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 #define MIN_IOS         16
  23 #define MIN_BIOS        16
  24
  25 struct dm_io_client {
  26         mempool_t *pool;
  27         struct bio_set *bios;
  28 };
  29
  30 /*
  31  * Aligning 'struct io' reduces the number of bits required to store
  32  * its address.  Refer to store_io_and_region_in_bio() below.
  33  */
  34 struct io {
  35         unsigned long error_bits;
  36         atomic_t count;
  37         struct task_struct *sleeper;
  38         struct dm_io_client *client;
  39         io_notify_fn callback;
  40         void *context;
  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
  52         client = kmalloc(sizeof(*client), GFP_KERNEL);
  53         if (!client)
  54                 return ERR_PTR(-ENOMEM);
  55
  56         client->pool = mempool_create_slab_pool(MIN_IOS, _dm_io_cache);
  57         if (!client->pool)
  58                 goto bad;
  59
  60         client->bios = bioset_create(MIN_BIOS, 0);
  61         if (!client->bios)
  62                 goto bad;
  63
  64         return client;
  65
  66    bad:
  67         if (client->pool)
  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 dec_count(struct io *io, unsigned int region, int error)
 114 {
 115         if (error)
 116                 set_bit(region, &io->error_bits);
 117
 118         if (atomic_dec_and_test(&io->count)) {
 119                 if (io->sleeper)
 120                         wake_up_process(io->sleeper);
 121
 122                 else {
 123                         unsigned long r = io->error_bits;
 124                         io_notify_fn fn = io->callback;
 125                         void *context = io->context;
 126
 127                         mempool_free(io, io->client->pool);
 128                         fn(r, context);
 129                 }
 130         }
 131 }
 132
 133 static void endio(struct bio *bio, int error)
 134 {
 135         struct io *io;
 136         unsigned region;
 137
 138         if (error && bio_data_dir(bio) == READ)
 139                 zero_fill_bio(bio);
 140
 141         /*
 142          * The bio destructor in bio_put() may use the io object.
 143          */
 144         retrieve_io_and_region_from_bio(bio, &io, &region);
 145
 146         bio_put(bio);
 147
 148         dec_count(io, region, error);
 149 }
 150
 151 /*-----------------------------------------------------------------
 152  * These little objects provide an abstraction for getting a new
 153  * destination page for io.
 154  *---------------------------------------------------------------*/
 155 struct dpages {
 156         void (*get_page)(struct dpages *dp,
 157                          struct page **p, unsigned long *len, unsigned *offset);
 158         void (*next_page)(struct dpages *dp);
 159
 160         unsigned context_u;
 161         void *context_ptr;
 162 };
 163
 164 /*
 165  * Functions for getting the pages from a list.
 166  */
 167 static void list_get_page(struct dpages *dp,
 168                   struct page **p, unsigned long *len, unsigned *offset)
 169 {
 170         unsigned o = dp->context_u;
 171         struct page_list *pl = (struct page_list *) dp->context_ptr;
 172
 173         *p = pl->page;
 174         *len = PAGE_SIZE - o;
 175         *offset = o;
 176 }
 177
 178 static void list_next_page(struct dpages *dp)
 179 {
 180         struct page_list *pl = (struct page_list *) dp->context_ptr;
 181         dp->context_ptr = pl->next;
 182         dp->context_u = 0;
 183 }
 184
 185 static void list_dp_init(struct dpages *dp, struct page_list *pl, unsigned offset)
 186 {
 187         dp->get_page = list_get_page;
 188         dp->next_page = list_next_page;
 189         dp->context_u = offset;
 190         dp->context_ptr = pl;
 191 }
 192
 193 /*
 194  * Functions for getting the pages from a bvec.
 195  */
 196 static void bvec_get_page(struct dpages *dp,
 197                   struct page **p, unsigned long *len, unsigned *offset)
 198 {
 199         struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
 200         *p = bvec->bv_page;
 201         *len = bvec->bv_len;
 202         *offset = bvec->bv_offset;
 203 }
 204
 205 static void bvec_next_page(struct dpages *dp)
 206 {
 207         struct bio_vec *bvec = (struct bio_vec *) dp->context_ptr;
 208         dp->context_ptr = bvec + 1;
 209 }
 210
 211 static void bvec_dp_init(struct dpages *dp, struct bio_vec *bvec)
 212 {
 213         dp->get_page = bvec_get_page;
 214         dp->next_page = bvec_next_page;
 215         dp->context_ptr = bvec;
 216 }
 217
 218 /*
 219  * Functions for getting the pages from a VMA.
 220  */
 221 static void vm_get_page(struct dpages *dp,
 222                  struct page **p, unsigned long *len, unsigned *offset)
 223 {
 224         *p = vmalloc_to_page(dp->context_ptr);
 225         *offset = dp->context_u;
 226         *len = PAGE_SIZE - dp->context_u;
 227 }
 228
 229 static void vm_next_page(struct dpages *dp)
 230 {
 231         dp->context_ptr += PAGE_SIZE - dp->context_u;
 232         dp->context_u = 0;
 233 }
 234
 235 static void vm_dp_init(struct dpages *dp, void *data)
 236 {
 237         dp->get_page = vm_get_page;
 238         dp->next_page = vm_next_page;
 239         dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
 240         dp->context_ptr = data;
 241 }
 242
 243 static void dm_bio_destructor(struct bio *bio)
 244 {
 245         unsigned region;
 246         struct io *io;
 247
 248         retrieve_io_and_region_from_bio(bio, &io, &region);
 249
 250         bio_free(bio, io->client->bios);
 251 }
 252
 253 /*
 254  * Functions for getting the pages from kernel memory.
 255  */
 256 static void km_get_page(struct dpages *dp, struct page **p, unsigned long *len,
 257                         unsigned *offset)
 258 {
 259         *p = virt_to_page(dp->context_ptr);
 260         *offset = dp->context_u;
 261         *len = PAGE_SIZE - dp->context_u;
 262 }
 263
 264 static void km_next_page(struct dpages *dp)
 265 {
 266         dp->context_ptr += PAGE_SIZE - dp->context_u;
 267         dp->context_u = 0;
 268 }
 269
 270 static void km_dp_init(struct dpages *dp, void *data)
 271 {
 272         dp->get_page = km_get_page;
 273         dp->next_page = km_next_page;
 274         dp->context_u = ((unsigned long) data) & (PAGE_SIZE - 1);
 275         dp->context_ptr = data;
 276 }
 277
 278 /*-----------------------------------------------------------------
 279  * IO routines that accept a list of pages.
 280  *---------------------------------------------------------------*/
 281 static void do_region(int rw, unsigned region, struct dm_io_region *where,
 282                       struct dpages *dp, struct io *io)
 283 {
 284         struct bio *bio;
 285         struct page *page;
 286         unsigned long len;
 287         unsigned offset;
 288         unsigned num_bvecs;
 289         sector_t remaining = where->count;
 290
 291         /*
 292          * where->count may be zero if rw holds a flush and we need to
 293          * send a zero-sized flush.
 294          */
 295         do {
 296                 /*
 297                  * Allocate a suitably sized-bio.
 298                  */
 299                 num_bvecs = dm_sector_div_up(remaining,
 300                                              (PAGE_SIZE >> SECTOR_SHIFT));
 301                 num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev), num_bvecs);
 302                 bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
 303                 bio->bi_sector = where->sector + (where->count - remaining);
 304                 bio->bi_bdev = where->bdev;
 305                 bio->bi_end_io = endio;
 306                 bio->bi_destructor = dm_bio_destructor;
 307                 store_io_and_region_in_bio(bio, io, region);
 308
 309                 /*
 310                  * Try and add as many pages as possible.
 311                  */
 312                 while (remaining) {
 313                         dp->get_page(dp, &page, &len, &offset);
 314                         len = min(len, to_bytes(remaining));
 315                         if (!bio_add_page(bio, page, len, offset))
 316                                 break;
 317
 318                         offset = 0;
 319                         remaining -= to_sector(len);
 320                         dp->next_page(dp);
 321                 }
 322
 323                 atomic_inc(&io->count);
 324                 submit_bio(rw, bio);
 325         } while (remaining);
 326 }
 327
 328 static void dispatch_io(int rw, unsigned int num_regions,
 329                         struct dm_io_region *where, struct dpages *dp,
 330                         struct io *io, int sync)
 331 {
 332         int i;
 333         struct dpages old_pages = *dp;
 334
 335         BUG_ON(num_regions > DM_IO_MAX_REGIONS);
 336
 337         if (sync)
 338                 rw |= REQ_SYNC;
 339
 340         /*
 341          * For multiple regions we need to be careful to rewind
 342          * the dp object for each call to do_region.
 343          */
 344         for (i = 0; i < num_regions; i++) {
 345                 *dp = old_pages;
 346                 if (where[i].count || (rw & REQ_FLUSH))
 347                         do_region(rw, i, where + i, dp, io);
 348         }
 349
 350         /*
 351          * Drop the extra reference that we were holding to avoid
 352          * the io being completed too early.
 353          */
 354         dec_count(io, 0, 0);
 355 }
 356
 357 static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 358                    struct dm_io_region *where, int rw, struct dpages *dp,
 359                    unsigned long *error_bits)
 360 {
 361         /*
 362          * gcc <= 4.3 can't do the alignment for stack variables, so we must
 363          * align it on our own.
 364          * volatile prevents the optimizer from removing or reusing
 365          * "io_" field from the stack frame (allowed in ANSI C).
 366          */
 367         volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
 368         struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
 369
 370         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 371                 WARN_ON(1);
 372                 return -EIO;
 373         }
 374
 375         io->error_bits = 0;
 376         atomic_set(&io->count, 1); /* see dispatch_io() */
 377         io->sleeper = current;
 378         io->client = client;
 379
 380         dispatch_io(rw, num_regions, where, dp, io, 1);
 381
 382         while (1) {
 383                 set_current_state(TASK_UNINTERRUPTIBLE);
 384
 385                 if (!atomic_read(&io->count))
 386                         break;
 387
 388                 io_schedule();
 389         }
 390         set_current_state(TASK_RUNNING);
 391
 392         if (error_bits)
 393                 *error_bits = io->error_bits;
 394
 395         return io->error_bits ? -EIO : 0;
 396 }
 397
 398 static int async_io(struct dm_io_client *client, unsigned int num_regions,
 399                     struct dm_io_region *where, int rw, struct dpages *dp,
 400                     io_notify_fn fn, void *context)
 401 {
 402         struct io *io;
 403
 404         if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 405                 WARN_ON(1);
 406                 fn(1, context);
 407                 return -EIO;
 408         }
 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->sleeper = NULL;
 414         io->client = client;
 415         io->callback = fn;
 416         io->context = context;
 417
 418         dispatch_io(rw, num_regions, where, dp, io, 0);
 419         return 0;
 420 }
 421
 422 static int dp_init(struct dm_io_request *io_req, struct dpages *dp)
 423 {
 424         /* Set up dpages based on memory type */
 425         switch (io_req->mem.type) {
 426         case DM_IO_PAGE_LIST:
 427                 list_dp_init(dp, io_req->mem.ptr.pl, io_req->mem.offset);
 428                 break;
 429
 430         case DM_IO_BVEC:
 431                 bvec_dp_init(dp, io_req->mem.ptr.bvec);
 432                 break;
 433
 434         case DM_IO_VMA:
 435                 vm_dp_init(dp, io_req->mem.ptr.vma);
 436                 break;
 437
 438         case DM_IO_KMEM:
 439                 km_dp_init(dp, io_req->mem.ptr.addr);
 440                 break;
 441
 442         default:
 443                 return -EINVAL;
 444         }
 445
 446         return 0;
 447 }
 448
 449 /*
 450  * New collapsed (a)synchronous interface.
 451  *
 452  * If the IO is asynchronous (i.e. it has notify.fn), you must either unplug
 453  * the queue with blk_unplug() some time later or set REQ_SYNC in
 454 io_req->bi_rw. If you fail to do one of these, the IO will be submitted to
 455  * the disk after q->unplug_delay, which defaults to 3ms in blk-settings.c.
 456  */
 457 int dm_io(struct dm_io_request *io_req, unsigned num_regions,
 458           struct dm_io_region *where, unsigned long *sync_error_bits)
 459 {
 460         int r;
 461         struct dpages dp;
 462
 463         r = dp_init(io_req, &dp);
 464         if (r)
 465                 return r;
 466
 467         if (!io_req->notify.fn)
 468                 return sync_io(io_req->client, num_regions, where,
 469                                io_req->bi_rw, &dp, sync_error_bits);
 470
 471         return async_io(io_req->client, num_regions, where, io_req->bi_rw,
 472                         &dp, io_req->notify.fn, io_req->notify.context);
 473 }
 474 EXPORT_SYMBOL(dm_io);
 475
 476 int __init dm_io_init(void)
 477 {
 478         _dm_io_cache = KMEM_CACHE(io, 0);
 479         if (!_dm_io_cache)
 480                 return -ENOMEM;
 481
 482         return 0;
 483 }
 484
 485 void dm_io_exit(void)
 486 {
 487         kmem_cache_destroy(_dm_io_cache);
 488         _dm_io_cache = NULL;
 489 }