Commit | Line | Data |
---|---|---|
40bafa33 JA |
1 | #include <stdio.h> |
2 | #include <stdio.h> | |
3 | #include <unistd.h> | |
4 | #include <inttypes.h> | |
5 | #include <assert.h> | |
6 | ||
7 | #include "../io_ddir.h" | |
8 | #include "../flist.h" | |
9 | #include "../hash.h" | |
10 | #include "../fifo.h" | |
11 | #include "../blktrace_api.h" | |
12 | #include "../os/os.h" | |
13 | #include "../log.h" | |
14 | #include "../lib/linux-dev-lookup.h" | |
15 | ||
16 | #define TRACE_FIFO_SIZE 8192 | |
17 | ||
18 | static unsigned int rt_threshold = 1000000; | |
19 | static unsigned int ios_threshold = 10; | |
8fc46b4d JA |
20 | static unsigned int rate_threshold; |
21 | static unsigned int set_rate; | |
40bafa33 | 22 | static int output_ascii = 1; |
35661615 | 23 | static char *filename; |
40bafa33 JA |
24 | |
25 | struct bs { | |
26 | unsigned int bs; | |
27 | unsigned int nr; | |
28 | int merges; | |
29 | }; | |
30 | ||
31 | struct trace_file { | |
32 | char *name; | |
33 | int major, minor; | |
34 | }; | |
35 | ||
36 | struct btrace_out { | |
37 | unsigned long ios[DDIR_RWDIR_CNT]; | |
40bafa33 JA |
38 | unsigned long merges[DDIR_RWDIR_CNT]; |
39 | ||
40 | uint64_t last_end[DDIR_RWDIR_CNT]; | |
41 | uint64_t seq[DDIR_RWDIR_CNT]; | |
42 | ||
43 | struct bs *bs[DDIR_RWDIR_CNT]; | |
44 | unsigned int nr_bs[DDIR_RWDIR_CNT]; | |
45 | ||
46 | int inflight; | |
47 | unsigned int depth; | |
8fc46b4d JA |
48 | uint64_t first_ttime[DDIR_RWDIR_CNT]; |
49 | uint64_t last_ttime[DDIR_RWDIR_CNT]; | |
50 | uint64_t kb[DDIR_RWDIR_CNT]; | |
40bafa33 | 51 | |
40bafa33 JA |
52 | uint64_t start_delay; |
53 | }; | |
54 | ||
55 | struct btrace_pid { | |
56 | struct flist_head hash_list; | |
57 | struct flist_head pid_list; | |
58 | pid_t pid; | |
65ff9cd5 JA |
59 | |
60 | struct trace_file *files; | |
61 | int nr_files; | |
62 | unsigned int last_major, last_minor; | |
63 | ||
40bafa33 JA |
64 | struct btrace_out o; |
65 | }; | |
66 | ||
67 | struct inflight { | |
68 | struct flist_head list; | |
69 | struct btrace_pid *p; | |
70 | uint64_t end_sector; | |
71 | }; | |
72 | ||
73 | #define PID_HASH_BITS 10 | |
74 | #define PID_HASH_SIZE (1U << PID_HASH_BITS) | |
75 | ||
76 | static struct flist_head pid_hash[PID_HASH_SIZE]; | |
77 | static FLIST_HEAD(pid_list); | |
78 | ||
17bf0853 JA |
79 | #define INFLIGHT_HASH_BITS 8 |
80 | #define INFLIGHT_HASH_SIZE (1U << INFLIGHT_HASH_BITS) | |
81 | static struct flist_head inflight_hash[INFLIGHT_HASH_SIZE]; | |
40bafa33 JA |
82 | |
83 | static uint64_t first_ttime = -1ULL; | |
84 | ||
85 | static struct inflight *inflight_find(uint64_t sector) | |
86 | { | |
17bf0853 | 87 | struct flist_head *inflight_list; |
40bafa33 JA |
88 | struct flist_head *e; |
89 | ||
17bf0853 JA |
90 | inflight_list = &inflight_hash[hash_long(sector, INFLIGHT_HASH_BITS)]; |
91 | ||
92 | flist_for_each(e, inflight_list) { | |
40bafa33 JA |
93 | struct inflight *i = flist_entry(e, struct inflight, list); |
94 | ||
95 | if (i->end_sector == sector) | |
96 | return i; | |
97 | } | |
98 | ||
99 | return NULL; | |
100 | } | |
101 | ||
102 | static void inflight_remove(struct inflight *i) | |
103 | { | |
104 | struct btrace_out *o = &i->p->o; | |
105 | ||
106 | o->inflight--; | |
107 | assert(o->inflight >= 0); | |
108 | flist_del(&i->list); | |
109 | free(i); | |
110 | } | |
111 | ||
17bf0853 | 112 | static void __inflight_add(struct inflight *i) |
40bafa33 | 113 | { |
17bf0853 JA |
114 | struct flist_head *list; |
115 | ||
116 | list = &inflight_hash[hash_long(i->end_sector, INFLIGHT_HASH_BITS)]; | |
117 | flist_add_tail(&i->list, list); | |
40bafa33 JA |
118 | } |
119 | ||
120 | static void inflight_add(struct btrace_pid *p, uint64_t sector, uint32_t len) | |
121 | { | |
122 | struct btrace_out *o = &p->o; | |
123 | struct inflight *i; | |
124 | ||
125 | i = calloc(1, sizeof(*i)); | |
126 | i->p = p; | |
127 | o->inflight++; | |
128 | o->depth = max((int) o->depth, o->inflight); | |
129 | i->end_sector = sector + (len >> 9); | |
17bf0853 JA |
130 | __inflight_add(i); |
131 | } | |
132 | ||
133 | static void inflight_merge(struct inflight *i, int rw, unsigned int size) | |
134 | { | |
135 | i->p->o.merges[rw]++; | |
136 | if (size) { | |
137 | i->end_sector += (size >> 9); | |
138 | flist_del(&i->list); | |
139 | __inflight_add(i); | |
140 | } | |
40bafa33 JA |
141 | } |
142 | ||
143 | /* | |
144 | * fifo refill frontend, to avoid reading data in trace sized bites | |
145 | */ | |
146 | static int refill_fifo(struct fifo *fifo, int fd) | |
147 | { | |
148 | char buf[TRACE_FIFO_SIZE]; | |
149 | unsigned int total; | |
150 | int ret; | |
151 | ||
152 | total = sizeof(buf); | |
153 | if (total > fifo_room(fifo)) | |
154 | total = fifo_room(fifo); | |
155 | ||
156 | ret = read(fd, buf, total); | |
157 | if (ret < 0) { | |
158 | perror("read refill"); | |
159 | return -1; | |
160 | } | |
161 | ||
162 | if (ret > 0) | |
163 | ret = fifo_put(fifo, buf, ret); | |
164 | ||
165 | return ret; | |
166 | } | |
167 | ||
168 | /* | |
169 | * Retrieve 'len' bytes from the fifo, refilling if necessary. | |
170 | */ | |
171 | static int trace_fifo_get(struct fifo *fifo, int fd, void *buf, | |
172 | unsigned int len) | |
173 | { | |
174 | if (fifo_len(fifo) < len) { | |
175 | int ret = refill_fifo(fifo, fd); | |
176 | ||
177 | if (ret < 0) | |
178 | return ret; | |
179 | } | |
180 | ||
181 | return fifo_get(fifo, buf, len); | |
182 | } | |
183 | ||
184 | /* | |
185 | * Just discard the pdu by seeking past it. | |
186 | */ | |
187 | static int discard_pdu(struct fifo *fifo, int fd, struct blk_io_trace *t) | |
188 | { | |
189 | if (t->pdu_len == 0) | |
190 | return 0; | |
191 | ||
192 | return trace_fifo_get(fifo, fd, NULL, t->pdu_len); | |
193 | } | |
194 | ||
8fc46b4d | 195 | static int handle_trace_notify(struct blk_io_trace *t) |
40bafa33 JA |
196 | { |
197 | switch (t->action) { | |
198 | case BLK_TN_PROCESS: | |
199 | //printf("got process notify: %x, %d\n", t->action, t->pid); | |
200 | break; | |
201 | case BLK_TN_TIMESTAMP: | |
202 | //printf("got timestamp notify: %x, %d\n", t->action, t->pid); | |
203 | break; | |
204 | case BLK_TN_MESSAGE: | |
205 | break; | |
206 | default: | |
17bf0853 | 207 | log_err("unknown trace act %x\n", t->action); |
8fc46b4d | 208 | return 1; |
40bafa33 | 209 | } |
8fc46b4d JA |
210 | |
211 | return 0; | |
40bafa33 JA |
212 | } |
213 | ||
214 | static void __add_bs(struct btrace_out *o, unsigned int len, int rw) | |
215 | { | |
216 | o->bs[rw] = realloc(o->bs[rw], (o->nr_bs[rw] + 1) * sizeof(struct bs)); | |
217 | o->bs[rw][o->nr_bs[rw]].bs = len; | |
218 | o->bs[rw][o->nr_bs[rw]].nr = 1; | |
219 | o->nr_bs[rw]++; | |
220 | } | |
221 | ||
222 | static void add_bs(struct btrace_out *o, unsigned int len, int rw) | |
223 | { | |
224 | struct bs *bs = o->bs[rw]; | |
225 | int i; | |
226 | ||
227 | if (!o->nr_bs[rw]) { | |
228 | __add_bs(o, len, rw); | |
229 | return; | |
230 | } | |
231 | ||
232 | for (i = 0; i < o->nr_bs[rw]; i++) { | |
233 | if (bs[i].bs == len) { | |
234 | bs[i].nr++; | |
235 | return; | |
236 | } | |
237 | } | |
238 | ||
239 | __add_bs(o, len, rw); | |
240 | } | |
241 | ||
242 | #define FMINORBITS 20 | |
243 | #define FMINORMASK ((1U << FMINORBITS) - 1) | |
244 | #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) | |
245 | #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK)) | |
246 | ||
8fc46b4d | 247 | static int btrace_add_file(struct btrace_pid *p, uint32_t devno) |
40bafa33 JA |
248 | { |
249 | unsigned int maj = FMAJOR(devno); | |
250 | unsigned int min = FMINOR(devno); | |
251 | struct trace_file *f; | |
252 | unsigned int i; | |
253 | char dev[256]; | |
254 | ||
35661615 | 255 | if (filename) |
8fc46b4d | 256 | return 0; |
65ff9cd5 | 257 | if (p->last_major == maj && p->last_minor == min) |
8fc46b4d | 258 | return 0; |
40bafa33 | 259 | |
65ff9cd5 JA |
260 | p->last_major = maj; |
261 | p->last_minor = min; | |
40bafa33 JA |
262 | |
263 | /* | |
264 | * check for this file in our list | |
265 | */ | |
65ff9cd5 JA |
266 | for (i = 0; i < p->nr_files; i++) { |
267 | f = &p->files[i]; | |
40bafa33 JA |
268 | |
269 | if (f->major == maj && f->minor == min) | |
8fc46b4d | 270 | return 0; |
40bafa33 JA |
271 | } |
272 | ||
273 | strcpy(dev, "/dev"); | |
274 | if (!blktrace_lookup_device(NULL, dev, maj, min)) { | |
275 | log_err("fio: failed to find device %u/%u\n", maj, min); | |
8fc46b4d JA |
276 | if (!output_ascii) { |
277 | log_err("fio: use -d to specify device\n"); | |
278 | return 1; | |
279 | } | |
280 | return 0; | |
40bafa33 JA |
281 | } |
282 | ||
65ff9cd5 JA |
283 | p->files = realloc(p->files, (p->nr_files + 1) * sizeof(*f)); |
284 | f = &p->files[p->nr_files]; | |
40bafa33 JA |
285 | f->name = strdup(dev); |
286 | f->major = maj; | |
287 | f->minor = min; | |
65ff9cd5 | 288 | p->nr_files++; |
8fc46b4d JA |
289 | return 0; |
290 | } | |
291 | ||
292 | static int t_to_rwdir(struct blk_io_trace *t) | |
293 | { | |
294 | if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) | |
295 | return DDIR_TRIM; | |
296 | ||
297 | return (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; | |
40bafa33 JA |
298 | } |
299 | ||
8fc46b4d | 300 | static int handle_trace_discard(struct blk_io_trace *t, struct btrace_pid *p) |
40bafa33 | 301 | { |
65ff9cd5 JA |
302 | struct btrace_out *o = &p->o; |
303 | ||
8fc46b4d JA |
304 | if (btrace_add_file(p, t->device)) |
305 | return 1; | |
40bafa33 | 306 | |
8fc46b4d JA |
307 | if (o->first_ttime[2] == -1ULL) |
308 | o->first_ttime[2] = t->time; | |
40bafa33 JA |
309 | |
310 | o->ios[DDIR_TRIM]++; | |
311 | add_bs(o, t->bytes, DDIR_TRIM); | |
8fc46b4d | 312 | return 0; |
40bafa33 JA |
313 | } |
314 | ||
8fc46b4d | 315 | static int handle_trace_fs(struct blk_io_trace *t, struct btrace_pid *p) |
40bafa33 | 316 | { |
65ff9cd5 | 317 | struct btrace_out *o = &p->o; |
40bafa33 JA |
318 | int rw; |
319 | ||
8fc46b4d JA |
320 | if (btrace_add_file(p, t->device)) |
321 | return 1; | |
40bafa33 JA |
322 | |
323 | first_ttime = min(first_ttime, (uint64_t) t->time); | |
324 | ||
40bafa33 JA |
325 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
326 | ||
8fc46b4d JA |
327 | if (o->first_ttime[rw] == -1ULL) |
328 | o->first_ttime[rw] = t->time; | |
329 | ||
40bafa33 JA |
330 | add_bs(o, t->bytes, rw); |
331 | o->ios[rw]++; | |
332 | ||
333 | if (t->sector == o->last_end[rw] || o->last_end[rw] == -1ULL) | |
334 | o->seq[rw]++; | |
335 | ||
336 | o->last_end[rw] = t->sector + (t->bytes >> 9); | |
8fc46b4d | 337 | return 0; |
40bafa33 JA |
338 | } |
339 | ||
8fc46b4d | 340 | static int handle_queue_trace(struct blk_io_trace *t, struct btrace_pid *p) |
40bafa33 JA |
341 | { |
342 | if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY)) | |
8fc46b4d | 343 | return handle_trace_notify(t); |
40bafa33 | 344 | else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) |
8fc46b4d | 345 | return handle_trace_discard(t, p); |
40bafa33 | 346 | else |
8fc46b4d | 347 | return handle_trace_fs(t, p); |
40bafa33 JA |
348 | } |
349 | ||
8fc46b4d | 350 | static int handle_trace(struct blk_io_trace *t, struct btrace_pid *p) |
40bafa33 JA |
351 | { |
352 | unsigned int act = t->action & 0xffff; | |
8fc46b4d | 353 | int ret = 0; |
40bafa33 JA |
354 | |
355 | if (act == __BLK_TA_QUEUE) { | |
356 | inflight_add(p, t->sector, t->bytes); | |
8fc46b4d | 357 | ret = handle_queue_trace(t, p); |
40bafa33 JA |
358 | } else if (act == __BLK_TA_REQUEUE) { |
359 | p->o.inflight--; | |
360 | } else if (act == __BLK_TA_BACKMERGE) { | |
361 | struct inflight *i; | |
362 | ||
363 | i = inflight_find(t->sector + (t->bytes >> 9)); | |
364 | if (i) | |
365 | inflight_remove(i); | |
366 | ||
367 | i = inflight_find(t->sector); | |
8fc46b4d JA |
368 | if (i) |
369 | inflight_merge(i, t_to_rwdir(t), t->bytes); | |
40bafa33 JA |
370 | } else if (act == __BLK_TA_FRONTMERGE) { |
371 | struct inflight *i; | |
372 | ||
373 | i = inflight_find(t->sector + (t->bytes >> 9)); | |
374 | if (i) | |
375 | inflight_remove(i); | |
376 | ||
377 | i = inflight_find(t->sector); | |
8fc46b4d JA |
378 | if (i) |
379 | inflight_merge(i, t_to_rwdir(t), 0); | |
40bafa33 JA |
380 | } else if (act == __BLK_TA_COMPLETE) { |
381 | struct inflight *i; | |
382 | ||
383 | i = inflight_find(t->sector + (t->bytes >> 9)); | |
cea475cd | 384 | if (i) { |
8fc46b4d | 385 | i->p->o.kb[t_to_rwdir(t)] += (t->bytes >> 10); |
40bafa33 | 386 | inflight_remove(i); |
cea475cd | 387 | } |
40bafa33 | 388 | } |
8fc46b4d JA |
389 | |
390 | return ret; | |
40bafa33 JA |
391 | } |
392 | ||
393 | static void byteswap_trace(struct blk_io_trace *t) | |
394 | { | |
395 | t->magic = fio_swap32(t->magic); | |
396 | t->sequence = fio_swap32(t->sequence); | |
397 | t->time = fio_swap64(t->time); | |
398 | t->sector = fio_swap64(t->sector); | |
399 | t->bytes = fio_swap32(t->bytes); | |
400 | t->action = fio_swap32(t->action); | |
401 | t->pid = fio_swap32(t->pid); | |
402 | t->device = fio_swap32(t->device); | |
403 | t->cpu = fio_swap32(t->cpu); | |
404 | t->error = fio_swap16(t->error); | |
405 | t->pdu_len = fio_swap16(t->pdu_len); | |
406 | } | |
407 | ||
408 | static struct btrace_pid *pid_hash_find(pid_t pid, struct flist_head *list) | |
409 | { | |
410 | struct flist_head *e; | |
411 | struct btrace_pid *p; | |
412 | ||
413 | flist_for_each(e, list) { | |
414 | p = flist_entry(e, struct btrace_pid, hash_list); | |
415 | if (p->pid == pid) | |
416 | return p; | |
417 | } | |
418 | ||
419 | return NULL; | |
420 | } | |
421 | ||
422 | static struct btrace_pid *pid_hash_get(pid_t pid) | |
423 | { | |
424 | struct flist_head *hash_list; | |
425 | struct btrace_pid *p; | |
426 | ||
427 | hash_list = &pid_hash[hash_long(pid, PID_HASH_BITS)]; | |
428 | ||
429 | p = pid_hash_find(pid, hash_list); | |
430 | if (!p) { | |
431 | int i; | |
432 | ||
433 | p = calloc(1, sizeof(*p)); | |
40bafa33 | 434 | |
8fc46b4d JA |
435 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
436 | p->o.first_ttime[i] = -1ULL; | |
437 | p->o.last_ttime[i] = -1ULL; | |
40bafa33 | 438 | p->o.last_end[i] = -1ULL; |
8fc46b4d | 439 | } |
40bafa33 JA |
440 | |
441 | p->pid = pid; | |
442 | flist_add_tail(&p->hash_list, hash_list); | |
443 | flist_add_tail(&p->pid_list, &pid_list); | |
444 | } | |
445 | ||
446 | return p; | |
447 | } | |
448 | ||
449 | /* | |
450 | * Load a blktrace file by reading all the blk_io_trace entries, and storing | |
451 | * them as io_pieces like the fio text version would do. | |
452 | */ | |
453 | static int load_blktrace(const char *filename, int need_swap) | |
454 | { | |
455 | struct btrace_pid *p; | |
456 | unsigned long traces; | |
457 | struct blk_io_trace t; | |
458 | struct fifo *fifo; | |
8fc46b4d | 459 | int fd, ret = 0; |
40bafa33 JA |
460 | |
461 | fd = open(filename, O_RDONLY); | |
462 | if (fd < 0) { | |
463 | perror("open trace file\n"); | |
464 | return 1; | |
465 | } | |
466 | ||
467 | fifo = fifo_alloc(TRACE_FIFO_SIZE); | |
468 | ||
469 | traces = 0; | |
470 | do { | |
471 | int ret = trace_fifo_get(fifo, fd, &t, sizeof(t)); | |
472 | ||
473 | if (ret < 0) | |
474 | goto err; | |
475 | else if (!ret) | |
476 | break; | |
477 | else if (ret < (int) sizeof(t)) { | |
17bf0853 | 478 | log_err("fio: short fifo get\n"); |
40bafa33 JA |
479 | break; |
480 | } | |
481 | ||
482 | if (need_swap) | |
483 | byteswap_trace(&t); | |
484 | ||
485 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { | |
17bf0853 | 486 | log_err("fio: bad magic in blktrace data: %x\n", t.magic); |
40bafa33 JA |
487 | goto err; |
488 | } | |
489 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { | |
17bf0853 | 490 | log_err("fio: bad blktrace version %d\n", t.magic & 0xff); |
40bafa33 JA |
491 | goto err; |
492 | } | |
493 | ret = discard_pdu(fifo, fd, &t); | |
494 | if (ret < 0) { | |
17bf0853 | 495 | log_err("blktrace lseek\n"); |
40bafa33 JA |
496 | goto err; |
497 | } else if (t.pdu_len != ret) { | |
17bf0853 | 498 | log_err("fio: discarded %d of %d\n", ret, t.pdu_len); |
40bafa33 JA |
499 | goto err; |
500 | } | |
501 | ||
502 | p = pid_hash_get(t.pid); | |
8fc46b4d JA |
503 | ret = handle_trace(&t, p); |
504 | if (ret) | |
505 | break; | |
506 | p->o.last_ttime[t_to_rwdir(&t)] = t.time; | |
40bafa33 JA |
507 | traces++; |
508 | } while (1); | |
509 | ||
510 | fifo_free(fifo); | |
511 | close(fd); | |
512 | ||
8fc46b4d JA |
513 | if (ret) |
514 | return ret; | |
515 | ||
40bafa33 JA |
516 | if (output_ascii) |
517 | printf("Traces loaded: %lu\n", traces); | |
518 | ||
519 | return 0; | |
520 | err: | |
521 | close(fd); | |
522 | fifo_free(fifo); | |
523 | return 1; | |
524 | } | |
525 | ||
526 | static int bs_cmp(const void *ba, const void *bb) | |
527 | { | |
528 | const struct bs *bsa = ba; | |
529 | const struct bs *bsb = bb; | |
530 | ||
531 | return bsb->nr - bsa->nr; | |
532 | } | |
533 | ||
8fc46b4d JA |
534 | static unsigned long o_to_kb_rate(struct btrace_out *o, int rw) |
535 | { | |
536 | uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL; | |
537 | uint64_t val; | |
538 | ||
539 | if (!usec) | |
540 | return 0; | |
541 | ||
542 | val = o->kb[rw] * 1000ULL; | |
543 | return val / (usec / 1000ULL); | |
544 | } | |
545 | ||
546 | static uint64_t o_first_ttime(struct btrace_out *o) | |
547 | { | |
548 | uint64_t first; | |
549 | ||
550 | first = min(o->first_ttime[0], o->first_ttime[1]); | |
551 | return min(first, o->first_ttime[2]); | |
552 | } | |
553 | ||
554 | static uint64_t o_longest_ttime(struct btrace_out *o) | |
555 | { | |
556 | uint64_t ret = 0; | |
557 | int i; | |
558 | ||
559 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
560 | uint64_t diff; | |
561 | ||
562 | diff = o->last_ttime[i] - o->first_ttime[i]; | |
563 | ret = max(diff, ret); | |
564 | } | |
565 | ||
566 | return ret; | |
567 | } | |
568 | ||
40bafa33 JA |
569 | static void __output_p_ascii(struct btrace_pid *p, unsigned long *ios) |
570 | { | |
571 | const char *msg[] = { "reads", "writes", "trims" }; | |
572 | struct btrace_out *o = &p->o; | |
cea475cd | 573 | unsigned long total, usec; |
40bafa33 JA |
574 | int i, j; |
575 | ||
576 | printf("[pid:\t%u]\n", p->pid); | |
577 | ||
578 | total = ddir_rw_sum(o->ios); | |
579 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
580 | float perc; | |
581 | ||
582 | if (!o->ios[i]) | |
583 | continue; | |
584 | ||
585 | ios[i] += o->ios[i] + o->merges[i]; | |
586 | printf("%s\n", msg[i]); | |
587 | perc = ((float) o->ios[i] * 100.0) / (float) total; | |
588 | printf("\tios: %lu (perc=%3.2f%%)\n", o->ios[i], perc); | |
589 | perc = ((float) o->merges[i] * 100.0) / (float) total; | |
590 | printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc); | |
591 | perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i]; | |
761c2729 | 592 | printf("\tseq: %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc); |
8fc46b4d | 593 | printf("\trate: %lu KB/sec\n", o_to_kb_rate(o, i)); |
40bafa33 JA |
594 | |
595 | for (j = 0; j < o->nr_bs[i]; j++) { | |
596 | struct bs *bs = &o->bs[i][j]; | |
597 | ||
598 | perc = (((float) bs->nr * 100.0) / (float) o->ios[i]); | |
599 | printf("\tbs=%u, perc=%3.2f%%\n", bs->bs, perc); | |
600 | } | |
601 | } | |
602 | ||
603 | printf("depth:\t%u\n", o->depth); | |
8fc46b4d | 604 | usec = o_longest_ttime(o) / 1000ULL; |
cea475cd | 605 | printf("usec:\t%lu (delay=%llu)\n", usec, (unsigned long long) o->start_delay); |
40bafa33 JA |
606 | |
607 | printf("files:\t"); | |
65ff9cd5 JA |
608 | for (i = 0; i < p->nr_files; i++) |
609 | printf("%s,", p->files[i].name); | |
40bafa33 JA |
610 | printf("\n"); |
611 | ||
612 | printf("\n"); | |
613 | } | |
614 | ||
615 | static int __output_p_fio(struct btrace_pid *p, unsigned long *ios) | |
616 | { | |
617 | struct btrace_out *o = &p->o; | |
618 | unsigned long total; | |
1a8cad44 | 619 | unsigned long long time; |
40bafa33 JA |
620 | float perc; |
621 | int i, j; | |
622 | ||
623 | if ((o->ios[0] + o->ios[1]) && o->ios[2]) { | |
624 | log_err("fio: trace has both read/write and trim\n"); | |
625 | return 1; | |
626 | } | |
8fc46b4d JA |
627 | if (!p->nr_files) { |
628 | log_err("fio: no devices found\n"); | |
629 | return 1; | |
630 | } | |
40bafa33 JA |
631 | |
632 | printf("[pid%u]\n", p->pid); | |
633 | printf("direct=1\n"); | |
634 | if (o->depth == 1) | |
635 | printf("ioengine=sync\n"); | |
636 | else | |
637 | printf("ioengine=libaio\niodepth=%u\n", o->depth); | |
638 | ||
639 | if (o->ios[0] && !o->ios[1]) | |
640 | printf("rw=randread\n"); | |
641 | else if (!o->ios[0] && o->ios[1]) | |
642 | printf("rw=randwrite\n"); | |
643 | else if (o->ios[2]) | |
644 | printf("rw=randtrim\n"); | |
645 | else { | |
646 | printf("rw=randrw\n"); | |
647 | total = ddir_rw_sum(o->ios); | |
648 | perc = ((float) o->ios[0] * 100.0) / (float) total; | |
649 | printf("rwmixread=%u\n", (int) (perc + 0.99)); | |
650 | } | |
651 | ||
35661615 | 652 | printf("percentage_random="); |
40bafa33 JA |
653 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
654 | if (o->seq[i] && o->ios[i]) { | |
655 | perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i]; | |
656 | if (perc >= 99.0) | |
657 | perc = 100.0; | |
658 | } else | |
659 | perc = 100.0; | |
660 | ||
661 | if (i) | |
662 | printf(","); | |
35661615 | 663 | perc = 100.0 - perc; |
40bafa33 JA |
664 | printf("%u", (int) perc); |
665 | } | |
666 | printf("\n"); | |
667 | ||
668 | printf("filename="); | |
65ff9cd5 | 669 | for (i = 0; i < p->nr_files; i++) { |
40bafa33 JA |
670 | if (i) |
671 | printf(":"); | |
65ff9cd5 | 672 | printf("%s", p->files[i].name); |
40bafa33 JA |
673 | } |
674 | printf("\n"); | |
675 | ||
676 | printf("startdelay=%llus\n", o->start_delay / 1000000ULL); | |
677 | ||
8fc46b4d | 678 | time = o_longest_ttime(o); |
1a8cad44 JA |
679 | time = (time + 1000000000ULL - 1) / 1000000000ULL; |
680 | printf("runtime=%llus\n", time); | |
681 | ||
40bafa33 JA |
682 | printf("bssplit="); |
683 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
684 | ||
685 | if (i && o->nr_bs[i - 1] && o->nr_bs[i]) | |
686 | printf(","); | |
687 | ||
688 | for (j = 0; j < o->nr_bs[i]; j++) { | |
689 | struct bs *bs = &o->bs[i][j]; | |
690 | ||
691 | perc = (((float) bs->nr * 100.0) / (float) o->ios[i]); | |
692 | if (perc < 1.00) | |
693 | continue; | |
694 | if (j) | |
695 | printf(":"); | |
696 | if (j + 1 == o->nr_bs[i]) | |
697 | printf("%u/", bs->bs); | |
698 | else | |
699 | printf("%u/%u", bs->bs, (int) perc); | |
700 | } | |
701 | } | |
8fc46b4d | 702 | printf("\n"); |
40bafa33 | 703 | |
8fc46b4d JA |
704 | if (set_rate) { |
705 | printf("rate="); | |
706 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
707 | unsigned long rate; | |
708 | ||
709 | rate = o_to_kb_rate(o, i); | |
710 | if (i) | |
711 | printf(","); | |
712 | if (rate) | |
713 | printf("%luk", rate); | |
714 | } | |
715 | printf("\n"); | |
716 | } | |
717 | ||
718 | printf("\n"); | |
40bafa33 JA |
719 | return 0; |
720 | } | |
721 | ||
722 | static int __output_p(struct btrace_pid *p, unsigned long *ios) | |
723 | { | |
724 | struct btrace_out *o = &p->o; | |
725 | int i, ret = 0; | |
726 | ||
727 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
728 | if (o->nr_bs[i] <= 1) | |
729 | continue; | |
730 | qsort(o->bs[i], o->nr_bs[i], sizeof(struct bs), bs_cmp); | |
731 | } | |
732 | ||
35661615 | 733 | if (filename) { |
65ff9cd5 JA |
734 | p->files = malloc(sizeof(struct trace_file)); |
735 | p->nr_files++; | |
736 | p->files[0].name = filename; | |
35661615 JA |
737 | } |
738 | ||
40bafa33 JA |
739 | if (output_ascii) |
740 | __output_p_ascii(p, ios); | |
741 | else | |
742 | ret = __output_p_fio(p, ios); | |
743 | ||
744 | return ret; | |
745 | } | |
746 | ||
8fc46b4d JA |
747 | static void remove_ddir(struct btrace_out *o, int rw) |
748 | { | |
749 | o->ios[rw] = 0; | |
750 | } | |
751 | ||
40bafa33 JA |
752 | static int prune_entry(struct btrace_out *o) |
753 | { | |
8fc46b4d | 754 | unsigned long rate; |
40bafa33 | 755 | uint64_t time; |
8fc46b4d | 756 | int i; |
40bafa33 JA |
757 | |
758 | if (ddir_rw_sum(o->ios) < ios_threshold) | |
759 | return 1; | |
760 | ||
8fc46b4d | 761 | time = o_longest_ttime(o) / 1000ULL; |
40bafa33 JA |
762 | if (time < rt_threshold) |
763 | return 1; | |
764 | ||
8fc46b4d JA |
765 | rate = 0; |
766 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { | |
767 | unsigned long this_rate; | |
768 | ||
769 | this_rate = o_to_kb_rate(o, i); | |
770 | if (this_rate < rate_threshold) { | |
771 | remove_ddir(o, i); | |
772 | this_rate = 0; | |
773 | } | |
774 | rate += this_rate; | |
775 | } | |
776 | ||
777 | if (rate < rate_threshold) | |
778 | return 1; | |
779 | ||
40bafa33 JA |
780 | return 0; |
781 | } | |
782 | ||
783 | static int entry_cmp(void *priv, struct flist_head *a, struct flist_head *b) | |
784 | { | |
785 | struct btrace_pid *pa = flist_entry(a, struct btrace_pid, pid_list); | |
786 | struct btrace_pid *pb = flist_entry(b, struct btrace_pid, pid_list); | |
787 | ||
788 | return ddir_rw_sum(pb->o.ios) - ddir_rw_sum(pa->o.ios); | |
789 | } | |
790 | ||
17bf0853 JA |
791 | static void free_p(struct btrace_pid *p) |
792 | { | |
793 | struct btrace_out *o = &p->o; | |
794 | int i; | |
795 | ||
65ff9cd5 JA |
796 | for (i = 0; i < p->nr_files; i++) { |
797 | if (p->files[i].name && p->files[i].name != filename) | |
798 | free(p->files[i].name); | |
17bf0853 JA |
799 | } |
800 | ||
801 | for (i = 0; i < DDIR_RWDIR_CNT; i++) | |
802 | free(o->bs[i]); | |
803 | ||
65ff9cd5 | 804 | free(p->files); |
17bf0853 JA |
805 | flist_del(&p->pid_list); |
806 | flist_del(&p->hash_list); | |
807 | free(p); | |
808 | } | |
809 | ||
40bafa33 JA |
810 | static int output_p(void) |
811 | { | |
812 | unsigned long ios[DDIR_RWDIR_CNT]; | |
813 | struct flist_head *e, *tmp; | |
814 | int ret = 0; | |
815 | ||
816 | flist_for_each_safe(e, tmp, &pid_list) { | |
817 | struct btrace_pid *p; | |
818 | ||
819 | p = flist_entry(e, struct btrace_pid, pid_list); | |
820 | if (prune_entry(&p->o)) { | |
17bf0853 | 821 | free_p(p); |
40bafa33 JA |
822 | continue; |
823 | } | |
8fc46b4d | 824 | p->o.start_delay = (o_first_ttime(&p->o) / 1000ULL) - first_ttime; |
40bafa33 JA |
825 | } |
826 | ||
827 | memset(ios, 0, sizeof(ios)); | |
828 | ||
829 | flist_sort(NULL, &pid_list, entry_cmp); | |
830 | ||
831 | flist_for_each(e, &pid_list) { | |
832 | struct btrace_pid *p; | |
833 | ||
834 | p = flist_entry(e, struct btrace_pid, pid_list); | |
835 | ret |= __output_p(p, ios); | |
8fc46b4d JA |
836 | if (ret && !output_ascii) |
837 | break; | |
40bafa33 JA |
838 | } |
839 | ||
840 | if (output_ascii) | |
841 | printf("Total: reads=%lu, writes=%lu\n", ios[0], ios[1]); | |
842 | ||
843 | return ret; | |
844 | } | |
845 | ||
846 | static int usage(char *argv[]) | |
847 | { | |
17bf0853 JA |
848 | log_err("%s: <blktrace bin file>\n", argv[0]); |
849 | log_err("\t-t\tUsec threshold to ignore task\n"); | |
850 | log_err("\t-n\tNumber IOS threshold to ignore task\n"); | |
851 | log_err("\t-f\tFio job file output\n"); | |
852 | log_err("\t-d\tUse this file/device for replay\n"); | |
8fc46b4d JA |
853 | log_err("\t-r\tIgnore jobs with less than this KB/sec rate\n"); |
854 | log_err("\t-R\tSet rate in fio job\n"); | |
40bafa33 JA |
855 | return 1; |
856 | } | |
857 | ||
17bf0853 | 858 | static int trace_needs_swap(const char *trace_file, int *swap) |
40bafa33 | 859 | { |
40bafa33 | 860 | struct blk_io_trace t; |
17bf0853 JA |
861 | int fd, ret; |
862 | ||
863 | *swap = -1; | |
864 | ||
865 | fd = open(trace_file, O_RDONLY); | |
866 | if (fd < 0) { | |
867 | perror("open"); | |
868 | return 1; | |
869 | } | |
870 | ||
871 | ret = read(fd, &t, sizeof(t)); | |
872 | if (ret < 0) { | |
873 | perror("read"); | |
874 | return 1; | |
875 | } else if (ret != sizeof(t)) { | |
876 | log_err("fio: short read on trace file\n"); | |
877 | return 1; | |
878 | } | |
879 | ||
880 | close(fd); | |
881 | ||
882 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) | |
883 | *swap = 0; | |
884 | else { | |
885 | /* | |
886 | * Maybe it needs to be endian swapped... | |
887 | */ | |
888 | t.magic = fio_swap32(t.magic); | |
889 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) | |
890 | *swap = 1; | |
891 | } | |
892 | ||
893 | if (*swap == -1) { | |
894 | log_err("fio: blktrace appears corrupt\n"); | |
895 | return 1; | |
896 | } | |
897 | ||
898 | return 0; | |
899 | } | |
900 | ||
901 | int main(int argc, char *argv[]) | |
902 | { | |
903 | int need_swap, i, c; | |
40bafa33 JA |
904 | |
905 | if (argc < 2) | |
906 | return usage(argv); | |
907 | ||
8fc46b4d | 908 | while ((c = getopt(argc, argv, "t:n:fd:r:R")) != -1) { |
40bafa33 | 909 | switch (c) { |
8fc46b4d JA |
910 | case 'R': |
911 | set_rate = 1; | |
912 | break; | |
913 | case 'r': | |
914 | rate_threshold = atoi(optarg); | |
915 | break; | |
40bafa33 JA |
916 | case 't': |
917 | rt_threshold = atoi(optarg); | |
918 | break; | |
919 | case 'n': | |
920 | ios_threshold = atoi(optarg); | |
921 | break; | |
922 | case 'f': | |
923 | output_ascii = 0; | |
924 | break; | |
35661615 JA |
925 | case 'd': |
926 | filename = strdup(optarg); | |
927 | break; | |
40bafa33 JA |
928 | case '?': |
929 | default: | |
930 | return usage(argv); | |
931 | } | |
932 | } | |
933 | ||
934 | if (argc == optind) | |
935 | return usage(argv); | |
936 | ||
17bf0853 | 937 | if (trace_needs_swap(argv[optind], &need_swap)) |
40bafa33 | 938 | return 1; |
40bafa33 JA |
939 | |
940 | for (i = 0; i < PID_HASH_SIZE; i++) | |
941 | INIT_FLIST_HEAD(&pid_hash[i]); | |
17bf0853 JA |
942 | for (i = 0; i < INFLIGHT_HASH_SIZE; i++) |
943 | INIT_FLIST_HEAD(&inflight_hash[i]); | |
40bafa33 JA |
944 | |
945 | load_blktrace(argv[optind], need_swap); | |
946 | first_ttime /= 1000ULL; | |
947 | ||
948 | return output_p(); | |
949 | } |