| 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 (sizeof(struct blk_io_trace) * 1000) |
| 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, left; |
| 20 | void *ptr; |
| 21 | int ret; |
| 22 | |
| 23 | total = 0; |
| 24 | ptr = buf; |
| 25 | while (total < TRACE_FIFO_SIZE) { |
| 26 | left = TRACE_FIFO_SIZE - total; |
| 27 | |
| 28 | ret = read(fd, ptr, left); |
| 29 | if (ret < 0) { |
| 30 | td_verror(td, errno, "read blktrace file"); |
| 31 | return -1; |
| 32 | } else if (!ret) |
| 33 | break; |
| 34 | |
| 35 | fifo_put(fifo, ptr, ret); |
| 36 | ptr += ret; |
| 37 | total += ret; |
| 38 | } |
| 39 | |
| 40 | return 0; |
| 41 | } |
| 42 | |
| 43 | /* |
| 44 | * Retrieve 'len' bytes from the fifo, refilling if necessary. |
| 45 | */ |
| 46 | static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd, |
| 47 | void *buf, unsigned int len) |
| 48 | { |
| 49 | int ret; |
| 50 | |
| 51 | if (fifo_len(fifo) >= len) |
| 52 | return fifo_get(fifo, buf, len); |
| 53 | |
| 54 | ret = refill_fifo(td, fifo, fd); |
| 55 | if (ret < 0) |
| 56 | return ret; |
| 57 | |
| 58 | if (fifo_len(fifo) < len) |
| 59 | return 0; |
| 60 | |
| 61 | return fifo_get(fifo, buf, len); |
| 62 | } |
| 63 | |
| 64 | /* |
| 65 | * Just discard the pdu by seeking past it. |
| 66 | */ |
| 67 | static int discard_pdu(int fd, struct blk_io_trace *t) |
| 68 | { |
| 69 | if (t->pdu_len == 0) |
| 70 | return 0; |
| 71 | |
| 72 | if (lseek(fd, t->pdu_len, SEEK_CUR) < 0) |
| 73 | return errno; |
| 74 | |
| 75 | return 0; |
| 76 | } |
| 77 | |
| 78 | /* |
| 79 | * Check if this is a blktrace binary data file. We read a single trace |
| 80 | * into memory and check for the magic signature. |
| 81 | */ |
| 82 | int is_blktrace(const char *filename) |
| 83 | { |
| 84 | struct blk_io_trace t; |
| 85 | int fd, ret; |
| 86 | |
| 87 | fd = open(filename, O_RDONLY); |
| 88 | if (fd < 0) { |
| 89 | perror("open blktrace"); |
| 90 | return 0; |
| 91 | } |
| 92 | |
| 93 | ret = read(fd, &t, sizeof(t)); |
| 94 | close(fd); |
| 95 | |
| 96 | if (ret < 0) { |
| 97 | perror("read blktrace"); |
| 98 | return 0; |
| 99 | } else if (ret != sizeof(t)) { |
| 100 | log_err("fio: short read on blktrace file\n"); |
| 101 | return 0; |
| 102 | } |
| 103 | |
| 104 | if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) |
| 105 | return 1; |
| 106 | |
| 107 | return 0; |
| 108 | } |
| 109 | |
| 110 | /* |
| 111 | * Store blk_io_trace data in an ipo for later retrieval. |
| 112 | */ |
| 113 | static void store_ipo(struct thread_data *td, unsigned long long offset, |
| 114 | unsigned int bytes, int rw, unsigned long long ttime) |
| 115 | { |
| 116 | struct io_piece *ipo = malloc(sizeof(*ipo)); |
| 117 | |
| 118 | memset(ipo, 0, sizeof(*ipo)); |
| 119 | INIT_LIST_HEAD(&ipo->list); |
| 120 | /* |
| 121 | * the 512 is wrong here, it should be the hardware sector size... |
| 122 | */ |
| 123 | ipo->offset = offset * 512; |
| 124 | ipo->len = bytes; |
| 125 | ipo->delay = ttime / 1000; |
| 126 | if (rw) |
| 127 | ipo->ddir = DDIR_WRITE; |
| 128 | else |
| 129 | ipo->ddir = DDIR_READ; |
| 130 | |
| 131 | list_add_tail(&ipo->list, &td->io_log_list); |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * We only care for queue traces, most of the others are side effects |
| 136 | * due to internal workings of the block layer. |
| 137 | */ |
| 138 | static void handle_trace(struct thread_data *td, struct blk_io_trace *t, |
| 139 | unsigned long long ttime, unsigned long *ios, |
| 140 | unsigned int *bs) |
| 141 | { |
| 142 | int rw; |
| 143 | |
| 144 | if ((t->action & 0xffff) != __BLK_TA_QUEUE) |
| 145 | return; |
| 146 | if (t->action & BLK_TC_ACT(BLK_TC_PC)) |
| 147 | return; |
| 148 | |
| 149 | /* |
| 150 | * should not happen, need to look into that... |
| 151 | */ |
| 152 | if (!t->bytes) |
| 153 | return; |
| 154 | |
| 155 | rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0; |
| 156 | |
| 157 | if (t->bytes > bs[rw]) |
| 158 | bs[rw] = t->bytes; |
| 159 | |
| 160 | ios[rw]++; |
| 161 | td->o.size += t->bytes; |
| 162 | store_ipo(td, t->sector, t->bytes, rw, ttime); |
| 163 | } |
| 164 | |
| 165 | /* |
| 166 | * Load a blktrace file by reading all the blk_io_trace entries, and storing |
| 167 | * them as io_pieces like the fio text version would do. |
| 168 | */ |
| 169 | int load_blktrace(struct thread_data *td, const char *filename) |
| 170 | { |
| 171 | unsigned long long ttime, delay; |
| 172 | struct blk_io_trace t; |
| 173 | unsigned long ios[2]; |
| 174 | unsigned int cpu; |
| 175 | unsigned int rw_bs[2]; |
| 176 | struct fifo *fifo; |
| 177 | int fd; |
| 178 | |
| 179 | fd = open(filename, O_RDONLY); |
| 180 | if (fd < 0) { |
| 181 | td_verror(td, errno, "open blktrace file"); |
| 182 | return 1; |
| 183 | } |
| 184 | |
| 185 | fifo = fifo_alloc(TRACE_FIFO_SIZE); |
| 186 | |
| 187 | td->o.size = 0; |
| 188 | |
| 189 | cpu = 0; |
| 190 | ttime = 0; |
| 191 | ios[0] = ios[1] = 0; |
| 192 | rw_bs[0] = rw_bs[1] = 0; |
| 193 | do { |
| 194 | int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); |
| 195 | |
| 196 | if (ret < 0) |
| 197 | goto err; |
| 198 | else if (!ret) |
| 199 | break; |
| 200 | else if (ret < (int) sizeof(t)) { |
| 201 | log_err("fio: short fifo get\n"); |
| 202 | break; |
| 203 | } |
| 204 | |
| 205 | if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { |
| 206 | log_err("fio: bad magic in blktrace data\n"); |
| 207 | goto err; |
| 208 | } |
| 209 | if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { |
| 210 | log_err("fio: bad blktrace version %d\n", t.magic & 0xff); |
| 211 | goto err; |
| 212 | } |
| 213 | ret = discard_pdu(fd, &t); |
| 214 | if (ret) { |
| 215 | td_verror(td, ret, "blktrace lseek"); |
| 216 | goto err; |
| 217 | } |
| 218 | if (!ttime) { |
| 219 | ttime = t.time; |
| 220 | cpu = t.cpu; |
| 221 | } |
| 222 | delay = 0; |
| 223 | if (cpu == t.cpu) |
| 224 | delay = t.time - ttime; |
| 225 | handle_trace(td, &t, delay, ios, rw_bs); |
| 226 | ttime = t.time; |
| 227 | cpu = t.cpu; |
| 228 | } while (1); |
| 229 | |
| 230 | fifo_free(fifo); |
| 231 | close(fd); |
| 232 | |
| 233 | if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { |
| 234 | log_err("fio: found no ios in blktrace data\n"); |
| 235 | return 1; |
| 236 | } else if (ios[DDIR_READ] && !ios[DDIR_READ]) { |
| 237 | td->o.td_ddir = TD_DDIR_READ; |
| 238 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
| 239 | } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { |
| 240 | td->o.td_ddir = TD_DDIR_WRITE; |
| 241 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
| 242 | } else { |
| 243 | td->o.td_ddir = TD_DDIR_RW; |
| 244 | td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; |
| 245 | td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; |
| 246 | } |
| 247 | |
| 248 | /* |
| 249 | * We need to do direct/raw ios to the device, to avoid getting |
| 250 | * read-ahead in our way. |
| 251 | */ |
| 252 | td->o.odirect = 1; |
| 253 | |
| 254 | return 0; |
| 255 | err: |
| 256 | close(fd); |
| 257 | fifo_free(fifo); |
| 258 | return 1; |
| 259 | } |