1 /*
2 * blktrace support code for fio
3 */
4 #include <stdio.h>
5 #include <stdlib.h>
6 #include <sys/stat.h>
7 #include <sys/ioctl.h>
8 #include <linux/fs.h>
9 #include <dirent.h>
10
11 #include "flist.h"
12 #include "fio.h"
13 #include "blktrace_api.h"
14 #include "oslib/linux-dev-lookup.h"
15
16 #define TRACE_FIFO_SIZE 8192
17
18 /*
19 * fifo refill frontend, to avoid reading data in trace sized bites
20 */
refill_fifo(struct thread_data * td,struct fifo * fifo,int fd)21 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
22 {
23 char buf[TRACE_FIFO_SIZE];
24 unsigned int total;
25 int ret;
26
27 total = sizeof(buf);
28 if (total > fifo_room(fifo))
29 total = fifo_room(fifo);
30
31 ret = read(fd, buf, total);
32 if (ret < 0) {
33 td_verror(td, errno, "read blktrace file");
34 return -1;
35 }
36
37 if (ret > 0)
38 ret = fifo_put(fifo, buf, ret);
39
40 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
41 return ret;
42 }
43
44 /*
45 * Retrieve 'len' bytes from the fifo, refilling if necessary.
46 */
trace_fifo_get(struct thread_data * td,struct fifo * fifo,int fd,void * buf,unsigned int len)47 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
48 void *buf, unsigned int len)
49 {
50 if (fifo_len(fifo) < len) {
51 int ret = refill_fifo(td, fifo, fd);
52
53 if (ret < 0)
54 return ret;
55 }
56
57 return fifo_get(fifo, buf, len);
58 }
59
60 /*
61 * Just discard the pdu by seeking past it.
62 */
discard_pdu(struct thread_data * td,struct fifo * fifo,int fd,struct blk_io_trace * t)63 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
64 struct blk_io_trace *t)
65 {
66 if (t->pdu_len == 0)
67 return 0;
68
69 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
70 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
71 }
72
73 /*
74 * Check if this is a blktrace binary data file. We read a single trace
75 * into memory and check for the magic signature.
76 */
is_blktrace(const char * filename,int * need_swap)77 int is_blktrace(const char *filename, int *need_swap)
78 {
79 struct blk_io_trace t;
80 int fd, ret;
81
82 fd = open(filename, O_RDONLY);
83 if (fd < 0)
84 return 0;
85
86 ret = read(fd, &t, sizeof(t));
87 close(fd);
88
89 if (ret < 0) {
90 perror("read blktrace");
91 return 0;
92 } else if (ret != sizeof(t)) {
93 log_err("fio: short read on blktrace file\n");
94 return 0;
95 }
96
97 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
98 *need_swap = 0;
99 return 1;
100 }
101
102 /*
103 * Maybe it needs to be endian swapped...
104 */
105 t.magic = fio_swap32(t.magic);
106 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
107 *need_swap = 1;
108 return 1;
109 }
110
111 return 0;
112 }
113
114 #define FMINORBITS 20
115 #define FMINORMASK ((1U << FMINORBITS) - 1)
116 #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
117 #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
118
trace_add_open_close_event(struct thread_data * td,int fileno,enum file_log_act action)119 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
120 {
121 struct io_piece *ipo;
122
123 ipo = calloc(1, sizeof(*ipo));
124 init_ipo(ipo);
125
126 ipo->ddir = DDIR_INVAL;
127 ipo->fileno = fileno;
128 ipo->file_action = action;
129 flist_add_tail(&ipo->list, &td->io_log_list);
130 }
131
get_dev_blocksize(const char * dev,unsigned int * bs)132 static int get_dev_blocksize(const char *dev, unsigned int *bs)
133 {
134 int fd;
135
136 fd = open(dev, O_RDONLY);
137 if (fd < 0)
138 return 1;
139
140 if (ioctl(fd, BLKSSZGET, bs) < 0) {
141 close(fd);
142 return 1;
143 }
144
145 close(fd);
146 return 0;
147 }
148
trace_add_file(struct thread_data * td,__u32 device,unsigned int * bs)149 static int trace_add_file(struct thread_data *td, __u32 device,
150 unsigned int *bs)
151 {
152 static unsigned int last_maj, last_min, last_fileno, last_bs;
153 unsigned int maj = FMAJOR(device);
154 unsigned int min = FMINOR(device);
155 struct fio_file *f;
156 unsigned int i;
157 char dev[256];
158
159 if (last_maj == maj && last_min == min) {
160 *bs = last_bs;
161 return last_fileno;
162 }
163
164 last_maj = maj;
165 last_min = min;
166
167 /*
168 * check for this file in our list
169 */
170 for_each_file(td, f, i) {
171 if (f->major == maj && f->minor == min) {
172 last_fileno = f->fileno;
173 last_bs = f->bs;
174 goto out;
175 }
176 }
177
178 strcpy(dev, "/dev");
179 if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
180 unsigned int this_bs;
181 int fileno;
182
183 if (td->o.replay_redirect)
184 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
185 " with: %s\n", maj, min,
186 td->o.replay_redirect);
187 else
188 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
189
190 dprint(FD_BLKTRACE, "add devices %s\n", dev);
191 fileno = add_file_exclusive(td, dev);
192
193 if (get_dev_blocksize(dev, &this_bs))
194 this_bs = 512;
195
196 td->o.open_files++;
197 td->files[fileno]->major = maj;
198 td->files[fileno]->minor = min;
199 td->files[fileno]->bs = this_bs;
200 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
201
202 last_fileno = fileno;
203 last_bs = this_bs;
204 }
205
206 out:
207 *bs = last_bs;
208 return last_fileno;
209 }
210
t_bytes_align(struct thread_options * o,struct blk_io_trace * t)211 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
212 {
213 if (!o->replay_align)
214 return;
215
216 t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
217 }
218
219 /*
220 * Store blk_io_trace data in an ipo for later retrieval.
221 */
store_ipo(struct thread_data * td,unsigned long long offset,unsigned int bytes,int rw,unsigned long long ttime,int fileno,unsigned int bs)222 static void store_ipo(struct thread_data *td, unsigned long long offset,
223 unsigned int bytes, int rw, unsigned long long ttime,
224 int fileno, unsigned int bs)
225 {
226 struct io_piece *ipo = malloc(sizeof(*ipo));
227
228 init_ipo(ipo);
229
230 ipo->offset = offset * bs;
231 if (td->o.replay_scale)
232 ipo->offset = ipo->offset / td->o.replay_scale;
233 ipo_bytes_align(td->o.replay_align, ipo);
234 ipo->len = bytes;
235 ipo->delay = ttime / 1000;
236 if (rw)
237 ipo->ddir = DDIR_WRITE;
238 else
239 ipo->ddir = DDIR_READ;
240 ipo->fileno = fileno;
241
242 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
243 ipo->ddir, ipo->offset,
244 ipo->len, ipo->delay);
245 queue_io_piece(td, ipo);
246 }
247
handle_trace_notify(struct blk_io_trace * t)248 static void handle_trace_notify(struct blk_io_trace *t)
249 {
250 switch (t->action) {
251 case BLK_TN_PROCESS:
252 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
253 t->action, t->pid);
254 break;
255 case BLK_TN_TIMESTAMP:
256 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
257 t->action, t->pid);
258 break;
259 case BLK_TN_MESSAGE:
260 break;
261 default:
262 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
263 break;
264 }
265 }
266
handle_trace_discard(struct thread_data * td,struct blk_io_trace * t,unsigned long long ttime,unsigned long * ios,unsigned int * rw_bs)267 static void handle_trace_discard(struct thread_data *td,
268 struct blk_io_trace *t,
269 unsigned long long ttime,
270 unsigned long *ios, unsigned int *rw_bs)
271 {
272 struct io_piece *ipo = malloc(sizeof(*ipo));
273 unsigned int bs;
274 int fileno;
275
276 init_ipo(ipo);
277 fileno = trace_add_file(td, t->device, &bs);
278
279 ios[DDIR_TRIM]++;
280 if (t->bytes > rw_bs[DDIR_TRIM])
281 rw_bs[DDIR_TRIM] = t->bytes;
282
283 td->o.size += t->bytes;
284
285 memset(ipo, 0, sizeof(*ipo));
286 INIT_FLIST_HEAD(&ipo->list);
287
288 ipo->offset = t->sector * bs;
289 if (td->o.replay_scale)
290 ipo->offset = ipo->offset / td->o.replay_scale;
291 ipo_bytes_align(td->o.replay_align, ipo);
292 ipo->len = t->bytes;
293 ipo->delay = ttime / 1000;
294 ipo->ddir = DDIR_TRIM;
295 ipo->fileno = fileno;
296
297 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
298 ipo->offset, ipo->len,
299 ipo->delay);
300 queue_io_piece(td, ipo);
301 }
302
handle_trace_fs(struct thread_data * td,struct blk_io_trace * t,unsigned long long ttime,unsigned long * ios,unsigned int * rw_bs)303 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
304 unsigned long long ttime, unsigned long *ios,
305 unsigned int *rw_bs)
306 {
307 unsigned int bs;
308 int rw;
309 int fileno;
310
311 fileno = trace_add_file(td, t->device, &bs);
312
313 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
314
315 if (t->bytes > rw_bs[rw])
316 rw_bs[rw] = t->bytes;
317
318 ios[rw]++;
319 td->o.size += t->bytes;
320 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
321 }
322
323 /*
324 * We only care for queue traces, most of the others are side effects
325 * due to internal workings of the block layer.
326 */
handle_trace(struct thread_data * td,struct blk_io_trace * t,unsigned long * ios,unsigned int * bs)327 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
328 unsigned long *ios, unsigned int *bs)
329 {
330 static unsigned long long last_ttime;
331 unsigned long long delay = 0;
332
333 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
334 return;
335
336 if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
337 if (!last_ttime || td->o.no_stall) {
338 last_ttime = t->time;
339 delay = 0;
340 } else {
341 delay = t->time - last_ttime;
342 last_ttime = t->time;
343 }
344 }
345
346 t_bytes_align(&td->o, t);
347
348 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
349 handle_trace_notify(t);
350 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
351 handle_trace_discard(td, t, delay, ios, bs);
352 else
353 handle_trace_fs(td, t, delay, ios, bs);
354 }
355
byteswap_trace(struct blk_io_trace * t)356 static void byteswap_trace(struct blk_io_trace *t)
357 {
358 t->magic = fio_swap32(t->magic);
359 t->sequence = fio_swap32(t->sequence);
360 t->time = fio_swap64(t->time);
361 t->sector = fio_swap64(t->sector);
362 t->bytes = fio_swap32(t->bytes);
363 t->action = fio_swap32(t->action);
364 t->pid = fio_swap32(t->pid);
365 t->device = fio_swap32(t->device);
366 t->cpu = fio_swap32(t->cpu);
367 t->error = fio_swap16(t->error);
368 t->pdu_len = fio_swap16(t->pdu_len);
369 }
370
t_is_write(struct blk_io_trace * t)371 static int t_is_write(struct blk_io_trace *t)
372 {
373 return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
374 }
375
t_get_ddir(struct blk_io_trace * t)376 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
377 {
378 if (t->action & BLK_TC_ACT(BLK_TC_READ))
379 return DDIR_READ;
380 else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
381 return DDIR_WRITE;
382 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
383 return DDIR_TRIM;
384
385 return DDIR_INVAL;
386 }
387
depth_inc(struct blk_io_trace * t,int * depth)388 static void depth_inc(struct blk_io_trace *t, int *depth)
389 {
390 enum fio_ddir ddir;
391
392 ddir = t_get_ddir(t);
393 if (ddir != DDIR_INVAL)
394 depth[ddir]++;
395 }
396
depth_dec(struct blk_io_trace * t,int * depth)397 static void depth_dec(struct blk_io_trace *t, int *depth)
398 {
399 enum fio_ddir ddir;
400
401 ddir = t_get_ddir(t);
402 if (ddir != DDIR_INVAL)
403 depth[ddir]--;
404 }
405
depth_end(struct blk_io_trace * t,int * this_depth,int * depth)406 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
407 {
408 enum fio_ddir ddir = DDIR_INVAL;
409
410 ddir = t_get_ddir(t);
411 if (ddir != DDIR_INVAL) {
412 depth[ddir] = max(depth[ddir], this_depth[ddir]);
413 this_depth[ddir] = 0;
414 }
415 }
416
417 /*
418 * Load a blktrace file by reading all the blk_io_trace entries, and storing
419 * them as io_pieces like the fio text version would do.
420 */
load_blktrace(struct thread_data * td,const char * filename,int need_swap)421 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
422 {
423 struct blk_io_trace t;
424 unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
425 unsigned int rw_bs[DDIR_RWDIR_CNT];
426 struct fifo *fifo;
427 int fd, i, old_state;
428 struct fio_file *f;
429 int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
430
431 fd = open(filename, O_RDONLY);
432 if (fd < 0) {
433 td_verror(td, errno, "open blktrace file");
434 return 1;
435 }
436
437 fifo = fifo_alloc(TRACE_FIFO_SIZE);
438
439 old_state = td_bump_runstate(td, TD_SETTING_UP);
440
441 td->o.size = 0;
442
443 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
444 ios[i] = 0;
445 rw_bs[i] = 0;
446 this_depth[i] = 0;
447 depth[i] = 0;
448 }
449
450 skipped_writes = 0;
451 do {
452 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
453
454 if (ret < 0)
455 goto err;
456 else if (!ret)
457 break;
458 else if (ret < (int) sizeof(t)) {
459 log_err("fio: short fifo get\n");
460 break;
461 }
462
463 if (need_swap)
464 byteswap_trace(&t);
465
466 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
467 log_err("fio: bad magic in blktrace data: %x\n",
468 t.magic);
469 goto err;
470 }
471 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
472 log_err("fio: bad blktrace version %d\n",
473 t.magic & 0xff);
474 goto err;
475 }
476 ret = discard_pdu(td, fifo, fd, &t);
477 if (ret < 0) {
478 td_verror(td, ret, "blktrace lseek");
479 goto err;
480 } else if (t.pdu_len != ret) {
481 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
482 goto err;
483 }
484 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
485 if ((t.action & 0xffff) == __BLK_TA_QUEUE)
486 depth_inc(&t, this_depth);
487 else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
488 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
489 depth_dec(&t, this_depth);
490 else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
491 depth_end(&t, this_depth, depth);
492
493 if (t_is_write(&t) && read_only) {
494 skipped_writes++;
495 continue;
496 }
497 }
498
499 handle_trace(td, &t, ios, rw_bs);
500 } while (1);
501
502 for (i = 0; i < td->files_index; i++) {
503 f = td->files[i];
504 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
505 }
506
507 fifo_free(fifo);
508 close(fd);
509
510 td_restore_runstate(td, old_state);
511
512 if (!td->files_index) {
513 log_err("fio: did not find replay device(s)\n");
514 return 1;
515 }
516
517 /*
518 * For stacked devices, we don't always get a COMPLETE event so
519 * the depth grows to insane values. Limit it to something sane(r).
520 */
521 max_depth = 0;
522 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
523 if (depth[i] > 1024)
524 depth[i] = 1024;
525 else if (!depth[i] && ios[i])
526 depth[i] = 1;
527 max_depth = max(depth[i], max_depth);
528 }
529
530 if (skipped_writes)
531 log_err("fio: %s skips replay of %lu writes due to read-only\n",
532 td->o.name, skipped_writes);
533
534 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
535 log_err("fio: found no ios in blktrace data\n");
536 return 1;
537 } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
538 td->o.td_ddir = TD_DDIR_READ;
539 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
540 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
541 td->o.td_ddir = TD_DDIR_WRITE;
542 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
543 } else {
544 td->o.td_ddir = TD_DDIR_RW;
545 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
546 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
547 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
548 }
549
550 /*
551 * We need to do direct/raw ios to the device, to avoid getting
552 * read-ahead in our way. But only do so if the minimum block size
553 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
554 */
555 if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
556 td->o.odirect = 1;
557
558 /*
559 * If depth wasn't manually set, use probed depth
560 */
561 if (!fio_option_is_set(&td->o, iodepth))
562 td->o.iodepth = td->o.iodepth_low = max_depth;
563
564 return 0;
565 err:
566 close(fd);
567 fifo_free(fifo);
568 return 1;
569 }
570