if (!strcmp(optarg, "-") && !pipeline) |
1900 | pipeline = 1; | |
1901 | else if (resize_devices(optarg) != 0) | |
1902 | return 1; | |
d5396421 | 1903 | break; |
d1d7f15f JA |
1904 | case 'D': |
1905 | input_dir = optarg; | |
1906 | break; | |
d5396421 | 1907 | case 'o': |
66efebf8 | 1908 | output_name = optarg; |
d5396421 | 1909 | break; |
79f19470 JA |
1910 | case 'b': |
1911 | rb_batch = atoi(optarg); | |
1912 | if (rb_batch <= 0) | |
1913 | rb_batch = RB_BATCH_DEFAULT; | |
1914 | break; | |
152f6476 JA |
1915 | case 's': |
1916 | per_process_stats = 1; | |
1917 | break; | |
7997c5b0 JA |
1918 | case 't': |
1919 | track_ios = 1; | |
1920 | break; | |
1e1c60f1 NS |
1921 | case 'q': |
1922 | per_device_and_cpu_stats = 0; | |
1923 | break; | |
46e6968b NS |
1924 | case 'w': |
1925 | if (find_stopwatch_interval(optarg) != 0) | |
1926 | return 1; | |
1927 | break; | |
ab197ca7 AB |
1928 | case 'f': |
1929 | set_all_format_specs(optarg); | |
1930 | break; | |
1931 | case 'F': | |
1932 | if (add_format_spec(optarg) != 0) | |
1933 | return 1; | |
1934 | break; | |
d915dee6 | 1935 | case 'h': |
715d8021 | 1936 | ppi_hash_by_pid = 0; |
bf0720af | 1937 | break; |
52724a0e | 1938 | case 'v': |
57ea8602 JA |
1939 | verbose++; |
1940 | break; | |
1941 | case 'V': | |
52724a0e JA |
1942 | printf("%s version %s\n", argv[0], blkparse_version); |
1943 | return 0; | |
d5396421 | 1944 | default: |
1f79c4a0 | 1945 | usage(argv[0]); |
d5396421 JA |
1946 | return 1; |
1947 | } | |
d0ca268b JA |
1948 | } |
1949 | ||
e7c9f3ff NS |
1950 | while (optind < argc) { |
1951 | if (!strcmp(argv[optind], "-") && !pipeline) | |
1952 | pipeline = 1; | |
1953 | else if (resize_devices(argv[optind]) != 0) | |
1954 | return 1; | |
1955 | optind++; | |
1956 | } | |
1957 | ||
1958 | if (!pipeline && !ndevices) { | |
1f79c4a0 | 1959 | usage(argv[0]); |
d5396421 JA |
1960 | return 1; |
1961 | } | |
1962 | ||
98f8386b AB |
1963 | if (act_mask_tmp != 0) |
1964 | act_mask = act_mask_tmp; | |
1965 | ||
7997c5b0 | 1966 | memset(&rb_sort_root, 0, sizeof(rb_sort_root)); |
412819ce JA |
1967 | |
1968 | signal(SIGINT, handle_sigint); | |
1969 | signal(SIGHUP, handle_sigint); | |
1970 | signal(SIGTERM, handle_sigint); | |
d5396421 | 1971 | |
d69db225 JA |
1972 | setlocale(LC_NUMERIC, "en_US"); |
1973 | ||
a66877e6 | 1974 | if (!output_name) { |
152f6476 | 1975 | ofp = fdopen(STDOUT_FILENO, "w"); |
a66877e6 JA |
1976 | mode = _IOLBF; |
1977 | } else { | |
152f6476 JA |
1978 | char ofname[128]; |
1979 | ||
7a9690c0 | 1980 | snprintf(ofname, sizeof(ofname) - 1, "%s", output_name); |
152f6476 | 1981 | ofp = fopen(ofname, "w"); |
a66877e6 | 1982 | mode = _IOFBF; |
152f6476 JA |
1983 | } |
1984 | ||
1985 | if (!ofp) { | |
1986 | perror("fopen"); | |
1987 | return 1; | |
1988 | } | |
1989 | ||
1990 | ofp_buffer = malloc(4096); | |
a66877e6 | 1991 | if (setvbuf(ofp, ofp_buffer, mode, 4096)) { |
152f6476 JA |
1992 | perror("setvbuf"); |
1993 | return 1; | |
1994 | } | |
1995 | ||
e7c9f3ff | 1996 | if (pipeline) |
d5396421 JA |
1997 | ret = do_stdin(); |
1998 | else | |
1999 | ret = do_file(); | |
2000 | ||
cbc927b6 | 2001 | show_stats(); |
eb9bd4e9 | 2002 | free(ofp_buffer); |
d5396421 | 2003 | return ret; |
d0ca268b | 2004 | } |