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.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 completion *wait;
  37         struct dm_io_client *client;
  38         io_notify_fn callback;
  39         void *context;
  40         void *vma_invalidate_address;
  41         unsigned long vma_invalidate_size;
  42 } __attribute__((aligned(DM_IO_MAX_REGIONS)));
  43
  44 static struct kmem_cache *_dm_io_cache;
  45
  46 /*
  47  * Create a client with mempool and bioset.
  48  */
  49 struct dm_io_client *dm_io_client_create(void)
  50 {
  51         struct dm_io_client *client;
  52         unsigned min_ios = dm_get_reserved_bio_based_ios();
  53
  54         client = kmalloc(sizeof(*client), GFP_KERNEL);
  55         if (!client)
  56                 return ERR_PTR(-ENOMEM);
  57
  58         client->pool = mempool_create_slab_pool(min_ios, _dm_io_cache);
  59         if (!client->pool)
  60                 goto bad;
  61
  62         client->bios = bioset_create(min_ios, 0);
  63         if (!client->bios)
  64                 goto bad;
  65
  66         return client;
  67
  68    bad:
  69         if (client->pool)
  70                 mempool_destroy(client->pool);
  71         kfree(client);
  72         return ERR_PTR(-ENOMEM);
  73 }
  74 EXPORT_SYMBOL(dm_io_client_create);
  75
  76 void dm_io_client_destroy(struct dm_io_client *client)
  77 {
  78         mempool_destroy(client->pool);
  79         bioset_free(client->bios);
  80         kfree(client);
  81 }
  82 EXPORT_SYMBOL(dm_io_client_destroy);
  83
  84 /*-----------------------------------------------------------------
  85  * We need to keep track of which region a bio is doing io for.
  86  * To avoid a memory allocation to store just 5 or 6 bits, we
  87  * ensure the 'struct io' pointer is aligned so enough low bits are
  88  * always zero and then combine it with the region number directly in
  89  * bi_private.
  90  *---------------------------------------------------------------*/
  91 static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
  92                                        unsigned region)
  93 {
  94         if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
  95                 DMCRIT("Unaligned struct io pointer %p", io);
  96                 BUG();
  97         }
  98
  99         bio->bi_private = (void *)((unsigned long)io | region);
 100 }
 101
 102 static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
 103                                        unsigned *region)
 104 {
 105         unsigned long val = (unsigned long)bio->bi_private;
 106
 107         *io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
 108         *region = val & (DM_IO_MAX_REGIONS - 1);
 109 }
 110
 111 /*-----------------------------------------------------------------
 112  * We need an io object to keep track of the number of bios that
 113  * have been dispatched for a particular io.
 114  *---------------------------------------------------------------*/
 115 static void dec_count(struct io *io, unsigned int region, int error)
 116 {
 117         if (error)
 118                 set_bit(region, &io->error_bits);
 119
 120         if (atomic_dec_and_test(&io->count)) {
 121                 if (io->vma_invalidate_size)
 122                         invalidate_kernel_vmap_range(io->vma_invalidate_address,
 123                                                      io->vma_invalidate_size);
 124
 125                 if (io->wait)
 126                         complete(io->wait);
 127
 128                 else {
 129                         unsigned long r = io->error_bits;
 130                         io_notify_fn fn = io->callback;
 131                         void *context = io->context;
 132
 133                         mempool_free(io, io->client->pool);
 134                         fn(r, context);
 135                 }
 136         }
 137 }
 138
 139 static void endio(struct bio *bio, int error)
 140 {
 141         struct io *io;
 142         unsigned region;
 143
 144         if (error && bio_data_dir(bio) == READ)
 145                 zero_fill_bio(bio);
 146
 147         /*
 148          * The bio destructor in bio_put() may use the io object.
 149          */
 150         retrieve_io_and_region_from_bio(bio, &io, &region);
 151
 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
 295         /*
 296          * where->count may be zero if rw holds a flush and we need to
 297          * send a zero-sized flush.
 298          */
 299         do {
 300                 /*
 301                  * Allocate a suitably sized-bio.
 302                  */
 303                 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME))
 304                         num_bvecs = 1;
 305                 else
 306                         num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev),
 307                                           dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
 308
 309                 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
 310                 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
 311                 bio->bi_bdev = where->bdev;
 312                 bio->bi_end_io = endio;
 313                 store_io_and_region_in_bio(bio, io, region);
 314
 315                 if (rw & REQ_DISCARD) {
 316                         num_sectors = min_t(sector_t, q->limits.max_discard_sectors, remaining);
 317                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 318                         remaining -= num_sectors;
 319                 } else if (rw & REQ_WRITE_SAME) {
 320                         /*
 321                          * WRITE SAME only uses a single page.
 322                          */
 323                         dp->get_page(dp, &page, &len, &offset);
 324                         bio_add_page(bio, page, logical_block_size, offset);
 325                         num_sectors = min_t(sector_t, q->limits.max_write_same_sectors, remaining);
 326                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 327
 328                         offset = 0;
 329                         remaining -= num_sectors;
 330                         dp->next_page(dp);
 331                 } else while (remaining) {
 332                         /*
 333                          * Try and add as many pages as possible.
 334                          */
 335                         dp->get_page(dp, &page, &len, &offset);
 336                         len = min(len, to_bytes(remaining));
 337                         if (!bio_add_page(bio, page, len, offset))
 338                                 break;
 339
 340                         offset = 0;
 341                         remaining -= to_sector(len);
 342                         dp->next_page(dp);
 343                 }
 344
 345                 atomic_inc(&io->count);
 346                 submit_bio(rw, bio);
 347         } while (remaining);
 348 }
 349
 350 static void dispatch_io(int rw, unsigned int num_regions,
 351                         struct dm_io_region *where, struct dpages *dp,
 352                         struct io *io, int sync)
 353 {
 354         int i;
 355         struct dpages old_pages = *dp;
 356
 357         BUG_ON(num_regions > DM_IO_MAX_REGIONS);
 358
 359         if (sync)
 360                 rw |= REQ_SYNC;
 361
 362         /*
 363          * For multiple regions we need to be careful to rewind
 364          * the dp object for each call to do_region.
 365          */
 366         for (i = 0; i < num_regions; i++) {
 367                 *dp = old_pages;
 368                 if (where[i].count || (rw & REQ_FLUSH))
 369                         do_region(rw, i, where + i, dp, io);
 370         }
 371
 372         /*
 373          * Drop the extra reference that we were holding to avoid
 374          * the io being completed too early.
 375          */
 376         dec_count(io, 0, 0);
 377 }
 378
 379 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 380                    struct dm_io_region *where, int rw, struct dpages *dp,
 381                    unsigned long *error_bits)
 382 {
 383         /*
 384          * gcc <= 4.3 can't do the alignment for stack variables, so we must
 385          * align it on our own.
 386          * volatile prevents the optimizer from removing or reusing
 387          * "io_" field from the stack frame (allowed in ANSI C).
 388          */
 389         volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
 390         struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
 391         DECLARE_COMPLETION_ONSTACK(wait);
 392
 393         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 394                 WARN_ON(1);
 395                 return -EIO;
 396         }
 397
 398         io->error_bits = 0;
 399         atomic_set(&io->count, 1); /* see dispatch_io() */
 400         io->wait = &wait;
 401         io->client = client;
 402
 403         io->vma_invalidate_address = dp->vma_invalidate_address;
 404         io->vma_invalidate_size = dp->vma_invalidate_size;
 405
 406         dispatch_io(rw, num_regions, where, dp, io, 1);
 407
 408         wait_for_completion_io(&wait);
 409
 410         if (error_bits)
 411                 *error_bits = io->error_bits;
 412
 413         return io->error_bits ? -EIO : 0;
 414 }
 415
 416 static int async_io(struct dm_io_client *client, unsigned int num_regions,
 417                     struct dm_io_region *where, int rw, struct dpages *dp,
 418                     io_notify_fn fn, void *context)
 419 {
 420         struct io *io;
 421
 422         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 423                 WARN_ON(1);
 424                 fn(1, context);
 425                 return -EIO;
 426         }
 427
 428         io = mempool_alloc(client->pool, GFP_NOIO);
 429         io->error_bits = 0;
 430         atomic_set(&io->count, 1); /* see dispatch_io() */
 431         io->wait = NULL;
 432         io->client = client;
 433         io->callback = fn;
 434         io->context = context;
 435
 436         io->vma_invalidate_address = dp->vma_invalidate_address;
 437         io->vma_invalidate_size = dp->vma_invalidate_size;
 438
 439         dispatch_io(rw, num_regions, where, dp, io, 0);
 440         return 0;
 441 }
 442
 443 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
 444                    unsigned long size)
 445 {
 446         /* Set up dpages based on memory type */
 447
 448         dp->vma_invalidate_address = NULL;
 449         dp->vma_invalidate_size = 0;
 450
 451         switch (io_req->mem.type) {
 452         case DM_IO_PAGE_LIST:
 453                 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
 454                 break;
 455
 456         case DM_IO_BIO:
 457                 bio_dp_init(dp, io_req->mem.ptr.bio);
 458                 break;
 459
 460         case DM_IO_VMA:
 461                 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
 462                 if ((io_req->bi_rw & RW_MASK) == READ) {
 463                         dp->vma_invalidate_address = io_req->mem.ptr.vma;
 464                         dp->vma_invalidate_size = size;
 465                 }
 466                 vm_dp_init(dp, io_req->mem.ptr.vma);
 467                 break;
 468
 469         case DM_IO_KMEM:
 470                 km_dp_init(dp, io_req->mem.ptr.addr);
 471                 break;
 472
 473         default:
 474                 return -EINVAL;
 475         }
 476
 477         return 0;
 478 }
 479
 480 /*
 481  * New collapsed (a)synchronous interface.
 482  *
 483  * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
 484  * the queue with blk_unplug() some time later or set REQ_SYNC in
 485 io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
 486  * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
 487  */
 488 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
 489           struct dm_io_region *where, unsigned long *sync_error_bits)
 490 {
 491         int r;
 492         struct dpages dp;
 493
 494         r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
 495         if (r)
 496                 return r;
 497
 498         if (!io_req->notify.fn)
 499                 return sync_io(io_req->client, num_regions, where,
 500                                io_req->bi_rw, &dp, sync_error_bits);
 501
 502         return async_io(io_req->client, num_regions, where, io_req->bi_rw,
 503                         &dp, io_req->notify.fn, io_req->notify.context);
 504 }
 505 EXPORT_SYMBOL(dm_io);
 506
 507 int __init dm_io_init(void)
 508 {
 509         _dm_io_cache = KMEM_CACHE(io, 0);
 510         if (!_dm_io_cache)
 511                 return -ENOMEM;
 512
 513         return 0;
 514 }
 515
 516 void dm_io_exit(void)
 517 {
 518         kmem_cache_destroy(_dm_io_cache);
 519         _dm_io_cache = NULL;
 520 }