xxhash: dos2unix'ize
[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                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
221                 last_fileno = fileno;
222         }
223
224         return last_fileno;
225 }
226
227 /*
228  * Store blk_io_trace data in an ipo for later retrieval.
229  */
230 static void store_ipo(struct thread_data *td, unsigned long long offset,
231                       unsigned int bytes, int rw, unsigned long long ttime,
232                       int fileno)
233 {
234         struct io_piece *ipo = malloc(sizeof(*ipo));
235
236         init_ipo(ipo);
237
238         /*
239          * the 512 is wrong here, it should be the hardware sector size...
240          */
241         ipo->offset = offset * 512;
242         ipo->len = bytes;
243         ipo->delay = ttime / 1000;
244         if (rw)
245                 ipo->ddir = DDIR_WRITE;
246         else
247                 ipo->ddir = DDIR_READ;
248         ipo->fileno = fileno;
249
250         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
251                                                         ipo->ddir, ipo->offset,
252                                                         ipo->len, ipo->delay);
253         queue_io_piece(td, ipo);
254 }
255
256 static void handle_trace_notify(struct blk_io_trace *t)
257 {
258         switch (t->action) {
259         case BLK_TN_PROCESS:
260                 log_info("blktrace: got process notify: %x, %d\n",
261                                 t->action, t->pid);
262                 break;
263         case BLK_TN_TIMESTAMP:
264                 log_info("blktrace: got timestamp notify: %x, %d\n",
265                                 t->action, t->pid);
266                 break;
267         case BLK_TN_MESSAGE:
268                 break;
269         default:
270                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
271                 break;
272         }
273 }
274
275 static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t,
276                                  unsigned long long ttime, unsigned long *ios)
277 {
278         struct io_piece *ipo = malloc(sizeof(*ipo));
279         int fileno;
280
281         init_ipo(ipo);
282         fileno = trace_add_file(td, t->device);
283
284         ios[DDIR_WRITE]++;
285         td->o.size += t->bytes;
286
287         memset(ipo, 0, sizeof(*ipo));
288         INIT_FLIST_HEAD(&ipo->list);
289
290         /*
291          * the 512 is wrong here, it should be the hardware sector size...
292          */
293         ipo->offset = t->sector * 512;
294         ipo->len = t->bytes;
295         ipo->delay = ttime / 1000;
296         ipo->ddir = DDIR_TRIM;
297         ipo->fileno = fileno;
298
299         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
300                                                         ipo->offset, ipo->len,
301                                                         ipo->delay);
302         queue_io_piece(td, ipo);
303 }
304
305 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
306                             unsigned long long ttime, unsigned long *ios,
307                             unsigned int *bs)
308 {
309         int rw;
310         int fileno;
311
312         fileno = trace_add_file(td, t->device);
313
314         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
315
316         if (t->bytes > bs[rw])
317                 bs[rw] = t->bytes;
318
319         ios[rw]++;
320         td->o.size += t->bytes;
321         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
322 }
323
324 /*
325  * We only care for queue traces, most of the others are side effects
326  * due to internal workings of the block layer.
327  */
328 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
329                          unsigned long long ttime, unsigned long *ios,
330                          unsigned int *bs)
331 {
332         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
333                 return;
334         if (t->action & BLK_TC_ACT(BLK_TC_PC))
335                 return;
336
337         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
338                 handle_trace_notify(t);
339         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
340                 handle_trace_discard(td, t, ttime, ios);
341         else
342                 handle_trace_fs(td, t, ttime, ios, bs);
343 }
344
345 static void byteswap_trace(struct blk_io_trace *t)
346 {
347         t->magic = fio_swap32(t->magic);
348         t->sequence = fio_swap32(t->sequence);
349         t->time = fio_swap64(t->time);
350         t->sector = fio_swap64(t->sector);
351         t->bytes = fio_swap32(t->bytes);
352         t->action = fio_swap32(t->action);
353         t->pid = fio_swap32(t->pid);
354         t->device = fio_swap32(t->device);
355         t->cpu = fio_swap32(t->cpu);
356         t->error = fio_swap16(t->error);
357         t->pdu_len = fio_swap16(t->pdu_len);
358 }
359
360 /*
361  * Load a blktrace file by reading all the blk_io_trace entries, and storing
362  * them as io_pieces like the fio text version would do.
363  */
364 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
365 {
366         unsigned long long ttime, delay;
367         struct blk_io_trace t;
368         unsigned long ios[2], skipped_writes;
369         unsigned int cpu;
370         unsigned int rw_bs[2];
371         struct fifo *fifo;
372         int fd, i;
373         struct fio_file *f;
374
375         fd = open(filename, O_RDONLY);
376         if (fd < 0) {
377                 td_verror(td, errno, "open blktrace file");
378                 return 1;
379         }
380
381         fifo = fifo_alloc(TRACE_FIFO_SIZE);
382
383         td->o.size = 0;
384
385         cpu = 0;
386         ttime = 0;
387         ios[0] = ios[1] = 0;
388         rw_bs[0] = rw_bs[1] = 0;
389         skipped_writes = 0;
390         do {
391                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
392
393                 if (ret < 0)
394                         goto err;
395                 else if (!ret)
396                         break;
397                 else if (ret < (int) sizeof(t)) {
398                         log_err("fio: short fifo get\n");
399                         break;
400                 }
401
402                 if (need_swap)
403                         byteswap_trace(&t);
404
405                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
406                         log_err("fio: bad magic in blktrace data: %x\n",
407                                                                 t.magic);
408                         goto err;
409                 }
410                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
411                         log_err("fio: bad blktrace version %d\n",
412                                                                 t.magic & 0xff);
413                         goto err;
414                 }
415                 ret = discard_pdu(td, fifo, fd, &t);
416                 if (ret < 0) {
417                         td_verror(td, ret, "blktrace lseek");
418                         goto err;
419                 } else if (t.pdu_len != ret) {
420                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
421                         goto err;
422                 }
423                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
424                         if (!ttime) {
425                                 ttime = t.time;
426                                 cpu = t.cpu;
427                         }
428
429                         delay = 0;
430                         if (cpu == t.cpu)
431                                 delay = t.time - ttime;
432                         if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only)
433                                 skipped_writes++;
434                         else {
435                                 /*
436                                  * set delay to zero if no_stall enabled for
437                                  * fast replay
438                                  */
439                                 if (td->o.no_stall)
440                                         delay = 0;
441
442                                 handle_trace(td, &t, delay, ios, rw_bs);
443                         }
444
445                         ttime = t.time;
446                         cpu = t.cpu;
447                 } else {
448                         delay = 0;
449                         handle_trace(td, &t, delay, ios, rw_bs);
450                 }
451         } while (1);
452
453         for (i = 0; i < td->files_index; i++) {
454                 f = td->files[i];
455                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
456         }
457
458         fifo_free(fifo);
459         close(fd);
460
461         if (!td->files_index) {
462                 log_err("fio: did not find replay device(s)\n");
463                 return 1;
464         }
465
466         if (skipped_writes)
467                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
468                                                 td->o.name, skipped_writes);
469
470         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
471                 log_err("fio: found no ios in blktrace data\n");
472                 return 1;
473         } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
474                 td->o.td_ddir = TD_DDIR_READ;
475                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
476         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
477                 td->o.td_ddir = TD_DDIR_WRITE;
478                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
479         } else {
480                 td->o.td_ddir = TD_DDIR_RW;
481                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
482                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
483         }
484
485         /*
486          * We need to do direct/raw ios to the device, to avoid getting
487          * read-ahead in our way.
488          */
489         td->o.odirect = 1;
490
491         return 0;
492 err:
493         close(fd);
494         fifo_free(fifo);
495         return 1;
496 }