drivers/block/brd.c

   1 /*
   2  * Ram backed block device driver.
   3  *
   4  * Copyright (C) 2007 Nick Piggin
   5  * Copyright (C) 2007 Novell Inc.
   6  *
   7  * Parts derived from drivers/block/rd.c, and drivers/block/loop.c, copyright
   8  * of their respective owners.
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/moduleparam.h>
  14 #include <linux/major.h>
  15 #include <linux/blkdev.h>
  16 #include <linux/bio.h>
  17 #include <linux/highmem.h>
  18 #include <linux/smp_lock.h>
  19 #include <linux/radix-tree.h>
  20 #include <linux/buffer_head.h> /* invalidate_bh_lrus() */
  21 #include <linux/slab.h>
  22
  23 #include <asm/uaccess.h>
  24
  25 #define SECTOR_SHIFT            9
  26 #define PAGE_SECTORS_SHIFT      (PAGE_SHIFT - SECTOR_SHIFT)
  27 #define PAGE_SECTORS            (1 << PAGE_SECTORS_SHIFT)
  28
  29 /*
  30  * Each block ramdisk device has a radix_tree brd_pages of pages that stores
  31  * the pages containing the block device's contents. A brd page's ->index is
  32  * its offset in PAGE_SIZE units. This is similar to, but in no way connected
  33  * with, the kernel's pagecache or buffer cache (which sit above our block
  34  * device).
  35  */
  36 struct brd_device {
  37         int             brd_number;
  38         int             brd_refcnt;
  39         loff_t          brd_offset;
  40         loff_t          brd_sizelimit;
  41         unsigned        brd_blocksize;
  42
  43         struct request_queue    *brd_queue;
  44         struct gendisk          *brd_disk;
  45         struct list_head        brd_list;
  46
  47         /*
  48          * Backing store of pages and lock to protect it. This is the contents
  49          * of the block device.
  50          */
  51         spinlock_t              brd_lock;
  52         struct radix_tree_root  brd_pages;
  53 };
  54
  55 /*
  56  * Look up and return a brd's page for a given sector.
  57  */
  58 static struct page *brd_lookup_page(struct brd_device *brd, sector_t sector)
  59 {
  60         pgoff_t idx;
  61         struct page *page;
  62
  63         /*
  64          * The page lifetime is protected by the fact that we have opened the
  65          * device node -- brd pages will never be deleted under us, so we
  66          * don't need any further locking or refcounting.
  67          *
  68          * This is strictly true for the radix-tree nodes as well (ie. we
  69          * don't actually need the rcu_read_lock()), however that is not a
  70          * documented feature of the radix-tree API so it is better to be
  71          * safe here (we don't have total exclusion from radix tree updates
  72          * here, only deletes).
  73          */
  74         rcu_read_lock();
  75         idx = sector >> PAGE_SECTORS_SHIFT; /* sector to page index */
  76         page = radix_tree_lookup(&brd->brd_pages, idx);
  77         rcu_read_unlock();
  78
  79         BUG_ON(page && page->index != idx);
  80
  81         return page;
  82 }
  83
  84 /*
  85  * Look up and return a brd's page for a given sector.
  86  * If one does not exist, allocate an empty page, and insert that. Then
  87  * return it.
  88  */
  89 static struct page *brd_insert_page(struct brd_device *brd, sector_t sector)
  90 {
  91         pgoff_t idx;
  92         struct page *page;
  93         gfp_t gfp_flags;
  94
  95         page = brd_lookup_page(brd, sector);
  96         if (page)
  97                 return page;
  98
  99         /*
 100          * Must use NOIO because we don't want to recurse back into the
 101          * block or filesystem layers from page reclaim.
 102          *
 103          * Cannot support XIP and highmem, because our ->direct_access
 104          * routine for XIP must return memory that is always addressable.
 105          * If XIP was reworked to use pfns and kmap throughout, this
 106          * restriction might be able to be lifted.
 107          */
 108         gfp_flags = GFP_NOIO | __GFP_ZERO;
 109 #ifndef CONFIG_BLK_DEV_XIP
 110         gfp_flags |= __GFP_HIGHMEM;
 111 #endif
 112         page = alloc_page(gfp_flags);
 113         if (!page)
 114                 return NULL;
 115
 116         if (radix_tree_preload(GFP_NOIO)) {
 117                 __free_page(page);
 118                 return NULL;
 119         }
 120
 121         spin_lock(&brd->brd_lock);
 122         idx = sector >> PAGE_SECTORS_SHIFT;
 123         if (radix_tree_insert(&brd->brd_pages, idx, page)) {
 124                 __free_page(page);
 125                 page = radix_tree_lookup(&brd->brd_pages, idx);
 126                 BUG_ON(!page);
 127                 BUG_ON(page->index != idx);
 128         } else
 129                 page->index = idx;
 130         spin_unlock(&brd->brd_lock);
 131
 132         radix_tree_preload_end();
 133
 134         return page;
 135 }
 136
 137 static void brd_free_page(struct brd_device *brd, sector_t sector)
 138 {
 139         struct page *page;
 140         pgoff_t idx;
 141
 142         spin_lock(&brd->brd_lock);
 143         idx = sector >> PAGE_SECTORS_SHIFT;
 144         page = radix_tree_delete(&brd->brd_pages, idx);
 145         spin_unlock(&brd->brd_lock);
 146         if (page)
 147                 __free_page(page);
 148 }
 149
 150 static void brd_zero_page(struct brd_device *brd, sector_t sector)
 151 {
 152         struct page *page;
 153
 154         page = brd_lookup_page(brd, sector);
 155         if (page)
 156                 clear_highpage(page);
 157 }
 158
 159 /*
 160  * Free all backing store pages and radix tree. This must only be called when
 161  * there are no other users of the device.
 162  */
 163 #define FREE_BATCH 16
 164 static void brd_free_pages(struct brd_device *brd)
 165 {
 166         unsigned long pos = 0;
 167         struct page *pages[FREE_BATCH];
 168         int nr_pages;
 169
 170         do {
 171                 int i;
 172
 173                 nr_pages = radix_tree_gang_lookup(&brd->brd_pages,
 174                                 (void **)pages, pos, FREE_BATCH);
 175
 176                 for (i = 0; i < nr_pages; i++) {
 177                         void *ret;
 178
 179                         BUG_ON(pages[i]->index < pos);
 180                         pos = pages[i]->index;
 181                         ret = radix_tree_delete(&brd->brd_pages, pos);
 182                         BUG_ON(!ret || ret != pages[i]);
 183                         __free_page(pages[i]);
 184                 }
 185
 186                 pos++;
 187
 188                 /*
 189                  * This assumes radix_tree_gang_lookup always returns as
 190                  * many pages as possible. If the radix-tree code changes,
 191                  * so will this have to.
 192                  */
 193         } while (nr_pages == FREE_BATCH);
 194 }
 195
 196 /*
 197  * copy_to_brd_setup must be called before copy_to_brd. It may sleep.
 198  */
 199 static int copy_to_brd_setup(struct brd_device *brd, sector_t sector, size_t n)
 200 {
 201         unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
 202         size_t copy;
 203
 204         copy = min_t(size_t, n, PAGE_SIZE - offset);
 205         if (!brd_insert_page(brd, sector))
 206                 return -ENOMEM;
 207         if (copy < n) {
 208                 sector += copy >> SECTOR_SHIFT;
 209                 if (!brd_insert_page(brd, sector))
 210                         return -ENOMEM;
 211         }
 212         return 0;
 213 }
 214
 215 static void discard_from_brd(struct brd_device *brd,
 216                         sector_t sector, size_t n)
 217 {
 218         while (n >= PAGE_SIZE) {
 219                 /*
 220                  * Don't want to actually discard pages here because
 221                  * re-allocating the pages can result in writeback
 222                  * deadlocks under heavy load.
 223                  */
 224                 if (0)
 225                         brd_free_page(brd, sector);
 226                 else
 227                         brd_zero_page(brd, sector);
 228                 sector += PAGE_SIZE >> SECTOR_SHIFT;
 229                 n -= PAGE_SIZE;
 230         }
 231 }
 232
 233 /*
 234  * Copy n bytes from src to the brd starting at sector. Does not sleep.
 235  */
 236 static void copy_to_brd(struct brd_device *brd, const void *src,
 237                         sector_t sector, size_t n)
 238 {
 239         struct page *page;
 240         void *dst;
 241         unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
 242         size_t copy;
 243
 244         copy = min_t(size_t, n, PAGE_SIZE - offset);
 245         page = brd_lookup_page(brd, sector);
 246         BUG_ON(!page);
 247
 248         dst = kmap_atomic(page, KM_USER1);
 249         memcpy(dst + offset, src, copy);
 250         kunmap_atomic(dst, KM_USER1);
 251
 252         if (copy < n) {
 253                 src += copy;
 254                 sector += copy >> SECTOR_SHIFT;
 255                 copy = n - copy;
 256                 page = brd_lookup_page(brd, sector);
 257                 BUG_ON(!page);
 258
 259                 dst = kmap_atomic(page, KM_USER1);
 260                 memcpy(dst, src, copy);
 261                 kunmap_atomic(dst, KM_USER1);
 262         }
 263 }
 264
 265 /*
 266  * Copy n bytes to dst from the brd starting at sector. Does not sleep.
 267  */
 268 static void copy_from_brd(void *dst, struct brd_device *brd,
 269                         sector_t sector, size_t n)
 270 {
 271         struct page *page;
 272         void *src;
 273         unsigned int offset = (sector & (PAGE_SECTORS-1)) << SECTOR_SHIFT;
 274         size_t copy;
 275
 276         copy = min_t(size_t, n, PAGE_SIZE - offset);
 277         page = brd_lookup_page(brd, sector);
 278         if (page) {
 279                 src = kmap_atomic(page, KM_USER1);
 280                 memcpy(dst, src + offset, copy);
 281                 kunmap_atomic(src, KM_USER1);
 282         } else
 283                 memset(dst, 0, copy);
 284
 285         if (copy < n) {
 286                 dst += copy;
 287                 sector += copy >> SECTOR_SHIFT;
 288                 copy = n - copy;
 289                 page = brd_lookup_page(brd, sector);
 290                 if (page) {
 291                         src = kmap_atomic(page, KM_USER1);
 292                         memcpy(dst, src, copy);
 293                         kunmap_atomic(src, KM_USER1);
 294                 } else
 295                         memset(dst, 0, copy);
 296         }
 297 }
 298
 299 /*
 300  * Process a single bvec of a bio.
 301  */
 302 static int brd_do_bvec(struct brd_device *brd, struct page *page,
 303                         unsigned int len, unsigned int off, int rw,
 304                         sector_t sector)
 305 {
 306         void *mem;
 307         int err = 0;
 308
 309         if (rw != READ) {
 310                 err = copy_to_brd_setup(brd, sector, len);
 311                 if (err)
 312                         goto out;
 313         }
 314
 315         mem = kmap_atomic(page, KM_USER0);
 316         if (rw == READ) {
 317                 copy_from_brd(mem + off, brd, sector, len);
 318                 flush_dcache_page(page);
 319         } else {
 320                 flush_dcache_page(page);
 321                 copy_to_brd(brd, mem + off, sector, len);
 322         }
 323         kunmap_atomic(mem, KM_USER0);
 324
 325 out:
 326         return err;
 327 }
 328
 329 static int brd_make_request(struct request_queue *q, struct bio *bio)
 330 {
 331         struct block_device *bdev = bio->bi_bdev;
 332         struct brd_device *brd = bdev->bd_disk->private_data;
 333         int rw;
 334         struct bio_vec *bvec;
 335         sector_t sector;
 336         int i;
 337         int err = -EIO;
 338
 339         sector = bio->bi_sector;
 340         if (sector + (bio->bi_size >> SECTOR_SHIFT) >
 341                                                 get_capacity(bdev->bd_disk))
 342                 goto out;
 343
 344         if (unlikely(bio->bi_rw & REQ_DISCARD)) {
 345                 err = 0;
 346                 discard_from_brd(brd, sector, bio->bi_size);
 347                 goto out;
 348         }
 349
 350         rw = bio_rw(bio);
 351         if (rw == READA)
 352                 rw = READ;
 353
 354         bio_for_each_segment(bvec, bio, i) {
 355                 unsigned int len = bvec->bv_len;
 356                 err = brd_do_bvec(brd, bvec->bv_page, len,
 357                                         bvec->bv_offset, rw, sector);
 358                 if (err)
 359                         break;
 360                 sector += len >> SECTOR_SHIFT;
 361         }
 362
 363 out:
 364         bio_endio(bio, err);
 365
 366         return 0;
 367 }
 368
 369 #ifdef CONFIG_BLK_DEV_XIP
 370 static int brd_direct_access(struct block_device *bdev, sector_t sector,
 371                         void **kaddr, unsigned long *pfn)
 372 {
 373         struct brd_device *brd = bdev->bd_disk->private_data;
 374         struct page *page;
 375
 376         if (!brd)
 377                 return -ENODEV;
 378         if (sector & (PAGE_SECTORS-1))
 379                 return -EINVAL;
 380         if (sector + PAGE_SECTORS > get_capacity(bdev->bd_disk))
 381                 return -ERANGE;
 382         page = brd_insert_page(brd, sector);
 383         if (!page)
 384                 return -ENOMEM;
 385         *kaddr = page_address(page);
 386         *pfn = page_to_pfn(page);
 387
 388         return 0;
 389 }
 390 #endif
 391
 392 static int brd_ioctl(struct block_device *bdev, fmode_t mode,
 393                         unsigned int cmd, unsigned long arg)
 394 {
 395         int error;
 396         struct brd_device *brd = bdev->bd_disk->private_data;
 397
 398         if (cmd != BLKFLSBUF)
 399                 return -ENOTTY;
 400
 401         /*
 402          * ram device BLKFLSBUF has special semantics, we want to actually
 403          * release and destroy the ramdisk data.
 404          */
 405         lock_kernel();
 406         mutex_lock(&bdev->bd_mutex);
 407         error = -EBUSY;
 408         if (bdev->bd_openers <= 1) {
 409                 /*
 410                  * Invalidate the cache first, so it isn't written
 411                  * back to the device.
 412                  *
 413                  * Another thread might instantiate more buffercache here,
 414                  * but there is not much we can do to close that race.
 415                  */
 416                 invalidate_bh_lrus();
 417                 truncate_inode_pages(bdev->bd_inode->i_mapping, 0);
 418                 brd_free_pages(brd);
 419                 error = 0;
 420         }
 421         mutex_unlock(&bdev->bd_mutex);
 422         unlock_kernel();
 423
 424         return error;
 425 }
 426
 427 static const struct block_device_operations brd_fops = {
 428         .owner =                THIS_MODULE,
 429         .ioctl =                brd_ioctl,
 430 #ifdef CONFIG_BLK_DEV_XIP
 431         .direct_access =        brd_direct_access,
 432 #endif
 433 };
 434
 435 /*
 436  * And now the modules code and kernel interface.
 437  */
 438 static int rd_nr;
 439 int rd_size = CONFIG_BLK_DEV_RAM_SIZE;
 440 static int max_part;
 441 static int part_shift;
 442 module_param(rd_nr, int, 0);
 443 MODULE_PARM_DESC(rd_nr, "Maximum number of brd devices");
 444 module_param(rd_size, int, 0);
 445 MODULE_PARM_DESC(rd_size, "Size of each RAM disk in kbytes.");
 446 module_param(max_part, int, 0);
 447 MODULE_PARM_DESC(max_part, "Maximum number of partitions per RAM disk");
 448 MODULE_LICENSE("GPL");
 449 MODULE_ALIAS_BLOCKDEV_MAJOR(RAMDISK_MAJOR);
 450 MODULE_ALIAS("rd");
 451
 452 #ifndef MODULE
 453 /* Legacy boot options - nonmodular */
 454 static int __init ramdisk_size(char *str)
 455 {
 456         rd_size = simple_strtol(str, NULL, 0);
 457         return 1;
 458 }
 459 __setup("ramdisk_size=", ramdisk_size);
 460 #endif
 461
 462 /*
 463  * The device scheme is derived from loop.c. Keep them in synch where possible
 464  * (should share code eventually).
 465  */
 466 static LIST_HEAD(brd_devices);
 467 static DEFINE_MUTEX(brd_devices_mutex);
 468
 469 static struct brd_device *brd_alloc(int i)
 470 {
 471         struct brd_device *brd;
 472         struct gendisk *disk;
 473
 474         brd = kzalloc(sizeof(*brd), GFP_KERNEL);
 475         if (!brd)
 476                 goto out;
 477         brd->brd_number         = i;
 478         spin_lock_init(&brd->brd_lock);
 479         INIT_RADIX_TREE(&brd->brd_pages, GFP_ATOMIC);
 480
 481         brd->brd_queue = blk_alloc_queue(GFP_KERNEL);
 482         if (!brd->brd_queue)
 483                 goto out_free_dev;
 484         blk_queue_make_request(brd->brd_queue, brd_make_request);
 485         blk_queue_ordered(brd->brd_queue, QUEUE_ORDERED_TAG);
 486         blk_queue_max_hw_sectors(brd->brd_queue, 1024);
 487         blk_queue_bounce_limit(brd->brd_queue, BLK_BOUNCE_ANY);
 488
 489         brd->brd_queue->limits.discard_granularity = PAGE_SIZE;
 490         brd->brd_queue->limits.max_discard_sectors = UINT_MAX;
 491         brd->brd_queue->limits.discard_zeroes_data = 1;
 492         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, brd->brd_queue);
 493
 494         disk = brd->brd_disk = alloc_disk(1 << part_shift);
 495         if (!disk)
 496                 goto out_free_queue;
 497         disk->major             = RAMDISK_MAJOR;
 498         disk->first_minor       = i << part_shift;
 499         disk->fops              = &brd_fops;
 500         disk->private_data      = brd;
 501         disk->queue             = brd->brd_queue;
 502         disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
 503         sprintf(disk->disk_name, "ram%d", i);
 504         set_capacity(disk, rd_size * 2);
 505
 506         return brd;
 507
 508 out_free_queue:
 509         blk_cleanup_queue(brd->brd_queue);
 510 out_free_dev:
 511         kfree(brd);
 512 out:
 513         return NULL;
 514 }
 515
 516 static void brd_free(struct brd_device *brd)
 517 {
 518         put_disk(brd->brd_disk);
 519         blk_cleanup_queue(brd->brd_queue);
 520         brd_free_pages(brd);
 521         kfree(brd);
 522 }
 523
 524 static struct brd_device *brd_init_one(int i)
 525 {
 526         struct brd_device *brd;
 527
 528         list_for_each_entry(brd, &brd_devices, brd_list) {
 529                 if (brd->brd_number == i)
 530                         goto out;
 531         }
 532
 533         brd = brd_alloc(i);
 534         if (brd) {
 535                 add_disk(brd->brd_disk);
 536                 list_add_tail(&brd->brd_list, &brd_devices);
 537         }
 538 out:
 539         return brd;
 540 }
 541
 542 static void brd_del_one(struct brd_device *brd)
 543 {
 544         list_del(&brd->brd_list);
 545         del_gendisk(brd->brd_disk);
 546         brd_free(brd);
 547 }
 548
 549 static struct kobject *brd_probe(dev_t dev, int *part, void *data)
 550 {
 551         struct brd_device *brd;
 552         struct kobject *kobj;
 553
 554         mutex_lock(&brd_devices_mutex);
 555         brd = brd_init_one(dev & MINORMASK);
 556         kobj = brd ? get_disk(brd->brd_disk) : ERR_PTR(-ENOMEM);
 557         mutex_unlock(&brd_devices_mutex);
 558
 559         *part = 0;
 560         return kobj;
 561 }
 562
 563 static int __init brd_init(void)
 564 {
 565         int i, nr;
 566         unsigned long range;
 567         struct brd_device *brd, *next;
 568
 569         /*
 570          * brd module now has a feature to instantiate underlying device
 571          * structure on-demand, provided that there is an access dev node.
 572          * However, this will not work well with user space tool that doesn't
 573          * know about such "feature".  In order to not break any existing
 574          * tool, we do the following:
 575          *
 576          * (1) if rd_nr is specified, create that many upfront, and this
 577          *     also becomes a hard limit.
 578          * (2) if rd_nr is not specified, create 1 rd device on module
 579          *     load, user can further extend brd device by create dev node
 580          *     themselves and have kernel automatically instantiate actual
 581          *     device on-demand.
 582          */
 583
 584         part_shift = 0;
 585         if (max_part > 0)
 586                 part_shift = fls(max_part);
 587
 588         if (rd_nr > 1UL << (MINORBITS - part_shift))
 589                 return -EINVAL;
 590
 591         if (rd_nr) {
 592                 nr = rd_nr;
 593                 range = rd_nr;
 594         } else {
 595                 nr = CONFIG_BLK_DEV_RAM_COUNT;
 596                 range = 1UL << (MINORBITS - part_shift);
 597         }
 598
 599         if (register_blkdev(RAMDISK_MAJOR, "ramdisk"))
 600                 return -EIO;
 601
 602         for (i = 0; i < nr; i++) {
 603                 brd = brd_alloc(i);
 604                 if (!brd)
 605                         goto out_free;
 606                 list_add_tail(&brd->brd_list, &brd_devices);
 607         }
 608
 609         /* point of no return */
 610
 611         list_for_each_entry(brd, &brd_devices, brd_list)
 612                 add_disk(brd->brd_disk);
 613
 614         blk_register_region(MKDEV(RAMDISK_MAJOR, 0), range,
 615                                   THIS_MODULE, brd_probe, NULL, NULL);
 616
 617         printk(KERN_INFO "brd: module loaded\n");
 618         return 0;
 619
 620 out_free:
 621         list_for_each_entry_safe(brd, next, &brd_devices, brd_list) {
 622                 list_del(&brd->brd_list);
 623                 brd_free(brd);
 624         }
 625         unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
 626
 627         return -ENOMEM;
 628 }
 629
 630 static void __exit brd_exit(void)
 631 {
 632         unsigned long range;
 633         struct brd_device *brd, *next;
 634
 635         range = rd_nr ? rd_nr :  1UL << (MINORBITS - part_shift);
 636
 637         list_for_each_entry_safe(brd, next, &brd_devices, brd_list)
 638                 brd_del_one(brd);
 639
 640         blk_unregister_region(MKDEV(RAMDISK_MAJOR, 0), range);
 641         unregister_blkdev(RAMDISK_MAJOR, "ramdisk");
 642 }
 643
 644 module_init(brd_init);
 645 module_exit(brd_exit);
 646