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, int error) 138 { 139 struct io *io; 140 unsigned region; 141 142 if (error && bio_data_dir(bio) == READ) 143 zero_fill_bio(bio); 144 145 /* 146 * The bio destructor in bio_put() may use the io object. 147 */ 148 retrieve_io_and_region_from_bio(bio, &io, ®ion); 149 150 bio_put(bio); 151 152 dec_count(io, region, error); 153 } 154 155 /*----------------------------------------------------------------- 156 * These little objects provide an abstraction for getting a new 157 * destination page for io. 158 *---------------------------------------------------------------*/ 159 struct dpages { 160 void (*get_page)(struct dpages *dp, 161 struct page **p, unsigned long *len, unsigned *offset); 162 void (*next_page)(struct dpages *dp); 163 164 unsigned context_u; 165 void *context_ptr; 166 167 void *vma_invalidate_address; 168 unsigned long vma_invalidate_size; 169 }; 170 171 /* 172 * Functions for getting the pages from a list. 173 */ 174 static void list_get_page(struct dpages *dp, 175 struct page **p, unsigned long *len, unsigned *offset) 176 { 177 unsigned o = dp->context_u; 178 struct page_list *pl = (struct page_list *) dp->context_ptr; 179 180 *p = pl->page; 181 *len = PAGE_SIZE - o; 182 *offset = o; 183 } 184 185 static void list_next_page(struct dpages *dp) 186 { 187 struct page_list *pl = (struct page_list *) dp->context_ptr; 188 dp->context_ptr = pl->next; 189 dp->context_u = 0; 190 } 191 192 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset) 193 { 194 dp->get_page = list_get_page; 195 dp->next_page = list_next_page; 196 dp->context_u = offset; 197 dp->context_ptr = pl; 198 } 199 200 /* 201 * Functions for getting the pages from a bvec. 202 */ 203 static void bio_get_page(struct dpages *dp, struct page **p, 204 unsigned long *len, unsigned *offset) 205 { 206 struct bio_vec *bvec = dp->context_ptr; 207 *p = bvec->bv_page; 208 *len = bvec->bv_len - dp->context_u; 209 *offset = bvec->bv_offset + dp->context_u; 210 } 211 212 static void bio_next_page(struct dpages *dp) 213 { 214 struct bio_vec *bvec = dp->context_ptr; 215 dp->context_ptr = bvec + 1; 216 dp->context_u = 0; 217 } 218 219 static void bio_dp_init(struct dpages *dp, struct bio *bio) 220 { 221 dp->get_page = bio_get_page; 222 dp->next_page = bio_next_page; 223 dp->context_ptr = __bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter); 224 dp->context_u = bio->bi_iter.bi_bvec_done; 225 } 226 227 /* 228 * Functions for getting the pages from a VMA. 229 */ 230 static void vm_get_page(struct dpages *dp, 231 struct page **p, unsigned long *len, unsigned *offset) 232 { 233 *p = vmalloc_to_page(dp->context_ptr); 234 *offset = dp->context_u; 235 *len = PAGE_SIZE - dp->context_u; 236 } 237 238 static void vm_next_page(struct dpages *dp) 239 { 240 dp->context_ptr += PAGE_SIZE - dp->context_u; 241 dp->context_u = 0; 242 } 243 244 static void vm_dp_init(struct dpages *dp, void *data) 245 { 246 dp->get_page = vm_get_page; 247 dp->next_page = vm_next_page; 248 dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); 249 dp->context_ptr = data; 250 } 251 252 /* 253 * Functions for getting the pages from kernel memory. 254 */ 255 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len, 256 unsigned *offset) 257 { 258 *p = virt_to_page(dp->context_ptr); 259 *offset = dp->context_u; 260 *len = PAGE_SIZE - dp->context_u; 261 } 262 263 static void km_next_page(struct dpages *dp) 264 { 265 dp->context_ptr += PAGE_SIZE - dp->context_u; 266 dp->context_u = 0; 267 } 268 269 static void km_dp_init(struct dpages *dp, void *data) 270 { 271 dp->get_page = km_get_page; 272 dp->next_page = km_next_page; 273 dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1); 274 dp->context_ptr = data; 275 } 276 277 /*----------------------------------------------------------------- 278 * IO routines that accept a list of pages. 279 *---------------------------------------------------------------*/ 280 static void do_region(int rw, unsigned region, struct dm_io_region *where, 281 struct dpages *dp, struct io *io) 282 { 283 struct bio *bio; 284 struct page *page; 285 unsigned long len; 286 unsigned offset; 287 unsigned num_bvecs; 288 sector_t remaining = where->count; 289 struct request_queue *q = bdev_get_queue(where->bdev); 290 unsigned short logical_block_size = queue_logical_block_size(q); 291 sector_t num_sectors; 292 293 /* Reject unsupported discard requests */ 294 if ((rw & REQ_DISCARD) && !blk_queue_discard(q)) { 295 dec_count(io, region, -EOPNOTSUPP); 296 return; 297 } 298 299 /* 300 * where->count may be zero if rw holds a flush and we need to 301 * send a zero-sized flush. 302 */ 303 do { 304 /* 305 * Allocate a suitably sized-bio. 306 */ 307 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME)) 308 num_bvecs = 1; 309 else 310 num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev), 311 dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT))); 312 313 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios); 314 bio->bi_iter.bi_sector = where->sector + (where->count - remaining); 315 bio->bi_bdev = where->bdev; 316 bio->bi_end_io = endio; 317 store_io_and_region_in_bio(bio, io, region); 318 319 if (rw & REQ_DISCARD) { 320 num_sectors = min_t(sector_t, q->limits.max_discard_sectors, remaining); 321 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT; 322 remaining -= num_sectors; 323 } else if (rw & REQ_WRITE_SAME) { 324 /* 325 * WRITE SAME only uses a single page. 326 */ 327 dp->get_page(dp, &page, &len, &offset); 328 bio_add_page(bio, page, logical_block_size, offset); 329 num_sectors = min_t(sector_t, q->limits.max_write_same_sectors, remaining); 330 bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT; 331 332 offset = 0; 333 remaining -= num_sectors; 334 dp->next_page(dp); 335 } else while (remaining) { 336 /* 337 * Try and add as many pages as possible. 338 */ 339 dp->get_page(dp, &page, &len, &offset); 340 len = min(len, to_bytes(remaining)); 341 if (!bio_add_page(bio, page, len, offset)) 342 break; 343 344 offset = 0; 345 remaining -= to_sector(len); 346 dp->next_page(dp); 347 } 348 349 atomic_inc(&io->count); 350 submit_bio(rw, bio); 351 } while (remaining); 352 } 353 354 static void dispatch_io(int rw, unsigned int num_regions, 355 struct dm_io_region *where, struct dpages *dp, 356 struct io *io, int sync) 357 { 358 int i; 359 struct dpages old_pages = *dp; 360 361 BUG_ON(num_regions > DM_IO_MAX_REGIONS); 362 363 if (sync) 364 rw |= REQ_SYNC; 365 366 /* 367 * For multiple regions we need to be careful to rewind 368 * the dp object for each call to do_region. 369 */ 370 for (i = 0; i < num_regions; i++) { 371 *dp = old_pages; 372 if (where[i].count || (rw & REQ_FLUSH)) 373 do_region(rw, i, where + i, dp, io); 374 } 375 376 /* 377 * Drop the extra reference that we were holding to avoid 378 * the io being completed too early. 379 */ 380 dec_count(io, 0, 0); 381 } 382 383 struct sync_io { 384 unsigned long error_bits; 385 struct completion wait; 386 }; 387 388 static void sync_io_complete(unsigned long error, void *context) 389 { 390 struct sync_io *sio = context; 391 392 sio->error_bits = error; 393 complete(&sio->wait); 394 } 395 396 static int sync_io(struct dm_io_client *client, unsigned int num_regions, 397 struct dm_io_region *where, int rw, struct dpages *dp, 398 unsigned long *error_bits) 399 { 400 struct io *io; 401 struct sync_io sio; 402 403 if (num_regions > 1 && (rw & RW_MASK) != WRITE) { 404 WARN_ON(1); 405 return -EIO; 406 } 407 408 init_completion(&sio.wait); 409 410 io = mempool_alloc(client->pool, GFP_NOIO); 411 io->error_bits = 0; 412 atomic_set(&io->count, 1); /* see dispatch_io() */ 413 io->client = client; 414 io->callback = sync_io_complete; 415 io->context = &sio; 416 417 io->vma_invalidate_address = dp->vma_invalidate_address; 418 io->vma_invalidate_size = dp->vma_invalidate_size; 419 420 dispatch_io(rw, num_regions, where, dp, io, 1); 421 422 wait_for_completion_io(&sio.wait); 423 424 if (error_bits) 425 *error_bits = sio.error_bits; 426 427 return sio.error_bits ? -EIO : 0; 428 } 429 430 static int async_io(struct dm_io_client *client, unsigned int num_regions, 431 struct dm_io_region *where, int rw, struct dpages *dp, 432 io_notify_fn fn, void *context) 433 { 434 struct io *io; 435 436 if (num_regions > 1 && (rw & RW_MASK) != WRITE) { 437 WARN_ON(1); 438 fn(1, context); 439 return -EIO; 440 } 441 442 io = mempool_alloc(client->pool, GFP_NOIO); 443 io->error_bits = 0; 444 atomic_set(&io->count, 1); /* see dispatch_io() */ 445 io->client = client; 446 io->callback = fn; 447 io->context = context; 448 449 io->vma_invalidate_address = dp->vma_invalidate_address; 450 io->vma_invalidate_size = dp->vma_invalidate_size; 451 452 dispatch_io(rw, num_regions, where, dp, io, 0); 453 return 0; 454 } 455 456 static int dp_init(struct dm_io_request *io_req, struct dpages *dp, 457 unsigned long size) 458 { 459 /* Set up dpages based on memory type */ 460 461 dp->vma_invalidate_address = NULL; 462 dp->vma_invalidate_size = 0; 463 464 switch (io_req->mem.type) { 465 case DM_IO_PAGE_LIST: 466 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset); 467 break; 468 469 case DM_IO_BIO: 470 bio_dp_init(dp, io_req->mem.ptr.bio); 471 break; 472 473 case DM_IO_VMA: 474 flush_kernel_vmap_range(io_req->mem.ptr.vma, size); 475 if ((io_req->bi_rw & RW_MASK) == READ) { 476 dp->vma_invalidate_address = io_req->mem.ptr.vma; 477 dp->vma_invalidate_size = size; 478 } 479 vm_dp_init(dp, io_req->mem.ptr.vma); 480 break; 481 482 case DM_IO_KMEM: 483 km_dp_init(dp, io_req->mem.ptr.addr); 484 break; 485 486 default: 487 return -EINVAL; 488 } 489 490 return 0; 491 } 492 493 /* 494 * New collapsed (a)synchronous interface. 495 * 496 * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug 497 * the queue with blk_unplug() some time later or set REQ_SYNC in io_req->bi_rw. 498 * If you fail to do one of these, the IO will be submitted to the disk after 499 * q->unplug_delay, which defaults to 3ms in blk-settings.c. 500 */ 501 int dm_io(struct dm_io_request *io_req, unsigned num_regions, 502 struct dm_io_region *where, unsigned long *sync_error_bits) 503 { 504 int r; 505 struct dpages dp; 506 507 r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT); 508 if (r) 509 return r; 510 511 if (!io_req->notify.fn) 512 return sync_io(io_req->client, num_regions, where, 513 io_req->bi_rw, &dp, sync_error_bits); 514 515 return async_io(io_req->client, num_regions, where, io_req->bi_rw, 516 &dp, io_req->notify.fn, io_req->notify.context); 517 } 518 EXPORT_SYMBOL(dm_io); 519 520 int __init dm_io_init(void) 521 { 522 _dm_io_cache = KMEM_CACHE(io, 0); 523 if (!_dm_io_cache) 524 return -ENOMEM; 525 526 return 0; 527 } 528 529 void dm_io_exit(void) 530 { 531 kmem_cache_destroy(_dm_io_cache); 532 _dm_io_cache = NULL; 533 } 534