1 /* 2 * Copyright (C) 2010-2012 by Dell Inc. All rights reserved. 3 * Copyright (C) 2011-2013 Red Hat, Inc. 4 * 5 * This file is released under the GPL. 6 * 7 * dm-switch is a device-mapper target that maps IO to underlying block 8 * devices efficiently when there are a large number of fixed-sized 9 * address regions but there is no simple pattern to allow for a compact 10 * mapping representation such as dm-stripe. 11 */ 12 13 #include <linux/device-mapper.h> 14 15 #include <linux/module.h> 16 #include <linux/init.h> 17 #include <linux/vmalloc.h> 18 19 #define DM_MSG_PREFIX "switch" 20 21 /* 22 * One region_table_slot_t holds <region_entries_per_slot> region table 23 * entries each of which is <region_table_entry_bits> in size. 24 */ 25 typedef unsigned long region_table_slot_t; 26 27 /* 28 * A device with the offset to its start sector. 29 */ 30 struct switch_path { 31 struct dm_dev *dmdev; 32 sector_t start; 33 }; 34 35 /* 36 * Context block for a dm switch device. 37 */ 38 struct switch_ctx { 39 struct dm_target *ti; 40 41 unsigned nr_paths; /* Number of paths in path_list. */ 42 43 unsigned region_size; /* Region size in 512-byte sectors */ 44 unsigned long nr_regions; /* Number of regions making up the device */ 45 signed char region_size_bits; /* log2 of region_size or -1 */ 46 47 unsigned char region_table_entry_bits; /* Number of bits in one region table entry */ 48 unsigned char region_entries_per_slot; /* Number of entries in one region table slot */ 49 signed char region_entries_per_slot_bits; /* log2 of region_entries_per_slot or -1 */ 50 51 region_table_slot_t *region_table; /* Region table */ 52 53 /* 54 * Array of dm devices to switch between. 55 */ 56 struct switch_path path_list[0]; 57 }; 58 59 static struct switch_ctx *alloc_switch_ctx(struct dm_target *ti, unsigned nr_paths, 60 unsigned region_size) 61 { 62 struct switch_ctx *sctx; 63 64 sctx = kzalloc(sizeof(struct switch_ctx) + nr_paths * sizeof(struct switch_path), 65 GFP_KERNEL); 66 if (!sctx) 67 return NULL; 68 69 sctx->ti = ti; 70 sctx->region_size = region_size; 71 72 ti->private = sctx; 73 74 return sctx; 75 } 76 77 static int alloc_region_table(struct dm_target *ti, unsigned nr_paths) 78 { 79 struct switch_ctx *sctx = ti->private; 80 sector_t nr_regions = ti->len; 81 sector_t nr_slots; 82 83 if (!(sctx->region_size & (sctx->region_size - 1))) 84 sctx->region_size_bits = __ffs(sctx->region_size); 85 else 86 sctx->region_size_bits = -1; 87 88 sctx->region_table_entry_bits = 1; 89 while (sctx->region_table_entry_bits < sizeof(region_table_slot_t) * 8 && 90 (region_table_slot_t)1 << sctx->region_table_entry_bits < nr_paths) 91 sctx->region_table_entry_bits++; 92 93 sctx->region_entries_per_slot = (sizeof(region_table_slot_t) * 8) / sctx->region_table_entry_bits; 94 if (!(sctx->region_entries_per_slot & (sctx->region_entries_per_slot - 1))) 95 sctx->region_entries_per_slot_bits = __ffs(sctx->region_entries_per_slot); 96 else 97 sctx->region_entries_per_slot_bits = -1; 98 99 if (sector_div(nr_regions, sctx->region_size)) 100 nr_regions++; 101 102 sctx->nr_regions = nr_regions; 103 if (sctx->nr_regions != nr_regions || sctx->nr_regions >= ULONG_MAX) { 104 ti->error = "Region table too large"; 105 return -EINVAL; 106 } 107 108 nr_slots = nr_regions; 109 if (sector_div(nr_slots, sctx->region_entries_per_slot)) 110 nr_slots++; 111 112 if (nr_slots > ULONG_MAX / sizeof(region_table_slot_t)) { 113 ti->error = "Region table too large"; 114 return -EINVAL; 115 } 116 117 sctx->region_table = vmalloc(nr_slots * sizeof(region_table_slot_t)); 118 if (!sctx->region_table) { 119 ti->error = "Cannot allocate region table"; 120 return -ENOMEM; 121 } 122 123 return 0; 124 } 125 126 static void switch_get_position(struct switch_ctx *sctx, unsigned long region_nr, 127 unsigned long *region_index, unsigned *bit) 128 { 129 if (sctx->region_entries_per_slot_bits >= 0) { 130 *region_index = region_nr >> sctx->region_entries_per_slot_bits; 131 *bit = region_nr & (sctx->region_entries_per_slot - 1); 132 } else { 133 *region_index = region_nr / sctx->region_entries_per_slot; 134 *bit = region_nr % sctx->region_entries_per_slot; 135 } 136 137 *bit *= sctx->region_table_entry_bits; 138 } 139 140 /* 141 * Find which path to use at given offset. 142 */ 143 static unsigned switch_get_path_nr(struct switch_ctx *sctx, sector_t offset) 144 { 145 unsigned long region_index; 146 unsigned bit, path_nr; 147 sector_t p; 148 149 p = offset; 150 if (sctx->region_size_bits >= 0) 151 p >>= sctx->region_size_bits; 152 else 153 sector_div(p, sctx->region_size); 154 155 switch_get_position(sctx, p, ®ion_index, &bit); 156 path_nr = (ACCESS_ONCE(sctx->region_table[region_index]) >> bit) & 157 ((1 << sctx->region_table_entry_bits) - 1); 158 159 /* This can only happen if the processor uses non-atomic stores. */ 160 if (unlikely(path_nr >= sctx->nr_paths)) 161 path_nr = 0; 162 163 return path_nr; 164 } 165 166 static void switch_region_table_write(struct switch_ctx *sctx, unsigned long region_nr, 167 unsigned value) 168 { 169 unsigned long region_index; 170 unsigned bit; 171 region_table_slot_t pte; 172 173 switch_get_position(sctx, region_nr, ®ion_index, &bit); 174 175 pte = sctx->region_table[region_index]; 176 pte &= ~((((region_table_slot_t)1 << sctx->region_table_entry_bits) - 1) << bit); 177 pte |= (region_table_slot_t)value << bit; 178 sctx->region_table[region_index] = pte; 179 } 180 181 /* 182 * Fill the region table with an initial round robin pattern. 183 */ 184 static void initialise_region_table(struct switch_ctx *sctx) 185 { 186 unsigned path_nr = 0; 187 unsigned long region_nr; 188 189 for (region_nr = 0; region_nr < sctx->nr_regions; region_nr++) { 190 switch_region_table_write(sctx, region_nr, path_nr); 191 if (++path_nr >= sctx->nr_paths) 192 path_nr = 0; 193 } 194 } 195 196 static int parse_path(struct dm_arg_set *as, struct dm_target *ti) 197 { 198 struct switch_ctx *sctx = ti->private; 199 unsigned long long start; 200 int r; 201 202 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table), 203 &sctx->path_list[sctx->nr_paths].dmdev); 204 if (r) { 205 ti->error = "Device lookup failed"; 206 return r; 207 } 208 209 if (kstrtoull(dm_shift_arg(as), 10, &start) || start != (sector_t)start) { 210 ti->error = "Invalid device starting offset"; 211 dm_put_device(ti, sctx->path_list[sctx->nr_paths].dmdev); 212 return -EINVAL; 213 } 214 215 sctx->path_list[sctx->nr_paths].start = start; 216 217 sctx->nr_paths++; 218 219 return 0; 220 } 221 222 /* 223 * Destructor: Don't free the dm_target, just the ti->private data (if any). 224 */ 225 static void switch_dtr(struct dm_target *ti) 226 { 227 struct switch_ctx *sctx = ti->private; 228 229 while (sctx->nr_paths--) 230 dm_put_device(ti, sctx->path_list[sctx->nr_paths].dmdev); 231 232 vfree(sctx->region_table); 233 kfree(sctx); 234 } 235 236 /* 237 * Constructor arguments: 238 * <num_paths> <region_size> <num_optional_args> [<optional_args>...] 239 * [<dev_path> <offset>]+ 240 * 241 * Optional args are to allow for future extension: currently this 242 * parameter must be 0. 243 */ 244 static int switch_ctr(struct dm_target *ti, unsigned argc, char **argv) 245 { 246 static struct dm_arg _args[] = { 247 {1, (KMALLOC_MAX_SIZE - sizeof(struct switch_ctx)) / sizeof(struct switch_path), "Invalid number of paths"}, 248 {1, UINT_MAX, "Invalid region size"}, 249 {0, 0, "Invalid number of optional args"}, 250 }; 251 252 struct switch_ctx *sctx; 253 struct dm_arg_set as; 254 unsigned nr_paths, region_size, nr_optional_args; 255 int r; 256 257 as.argc = argc; 258 as.argv = argv; 259 260 r = dm_read_arg(_args, &as, &nr_paths, &ti->error); 261 if (r) 262 return -EINVAL; 263 264 r = dm_read_arg(_args + 1, &as, ®ion_size, &ti->error); 265 if (r) 266 return r; 267 268 r = dm_read_arg_group(_args + 2, &as, &nr_optional_args, &ti->error); 269 if (r) 270 return r; 271 /* parse optional arguments here, if we add any */ 272 273 if (as.argc != nr_paths * 2) { 274 ti->error = "Incorrect number of path arguments"; 275 return -EINVAL; 276 } 277 278 sctx = alloc_switch_ctx(ti, nr_paths, region_size); 279 if (!sctx) { 280 ti->error = "Cannot allocate redirection context"; 281 return -ENOMEM; 282 } 283 284 r = dm_set_target_max_io_len(ti, region_size); 285 if (r) 286 goto error; 287 288 while (as.argc) { 289 r = parse_path(&as, ti); 290 if (r) 291 goto error; 292 } 293 294 r = alloc_region_table(ti, nr_paths); 295 if (r) 296 goto error; 297 298 initialise_region_table(sctx); 299 300 /* For UNMAP, sending the request down any path is sufficient */ 301 ti->num_discard_bios = 1; 302 303 return 0; 304 305 error: 306 switch_dtr(ti); 307 308 return r; 309 } 310 311 static int switch_map(struct dm_target *ti, struct bio *bio) 312 { 313 struct switch_ctx *sctx = ti->private; 314 sector_t offset = dm_target_offset(ti, bio->bi_iter.bi_sector); 315 unsigned path_nr = switch_get_path_nr(sctx, offset); 316 317 bio->bi_bdev = sctx->path_list[path_nr].dmdev->bdev; 318 bio->bi_iter.bi_sector = sctx->path_list[path_nr].start + offset; 319 320 return DM_MAPIO_REMAPPED; 321 } 322 323 /* 324 * We need to parse hex numbers in the message as quickly as possible. 325 * 326 * This table-based hex parser improves performance. 327 * It improves a time to load 1000000 entries compared to the condition-based 328 * parser. 329 * table-based parser condition-based parser 330 * PA-RISC 0.29s 0.31s 331 * Opteron 0.0495s 0.0498s 332 */ 333 static const unsigned char hex_table[256] = { 334 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 335 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 336 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 337 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 255, 255, 255, 255, 255, 255, 338 255, 10, 11, 12, 13, 14, 15, 255, 255, 255, 255, 255, 255, 255, 255, 255, 339 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 340 255, 10, 11, 12, 13, 14, 15, 255, 255, 255, 255, 255, 255, 255, 255, 255, 341 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 342 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 343 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 344 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 345 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 346 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 347 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 348 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 349 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255 350 }; 351 352 static __always_inline unsigned long parse_hex(const char **string) 353 { 354 unsigned char d; 355 unsigned long r = 0; 356 357 while ((d = hex_table[(unsigned char)**string]) < 16) { 358 r = (r << 4) | d; 359 (*string)++; 360 } 361 362 return r; 363 } 364 365 static int process_set_region_mappings(struct switch_ctx *sctx, 366 unsigned argc, char **argv) 367 { 368 unsigned i; 369 unsigned long region_index = 0; 370 371 for (i = 1; i < argc; i++) { 372 unsigned long path_nr; 373 const char *string = argv[i]; 374 375 if (*string == ':') 376 region_index++; 377 else { 378 region_index = parse_hex(&string); 379 if (unlikely(*string != ':')) { 380 DMWARN("invalid set_region_mappings argument: '%s'", argv[i]); 381 return -EINVAL; 382 } 383 } 384 385 string++; 386 if (unlikely(!*string)) { 387 DMWARN("invalid set_region_mappings argument: '%s'", argv[i]); 388 return -EINVAL; 389 } 390 391 path_nr = parse_hex(&string); 392 if (unlikely(*string)) { 393 DMWARN("invalid set_region_mappings argument: '%s'", argv[i]); 394 return -EINVAL; 395 } 396 if (unlikely(region_index >= sctx->nr_regions)) { 397 DMWARN("invalid set_region_mappings region number: %lu >= %lu", region_index, sctx->nr_regions); 398 return -EINVAL; 399 } 400 if (unlikely(path_nr >= sctx->nr_paths)) { 401 DMWARN("invalid set_region_mappings device: %lu >= %u", path_nr, sctx->nr_paths); 402 return -EINVAL; 403 } 404 405 switch_region_table_write(sctx, region_index, path_nr); 406 } 407 408 return 0; 409 } 410 411 /* 412 * Messages are processed one-at-a-time. 413 * 414 * Only set_region_mappings is supported. 415 */ 416 static int switch_message(struct dm_target *ti, unsigned argc, char **argv) 417 { 418 static DEFINE_MUTEX(message_mutex); 419 420 struct switch_ctx *sctx = ti->private; 421 int r = -EINVAL; 422 423 mutex_lock(&message_mutex); 424 425 if (!strcasecmp(argv[0], "set_region_mappings")) 426 r = process_set_region_mappings(sctx, argc, argv); 427 else 428 DMWARN("Unrecognised message received."); 429 430 mutex_unlock(&message_mutex); 431 432 return r; 433 } 434 435 static void switch_status(struct dm_target *ti, status_type_t type, 436 unsigned status_flags, char *result, unsigned maxlen) 437 { 438 struct switch_ctx *sctx = ti->private; 439 unsigned sz = 0; 440 int path_nr; 441 442 switch (type) { 443 case STATUSTYPE_INFO: 444 result[0] = '\0'; 445 break; 446 447 case STATUSTYPE_TABLE: 448 DMEMIT("%u %u 0", sctx->nr_paths, sctx->region_size); 449 for (path_nr = 0; path_nr < sctx->nr_paths; path_nr++) 450 DMEMIT(" %s %llu", sctx->path_list[path_nr].dmdev->name, 451 (unsigned long long)sctx->path_list[path_nr].start); 452 break; 453 } 454 } 455 456 /* 457 * Switch ioctl: 458 * 459 * Passthrough all ioctls to the path for sector 0 460 */ 461 static int switch_ioctl(struct dm_target *ti, unsigned cmd, 462 unsigned long arg) 463 { 464 struct switch_ctx *sctx = ti->private; 465 struct block_device *bdev; 466 fmode_t mode; 467 unsigned path_nr; 468 int r = 0; 469 470 path_nr = switch_get_path_nr(sctx, 0); 471 472 bdev = sctx->path_list[path_nr].dmdev->bdev; 473 mode = sctx->path_list[path_nr].dmdev->mode; 474 475 /* 476 * Only pass ioctls through if the device sizes match exactly. 477 */ 478 if (ti->len + sctx->path_list[path_nr].start != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) 479 r = scsi_verify_blk_ioctl(NULL, cmd); 480 481 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg); 482 } 483 484 static int switch_iterate_devices(struct dm_target *ti, 485 iterate_devices_callout_fn fn, void *data) 486 { 487 struct switch_ctx *sctx = ti->private; 488 int path_nr; 489 int r; 490 491 for (path_nr = 0; path_nr < sctx->nr_paths; path_nr++) { 492 r = fn(ti, sctx->path_list[path_nr].dmdev, 493 sctx->path_list[path_nr].start, ti->len, data); 494 if (r) 495 return r; 496 } 497 498 return 0; 499 } 500 501 static struct target_type switch_target = { 502 .name = "switch", 503 .version = {1, 0, 0}, 504 .module = THIS_MODULE, 505 .ctr = switch_ctr, 506 .dtr = switch_dtr, 507 .map = switch_map, 508 .message = switch_message, 509 .status = switch_status, 510 .ioctl = switch_ioctl, 511 .iterate_devices = switch_iterate_devices, 512 }; 513 514 static int __init dm_switch_init(void) 515 { 516 int r; 517 518 r = dm_register_target(&switch_target); 519 if (r < 0) 520 DMERR("dm_register_target() failed %d", r); 521 522 return r; 523 } 524 525 static void __exit dm_switch_exit(void) 526 { 527 dm_unregister_target(&switch_target); 528 } 529 530 module_init(dm_switch_init); 531 module_exit(dm_switch_exit); 532 533 MODULE_DESCRIPTION(DM_NAME " dynamic path switching target"); 534 MODULE_AUTHOR("Kevin D. O'Kelley <Kevin_OKelley@dell.com>"); 535 MODULE_AUTHOR("Narendran Ganapathy <Narendran_Ganapathy@dell.com>"); 536 MODULE_AUTHOR("Jim Ramsay <Jim_Ramsay@dell.com>"); 537 MODULE_AUTHOR("Mikulas Patocka <mpatocka@redhat.com>"); 538 MODULE_LICENSE("GPL"); 539