| 1 | /* |
| 2 | * blktrace support code for fio |
| 3 | */ |
| 4 | #include <stdio.h> |
| 5 | #include <stdlib.h> |
| 6 | |
| 7 | #include "list.h" |
| 8 | #include "fio.h" |
| 9 | #include "blktrace_api.h" |
| 10 | |
| 11 | #define TRACE_FIFO_SIZE 65536 |
| 12 | |
| 13 | /* |
| 14 | * fifo refill frontend, to avoid reading data in trace sized bites |
| 15 | */ |
| 16 | static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd) |
| 17 | { |
| 18 | char buf[TRACE_FIFO_SIZE]; |
| 19 | unsigned int total; |
| 20 | int ret; |
| 21 | |
| 22 | total = sizeof(buf); |
| 23 | if (total > fifo_room(fifo)) |
| 24 | total = fifo_room(fifo); |
| 25 | |
| 26 | ret = read(fd, buf, total); |
| 27 | if (ret < 0) { |
| 28 | td_verror(td, errno, "read blktrace file"); |
| 29 | return -1; |
| 30 | } |
| 31 | |
| 32 | if (ret > 0) |
| 33 | ret = fifo_put(fifo, buf, ret); |
| 34 | |
| 35 | return ret; |
| 36 | } |
| 37 | |
| 38 | /* |
| 39 | * Retrieve 'len' bytes from the fifo, refilling if necessary. |
| 40 | */ |
| 41 | static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd, |
| 42 | void *buf, unsigned int len) |
| 43 | { |
| 44 | if (fifo_len(fifo) < len) { |
| 45 | int ret = refill_fifo(td, fifo, fd); |
| 46 | |
| 47 | if (ret < 0) |
| 48 | return ret; |
| 49 | } |
| 50 | |
| 51 | return fifo_get(fifo, buf, len); |
| 52 | } |
| 53 | |
| 54 | /* |
| 55 | * Just discard the pdu by seeking past it. |
| 56 | */ |
| 57 | static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd, |
| 58 | struct blk_io_trace *t) |
| 59 | { |
| 60 | if (t->pdu_len == 0) |
| 61 | return 0; |
| 62 | |
| 63 | return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len); |
| 64 | } |
| 65 | |
| 66 | /* |
| 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 | */ |
| 70 | int 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 | |
| 98 | /* |
| 99 | * Store blk_io_trace data in an ipo for later retrieval. |
| 100 | */ |
| 101 | static void store_ipo(struct thread_data *td, unsigned long long offset, |
| 102 | unsigned int bytes, int rw, unsigned long long ttime) |
| 103 | { |
| 104 | struct io_piece *ipo = malloc(sizeof(*ipo)); |
| 105 | |
| 106 | memset(ipo, 0, sizeof(*ipo)); |
| 107 | INIT_LIST_HEAD(&ipo->list); |
| 108 | /* |
| 109 | * the 512 is wrong here, it should be the hardware sector size... |
| 110 | */ |
| 111 | ipo->offset = offset * 512; |
| 112 | ipo->len = bytes; |
| 113 | ipo->delay = ttime / 1000; |
| 114 | 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 | |
| 122 | /* |
| 123 | * We only care for queue traces, most of the others are side effects |
| 124 | * due to internal workings of the block layer. |
| 125 | */ |
| 126 | static void handle_trace(struct thread_data *td, struct blk_io_trace *t, |
| 127 | unsigned long long ttime, unsigned long *ios, |
| 128 | unsigned int *bs) |
| 129 | { |
| 130 | int rw; |
| 131 | |
| 132 | if ((t->action & 0xffff) != __BLK_TA_QUEUE) |
| 133 | return; |
| 134 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) |
| 135 | return; |
| 136 | |
| 137 | /* |
| 138 | * should not happen, need to look into that... |
| 139 | */ |
| 140 | if (!t->bytes) |
| 141 | return; |
| 142 | |
| 143 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
| 144 | |
| 145 | if (t->bytes > bs[rw]) |
| 146 | bs[rw] = t->bytes; |
| 147 | |
| 148 | ios[rw]++; |
| 149 | td->o.size += t->bytes; |
| 150 | store_ipo(td, t->sector, t->bytes, rw, ttime); |
| 151 | } |
| 152 | |
| 153 | /* |
| 154 | * Load a blktrace file by reading all the blk_io_trace entries, and storing |
| 155 | * them as io_pieces like the fio text version would do. |
| 156 | */ |
| 157 | int load_blktrace(struct thread_data *td, const char *filename) |
| 158 | { |
| 159 | unsigned long long ttime, delay; |
| 160 | struct blk_io_trace t; |
| 161 | unsigned long ios[2]; |
| 162 | unsigned int cpu; |
| 163 | unsigned int rw_bs[2]; |
| 164 | struct fifo *fifo; |
| 165 | int fd; |
| 166 | |
| 167 | fd = open(filename, O_RDONLY); |
| 168 | if (fd < 0) { |
| 169 | td_verror(td, errno, "open blktrace file"); |
| 170 | return 1; |
| 171 | } |
| 172 | |
| 173 | fifo = fifo_alloc(TRACE_FIFO_SIZE); |
| 174 | |
| 175 | td->o.size = 0; |
| 176 | |
| 177 | cpu = 0; |
| 178 | ttime = 0; |
| 179 | ios[0] = ios[1] = 0; |
| 180 | rw_bs[0] = rw_bs[1] = 0; |
| 181 | do { |
| 182 | int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); |
| 183 | |
| 184 | if (ret < 0) |
| 185 | goto err; |
| 186 | else if (!ret) |
| 187 | break; |
| 188 | else if (ret < (int) sizeof(t)) { |
| 189 | log_err("fio: short fifo get\n"); |
| 190 | break; |
| 191 | } |
| 192 | |
| 193 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { |
| 194 | log_err("fio: bad magic in blktrace data: %x\n", t.magic); |
| 195 | goto err; |
| 196 | } |
| 197 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { |
| 198 | log_err("fio: bad blktrace version %d\n", t.magic & 0xff); |
| 199 | goto err; |
| 200 | } |
| 201 | ret = discard_pdu(td, fifo, fd, &t); |
| 202 | if (ret < 0) { |
| 203 | td_verror(td, ret, "blktrace lseek"); |
| 204 | goto err; |
| 205 | } else if (t.pdu_len != ret) { |
| 206 | log_err("fio: discarded %d of %d\n", ret, t.pdu_len); |
| 207 | goto err; |
| 208 | } |
| 209 | if (!ttime) { |
| 210 | ttime = t.time; |
| 211 | cpu = t.cpu; |
| 212 | } |
| 213 | delay = 0; |
| 214 | if (cpu == t.cpu) |
| 215 | delay = t.time - ttime; |
| 216 | handle_trace(td, &t, delay, ios, rw_bs); |
| 217 | ttime = t.time; |
| 218 | cpu = t.cpu; |
| 219 | } while (1); |
| 220 | |
| 221 | fifo_free(fifo); |
| 222 | close(fd); |
| 223 | |
| 224 | if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
| 225 | log_err("fio: found no ios in blktrace data\n"); |
| 226 | return 1; |
| 227 | } else if (ios[DDIR_READ] && !ios[DDIR_READ]) { |
| 228 | td->o.td_ddir = TD_DDIR_READ; |
| 229 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
| 230 | } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { |
| 231 | td->o.td_ddir = TD_DDIR_WRITE; |
| 232 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
| 233 | } else { |
| 234 | td->o.td_ddir = TD_DDIR_RW; |
| 235 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
| 236 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * We need to do direct/raw ios to the device, to avoid getting |
| 241 | * read-ahead in our way. |
| 242 | */ |
| 243 | td->o.odirect = 1; |
| 244 | |
| 245 | return 0; |
| 246 | err: |
| 247 | close(fd); |
| 248 | fifo_free(fifo); |
| 249 | return 1; |
| 250 | } |