fs/dax.c

   1 /*
   2  * fs/dax.c - Direct Access filesystem code
   3  * Copyright (c) 2013-2014 Intel Corporation
   4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  */
  16
  17 #include <linux/atomic.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/buffer_head.h>
  20 #include <linux/dax.h>
  21 #include <linux/fs.h>
  22 #include <linux/genhd.h>
  23 #include <linux/highmem.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm.h>
  26 #include <linux/mutex.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/pmem.h>
  29 #include <linux/sched.h>
  30 #include <linux/uio.h>
  31 #include <linux/vmstat.h>
  32 #include <linux/pfn_t.h>
  33 #include <linux/sizes.h>
  34
  35 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  36 {
  37         struct request_queue *q = bdev->bd_queue;
  38         long rc = -EIO;
  39
  40         dax->addr = (void __pmem *) ERR_PTR(-EIO);
  41         if (blk_queue_enter(q, true) != 0)
  42                 return rc;
  43
  44         rc = bdev_direct_access(bdev, dax);
  45         if (rc < 0) {
  46                 dax->addr = (void __pmem *) ERR_PTR(rc);
  47                 blk_queue_exit(q);
  48                 return rc;
  49         }
  50         return rc;
  51 }
  52
  53 static void dax_unmap_atomic(struct block_device *bdev,
  54                 const struct blk_dax_ctl *dax)
  55 {
  56         if (IS_ERR(dax->addr))
  57                 return;
  58         blk_queue_exit(bdev->bd_queue);
  59 }
  60
  61 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
  62 {
  63         struct page *page = alloc_pages(GFP_KERNEL, 0);
  64         struct blk_dax_ctl dax = {
  65                 .size = PAGE_SIZE,
  66                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
  67         };
  68         long rc;
  69
  70         if (!page)
  71                 return ERR_PTR(-ENOMEM);
  72
  73         rc = dax_map_atomic(bdev, &dax);
  74         if (rc < 0)
  75                 return ERR_PTR(rc);
  76         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
  77         dax_unmap_atomic(bdev, &dax);
  78         return page;
  79 }
  80
  81 /*
  82  * dax_clear_sectors() is called from within transaction context from XFS,
  83  * and hence this means the stack from this point must follow GFP_NOFS
  84  * semantics for all operations.
  85  */
  86 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
  87 {
  88         struct blk_dax_ctl dax = {
  89                 .sector = _sector,
  90                 .size = _size,
  91         };
  92
  93         might_sleep();
  94         do {
  95                 long count, sz;
  96
  97                 count = dax_map_atomic(bdev, &dax);
  98                 if (count < 0)
  99                         return count;
 100                 sz = min_t(long, count, SZ_128K);
 101                 clear_pmem(dax.addr, sz);
 102                 dax.size -= sz;
 103                 dax.sector += sz / 512;
 104                 dax_unmap_atomic(bdev, &dax);
 105                 cond_resched();
 106         } while (dax.size);
 107
 108         wmb_pmem();
 109         return 0;
 110 }
 111 EXPORT_SYMBOL_GPL(dax_clear_sectors);
 112
 113 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
 114 static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
 115                 loff_t pos, loff_t end)
 116 {
 117         loff_t final = end - pos + first; /* The final byte of the buffer */
 118
 119         if (first > 0)
 120                 clear_pmem(addr, first);
 121         if (final < size)
 122                 clear_pmem(addr + final, size - final);
 123 }
 124
 125 static bool buffer_written(struct buffer_head *bh)
 126 {
 127         return buffer_mapped(bh) && !buffer_unwritten(bh);
 128 }
 129
 130 /*
 131  * When ext4 encounters a hole, it returns without modifying the buffer_head
 132  * which means that we can't trust b_size.  To cope with this, we set b_state
 133  * to 0 before calling get_block and, if any bit is set, we know we can trust
 134  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 135  * and would save us time calling get_block repeatedly.
 136  */
 137 static bool buffer_size_valid(struct buffer_head *bh)
 138 {
 139         return bh->b_state != 0;
 140 }
 141
 142
 143 static sector_t to_sector(const struct buffer_head *bh,
 144                 const struct inode *inode)
 145 {
 146         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 147
 148         return sector;
 149 }
 150
 151 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 152                       loff_t start, loff_t end, get_block_t get_block,
 153                       struct buffer_head *bh)
 154 {
 155         loff_t pos = start, max = start, bh_max = start;
 156         bool hole = false, need_wmb = false;
 157         struct block_device *bdev = NULL;
 158         int rw = iov_iter_rw(iter), rc;
 159         long map_len = 0;
 160         struct blk_dax_ctl dax = {
 161                 .addr = (void __pmem *) ERR_PTR(-EIO),
 162         };
 163
 164         if (rw == READ)
 165                 end = min(end, i_size_read(inode));
 166
 167         while (pos < end) {
 168                 size_t len;
 169                 if (pos == max) {
 170                         unsigned blkbits = inode->i_blkbits;
 171                         long page = pos >> PAGE_SHIFT;
 172                         sector_t block = page << (PAGE_SHIFT - blkbits);
 173                         unsigned first = pos - (block << blkbits);
 174                         long size;
 175
 176                         if (pos == bh_max) {
 177                                 bh->b_size = PAGE_ALIGN(end - pos);
 178                                 bh->b_state = 0;
 179                                 rc = get_block(inode, block, bh, rw == WRITE);
 180                                 if (rc)
 181                                         break;
 182                                 if (!buffer_size_valid(bh))
 183                                         bh->b_size = 1 << blkbits;
 184                                 bh_max = pos - first + bh->b_size;
 185                                 bdev = bh->b_bdev;
 186                         } else {
 187                                 unsigned done = bh->b_size -
 188                                                 (bh_max - (pos - first));
 189                                 bh->b_blocknr += done >> blkbits;
 190                                 bh->b_size -= done;
 191                         }
 192
 193                         hole = rw == READ && !buffer_written(bh);
 194                         if (hole) {
 195                                 size = bh->b_size - first;
 196                         } else {
 197                                 dax_unmap_atomic(bdev, &dax);
 198                                 dax.sector = to_sector(bh, inode);
 199                                 dax.size = bh->b_size;
 200                                 map_len = dax_map_atomic(bdev, &dax);
 201                                 if (map_len < 0) {
 202                                         rc = map_len;
 203                                         break;
 204                                 }
 205                                 if (buffer_unwritten(bh) || buffer_new(bh)) {
 206                                         dax_new_buf(dax.addr, map_len, first,
 207                                                         pos, end);
 208                                         need_wmb = true;
 209                                 }
 210                                 dax.addr += first;
 211                                 size = map_len - first;
 212                         }
 213                         max = min(pos + size, end);
 214                 }
 215
 216                 if (iov_iter_rw(iter) == WRITE) {
 217                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 218                         need_wmb = true;
 219                 } else if (!hole)
 220                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 221                                         iter);
 222                 else
 223                         len = iov_iter_zero(max - pos, iter);
 224
 225                 if (!len) {
 226                         rc = -EFAULT;
 227                         break;
 228                 }
 229
 230                 pos += len;
 231                 if (!IS_ERR(dax.addr))
 232                         dax.addr += len;
 233         }
 234
 235         if (need_wmb)
 236                 wmb_pmem();
 237         dax_unmap_atomic(bdev, &dax);
 238
 239         return (pos == start) ? rc : pos - start;
 240 }
 241
 242 /**
 243  * dax_do_io - Perform I/O to a DAX file
 244  * @iocb: The control block for this I/O
 245  * @inode: The file which the I/O is directed at
 246  * @iter: The addresses to do I/O from or to
 247  * @pos: The file offset where the I/O starts
 248  * @get_block: The filesystem method used to translate file offsets to blocks
 249  * @end_io: A filesystem callback for I/O completion
 250  * @flags: See below
 251  *
 252  * This function uses the same locking scheme as do_blockdev_direct_IO:
 253  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 254  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 255  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 256  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 257  * is in progress.
 258  */
 259 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 260                   struct iov_iter *iter, loff_t pos, get_block_t get_block,
 261                   dio_iodone_t end_io, int flags)
 262 {
 263         struct buffer_head bh;
 264         ssize_t retval = -EINVAL;
 265         loff_t end = pos + iov_iter_count(iter);
 266
 267         memset(&bh, 0, sizeof(bh));
 268         bh.b_bdev = inode->i_sb->s_bdev;
 269
 270         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
 271                 struct address_space *mapping = inode->i_mapping;
 272                 inode_lock(inode);
 273                 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
 274                 if (retval) {
 275                         inode_unlock(inode);
 276                         goto out;
 277                 }
 278         }
 279
 280         /* Protects against truncate */
 281         if (!(flags & DIO_SKIP_DIO_COUNT))
 282                 inode_dio_begin(inode);
 283
 284         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 285
 286         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 287                 inode_unlock(inode);
 288
 289         if ((retval > 0) && end_io)
 290                 end_io(iocb, pos, retval, bh.b_private);
 291
 292         if (!(flags & DIO_SKIP_DIO_COUNT))
 293                 inode_dio_end(inode);
 294  out:
 295         return retval;
 296 }
 297 EXPORT_SYMBOL_GPL(dax_do_io);
 298
 299 /*
 300  * The user has performed a load from a hole in the file.  Allocating
 301  * a new page in the file would cause excessive storage usage for
 302  * workloads with sparse files.  We allocate a page cache page instead.
 303  * We'll kick it out of the page cache if it's ever written to,
 304  * otherwise it will simply fall out of the page cache under memory
 305  * pressure without ever having been dirtied.
 306  */
 307 static int dax_load_hole(struct address_space *mapping, struct page *page,
 308                                                         struct vm_fault *vmf)
 309 {
 310         unsigned long size;
 311         struct inode *inode = mapping->host;
 312         if (!page)
 313                 page = find_or_create_page(mapping, vmf->pgoff,
 314                                                 GFP_KERNEL | __GFP_ZERO);
 315         if (!page)
 316                 return VM_FAULT_OOM;
 317         /* Recheck i_size under page lock to avoid truncate race */
 318         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 319         if (vmf->pgoff >= size) {
 320                 unlock_page(page);
 321                 page_cache_release(page);
 322                 return VM_FAULT_SIGBUS;
 323         }
 324
 325         vmf->page = page;
 326         return VM_FAULT_LOCKED;
 327 }
 328
 329 static int copy_user_bh(struct page *to, struct inode *inode,
 330                 struct buffer_head *bh, unsigned long vaddr)
 331 {
 332         struct blk_dax_ctl dax = {
 333                 .sector = to_sector(bh, inode),
 334                 .size = bh->b_size,
 335         };
 336         struct block_device *bdev = bh->b_bdev;
 337         void *vto;
 338
 339         if (dax_map_atomic(bdev, &dax) < 0)
 340                 return PTR_ERR(dax.addr);
 341         vto = kmap_atomic(to);
 342         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 343         kunmap_atomic(vto);
 344         dax_unmap_atomic(bdev, &dax);
 345         return 0;
 346 }
 347
 348 #define NO_SECTOR -1
 349 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_CACHE_SHIFT))
 350
 351 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 352                 sector_t sector, bool pmd_entry, bool dirty)
 353 {
 354         struct radix_tree_root *page_tree = &mapping->page_tree;
 355         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 356         int type, error = 0;
 357         void *entry;
 358
 359         WARN_ON_ONCE(pmd_entry && !dirty);
 360         if (dirty)
 361                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 362
 363         spin_lock_irq(&mapping->tree_lock);
 364
 365         entry = radix_tree_lookup(page_tree, pmd_index);
 366         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 367                 index = pmd_index;
 368                 goto dirty;
 369         }
 370
 371         entry = radix_tree_lookup(page_tree, index);
 372         if (entry) {
 373                 type = RADIX_DAX_TYPE(entry);
 374                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 375                                         type != RADIX_DAX_PMD)) {
 376                         error = -EIO;
 377                         goto unlock;
 378                 }
 379
 380                 if (!pmd_entry || type == RADIX_DAX_PMD)
 381                         goto dirty;
 382
 383                 /*
 384                  * We only insert dirty PMD entries into the radix tree.  This
 385                  * means we don't need to worry about removing a dirty PTE
 386                  * entry and inserting a clean PMD entry, thus reducing the
 387                  * range we would flush with a follow-up fsync/msync call.
 388                  */
 389                 radix_tree_delete(&mapping->page_tree, index);
 390                 mapping->nrexceptional--;
 391         }
 392
 393         if (sector == NO_SECTOR) {
 394                 /*
 395                  * This can happen during correct operation if our pfn_mkwrite
 396                  * fault raced against a hole punch operation.  If this
 397                  * happens the pte that was hole punched will have been
 398                  * unmapped and the radix tree entry will have been removed by
 399                  * the time we are called, but the call will still happen.  We
 400                  * will return all the way up to wp_pfn_shared(), where the
 401                  * pte_same() check will fail, eventually causing page fault
 402                  * to be retried by the CPU.
 403                  */
 404                 goto unlock;
 405         }
 406
 407         error = radix_tree_insert(page_tree, index,
 408                         RADIX_DAX_ENTRY(sector, pmd_entry));
 409         if (error)
 410                 goto unlock;
 411
 412         mapping->nrexceptional++;
 413  dirty:
 414         if (dirty)
 415                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 416  unlock:
 417         spin_unlock_irq(&mapping->tree_lock);
 418         return error;
 419 }
 420
 421 static int dax_writeback_one(struct block_device *bdev,
 422                 struct address_space *mapping, pgoff_t index, void *entry)
 423 {
 424         struct radix_tree_root *page_tree = &mapping->page_tree;
 425         int type = RADIX_DAX_TYPE(entry);
 426         struct radix_tree_node *node;
 427         struct blk_dax_ctl dax;
 428         void **slot;
 429         int ret = 0;
 430
 431         spin_lock_irq(&mapping->tree_lock);
 432         /*
 433          * Regular page slots are stabilized by the page lock even
 434          * without the tree itself locked.  These unlocked entries
 435          * need verification under the tree lock.
 436          */
 437         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 438                 goto unlock;
 439         if (*slot != entry)
 440                 goto unlock;
 441
 442         /* another fsync thread may have already written back this entry */
 443         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 444                 goto unlock;
 445
 446         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 447                 ret = -EIO;
 448                 goto unlock;
 449         }
 450
 451         dax.sector = RADIX_DAX_SECTOR(entry);
 452         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 453         spin_unlock_irq(&mapping->tree_lock);
 454
 455         /*
 456          * We cannot hold tree_lock while calling dax_map_atomic() because it
 457          * eventually calls cond_resched().
 458          */
 459         ret = dax_map_atomic(bdev, &dax);
 460         if (ret < 0)
 461                 return ret;
 462
 463         if (WARN_ON_ONCE(ret < dax.size)) {
 464                 ret = -EIO;
 465                 goto unmap;
 466         }
 467
 468         wb_cache_pmem(dax.addr, dax.size);
 469
 470         spin_lock_irq(&mapping->tree_lock);
 471         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 472         spin_unlock_irq(&mapping->tree_lock);
 473  unmap:
 474         dax_unmap_atomic(bdev, &dax);
 475         return ret;
 476
 477  unlock:
 478         spin_unlock_irq(&mapping->tree_lock);
 479         return ret;
 480 }
 481
 482 /*
 483  * Flush the mapping to the persistent domain within the byte range of [start,
 484  * end]. This is required by data integrity operations to ensure file data is
 485  * on persistent storage prior to completion of the operation.
 486  */
 487 int dax_writeback_mapping_range(struct address_space *mapping,
 488                 struct block_device *bdev, struct writeback_control *wbc)
 489 {
 490         struct inode *inode = mapping->host;
 491         pgoff_t start_index, end_index, pmd_index;
 492         pgoff_t indices[PAGEVEC_SIZE];
 493         struct pagevec pvec;
 494         bool done = false;
 495         int i, ret = 0;
 496         void *entry;
 497
 498         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 499                 return -EIO;
 500
 501         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 502                 return 0;
 503
 504         start_index = wbc->range_start >> PAGE_CACHE_SHIFT;
 505         end_index = wbc->range_end >> PAGE_CACHE_SHIFT;
 506         pmd_index = DAX_PMD_INDEX(start_index);
 507
 508         rcu_read_lock();
 509         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 510         rcu_read_unlock();
 511
 512         /* see if the start of our range is covered by a PMD entry */
 513         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 514                 start_index = pmd_index;
 515
 516         tag_pages_for_writeback(mapping, start_index, end_index);
 517
 518         pagevec_init(&pvec, 0);
 519         while (!done) {
 520                 pvec.nr = find_get_entries_tag(mapping, start_index,
 521                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 522                                 pvec.pages, indices);
 523
 524                 if (pvec.nr == 0)
 525                         break;
 526
 527                 for (i = 0; i < pvec.nr; i++) {
 528                         if (indices[i] > end_index) {
 529                                 done = true;
 530                                 break;
 531                         }
 532
 533                         ret = dax_writeback_one(bdev, mapping, indices[i],
 534                                         pvec.pages[i]);
 535                         if (ret < 0)
 536                                 return ret;
 537                 }
 538         }
 539         wmb_pmem();
 540         return 0;
 541 }
 542 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 543
 544 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 545                         struct vm_area_struct *vma, struct vm_fault *vmf)
 546 {
 547         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 548         struct address_space *mapping = inode->i_mapping;
 549         struct block_device *bdev = bh->b_bdev;
 550         struct blk_dax_ctl dax = {
 551                 .sector = to_sector(bh, inode),
 552                 .size = bh->b_size,
 553         };
 554         pgoff_t size;
 555         int error;
 556
 557         i_mmap_lock_read(mapping);
 558
 559         /*
 560          * Check truncate didn't happen while we were allocating a block.
 561          * If it did, this block may or may not be still allocated to the
 562          * file.  We can't tell the filesystem to free it because we can't
 563          * take i_mutex here.  In the worst case, the file still has blocks
 564          * allocated past the end of the file.
 565          */
 566         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 567         if (unlikely(vmf->pgoff >= size)) {
 568                 error = -EIO;
 569                 goto out;
 570         }
 571
 572         if (dax_map_atomic(bdev, &dax) < 0) {
 573                 error = PTR_ERR(dax.addr);
 574                 goto out;
 575         }
 576
 577         if (buffer_unwritten(bh) || buffer_new(bh)) {
 578                 clear_pmem(dax.addr, PAGE_SIZE);
 579                 wmb_pmem();
 580         }
 581         dax_unmap_atomic(bdev, &dax);
 582
 583         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 584                         vmf->flags & FAULT_FLAG_WRITE);
 585         if (error)
 586                 goto out;
 587
 588         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 589
 590  out:
 591         i_mmap_unlock_read(mapping);
 592
 593         return error;
 594 }
 595
 596 /**
 597  * __dax_fault - handle a page fault on a DAX file
 598  * @vma: The virtual memory area where the fault occurred
 599  * @vmf: The description of the fault
 600  * @get_block: The filesystem method used to translate file offsets to blocks
 601  * @complete_unwritten: The filesystem method used to convert unwritten blocks
 602  *      to written so the data written to them is exposed. This is required for
 603  *      required by write faults for filesystems that will return unwritten
 604  *      extent mappings from @get_block, but it is optional for reads as
 605  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
 606  *      not support unwritten extents, the it should pass NULL.
 607  *
 608  * When a page fault occurs, filesystems may call this helper in their
 609  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 610  * the necessary locking for the page fault to proceed successfully.
 611  */
 612 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 613                         get_block_t get_block, dax_iodone_t complete_unwritten)
 614 {
 615         struct file *file = vma->vm_file;
 616         struct address_space *mapping = file->f_mapping;
 617         struct inode *inode = mapping->host;
 618         struct page *page;
 619         struct buffer_head bh;
 620         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 621         unsigned blkbits = inode->i_blkbits;
 622         sector_t block;
 623         pgoff_t size;
 624         int error;
 625         int major = 0;
 626
 627         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 628         if (vmf->pgoff >= size)
 629                 return VM_FAULT_SIGBUS;
 630
 631         memset(&bh, 0, sizeof(bh));
 632         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 633         bh.b_bdev = inode->i_sb->s_bdev;
 634         bh.b_size = PAGE_SIZE;
 635
 636  repeat:
 637         page = find_get_page(mapping, vmf->pgoff);
 638         if (page) {
 639                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 640                         page_cache_release(page);
 641                         return VM_FAULT_RETRY;
 642                 }
 643                 if (unlikely(page->mapping != mapping)) {
 644                         unlock_page(page);
 645                         page_cache_release(page);
 646                         goto repeat;
 647                 }
 648                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 649                 if (unlikely(vmf->pgoff >= size)) {
 650                         /*
 651                          * We have a struct page covering a hole in the file
 652                          * from a read fault and we've raced with a truncate
 653                          */
 654                         error = -EIO;
 655                         goto unlock_page;
 656                 }
 657         }
 658
 659         error = get_block(inode, block, &bh, 0);
 660         if (!error && (bh.b_size < PAGE_SIZE))
 661                 error = -EIO;           /* fs corruption? */
 662         if (error)
 663                 goto unlock_page;
 664
 665         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 666                 if (vmf->flags & FAULT_FLAG_WRITE) {
 667                         error = get_block(inode, block, &bh, 1);
 668                         count_vm_event(PGMAJFAULT);
 669                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 670                         major = VM_FAULT_MAJOR;
 671                         if (!error && (bh.b_size < PAGE_SIZE))
 672                                 error = -EIO;
 673                         if (error)
 674                                 goto unlock_page;
 675                 } else {
 676                         return dax_load_hole(mapping, page, vmf);
 677                 }
 678         }
 679
 680         if (vmf->cow_page) {
 681                 struct page *new_page = vmf->cow_page;
 682                 if (buffer_written(&bh))
 683                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 684                 else
 685                         clear_user_highpage(new_page, vaddr);
 686                 if (error)
 687                         goto unlock_page;
 688                 vmf->page = page;
 689                 if (!page) {
 690                         i_mmap_lock_read(mapping);
 691                         /* Check we didn't race with truncate */
 692                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 693                                                                 PAGE_SHIFT;
 694                         if (vmf->pgoff >= size) {
 695                                 i_mmap_unlock_read(mapping);
 696                                 error = -EIO;
 697                                 goto out;
 698                         }
 699                 }
 700                 return VM_FAULT_LOCKED;
 701         }
 702
 703         /* Check we didn't race with a read fault installing a new page */
 704         if (!page && major)
 705                 page = find_lock_page(mapping, vmf->pgoff);
 706
 707         if (page) {
 708                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 709                                                         PAGE_CACHE_SIZE, 0);
 710                 delete_from_page_cache(page);
 711                 unlock_page(page);
 712                 page_cache_release(page);
 713                 page = NULL;
 714         }
 715
 716         /*
 717          * If we successfully insert the new mapping over an unwritten extent,
 718          * we need to ensure we convert the unwritten extent. If there is an
 719          * error inserting the mapping, the filesystem needs to leave it as
 720          * unwritten to prevent exposure of the stale underlying data to
 721          * userspace, but we still need to call the completion function so
 722          * the private resources on the mapping buffer can be released. We
 723          * indicate what the callback should do via the uptodate variable, same
 724          * as for normal BH based IO completions.
 725          */
 726         error = dax_insert_mapping(inode, &bh, vma, vmf);
 727         if (buffer_unwritten(&bh)) {
 728                 if (complete_unwritten)
 729                         complete_unwritten(&bh, !error);
 730                 else
 731                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 732         }
 733
 734  out:
 735         if (error == -ENOMEM)
 736                 return VM_FAULT_OOM | major;
 737         /* -EBUSY is fine, somebody else faulted on the same PTE */
 738         if ((error < 0) && (error != -EBUSY))
 739                 return VM_FAULT_SIGBUS | major;
 740         return VM_FAULT_NOPAGE | major;
 741
 742  unlock_page:
 743         if (page) {
 744                 unlock_page(page);
 745                 page_cache_release(page);
 746         }
 747         goto out;
 748 }
 749 EXPORT_SYMBOL(__dax_fault);
 750
 751 /**
 752  * dax_fault - handle a page fault on a DAX file
 753  * @vma: The virtual memory area where the fault occurred
 754  * @vmf: The description of the fault
 755  * @get_block: The filesystem method used to translate file offsets to blocks
 756  *
 757  * When a page fault occurs, filesystems may call this helper in their
 758  * fault handler for DAX files.
 759  */
 760 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 761               get_block_t get_block, dax_iodone_t complete_unwritten)
 762 {
 763         int result;
 764         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 765
 766         if (vmf->flags & FAULT_FLAG_WRITE) {
 767                 sb_start_pagefault(sb);
 768                 file_update_time(vma->vm_file);
 769         }
 770         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 771         if (vmf->flags & FAULT_FLAG_WRITE)
 772                 sb_end_pagefault(sb);
 773
 774         return result;
 775 }
 776 EXPORT_SYMBOL_GPL(dax_fault);
 777
 778 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 779 /*
 780  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 781  * more often than one might expect in the below function.
 782  */
 783 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 784
 785 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 786                 const char *reason, const char *fn)
 787 {
 788         if (bh) {
 789                 char bname[BDEVNAME_SIZE];
 790                 bdevname(bh->b_bdev, bname);
 791                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 792                         "length %zd fallback: %s\n", fn, current->comm,
 793                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 794                         bh->b_size, reason);
 795         } else {
 796                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 797                         current->comm, address, reason);
 798         }
 799 }
 800
 801 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 802
 803 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 804                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
 805                 dax_iodone_t complete_unwritten)
 806 {
 807         struct file *file = vma->vm_file;
 808         struct address_space *mapping = file->f_mapping;
 809         struct inode *inode = mapping->host;
 810         struct buffer_head bh;
 811         unsigned blkbits = inode->i_blkbits;
 812         unsigned long pmd_addr = address & PMD_MASK;
 813         bool write = flags & FAULT_FLAG_WRITE;
 814         struct block_device *bdev;
 815         pgoff_t size, pgoff;
 816         sector_t block;
 817         int error, result = 0;
 818         bool alloc = false;
 819
 820         /* dax pmd mappings require pfn_t_devmap() */
 821         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 822                 return VM_FAULT_FALLBACK;
 823
 824         /* Fall back to PTEs if we're going to COW */
 825         if (write && !(vma->vm_flags & VM_SHARED)) {
 826                 split_huge_pmd(vma, pmd, address);
 827                 dax_pmd_dbg(NULL, address, "cow write");
 828                 return VM_FAULT_FALLBACK;
 829         }
 830         /* If the PMD would extend outside the VMA */
 831         if (pmd_addr < vma->vm_start) {
 832                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 833                 return VM_FAULT_FALLBACK;
 834         }
 835         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 836                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 837                 return VM_FAULT_FALLBACK;
 838         }
 839
 840         pgoff = linear_page_index(vma, pmd_addr);
 841         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 842         if (pgoff >= size)
 843                 return VM_FAULT_SIGBUS;
 844         /* If the PMD would cover blocks out of the file */
 845         if ((pgoff | PG_PMD_COLOUR) >= size) {
 846                 dax_pmd_dbg(NULL, address,
 847                                 "offset + huge page size > file size");
 848                 return VM_FAULT_FALLBACK;
 849         }
 850
 851         memset(&bh, 0, sizeof(bh));
 852         bh.b_bdev = inode->i_sb->s_bdev;
 853         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 854
 855         bh.b_size = PMD_SIZE;
 856
 857         if (get_block(inode, block, &bh, 0) != 0)
 858                 return VM_FAULT_SIGBUS;
 859
 860         if (!buffer_mapped(&bh) && write) {
 861                 if (get_block(inode, block, &bh, 1) != 0)
 862                         return VM_FAULT_SIGBUS;
 863                 alloc = true;
 864         }
 865
 866         bdev = bh.b_bdev;
 867
 868         /*
 869          * If the filesystem isn't willing to tell us the length of a hole,
 870          * just fall back to PTEs.  Calling get_block 512 times in a loop
 871          * would be silly.
 872          */
 873         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 874                 dax_pmd_dbg(&bh, address, "allocated block too small");
 875                 return VM_FAULT_FALLBACK;
 876         }
 877
 878         /*
 879          * If we allocated new storage, make sure no process has any
 880          * zero pages covering this hole
 881          */
 882         if (alloc) {
 883                 loff_t lstart = pgoff << PAGE_SHIFT;
 884                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 885
 886                 truncate_pagecache_range(inode, lstart, lend);
 887         }
 888
 889         i_mmap_lock_read(mapping);
 890
 891         /*
 892          * If a truncate happened while we were allocating blocks, we may
 893          * leave blocks allocated to the file that are beyond EOF.  We can't
 894          * take i_mutex here, so just leave them hanging; they'll be freed
 895          * when the file is deleted.
 896          */
 897         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 898         if (pgoff >= size) {
 899                 result = VM_FAULT_SIGBUS;
 900                 goto out;
 901         }
 902         if ((pgoff | PG_PMD_COLOUR) >= size) {
 903                 dax_pmd_dbg(&bh, address,
 904                                 "offset + huge page size > file size");
 905                 goto fallback;
 906         }
 907
 908         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 909                 spinlock_t *ptl;
 910                 pmd_t entry;
 911                 struct page *zero_page = get_huge_zero_page();
 912
 913                 if (unlikely(!zero_page)) {
 914                         dax_pmd_dbg(&bh, address, "no zero page");
 915                         goto fallback;
 916                 }
 917
 918                 ptl = pmd_lock(vma->vm_mm, pmd);
 919                 if (!pmd_none(*pmd)) {
 920                         spin_unlock(ptl);
 921                         dax_pmd_dbg(&bh, address, "pmd already present");
 922                         goto fallback;
 923                 }
 924
 925                 dev_dbg(part_to_dev(bdev->bd_part),
 926                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 927                                 __func__, current->comm, address,
 928                                 (unsigned long long) to_sector(&bh, inode));
 929
 930                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 931                 entry = pmd_mkhuge(entry);
 932                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 933                 result = VM_FAULT_NOPAGE;
 934                 spin_unlock(ptl);
 935         } else {
 936                 struct blk_dax_ctl dax = {
 937                         .sector = to_sector(&bh, inode),
 938                         .size = PMD_SIZE,
 939                 };
 940                 long length = dax_map_atomic(bdev, &dax);
 941
 942                 if (length < 0) {
 943                         result = VM_FAULT_SIGBUS;
 944                         goto out;
 945                 }
 946                 if (length < PMD_SIZE) {
 947                         dax_pmd_dbg(&bh, address, "dax-length too small");
 948                         dax_unmap_atomic(bdev, &dax);
 949                         goto fallback;
 950                 }
 951                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 952                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 953                         dax_unmap_atomic(bdev, &dax);
 954                         goto fallback;
 955                 }
 956
 957                 if (!pfn_t_devmap(dax.pfn)) {
 958                         dax_unmap_atomic(bdev, &dax);
 959                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 960                         goto fallback;
 961                 }
 962
 963                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 964                         clear_pmem(dax.addr, PMD_SIZE);
 965                         wmb_pmem();
 966                         count_vm_event(PGMAJFAULT);
 967                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 968                         result |= VM_FAULT_MAJOR;
 969                 }
 970                 dax_unmap_atomic(bdev, &dax);
 971
 972                 /*
 973                  * For PTE faults we insert a radix tree entry for reads, and
 974                  * leave it clean.  Then on the first write we dirty the radix
 975                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 976                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 977                  * call into get_block() to translate the pgoff to a sector in
 978                  * order to be able to create a new radix tree entry.
 979                  *
 980                  * The PMD path doesn't have an equivalent to
 981                  * dax_pfn_mkwrite(), though, so for a read followed by a
 982                  * write we traverse all the way through __dax_pmd_fault()
 983                  * twice.  This means we can just skip inserting a radix tree
 984                  * entry completely on the initial read and just wait until
 985                  * the write to insert a dirty entry.
 986                  */
 987                 if (write) {
 988                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 989                                         true, true);
 990                         if (error) {
 991                                 dax_pmd_dbg(&bh, address,
 992                                                 "PMD radix insertion failed");
 993                                 goto fallback;
 994                         }
 995                 }
 996
 997                 dev_dbg(part_to_dev(bdev->bd_part),
 998                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 999                                 __func__, current->comm, address,
1000                                 pfn_t_to_pfn(dax.pfn),
1001                                 (unsigned long long) dax.sector);
1002                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1003                                 dax.pfn, write);
1004         }
1005
1006  out:
1007         i_mmap_unlock_read(mapping);
1008
1009         if (buffer_unwritten(&bh))
1010                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1011
1012         return result;
1013
1014  fallback:
1015         count_vm_event(THP_FAULT_FALLBACK);
1016         result = VM_FAULT_FALLBACK;
1017         goto out;
1018 }
1019 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1020
1021 /**
1022  * dax_pmd_fault - handle a PMD fault on a DAX file
1023  * @vma: The virtual memory area where the fault occurred
1024  * @vmf: The description of the fault
1025  * @get_block: The filesystem method used to translate file offsets to blocks
1026  *
1027  * When a page fault occurs, filesystems may call this helper in their
1028  * pmd_fault handler for DAX files.
1029  */
1030 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1031                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
1032                         dax_iodone_t complete_unwritten)
1033 {
1034         int result;
1035         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1036
1037         if (flags & FAULT_FLAG_WRITE) {
1038                 sb_start_pagefault(sb);
1039                 file_update_time(vma->vm_file);
1040         }
1041         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1042                                 complete_unwritten);
1043         if (flags & FAULT_FLAG_WRITE)
1044                 sb_end_pagefault(sb);
1045
1046         return result;
1047 }
1048 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1049 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1050
1051 /**
1052  * dax_pfn_mkwrite - handle first write to DAX page
1053  * @vma: The virtual memory area where the fault occurred
1054  * @vmf: The description of the fault
1055  */
1056 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1057 {
1058         struct file *file = vma->vm_file;
1059         int error;
1060
1061         /*
1062          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1063          * RADIX_DAX_PTE entry already exists in the radix tree from a
1064          * previous call to __dax_fault().  We just want to look up that PTE
1065          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1066          * saves us from having to make a call to get_block() here to look
1067          * up the sector.
1068          */
1069         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1070                         true);
1071
1072         if (error == -ENOMEM)
1073                 return VM_FAULT_OOM;
1074         if (error)
1075                 return VM_FAULT_SIGBUS;
1076         return VM_FAULT_NOPAGE;
1077 }
1078 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1079
1080 /**
1081  * dax_zero_page_range - zero a range within a page of a DAX file
1082  * @inode: The file being truncated
1083  * @from: The file offset that is being truncated to
1084  * @length: The number of bytes to zero
1085  * @get_block: The filesystem method used to translate file offsets to blocks
1086  *
1087  * This function can be called by a filesystem when it is zeroing part of a
1088  * page in a DAX file.  This is intended for hole-punch operations.  If
1089  * you are truncating a file, the helper function dax_truncate_page() may be
1090  * more convenient.
1091  *
1092  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1093  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1094  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1095  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1096  * since the file might be mmapped.
1097  */
1098 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1099                                                         get_block_t get_block)
1100 {
1101         struct buffer_head bh;
1102         pgoff_t index = from >> PAGE_CACHE_SHIFT;
1103         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1104         int err;
1105
1106         /* Block boundary? Nothing to do */
1107         if (!length)
1108                 return 0;
1109         BUG_ON((offset + length) > PAGE_CACHE_SIZE);
1110
1111         memset(&bh, 0, sizeof(bh));
1112         bh.b_bdev = inode->i_sb->s_bdev;
1113         bh.b_size = PAGE_CACHE_SIZE;
1114         err = get_block(inode, index, &bh, 0);
1115         if (err < 0)
1116                 return err;
1117         if (buffer_written(&bh)) {
1118                 struct block_device *bdev = bh.b_bdev;
1119                 struct blk_dax_ctl dax = {
1120                         .sector = to_sector(&bh, inode),
1121                         .size = PAGE_CACHE_SIZE,
1122                 };
1123
1124                 if (dax_map_atomic(bdev, &dax) < 0)
1125                         return PTR_ERR(dax.addr);
1126                 clear_pmem(dax.addr + offset, length);
1127                 wmb_pmem();
1128                 dax_unmap_atomic(bdev, &dax);
1129         }
1130
1131         return 0;
1132 }
1133 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1134
1135 /**
1136  * dax_truncate_page - handle a partial page being truncated in a DAX file
1137  * @inode: The file being truncated
1138  * @from: The file offset that is being truncated to
1139  * @get_block: The filesystem method used to translate file offsets to blocks
1140  *
1141  * Similar to block_truncate_page(), this function can be called by a
1142  * filesystem when it is truncating a DAX file to handle the partial page.
1143  *
1144  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1145  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1146  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1147  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1148  * since the file might be mmapped.
1149  */
1150 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1151 {
1152         unsigned length = PAGE_CACHE_ALIGN(from) - from;
1153         return dax_zero_page_range(inode, from, length, get_block);
1154 }
1155 EXPORT_SYMBOL_GPL(dax_truncate_page);