1 /* 2 * Copyright (C) 2003 Sistina Software 3 * Copyright (C) 2006 Red Hat GmbH 4 * 5 * This file is released under the GPL. 6 */ 7 8 #include "dm.h" 9 10 #include <linux/device-mapper.h> 11 12 #include <linux/bio.h> 13 #include <linux/completion.h> 14 #include <linux/mempool.h> 15 #include <linux/module.h> 16 #include <linux/sched.h> 17 #include <linux/slab.h> 18 #include <linux/dm-io.h> 19 20 #define DM_MSG_PREFIX "io" 21 22 #define DM_IO_MAX_REGIONS BITS_PER_LONG 23 24 struct dm_io_client { 25 mempool_t *pool; 26 struct bio_set *bios; 27 }; 28 29 /* 30 * Aligning 'struct io' reduces the number of bits required to store 31 * its address. Refer to store_io_and_region_in_bio() below. 32 */ 33 struct io { 34 unsigned long error_bits; 35 atomic_t count; 36 struct dm_io_client *client; 37 io_notify_fn callback; 38 void *context; 39 void *vma_invalidate_address; 40 unsigned long vma_invalidate_size; 41 } __attribute__((aligned(DM_IO_MAX_REGIONS))); 42 43 static struct kmem_cache *_dm_io_cache; 44 45 /* 46 * Create a client with mempool and bioset. 47 */ 48 struct dm_io_client *dm_io_client_create(void) 49 { 50 struct dm_io_client *client; 51 unsigned min_ios = dm_get_reserved_bio_based_ios(); 52 53 client = kmalloc(sizeof(*client), GFP_KERNEL); 54 if (!client) 55 return ERR_PTR(-ENOMEM); 56 57 client->pool = mempool_create_slab_pool(min_ios, _dm_io_cache); 58 if (!client->pool) 59 goto bad; 60 61 client->bios = bioset_create(min_ios, 0); 62 if (!client->bios) 63 goto bad; 64 65 return client; 66 67 bad: 68 if (client->pool) 69 mempool_destroy(client->pool); 70 kfree(client); 71 return ERR_PTR(-ENOMEM); 72 } 73 EXPORT_SYMBOL(dm_io_client_create); 74 75 void dm_io_client_destroy(struct dm_io_client *client) 76 { 77 mempool_destroy(client->pool); 78 bioset_free(client->bios); 79 kfree(client); 80 } 81 EXPORT_SYMBOL(dm_io_client_destroy); 82 83 /*----------------------------------------------------------------- 84 * We need to keep track of which region a bio is doing io for. 85 * To avoid a memory allocation to store just 5 or 6 bits, we 86 * ensure the 'struct io' pointer is aligned so enough low bits are 87 * always zero and then combine it with the region number directly in 88 * bi_private. 89 *---------------------------------------------------------------*/ 90 static void store_io_and_region_in_bio(struct bio *bio, struct io *io, 91 unsigned region) 92 { 93 if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) { 94 DMCRIT("Unaligned struct io pointer %p", io); 95 BUG(); 96 } 97 98 bio->bi_private = (void *)((unsigned long)io | region); 99 } 100 101 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io, 102 unsigned *region) 103 { 104 unsigned long val = (unsigned long)bio->bi_private; 105 106 *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS); 107 *region = val & (DM_IO_MAX_REGIONS - 1); 108 } 109 110 /*----------------------------------------------------------------- 111 * We need an io object to keep track of the number of bios that 112 * have been dispatched for a particular io. 113 *---------------------------------------------------------------*/ 114 static void complete_io(struct io *io) 115 { 116 unsigned long error_bits = io->error_bits; 117 io_notify_fn fn = io->callback; 118 void *context = io->context; 119 120 if (io->vma_invalidate_size) 121 invalidate_kernel_vmap_range(io->vma_invalidate_address, 122 io->vma_invalidate_size); 123 124 mempool_free(io, io->client->pool); 125 fn(error_bits, context); 126 } 127 128 static void dec_count(struct io *io, unsigned int region, int error) 129 { 130 if (error) 131 set_bit(region, &io->error_bits); 132 133 if (atomic_dec_and_test(&io->count)) 134 complete_io(io); 135 } 136 137 static void endio(struct bio *bio) 138 { 139 struct io *io; 140 unsigned region; 141 int error; 142 143 if (bio->bi_error && bio_data_dir(bio) == READ) 144 zero_fill_bio(bio); 145 146 /* 147 * The bio destructor in bio_put() may use the io object. 148 */ 149 retrieve_io_and_region_from_bio(bio, &io, ®ion); 150 151 error = bio->bi_error; 152 bio_put(bio); 153 154 dec_count(io, region, error); 155 } 156 157 /*----------------------------------------------------------------- 158 * These little objects provide an abstraction for getting a new 159 * destination page for io. 160 *---------------------------------------------------------------*/ 161 struct dpages { 162 void (*get_page)(struct dpages *dp, 163 struct page **p, unsigned long *len, unsigned *offset); 164 void (*next_page)(struct dpages *dp); 165 166 unsigned context_u; 167 void *context_ptr; 168 169 void *vma_invalidate_address; 170 unsigned long vma_invalidate_size; 171 }; 172 173 /* 174 * Functions for getting the pages from a list. 175 */ 176 static void list_get_page(struct dpages *dp, 177 struct page **p, unsigned long *len, unsigned *offset) 178 { 179 unsigned o = dp->context_u; 180 struct page_list *pl = (struct page_list *) dp->context_ptr; 181 182 *p = pl->page; 183 *len = PAGE_SIZE - o; 184 *offset = o; 185 } 186 187 static void list_next_page(struct dpages *dp) 188 { 189 struct page_list *pl = (struct page_list *) dp->context_ptr; 190 dp->context_ptr = pl->next; 191 dp->context_u = 0; 192 } 193 194 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset) 195 { 196 dp->get_page = list_get_page; 197 dp->next_page = list_next_page; 198 dp->context_u = offset; 199 dp->context_ptr = pl; 200 } 201 202 /* 203 * Functions for getting the pages from a bvec. 204 */ 205 static void bio_get_page(struct dpages *dp, struct page **p, 206 unsigned long *len, unsigned *offset) 207 { 208 struct bio_vec *bvec = dp->context_ptr; 209 *p = bvec->bv_page; 210 *len = bvec->bv_len - dp->context_u; 211 *offset = bvec->bv_offset + dp->context_u; 212 } 213 214 static void bio_next_page(struct dpages *dp) 215 { 216 struct bio_vec *bvec = dp->context_ptr; 217 dp->context_ptr = bvec + 1; 218 dp->context_u = 0; 219 } 220 221 static void bio_dp_init(struct dpages *dp, struct bio *bio) 222 { 223 dp->get_page = bio_get_page; 224 dp->next_page = bio_next_page; 225 dp->context_ptr = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 226 dp->context_u = bio->bi_iter.bi_bvec_done; 227 } 228 229 /* 230 * Functions for getting the pages from a VMA. 231 */ 232 static void vm_get_page(struct dpages *dp, 233 struct page **p, unsigned long *len, unsigned *offset) 234 { 235 *p = vmalloc_to_page(dp->context_ptr); 236 *offset = dp->context_u; 237 *len = PAGE_SIZE - dp->context_u; 238 } 239 240 static void vm_next_page(struct dpages *dp) 241 { 242 dp->context_ptr += PAGE_SIZE - dp->context_u; 243 dp->context_u = 0; 244 } 245 246 static void vm_dp_init(struct dpages *dp, void *data) 247 { 248 dp->get_page = vm_get_page; 249 dp->next_page = vm_next_page; 250 dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); 251 dp->context_ptr = data; 252 } 253 254 /* 255 * Functions for getting the pages from kernel memory. 256 */ 257 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len, 258 unsigned *offset) 259 { 260 *p = virt_to_page(dp->context_ptr); 261 *offset = dp->context_u; 262 *len = PAGE_SIZE - dp->context_u; 263 } 264 265 static void km_next_page(struct dpages *dp) 266 { 267 dp->context_ptr += PAGE_SIZE - dp->context_u; 268 dp->context_u = 0; 269 } 270 271 static void km_dp_init(struct dpages *dp, void *data) 272 { 273 dp->get_page = km_get_page; 274 dp->next_page = km_next_page; 275 dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); 276 dp->context_ptr = data; 277 } 278 279 /*----------------------------------------------------------------- 280 * IO routines that accept a list of pages. 281 *---------------------------------------------------------------*/ 282 static void do_region(int rw, unsigned region, struct dm_io_region *where, 283 struct dpages *dp, struct io *io) 284 { 285 struct bio *bio; 286 struct page *page; 287 unsigned long len; 288 unsigned offset; 289 unsigned num_bvecs; 290 sector_t remaining = where->count; 291 struct request_queue *q = bdev_get_queue(where->bdev); 292 unsigned short logical_block_size = queue_logical_block_size(q); 293 sector_t num_sectors; 294 unsigned int uninitialized_var(special_cmd_max_sectors); 295 296 /* 297 * Reject unsupported discard and write same requests. 298 */ 299 if (rw & REQ_DISCARD) 300 special_cmd_max_sectors = q->limits.max_discard_sectors; 301 else if (rw & REQ_WRITE_SAME) 302 special_cmd_max_sectors = q->limits.max_write_same_sectors; 303 if ((rw & (REQ_DISCARD | REQ_WRITE_SAME)) && special_cmd_max_sectors == 0) { 304 dec_count(io, region, -EOPNOTSUPP); 305 return; 306 } 307 308 /* 309 * where->count may be zero if rw holds a flush and we need to 310 * send a zero-sized flush. 311 */ 312 do { 313 /* 314 * Allocate a suitably sized-bio. 315 */ 316 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME)) 317 num_bvecs = 1; 318 else 319 num_bvecs = min_t(int, BIO_MAX_PAGES, 320 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT))); 321 322 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios); 323 bio->bi_iter.bi_sector = where->sector + (where->count - remaining); 324 bio->bi_bdev = where->bdev; 325 bio->bi_end_io = endio; 326 store_io_and_region_in_bio(bio, io, region); 327 328 if (rw & REQ_DISCARD) { 329 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining); 330 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT; 331 remaining -= num_sectors; 332 } else if (rw & REQ_WRITE_SAME) { 333 /* 334 * WRITE SAME only uses a single page. 335 */ 336 dp->get_page(dp, &page, &len, &offset); 337 bio_add_page(bio, page, logical_block_size, offset); 338 num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining); 339 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT; 340 341 offset = 0; 342 remaining -= num_sectors; 343 dp->next_page(dp); 344 } else while (remaining) { 345 /* 346 * Try and add as many pages as possible. 347 */ 348 dp->get_page(dp, &page, &len, &offset); 349 len = min(len, to_bytes(remaining)); 350 if (!bio_add_page(bio, page, len, offset)) 351 break; 352 353 offset = 0; 354 remaining -= to_sector(len); 355 dp->next_page(dp); 356 } 357 358 atomic_inc(&io->count); 359 submit_bio(rw, bio); 360 } while (remaining); 361 } 362 363 static void dispatch_io(int rw, unsigned int num_regions, 364 struct dm_io_region *where, struct dpages *dp, 365 struct io *io, int sync) 366 { 367 int i; 368 struct dpages old_pages = *dp; 369 370 BUG_ON(num_regions > DM_IO_MAX_REGIONS); 371 372 if (sync) 373 rw |= REQ_SYNC; 374 375 /* 376 * For multiple regions we need to be careful to rewind 377 * the dp object for each call to do_region. 378 */ 379 for (i = 0; i < num_regions; i++) { 380 *dp = old_pages; 381 if (where[i].count || (rw & REQ_FLUSH)) 382 do_region(rw, i, where + i, dp, io); 383 } 384 385 /* 386 * Drop the extra reference that we were holding to avoid 387 * the io being completed too early. 388 */ 389 dec_count(io, 0, 0); 390 } 391 392 struct sync_io { 393 unsigned long error_bits; 394 struct completion wait; 395 }; 396 397 static void sync_io_complete(unsigned long error, void *context) 398 { 399 struct sync_io *sio = context; 400 401 sio->error_bits = error; 402 complete(&sio->wait); 403 } 404 405 static int sync_io(struct dm_io_client *client, unsigned int num_regions, 406 struct dm_io_region *where, int rw, struct dpages *dp, 407 unsigned long *error_bits) 408 { 409 struct io *io; 410 struct sync_io sio; 411 412 if (num_regions > 1 && (rw & RW_MASK) != WRITE) { 413 WARN_ON(1); 414 return -EIO; 415 } 416 417 init_completion(&sio.wait); 418 419 io = mempool_alloc(client->pool, GFP_NOIO); 420 io->error_bits = 0; 421 atomic_set(&io->count, 1); /* see dispatch_io() */ 422 io->client = client; 423 io->callback = sync_io_complete; 424 io->context = &sio; 425 426 io->vma_invalidate_address = dp->vma_invalidate_address; 427 io->vma_invalidate_size = dp->vma_invalidate_size; 428 429 dispatch_io(rw, num_regions, where, dp, io, 1); 430 431 wait_for_completion_io(&sio.wait); 432 433 if (error_bits) 434 *error_bits = sio.error_bits; 435 436 return sio.error_bits ? -EIO : 0; 437 } 438 439 static int async_io(struct dm_io_client *client, unsigned int num_regions, 440 struct dm_io_region *where, int rw, struct dpages *dp, 441 io_notify_fn fn, void *context) 442 { 443 struct io *io; 444 445 if (num_regions > 1 && (rw & RW_MASK) != WRITE) { 446 WARN_ON(1); 447 fn(1, context); 448 return -EIO; 449 } 450 451 io = mempool_alloc(client->pool, GFP_NOIO); 452 io->error_bits = 0; 453 atomic_set(&io->count, 1); /* see dispatch_io() */ 454 io->client = client; 455 io->callback = fn; 456 io->context = context; 457 458 io->vma_invalidate_address = dp->vma_invalidate_address; 459 io->vma_invalidate_size = dp->vma_invalidate_size; 460 461 dispatch_io(rw, num_regions, where, dp, io, 0); 462 return 0; 463 } 464 465 static int dp_init(struct dm_io_request *io_req, struct dpages *dp, 466 unsigned long size) 467 { 468 /* Set up dpages based on memory type */ 469 470 dp->vma_invalidate_address = NULL; 471 dp->vma_invalidate_size = 0; 472 473 switch (io_req->mem.type) { 474 case DM_IO_PAGE_LIST: 475 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset); 476 break; 477 478 case DM_IO_BIO: 479 bio_dp_init(dp, io_req->mem.ptr.bio); 480 break; 481 482 case DM_IO_VMA: 483 flush_kernel_vmap_range(io_req->mem.ptr.vma, size); 484 if ((io_req->bi_rw & RW_MASK) == READ) { 485 dp->vma_invalidate_address = io_req->mem.ptr.vma; 486 dp->vma_invalidate_size = size; 487 } 488 vm_dp_init(dp, io_req->mem.ptr.vma); 489 break; 490 491 case DM_IO_KMEM: 492 km_dp_init(dp, io_req->mem.ptr.addr); 493 break; 494 495 default: 496 return -EINVAL; 497 } 498 499 return 0; 500 } 501 502 /* 503 * New collapsed (a)synchronous interface. 504 * 505 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug 506 * the queue with blk_unplug() some time later or set REQ_SYNC in io_req->bi_rw. 507 * If you fail to do one of these, the IO will be submitted to the disk after 508 * q->unplug_delay, which defaults to 3ms in blk-settings.c. 509 */ 510 int dm_io(struct dm_io_request *io_req, unsigned num_regions, 511 struct dm_io_region *where, unsigned long *sync_error_bits) 512 { 513 int r; 514 struct dpages dp; 515 516 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT); 517 if (r) 518 return r; 519 520 if (!io_req->notify.fn) 521 return sync_io(io_req->client, num_regions, where, 522 io_req->bi_rw, &dp, sync_error_bits); 523 524 return async_io(io_req->client, num_regions, where, io_req->bi_rw, 525 &dp, io_req->notify.fn, io_req->notify.context); 526 } 527 EXPORT_SYMBOL(dm_io); 528 529 int __init dm_io_init(void) 530 { 531 _dm_io_cache = KMEM_CACHE(io, 0); 532 if (!_dm_io_cache) 533 return -ENOMEM; 534 535 return 0; 536 } 537 538 void dm_io_exit(void) 539 { 540 kmem_cache_destroy(_dm_io_cache); 541 _dm_io_cache = NULL; 542 } 543