Windows: add install directory to the Path environment variable
[fio.git] / blktrace.c
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  */
18 static 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  */
44 static 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  */
60 static 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  */
74 int is_blktrace(const char *filename, int *need_swap)
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                 *need_swap = 0;
96                 return 1;
97         }
98
99         /*
100          * Maybe it needs to be endian swapped...
101          */
102         t.magic = fio_swap32(t.magic);
103         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
104                 *need_swap = 1;
105                 return 1;
106         }
107
108         return 0;
109 }
110
111 static int lookup_device(struct thread_data *td, char *path, unsigned int maj,
112                          unsigned int min)
113 {
114         struct dirent *dir;
115         struct stat st;
116         int found = 0;
117         DIR *D;
118
119         D = opendir(path);
120         if (!D)
121                 return 0;
122
123         while ((dir = readdir(D)) != NULL) {
124                 char full_path[256];
125
126                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
127                         continue;
128
129                 sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name);
130                 if (lstat(full_path, &st) == -1) {
131                         perror("lstat");
132                         break;
133                 }
134
135                 if (S_ISDIR(st.st_mode)) {
136                         found = lookup_device(td, full_path, maj, min);
137                         if (found) {
138                                 strcpy(path, full_path);
139                                 break;
140                         }
141                 }
142
143                 if (!S_ISBLK(st.st_mode))
144                         continue;
145
146                 /*
147                  * If replay_redirect is set then always return this device
148                  * upon lookup which overrides the device lookup based on
149                  * major minor in the actual blktrace
150                  */
151                 if (td->o.replay_redirect) {
152                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
153                                         " with: %s", maj, min,
154                                         td->o.replay_redirect);
155                         strcpy(path, td->o.replay_redirect);
156                         found = 1;
157                         break;
158                 }
159
160                 if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
161                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
162                         strcpy(path, full_path);
163                         found = 1;
164                         break;
165                 }
166         }
167
168         closedir(D);
169         return found;
170 }
171
172 #define FMINORBITS      20
173 #define FMINORMASK      ((1U << FMINORBITS) - 1)
174 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
175 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
176
177 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
178 {
179         struct io_piece *ipo;
180
181         ipo = calloc(1, sizeof(*ipo));
182         init_ipo(ipo);
183
184         ipo->ddir = DDIR_INVAL;
185         ipo->fileno = fileno;
186         ipo->file_action = action;
187         flist_add_tail(&ipo->list, &td->io_log_list);
188 }
189
190 static int trace_add_file(struct thread_data *td, __u32 device)
191 {
192         static unsigned int last_maj, last_min, last_fileno;
193         unsigned int maj = FMAJOR(device);
194         unsigned int min = FMINOR(device);
195         struct fio_file *f;
196         char dev[256];
197         unsigned int i;
198
199         if (last_maj == maj && last_min == min)
200                 return last_fileno;
201
202         last_maj = maj;
203         last_min = min;
204
205         /*
206          * check for this file in our list
207          */
208         for_each_file(td, f, i)
209                 if (f->major == maj && f->minor == min) {
210                         last_fileno = f->fileno;
211                         return last_fileno;
212                 }
213
214         strcpy(dev, "/dev");
215         if (lookup_device(td, dev, maj, min)) {
216                 int fileno;
217
218                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
219                 fileno = add_file_exclusive(td, dev);
220                 td->files[fileno]->major = maj;
221                 td->files[fileno]->minor = min;
222                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
223                 last_fileno = fileno;
224         }
225
226         return last_fileno;
227 }
228
229 /*
230  * Store blk_io_trace data in an ipo for later retrieval.
231  */
232 static void store_ipo(struct thread_data *td, unsigned long long offset,
233                       unsigned int bytes, int rw, unsigned long long ttime,
234                       int fileno)
235 {
236         struct io_piece *ipo = malloc(sizeof(*ipo));
237
238         init_ipo(ipo);
239
240         /*
241          * the 512 is wrong here, it should be the hardware sector size...
242          */
243         ipo->offset = offset * 512;
244         ipo->len = bytes;
245         ipo->delay = ttime / 1000;
246         if (rw)
247                 ipo->ddir = DDIR_WRITE;
248         else
249                 ipo->ddir = DDIR_READ;
250         ipo->fileno = fileno;
251
252         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
253                                                         ipo->ddir, ipo->offset,
254                                                         ipo->len, ipo->delay);
255         queue_io_piece(td, ipo);
256 }
257
258 static void handle_trace_notify(struct blk_io_trace *t)
259 {
260         switch (t->action) {
261         case BLK_TN_PROCESS:
262                 log_info("blktrace: got process notify: %x, %d\n",
263                                 t->action, t->pid);
264                 break;
265         case BLK_TN_TIMESTAMP:
266                 log_info("blktrace: got timestamp notify: %x, %d\n",
267                                 t->action, t->pid);
268                 break;
269         case BLK_TN_MESSAGE:
270                 break;
271         default:
272                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
273                 break;
274         }
275 }
276
277 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
278                                  unsigned long long ttime, unsigned long *ios)
279 {
280         struct io_piece *ipo = malloc(sizeof(*ipo));
281         int fileno;
282
283         init_ipo(ipo);
284         fileno = trace_add_file(td, t->device);
285
286         ios[DDIR_WRITE]++;
287         td->o.size += t->bytes;
288
289         memset(ipo, 0, sizeof(*ipo));
290         INIT_FLIST_HEAD(&ipo->list);
291
292         /*
293          * the 512 is wrong here, it should be the hardware sector size...
294          */
295         ipo->offset = t->sector * 512;
296         ipo->len = t->bytes;
297         ipo->delay = ttime / 1000;
298         ipo->ddir = DDIR_TRIM;
299         ipo->fileno = fileno;
300
301         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
302                                                         ipo->offset, ipo->len,
303                                                         ipo->delay);
304         queue_io_piece(td, ipo);
305 }
306
307 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
308                             unsigned long long ttime, unsigned long *ios,
309                             unsigned int *bs)
310 {
311         int rw;
312         int fileno;
313
314         fileno = trace_add_file(td, t->device);
315
316         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
317
318         if (t->bytes > bs[rw])
319                 bs[rw] = t->bytes;
320
321         ios[rw]++;
322         td->o.size += t->bytes;
323         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
324 }
325
326 /*
327  * We only care for queue traces, most of the others are side effects
328  * due to internal workings of the block layer.
329  */
330 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
331                          unsigned long long ttime, unsigned long *ios,
332                          unsigned int *bs)
333 {
334         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
335                 return;
336         if (t->action & BLK_TC_ACT(BLK_TC_PC))
337                 return;
338
339         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
340                 handle_trace_notify(t);
341         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
342                 handle_trace_discard(td, t, ttime, ios);
343         else
344                 handle_trace_fs(td, t, ttime, ios, bs);
345 }
346
347 static void byteswap_trace(struct blk_io_trace *t)
348 {
349         t->magic = fio_swap32(t->magic);
350         t->sequence = fio_swap32(t->sequence);
351         t->time = fio_swap64(t->time);
352         t->sector = fio_swap64(t->sector);
353         t->bytes = fio_swap32(t->bytes);
354         t->action = fio_swap32(t->action);
355         t->pid = fio_swap32(t->pid);
356         t->device = fio_swap32(t->device);
357         t->cpu = fio_swap32(t->cpu);
358         t->error = fio_swap16(t->error);
359         t->pdu_len = fio_swap16(t->pdu_len);
360 }
361
362 /*
363  * Load a blktrace file by reading all the blk_io_trace entries, and storing
364  * them as io_pieces like the fio text version would do.
365  */
366 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
367 {
368         unsigned long long ttime, delay;
369         struct blk_io_trace t;
370         unsigned long ios[2], skipped_writes;
371         unsigned int cpu;
372         unsigned int rw_bs[2];
373         struct fifo *fifo;
374         int fd, i, old_state;
375         struct fio_file *f;
376
377         fd = open(filename, O_RDONLY);
378         if (fd < 0) {
379                 td_verror(td, errno, "open blktrace file");
380                 return 1;
381         }
382
383         fifo = fifo_alloc(TRACE_FIFO_SIZE);
384
385         old_state = td->runstate;
386         td_set_runstate(td, TD_SETTING_UP);
387
388         td->o.size = 0;
389
390         cpu = 0;
391         ttime = 0;
392         ios[0] = ios[1] = 0;
393         rw_bs[0] = rw_bs[1] = 0;
394         skipped_writes = 0;
395         do {
396                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
397
398                 if (ret < 0)
399                         goto err;
400                 else if (!ret)
401                         break;
402                 else if (ret < (int) sizeof(t)) {
403                         log_err("fio: short fifo get\n");
404                         break;
405                 }
406
407                 if (need_swap)
408                         byteswap_trace(&t);
409
410                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
411                         log_err("fio: bad magic in blktrace data: %x\n",
412                                                                 t.magic);
413                         goto err;
414                 }
415                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
416                         log_err("fio: bad blktrace version %d\n",
417                                                                 t.magic & 0xff);
418                         goto err;
419                 }
420                 ret = discard_pdu(td, fifo, fd, &t);
421                 if (ret < 0) {
422                         td_verror(td, ret, "blktrace lseek");
423                         goto err;
424                 } else if (t.pdu_len != ret) {
425                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
426                         goto err;
427                 }
428                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
429                         if (!ttime) {
430                                 ttime = t.time;
431                                 cpu = t.cpu;
432                         }
433
434                         delay = 0;
435                         if (cpu == t.cpu)
436                                 delay = t.time - ttime;
437                         if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
438                                 skipped_writes++;
439                         else {
440                                 /*
441                                  * set delay to zero if no_stall enabled for
442                                  * fast replay
443                                  */
444                                 if (td->o.no_stall)
445                                         delay = 0;
446
447                                 handle_trace(td, &t, delay, ios, rw_bs);
448                         }
449
450                         ttime = t.time;
451                         cpu = t.cpu;
452                 } else {
453                         delay = 0;
454                         handle_trace(td, &t, delay, ios, rw_bs);
455                 }
456         } while (1);
457
458         for (i = 0; i < td->files_index; i++) {
459                 f = td->files[i];
460                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
461         }
462
463         fifo_free(fifo);
464         close(fd);
465
466         td_set_runstate(td, old_state);
467
468         if (!td->files_index) {
469                 log_err("fio: did not find replay device(s)\n");
470                 return 1;
471         }
472
473         if (skipped_writes)
474                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
475                                                 td->o.name, skipped_writes);
476
477         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
478                 log_err("fio: found no ios in blktrace data\n");
479                 return 1;
480         } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
481                 td->o.td_ddir = TD_DDIR_READ;
482                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
483         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
484                 td->o.td_ddir = TD_DDIR_WRITE;
485                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
486         } else {
487                 td->o.td_ddir = TD_DDIR_RW;
488                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
489                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
490         }
491
492         /*
493          * We need to do direct/raw ios to the device, to avoid getting
494          * read-ahead in our way.
495          */
496         td->o.odirect = 1;
497
498         return 0;
499 err:
500         close(fd);
501         fifo_free(fifo);
502         return 1;
503 }