| 1 | /* |
| 2 | * mmap engine |
| 3 | * |
| 4 | * IO engine that reads/writes from files by doing memcpy to/from |
| 5 | * a memory mapped region of the file. |
| 6 | * |
| 7 | */ |
| 8 | #include <stdio.h> |
| 9 | #include <stdlib.h> |
| 10 | #include <errno.h> |
| 11 | #include <sys/mman.h> |
| 12 | |
| 13 | #include "../fio.h" |
| 14 | #include "../optgroup.h" |
| 15 | #include "../verify.h" |
| 16 | |
| 17 | /* |
| 18 | * Limits us to 1GiB of mapped files in total |
| 19 | */ |
| 20 | #define MMAP_TOTAL_SZ (1 * 1024 * 1024 * 1024UL) |
| 21 | |
| 22 | static unsigned long mmap_map_size; |
| 23 | |
| 24 | struct fio_mmap_data { |
| 25 | void *mmap_ptr; |
| 26 | size_t mmap_sz; |
| 27 | off_t mmap_off; |
| 28 | }; |
| 29 | |
| 30 | #ifdef CONFIG_HAVE_THP |
| 31 | struct mmap_options { |
| 32 | void *pad; |
| 33 | unsigned int thp; |
| 34 | }; |
| 35 | |
| 36 | static struct fio_option options[] = { |
| 37 | { |
| 38 | .name = "thp", |
| 39 | .lname = "Transparent Huge Pages", |
| 40 | .type = FIO_OPT_INT, |
| 41 | .off1 = offsetof(struct mmap_options, thp), |
| 42 | .help = "Memory Advise Huge Page", |
| 43 | .category = FIO_OPT_C_ENGINE, |
| 44 | .group = FIO_OPT_G_MMAP, |
| 45 | }, |
| 46 | { |
| 47 | .name = NULL, |
| 48 | }, |
| 49 | }; |
| 50 | #endif |
| 51 | |
| 52 | static bool fio_madvise_file(struct thread_data *td, struct fio_file *f, |
| 53 | size_t length) |
| 54 | |
| 55 | { |
| 56 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 57 | #ifdef CONFIG_HAVE_THP |
| 58 | struct mmap_options *o = td->eo; |
| 59 | |
| 60 | /* Ignore errors on this optional advisory */ |
| 61 | if (o->thp) |
| 62 | madvise(fmd->mmap_ptr, length, MADV_HUGEPAGE); |
| 63 | #endif |
| 64 | |
| 65 | if (!td->o.fadvise_hint) |
| 66 | return true; |
| 67 | |
| 68 | if (!td_random(td)) { |
| 69 | if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_SEQUENTIAL) < 0) { |
| 70 | td_verror(td, errno, "madvise"); |
| 71 | return false; |
| 72 | } |
| 73 | } else { |
| 74 | if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_RANDOM) < 0) { |
| 75 | td_verror(td, errno, "madvise"); |
| 76 | return false; |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | return true; |
| 81 | } |
| 82 | |
| 83 | #ifdef CONFIG_HAVE_THP |
| 84 | static int fio_mmap_get_shared(struct thread_data *td) |
| 85 | { |
| 86 | struct mmap_options *o = td->eo; |
| 87 | |
| 88 | if (o->thp) |
| 89 | return MAP_PRIVATE; |
| 90 | return MAP_SHARED; |
| 91 | } |
| 92 | #else |
| 93 | static int fio_mmap_get_shared(struct thread_data *td) |
| 94 | { |
| 95 | return MAP_SHARED; |
| 96 | } |
| 97 | #endif |
| 98 | |
| 99 | static int fio_mmap_file(struct thread_data *td, struct fio_file *f, |
| 100 | size_t length, off_t off) |
| 101 | { |
| 102 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 103 | int flags = 0, shared = fio_mmap_get_shared(td); |
| 104 | |
| 105 | if (td_rw(td) && !td->o.verify_only) |
| 106 | flags = PROT_READ | PROT_WRITE; |
| 107 | else if (td_write(td) && !td->o.verify_only) { |
| 108 | flags = PROT_WRITE; |
| 109 | |
| 110 | if (td->o.verify != VERIFY_NONE) |
| 111 | flags |= PROT_READ; |
| 112 | } else |
| 113 | flags = PROT_READ; |
| 114 | |
| 115 | fmd->mmap_ptr = mmap(NULL, length, flags, shared, f->fd, off); |
| 116 | if (fmd->mmap_ptr == MAP_FAILED) { |
| 117 | fmd->mmap_ptr = NULL; |
| 118 | td_verror(td, errno, "mmap"); |
| 119 | goto err; |
| 120 | } |
| 121 | |
| 122 | if (!fio_madvise_file(td, f, length)) |
| 123 | goto err; |
| 124 | |
| 125 | if (posix_madvise(fmd->mmap_ptr, length, POSIX_MADV_DONTNEED) < 0) { |
| 126 | td_verror(td, errno, "madvise"); |
| 127 | goto err; |
| 128 | } |
| 129 | |
| 130 | #ifdef FIO_MADV_FREE |
| 131 | if (f->filetype == FIO_TYPE_BLOCK) |
| 132 | (void) posix_madvise(fmd->mmap_ptr, fmd->mmap_sz, FIO_MADV_FREE); |
| 133 | #endif |
| 134 | |
| 135 | err: |
| 136 | if (td->error && fmd->mmap_ptr) |
| 137 | munmap(fmd->mmap_ptr, length); |
| 138 | |
| 139 | return td->error; |
| 140 | } |
| 141 | |
| 142 | /* |
| 143 | * Just mmap an appropriate portion, we cannot mmap the full extent |
| 144 | */ |
| 145 | static int fio_mmapio_prep_limited(struct thread_data *td, struct io_u *io_u) |
| 146 | { |
| 147 | struct fio_file *f = io_u->file; |
| 148 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 149 | |
| 150 | if (io_u->buflen > mmap_map_size) { |
| 151 | log_err("fio: bs too big for mmap engine\n"); |
| 152 | return EIO; |
| 153 | } |
| 154 | |
| 155 | fmd->mmap_sz = mmap_map_size; |
| 156 | if (fmd->mmap_sz > f->io_size) |
| 157 | fmd->mmap_sz = f->io_size; |
| 158 | |
| 159 | fmd->mmap_off = io_u->offset; |
| 160 | |
| 161 | return fio_mmap_file(td, f, fmd->mmap_sz, fmd->mmap_off); |
| 162 | } |
| 163 | |
| 164 | /* |
| 165 | * Attempt to mmap the entire file |
| 166 | */ |
| 167 | static int fio_mmapio_prep_full(struct thread_data *td, struct io_u *io_u) |
| 168 | { |
| 169 | struct fio_file *f = io_u->file; |
| 170 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 171 | int ret; |
| 172 | |
| 173 | if (fio_file_partial_mmap(f)) |
| 174 | return EINVAL; |
| 175 | if (io_u->offset != (size_t) io_u->offset || |
| 176 | f->io_size != (size_t) f->io_size) { |
| 177 | fio_file_set_partial_mmap(f); |
| 178 | return EINVAL; |
| 179 | } |
| 180 | |
| 181 | fmd->mmap_sz = f->io_size; |
| 182 | fmd->mmap_off = 0; |
| 183 | |
| 184 | ret = fio_mmap_file(td, f, fmd->mmap_sz, fmd->mmap_off); |
| 185 | if (ret) |
| 186 | fio_file_set_partial_mmap(f); |
| 187 | |
| 188 | return ret; |
| 189 | } |
| 190 | |
| 191 | static int fio_mmapio_prep(struct thread_data *td, struct io_u *io_u) |
| 192 | { |
| 193 | struct fio_file *f = io_u->file; |
| 194 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 195 | int ret; |
| 196 | |
| 197 | /* |
| 198 | * It fits within existing mapping, use it |
| 199 | */ |
| 200 | if (io_u->offset >= fmd->mmap_off && |
| 201 | io_u->offset + io_u->buflen <= fmd->mmap_off + fmd->mmap_sz) |
| 202 | goto done; |
| 203 | |
| 204 | /* |
| 205 | * unmap any existing mapping |
| 206 | */ |
| 207 | if (fmd->mmap_ptr) { |
| 208 | if (munmap(fmd->mmap_ptr, fmd->mmap_sz) < 0) |
| 209 | return errno; |
| 210 | fmd->mmap_ptr = NULL; |
| 211 | } |
| 212 | |
| 213 | if (fio_mmapio_prep_full(td, io_u)) { |
| 214 | td_clear_error(td); |
| 215 | ret = fio_mmapio_prep_limited(td, io_u); |
| 216 | if (ret) |
| 217 | return ret; |
| 218 | } |
| 219 | |
| 220 | done: |
| 221 | io_u->mmap_data = fmd->mmap_ptr + io_u->offset - fmd->mmap_off - |
| 222 | f->file_offset; |
| 223 | return 0; |
| 224 | } |
| 225 | |
| 226 | static enum fio_q_status fio_mmapio_queue(struct thread_data *td, |
| 227 | struct io_u *io_u) |
| 228 | { |
| 229 | struct fio_file *f = io_u->file; |
| 230 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 231 | |
| 232 | fio_ro_check(td, io_u); |
| 233 | |
| 234 | if (io_u->ddir == DDIR_READ) |
| 235 | memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen); |
| 236 | else if (io_u->ddir == DDIR_WRITE) |
| 237 | memcpy(io_u->mmap_data, io_u->xfer_buf, io_u->xfer_buflen); |
| 238 | else if (ddir_sync(io_u->ddir)) { |
| 239 | if (msync(fmd->mmap_ptr, fmd->mmap_sz, MS_SYNC)) { |
| 240 | io_u->error = errno; |
| 241 | td_verror(td, io_u->error, "msync"); |
| 242 | } |
| 243 | } else if (io_u->ddir == DDIR_TRIM) { |
| 244 | int ret = do_io_u_trim(td, io_u); |
| 245 | |
| 246 | if (!ret) |
| 247 | td_verror(td, io_u->error, "trim"); |
| 248 | } |
| 249 | |
| 250 | |
| 251 | /* |
| 252 | * not really direct, but should drop the pages from the cache |
| 253 | */ |
| 254 | if (td->o.odirect && ddir_rw(io_u->ddir)) { |
| 255 | if (msync(io_u->mmap_data, io_u->xfer_buflen, MS_SYNC) < 0) { |
| 256 | io_u->error = errno; |
| 257 | td_verror(td, io_u->error, "msync"); |
| 258 | } |
| 259 | if (posix_madvise(io_u->mmap_data, io_u->xfer_buflen, POSIX_MADV_DONTNEED) < 0) { |
| 260 | io_u->error = errno; |
| 261 | td_verror(td, io_u->error, "madvise"); |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | return FIO_Q_COMPLETED; |
| 266 | } |
| 267 | |
| 268 | static int fio_mmapio_init(struct thread_data *td) |
| 269 | { |
| 270 | struct thread_options *o = &td->o; |
| 271 | |
| 272 | if ((o->rw_min_bs & page_mask) && |
| 273 | (o->odirect || o->fsync_blocks || o->fdatasync_blocks)) { |
| 274 | log_err("fio: mmap options dictate a minimum block size of " |
| 275 | "%llu bytes\n", (unsigned long long) page_size); |
| 276 | return 1; |
| 277 | } |
| 278 | |
| 279 | mmap_map_size = MMAP_TOTAL_SZ / o->nr_files; |
| 280 | return 0; |
| 281 | } |
| 282 | |
| 283 | static int fio_mmapio_open_file(struct thread_data *td, struct fio_file *f) |
| 284 | { |
| 285 | struct fio_mmap_data *fmd; |
| 286 | int ret; |
| 287 | |
| 288 | ret = generic_open_file(td, f); |
| 289 | if (ret) |
| 290 | return ret; |
| 291 | |
| 292 | fmd = calloc(1, sizeof(*fmd)); |
| 293 | if (!fmd) { |
| 294 | int fio_unused __ret; |
| 295 | __ret = generic_close_file(td, f); |
| 296 | return 1; |
| 297 | } |
| 298 | |
| 299 | FILE_SET_ENG_DATA(f, fmd); |
| 300 | return 0; |
| 301 | } |
| 302 | |
| 303 | static int fio_mmapio_close_file(struct thread_data *td, struct fio_file *f) |
| 304 | { |
| 305 | struct fio_mmap_data *fmd = FILE_ENG_DATA(f); |
| 306 | |
| 307 | FILE_SET_ENG_DATA(f, NULL); |
| 308 | free(fmd); |
| 309 | fio_file_clear_partial_mmap(f); |
| 310 | |
| 311 | return generic_close_file(td, f); |
| 312 | } |
| 313 | |
| 314 | static struct ioengine_ops ioengine = { |
| 315 | .name = "mmap", |
| 316 | .version = FIO_IOOPS_VERSION, |
| 317 | .init = fio_mmapio_init, |
| 318 | .prep = fio_mmapio_prep, |
| 319 | .queue = fio_mmapio_queue, |
| 320 | .open_file = fio_mmapio_open_file, |
| 321 | .close_file = fio_mmapio_close_file, |
| 322 | .get_file_size = generic_get_file_size, |
| 323 | .flags = FIO_SYNCIO | FIO_NOEXTEND, |
| 324 | #ifdef CONFIG_HAVE_THP |
| 325 | .options = options, |
| 326 | .option_struct_size = sizeof(struct mmap_options), |
| 327 | #endif |
| 328 | }; |
| 329 | |
| 330 | static void fio_init fio_mmapio_register(void) |
| 331 | { |
| 332 | register_ioengine(&ioengine); |
| 333 | } |
| 334 | |
| 335 | static void fio_exit fio_mmapio_unregister(void) |
| 336 | { |
| 337 | unregister_ioengine(&ioengine); |
| 338 | } |