| 1 | /* |
| 2 | * net engine |
| 3 | * |
| 4 | * IO engine that reads/writes to/from sockets. |
| 5 | * |
| 6 | */ |
| 7 | #include <stdio.h> |
| 8 | #include <stdlib.h> |
| 9 | #include <unistd.h> |
| 10 | #include <errno.h> |
| 11 | #include <assert.h> |
| 12 | #include <netinet/in.h> |
| 13 | #include <arpa/inet.h> |
| 14 | #include <netdb.h> |
| 15 | #include <sys/poll.h> |
| 16 | |
| 17 | #include "../fio.h" |
| 18 | |
| 19 | struct netio_data { |
| 20 | int listenfd; |
| 21 | int send_to_net; |
| 22 | int use_splice; |
| 23 | int pipes[2]; |
| 24 | char host[64]; |
| 25 | struct sockaddr_in addr; |
| 26 | }; |
| 27 | |
| 28 | static int fio_netio_prep(struct thread_data *td, struct io_u *io_u) |
| 29 | { |
| 30 | struct netio_data *nd = td->io_ops->data; |
| 31 | struct fio_file *f = io_u->file; |
| 32 | |
| 33 | /* |
| 34 | * Make sure we don't see spurious reads to a receiver, and vice versa |
| 35 | */ |
| 36 | if ((nd->send_to_net && io_u->ddir == DDIR_READ) || |
| 37 | (!nd->send_to_net && io_u->ddir == DDIR_WRITE)) { |
| 38 | td_verror(td, EINVAL, "bad direction"); |
| 39 | return 1; |
| 40 | } |
| 41 | |
| 42 | if (io_u->ddir == DDIR_SYNC) |
| 43 | return 0; |
| 44 | if (io_u->offset == f->last_completed_pos) |
| 45 | return 0; |
| 46 | |
| 47 | /* |
| 48 | * If offset is different from last end position, it's a seek. |
| 49 | * As network io is purely sequential, we don't allow seeks. |
| 50 | */ |
| 51 | td_verror(td, EINVAL, "cannot seek"); |
| 52 | return 1; |
| 53 | } |
| 54 | |
| 55 | static int splice_io_u(int fdin, int fdout, unsigned int len) |
| 56 | { |
| 57 | int bytes = 0; |
| 58 | |
| 59 | while (len) { |
| 60 | int ret = splice(fdin, NULL, fdout, NULL, len, 0); |
| 61 | |
| 62 | if (ret < 0) { |
| 63 | if (!bytes) |
| 64 | bytes = ret; |
| 65 | |
| 66 | break; |
| 67 | } else if (!ret) |
| 68 | break; |
| 69 | |
| 70 | bytes += ret; |
| 71 | len -= ret; |
| 72 | } |
| 73 | |
| 74 | return bytes; |
| 75 | } |
| 76 | |
| 77 | /* |
| 78 | * Receive bytes from a socket and fill them into the internal pipe |
| 79 | */ |
| 80 | static int splice_in(struct thread_data *td, struct io_u *io_u) |
| 81 | { |
| 82 | struct netio_data *nd = td->io_ops->data; |
| 83 | |
| 84 | return splice_io_u(io_u->file->fd, nd->pipes[1], io_u->xfer_buflen); |
| 85 | } |
| 86 | |
| 87 | /* |
| 88 | * Transmit 'len' bytes from the internal pipe |
| 89 | */ |
| 90 | static int splice_out(struct thread_data *td, struct io_u *io_u, |
| 91 | unsigned int len) |
| 92 | { |
| 93 | struct netio_data *nd = td->io_ops->data; |
| 94 | |
| 95 | return splice_io_u(nd->pipes[0], io_u->file->fd, len); |
| 96 | } |
| 97 | |
| 98 | static int vmsplice_io_u(struct io_u *io_u, int fd, unsigned int len) |
| 99 | { |
| 100 | struct iovec iov = { |
| 101 | .iov_base = io_u->xfer_buf, |
| 102 | .iov_len = len, |
| 103 | }; |
| 104 | int bytes = 0; |
| 105 | |
| 106 | while (iov.iov_len) { |
| 107 | int ret = vmsplice(fd, &iov, 1, SPLICE_F_MOVE); |
| 108 | |
| 109 | if (ret < 0) { |
| 110 | if (!bytes) |
| 111 | bytes = ret; |
| 112 | break; |
| 113 | } else if (!ret) |
| 114 | break; |
| 115 | |
| 116 | iov.iov_len -= ret; |
| 117 | iov.iov_base += ret; |
| 118 | bytes += ret; |
| 119 | } |
| 120 | |
| 121 | return bytes; |
| 122 | |
| 123 | } |
| 124 | |
| 125 | /* |
| 126 | * vmsplice() pipe to io_u buffer |
| 127 | */ |
| 128 | static int vmsplice_io_u_out(struct thread_data *td, struct io_u *io_u, |
| 129 | unsigned int len) |
| 130 | { |
| 131 | struct netio_data *nd = td->io_ops->data; |
| 132 | |
| 133 | return vmsplice_io_u(io_u, nd->pipes[0], len); |
| 134 | } |
| 135 | |
| 136 | /* |
| 137 | * vmsplice() io_u to pipe |
| 138 | */ |
| 139 | static int vmsplice_io_u_in(struct thread_data *td, struct io_u *io_u) |
| 140 | { |
| 141 | struct netio_data *nd = td->io_ops->data; |
| 142 | |
| 143 | return vmsplice_io_u(io_u, nd->pipes[1], io_u->xfer_buflen); |
| 144 | } |
| 145 | |
| 146 | /* |
| 147 | * splice receive - transfer socket data into a pipe using splice, then map |
| 148 | * that pipe data into the io_u using vmsplice. |
| 149 | */ |
| 150 | static int fio_netio_splice_in(struct thread_data *td, struct io_u *io_u) |
| 151 | { |
| 152 | int ret; |
| 153 | |
| 154 | ret = splice_in(td, io_u); |
| 155 | if (ret > 0) |
| 156 | return vmsplice_io_u_out(td, io_u, ret); |
| 157 | |
| 158 | return ret; |
| 159 | } |
| 160 | |
| 161 | /* |
| 162 | * splice transmit - map data from the io_u into a pipe by using vmsplice, |
| 163 | * then transfer that pipe to a socket using splice. |
| 164 | */ |
| 165 | static int fio_netio_splice_out(struct thread_data *td, struct io_u *io_u) |
| 166 | { |
| 167 | int ret; |
| 168 | |
| 169 | ret = vmsplice_io_u_in(td, io_u); |
| 170 | if (ret > 0) |
| 171 | return splice_out(td, io_u, ret); |
| 172 | |
| 173 | return ret; |
| 174 | } |
| 175 | |
| 176 | static int fio_netio_send(struct thread_data *td, struct io_u *io_u) |
| 177 | { |
| 178 | int flags = 0; |
| 179 | |
| 180 | /* |
| 181 | * if we are going to write more, set MSG_MORE |
| 182 | */ |
| 183 | if (td->this_io_bytes[DDIR_WRITE] + io_u->xfer_buflen < td->o.size) |
| 184 | flags = MSG_MORE; |
| 185 | |
| 186 | return send(io_u->file->fd, io_u->xfer_buf, io_u->xfer_buflen, flags); |
| 187 | } |
| 188 | |
| 189 | static int fio_netio_recv(struct io_u *io_u) |
| 190 | { |
| 191 | int flags = MSG_WAITALL; |
| 192 | |
| 193 | return recv(io_u->file->fd, io_u->xfer_buf, io_u->xfer_buflen, flags); |
| 194 | } |
| 195 | |
| 196 | static int fio_netio_queue(struct thread_data *td, struct io_u *io_u) |
| 197 | { |
| 198 | struct netio_data *nd = td->io_ops->data; |
| 199 | int ret; |
| 200 | |
| 201 | fio_ro_check(td, io_u); |
| 202 | |
| 203 | if (io_u->ddir == DDIR_WRITE) { |
| 204 | if (nd->use_splice) |
| 205 | ret = fio_netio_splice_out(td, io_u); |
| 206 | else |
| 207 | ret = fio_netio_send(td, io_u); |
| 208 | } else if (io_u->ddir == DDIR_READ) { |
| 209 | if (nd->use_splice) |
| 210 | ret = fio_netio_splice_in(td, io_u); |
| 211 | else |
| 212 | ret = fio_netio_recv(io_u); |
| 213 | } else |
| 214 | ret = 0; /* must be a SYNC */ |
| 215 | |
| 216 | if (ret != (int) io_u->xfer_buflen) { |
| 217 | if (ret >= 0) { |
| 218 | io_u->resid = io_u->xfer_buflen - ret; |
| 219 | io_u->error = 0; |
| 220 | return FIO_Q_COMPLETED; |
| 221 | } else |
| 222 | io_u->error = errno; |
| 223 | } |
| 224 | |
| 225 | if (io_u->error) |
| 226 | td_verror(td, io_u->error, "xfer"); |
| 227 | |
| 228 | return FIO_Q_COMPLETED; |
| 229 | } |
| 230 | |
| 231 | static int fio_netio_connect(struct thread_data *td, struct fio_file *f) |
| 232 | { |
| 233 | struct netio_data *nd = td->io_ops->data; |
| 234 | |
| 235 | f->fd = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP); |
| 236 | if (f->fd < 0) { |
| 237 | td_verror(td, errno, "socket"); |
| 238 | return 1; |
| 239 | } |
| 240 | |
| 241 | if (connect(f->fd, (struct sockaddr *) &nd->addr, sizeof(nd->addr)) < 0) { |
| 242 | td_verror(td, errno, "connect"); |
| 243 | return 1; |
| 244 | } |
| 245 | |
| 246 | return 0; |
| 247 | } |
| 248 | |
| 249 | static int fio_netio_accept(struct thread_data *td, struct fio_file *f) |
| 250 | { |
| 251 | struct netio_data *nd = td->io_ops->data; |
| 252 | socklen_t socklen = sizeof(nd->addr); |
| 253 | struct pollfd pfd; |
| 254 | int ret; |
| 255 | |
| 256 | log_info("fio: waiting for connection\n"); |
| 257 | |
| 258 | /* |
| 259 | * Accept loop. poll for incoming events, accept them. Repeat until we |
| 260 | * have all connections. |
| 261 | */ |
| 262 | while (!td->terminate) { |
| 263 | pfd.fd = nd->listenfd; |
| 264 | pfd.events = POLLIN; |
| 265 | |
| 266 | ret = poll(&pfd, 1, -1); |
| 267 | if (ret < 0) { |
| 268 | if (errno == EINTR) |
| 269 | continue; |
| 270 | |
| 271 | td_verror(td, errno, "poll"); |
| 272 | break; |
| 273 | } else if (!ret) |
| 274 | continue; |
| 275 | |
| 276 | /* |
| 277 | * should be impossible |
| 278 | */ |
| 279 | if (!(pfd.revents & POLLIN)) |
| 280 | continue; |
| 281 | |
| 282 | f->fd = accept(nd->listenfd, (struct sockaddr *) &nd->addr, &socklen); |
| 283 | if (f->fd < 0) { |
| 284 | td_verror(td, errno, "accept"); |
| 285 | return 1; |
| 286 | } |
| 287 | break; |
| 288 | } |
| 289 | |
| 290 | return 0; |
| 291 | } |
| 292 | |
| 293 | static int fio_netio_open_file(struct thread_data *td, struct fio_file *f) |
| 294 | { |
| 295 | if (td_read(td)) |
| 296 | return fio_netio_accept(td, f); |
| 297 | else |
| 298 | return fio_netio_connect(td, f); |
| 299 | } |
| 300 | |
| 301 | static int fio_netio_setup_connect(struct thread_data *td, const char *host, |
| 302 | unsigned short port) |
| 303 | { |
| 304 | struct netio_data *nd = td->io_ops->data; |
| 305 | |
| 306 | nd->addr.sin_family = AF_INET; |
| 307 | nd->addr.sin_port = htons(port); |
| 308 | |
| 309 | if (inet_aton(host, &nd->addr.sin_addr) != 1) { |
| 310 | struct hostent *hent; |
| 311 | |
| 312 | hent = gethostbyname(host); |
| 313 | if (!hent) { |
| 314 | td_verror(td, errno, "gethostbyname"); |
| 315 | return 1; |
| 316 | } |
| 317 | |
| 318 | memcpy(&nd->addr.sin_addr, hent->h_addr, 4); |
| 319 | } |
| 320 | |
| 321 | return 0; |
| 322 | } |
| 323 | |
| 324 | static int fio_netio_setup_listen(struct thread_data *td, short port) |
| 325 | { |
| 326 | struct netio_data *nd = td->io_ops->data; |
| 327 | int fd, opt; |
| 328 | |
| 329 | fd = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP); |
| 330 | if (fd < 0) { |
| 331 | td_verror(td, errno, "socket"); |
| 332 | return 1; |
| 333 | } |
| 334 | |
| 335 | opt = 1; |
| 336 | if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) { |
| 337 | td_verror(td, errno, "setsockopt"); |
| 338 | return 1; |
| 339 | } |
| 340 | #ifdef SO_REUSEPORT |
| 341 | if (setsockopt(fd, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt)) < 0) { |
| 342 | td_verror(td, errno, "setsockopt"); |
| 343 | return 1; |
| 344 | } |
| 345 | #endif |
| 346 | |
| 347 | nd->addr.sin_family = AF_INET; |
| 348 | nd->addr.sin_addr.s_addr = htonl(INADDR_ANY); |
| 349 | nd->addr.sin_port = htons(port); |
| 350 | |
| 351 | if (bind(fd, (struct sockaddr *) &nd->addr, sizeof(nd->addr)) < 0) { |
| 352 | td_verror(td, errno, "bind"); |
| 353 | return 1; |
| 354 | } |
| 355 | if (listen(fd, 1) < 0) { |
| 356 | td_verror(td, errno, "listen"); |
| 357 | return 1; |
| 358 | } |
| 359 | |
| 360 | nd->listenfd = fd; |
| 361 | return 0; |
| 362 | } |
| 363 | |
| 364 | static int fio_netio_init(struct thread_data *td) |
| 365 | { |
| 366 | struct netio_data *nd = td->io_ops->data; |
| 367 | unsigned short port; |
| 368 | char host[64], buf[128]; |
| 369 | char *sep; |
| 370 | int ret; |
| 371 | |
| 372 | if (td_rw(td)) { |
| 373 | log_err("fio: network connections must be read OR write\n"); |
| 374 | return 1; |
| 375 | } |
| 376 | if (td_random(td)) { |
| 377 | log_err("fio: network IO can't be random\n"); |
| 378 | return 1; |
| 379 | } |
| 380 | |
| 381 | strcpy(buf, td->o.filename); |
| 382 | |
| 383 | sep = strchr(buf, '/'); |
| 384 | if (!sep) { |
| 385 | log_err("fio: bad network host/port <<%s>>\n", td->o.filename); |
| 386 | return 1; |
| 387 | } |
| 388 | |
| 389 | *sep = '\0'; |
| 390 | sep++; |
| 391 | strcpy(host, buf); |
| 392 | port = atoi(sep); |
| 393 | |
| 394 | if (td_read(td)) { |
| 395 | nd->send_to_net = 0; |
| 396 | ret = fio_netio_setup_listen(td, port); |
| 397 | } else { |
| 398 | nd->send_to_net = 1; |
| 399 | ret = fio_netio_setup_connect(td, host, port); |
| 400 | } |
| 401 | |
| 402 | return ret; |
| 403 | } |
| 404 | |
| 405 | static void fio_netio_cleanup(struct thread_data *td) |
| 406 | { |
| 407 | struct netio_data *nd = td->io_ops->data; |
| 408 | |
| 409 | if (nd) { |
| 410 | if (nd->listenfd != -1) |
| 411 | close(nd->listenfd); |
| 412 | if (nd->pipes[0] != -1) |
| 413 | close(nd->pipes[0]); |
| 414 | if (nd->pipes[1] != -1) |
| 415 | close(nd->pipes[1]); |
| 416 | |
| 417 | free(nd); |
| 418 | td->io_ops->data = NULL; |
| 419 | } |
| 420 | } |
| 421 | |
| 422 | static int fio_netio_setup(struct thread_data *td) |
| 423 | { |
| 424 | struct netio_data *nd; |
| 425 | |
| 426 | if (!td->io_ops->data) { |
| 427 | nd = malloc(sizeof(*nd));; |
| 428 | |
| 429 | memset(nd, 0, sizeof(*nd)); |
| 430 | nd->listenfd = -1; |
| 431 | nd->pipes[0] = nd->pipes[1] = -1; |
| 432 | td->io_ops->data = nd; |
| 433 | } |
| 434 | |
| 435 | return 0; |
| 436 | } |
| 437 | |
| 438 | static int fio_netio_setup_splice(struct thread_data *td) |
| 439 | { |
| 440 | struct netio_data *nd; |
| 441 | |
| 442 | fio_netio_setup(td); |
| 443 | |
| 444 | nd = td->io_ops->data; |
| 445 | if (nd) { |
| 446 | if (pipe(nd->pipes) < 0) |
| 447 | return 1; |
| 448 | |
| 449 | nd->use_splice = 1; |
| 450 | return 0; |
| 451 | } |
| 452 | |
| 453 | return 1; |
| 454 | } |
| 455 | |
| 456 | static struct ioengine_ops ioengine_rw = { |
| 457 | .name = "net", |
| 458 | .version = FIO_IOOPS_VERSION, |
| 459 | .prep = fio_netio_prep, |
| 460 | .queue = fio_netio_queue, |
| 461 | .setup = fio_netio_setup, |
| 462 | .init = fio_netio_init, |
| 463 | .cleanup = fio_netio_cleanup, |
| 464 | .open_file = fio_netio_open_file, |
| 465 | .close_file = generic_close_file, |
| 466 | .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR, |
| 467 | }; |
| 468 | |
| 469 | static struct ioengine_ops ioengine_splice = { |
| 470 | .name = "netsplice", |
| 471 | .version = FIO_IOOPS_VERSION, |
| 472 | .prep = fio_netio_prep, |
| 473 | .queue = fio_netio_queue, |
| 474 | .setup = fio_netio_setup_splice, |
| 475 | .init = fio_netio_init, |
| 476 | .cleanup = fio_netio_cleanup, |
| 477 | .open_file = fio_netio_open_file, |
| 478 | .close_file = generic_close_file, |
| 479 | .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR, |
| 480 | }; |
| 481 | |
| 482 | static void fio_init fio_netio_register(void) |
| 483 | { |
| 484 | register_ioengine(&ioengine_rw); |
| 485 | register_ioengine(&ioengine_splice); |
| 486 | } |
| 487 | |
| 488 | static void fio_exit fio_netio_unregister(void) |
| 489 | { |
| 490 | unregister_ioengine(&ioengine_rw); |
| 491 | unregister_ioengine(&ioengine_splice); |
| 492 | } |