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 | |
fb7b71a3 JA |
9 | #include "list.h" |
10 | #include "fio.h" | |
11 | #include "blktrace_api.h" | |
12 | ||
f12b323f | 13 | #define TRACE_FIFO_SIZE 65536 |
e2887563 JA |
14 | |
15 | /* | |
16 | * fifo refill frontend, to avoid reading data in trace sized bites | |
17 | */ | |
18 | static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd) | |
19 | { | |
20 | char buf[TRACE_FIFO_SIZE]; | |
f12b323f | 21 | unsigned int total; |
e2887563 JA |
22 | int ret; |
23 | ||
f12b323f JA |
24 | total = sizeof(buf); |
25 | if (total > fifo_room(fifo)) | |
26 | total = fifo_room(fifo); | |
e2887563 | 27 | |
f12b323f JA |
28 | ret = read(fd, buf, total); |
29 | if (ret < 0) { | |
30 | td_verror(td, errno, "read blktrace file"); | |
31 | return -1; | |
e2887563 JA |
32 | } |
33 | ||
f12b323f JA |
34 | if (ret > 0) |
35 | ret = fifo_put(fifo, buf, ret); | |
36 | ||
bd6f78b2 | 37 | dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret); |
f12b323f | 38 | return ret; |
e2887563 JA |
39 | } |
40 | ||
41 | /* | |
42 | * Retrieve 'len' bytes from the fifo, refilling if necessary. | |
43 | */ | |
44 | static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd, | |
45 | void *buf, unsigned int len) | |
46 | { | |
f12b323f JA |
47 | if (fifo_len(fifo) < len) { |
48 | int ret = refill_fifo(td, fifo, fd); | |
e2887563 | 49 | |
f12b323f JA |
50 | if (ret < 0) |
51 | return ret; | |
52 | } | |
e2887563 JA |
53 | |
54 | return fifo_get(fifo, buf, len); | |
55 | } | |
56 | ||
8c1fdf04 JA |
57 | /* |
58 | * Just discard the pdu by seeking past it. | |
59 | */ | |
f12b323f JA |
60 | static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd, |
61 | struct blk_io_trace *t) | |
fb7b71a3 JA |
62 | { |
63 | if (t->pdu_len == 0) | |
64 | return 0; | |
65 | ||
bd6f78b2 | 66 | dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len); |
f12b323f | 67 | return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len); |
fb7b71a3 JA |
68 | } |
69 | ||
8c1fdf04 JA |
70 | /* |
71 | * Check if this is a blktrace binary data file. We read a single trace | |
72 | * into memory and check for the magic signature. | |
73 | */ | |
fb7b71a3 JA |
74 | int is_blktrace(const char *filename) |
75 | { | |
76 | struct blk_io_trace t; | |
77 | int fd, ret; | |
78 | ||
79 | fd = open(filename, O_RDONLY); | |
4dced407 | 80 | if (fd < 0) |
fb7b71a3 | 81 | return 0; |
fb7b71a3 JA |
82 | |
83 | ret = read(fd, &t, sizeof(t)); | |
84 | close(fd); | |
85 | ||
86 | if (ret < 0) { | |
87 | perror("read blktrace"); | |
88 | return 0; | |
89 | } else if (ret != sizeof(t)) { | |
90 | log_err("fio: short read on blktrace file\n"); | |
91 | return 0; | |
92 | } | |
93 | ||
94 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) | |
95 | return 1; | |
96 | ||
97 | return 0; | |
98 | } | |
99 | ||
5e6c2067 JA |
100 | static int lookup_device(char *path, unsigned int maj, unsigned int min) |
101 | { | |
102 | struct dirent *dir; | |
103 | struct stat st; | |
104 | int found = 0; | |
105 | DIR *D; | |
106 | ||
107 | D = opendir(path); | |
108 | if (!D) | |
109 | return 0; | |
110 | ||
111 | while ((dir = readdir(D)) != NULL) { | |
112 | char full_path[256]; | |
113 | ||
114 | if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) | |
115 | continue; | |
116 | ||
117 | sprintf(full_path, "%s/%s", path, dir->d_name); | |
118 | if (lstat(full_path, &st) == -1) { | |
119 | perror("lstat"); | |
120 | break; | |
121 | } | |
122 | ||
123 | if (S_ISDIR(st.st_mode)) { | |
124 | found = lookup_device(full_path, maj, min); | |
125 | if (found) { | |
126 | strcpy(path, full_path); | |
127 | break; | |
128 | } | |
129 | } | |
130 | ||
131 | if (!S_ISBLK(st.st_mode)) | |
132 | continue; | |
133 | ||
134 | if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) { | |
5ec10eaa | 135 | dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min); |
5e6c2067 JA |
136 | strcpy(path, full_path); |
137 | found = 1; | |
138 | break; | |
139 | } | |
140 | } | |
141 | ||
142 | closedir(D); | |
143 | return found; | |
144 | } | |
145 | ||
c69aa91f JA |
146 | #define FMINORBITS 20 |
147 | #define FMINORMASK ((1U << FMINORBITS) - 1) | |
148 | #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) | |
149 | #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK)) | |
eeb9c2aa | 150 | |
5e6c2067 JA |
151 | static void trace_add_file(struct thread_data *td, __u32 device) |
152 | { | |
153 | static unsigned int last_maj, last_min; | |
c69aa91f JA |
154 | unsigned int maj = FMAJOR(device); |
155 | unsigned int min = FMINOR(device); | |
5e6c2067 JA |
156 | struct fio_file *f; |
157 | char dev[256]; | |
158 | unsigned int i; | |
159 | ||
160 | if (last_maj == maj && last_min == min) | |
161 | return; | |
162 | ||
163 | last_maj = maj; | |
164 | last_min = min; | |
165 | ||
166 | /* | |
167 | * check for this file in our list | |
168 | */ | |
169 | for_each_file(td, f, i) | |
170 | if (f->major == maj && f->minor == min) | |
171 | return; | |
172 | ||
173 | strcpy(dev, "/dev"); | |
bd6f78b2 JA |
174 | if (lookup_device(dev, maj, min)) { |
175 | dprint(FD_BLKTRACE, "add devices %s\n", dev); | |
5e6c2067 | 176 | add_file(td, dev); |
bd6f78b2 | 177 | } |
5e6c2067 JA |
178 | } |
179 | ||
8c1fdf04 JA |
180 | /* |
181 | * Store blk_io_trace data in an ipo for later retrieval. | |
182 | */ | |
fdefd987 | 183 | static void store_ipo(struct thread_data *td, unsigned long long offset, |
8c1fdf04 | 184 | unsigned int bytes, int rw, unsigned long long ttime) |
fdefd987 JA |
185 | { |
186 | struct io_piece *ipo = malloc(sizeof(*ipo)); | |
187 | ||
188 | memset(ipo, 0, sizeof(*ipo)); | |
189 | INIT_LIST_HEAD(&ipo->list); | |
a2eea81b JA |
190 | /* |
191 | * the 512 is wrong here, it should be the hardware sector size... | |
192 | */ | |
193 | ipo->offset = offset * 512; | |
fdefd987 | 194 | ipo->len = bytes; |
8c1fdf04 | 195 | ipo->delay = ttime / 1000; |
fdefd987 JA |
196 | if (rw) |
197 | ipo->ddir = DDIR_WRITE; | |
198 | else | |
199 | ipo->ddir = DDIR_READ; | |
200 | ||
bd6f78b2 JA |
201 | dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n", |
202 | ipo->ddir, ipo->offset, | |
203 | ipo->len, ipo->delay); | |
fdefd987 JA |
204 | list_add_tail(&ipo->list, &td->io_log_list); |
205 | } | |
206 | ||
cd991b9e | 207 | static void handle_trace_notify(struct thread_data *td, struct blk_io_trace *t) |
fb7b71a3 | 208 | { |
cd991b9e JA |
209 | printf("got notify: %x, %d\n", t->action, t->pid); |
210 | } | |
fdefd987 | 211 | |
cd991b9e JA |
212 | static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t, |
213 | unsigned long long ttime, unsigned long *ios, | |
214 | unsigned int *bs) | |
215 | { | |
216 | int rw; | |
fdefd987 | 217 | |
5e6c2067 JA |
218 | trace_add_file(td, t->device); |
219 | ||
e7a7d70b | 220 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
d84f8d49 JA |
221 | |
222 | if (t->bytes > bs[rw]) | |
223 | bs[rw] = t->bytes; | |
224 | ||
8c1fdf04 | 225 | ios[rw]++; |
6df8adaa | 226 | td->o.size += t->bytes; |
8c1fdf04 | 227 | store_ipo(td, t->sector, t->bytes, rw, ttime); |
fb7b71a3 JA |
228 | } |
229 | ||
cd991b9e JA |
230 | /* |
231 | * We only care for queue traces, most of the others are side effects | |
232 | * due to internal workings of the block layer. | |
233 | */ | |
234 | static void handle_trace(struct thread_data *td, struct blk_io_trace *t, | |
235 | unsigned long long ttime, unsigned long *ios, | |
236 | unsigned int *bs) | |
237 | { | |
238 | if ((t->action & 0xffff) != __BLK_TA_QUEUE) | |
239 | return; | |
240 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) | |
241 | return; | |
242 | ||
243 | if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY)) | |
244 | handle_trace_notify(td, t); | |
245 | else | |
246 | handle_trace_fs(td, t, ttime, ios, bs); | |
247 | } | |
248 | ||
8c1fdf04 JA |
249 | /* |
250 | * Load a blktrace file by reading all the blk_io_trace entries, and storing | |
251 | * them as io_pieces like the fio text version would do. | |
252 | */ | |
fb7b71a3 JA |
253 | int load_blktrace(struct thread_data *td, const char *filename) |
254 | { | |
a61eddec | 255 | unsigned long long ttime, delay; |
fb7b71a3 | 256 | struct blk_io_trace t; |
4241ea8f | 257 | unsigned long ios[2], skipped_writes; |
a61eddec | 258 | unsigned int cpu; |
d84f8d49 | 259 | unsigned int rw_bs[2]; |
e2887563 | 260 | struct fifo *fifo; |
fb7b71a3 JA |
261 | int fd; |
262 | ||
263 | fd = open(filename, O_RDONLY); | |
264 | if (fd < 0) { | |
265 | td_verror(td, errno, "open blktrace file"); | |
266 | return 1; | |
267 | } | |
268 | ||
e2887563 JA |
269 | fifo = fifo_alloc(TRACE_FIFO_SIZE); |
270 | ||
6df8adaa JA |
271 | td->o.size = 0; |
272 | ||
a61eddec | 273 | cpu = 0; |
d84f8d49 JA |
274 | ttime = 0; |
275 | ios[0] = ios[1] = 0; | |
276 | rw_bs[0] = rw_bs[1] = 0; | |
4241ea8f | 277 | skipped_writes = 0; |
fb7b71a3 | 278 | do { |
e2887563 | 279 | int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); |
fb7b71a3 | 280 | |
e2887563 | 281 | if (ret < 0) |
8c1fdf04 | 282 | goto err; |
e2887563 JA |
283 | else if (!ret) |
284 | break; | |
285 | else if (ret < (int) sizeof(t)) { | |
286 | log_err("fio: short fifo get\n"); | |
fb7b71a3 | 287 | break; |
fb7b71a3 JA |
288 | } |
289 | ||
290 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { | |
5ec10eaa JA |
291 | log_err("fio: bad magic in blktrace data: %x\n", |
292 | t.magic); | |
8c1fdf04 | 293 | goto err; |
fb7b71a3 JA |
294 | } |
295 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { | |
5ec10eaa JA |
296 | log_err("fio: bad blktrace version %d\n", |
297 | t.magic & 0xff); | |
8c1fdf04 | 298 | goto err; |
fb7b71a3 | 299 | } |
f12b323f JA |
300 | ret = discard_pdu(td, fifo, fd, &t); |
301 | if (ret < 0) { | |
fb7b71a3 | 302 | td_verror(td, ret, "blktrace lseek"); |
8c1fdf04 | 303 | goto err; |
f12b323f JA |
304 | } else if (t.pdu_len != ret) { |
305 | log_err("fio: discarded %d of %d\n", ret, t.pdu_len); | |
306 | goto err; | |
fb7b71a3 | 307 | } |
f3f72c5e JA |
308 | if (t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) |
309 | continue; | |
a61eddec | 310 | if (!ttime) { |
8c1fdf04 | 311 | ttime = t.time; |
a61eddec JA |
312 | cpu = t.cpu; |
313 | } | |
314 | delay = 0; | |
315 | if (cpu == t.cpu) | |
316 | delay = t.time - ttime; | |
4241ea8f JA |
317 | if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only) |
318 | skipped_writes++; | |
319 | else | |
320 | handle_trace(td, &t, delay, ios, rw_bs); | |
8c1fdf04 | 321 | ttime = t.time; |
a61eddec | 322 | cpu = t.cpu; |
fb7b71a3 JA |
323 | } while (1); |
324 | ||
38470f85 | 325 | fifo_free(fifo); |
fb7b71a3 | 326 | close(fd); |
8c1fdf04 | 327 | |
4241ea8f | 328 | if (skipped_writes) |
5ec10eaa JA |
329 | log_err("fio: %s skips replay of %lu writes due to read-only\n", |
330 | td->o.name, skipped_writes); | |
4241ea8f | 331 | |
8c1fdf04 JA |
332 | if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
333 | log_err("fio: found no ios in blktrace data\n"); | |
334 | return 1; | |
d84f8d49 | 335 | } else if (ios[DDIR_READ] && !ios[DDIR_READ]) { |
8c1fdf04 | 336 | td->o.td_ddir = TD_DDIR_READ; |
d84f8d49 JA |
337 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
338 | } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { | |
8c1fdf04 | 339 | td->o.td_ddir = TD_DDIR_WRITE; |
d84f8d49 JA |
340 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
341 | } else { | |
8c1fdf04 | 342 | td->o.td_ddir = TD_DDIR_RW; |
d84f8d49 JA |
343 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
344 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; | |
345 | } | |
8c1fdf04 JA |
346 | |
347 | /* | |
348 | * We need to do direct/raw ios to the device, to avoid getting | |
349 | * read-ahead in our way. | |
350 | */ | |
351 | td->o.odirect = 1; | |
352 | ||
fb7b71a3 | 353 | return 0; |
8c1fdf04 JA |
354 | err: |
355 | close(fd); | |
38470f85 | 356 | fifo_free(fifo); |
8c1fdf04 | 357 | return 1; |
fb7b71a3 | 358 | } |