Add missing types to io_u_log_error()
[fio.git] / blktrace.c
... / ...
CommitLineData
1/*
2 * blktrace support code for fio
3 */
4#include <stdio.h>
5#include <stdlib.h>
6#include <sys/stat.h>
7#include <dirent.h>
8
9#include "flist.h"
10#include "fio.h"
11#include "blktrace_api.h"
12
13#define TRACE_FIFO_SIZE 8192
14
15/*
16 * fifo refill frontend, to avoid reading data in trace sized bites
17 */
18static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
19{
20 char buf[TRACE_FIFO_SIZE];
21 unsigned int total;
22 int ret;
23
24 total = sizeof(buf);
25 if (total > fifo_room(fifo))
26 total = fifo_room(fifo);
27
28 ret = read(fd, buf, total);
29 if (ret < 0) {
30 td_verror(td, errno, "read blktrace file");
31 return -1;
32 }
33
34 if (ret > 0)
35 ret = fifo_put(fifo, buf, ret);
36
37 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
38 return ret;
39}
40
41/*
42 * Retrieve 'len' bytes from the fifo, refilling if necessary.
43 */
44static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
45 void *buf, unsigned int len)
46{
47 if (fifo_len(fifo) < len) {
48 int ret = refill_fifo(td, fifo, fd);
49
50 if (ret < 0)
51 return ret;
52 }
53
54 return fifo_get(fifo, buf, len);
55}
56
57/*
58 * Just discard the pdu by seeking past it.
59 */
60static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
61 struct blk_io_trace *t)
62{
63 if (t->pdu_len == 0)
64 return 0;
65
66 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
67 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
68}
69
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 */
74int is_blktrace(const char *filename)
75{
76 struct blk_io_trace t;
77 int fd, ret;
78
79 fd = open(filename, O_RDONLY);
80 if (fd < 0)
81 return 0;
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
100static 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)) {
135 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
136 strcpy(path, full_path);
137 found = 1;
138 break;
139 }
140 }
141
142 closedir(D);
143 return found;
144}
145
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))
150
151static void trace_add_open_event(struct thread_data *td, int fileno)
152{
153 struct io_piece *ipo;
154
155 ipo = calloc(1, sizeof(*ipo));
156
157 ipo->ddir = DDIR_INVAL;
158 ipo->fileno = fileno;
159 ipo->file_action = FIO_LOG_OPEN_FILE;
160 flist_add_tail(&ipo->list, &td->io_log_list);
161}
162
163static void trace_add_file(struct thread_data *td, __u32 device)
164{
165 static unsigned int last_maj, last_min;
166 unsigned int maj = FMAJOR(device);
167 unsigned int min = FMINOR(device);
168 struct fio_file *f;
169 char dev[256];
170 unsigned int i;
171
172 if (last_maj == maj && last_min == min)
173 return;
174
175 last_maj = maj;
176 last_min = min;
177
178 /*
179 * check for this file in our list
180 */
181 for_each_file(td, f, i)
182 if (f->major == maj && f->minor == min)
183 return;
184
185 strcpy(dev, "/dev");
186 if (lookup_device(dev, maj, min)) {
187 int fileno;
188
189 dprint(FD_BLKTRACE, "add devices %s\n", dev);
190 fileno = add_file_exclusive(td, dev);
191 trace_add_open_event(td, fileno);
192 }
193}
194
195/*
196 * Store blk_io_trace data in an ipo for later retrieval.
197 */
198static void store_ipo(struct thread_data *td, unsigned long long offset,
199 unsigned int bytes, int rw, unsigned long long ttime)
200{
201 struct io_piece *ipo = malloc(sizeof(*ipo));
202
203 memset(ipo, 0, sizeof(*ipo));
204 INIT_FLIST_HEAD(&ipo->list);
205 /*
206 * the 512 is wrong here, it should be the hardware sector size...
207 */
208 ipo->offset = offset * 512;
209 ipo->len = bytes;
210 ipo->delay = ttime / 1000;
211 if (rw)
212 ipo->ddir = DDIR_WRITE;
213 else
214 ipo->ddir = DDIR_READ;
215
216 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
217 ipo->ddir, ipo->offset,
218 ipo->len, ipo->delay);
219 queue_io_piece(td, ipo);
220}
221
222static void handle_trace_notify(struct blk_io_trace *t)
223{
224 switch (t->action) {
225 case BLK_TN_PROCESS:
226 printf("got process notify: %x, %d\n", t->action, t->pid);
227 break;
228 case BLK_TN_TIMESTAMP:
229 printf("got timestamp notify: %x, %d\n", t->action, t->pid);
230 break;
231 default:
232 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
233 break;
234 }
235}
236
237static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
238 unsigned long long ttime, unsigned long *ios,
239 unsigned int *bs)
240{
241 int rw;
242
243 trace_add_file(td, t->device);
244
245 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
246
247 if (t->bytes > bs[rw])
248 bs[rw] = t->bytes;
249
250 ios[rw]++;
251 td->o.size += t->bytes;
252 store_ipo(td, t->sector, t->bytes, rw, ttime);
253}
254
255/*
256 * We only care for queue traces, most of the others are side effects
257 * due to internal workings of the block layer.
258 */
259static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
260 unsigned long long ttime, unsigned long *ios,
261 unsigned int *bs)
262{
263 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
264 return;
265 if (t->action & BLK_TC_ACT(BLK_TC_PC))
266 return;
267
268 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
269 handle_trace_notify(t);
270 else
271 handle_trace_fs(td, t, ttime, ios, bs);
272}
273
274/*
275 * Load a blktrace file by reading all the blk_io_trace entries, and storing
276 * them as io_pieces like the fio text version would do.
277 */
278int load_blktrace(struct thread_data *td, const char *filename)
279{
280 unsigned long long ttime, delay;
281 struct blk_io_trace t;
282 unsigned long ios[2], skipped_writes;
283 unsigned int cpu;
284 unsigned int rw_bs[2];
285 struct fifo *fifo;
286 int fd;
287
288 fd = open(filename, O_RDONLY);
289 if (fd < 0) {
290 td_verror(td, errno, "open blktrace file");
291 return 1;
292 }
293
294 fifo = fifo_alloc(TRACE_FIFO_SIZE);
295
296 td->o.size = 0;
297
298 cpu = 0;
299 ttime = 0;
300 ios[0] = ios[1] = 0;
301 rw_bs[0] = rw_bs[1] = 0;
302 skipped_writes = 0;
303 do {
304 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
305
306 if (ret < 0)
307 goto err;
308 else if (!ret)
309 break;
310 else if (ret < (int) sizeof(t)) {
311 log_err("fio: short fifo get\n");
312 break;
313 }
314
315 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
316 log_err("fio: bad magic in blktrace data: %x\n",
317 t.magic);
318 goto err;
319 }
320 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
321 log_err("fio: bad blktrace version %d\n",
322 t.magic & 0xff);
323 goto err;
324 }
325 ret = discard_pdu(td, fifo, fd, &t);
326 if (ret < 0) {
327 td_verror(td, ret, "blktrace lseek");
328 goto err;
329 } else if (t.pdu_len != ret) {
330 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
331 goto err;
332 }
333 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
334 if (!ttime) {
335 ttime = t.time;
336 cpu = t.cpu;
337 }
338
339 delay = 0;
340 if (cpu == t.cpu)
341 delay = t.time - ttime;
342 if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
343 skipped_writes++;
344 else {
345 /*
346 * set delay to zero if no_stall enabled for
347 * fast replay
348 */
349 if (td->o.no_stall)
350 delay = 0;
351
352 handle_trace(td, &t, delay, ios, rw_bs);
353 }
354
355 ttime = t.time;
356 cpu = t.cpu;
357 } else {
358 delay = 0;
359 handle_trace(td, &t, delay, ios, rw_bs);
360 }
361 } while (1);
362
363 fifo_free(fifo);
364 close(fd);
365
366 if (skipped_writes)
367 log_err("fio: %s skips replay of %lu writes due to read-only\n",
368 td->o.name, skipped_writes);
369
370 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
371 log_err("fio: found no ios in blktrace data\n");
372 return 1;
373 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
374 td->o.td_ddir = TD_DDIR_READ;
375 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
376 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
377 td->o.td_ddir = TD_DDIR_WRITE;
378 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
379 } else {
380 td->o.td_ddir = TD_DDIR_RW;
381 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
382 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
383 }
384
385 /*
386 * We need to do direct/raw ios to the device, to avoid getting
387 * read-ahead in our way.
388 */
389 td->o.odirect = 1;
390
391 return 0;
392err:
393 close(fd);
394 fifo_free(fifo);
395 return 1;
396}