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" };
25 assert(io_u->ddir < 3);
27 if (!td->o.write_iolog_file)
30 fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
31 act[io_u->ddir], io_u->offset,
35 void log_file(struct thread_data *td, struct fio_file *f,
36 enum file_log_act what)
38 const char *act[] = { "add", "open", "close" };
42 if (!td->o.write_iolog_file)
47 * this happens on the pre-open/close done before the job starts
52 fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
55 static void iolog_delay(struct thread_data *td, unsigned long delay)
57 unsigned long usec = utime_since_now(&td->last_issue);
65 * less than 100 usec delay, just regard it as noise
70 usec_sleep(td, delay);
73 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
81 if (ipo->ddir != DDIR_INVAL)
84 f = td->files[ipo->fileno];
86 switch (ipo->file_action) {
87 case FIO_LOG_OPEN_FILE:
88 ret = td_io_open_file(td, f);
91 td_verror(td, ret, "iolog open file");
93 case FIO_LOG_CLOSE_FILE:
94 td_io_close_file(td, f);
96 case FIO_LOG_UNLINK_FILE:
100 log_err("fio: bad file action %d\n", ipo->file_action);
107 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
109 struct io_piece *ipo;
110 unsigned long elapsed;
112 while (!flist_empty(&td->io_log_list)) {
115 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
116 flist_del(&ipo->list);
118 ret = ipo_special(td, ipo);
122 } else if (ret > 0) {
127 io_u->ddir = ipo->ddir;
128 if (ipo->ddir != DDIR_WAIT) {
129 io_u->offset = ipo->offset;
130 io_u->buflen = ipo->len;
131 io_u->file = td->files[ipo->fileno];
132 get_file(io_u->file);
134 dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
135 io_u->buflen, io_u->file->file_name);
136 if (ipo->delay) iolog_delay(td, ipo->delay);
138 elapsed = mtime_since_genesis();
139 if (ipo->delay > elapsed)
140 usec_sleep(td, (ipo->delay - elapsed) * 1000);
146 if (ipo->ddir != DDIR_WAIT)
154 void prune_io_piece_log(struct thread_data *td)
156 struct io_piece *ipo;
159 while ((n = rb_first(&td->io_hist_tree)) != NULL) {
160 ipo = rb_entry(n, struct io_piece, rb_node);
161 rb_erase(n, &td->io_hist_tree);
165 while (!flist_empty(&td->io_hist_list)) {
166 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
167 flist_del(&ipo->list);
173 * log a successful write, so we can unwind the log for verify
175 void log_io_piece(struct thread_data *td, struct io_u *io_u)
177 struct rb_node **p, *parent;
178 struct io_piece *ipo, *__ipo;
180 ipo = malloc(sizeof(struct io_piece));
181 ipo->file = io_u->file;
182 ipo->offset = io_u->offset;
183 ipo->len = io_u->buflen;
186 * We don't need to sort the entries, if:
188 * Sequential writes, or
189 * Random writes that lay out the file as it goes along
191 * For both these cases, just reading back data in the order we
192 * wrote it out is the fastest.
194 * One exception is if we don't have a random map AND we are doing
195 * verifies, in that case we need to check for duplicate blocks and
196 * drop the old one, which we rely on the rb insert/lookup for
199 if ((!td_random(td) || !td->o.overwrite) &&
200 (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
201 INIT_FLIST_HEAD(&ipo->list);
202 flist_add_tail(&ipo->list, &td->io_hist_list);
206 RB_CLEAR_NODE(&ipo->rb_node);
209 * Sort the entry into the verification list
212 p = &td->io_hist_tree.rb_node;
217 __ipo = rb_entry(parent, struct io_piece, rb_node);
218 if (ipo->offset < __ipo->offset)
220 else if (ipo->offset > __ipo->offset)
223 assert(ipo->len == __ipo->len);
224 rb_erase(parent, &td->io_hist_tree);
229 rb_link_node(&ipo->rb_node, parent, p);
230 rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
233 void write_iolog_close(struct thread_data *td)
239 td->iolog_buf = NULL;
243 * Read version 2 iolog data. It is enhanced to include per-file logging,
246 static int read_iolog2(struct thread_data *td, FILE *f)
248 unsigned long long offset;
250 int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
255 free_release_files(td);
258 * Read in the read iolog and store it, reuse the infrastructure
259 * for doing verifications.
262 fname = malloc(256+16);
263 act = malloc(256+16);
265 reads = writes = waits = 0;
266 while ((p = fgets(str, 4096, f)) != NULL) {
267 struct io_piece *ipo;
270 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
276 if (!strcmp(act, "wait"))
278 else if (!strcmp(act, "read"))
280 else if (!strcmp(act, "write"))
282 else if (!strcmp(act, "sync"))
284 else if (!strcmp(act, "datasync"))
287 log_err("fio: bad iolog file action: %s\n",
293 if (!strcmp(act, "add")) {
295 fileno = add_file(td, fname);
296 file_action = FIO_LOG_ADD_FILE;
298 } else if (!strcmp(act, "open")) {
299 fileno = get_fileno(td, fname);
300 file_action = FIO_LOG_OPEN_FILE;
301 } else if (!strcmp(act, "close")) {
302 fileno = get_fileno(td, fname);
303 file_action = FIO_LOG_CLOSE_FILE;
305 log_err("fio: bad iolog file action: %s\n",
310 log_err("bad iolog2: %s", p);
316 else if (rw == DDIR_WRITE) {
318 * Don't add a write for ro mode
323 } else if (rw == DDIR_WAIT) {
325 } else if (rw == DDIR_INVAL) {
326 } else if (!ddir_sync(rw)) {
327 log_err("bad ddir: %d\n", rw);
334 ipo = malloc(sizeof(*ipo));
335 memset(ipo, 0, sizeof(*ipo));
336 INIT_FLIST_HEAD(&ipo->list);
338 if (rw == DDIR_WAIT) {
341 ipo->offset = offset;
343 if (bytes > td->o.max_bs[rw])
344 td->o.max_bs[rw] = bytes;
345 ipo->fileno = fileno;
346 ipo->file_action = file_action;
349 queue_io_piece(td, ipo);
356 if (writes && read_only) {
357 log_err("fio: <%s> skips replay of %d writes due to"
358 " read-only\n", td->o.name, writes);
362 if (!reads && !writes && !waits)
364 else if (reads && !writes)
365 td->o.td_ddir = TD_DDIR_READ;
366 else if (!reads && writes)
367 td->o.td_ddir = TD_DDIR_WRITE;
369 td->o.td_ddir = TD_DDIR_RW;
375 * open iolog, check version, and call appropriate parser
377 static int init_iolog_read(struct thread_data *td)
379 char buffer[256], *p;
383 f = fopen(td->o.read_iolog_file, "r");
385 perror("fopen read iolog");
389 p = fgets(buffer, sizeof(buffer), f);
391 td_verror(td, errno, "iolog read");
392 log_err("fio: unable to read iolog\n");
397 * version 2 of the iolog stores a specific string as the
398 * first line, check for that
400 if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
401 ret = read_iolog2(td, f);
403 log_err("fio: iolog version 1 is no longer supported\n");
412 * Setup a log for storing io patterns.
414 static int init_iolog_write(struct thread_data *td)
420 f = fopen(td->o.write_iolog_file, "a");
422 perror("fopen write iolog");
427 * That's it for writing, setup a log buffer and we're done.
430 td->iolog_buf = malloc(8192);
431 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
434 * write our version line
436 if (fprintf(f, "%s\n", iolog_ver2) < 0) {
437 perror("iolog init\n");
442 * add all known files
444 for_each_file(td, ff, i)
445 log_file(td, ff, FIO_LOG_ADD_FILE);
450 int init_iolog(struct thread_data *td)
454 if (td->o.read_iolog_file) {
456 * Check if it's a blktrace file and load that if possible.
457 * Otherwise assume it's a normal log file and load that.
459 if (is_blktrace(td->o.read_iolog_file))
460 ret = load_blktrace(td, td->o.read_iolog_file);
462 ret = init_iolog_read(td);
463 } else if (td->o.write_iolog_file)
464 ret = init_iolog_write(td);
469 void setup_log(struct io_log **log)
471 struct io_log *l = malloc(sizeof(*l));
474 l->max_samples = 1024;
475 l->log = malloc(l->max_samples * sizeof(struct io_sample));
479 void __finish_log(struct io_log *log, const char *name)
484 f = fopen(name, "a");
490 for (i = 0; i < log->nr_samples; i++) {
491 fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time,
502 void finish_log_named(struct thread_data *td, struct io_log *log,
503 const char *prefix, const char *postfix)
505 char file_name[256], *p;
507 snprintf(file_name, 200, "%s_%s.log", prefix, postfix);
508 p = basename(file_name);
509 __finish_log(log, p);
512 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
514 finish_log_named(td, log, td->o.name, name);