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/mempool.h>
  14 #include <linux/module.h>
  15 #include <linux/sched.h>
  16 #include <linux/slab.h>
  17 #include <linux/dm-io.h>
  18
  19 #define DM_MSG_PREFIX "io"
  20
  21 #define DM_IO_MAX_REGIONS       BITS_PER_LONG
  22
  23 struct dm_io_client {
  24         mempool_t *pool;
  25         struct bio_set *bios;
  26 };
  27
  28 /*
  29  * Aligning 'struct io' reduces the number of bits required to store
  30  * its address.  Refer to store_io_and_region_in_bio() below.
  31  */
  32 struct io {
  33         unsigned long error_bits;
  34         atomic_t count;
  35         struct task_struct *sleeper;
  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 dec_count(struct io *io, unsigned int region, int error)
 115 {
 116         if (error)
 117                 set_bit(region, &io->error_bits);
 118
 119         if (atomic_dec_and_test(&io->count)) {
 120                 if (io->vma_invalidate_size)
 121                         invalidate_kernel_vmap_range(io->vma_invalidate_address,
 122                                                      io->vma_invalidate_size);
 123
 124                 if (io->sleeper)
 125                         wake_up_process(io->sleeper);
 126
 127                 else {
 128                         unsigned long r = io->error_bits;
 129                         io_notify_fn fn = io->callback;
 130                         void *context = io->context;
 131
 132                         mempool_free(io, io->client->pool);
 133                         fn(r, context);
 134                 }
 135         }
 136 }
 137
 138 static void endio(struct bio *bio, int error)
 139 {
 140         struct io *io;
 141         unsigned region;
 142
 143         if (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, &region);
 150
 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         unsigned context_u;
 166         void *context_ptr;
 167
 168         void *vma_invalidate_address;
 169         unsigned long vma_invalidate_size;
 170 };
 171
 172 /*
 173  * Functions for getting the pages from a list.
 174  */
 175 static void list_get_page(struct dpages *dp,
 176                   struct page **p, unsigned long *len, unsigned *offset)
 177 {
 178         unsigned o = dp->context_u;
 179         struct page_list *pl = (struct page_list *) dp->context_ptr;
 180
 181         *p = pl->page;
 182         *len = PAGE_SIZE - o;
 183         *offset = o;
 184 }
 185
 186 static void list_next_page(struct dpages *dp)
 187 {
 188         struct page_list *pl = (struct page_list *) dp->context_ptr;
 189         dp->context_ptr = pl->next;
 190         dp->context_u = 0;
 191 }
 192
 193 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
 194 {
 195         dp->get_page = list_get_page;
 196         dp->next_page = list_next_page;
 197         dp->context_u = offset;
 198         dp->context_ptr = pl;
 199 }
 200
 201 /*
 202  * Functions for getting the pages from a bvec.
 203  */
 204 static void bvec_get_page(struct dpages *dp,
 205                   struct page **p, unsigned long *len, unsigned *offset)
 206 {
 207         struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
 208         *p = bvec->bv_page;
 209         *len = bvec->bv_len;
 210         *offset = bvec->bv_offset;
 211 }
 212
 213 static void bvec_next_page(struct dpages *dp)
 214 {
 215         struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
 216         dp->context_ptr = bvec + 1;
 217 }
 218
 219 static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
 220 {
 221         dp->get_page = bvec_get_page;
 222         dp->next_page = bvec_next_page;
 223         dp->context_ptr = bvec;
 224 }
 225
 226 /*
 227  * Functions for getting the pages from a VMA.
 228  */
 229 static void vm_get_page(struct dpages *dp,
 230                  struct page **p, unsigned long *len, unsigned *offset)
 231 {
 232         *p = vmalloc_to_page(dp->context_ptr);
 233         *offset = dp->context_u;
 234         *len = PAGE_SIZE - dp->context_u;
 235 }
 236
 237 static void vm_next_page(struct dpages *dp)
 238 {
 239         dp->context_ptr += PAGE_SIZE - dp->context_u;
 240         dp->context_u = 0;
 241 }
 242
 243 static void vm_dp_init(struct dpages *dp, void *data)
 244 {
 245         dp->get_page = vm_get_page;
 246         dp->next_page = vm_next_page;
 247         dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
 248         dp->context_ptr = data;
 249 }
 250
 251 /*
 252  * Functions for getting the pages from kernel memory.
 253  */
 254 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
 255                         unsigned *offset)
 256 {
 257         *p = virt_to_page(dp->context_ptr);
 258         *offset = dp->context_u;
 259         *len = PAGE_SIZE - dp->context_u;
 260 }
 261
 262 static void km_next_page(struct dpages *dp)
 263 {
 264         dp->context_ptr += PAGE_SIZE - dp->context_u;
 265         dp->context_u = 0;
 266 }
 267
 268 static void km_dp_init(struct dpages *dp, void *data)
 269 {
 270         dp->get_page = km_get_page;
 271         dp->next_page = km_next_page;
 272         dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
 273         dp->context_ptr = data;
 274 }
 275
 276 /*-----------------------------------------------------------------
 277  * IO routines that accept a list of pages.
 278  *---------------------------------------------------------------*/
 279 static void do_region(int rw, unsigned region, struct dm_io_region *where,
 280                       struct dpages *dp, struct io *io)
 281 {
 282         struct bio *bio;
 283         struct page *page;
 284         unsigned long len;
 285         unsigned offset;
 286         unsigned num_bvecs;
 287         sector_t remaining = where->count;
 288         struct request_queue *q = bdev_get_queue(where->bdev);
 289         unsigned short logical_block_size = queue_logical_block_size(q);
 290         sector_t num_sectors;
 291
 292         /*
 293          * where->count may be zero if rw holds a flush and we need to
 294          * send a zero-sized flush.
 295          */
 296         do {
 297                 /*
 298                  * Allocate a suitably sized-bio.
 299                  */
 300                 if ((rw & REQ_DISCARD) || (rw & REQ_WRITE_SAME))
 301                         num_bvecs = 1;
 302                 else
 303                         num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev),
 304                                           dm_sector_div_up(remaining, (PAGE_SIZE >> SECTOR_SHIFT)));
 305
 306                 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
 307                 bio->bi_iter.bi_sector = where->sector + (where->count - remaining);
 308                 bio->bi_bdev = where->bdev;
 309                 bio->bi_end_io = endio;
 310                 store_io_and_region_in_bio(bio, io, region);
 311
 312                 if (rw & REQ_DISCARD) {
 313                         num_sectors = min_t(sector_t, q->limits.max_discard_sectors, remaining);
 314                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 315                         remaining -= num_sectors;
 316                 } else if (rw & REQ_WRITE_SAME) {
 317                         /*
 318                          * WRITE SAME only uses a single page.
 319                          */
 320                         dp->get_page(dp, &page, &len, &offset);
 321                         bio_add_page(bio, page, logical_block_size, offset);
 322                         num_sectors = min_t(sector_t, q->limits.max_write_same_sectors, remaining);
 323                         bio->bi_iter.bi_size = num_sectors << SECTOR_SHIFT;
 324
 325                         offset = 0;
 326                         remaining -= num_sectors;
 327                         dp->next_page(dp);
 328                 } else while (remaining) {
 329                         /*
 330                          * Try and add as many pages as possible.
 331                          */
 332                         dp->get_page(dp, &page, &len, &offset);
 333                         len = min(len, to_bytes(remaining));
 334                         if (!bio_add_page(bio, page, len, offset))
 335                                 break;
 336
 337                         offset = 0;
 338                         remaining -= to_sector(len);
 339                         dp->next_page(dp);
 340                 }
 341
 342                 atomic_inc(&io->count);
 343                 submit_bio(rw, bio);
 344         } while (remaining);
 345 }
 346
 347 static void dispatch_io(int rw, unsigned int num_regions,
 348                         struct dm_io_region *where, struct dpages *dp,
 349                         struct io *io, int sync)
 350 {
 351         int i;
 352         struct dpages old_pages = *dp;
 353
 354         BUG_ON(num_regions > DM_IO_MAX_REGIONS);
 355
 356         if (sync)
 357                 rw |= REQ_SYNC;
 358
 359         /*
 360          * For multiple regions we need to be careful to rewind
 361          * the dp object for each call to do_region.
 362          */
 363         for (i = 0; i < num_regions; i++) {
 364                 *dp = old_pages;
 365                 if (where[i].count || (rw & REQ_FLUSH))
 366                         do_region(rw, i, where + i, dp, io);
 367         }
 368
 369         /*
 370          * Drop the extra reference that we were holding to avoid
 371          * the io being completed too early.
 372          */
 373         dec_count(io, 0, 0);
 374 }
 375
 376 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 377                    struct dm_io_region *where, int rw, struct dpages *dp,
 378                    unsigned long *error_bits)
 379 {
 380         /*
 381          * gcc <= 4.3 can't do the alignment for stack variables, so we must
 382          * align it on our own.
 383          * volatile prevents the optimizer from removing or reusing
 384          * "io_" field from the stack frame (allowed in ANSI C).
 385          */
 386         volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
 387         struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
 388
 389         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 390                 WARN_ON(1);
 391                 return -EIO;
 392         }
 393
 394         io->error_bits = 0;
 395         atomic_set(&io->count, 1); /* see dispatch_io() */
 396         io->sleeper = current;
 397         io->client = client;
 398
 399         io->vma_invalidate_address = dp->vma_invalidate_address;
 400         io->vma_invalidate_size = dp->vma_invalidate_size;
 401
 402         dispatch_io(rw, num_regions, where, dp, io, 1);
 403
 404         while (1) {
 405                 set_current_state(TASK_UNINTERRUPTIBLE);
 406
 407                 if (!atomic_read(&io->count))
 408                         break;
 409
 410                 io_schedule();
 411         }
 412         set_current_state(TASK_RUNNING);
 413
 414         if (error_bits)
 415                 *error_bits = io->error_bits;
 416
 417         return io->error_bits ? -EIO : 0;
 418 }
 419
 420 static int async_io(struct dm_io_client *client, unsigned int num_regions,
 421                     struct dm_io_region *where, int rw, struct dpages *dp,
 422                     io_notify_fn fn, void *context)
 423 {
 424         struct io *io;
 425
 426         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 427                 WARN_ON(1);
 428                 fn(1, context);
 429                 return -EIO;
 430         }
 431
 432         io = mempool_alloc(client->pool, GFP_NOIO);
 433         io->error_bits = 0;
 434         atomic_set(&io->count, 1); /* see dispatch_io() */
 435         io->sleeper = NULL;
 436         io->client = client;
 437         io->callback = fn;
 438         io->context = context;
 439
 440         io->vma_invalidate_address = dp->vma_invalidate_address;
 441         io->vma_invalidate_size = dp->vma_invalidate_size;
 442
 443         dispatch_io(rw, num_regions, where, dp, io, 0);
 444         return 0;
 445 }
 446
 447 static int dp_init(struct dm_io_request *io_req, struct dpages *dp,
 448                    unsigned long size)
 449 {
 450         /* Set up dpages based on memory type */
 451
 452         dp->vma_invalidate_address = NULL;
 453         dp->vma_invalidate_size = 0;
 454
 455         switch (io_req->mem.type) {
 456         case DM_IO_PAGE_LIST:
 457                 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
 458                 break;
 459
 460         case DM_IO_BVEC:
 461                 bvec_dp_init(dp, io_req->mem.ptr.bvec);
 462                 break;
 463
 464         case DM_IO_VMA:
 465                 flush_kernel_vmap_range(io_req->mem.ptr.vma, size);
 466                 if ((io_req->bi_rw & RW_MASK) == READ) {
 467                         dp->vma_invalidate_address = io_req->mem.ptr.vma;
 468                         dp->vma_invalidate_size = size;
 469                 }
 470                 vm_dp_init(dp, io_req->mem.ptr.vma);
 471                 break;
 472
 473         case DM_IO_KMEM:
 474                 km_dp_init(dp, io_req->mem.ptr.addr);
 475                 break;
 476
 477         default:
 478                 return -EINVAL;
 479         }
 480
 481         return 0;
 482 }
 483
 484 /*
 485  * New collapsed (a)synchronous interface.
 486  *
 487  * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
 488  * the queue with blk_unplug() some time later or set REQ_SYNC in
 489 io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
 490  * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
 491  */
 492 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
 493           struct dm_io_region *where, unsigned long *sync_error_bits)
 494 {
 495         int r;
 496         struct dpages dp;
 497
 498         r = dp_init(io_req, &dp, (unsigned long)where->count << SECTOR_SHIFT);
 499         if (r)
 500                 return r;
 501
 502         if (!io_req->notify.fn)
 503                 return sync_io(io_req->client, num_regions, where,
 504                                io_req->bi_rw, &dp, sync_error_bits);
 505
 506         return async_io(io_req->client, num_regions, where, io_req->bi_rw,
 507                         &dp, io_req->notify.fn, io_req->notify.context);
 508 }
 509 EXPORT_SYMBOL(dm_io);
 510
 511 int __init dm_io_init(void)
 512 {
 513         _dm_io_cache = KMEM_CACHE(io, 0);
 514         if (!_dm_io_cache)
 515                 return -ENOMEM;
 516
 517         return 0;
 518 }
 519
 520 void dm_io_exit(void)
 521 {
 522         kmem_cache_destroy(_dm_io_cache);
 523         _dm_io_cache = NULL;
 524 }