2 * Code related to writing an iolog of what a thread is doing, and to
3 * later read that back and replay
13 static const char iolog_ver2[] = "fio version 2 iolog";
15 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
17 flist_add_tail(&ipo->list, &td->io_log_list);
18 td->total_io_size += ipo->len;
21 void log_io_u(struct thread_data *td, struct io_u *io_u)
23 const char *act[] = { "read", "write", "sync", "datasync",
26 assert(io_u->ddir <= 4);
28 if (!td->o.write_iolog_file)
31 fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
32 act[io_u->ddir], io_u->offset,
36 void log_file(struct thread_data *td, struct fio_file *f,
37 enum file_log_act what)
39 const char *act[] = { "add", "open", "close" };
43 if (!td->o.write_iolog_file)
48 * this happens on the pre-open/close done before the job starts
53 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
56 static void iolog_delay(struct thread_data *td, unsigned long delay)
58 unsigned long usec = utime_since_now(&td->last_issue);
66 * less than 100 usec delay, just regard it as noise
71 usec_sleep(td, delay);
74 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
82 if (ipo->ddir != DDIR_INVAL)
85 f = td->files[ipo->fileno];
87 switch (ipo->file_action) {
88 case FIO_LOG_OPEN_FILE:
89 ret = td_io_open_file(td, f);
92 td_verror(td, ret, "iolog open file");
94 case FIO_LOG_CLOSE_FILE:
95 td_io_close_file(td, f);
97 case FIO_LOG_UNLINK_FILE:
101 log_err("fio: bad file action %d\n", ipo->file_action);
108 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
110 struct io_piece *ipo;
111 unsigned long elapsed;
113 while (!flist_empty(&td->io_log_list)) {
116 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
117 flist_del(&ipo->list);
119 ret = ipo_special(td, ipo);
123 } else if (ret > 0) {
128 io_u->ddir = ipo->ddir;
129 if (ipo->ddir != DDIR_WAIT) {
130 io_u->offset = ipo->offset;
131 io_u->buflen = ipo->len;
132 io_u->file = td->files[ipo->fileno];
133 get_file(io_u->file);
135 dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
136 io_u->buflen, io_u->file->file_name);
137 if (ipo->delay) iolog_delay(td, ipo->delay);
139 elapsed = mtime_since_genesis();
140 if (ipo->delay > elapsed)
141 usec_sleep(td, (ipo->delay - elapsed) * 1000);
147 if (ipo->ddir != DDIR_WAIT)
155 void prune_io_piece_log(struct thread_data *td)
157 struct io_piece *ipo;
160 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
161 ipo = rb_entry(n, struct io_piece, rb_node);
162 rb_erase(n, &td->io_hist_tree);
167 while (!flist_empty(&td->io_hist_list)) {
168 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
169 flist_del(&ipo->list);
176 * log a successful write, so we can unwind the log for verify
178 void log_io_piece(struct thread_data *td, struct io_u *io_u)
180 struct rb_node **p, *parent;
181 struct io_piece *ipo, *__ipo;
183 ipo = malloc(sizeof(struct io_piece));
184 ipo->file = io_u->file;
185 ipo->offset = io_u->offset;
186 ipo->len = io_u->buflen;
189 * We don't need to sort the entries, if:
191 * Sequential writes, or
192 * Random writes that lay out the file as it goes along
194 * For both these cases, just reading back data in the order we
195 * wrote it out is the fastest.
197 * One exception is if we don't have a random map AND we are doing
198 * verifies, in that case we need to check for duplicate blocks and
199 * drop the old one, which we rely on the rb insert/lookup for
202 if ((!td_random(td) || !td->o.overwrite) &&
203 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
204 INIT_FLIST_HEAD(&ipo->list);
205 flist_add_tail(&ipo->list, &td->io_hist_list);
210 RB_CLEAR_NODE(&ipo->rb_node);
213 * Sort the entry into the verification list
216 p = &td->io_hist_tree.rb_node;
221 __ipo = rb_entry(parent, struct io_piece, rb_node);
222 if (ipo->offset < __ipo->offset)
224 else if (ipo->offset > __ipo->offset)
227 assert(ipo->len == __ipo->len);
229 rb_erase(parent, &td->io_hist_tree);
234 rb_link_node(&ipo->rb_node, parent, p);
235 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
239 void write_iolog_close(struct thread_data *td)
245 td->iolog_buf = NULL;
249 * Read version 2 iolog data. It is enhanced to include per-file logging,
252 static int read_iolog2(struct thread_data *td, FILE *f)
254 unsigned long long offset;
256 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
261 free_release_files(td);
264 * Read in the read iolog and store it, reuse the infrastructure
265 * for doing verifications.
268 fname = malloc(256+16);
269 act = malloc(256+16);
271 reads = writes = waits = 0;
272 while ((p = fgets(str, 4096, f)) != NULL) {
273 struct io_piece *ipo;
276 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
282 if (!strcmp(act, "wait"))
284 else if (!strcmp(act, "read"))
286 else if (!strcmp(act, "write"))
288 else if (!strcmp(act, "sync"))
290 else if (!strcmp(act, "datasync"))
293 log_err("fio: bad iolog file action: %s\n",
299 if (!strcmp(act, "add")) {
301 fileno = add_file(td, fname);
302 file_action = FIO_LOG_ADD_FILE;
304 } else if (!strcmp(act, "open")) {
305 fileno = get_fileno(td, fname);
306 file_action = FIO_LOG_OPEN_FILE;
307 } else if (!strcmp(act, "close")) {
308 fileno = get_fileno(td, fname);
309 file_action = FIO_LOG_CLOSE_FILE;
311 log_err("fio: bad iolog file action: %s\n",
316 log_err("bad iolog2: %s", p);
322 else if (rw == DDIR_WRITE) {
324 * Don't add a write for ro mode
329 } else if (rw == DDIR_WAIT) {
331 } else if (rw == DDIR_INVAL) {
332 } else if (!ddir_sync(rw)) {
333 log_err("bad ddir: %d\n", rw);
340 ipo = malloc(sizeof(*ipo));
341 memset(ipo, 0, sizeof(*ipo));
342 INIT_FLIST_HEAD(&ipo->list);
344 if (rw == DDIR_WAIT) {
347 ipo->offset = offset;
349 if (bytes > td->o.max_bs[rw])
350 td->o.max_bs[rw] = bytes;
351 ipo->fileno = fileno;
352 ipo->file_action = file_action;
355 queue_io_piece(td, ipo);
362 if (writes && read_only) {
363 log_err("fio: <%s> skips replay of %d writes due to"
364 " read-only\n", td->o.name, writes);
368 if (!reads && !writes && !waits)
370 else if (reads && !writes)
371 td->o.td_ddir = TD_DDIR_READ;
372 else if (!reads && writes)
373 td->o.td_ddir = TD_DDIR_WRITE;
375 td->o.td_ddir = TD_DDIR_RW;
381 * open iolog, check version, and call appropriate parser
383 static int init_iolog_read(struct thread_data *td)
385 char buffer[256], *p;
389 f = fopen(td->o.read_iolog_file, "r");
391 perror("fopen read iolog");
395 p = fgets(buffer, sizeof(buffer), f);
397 td_verror(td, errno, "iolog read");
398 log_err("fio: unable to read iolog\n");
403 * version 2 of the iolog stores a specific string as the
404 * first line, check for that
406 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
407 ret = read_iolog2(td, f);
409 log_err("fio: iolog version 1 is no longer supported\n");
418 * Setup a log for storing io patterns.
420 static int init_iolog_write(struct thread_data *td)
426 f = fopen(td->o.write_iolog_file, "a");
428 perror("fopen write iolog");
433 * That's it for writing, setup a log buffer and we're done.
436 td->iolog_buf = malloc(8192);
437 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
440 * write our version line
442 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
443 perror("iolog init\n");
448 * add all known files
450 for_each_file(td, ff, i)
451 log_file(td, ff, FIO_LOG_ADD_FILE);
456 int init_iolog(struct thread_data *td)
460 if (td->o.read_iolog_file) {
462 * Check if it's a blktrace file and load that if possible.
463 * Otherwise assume it's a normal log file and load that.
465 if (is_blktrace(td->o.read_iolog_file))
466 ret = load_blktrace(td, td->o.read_iolog_file);
468 ret = init_iolog_read(td);
469 } else if (td->o.write_iolog_file)
470 ret = init_iolog_write(td);
475 void setup_log(struct io_log **log)
477 struct io_log *l = malloc(sizeof(*l));
480 l->max_samples = 1024;
481 l->log = malloc(l->max_samples * sizeof(struct io_sample));
485 void __finish_log(struct io_log *log, const char *name)
490 f = fopen(name, "a");
496 for (i = 0; i < log->nr_samples; i++) {
497 fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time,
508 void finish_log_named(struct thread_data *td, struct io_log *log,
509 const char *prefix, const char *postfix)
511 char file_name[256], *p;
513 snprintf(file_name, 200, "%s_%s.log", prefix, postfix);
514 p = basename(file_name);
515 __finish_log(log, p);
518 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
520 finish_log_named(td, log, td->o.name, name);