| 1 | /* |
| 2 | * pmemblk: IO engine that uses NVML libpmemblk to read and write data |
| 3 | * |
| 4 | * Copyright (C) 2016 Hewlett Packard Enterprise Development LP |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU General Public License, |
| 8 | * version 2 as published by the Free Software Foundation.. |
| 9 | * |
| 10 | * This program is distributed in the hope that it will be useful, |
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 13 | * GNU General Public License for more details. |
| 14 | * |
| 15 | * You should have received a copy of the GNU General Public |
| 16 | * License along with this program; if not, write to the Free |
| 17 | * Software Foundation, Inc., 59 Temple Place, Suite 330, |
| 18 | * Boston, MA 02111-1307 USA |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * pmemblk engine |
| 23 | * |
| 24 | * IO engine that uses libpmemblk to read and write data |
| 25 | * |
| 26 | * To use: |
| 27 | * ioengine=pmemblk |
| 28 | * |
| 29 | * Other relevant settings: |
| 30 | * thread=1 REQUIRED |
| 31 | * iodepth=1 |
| 32 | * direct=1 |
| 33 | * unlink=1 |
| 34 | * filename=/mnt/pmem0/fiotestfile,BSIZE,FSIZEMiB |
| 35 | * |
| 36 | * thread must be set to 1 for pmemblk as multiple processes cannot |
| 37 | * open the same block pool file. |
| 38 | * |
| 39 | * iodepth should be set to 1 as pmemblk is always synchronous. |
| 40 | * Use numjobs to scale up. |
| 41 | * |
| 42 | * direct=1 is implied as pmemblk is always direct. A warning message |
| 43 | * is printed if this is not specified. |
| 44 | * |
| 45 | * unlink=1 removes the block pool file after testing, and is optional. |
| 46 | * |
| 47 | * The pmem device must have a DAX-capable filesystem and be mounted |
| 48 | * with DAX enabled. filename must point to a file on that filesystem. |
| 49 | * |
| 50 | * Example: |
| 51 | * mkfs.xfs /dev/pmem0 |
| 52 | * mkdir /mnt/pmem0 |
| 53 | * mount -o dax /dev/pmem0 /mnt/pmem0 |
| 54 | * |
| 55 | * When specifying the filename, if the block pool file does not already |
| 56 | * exist, then the pmemblk engine creates the pool file if you specify |
| 57 | * the block and file sizes. BSIZE is the block size in bytes. |
| 58 | * FSIZEMB is the pool file size in MiB. |
| 59 | * |
| 60 | * See examples/pmemblk.fio for more. |
| 61 | * |
| 62 | */ |
| 63 | |
| 64 | #include <stdio.h> |
| 65 | #include <stdlib.h> |
| 66 | #include <unistd.h> |
| 67 | #include <sys/uio.h> |
| 68 | #include <errno.h> |
| 69 | #include <assert.h> |
| 70 | #include <string.h> |
| 71 | #include <libpmem.h> |
| 72 | #include <libpmemblk.h> |
| 73 | |
| 74 | #include "../fio.h" |
| 75 | |
| 76 | /* |
| 77 | * libpmemblk |
| 78 | */ |
| 79 | typedef struct fio_pmemblk_file *fio_pmemblk_file_t; |
| 80 | |
| 81 | struct fio_pmemblk_file { |
| 82 | fio_pmemblk_file_t pmb_next; |
| 83 | char *pmb_filename; |
| 84 | uint64_t pmb_refcnt; |
| 85 | PMEMblkpool *pmb_pool; |
| 86 | size_t pmb_bsize; |
| 87 | size_t pmb_nblocks; |
| 88 | }; |
| 89 | |
| 90 | static fio_pmemblk_file_t Cache; |
| 91 | |
| 92 | static pthread_mutex_t CacheLock = PTHREAD_MUTEX_INITIALIZER; |
| 93 | |
| 94 | #define PMB_CREATE (0x0001) /* should create file */ |
| 95 | |
| 96 | fio_pmemblk_file_t fio_pmemblk_cache_lookup(const char *filename) |
| 97 | { |
| 98 | fio_pmemblk_file_t i; |
| 99 | |
| 100 | for (i = Cache; i != NULL; i = i->pmb_next) |
| 101 | if (!strcmp(filename, i->pmb_filename)) |
| 102 | return i; |
| 103 | |
| 104 | return NULL; |
| 105 | } |
| 106 | |
| 107 | static void fio_pmemblk_cache_insert(fio_pmemblk_file_t pmb) |
| 108 | { |
| 109 | pmb->pmb_next = Cache; |
| 110 | Cache = pmb; |
| 111 | } |
| 112 | |
| 113 | static void fio_pmemblk_cache_remove(fio_pmemblk_file_t pmb) |
| 114 | { |
| 115 | fio_pmemblk_file_t i; |
| 116 | |
| 117 | if (pmb == Cache) { |
| 118 | Cache = Cache->pmb_next; |
| 119 | pmb->pmb_next = NULL; |
| 120 | return; |
| 121 | } |
| 122 | |
| 123 | for (i = Cache; i != NULL; i = i->pmb_next) |
| 124 | if (pmb == i->pmb_next) { |
| 125 | i->pmb_next = i->pmb_next->pmb_next; |
| 126 | pmb->pmb_next = NULL; |
| 127 | return; |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | /* |
| 132 | * to control block size and gross file size at the libpmemblk |
| 133 | * level, we allow the block size and file size to be appended |
| 134 | * to the file name: |
| 135 | * |
| 136 | * path[,bsize,fsizemib] |
| 137 | * |
| 138 | * note that we do not use the fio option "filesize" to dictate |
| 139 | * the file size because we can only give libpmemblk the gross |
| 140 | * file size, which is different from the net or usable file |
| 141 | * size (which is probably what fio wants). |
| 142 | * |
| 143 | * the final path without the parameters is returned in ppath. |
| 144 | * the block size and file size are returned in pbsize and fsize. |
| 145 | * |
| 146 | * note that the user specifies the file size in MiB, but |
| 147 | * we return bytes from here. |
| 148 | */ |
| 149 | static void pmb_parse_path(const char *pathspec, char **ppath, uint64_t *pbsize, |
| 150 | uint64_t *pfsize) |
| 151 | { |
| 152 | char *path; |
| 153 | char *s; |
| 154 | uint64_t bsize; |
| 155 | uint64_t fsizemib; |
| 156 | |
| 157 | path = strdup(pathspec); |
| 158 | if (!path) { |
| 159 | *ppath = NULL; |
| 160 | return; |
| 161 | } |
| 162 | |
| 163 | /* extract sizes, if given */ |
| 164 | s = strrchr(path, ','); |
| 165 | if (s && (fsizemib = strtoull(s + 1, NULL, 10))) { |
| 166 | *s = 0; |
| 167 | s = strrchr(path, ','); |
| 168 | if (s && (bsize = strtoull(s + 1, NULL, 10))) { |
| 169 | *s = 0; |
| 170 | *ppath = path; |
| 171 | *pbsize = bsize; |
| 172 | *pfsize = fsizemib << 20; |
| 173 | return; |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | /* size specs not found */ |
| 178 | strcpy(path, pathspec); |
| 179 | *ppath = path; |
| 180 | *pbsize = 0; |
| 181 | *pfsize = 0; |
| 182 | } |
| 183 | |
| 184 | static fio_pmemblk_file_t pmb_open(const char *pathspec, int flags) |
| 185 | { |
| 186 | fio_pmemblk_file_t pmb; |
| 187 | char *path = NULL; |
| 188 | uint64_t bsize = 0; |
| 189 | uint64_t fsize = 0; |
| 190 | |
| 191 | pmb_parse_path(pathspec, &path, &bsize, &fsize); |
| 192 | if (!path) |
| 193 | return NULL; |
| 194 | |
| 195 | pthread_mutex_lock(&CacheLock); |
| 196 | |
| 197 | pmb = fio_pmemblk_cache_lookup(path); |
| 198 | if (!pmb) { |
| 199 | pmb = malloc(sizeof(*pmb)); |
| 200 | if (!pmb) |
| 201 | goto error; |
| 202 | |
| 203 | /* try opening existing first, create it if needed */ |
| 204 | pmb->pmb_pool = pmemblk_open(path, bsize); |
| 205 | if (!pmb->pmb_pool && (errno == ENOENT) && |
| 206 | (flags & PMB_CREATE) && (0 < fsize) && (0 < bsize)) { |
| 207 | pmb->pmb_pool = |
| 208 | pmemblk_create(path, bsize, fsize, 0644); |
| 209 | } |
| 210 | if (!pmb->pmb_pool) { |
| 211 | log_err("pmemblk: unable to open pmemblk pool file %s (%s)\n", |
| 212 | path, strerror(errno)); |
| 213 | goto error; |
| 214 | } |
| 215 | |
| 216 | pmb->pmb_filename = path; |
| 217 | pmb->pmb_next = NULL; |
| 218 | pmb->pmb_refcnt = 0; |
| 219 | pmb->pmb_bsize = pmemblk_bsize(pmb->pmb_pool); |
| 220 | pmb->pmb_nblocks = pmemblk_nblock(pmb->pmb_pool); |
| 221 | |
| 222 | fio_pmemblk_cache_insert(pmb); |
| 223 | } |
| 224 | |
| 225 | pmb->pmb_refcnt += 1; |
| 226 | |
| 227 | pthread_mutex_unlock(&CacheLock); |
| 228 | |
| 229 | return pmb; |
| 230 | |
| 231 | error: |
| 232 | if (pmb) { |
| 233 | if (pmb->pmb_pool) |
| 234 | pmemblk_close(pmb->pmb_pool); |
| 235 | pmb->pmb_pool = NULL; |
| 236 | pmb->pmb_filename = NULL; |
| 237 | free(pmb); |
| 238 | } |
| 239 | if (path) |
| 240 | free(path); |
| 241 | |
| 242 | pthread_mutex_unlock(&CacheLock); |
| 243 | return NULL; |
| 244 | } |
| 245 | |
| 246 | static void pmb_close(fio_pmemblk_file_t pmb, const bool keep) |
| 247 | { |
| 248 | pthread_mutex_lock(&CacheLock); |
| 249 | |
| 250 | pmb->pmb_refcnt--; |
| 251 | |
| 252 | if (!keep && !pmb->pmb_refcnt) { |
| 253 | pmemblk_close(pmb->pmb_pool); |
| 254 | pmb->pmb_pool = NULL; |
| 255 | free(pmb->pmb_filename); |
| 256 | pmb->pmb_filename = NULL; |
| 257 | fio_pmemblk_cache_remove(pmb); |
| 258 | free(pmb); |
| 259 | } |
| 260 | |
| 261 | pthread_mutex_unlock(&CacheLock); |
| 262 | } |
| 263 | |
| 264 | static int pmb_get_flags(struct thread_data *td, uint64_t *pflags) |
| 265 | { |
| 266 | static int thread_warned = 0; |
| 267 | static int odirect_warned = 0; |
| 268 | |
| 269 | uint64_t flags = 0; |
| 270 | |
| 271 | if (!td->o.use_thread) { |
| 272 | if (!thread_warned) { |
| 273 | thread_warned = 1; |
| 274 | log_err("pmemblk: must set thread=1 for pmemblk engine\n"); |
| 275 | } |
| 276 | return 1; |
| 277 | } |
| 278 | |
| 279 | if (!td->o.odirect && !odirect_warned) { |
| 280 | odirect_warned = 1; |
| 281 | log_info("pmemblk: direct == 0, but pmemblk is always direct\n"); |
| 282 | } |
| 283 | |
| 284 | if (td->o.allow_create) |
| 285 | flags |= PMB_CREATE; |
| 286 | |
| 287 | (*pflags) = flags; |
| 288 | return 0; |
| 289 | } |
| 290 | |
| 291 | static int fio_pmemblk_open_file(struct thread_data *td, struct fio_file *f) |
| 292 | { |
| 293 | uint64_t flags = 0; |
| 294 | fio_pmemblk_file_t pmb; |
| 295 | |
| 296 | if (pmb_get_flags(td, &flags)) |
| 297 | return 1; |
| 298 | |
| 299 | pmb = pmb_open(f->file_name, flags); |
| 300 | if (!pmb) |
| 301 | return 1; |
| 302 | |
| 303 | FILE_SET_ENG_DATA(f, pmb); |
| 304 | return 0; |
| 305 | } |
| 306 | |
| 307 | static int fio_pmemblk_close_file(struct thread_data fio_unused *td, |
| 308 | struct fio_file *f) |
| 309 | { |
| 310 | fio_pmemblk_file_t pmb = FILE_ENG_DATA(f); |
| 311 | |
| 312 | if (pmb) |
| 313 | pmb_close(pmb, false); |
| 314 | |
| 315 | FILE_SET_ENG_DATA(f, NULL); |
| 316 | return 0; |
| 317 | } |
| 318 | |
| 319 | static int fio_pmemblk_get_file_size(struct thread_data *td, struct fio_file *f) |
| 320 | { |
| 321 | uint64_t flags = 0; |
| 322 | fio_pmemblk_file_t pmb = FILE_ENG_DATA(f); |
| 323 | |
| 324 | if (fio_file_size_known(f)) |
| 325 | return 0; |
| 326 | |
| 327 | if (!pmb) { |
| 328 | if (pmb_get_flags(td, &flags)) |
| 329 | return 1; |
| 330 | pmb = pmb_open(f->file_name, flags); |
| 331 | if (!pmb) |
| 332 | return 1; |
| 333 | } |
| 334 | |
| 335 | f->real_file_size = pmb->pmb_bsize * pmb->pmb_nblocks; |
| 336 | |
| 337 | fio_file_set_size_known(f); |
| 338 | |
| 339 | if (!FILE_ENG_DATA(f)) |
| 340 | pmb_close(pmb, true); |
| 341 | |
| 342 | return 0; |
| 343 | } |
| 344 | |
| 345 | static int fio_pmemblk_queue(struct thread_data *td, struct io_u *io_u) |
| 346 | { |
| 347 | struct fio_file *f = io_u->file; |
| 348 | fio_pmemblk_file_t pmb = FILE_ENG_DATA(f); |
| 349 | |
| 350 | unsigned long long off; |
| 351 | unsigned long len; |
| 352 | void *buf; |
| 353 | |
| 354 | fio_ro_check(td, io_u); |
| 355 | |
| 356 | switch (io_u->ddir) { |
| 357 | case DDIR_READ: |
| 358 | case DDIR_WRITE: |
| 359 | off = io_u->offset; |
| 360 | len = io_u->xfer_buflen; |
| 361 | |
| 362 | io_u->error = EINVAL; |
| 363 | if (off % pmb->pmb_bsize) |
| 364 | break; |
| 365 | if (len % pmb->pmb_bsize) |
| 366 | break; |
| 367 | if ((off + len) / pmb->pmb_bsize > pmb->pmb_nblocks) |
| 368 | break; |
| 369 | |
| 370 | io_u->error = 0; |
| 371 | buf = io_u->xfer_buf; |
| 372 | off /= pmb->pmb_bsize; |
| 373 | len /= pmb->pmb_bsize; |
| 374 | while (0 < len) { |
| 375 | if (io_u->ddir == DDIR_READ && |
| 376 | 0 != pmemblk_read(pmb->pmb_pool, buf, off)) { |
| 377 | io_u->error = errno; |
| 378 | break; |
| 379 | } else if (0 != pmemblk_write(pmb->pmb_pool, buf, off)) { |
| 380 | io_u->error = errno; |
| 381 | break; |
| 382 | } |
| 383 | buf += pmb->pmb_bsize; |
| 384 | off++; |
| 385 | len--; |
| 386 | } |
| 387 | off *= pmb->pmb_bsize; |
| 388 | len *= pmb->pmb_bsize; |
| 389 | io_u->resid = io_u->xfer_buflen - (off - io_u->offset); |
| 390 | break; |
| 391 | case DDIR_SYNC: |
| 392 | case DDIR_DATASYNC: |
| 393 | case DDIR_SYNC_FILE_RANGE: |
| 394 | /* we're always sync'd */ |
| 395 | io_u->error = 0; |
| 396 | break; |
| 397 | default: |
| 398 | io_u->error = EINVAL; |
| 399 | break; |
| 400 | } |
| 401 | |
| 402 | return FIO_Q_COMPLETED; |
| 403 | } |
| 404 | |
| 405 | static int fio_pmemblk_unlink_file(struct thread_data *td, struct fio_file *f) |
| 406 | { |
| 407 | char *path = NULL; |
| 408 | uint64_t bsize = 0; |
| 409 | uint64_t fsize = 0; |
| 410 | |
| 411 | /* |
| 412 | * we need our own unlink in case the user has specified |
| 413 | * the block and file sizes in the path name. we parse |
| 414 | * the file_name to determine the file name we actually used. |
| 415 | */ |
| 416 | |
| 417 | pmb_parse_path(f->file_name, &path, &bsize, &fsize); |
| 418 | if (!path) |
| 419 | return ENOENT; |
| 420 | |
| 421 | unlink(path); |
| 422 | free(path); |
| 423 | return 0; |
| 424 | } |
| 425 | |
| 426 | static struct ioengine_ops ioengine = { |
| 427 | .name = "pmemblk", |
| 428 | .version = FIO_IOOPS_VERSION, |
| 429 | .queue = fio_pmemblk_queue, |
| 430 | .open_file = fio_pmemblk_open_file, |
| 431 | .close_file = fio_pmemblk_close_file, |
| 432 | .get_file_size = fio_pmemblk_get_file_size, |
| 433 | .unlink_file = fio_pmemblk_unlink_file, |
| 434 | .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_NOEXTEND | FIO_NODISKUTIL, |
| 435 | }; |
| 436 | |
| 437 | static void fio_init fio_pmemblk_register(void) |
| 438 | { |
| 439 | register_ioengine(&ioengine); |
| 440 | } |
| 441 | |
| 442 | static void fio_exit fio_pmemblk_unregister(void) |
| 443 | { |
| 444 | unregister_ioengine(&ioengine); |
| 445 | } |