| 1 | /* |
| 2 | * NBD engine |
| 3 | * |
| 4 | * IO engine that talks to an NBD server. |
| 5 | * |
| 6 | * Copyright (C) 2019 Red Hat Inc. |
| 7 | * Written by Richard W.M. Jones <rjones@redhat.com> |
| 8 | * |
| 9 | */ |
| 10 | |
| 11 | #include <stdio.h> |
| 12 | #include <stdlib.h> |
| 13 | #include <stdint.h> |
| 14 | #include <errno.h> |
| 15 | |
| 16 | #include <libnbd.h> |
| 17 | |
| 18 | #include "../fio.h" |
| 19 | #include "../optgroup.h" |
| 20 | |
| 21 | /* Actually this differs across servers, but for nbdkit ... */ |
| 22 | #define NBD_MAX_REQUEST_SIZE (64 * 1024 * 1024) |
| 23 | |
| 24 | /* Storage for the NBD handle. */ |
| 25 | struct nbd_data { |
| 26 | struct nbd_handle *nbd; |
| 27 | int debug; |
| 28 | |
| 29 | /* The list of completed io_u structs. */ |
| 30 | struct io_u **completed; |
| 31 | size_t nr_completed; |
| 32 | }; |
| 33 | |
| 34 | /* Options. */ |
| 35 | struct nbd_options { |
| 36 | void *padding; |
| 37 | char *uri; |
| 38 | }; |
| 39 | |
| 40 | static struct fio_option options[] = { |
| 41 | { |
| 42 | .name = "uri", |
| 43 | .lname = "NBD URI", |
| 44 | .help = "Name of NBD URI", |
| 45 | .category = FIO_OPT_C_ENGINE, |
| 46 | .group = FIO_OPT_G_NBD, |
| 47 | .type = FIO_OPT_STR_STORE, |
| 48 | .off1 = offsetof(struct nbd_options, uri), |
| 49 | }, |
| 50 | { |
| 51 | .name = NULL, |
| 52 | }, |
| 53 | }; |
| 54 | |
| 55 | /* Alocates nbd_data. */ |
| 56 | static int nbd_setup(struct thread_data *td) |
| 57 | { |
| 58 | struct nbd_data *nbd_data; |
| 59 | struct nbd_options *o = td->eo; |
| 60 | struct fio_file *f; |
| 61 | int r; |
| 62 | int64_t size; |
| 63 | |
| 64 | nbd_data = calloc(1, sizeof(*nbd_data)); |
| 65 | if (!nbd_data) { |
| 66 | td_verror(td, errno, "calloc"); |
| 67 | return 1; |
| 68 | } |
| 69 | td->io_ops_data = nbd_data; |
| 70 | |
| 71 | /* Pretend to deal with files. See engines/rbd.c */ |
| 72 | if (!td->files_index) { |
| 73 | add_file(td, "nbd", 0, 0); |
| 74 | td->o.nr_files = td->o.nr_files ? : 1; |
| 75 | td->o.open_files++; |
| 76 | } |
| 77 | f = td->files[0]; |
| 78 | |
| 79 | nbd_data->nbd = nbd_create(); |
| 80 | if (!nbd_data->nbd) { |
| 81 | log_err("fio: nbd_create: %s\n", nbd_get_error()); |
| 82 | return 1; |
| 83 | } |
| 84 | |
| 85 | /* Get the debug flag which can be set through LIBNBD_DEBUG=1. */ |
| 86 | nbd_data->debug = nbd_get_debug(nbd_data->nbd); |
| 87 | |
| 88 | /* Connect synchronously here so we can check for the size and |
| 89 | * in future other properties of the server. |
| 90 | */ |
| 91 | if (!o->uri) { |
| 92 | log_err("fio: nbd: uri parameter was not specified\n"); |
| 93 | return 1; |
| 94 | } |
| 95 | r = nbd_connect_uri(nbd_data->nbd, o->uri); |
| 96 | if (r == -1) { |
| 97 | log_err("fio: nbd_connect_uri: %s\n", nbd_get_error()); |
| 98 | return 1; |
| 99 | } |
| 100 | size = nbd_get_size(nbd_data->nbd); |
| 101 | if (size == -1) { |
| 102 | log_err("fio: nbd_get_size: %s\n", nbd_get_error()); |
| 103 | return 1; |
| 104 | } |
| 105 | |
| 106 | f->real_file_size = size; |
| 107 | |
| 108 | nbd_close (nbd_data->nbd); |
| 109 | nbd_data->nbd = NULL; |
| 110 | |
| 111 | return 0; |
| 112 | } |
| 113 | |
| 114 | /* Closes socket and frees nbd_data -- the opposite of nbd_setup. */ |
| 115 | static void nbd_cleanup(struct thread_data *td) |
| 116 | { |
| 117 | struct nbd_data *nbd_data = td->io_ops_data; |
| 118 | |
| 119 | if (nbd_data) { |
| 120 | if (nbd_data->nbd) |
| 121 | nbd_close(nbd_data->nbd); |
| 122 | free(nbd_data); |
| 123 | } |
| 124 | } |
| 125 | |
| 126 | /* Connect to the server from each thread. */ |
| 127 | static int nbd_init(struct thread_data *td) |
| 128 | { |
| 129 | struct nbd_options *o = td->eo; |
| 130 | struct nbd_data *nbd_data = td->io_ops_data; |
| 131 | int r; |
| 132 | |
| 133 | if (!o->uri) { |
| 134 | log_err("fio: nbd: uri parameter was not specified\n"); |
| 135 | return 1; |
| 136 | } |
| 137 | |
| 138 | nbd_data->nbd = nbd_create(); |
| 139 | if (!nbd_data->nbd) { |
| 140 | log_err("fio: nbd_create: %s\n", nbd_get_error()); |
| 141 | return 1; |
| 142 | } |
| 143 | /* This is actually a synchronous connect and handshake. */ |
| 144 | r = nbd_connect_uri(nbd_data->nbd, o->uri); |
| 145 | if (r == -1) { |
| 146 | log_err("fio: nbd_connect_uri: %s\n", nbd_get_error()); |
| 147 | return 1; |
| 148 | } |
| 149 | |
| 150 | log_info("fio: connected to NBD server\n"); |
| 151 | return 0; |
| 152 | } |
| 153 | |
| 154 | /* A command in flight has been completed. */ |
| 155 | static int cmd_completed (void *vp, int *error) |
| 156 | { |
| 157 | struct io_u *io_u; |
| 158 | struct nbd_data *nbd_data; |
| 159 | struct io_u **completed; |
| 160 | |
| 161 | io_u = vp; |
| 162 | nbd_data = io_u->engine_data; |
| 163 | |
| 164 | if (nbd_data->debug) |
| 165 | log_info("fio: nbd: command completed\n"); |
| 166 | |
| 167 | if (*error != 0) |
| 168 | io_u->error = *error; |
| 169 | else |
| 170 | io_u->error = 0; |
| 171 | |
| 172 | /* Add this completion to the list so it can be picked up |
| 173 | * later by ->event. |
| 174 | */ |
| 175 | completed = realloc(nbd_data->completed, |
| 176 | sizeof(struct io_u *) * |
| 177 | (nbd_data->nr_completed+1)); |
| 178 | if (completed == NULL) { |
| 179 | io_u->error = errno; |
| 180 | return 0; |
| 181 | } |
| 182 | |
| 183 | nbd_data->completed = completed; |
| 184 | nbd_data->completed[nbd_data->nr_completed] = io_u; |
| 185 | nbd_data->nr_completed++; |
| 186 | |
| 187 | return 0; |
| 188 | } |
| 189 | |
| 190 | /* Begin read or write request. */ |
| 191 | static enum fio_q_status nbd_queue(struct thread_data *td, |
| 192 | struct io_u *io_u) |
| 193 | { |
| 194 | struct nbd_data *nbd_data = td->io_ops_data; |
| 195 | nbd_completion_callback completion = { .callback = cmd_completed, |
| 196 | .user_data = io_u }; |
| 197 | int r; |
| 198 | |
| 199 | fio_ro_check(td, io_u); |
| 200 | |
| 201 | io_u->engine_data = nbd_data; |
| 202 | |
| 203 | if (io_u->ddir == DDIR_WRITE || io_u->ddir == DDIR_READ) |
| 204 | assert(io_u->xfer_buflen <= NBD_MAX_REQUEST_SIZE); |
| 205 | |
| 206 | switch (io_u->ddir) { |
| 207 | case DDIR_READ: |
| 208 | r = nbd_aio_pread(nbd_data->nbd, |
| 209 | io_u->xfer_buf, io_u->xfer_buflen, |
| 210 | io_u->offset, completion, 0); |
| 211 | break; |
| 212 | case DDIR_WRITE: |
| 213 | r = nbd_aio_pwrite(nbd_data->nbd, |
| 214 | io_u->xfer_buf, io_u->xfer_buflen, |
| 215 | io_u->offset, completion, 0); |
| 216 | break; |
| 217 | case DDIR_TRIM: |
| 218 | r = nbd_aio_trim(nbd_data->nbd, io_u->xfer_buflen, |
| 219 | io_u->offset, completion, 0); |
| 220 | break; |
| 221 | case DDIR_SYNC: |
| 222 | /* XXX We could probably also handle |
| 223 | * DDIR_SYNC_FILE_RANGE with a bit of effort. |
| 224 | */ |
| 225 | r = nbd_aio_flush(nbd_data->nbd, completion, 0); |
| 226 | break; |
| 227 | default: |
| 228 | io_u->error = EINVAL; |
| 229 | return FIO_Q_COMPLETED; |
| 230 | } |
| 231 | |
| 232 | if (r == -1) { |
| 233 | /* errno is optional information on libnbd error path; |
| 234 | * if it's 0, set it to a default value |
| 235 | */ |
| 236 | io_u->error = nbd_get_errno(); |
| 237 | if (io_u->error == 0) |
| 238 | io_u->error = EIO; |
| 239 | return FIO_Q_COMPLETED; |
| 240 | } |
| 241 | |
| 242 | if (nbd_data->debug) |
| 243 | log_info("fio: nbd: command issued\n"); |
| 244 | io_u->error = 0; |
| 245 | return FIO_Q_QUEUED; |
| 246 | } |
| 247 | |
| 248 | static unsigned retire_commands(struct nbd_handle *nbd) |
| 249 | { |
| 250 | int64_t cookie; |
| 251 | unsigned r = 0; |
| 252 | |
| 253 | while ((cookie = nbd_aio_peek_command_completed(nbd)) > 0) { |
| 254 | /* Ignore the return value. cmd_completed has already |
| 255 | * checked for an error and set io_u->error. We only |
| 256 | * have to call this to retire the command. |
| 257 | */ |
| 258 | nbd_aio_command_completed(nbd, cookie); |
| 259 | r++; |
| 260 | } |
| 261 | |
| 262 | if (nbd_get_debug(nbd)) |
| 263 | log_info("fio: nbd: %u commands retired\n", r); |
| 264 | return r; |
| 265 | } |
| 266 | |
| 267 | static int nbd_getevents(struct thread_data *td, unsigned int min, |
| 268 | unsigned int max, const struct timespec *t) |
| 269 | { |
| 270 | struct nbd_data *nbd_data = td->io_ops_data; |
| 271 | int r; |
| 272 | unsigned events = 0; |
| 273 | int timeout; |
| 274 | |
| 275 | /* XXX This handling of timeout is wrong because it will wait |
| 276 | * for up to loop iterations * timeout. |
| 277 | */ |
| 278 | timeout = !t ? -1 : t->tv_sec * 1000 + t->tv_nsec / 1000000; |
| 279 | |
| 280 | while (events < min) { |
| 281 | r = nbd_poll(nbd_data->nbd, timeout); |
| 282 | if (r == -1) { |
| 283 | /* error in poll */ |
| 284 | log_err("fio: nbd_poll: %s\n", nbd_get_error()); |
| 285 | return -1; |
| 286 | } |
| 287 | else { |
| 288 | /* poll made progress */ |
| 289 | events += retire_commands(nbd_data->nbd); |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | return events; |
| 294 | } |
| 295 | |
| 296 | static struct io_u *nbd_event(struct thread_data *td, int event) |
| 297 | { |
| 298 | struct nbd_data *nbd_data = td->io_ops_data; |
| 299 | |
| 300 | if (nbd_data->nr_completed == 0) |
| 301 | return NULL; |
| 302 | |
| 303 | /* XXX We ignore the event number and assume fio calls us |
| 304 | * exactly once for [0..nr_events-1]. |
| 305 | */ |
| 306 | nbd_data->nr_completed--; |
| 307 | return nbd_data->completed[nbd_data->nr_completed]; |
| 308 | } |
| 309 | |
| 310 | static int nbd_io_u_init(struct thread_data *td, struct io_u *io_u) |
| 311 | { |
| 312 | io_u->engine_data = NULL; |
| 313 | return 0; |
| 314 | } |
| 315 | |
| 316 | static void nbd_io_u_free(struct thread_data *td, struct io_u *io_u) |
| 317 | { |
| 318 | /* Nothing needs to be done. */ |
| 319 | } |
| 320 | |
| 321 | static int nbd_open_file(struct thread_data *td, struct fio_file *f) |
| 322 | { |
| 323 | return 0; |
| 324 | } |
| 325 | |
| 326 | static int nbd_invalidate(struct thread_data *td, struct fio_file *f) |
| 327 | { |
| 328 | return 0; |
| 329 | } |
| 330 | |
| 331 | static struct ioengine_ops ioengine = { |
| 332 | .name = "nbd", |
| 333 | .version = FIO_IOOPS_VERSION, |
| 334 | .options = options, |
| 335 | .option_struct_size = sizeof(struct nbd_options), |
| 336 | .flags = FIO_DISKLESSIO | FIO_NOEXTEND, |
| 337 | |
| 338 | .setup = nbd_setup, |
| 339 | .init = nbd_init, |
| 340 | .cleanup = nbd_cleanup, |
| 341 | .queue = nbd_queue, |
| 342 | .getevents = nbd_getevents, |
| 343 | .event = nbd_event, |
| 344 | .io_u_init = nbd_io_u_init, |
| 345 | .io_u_free = nbd_io_u_free, |
| 346 | |
| 347 | .open_file = nbd_open_file, |
| 348 | .invalidate = nbd_invalidate, |
| 349 | }; |
| 350 | |
| 351 | static void fio_init fio_nbd_register(void) |
| 352 | { |
| 353 | register_ioengine(&ioengine); |
| 354 | } |
| 355 | |
| 356 | static void fio_exit fio_nbd_unregister(void) |
| 357 | { |
| 358 | unregister_ioengine(&ioengine); |
| 359 | } |