blob: a3474cb57ee9825077684f3b3f7de96b272a2e4c [file] [log] [blame]
Jens Axboefb7b71a2007-05-15 08:44:04 +02001/*
2 * blktrace support code for fio
3 */
4#include <stdio.h>
5#include <stdlib.h>
Jens Axboe5e6c2062007-05-16 14:40:29 +02006#include <sys/stat.h>
Elliott Hugheseda3a602017-05-19 18:53:02 -07007#include <sys/ioctl.h>
8#include <linux/fs.h>
Jens Axboe5e6c2062007-05-16 14:40:29 +02009#include <dirent.h>
Jens Axboe8c1fdf02007-05-15 11:54:21 +020010
Jens Axboe01743ee2008-06-02 12:19:19 +020011#include "flist.h"
Jens Axboefb7b71a2007-05-15 08:44:04 +020012#include "fio.h"
13#include "blktrace_api.h"
Elliott Hugheseda3a602017-05-19 18:53:02 -070014#include "oslib/linux-dev-lookup.h"
Jens Axboefb7b71a2007-05-15 08:44:04 +020015
Jens Axboe2da7df12010-08-01 21:27:28 +020016#define TRACE_FIFO_SIZE 8192
Jens Axboee2887562007-05-16 09:25:09 +020017
18/*
19 * fifo refill frontend, to avoid reading data in trace sized bites
20 */
21static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
22{
23 char buf[TRACE_FIFO_SIZE];
Jens Axboef12b3232007-05-16 12:16:44 +020024 unsigned int total;
Jens Axboee2887562007-05-16 09:25:09 +020025 int ret;
26
Jens Axboef12b3232007-05-16 12:16:44 +020027 total = sizeof(buf);
28 if (total > fifo_room(fifo))
29 total = fifo_room(fifo);
Jens Axboee2887562007-05-16 09:25:09 +020030
Jens Axboef12b3232007-05-16 12:16:44 +020031 ret = read(fd, buf, total);
32 if (ret < 0) {
33 td_verror(td, errno, "read blktrace file");
34 return -1;
Jens Axboee2887562007-05-16 09:25:09 +020035 }
36
Jens Axboef12b3232007-05-16 12:16:44 +020037 if (ret > 0)
38 ret = fifo_put(fifo, buf, ret);
39
Jens Axboebd6f78b2008-02-01 20:27:52 +010040 dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
Jens Axboef12b3232007-05-16 12:16:44 +020041 return ret;
Jens Axboee2887562007-05-16 09:25:09 +020042}
43
44/*
45 * Retrieve 'len' bytes from the fifo, refilling if necessary.
46 */
47static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
48 void *buf, unsigned int len)
49{
Jens Axboef12b3232007-05-16 12:16:44 +020050 if (fifo_len(fifo) < len) {
51 int ret = refill_fifo(td, fifo, fd);
Jens Axboee2887562007-05-16 09:25:09 +020052
Jens Axboef12b3232007-05-16 12:16:44 +020053 if (ret < 0)
54 return ret;
55 }
Jens Axboee2887562007-05-16 09:25:09 +020056
57 return fifo_get(fifo, buf, len);
58}
59
Jens Axboe8c1fdf02007-05-15 11:54:21 +020060/*
61 * Just discard the pdu by seeking past it.
62 */
Jens Axboef12b3232007-05-16 12:16:44 +020063static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
64 struct blk_io_trace *t)
Jens Axboefb7b71a2007-05-15 08:44:04 +020065{
66 if (t->pdu_len == 0)
67 return 0;
68
Jens Axboebd6f78b2008-02-01 20:27:52 +010069 dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
Jens Axboef12b3232007-05-16 12:16:44 +020070 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
Jens Axboefb7b71a2007-05-15 08:44:04 +020071}
72
Jens Axboe8c1fdf02007-05-15 11:54:21 +020073/*
74 * Check if this is a blktrace binary data file. We read a single trace
75 * into memory and check for the magic signature.
76 */
Jens Axboed95b34a2013-11-21 09:55:49 -070077int is_blktrace(const char *filename, int *need_swap)
Jens Axboefb7b71a2007-05-15 08:44:04 +020078{
79 struct blk_io_trace t;
80 int fd, ret;
81
82 fd = open(filename, O_RDONLY);
Jens Axboe4dced402007-07-23 14:16:55 +020083 if (fd < 0)
Jens Axboefb7b71a2007-05-15 08:44:04 +020084 return 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +020085
86 ret = read(fd, &t, sizeof(t));
87 close(fd);
88
89 if (ret < 0) {
90 perror("read blktrace");
91 return 0;
92 } else if (ret != sizeof(t)) {
93 log_err("fio: short read on blktrace file\n");
94 return 0;
95 }
96
Jens Axboed95b34a2013-11-21 09:55:49 -070097 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
98 *need_swap = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +020099 return 1;
Jens Axboed95b34a2013-11-21 09:55:49 -0700100 }
101
102 /*
103 * Maybe it needs to be endian swapped...
104 */
105 t.magic = fio_swap32(t.magic);
106 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
107 *need_swap = 1;
108 return 1;
109 }
Jens Axboefb7b71a2007-05-15 08:44:04 +0200110
111 return 0;
112}
113
Jens Axboec69aa912007-05-22 15:51:50 +0200114#define FMINORBITS 20
115#define FMINORMASK ((1U << FMINORBITS) - 1)
116#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
117#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
Jens Axboeeeb9c2a2007-05-16 18:28:47 +0200118
Shaohua Li89ac1d42012-06-11 08:56:32 +0200119static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
Jens Axboe691c8fb2008-03-07 14:26:26 +0100120{
121 struct io_piece *ipo;
122
123 ipo = calloc(1, sizeof(*ipo));
Jens Axboe0d29de82010-09-01 13:54:15 +0200124 init_ipo(ipo);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100125
126 ipo->ddir = DDIR_INVAL;
127 ipo->fileno = fileno;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200128 ipo->file_action = action;
Jens Axboe01743ee2008-06-02 12:19:19 +0200129 flist_add_tail(&ipo->list, &td->io_log_list);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100130}
131
Elliott Hugheseda3a602017-05-19 18:53:02 -0700132static int get_dev_blocksize(const char *dev, unsigned int *bs)
Jens Axboe5e6c2062007-05-16 14:40:29 +0200133{
Elliott Hugheseda3a602017-05-19 18:53:02 -0700134 int fd;
135
136 fd = open(dev, O_RDONLY);
137 if (fd < 0)
138 return 1;
139
140 if (ioctl(fd, BLKSSZGET, bs) < 0) {
141 close(fd);
142 return 1;
143 }
144
145 close(fd);
146 return 0;
147}
148
149static int trace_add_file(struct thread_data *td, __u32 device,
150 unsigned int *bs)
151{
152 static unsigned int last_maj, last_min, last_fileno, last_bs;
Jens Axboec69aa912007-05-22 15:51:50 +0200153 unsigned int maj = FMAJOR(device);
154 unsigned int min = FMINOR(device);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200155 struct fio_file *f;
Jens Axboe5e6c2062007-05-16 14:40:29 +0200156 unsigned int i;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700157 char dev[256];
Jens Axboe5e6c2062007-05-16 14:40:29 +0200158
Elliott Hugheseda3a602017-05-19 18:53:02 -0700159 if (last_maj == maj && last_min == min) {
160 *bs = last_bs;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200161 return last_fileno;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700162 }
Jens Axboe5e6c2062007-05-16 14:40:29 +0200163
164 last_maj = maj;
165 last_min = min;
166
167 /*
168 * check for this file in our list
169 */
Elliott Hugheseda3a602017-05-19 18:53:02 -0700170 for_each_file(td, f, i) {
Shaohua Li89ac1d42012-06-11 08:56:32 +0200171 if (f->major == maj && f->minor == min) {
172 last_fileno = f->fileno;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700173 last_bs = f->bs;
174 goto out;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200175 }
Elliott Hugheseda3a602017-05-19 18:53:02 -0700176 }
Jens Axboe5e6c2062007-05-16 14:40:29 +0200177
178 strcpy(dev, "/dev");
Jens Axboe42a80e32014-09-16 17:06:52 +0200179 if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
Elliott Hugheseda3a602017-05-19 18:53:02 -0700180 unsigned int this_bs;
Jens Axboe691c8fb2008-03-07 14:26:26 +0100181 int fileno;
182
Jens Axboe42a80e32014-09-16 17:06:52 +0200183 if (td->o.replay_redirect)
184 dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
185 " with: %s\n", maj, min,
186 td->o.replay_redirect);
187 else
188 dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
189
Jens Axboebd6f78b2008-02-01 20:27:52 +0100190 dprint(FD_BLKTRACE, "add devices %s\n", dev);
Jens Axboe49ffb4a2010-08-24 15:01:37 +0200191 fileno = add_file_exclusive(td, dev);
Elliott Hugheseda3a602017-05-19 18:53:02 -0700192
193 if (get_dev_blocksize(dev, &this_bs))
194 this_bs = 512;
195
Jens Axboeb53f2c52014-04-08 21:07:12 -0600196 td->o.open_files++;
Jens Axboe5903e7b2014-02-26 13:42:13 -0800197 td->files[fileno]->major = maj;
198 td->files[fileno]->minor = min;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700199 td->files[fileno]->bs = this_bs;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200200 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
Elliott Hugheseda3a602017-05-19 18:53:02 -0700201
Shaohua Li89ac1d42012-06-11 08:56:32 +0200202 last_fileno = fileno;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700203 last_bs = this_bs;
Jens Axboebd6f78b2008-02-01 20:27:52 +0100204 }
Jens Axboef01b34a2013-11-21 11:13:12 -0700205
Elliott Hugheseda3a602017-05-19 18:53:02 -0700206out:
207 *bs = last_bs;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200208 return last_fileno;
Jens Axboe5e6c2062007-05-16 14:40:29 +0200209}
210
Elliott Hugheseda3a602017-05-19 18:53:02 -0700211static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
212{
213 if (!o->replay_align)
214 return;
215
216 t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
217}
218
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200219/*
220 * Store blk_io_trace data in an ipo for later retrieval.
221 */
Jens Axboefdefd982007-05-15 10:12:26 +0200222static void store_ipo(struct thread_data *td, unsigned long long offset,
Shaohua Li89ac1d42012-06-11 08:56:32 +0200223 unsigned int bytes, int rw, unsigned long long ttime,
Elliott Hugheseda3a602017-05-19 18:53:02 -0700224 int fileno, unsigned int bs)
Jens Axboefdefd982007-05-15 10:12:26 +0200225{
226 struct io_piece *ipo = malloc(sizeof(*ipo));
227
Jens Axboe0d29de82010-09-01 13:54:15 +0200228 init_ipo(ipo);
229
Elliott Hugheseda3a602017-05-19 18:53:02 -0700230 ipo->offset = offset * bs;
231 if (td->o.replay_scale)
232 ipo->offset = ipo->offset / td->o.replay_scale;
233 ipo_bytes_align(td->o.replay_align, ipo);
Jens Axboefdefd982007-05-15 10:12:26 +0200234 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200235 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200236 if (rw)
237 ipo->ddir = DDIR_WRITE;
238 else
239 ipo->ddir = DDIR_READ;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200240 ipo->fileno = fileno;
Jens Axboefdefd982007-05-15 10:12:26 +0200241
Jens Axboebd6f78b2008-02-01 20:27:52 +0100242 dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
243 ipo->ddir, ipo->offset,
244 ipo->len, ipo->delay);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100245 queue_io_piece(td, ipo);
Jens Axboefdefd982007-05-15 10:12:26 +0200246}
247
Jens Axboe0b9d69e2009-09-11 22:29:54 +0200248static void handle_trace_notify(struct blk_io_trace *t)
Jens Axboe691c8fb2008-03-07 14:26:26 +0100249{
250 switch (t->action) {
251 case BLK_TN_PROCESS:
Jens Axboe5bca6f92014-07-14 15:07:11 +0200252 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
Jens Axboed95b34a2013-11-21 09:55:49 -0700253 t->action, t->pid);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100254 break;
255 case BLK_TN_TIMESTAMP:
Jens Axboe5bca6f92014-07-14 15:07:11 +0200256 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
Jens Axboed95b34a2013-11-21 09:55:49 -0700257 t->action, t->pid);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100258 break;
Jens Axboeff58fce2010-08-25 12:02:08 +0200259 case BLK_TN_MESSAGE:
260 break;
Jens Axboe691c8fb2008-03-07 14:26:26 +0100261 default:
262 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
263 break;
264 }
265}
266
Jens Axboe5bca6f92014-07-14 15:07:11 +0200267static void handle_trace_discard(struct thread_data *td,
268 struct blk_io_trace *t,
269 unsigned long long ttime,
Elliott Hugheseda3a602017-05-19 18:53:02 -0700270 unsigned long *ios, unsigned int *rw_bs)
Jens Axboeff58fce2010-08-25 12:02:08 +0200271{
272 struct io_piece *ipo = malloc(sizeof(*ipo));
Elliott Hugheseda3a602017-05-19 18:53:02 -0700273 unsigned int bs;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200274 int fileno;
Jens Axboeff58fce2010-08-25 12:02:08 +0200275
Jens Axboe0d29de82010-09-01 13:54:15 +0200276 init_ipo(ipo);
Elliott Hugheseda3a602017-05-19 18:53:02 -0700277 fileno = trace_add_file(td, t->device, &bs);
Jens Axboeff58fce2010-08-25 12:02:08 +0200278
Jens Axboe5bca6f92014-07-14 15:07:11 +0200279 ios[DDIR_TRIM]++;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700280 if (t->bytes > rw_bs[DDIR_TRIM])
281 rw_bs[DDIR_TRIM] = t->bytes;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200282
Jens Axboeff58fce2010-08-25 12:02:08 +0200283 td->o.size += t->bytes;
284
285 memset(ipo, 0, sizeof(*ipo));
286 INIT_FLIST_HEAD(&ipo->list);
287
Elliott Hugheseda3a602017-05-19 18:53:02 -0700288 ipo->offset = t->sector * bs;
289 if (td->o.replay_scale)
290 ipo->offset = ipo->offset / td->o.replay_scale;
291 ipo_bytes_align(td->o.replay_align, ipo);
Jens Axboeff58fce2010-08-25 12:02:08 +0200292 ipo->len = t->bytes;
293 ipo->delay = ttime / 1000;
294 ipo->ddir = DDIR_TRIM;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200295 ipo->fileno = fileno;
Jens Axboeff58fce2010-08-25 12:02:08 +0200296
297 dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
298 ipo->offset, ipo->len,
299 ipo->delay);
300 queue_io_piece(td, ipo);
301}
302
Jens Axboe691c8fb2008-03-07 14:26:26 +0100303static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
304 unsigned long long ttime, unsigned long *ios,
Elliott Hugheseda3a602017-05-19 18:53:02 -0700305 unsigned int *rw_bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200306{
Elliott Hugheseda3a602017-05-19 18:53:02 -0700307 unsigned int bs;
Jens Axboefdefd982007-05-15 10:12:26 +0200308 int rw;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200309 int fileno;
Jens Axboefdefd982007-05-15 10:12:26 +0200310
Elliott Hugheseda3a602017-05-19 18:53:02 -0700311 fileno = trace_add_file(td, t->device, &bs);
Jens Axboe5e6c2062007-05-16 14:40:29 +0200312
Jens Axboee7a7d702007-05-15 10:13:04 +0200313 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200314
Elliott Hugheseda3a602017-05-19 18:53:02 -0700315 if (t->bytes > rw_bs[rw])
316 rw_bs[rw] = t->bytes;
Jens Axboed84f8d42007-05-16 08:47:46 +0200317
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200318 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200319 td->o.size += t->bytes;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700320 store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200321}
322
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200323/*
Jens Axboe691c8fb2008-03-07 14:26:26 +0100324 * We only care for queue traces, most of the others are side effects
325 * due to internal workings of the block layer.
326 */
327static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
Jens Axboe5bca6f92014-07-14 15:07:11 +0200328 unsigned long *ios, unsigned int *bs)
Jens Axboe691c8fb2008-03-07 14:26:26 +0100329{
Jens Axboe5bca6f92014-07-14 15:07:11 +0200330 static unsigned long long last_ttime;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700331 unsigned long long delay = 0;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200332
Jens Axboe691c8fb2008-03-07 14:26:26 +0100333 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
334 return;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200335
336 if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
337 if (!last_ttime || td->o.no_stall) {
338 last_ttime = t->time;
339 delay = 0;
340 } else {
341 delay = t->time - last_ttime;
342 last_ttime = t->time;
343 }
344 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100345
Elliott Hugheseda3a602017-05-19 18:53:02 -0700346 t_bytes_align(&td->o, t);
347
Jens Axboe691c8fb2008-03-07 14:26:26 +0100348 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboeaec2de22008-04-24 12:44:42 +0200349 handle_trace_notify(t);
Jens Axboeff58fce2010-08-25 12:02:08 +0200350 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
Jens Axboe5bca6f92014-07-14 15:07:11 +0200351 handle_trace_discard(td, t, delay, ios, bs);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100352 else
Jens Axboe5bca6f92014-07-14 15:07:11 +0200353 handle_trace_fs(td, t, delay, ios, bs);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100354}
355
Jens Axboed95b34a2013-11-21 09:55:49 -0700356static void byteswap_trace(struct blk_io_trace *t)
357{
358 t->magic = fio_swap32(t->magic);
359 t->sequence = fio_swap32(t->sequence);
360 t->time = fio_swap64(t->time);
361 t->sector = fio_swap64(t->sector);
362 t->bytes = fio_swap32(t->bytes);
363 t->action = fio_swap32(t->action);
364 t->pid = fio_swap32(t->pid);
365 t->device = fio_swap32(t->device);
366 t->cpu = fio_swap32(t->cpu);
367 t->error = fio_swap16(t->error);
368 t->pdu_len = fio_swap16(t->pdu_len);
369}
370
Jens Axboe5bca6f92014-07-14 15:07:11 +0200371static int t_is_write(struct blk_io_trace *t)
372{
373 return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
374}
375
Elliott Hugheseda3a602017-05-19 18:53:02 -0700376static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
377{
378 if (t->action & BLK_TC_ACT(BLK_TC_READ))
379 return DDIR_READ;
380 else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
381 return DDIR_WRITE;
382 else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
383 return DDIR_TRIM;
384
385 return DDIR_INVAL;
386}
387
388static void depth_inc(struct blk_io_trace *t, int *depth)
389{
390 enum fio_ddir ddir;
391
392 ddir = t_get_ddir(t);
393 if (ddir != DDIR_INVAL)
394 depth[ddir]++;
395}
396
397static void depth_dec(struct blk_io_trace *t, int *depth)
398{
399 enum fio_ddir ddir;
400
401 ddir = t_get_ddir(t);
402 if (ddir != DDIR_INVAL)
403 depth[ddir]--;
404}
405
406static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
407{
408 enum fio_ddir ddir = DDIR_INVAL;
409
410 ddir = t_get_ddir(t);
411 if (ddir != DDIR_INVAL) {
412 depth[ddir] = max(depth[ddir], this_depth[ddir]);
413 this_depth[ddir] = 0;
414 }
415}
416
Jens Axboe691c8fb2008-03-07 14:26:26 +0100417/*
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200418 * Load a blktrace file by reading all the blk_io_trace entries, and storing
419 * them as io_pieces like the fio text version would do.
420 */
Jens Axboed95b34a2013-11-21 09:55:49 -0700421int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200422{
423 struct blk_io_trace t;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200424 unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
425 unsigned int rw_bs[DDIR_RWDIR_CNT];
Jens Axboee2887562007-05-16 09:25:09 +0200426 struct fifo *fifo;
Jens Axboe5903e7b2014-02-26 13:42:13 -0800427 int fd, i, old_state;
Shaohua Li89ac1d42012-06-11 08:56:32 +0200428 struct fio_file *f;
Elliott Hugheseda3a602017-05-19 18:53:02 -0700429 int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200430
431 fd = open(filename, O_RDONLY);
432 if (fd < 0) {
433 td_verror(td, errno, "open blktrace file");
434 return 1;
435 }
436
Jens Axboee2887562007-05-16 09:25:09 +0200437 fifo = fifo_alloc(TRACE_FIFO_SIZE);
438
Jens Axboe8edd9732014-03-01 08:24:03 -0700439 old_state = td_bump_runstate(td, TD_SETTING_UP);
Jens Axboe5903e7b2014-02-26 13:42:13 -0800440
Jens Axboe6df8ada2007-05-15 13:23:19 +0200441 td->o.size = 0;
442
Elliott Hugheseda3a602017-05-19 18:53:02 -0700443 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
444 ios[i] = 0;
445 rw_bs[i] = 0;
446 this_depth[i] = 0;
447 depth[i] = 0;
448 }
449
Jens Axboe4241ea82007-09-12 08:18:36 +0200450 skipped_writes = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200451 do {
Jens Axboee2887562007-05-16 09:25:09 +0200452 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200453
Jens Axboee2887562007-05-16 09:25:09 +0200454 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200455 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200456 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200457 break;
Jens Axboee2887562007-05-16 09:25:09 +0200458 else if (ret < (int) sizeof(t)) {
459 log_err("fio: short fifo get\n");
460 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200461 }
462
Jens Axboed95b34a2013-11-21 09:55:49 -0700463 if (need_swap)
464 byteswap_trace(&t);
465
Jens Axboefb7b71a2007-05-15 08:44:04 +0200466 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100467 log_err("fio: bad magic in blktrace data: %x\n",
468 t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200469 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200470 }
471 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100472 log_err("fio: bad blktrace version %d\n",
473 t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200474 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200475 }
Jens Axboef12b3232007-05-16 12:16:44 +0200476 ret = discard_pdu(td, fifo, fd, &t);
477 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200478 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200479 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200480 } else if (t.pdu_len != ret) {
481 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
482 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200483 }
Jens Axboe691c8fb2008-03-07 14:26:26 +0100484 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600485 if ((t.action & 0xffff) == __BLK_TA_QUEUE)
Elliott Hugheseda3a602017-05-19 18:53:02 -0700486 depth_inc(&t, this_depth);
487 else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
488 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
489 depth_dec(&t, this_depth);
490 else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
491 depth_end(&t, this_depth, depth);
Jens Axboe691c8fb2008-03-07 14:26:26 +0100492
Jens Axboe5bca6f92014-07-14 15:07:11 +0200493 if (t_is_write(&t) && read_only) {
Jens Axboe691c8fb2008-03-07 14:26:26 +0100494 skipped_writes++;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200495 continue;
David Nellans64bbb862010-08-24 22:13:30 +0200496 }
Jens Axboea6edd632008-03-07 21:41:54 +0100497 }
Jens Axboe5bca6f92014-07-14 15:07:11 +0200498
499 handle_trace(td, &t, ios, rw_bs);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200500 } while (1);
501
Jens Axboe85a47ca2012-06-11 13:19:13 +0200502 for (i = 0; i < td->files_index; i++) {
Jens Axboef01b34a2013-11-21 11:13:12 -0700503 f = td->files[i];
Shaohua Li89ac1d42012-06-11 08:56:32 +0200504 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
Jens Axboe85a47ca2012-06-11 13:19:13 +0200505 }
Shaohua Li89ac1d42012-06-11 08:56:32 +0200506
Jens Axboe38470f82007-05-16 09:26:23 +0200507 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200508 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200509
Jens Axboe8edd9732014-03-01 08:24:03 -0700510 td_restore_runstate(td, old_state);
Jens Axboe5903e7b2014-02-26 13:42:13 -0800511
Jens Axboef01b34a2013-11-21 11:13:12 -0700512 if (!td->files_index) {
513 log_err("fio: did not find replay device(s)\n");
514 return 1;
515 }
516
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600517 /*
518 * For stacked devices, we don't always get a COMPLETE event so
519 * the depth grows to insane values. Limit it to something sane(r).
520 */
Elliott Hugheseda3a602017-05-19 18:53:02 -0700521 max_depth = 0;
522 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
523 if (depth[i] > 1024)
524 depth[i] = 1024;
525 else if (!depth[i] && ios[i])
526 depth[i] = 1;
527 max_depth = max(depth[i], max_depth);
528 }
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600529
Jens Axboe4241ea82007-09-12 08:18:36 +0200530 if (skipped_writes)
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100531 log_err("fio: %s skips replay of %lu writes due to read-only\n",
532 td->o.name, skipped_writes);
Jens Axboe4241ea82007-09-12 08:18:36 +0200533
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200534 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
535 log_err("fio: found no ios in blktrace data\n");
536 return 1;
Jens Axboe5bca6f92014-07-14 15:07:11 +0200537 } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200538 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200539 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
540 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200541 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200542 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
543 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200544 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200545 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
546 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
Jens Axboe5bca6f92014-07-14 15:07:11 +0200547 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
Jens Axboed84f8d42007-05-16 08:47:46 +0200548 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200549
550 /*
551 * We need to do direct/raw ios to the device, to avoid getting
Elliott Hugheseda3a602017-05-19 18:53:02 -0700552 * read-ahead in our way. But only do so if the minimum block size
553 * is a multiple of 4k, otherwise we don't know if it's safe to do so.
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200554 */
Elliott Hugheseda3a602017-05-19 18:53:02 -0700555 if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
556 td->o.odirect = 1;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200557
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600558 /*
Elliott Hugheseda3a602017-05-19 18:53:02 -0700559 * If depth wasn't manually set, use probed depth
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600560 */
Elliott Hugheseda3a602017-05-19 18:53:02 -0700561 if (!fio_option_is_set(&td->o, iodepth))
562 td->o.iodepth = td->o.iodepth_low = max_depth;
Jens Axboeeb5fdcf2014-04-08 13:43:19 -0600563
Jens Axboefb7b71a2007-05-15 08:44:04 +0200564 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200565err:
566 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200567 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200568 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200569}