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