drivers/md/dm-io.c

   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-core.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         mempool_destroy(client->pool);
  69         kfree(client);
  70         return ERR_PTR(-ENOMEM);
  71 }
  72 EXPORT_SYMBOL(dm_io_client_create);
  73
  74 void dm_io_client_destroy(struct dm_io_client *client)
  75 {
  76         mempool_destroy(client->pool);
  77         bioset_free(client->bios);
  78         kfree(client);
  79 }
  80 EXPORT_SYMBOL(dm_io_client_destroy);
  81
  82 /*-----------------------------------------------------------------
  83  * We need to keep track of which region a bio is doing io for.
  84  * To avoid a memory allocation to store just 5 or 6 bits, we
  85  * ensure the 'struct io' pointer is aligned so enough low bits are
  86  * always zero and then combine it with the region number directly in
  87  * bi_private.
  88  *---------------------------------------------------------------*/
  89 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
  90                                        unsigned region)
  91 {
  92         if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
  93                 DMCRIT("Unaligned struct io pointer %p", io);
  94                 BUG();
  95         }
  96
  97         bio->bi_private = (void *)((unsigned long)io | region);
  98 }
  99
 100 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
 101                                        unsigned *region)
 102 {
 103         unsigned long val = (unsigned long)bio->bi_private;
 104
 105         *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
 106         *region = val & (DM_IO_MAX_REGIONS - 1);
 107 }
 108
 109 /*-----------------------------------------------------------------
 110  * We need an io object to keep track of the number of bios that
 111  * have been dispatched for a particular io.
 112  *---------------------------------------------------------------*/
 113 static void complete_io(struct io *io)
 114 {
 115         unsigned long error_bits = io->error_bits;
 116         io_notify_fn fn = io->callback;
 117         void *context = io->context;
 118
 119         if (io->vma_invalidate_size)
 120                 invalidate_kernel_vmap_range(io->vma_invalidate_address,
 121                                              io->vma_invalidate_size);
 122
 123         mempool_free(io, io->client->pool);
 124         fn(error_bits, context);
 125 }
 126
 127 static void dec_count(struct io *io, unsigned int region, int error)
 128 {
 129         if (error)
 130                 set_bit(region, &io->error_bits);
 131
 132         if (atomic_dec_and_test(&io->count))
 133                 complete_io(io);
 134 }
 135
 136 static void endio(struct bio *bio)
 137 {
 138         struct io *io;
 139         unsigned region;
 140         int error;
 141
 142         if (bio->bi_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, &region);
 149
 150         error = bio->bi_error;
 151         bio_put(bio);
 152
 153         dec_count(io, region, error);
 154 }
 155
 156 /*-----------------------------------------------------------------
 157  * These little objects provide an abstraction for getting a new
 158  * destination page for io.
 159  *---------------------------------------------------------------*/
 160 struct dpages {
 161         void (*get_page)(struct dpages *dp,
 162                          struct page **p, unsigned long *len, unsigned *offset);
 163         void (*next_page)(struct dpages *dp);
 164
 165         union {
 166                 unsigned context_u;
 167                 struct bvec_iter context_bi;
 168         };
 169         void *context_ptr;
 170
 171         void *vma_invalidate_address;
 172         unsigned long vma_invalidate_size;
 173 };
 174
 175 /*
 176  * Functions for getting the pages from a list.
 177  */
 178 static void list_get_page(struct dpages *dp,
 179                   struct page **p, unsigned long *len, unsigned *offset)
 180 {
 181         unsigned o = dp->context_u;
 182         struct page_list *pl = (struct page_list *) dp->context_ptr;
 183
 184         *p = pl->page;
 185         *len = PAGE_SIZE - o;
 186         *offset = o;
 187 }
 188
 189 static void list_next_page(struct dpages *dp)
 190 {
 191         struct page_list *pl = (struct page_list *) dp->context_ptr;
 192         dp->context_ptr = pl->next;
 193         dp->context_u = 0;
 194 }
 195
 196 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
 197 {
 198         dp->get_page = list_get_page;
 199         dp->next_page = list_next_page;
 200         dp->context_u = offset;
 201         dp->context_ptr = pl;
 202 }
 203
 204 /*
 205  * Functions for getting the pages from a bvec.
 206  */
 207 static void bio_get_page(struct dpages *dp, struct page **p,
 208                          unsigned long *len, unsigned *offset)
 209 {
 210         struct bio_vec bvec = bvec_iter_bvec((struct bio_vec *)dp->context_ptr,
 211                                              dp->context_bi);
 212
 213         *p = bvec.bv_page;
 214         *len = bvec.bv_len;
 215         *offset = bvec.bv_offset;
 216
 217         /* avoid figuring it out again in bio_next_page() */
 218         dp->context_bi.bi_sector = (sector_t)bvec.bv_len;
 219 }
 220
 221 static void bio_next_page(struct dpages *dp)
 222 {
 223         unsigned int len = (unsigned int)dp->context_bi.bi_sector;
 224
 225         bvec_iter_advance((struct bio_vec *)dp->context_ptr,
 226                           &dp->context_bi, len);
 227 }
 228
 229 static void bio_dp_init(struct dpages *dp, struct bio *bio)
 230 {
 231         dp->get_page = bio_get_page;
 232         dp->next_page = bio_next_page;
 233
 234         /*
 235          * We just use bvec iterator to retrieve pages, so it is ok to
 236          * access the bvec table directly here
 237          */
 238         dp->context_ptr = bio->bi_io_vec;
 239         dp->context_bi = bio->bi_iter;
 240 }
 241
 242 /*
 243  * Functions for getting the pages from a VMA.
 244  */
 245 static void vm_get_page(struct dpages *dp,
 246                  struct page **p, unsigned long *len, unsigned *offset)
 247 {
 248         *p = vmalloc_to_page(dp->context_ptr);
 249         *offset = dp->context_u;
 250         *len = PAGE_SIZE - dp->context_u;
 251 }
 252
 253 static void vm_next_page(struct dpages *dp)
 254 {
 255         dp->context_ptr += PAGE_SIZE - dp->context_u;
 256         dp->context_u = 0;
 257 }
 258
 259 static void vm_dp_init(struct dpages *dp, void *data)
 260 {
 261         dp->get_page = vm_get_page;
 262         dp->next_page = vm_next_page;
 263         dp->context_u = offset_in_page(data);
 264         dp->context_ptr = data;
 265 }
 266
 267 /*
 268  * Functions for getting the pages from kernel memory.
 269  */
 270 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
 271                         unsigned *offset)
 272 {
 273         *p = virt_to_page(dp->context_ptr);
 274         *offset = dp->context_u;
 275         *len = PAGE_SIZE - dp->context_u;
 276 }
 277
 278 static void km_next_page(struct dpages *dp)
 279 {
 280         dp->context_ptr += PAGE_SIZE - dp->context_u;
 281         dp->context_u = 0;
 282 }
 283
 284 static void km_dp_init(struct dpages *dp, void *data)
 285 {
 286         dp->get_page = km_get_page;
 287         dp->next_page = km_next_page;
 288         dp->context_u = offset_in_page(data);
 289         dp->context_ptr = data;
 290 }
 291
 292 /*-----------------------------------------------------------------
 293  * IO routines that accept a list of pages.
 294  *---------------------------------------------------------------*/
 295 static void do_region(int op, int op_flags, unsigned region,
 296                       struct dm_io_region *where, struct dpages *dp,
 297                       struct io *io)
 298 {
 299         struct bio *bio;
 300         struct page *page;
 301         unsigned long len;
 302         unsigned offset;
 303         unsigned num_bvecs;
 304         sector_t remaining = where->count;
 305         struct request_queue *q = bdev_get_queue(where->bdev);
 306         unsigned short logical_block_size = queue_logical_block_size(q);
 307         sector_t num_sectors;
 308         unsigned int uninitialized_var(special_cmd_max_sectors);
 309
 310         /*
 311          * Reject unsupported discard and write same requests.
 312          */
 313         if (op == REQ_OP_DISCARD)
 314                 special_cmd_max_sectors = q->limits.max_discard_sectors;
 315         else if (op == REQ_OP_WRITE_SAME)
 316                 special_cmd_max_sectors = q->limits.max_write_same_sectors;
 317         if ((op == REQ_OP_DISCARD || op == REQ_OP_WRITE_SAME) &&
 318             special_cmd_max_sectors == 0) {
 319                 dec_count(io, region, -EOPNOTSUPP);
 320                 return;
 321         }
 322
 323         /*
 324          * where->count may be zero if op holds a flush and we need to
 325          * send a zero-sized flush.
 326          */
 327         do {
 328                 /*
 329                  * Allocate a suitably sized-bio.
 330                  */
 331                 if ((op == REQ_OP_DISCARD) || (op == REQ_OP_WRITE_SAME))
 332                         num_bvecs = 1;
 333                 else
 334                         num_bvecs = min_t(int, BIO_MAX_PAGES,
 335                                           dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
 336
 337                 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
 338                 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
 339                 bio->bi_bdev = where->bdev;
 340                 bio->bi_end_io = endio;
 341                 bio_set_op_attrs(bio, op, op_flags);
 342                 store_io_and_region_in_bio(bio, io, region);
 343
 344                 if (op == REQ_OP_DISCARD) {
 345                         num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
 346                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 347                         remaining -= num_sectors;
 348                 } else if (op == REQ_OP_WRITE_SAME) {
 349                         /*
 350                          * WRITE SAME only uses a single page.
 351                          */
 352                         dp->get_page(dp, &page, &len, &offset);
 353                         bio_add_page(bio, page, logical_block_size, offset);
 354                         num_sectors = min_t(sector_t, special_cmd_max_sectors, remaining);
 355                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 356
 357                         offset = 0;
 358                         remaining -= num_sectors;
 359                         dp->next_page(dp);
 360                 } else while (remaining) {
 361                         /*
 362                          * Try and add as many pages as possible.
 363                          */
 364                         dp->get_page(dp, &page, &len, &offset);
 365                         len = min(len, to_bytes(remaining));
 366                         if (!bio_add_page(bio, page, len, offset))
 367                                 break;
 368
 369                         offset = 0;
 370                         remaining -= to_sector(len);
 371                         dp->next_page(dp);
 372                 }
 373
 374                 atomic_inc(&io->count);
 375                 submit_bio(bio);
 376         } while (remaining);
 377 }
 378
 379 static void dispatch_io(int op, int op_flags, unsigned int num_regions,
 380                         struct dm_io_region *where, struct dpages *dp,
 381                         struct io *io, int sync)
 382 {
 383         int i;
 384         struct dpages old_pages = *dp;
 385
 386         BUG_ON(num_regions > DM_IO_MAX_REGIONS);
 387
 388         if (sync)
 389                 op_flags |= REQ_SYNC;
 390
 391         /*
 392          * For multiple regions we need to be careful to rewind
 393          * the dp object for each call to do_region.
 394          */
 395         for (i = 0; i < num_regions; i++) {
 396                 *dp = old_pages;
 397                 if (where[i].count || (op_flags & REQ_PREFLUSH))
 398                         do_region(op, op_flags, i, where + i, dp, io);
 399         }
 400
 401         /*
 402          * Drop the extra reference that we were holding to avoid
 403          * the io being completed too early.
 404          */
 405         dec_count(io, 0, 0);
 406 }
 407
 408 struct sync_io {
 409         unsigned long error_bits;
 410         struct completion wait;
 411 };
 412
 413 static void sync_io_complete(unsigned long error, void *context)
 414 {
 415         struct sync_io *sio = context;
 416
 417         sio->error_bits = error;
 418         complete(&sio->wait);
 419 }
 420
 421 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 422                    struct dm_io_region *where, int op, int op_flags,
 423                    struct dpages *dp, unsigned long *error_bits)
 424 {
 425         struct io *io;
 426         struct sync_io sio;
 427
 428         if (num_regions > 1 && !op_is_write(op)) {
 429                 WARN_ON(1);
 430                 return -EIO;
 431         }
 432
 433         init_completion(&sio.wait);
 434
 435         io = mempool_alloc(client->pool, GFP_NOIO);
 436         io->error_bits = 0;
 437         atomic_set(&io->count, 1); /* see dispatch_io() */
 438         io->client = client;
 439         io->callback = sync_io_complete;
 440         io->context = &sio;
 441
 442         io->vma_invalidate_address = dp->vma_invalidate_address;
 443         io->vma_invalidate_size = dp->vma_invalidate_size;
 444
 445         dispatch_io(op, op_flags, num_regions, where, dp, io, 1);
 446
 447         wait_for_completion_io(&sio.wait);
 448
 449         if (error_bits)
 450                 *error_bits = sio.error_bits;
 451
 452         return sio.error_bits ? -EIO : 0;
 453 }
 454
 455 static int async_io(struct dm_io_client *client, unsigned int num_regions,
 456                     struct dm_io_region *where, int op, int op_flags,
 457                     struct dpages *dp, io_notify_fn fn, void *context)
 458 {
 459         struct io *io;
 460
 461         if (num_regions > 1 && !op_is_write(op)) {
 462                 WARN_ON(1);
 463                 fn(1, context);
 464                 return -EIO;
 465         }
 466
 467         io = mempool_alloc(client->pool, GFP_NOIO);
 468         io->error_bits = 0;
 469         atomic_set(&io->count, 1); /* see dispatch_io() */
 470         io->client = client;
 471         io->callback = fn;
 472         io->context = context;
 473
 474         io->vma_invalidate_address = dp->vma_invalidate_address;
 475         io->vma_invalidate_size = dp->vma_invalidate_size;
 476
 477         dispatch_io(op, op_flags, num_regions, where, dp, io, 0);
 478         return 0;
 479 }
 480
 481 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
 482                    unsigned long size)
 483 {
 484         /* Set up dpages based on memory type */
 485
 486         dp->vma_invalidate_address = NULL;
 487         dp->vma_invalidate_size = 0;
 488
 489         switch (io_req->mem.type) {
 490         case DM_IO_PAGE_LIST:
 491                 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
 492                 break;
 493
 494         case DM_IO_BIO:
 495                 bio_dp_init(dp, io_req->mem.ptr.bio);
 496                 break;
 497
 498         case DM_IO_VMA:
 499                 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
 500                 if (io_req->bi_op == REQ_OP_READ) {
 501                         dp->vma_invalidate_address = io_req->mem.ptr.vma;
 502                         dp->vma_invalidate_size = size;
 503                 }
 504                 vm_dp_init(dp, io_req->mem.ptr.vma);
 505                 break;
 506
 507         case DM_IO_KMEM:
 508                 km_dp_init(dp, io_req->mem.ptr.addr);
 509                 break;
 510
 511         default:
 512                 return -EINVAL;
 513         }
 514
 515         return 0;
 516 }
 517
 518 /*
 519  * New collapsed (a)synchronous interface.
 520  *
 521  * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
 522  * the queue with blk_unplug() some time later or set REQ_SYNC in
 523  * io_req->bi_opf. If you fail to do one of these, the IO will be submitted to
 524  * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
 525  */
 526 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
 527           struct dm_io_region *where, unsigned long *sync_error_bits)
 528 {
 529         int r;
 530         struct dpages dp;
 531
 532         r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
 533         if (r)
 534                 return r;
 535
 536         if (!io_req->notify.fn)
 537                 return sync_io(io_req->client, num_regions, where,
 538                                io_req->bi_op, io_req->bi_op_flags, &dp,
 539                                sync_error_bits);
 540
 541         return async_io(io_req->client, num_regions, where, io_req->bi_op,
 542                         io_req->bi_op_flags, &dp, io_req->notify.fn,
 543                         io_req->notify.context);
 544 }
 545 EXPORT_SYMBOL(dm_io);
 546
 547 int __init dm_io_init(void)
 548 {
 549         _dm_io_cache = KMEM_CACHE(io, 0);
 550         if (!_dm_io_cache)
 551                 return -ENOMEM;
 552
 553         return 0;
 554 }
 555
 556 void dm_io_exit(void)
 557 {
 558         kmem_cache_destroy(_dm_io_cache);
 559         _dm_io_cache = NULL;
 560 }