Commit | Line | Data |
---|---|---|
fb7b71a3 JA |
1 | /* |
2 | * blktrace support code for fio | |
3 | */ | |
4 | #include <stdio.h> | |
5 | #include <stdlib.h> | |
5e6c2067 JA |
6 | #include <sys/stat.h> |
7 | #include <dirent.h> | |
8c1fdf04 | 8 | |
01743ee1 | 9 | #include "flist.h" |
fb7b71a3 JA |
10 | #include "fio.h" |
11 | #include "blktrace_api.h" | |
40bafa33 | 12 | #include "lib/linux-dev-lookup.h" |
fb7b71a3 | 13 | |
2da7df1d | 14 | #define TRACE_FIFO_SIZE 8192 |
e2887563 JA |
15 | |
16 | /* | |
17 | * fifo refill frontend, to avoid reading data in trace sized bites | |
18 | */ | |
19 | static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd) | |
20 | { | |
21 | char buf[TRACE_FIFO_SIZE]; | |
f12b323f | 22 | unsigned int total; |
e2887563 JA |
23 | int ret; |
24 | ||
f12b323f JA |
25 | total = sizeof(buf); |
26 | if (total > fifo_room(fifo)) | |
27 | total = fifo_room(fifo); | |
e2887563 | 28 | |
f12b323f JA |
29 | ret = read(fd, buf, total); |
30 | if (ret < 0) { | |
31 | td_verror(td, errno, "read blktrace file"); | |
32 | return -1; | |
e2887563 JA |
33 | } |
34 | ||
f12b323f JA |
35 | if (ret > 0) |
36 | ret = fifo_put(fifo, buf, ret); | |
37 | ||
bd6f78b2 | 38 | dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret); |
f12b323f | 39 | return ret; |
e2887563 JA |
40 | } |
41 | ||
42 | /* | |
43 | * Retrieve 'len' bytes from the fifo, refilling if necessary. | |
44 | */ | |
45 | static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd, | |
46 | void *buf, unsigned int len) | |
47 | { | |
f12b323f JA |
48 | if (fifo_len(fifo) < len) { |
49 | int ret = refill_fifo(td, fifo, fd); | |
e2887563 | 50 | |
f12b323f JA |
51 | if (ret < 0) |
52 | return ret; | |
53 | } | |
e2887563 JA |
54 | |
55 | return fifo_get(fifo, buf, len); | |
56 | } | |
57 | ||
8c1fdf04 JA |
58 | /* |
59 | * Just discard the pdu by seeking past it. | |
60 | */ | |
f12b323f JA |
61 | static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd, |
62 | struct blk_io_trace *t) | |
fb7b71a3 JA |
63 | { |
64 | if (t->pdu_len == 0) | |
65 | return 0; | |
66 | ||
bd6f78b2 | 67 | dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len); |
f12b323f | 68 | return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len); |
fb7b71a3 JA |
69 | } |
70 | ||
8c1fdf04 JA |
71 | /* |
72 | * Check if this is a blktrace binary data file. We read a single trace | |
73 | * into memory and check for the magic signature. | |
74 | */ | |
d95b34a6 | 75 | int is_blktrace(const char *filename, int *need_swap) |
fb7b71a3 JA |
76 | { |
77 | struct blk_io_trace t; | |
78 | int fd, ret; | |
79 | ||
80 | fd = open(filename, O_RDONLY); | |
4dced407 | 81 | if (fd < 0) |
fb7b71a3 | 82 | return 0; |
fb7b71a3 JA |
83 | |
84 | ret = read(fd, &t, sizeof(t)); | |
85 | close(fd); | |
86 | ||
87 | if (ret < 0) { | |
88 | perror("read blktrace"); | |
89 | return 0; | |
90 | } else if (ret != sizeof(t)) { | |
91 | log_err("fio: short read on blktrace file\n"); | |
92 | return 0; | |
93 | } | |
94 | ||
d95b34a6 JA |
95 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) { |
96 | *need_swap = 0; | |
fb7b71a3 | 97 | return 1; |
d95b34a6 JA |
98 | } |
99 | ||
100 | /* | |
101 | * Maybe it needs to be endian swapped... | |
102 | */ | |
103 | t.magic = fio_swap32(t.magic); | |
104 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) { | |
105 | *need_swap = 1; | |
106 | return 1; | |
107 | } | |
fb7b71a3 JA |
108 | |
109 | return 0; | |
110 | } | |
111 | ||
c69aa91f JA |
112 | #define FMINORBITS 20 |
113 | #define FMINORMASK ((1U << FMINORBITS) - 1) | |
114 | #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) | |
115 | #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK)) | |
eeb9c2aa | 116 | |
89ac1d48 | 117 | static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action) |
691c8fb0 JA |
118 | { |
119 | struct io_piece *ipo; | |
120 | ||
121 | ipo = calloc(1, sizeof(*ipo)); | |
0d29de83 | 122 | init_ipo(ipo); |
691c8fb0 JA |
123 | |
124 | ipo->ddir = DDIR_INVAL; | |
125 | ipo->fileno = fileno; | |
89ac1d48 | 126 | ipo->file_action = action; |
01743ee1 | 127 | flist_add_tail(&ipo->list, &td->io_log_list); |
691c8fb0 JA |
128 | } |
129 | ||
89ac1d48 | 130 | static int trace_add_file(struct thread_data *td, __u32 device) |
5e6c2067 | 131 | { |
89ac1d48 | 132 | static unsigned int last_maj, last_min, last_fileno; |
c69aa91f JA |
133 | unsigned int maj = FMAJOR(device); |
134 | unsigned int min = FMINOR(device); | |
5e6c2067 JA |
135 | struct fio_file *f; |
136 | char dev[256]; | |
137 | unsigned int i; | |
138 | ||
139 | if (last_maj == maj && last_min == min) | |
89ac1d48 | 140 | return last_fileno; |
5e6c2067 JA |
141 | |
142 | last_maj = maj; | |
143 | last_min = min; | |
144 | ||
145 | /* | |
146 | * check for this file in our list | |
147 | */ | |
148 | for_each_file(td, f, i) | |
89ac1d48 SL |
149 | if (f->major == maj && f->minor == min) { |
150 | last_fileno = f->fileno; | |
151 | return last_fileno; | |
152 | } | |
5e6c2067 JA |
153 | |
154 | strcpy(dev, "/dev"); | |
40bafa33 | 155 | if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) { |
691c8fb0 JA |
156 | int fileno; |
157 | ||
40bafa33 JA |
158 | if (td->o.replay_redirect) |
159 | dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden" | |
160 | " with: %s\n", maj, min, | |
161 | td->o.replay_redirect); | |
162 | else | |
163 | dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min); | |
164 | ||
bd6f78b2 | 165 | dprint(FD_BLKTRACE, "add devices %s\n", dev); |
49ffb4a2 | 166 | fileno = add_file_exclusive(td, dev); |
b53f2c54 | 167 | td->o.open_files++; |
5903e7b7 JA |
168 | td->files[fileno]->major = maj; |
169 | td->files[fileno]->minor = min; | |
89ac1d48 SL |
170 | trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE); |
171 | last_fileno = fileno; | |
bd6f78b2 | 172 | } |
f01b34ae | 173 | |
89ac1d48 | 174 | return last_fileno; |
5e6c2067 JA |
175 | } |
176 | ||
8c1fdf04 JA |
177 | /* |
178 | * Store blk_io_trace data in an ipo for later retrieval. | |
179 | */ | |
fdefd987 | 180 | static void store_ipo(struct thread_data *td, unsigned long long offset, |
89ac1d48 SL |
181 | unsigned int bytes, int rw, unsigned long long ttime, |
182 | int fileno) | |
fdefd987 JA |
183 | { |
184 | struct io_piece *ipo = malloc(sizeof(*ipo)); | |
185 | ||
0d29de83 JA |
186 | init_ipo(ipo); |
187 | ||
a2eea81b JA |
188 | /* |
189 | * the 512 is wrong here, it should be the hardware sector size... | |
190 | */ | |
191 | ipo->offset = offset * 512; | |
fdefd987 | 192 | ipo->len = bytes; |
8c1fdf04 | 193 | ipo->delay = ttime / 1000; |
fdefd987 JA |
194 | if (rw) |
195 | ipo->ddir = DDIR_WRITE; | |
196 | else | |
197 | ipo->ddir = DDIR_READ; | |
89ac1d48 | 198 | ipo->fileno = fileno; |
fdefd987 | 199 | |
bd6f78b2 JA |
200 | dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n", |
201 | ipo->ddir, ipo->offset, | |
202 | ipo->len, ipo->delay); | |
691c8fb0 | 203 | queue_io_piece(td, ipo); |
fdefd987 JA |
204 | } |
205 | ||
0b9d69ec | 206 | static void handle_trace_notify(struct blk_io_trace *t) |
cd991b9e | 207 | { |
691c8fb0 JA |
208 | switch (t->action) { |
209 | case BLK_TN_PROCESS: | |
24653680 | 210 | dprint(FD_BLKTRACE, "got process notify: %x, %d\n", |
d95b34a6 | 211 | t->action, t->pid); |
691c8fb0 JA |
212 | break; |
213 | case BLK_TN_TIMESTAMP: | |
24653680 | 214 | dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n", |
d95b34a6 | 215 | t->action, t->pid); |
691c8fb0 | 216 | break; |
ff58fced JA |
217 | case BLK_TN_MESSAGE: |
218 | break; | |
691c8fb0 JA |
219 | default: |
220 | dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action); | |
221 | break; | |
222 | } | |
223 | } | |
5b3023b8 | 224 | |
24653680 JA |
225 | static void handle_trace_discard(struct thread_data *td, |
226 | struct blk_io_trace *t, | |
227 | unsigned long long ttime, | |
228 | unsigned long *ios, unsigned int *bs) | |
ff58fced JA |
229 | { |
230 | struct io_piece *ipo = malloc(sizeof(*ipo)); | |
89ac1d48 | 231 | int fileno; |
ff58fced | 232 | |
0d29de83 | 233 | init_ipo(ipo); |
89ac1d48 | 234 | fileno = trace_add_file(td, t->device); |
ff58fced | 235 | |
24653680 JA |
236 | ios[DDIR_TRIM]++; |
237 | if (t->bytes > bs[DDIR_TRIM]) | |
238 | bs[DDIR_TRIM] = t->bytes; | |
239 | ||
ff58fced JA |
240 | td->o.size += t->bytes; |
241 | ||
242 | memset(ipo, 0, sizeof(*ipo)); | |
243 | INIT_FLIST_HEAD(&ipo->list); | |
244 | ||
245 | /* | |
246 | * the 512 is wrong here, it should be the hardware sector size... | |
247 | */ | |
248 | ipo->offset = t->sector * 512; | |
249 | ipo->len = t->bytes; | |
250 | ipo->delay = ttime / 1000; | |
251 | ipo->ddir = DDIR_TRIM; | |
89ac1d48 | 252 | ipo->fileno = fileno; |
ff58fced JA |
253 | |
254 | dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n", | |
255 | ipo->offset, ipo->len, | |
256 | ipo->delay); | |
257 | queue_io_piece(td, ipo); | |
258 | } | |
259 | ||
691c8fb0 JA |
260 | static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t, |
261 | unsigned long long ttime, unsigned long *ios, | |
262 | unsigned int *bs) | |
263 | { | |
264 | int rw; | |
89ac1d48 | 265 | int fileno; |
5b3023b8 | 266 | |
89ac1d48 | 267 | fileno = trace_add_file(td, t->device); |
5b3023b8 JA |
268 | |
269 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; | |
270 | ||
271 | if (t->bytes > bs[rw]) | |
272 | bs[rw] = t->bytes; | |
273 | ||
274 | ios[rw]++; | |
275 | td->o.size += t->bytes; | |
89ac1d48 | 276 | store_ipo(td, t->sector, t->bytes, rw, ttime, fileno); |
cd991b9e JA |
277 | } |
278 | ||
691c8fb0 JA |
279 | /* |
280 | * We only care for queue traces, most of the others are side effects | |
281 | * due to internal workings of the block layer. | |
282 | */ | |
283 | static void handle_trace(struct thread_data *td, struct blk_io_trace *t, | |
24653680 | 284 | unsigned long *ios, unsigned int *bs) |
691c8fb0 | 285 | { |
24653680 JA |
286 | static unsigned long long last_ttime; |
287 | unsigned long long delay; | |
288 | ||
691c8fb0 JA |
289 | if ((t->action & 0xffff) != __BLK_TA_QUEUE) |
290 | return; | |
24653680 JA |
291 | |
292 | if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { | |
293 | if (!last_ttime || td->o.no_stall) { | |
294 | last_ttime = t->time; | |
295 | delay = 0; | |
296 | } else { | |
297 | delay = t->time - last_ttime; | |
298 | last_ttime = t->time; | |
299 | } | |
300 | } | |
691c8fb0 JA |
301 | |
302 | if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY)) | |
aec2de20 | 303 | handle_trace_notify(t); |
ff58fced | 304 | else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) |
24653680 | 305 | handle_trace_discard(td, t, delay, ios, bs); |
691c8fb0 | 306 | else |
24653680 | 307 | handle_trace_fs(td, t, delay, ios, bs); |
691c8fb0 JA |
308 | } |
309 | ||
d95b34a6 JA |
310 | static void byteswap_trace(struct blk_io_trace *t) |
311 | { | |
312 | t->magic = fio_swap32(t->magic); | |
313 | t->sequence = fio_swap32(t->sequence); | |
314 | t->time = fio_swap64(t->time); | |
315 | t->sector = fio_swap64(t->sector); | |
316 | t->bytes = fio_swap32(t->bytes); | |
317 | t->action = fio_swap32(t->action); | |
318 | t->pid = fio_swap32(t->pid); | |
319 | t->device = fio_swap32(t->device); | |
320 | t->cpu = fio_swap32(t->cpu); | |
321 | t->error = fio_swap16(t->error); | |
322 | t->pdu_len = fio_swap16(t->pdu_len); | |
323 | } | |
324 | ||
24653680 JA |
325 | static int t_is_write(struct blk_io_trace *t) |
326 | { | |
327 | return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0; | |
328 | } | |
329 | ||
8c1fdf04 JA |
330 | /* |
331 | * Load a blktrace file by reading all the blk_io_trace entries, and storing | |
332 | * them as io_pieces like the fio text version would do. | |
333 | */ | |
d95b34a6 | 334 | int load_blktrace(struct thread_data *td, const char *filename, int need_swap) |
fb7b71a3 JA |
335 | { |
336 | struct blk_io_trace t; | |
24653680 JA |
337 | unsigned long ios[DDIR_RWDIR_CNT], skipped_writes; |
338 | unsigned int rw_bs[DDIR_RWDIR_CNT]; | |
e2887563 | 339 | struct fifo *fifo; |
5903e7b7 | 340 | int fd, i, old_state; |
89ac1d48 | 341 | struct fio_file *f; |
eb5fdcf1 | 342 | int this_depth, depth; |
fb7b71a3 JA |
343 | |
344 | fd = open(filename, O_RDONLY); | |
345 | if (fd < 0) { | |
346 | td_verror(td, errno, "open blktrace file"); | |
347 | return 1; | |
348 | } | |
349 | ||
e2887563 JA |
350 | fifo = fifo_alloc(TRACE_FIFO_SIZE); |
351 | ||
8edd973d | 352 | old_state = td_bump_runstate(td, TD_SETTING_UP); |
5903e7b7 | 353 | |
6df8adaa JA |
354 | td->o.size = 0; |
355 | ||
d84f8d49 JA |
356 | ios[0] = ios[1] = 0; |
357 | rw_bs[0] = rw_bs[1] = 0; | |
4241ea8f | 358 | skipped_writes = 0; |
eb5fdcf1 | 359 | this_depth = depth = 0; |
fb7b71a3 | 360 | do { |
e2887563 | 361 | int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); |
fb7b71a3 | 362 | |
e2887563 | 363 | if (ret < 0) |
8c1fdf04 | 364 | goto err; |
e2887563 JA |
365 | else if (!ret) |
366 | break; | |
367 | else if (ret < (int) sizeof(t)) { | |
368 | log_err("fio: short fifo get\n"); | |
fb7b71a3 | 369 | break; |
fb7b71a3 JA |
370 | } |
371 | ||
d95b34a6 JA |
372 | if (need_swap) |
373 | byteswap_trace(&t); | |
374 | ||
fb7b71a3 | 375 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { |
5ec10eaa JA |
376 | log_err("fio: bad magic in blktrace data: %x\n", |
377 | t.magic); | |
8c1fdf04 | 378 | goto err; |
fb7b71a3 JA |
379 | } |
380 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { | |
5ec10eaa JA |
381 | log_err("fio: bad blktrace version %d\n", |
382 | t.magic & 0xff); | |
8c1fdf04 | 383 | goto err; |
fb7b71a3 | 384 | } |
f12b323f JA |
385 | ret = discard_pdu(td, fifo, fd, &t); |
386 | if (ret < 0) { | |
fb7b71a3 | 387 | td_verror(td, ret, "blktrace lseek"); |
8c1fdf04 | 388 | goto err; |
f12b323f JA |
389 | } else if (t.pdu_len != ret) { |
390 | log_err("fio: discarded %d of %d\n", ret, t.pdu_len); | |
391 | goto err; | |
fb7b71a3 | 392 | } |
691c8fb0 | 393 | if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) { |
eb5fdcf1 JA |
394 | if ((t.action & 0xffff) == __BLK_TA_QUEUE) |
395 | this_depth++; | |
396 | else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) { | |
397 | depth = max(depth, this_depth); | |
398 | this_depth = 0; | |
399 | } | |
691c8fb0 | 400 | |
24653680 | 401 | if (t_is_write(&t) && read_only) { |
691c8fb0 | 402 | skipped_writes++; |
24653680 | 403 | continue; |
64bbb865 | 404 | } |
a6edd638 | 405 | } |
24653680 JA |
406 | |
407 | handle_trace(td, &t, ios, rw_bs); | |
fb7b71a3 JA |
408 | } while (1); |
409 | ||
85a47ca2 | 410 | for (i = 0; i < td->files_index; i++) { |
f01b34ae | 411 | f = td->files[i]; |
89ac1d48 | 412 | trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE); |
85a47ca2 | 413 | } |
89ac1d48 | 414 | |
38470f85 | 415 | fifo_free(fifo); |
fb7b71a3 | 416 | close(fd); |
8c1fdf04 | 417 | |
8edd973d | 418 | td_restore_runstate(td, old_state); |
5903e7b7 | 419 | |
f01b34ae JA |
420 | if (!td->files_index) { |
421 | log_err("fio: did not find replay device(s)\n"); | |
422 | return 1; | |
423 | } | |
424 | ||
eb5fdcf1 JA |
425 | /* |
426 | * For stacked devices, we don't always get a COMPLETE event so | |
427 | * the depth grows to insane values. Limit it to something sane(r). | |
428 | */ | |
429 | if (!depth || depth > 1024) | |
430 | depth = 1024; | |
431 | ||
4241ea8f | 432 | if (skipped_writes) |
5ec10eaa JA |
433 | log_err("fio: %s skips replay of %lu writes due to read-only\n", |
434 | td->o.name, skipped_writes); | |
4241ea8f | 435 | |
8c1fdf04 JA |
436 | if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
437 | log_err("fio: found no ios in blktrace data\n"); | |
438 | return 1; | |
24653680 | 439 | } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
8c1fdf04 | 440 | td->o.td_ddir = TD_DDIR_READ; |
d84f8d49 JA |
441 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
442 | } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { | |
8c1fdf04 | 443 | td->o.td_ddir = TD_DDIR_WRITE; |
d84f8d49 JA |
444 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
445 | } else { | |
8c1fdf04 | 446 | td->o.td_ddir = TD_DDIR_RW; |
d84f8d49 JA |
447 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
448 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; | |
24653680 | 449 | td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM]; |
d84f8d49 | 450 | } |
8c1fdf04 JA |
451 | |
452 | /* | |
453 | * We need to do direct/raw ios to the device, to avoid getting | |
454 | * read-ahead in our way. | |
455 | */ | |
456 | td->o.odirect = 1; | |
457 | ||
eb5fdcf1 JA |
458 | /* |
459 | * we don't know if this option was set or not. it defaults to 1, | |
460 | * so we'll just guess that we should override it if it's still 1 | |
461 | */ | |
0a5d0ddc JA |
462 | if (td->o.iodepth == 1) |
463 | td->o.iodepth = td->o.iodepth_low = depth; | |
eb5fdcf1 | 464 | |
fb7b71a3 | 465 | return 0; |
8c1fdf04 JA |
466 | err: |
467 | close(fd); | |
38470f85 | 468 | fifo_free(fifo); |
8c1fdf04 | 469 | return 1; |
fb7b71a3 | 470 | } |