blob: 5825d29714e13627dd9712699a904f752445cd70 [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 Axboe8c1fdf02007-05-15 11:54:21 +02006
Jens Axboefb7b71a2007-05-15 08:44:04 +02007#include "list.h"
8#include "fio.h"
9#include "blktrace_api.h"
10
Jens Axboef12b3232007-05-16 12:16:44 +020011#define TRACE_FIFO_SIZE 65536
Jens Axboee2887562007-05-16 09:25:09 +020012
13/*
14 * fifo refill frontend, to avoid reading data in trace sized bites
15 */
16static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
17{
18 char buf[TRACE_FIFO_SIZE];
Jens Axboef12b3232007-05-16 12:16:44 +020019 unsigned int total;
Jens Axboee2887562007-05-16 09:25:09 +020020 int ret;
21
Jens Axboef12b3232007-05-16 12:16:44 +020022 total = sizeof(buf);
23 if (total > fifo_room(fifo))
24 total = fifo_room(fifo);
Jens Axboee2887562007-05-16 09:25:09 +020025
Jens Axboef12b3232007-05-16 12:16:44 +020026 ret = read(fd, buf, total);
27 if (ret < 0) {
28 td_verror(td, errno, "read blktrace file");
29 return -1;
Jens Axboee2887562007-05-16 09:25:09 +020030 }
31
Jens Axboef12b3232007-05-16 12:16:44 +020032 if (ret > 0)
33 ret = fifo_put(fifo, buf, ret);
34
35 return ret;
Jens Axboee2887562007-05-16 09:25:09 +020036}
37
38/*
39 * Retrieve 'len' bytes from the fifo, refilling if necessary.
40 */
41static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
42 void *buf, unsigned int len)
43{
Jens Axboef12b3232007-05-16 12:16:44 +020044 if (fifo_len(fifo) < len) {
45 int ret = refill_fifo(td, fifo, fd);
Jens Axboee2887562007-05-16 09:25:09 +020046
Jens Axboef12b3232007-05-16 12:16:44 +020047 if (ret < 0)
48 return ret;
49 }
Jens Axboee2887562007-05-16 09:25:09 +020050
51 return fifo_get(fifo, buf, len);
52}
53
Jens Axboe8c1fdf02007-05-15 11:54:21 +020054/*
55 * Just discard the pdu by seeking past it.
56 */
Jens Axboef12b3232007-05-16 12:16:44 +020057static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
58 struct blk_io_trace *t)
Jens Axboefb7b71a2007-05-15 08:44:04 +020059{
60 if (t->pdu_len == 0)
61 return 0;
62
Jens Axboef12b3232007-05-16 12:16:44 +020063 return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
Jens Axboefb7b71a2007-05-15 08:44:04 +020064}
65
Jens Axboe8c1fdf02007-05-15 11:54:21 +020066/*
67 * Check if this is a blktrace binary data file. We read a single trace
68 * into memory and check for the magic signature.
69 */
Jens Axboefb7b71a2007-05-15 08:44:04 +020070int is_blktrace(const char *filename)
71{
72 struct blk_io_trace t;
73 int fd, ret;
74
75 fd = open(filename, O_RDONLY);
76 if (fd < 0) {
77 perror("open blktrace");
78 return 0;
79 }
80
81 ret = read(fd, &t, sizeof(t));
82 close(fd);
83
84 if (ret < 0) {
85 perror("read blktrace");
86 return 0;
87 } else if (ret != sizeof(t)) {
88 log_err("fio: short read on blktrace file\n");
89 return 0;
90 }
91
92 if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
93 return 1;
94
95 return 0;
96}
97
Jens Axboe8c1fdf02007-05-15 11:54:21 +020098/*
99 * Store blk_io_trace data in an ipo for later retrieval.
100 */
Jens Axboefdefd982007-05-15 10:12:26 +0200101static void store_ipo(struct thread_data *td, unsigned long long offset,
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200102 unsigned int bytes, int rw, unsigned long long ttime)
Jens Axboefdefd982007-05-15 10:12:26 +0200103{
104 struct io_piece *ipo = malloc(sizeof(*ipo));
105
106 memset(ipo, 0, sizeof(*ipo));
107 INIT_LIST_HEAD(&ipo->list);
Jens Axboea2eea812007-05-15 13:10:41 +0200108 /*
109 * the 512 is wrong here, it should be the hardware sector size...
110 */
111 ipo->offset = offset * 512;
Jens Axboefdefd982007-05-15 10:12:26 +0200112 ipo->len = bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200113 ipo->delay = ttime / 1000;
Jens Axboefdefd982007-05-15 10:12:26 +0200114 if (rw)
115 ipo->ddir = DDIR_WRITE;
116 else
117 ipo->ddir = DDIR_READ;
118
119 list_add_tail(&ipo->list, &td->io_log_list);
120}
121
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200122/*
123 * We only care for queue traces, most of the others are side effects
124 * due to internal workings of the block layer.
125 */
126static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
Jens Axboed84f8d42007-05-16 08:47:46 +0200127 unsigned long long ttime, unsigned long *ios,
128 unsigned int *bs)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200129{
Jens Axboefdefd982007-05-15 10:12:26 +0200130 int rw;
131
132 if ((t->action & 0xffff) != __BLK_TA_QUEUE)
133 return;
Jens Axboea2eea812007-05-15 13:10:41 +0200134 if (t->action & BLK_TC_ACT(BLK_TC_PC))
135 return;
Jens Axboe3f3d3612007-05-16 12:29:06 +0200136 if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
Jens Axboea2eea812007-05-15 13:10:41 +0200137 return;
Jens Axboefdefd982007-05-15 10:12:26 +0200138
Jens Axboee7a7d702007-05-15 10:13:04 +0200139 rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200140
141 if (t->bytes > bs[rw])
142 bs[rw] = t->bytes;
143
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200144 ios[rw]++;
Jens Axboe6df8ada2007-05-15 13:23:19 +0200145 td->o.size += t->bytes;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200146 store_ipo(td, t->sector, t->bytes, rw, ttime);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200147}
148
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200149/*
150 * Load a blktrace file by reading all the blk_io_trace entries, and storing
151 * them as io_pieces like the fio text version would do.
152 */
Jens Axboefb7b71a2007-05-15 08:44:04 +0200153int load_blktrace(struct thread_data *td, const char *filename)
154{
Jens Axboea61edde2007-05-15 14:29:58 +0200155 unsigned long long ttime, delay;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200156 struct blk_io_trace t;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200157 unsigned long ios[2];
Jens Axboea61edde2007-05-15 14:29:58 +0200158 unsigned int cpu;
Jens Axboed84f8d42007-05-16 08:47:46 +0200159 unsigned int rw_bs[2];
Jens Axboee2887562007-05-16 09:25:09 +0200160 struct fifo *fifo;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200161 int fd;
162
163 fd = open(filename, O_RDONLY);
164 if (fd < 0) {
165 td_verror(td, errno, "open blktrace file");
166 return 1;
167 }
168
Jens Axboee2887562007-05-16 09:25:09 +0200169 fifo = fifo_alloc(TRACE_FIFO_SIZE);
170
Jens Axboe6df8ada2007-05-15 13:23:19 +0200171 td->o.size = 0;
172
Jens Axboea61edde2007-05-15 14:29:58 +0200173 cpu = 0;
Jens Axboed84f8d42007-05-16 08:47:46 +0200174 ttime = 0;
175 ios[0] = ios[1] = 0;
176 rw_bs[0] = rw_bs[1] = 0;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200177 do {
Jens Axboee2887562007-05-16 09:25:09 +0200178 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
Jens Axboefb7b71a2007-05-15 08:44:04 +0200179
Jens Axboee2887562007-05-16 09:25:09 +0200180 if (ret < 0)
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200181 goto err;
Jens Axboee2887562007-05-16 09:25:09 +0200182 else if (!ret)
Jens Axboefb7b71a2007-05-15 08:44:04 +0200183 break;
Jens Axboee2887562007-05-16 09:25:09 +0200184 else if (ret < (int) sizeof(t)) {
185 log_err("fio: short fifo get\n");
186 break;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200187 }
188
189 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
Jens Axboef12b3232007-05-16 12:16:44 +0200190 log_err("fio: bad magic in blktrace data: %x\n", t.magic);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200191 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200192 }
193 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
194 log_err("fio: bad blktrace version %d\n", t.magic & 0xff);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200195 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200196 }
Jens Axboef12b3232007-05-16 12:16:44 +0200197 ret = discard_pdu(td, fifo, fd, &t);
198 if (ret < 0) {
Jens Axboefb7b71a2007-05-15 08:44:04 +0200199 td_verror(td, ret, "blktrace lseek");
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200200 goto err;
Jens Axboef12b3232007-05-16 12:16:44 +0200201 } else if (t.pdu_len != ret) {
202 log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
203 goto err;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200204 }
Jens Axboea61edde2007-05-15 14:29:58 +0200205 if (!ttime) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200206 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200207 cpu = t.cpu;
208 }
209 delay = 0;
210 if (cpu == t.cpu)
211 delay = t.time - ttime;
Jens Axboed84f8d42007-05-16 08:47:46 +0200212 handle_trace(td, &t, delay, ios, rw_bs);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200213 ttime = t.time;
Jens Axboea61edde2007-05-15 14:29:58 +0200214 cpu = t.cpu;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200215 } while (1);
216
Jens Axboe38470f82007-05-16 09:26:23 +0200217 fifo_free(fifo);
Jens Axboefb7b71a2007-05-15 08:44:04 +0200218 close(fd);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200219
220 if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
221 log_err("fio: found no ios in blktrace data\n");
222 return 1;
Jens Axboed84f8d42007-05-16 08:47:46 +0200223 } else if (ios[DDIR_READ] && !ios[DDIR_READ]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200224 td->o.td_ddir = TD_DDIR_READ;
Jens Axboed84f8d42007-05-16 08:47:46 +0200225 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
226 } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200227 td->o.td_ddir = TD_DDIR_WRITE;
Jens Axboed84f8d42007-05-16 08:47:46 +0200228 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
229 } else {
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200230 td->o.td_ddir = TD_DDIR_RW;
Jens Axboed84f8d42007-05-16 08:47:46 +0200231 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
232 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
233 }
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200234
235 /*
236 * We need to do direct/raw ios to the device, to avoid getting
237 * read-ahead in our way.
238 */
239 td->o.odirect = 1;
240
Jens Axboefb7b71a2007-05-15 08:44:04 +0200241 return 0;
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200242err:
243 close(fd);
Jens Axboe38470f82007-05-16 09:26:23 +0200244 fifo_free(fifo);
Jens Axboe8c1fdf02007-05-15 11:54:21 +0200245 return 1;
Jens Axboefb7b71a2007-05-15 08:44:04 +0200246}