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 #define RADIX_DAX_MASK  0xf
  36 #define RADIX_DAX_SHIFT 4
  37 #define RADIX_DAX_PTE  (0x4 | RADIX_TREE_EXCEPTIONAL_ENTRY)
  38 #define RADIX_DAX_PMD  (0x8 | RADIX_TREE_EXCEPTIONAL_ENTRY)
  39 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_MASK)
  40 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
  41 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
  42                 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE)))
  43
  44 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  45 {
  46         struct request_queue *q = bdev->bd_queue;
  47         long rc = -EIO;
  48
  49         dax->addr = (void __pmem *) ERR_PTR(-EIO);
  50         if (blk_queue_enter(q, true) != 0)
  51                 return rc;
  52
  53         rc = bdev_direct_access(bdev, dax);
  54         if (rc < 0) {
  55                 dax->addr = (void __pmem *) ERR_PTR(rc);
  56                 blk_queue_exit(q);
  57                 return rc;
  58         }
  59         return rc;
  60 }
  61
  62 static void dax_unmap_atomic(struct block_device *bdev,
  63                 const struct blk_dax_ctl *dax)
  64 {
  65         if (IS_ERR(dax->addr))
  66                 return;
  67         blk_queue_exit(bdev->bd_queue);
  68 }
  69
  70 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
  71 {
  72         struct page *page = alloc_pages(GFP_KERNEL, 0);
  73         struct blk_dax_ctl dax = {
  74                 .size = PAGE_SIZE,
  75                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
  76         };
  77         long rc;
  78
  79         if (!page)
  80                 return ERR_PTR(-ENOMEM);
  81
  82         rc = dax_map_atomic(bdev, &dax);
  83         if (rc < 0)
  84                 return ERR_PTR(rc);
  85         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
  86         dax_unmap_atomic(bdev, &dax);
  87         return page;
  88 }
  89
  90 /*
  91  * dax_clear_sectors() is called from within transaction context from XFS,
  92  * and hence this means the stack from this point must follow GFP_NOFS
  93  * semantics for all operations.
  94  */
  95 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
  96 {
  97         struct blk_dax_ctl dax = {
  98                 .sector = _sector,
  99                 .size = _size,
 100         };
 101
 102         might_sleep();
 103         do {
 104                 long count, sz;
 105
 106                 count = dax_map_atomic(bdev, &dax);
 107                 if (count < 0)
 108                         return count;
 109                 sz = min_t(long, count, SZ_128K);
 110                 clear_pmem(dax.addr, sz);
 111                 dax.size -= sz;
 112                 dax.sector += sz / 512;
 113                 dax_unmap_atomic(bdev, &dax);
 114                 cond_resched();
 115         } while (dax.size);
 116
 117         wmb_pmem();
 118         return 0;
 119 }
 120 EXPORT_SYMBOL_GPL(dax_clear_sectors);
 121
 122 static bool buffer_written(struct buffer_head *bh)
 123 {
 124         return buffer_mapped(bh) && !buffer_unwritten(bh);
 125 }
 126
 127 /*
 128  * When ext4 encounters a hole, it returns without modifying the buffer_head
 129  * which means that we can't trust b_size.  To cope with this, we set b_state
 130  * to 0 before calling get_block and, if any bit is set, we know we can trust
 131  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 132  * and would save us time calling get_block repeatedly.
 133  */
 134 static bool buffer_size_valid(struct buffer_head *bh)
 135 {
 136         return bh->b_state != 0;
 137 }
 138
 139
 140 static sector_t to_sector(const struct buffer_head *bh,
 141                 const struct inode *inode)
 142 {
 143         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 144
 145         return sector;
 146 }
 147
 148 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 149                       loff_t start, loff_t end, get_block_t get_block,
 150                       struct buffer_head *bh)
 151 {
 152         loff_t pos = start, max = start, bh_max = start;
 153         bool hole = false, need_wmb = false;
 154         struct block_device *bdev = NULL;
 155         int rw = iov_iter_rw(iter), rc;
 156         long map_len = 0;
 157         struct blk_dax_ctl dax = {
 158                 .addr = (void __pmem *) ERR_PTR(-EIO),
 159         };
 160         unsigned blkbits = inode->i_blkbits;
 161         sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
 162                                                                 >> blkbits;
 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                         long page = pos >> PAGE_SHIFT;
 171                         sector_t block = page << (PAGE_SHIFT - blkbits);
 172                         unsigned first = pos - (block << blkbits);
 173                         long size;
 174
 175                         if (pos == bh_max) {
 176                                 bh->b_size = PAGE_ALIGN(end - pos);
 177                                 bh->b_state = 0;
 178                                 rc = get_block(inode, block, bh, rw == WRITE);
 179                                 if (rc)
 180                                         break;
 181                                 if (!buffer_size_valid(bh))
 182                                         bh->b_size = 1 << blkbits;
 183                                 bh_max = pos - first + bh->b_size;
 184                                 bdev = bh->b_bdev;
 185                                 /*
 186                                  * We allow uninitialized buffers for writes
 187                                  * beyond EOF as those cannot race with faults
 188                                  */
 189                                 WARN_ON_ONCE(
 190                                         (buffer_new(bh) && block < file_blks) ||
 191                                         (rw == WRITE && buffer_unwritten(bh)));
 192                         } else {
 193                                 unsigned done = bh->b_size -
 194                                                 (bh_max - (pos - first));
 195                                 bh->b_blocknr += done >> blkbits;
 196                                 bh->b_size -= done;
 197                         }
 198
 199                         hole = rw == READ && !buffer_written(bh);
 200                         if (hole) {
 201                                 size = bh->b_size - first;
 202                         } else {
 203                                 dax_unmap_atomic(bdev, &dax);
 204                                 dax.sector = to_sector(bh, inode);
 205                                 dax.size = bh->b_size;
 206                                 map_len = dax_map_atomic(bdev, &dax);
 207                                 if (map_len < 0) {
 208                                         rc = map_len;
 209                                         break;
 210                                 }
 211                                 dax.addr += first;
 212                                 size = map_len - first;
 213                         }
 214                         max = min(pos + size, end);
 215                 }
 216
 217                 if (iov_iter_rw(iter) == WRITE) {
 218                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 219                         need_wmb = true;
 220                 } else if (!hole)
 221                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 222                                         iter);
 223                 else
 224                         len = iov_iter_zero(max - pos, iter);
 225
 226                 if (!len) {
 227                         rc = -EFAULT;
 228                         break;
 229                 }
 230
 231                 pos += len;
 232                 if (!IS_ERR(dax.addr))
 233                         dax.addr += len;
 234         }
 235
 236         if (need_wmb)
 237                 wmb_pmem();
 238         dax_unmap_atomic(bdev, &dax);
 239
 240         return (pos == start) ? rc : pos - start;
 241 }
 242
 243 /**
 244  * dax_do_io - Perform I/O to a DAX file
 245  * @iocb: The control block for this I/O
 246  * @inode: The file which the I/O is directed at
 247  * @iter: The addresses to do I/O from or to
 248  * @pos: The file offset where the I/O starts
 249  * @get_block: The filesystem method used to translate file offsets to blocks
 250  * @end_io: A filesystem callback for I/O completion
 251  * @flags: See below
 252  *
 253  * This function uses the same locking scheme as do_blockdev_direct_IO:
 254  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 255  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 256  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 257  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 258  * is in progress.
 259  */
 260 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 261                   struct iov_iter *iter, loff_t pos, get_block_t get_block,
 262                   dio_iodone_t end_io, int flags)
 263 {
 264         struct buffer_head bh;
 265         ssize_t retval = -EINVAL;
 266         loff_t end = pos + iov_iter_count(iter);
 267
 268         memset(&bh, 0, sizeof(bh));
 269         bh.b_bdev = inode->i_sb->s_bdev;
 270
 271         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 272                 inode_lock(inode);
 273
 274         /* Protects against truncate */
 275         if (!(flags & DIO_SKIP_DIO_COUNT))
 276                 inode_dio_begin(inode);
 277
 278         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 279
 280         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 281                 inode_unlock(inode);
 282
 283         if (end_io) {
 284                 int err;
 285
 286                 err = end_io(iocb, pos, retval, bh.b_private);
 287                 if (err)
 288                         retval = err;
 289         }
 290
 291         if (!(flags & DIO_SKIP_DIO_COUNT))
 292                 inode_dio_end(inode);
 293         return retval;
 294 }
 295 EXPORT_SYMBOL_GPL(dax_do_io);
 296
 297 /*
 298  * The user has performed a load from a hole in the file.  Allocating
 299  * a new page in the file would cause excessive storage usage for
 300  * workloads with sparse files.  We allocate a page cache page instead.
 301  * We'll kick it out of the page cache if it's ever written to,
 302  * otherwise it will simply fall out of the page cache under memory
 303  * pressure without ever having been dirtied.
 304  */
 305 static int dax_load_hole(struct address_space *mapping, struct page *page,
 306                                                         struct vm_fault *vmf)
 307 {
 308         unsigned long size;
 309         struct inode *inode = mapping->host;
 310         if (!page)
 311                 page = find_or_create_page(mapping, vmf->pgoff,
 312                                                 GFP_KERNEL | __GFP_ZERO);
 313         if (!page)
 314                 return VM_FAULT_OOM;
 315         /* Recheck i_size under page lock to avoid truncate race */
 316         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 317         if (vmf->pgoff >= size) {
 318                 unlock_page(page);
 319                 put_page(page);
 320                 return VM_FAULT_SIGBUS;
 321         }
 322
 323         vmf->page = page;
 324         return VM_FAULT_LOCKED;
 325 }
 326
 327 static int copy_user_bh(struct page *to, struct inode *inode,
 328                 struct buffer_head *bh, unsigned long vaddr)
 329 {
 330         struct blk_dax_ctl dax = {
 331                 .sector = to_sector(bh, inode),
 332                 .size = bh->b_size,
 333         };
 334         struct block_device *bdev = bh->b_bdev;
 335         void *vto;
 336
 337         if (dax_map_atomic(bdev, &dax) < 0)
 338                 return PTR_ERR(dax.addr);
 339         vto = kmap_atomic(to);
 340         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 341         kunmap_atomic(vto);
 342         dax_unmap_atomic(bdev, &dax);
 343         return 0;
 344 }
 345
 346 #define NO_SECTOR -1
 347 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 348
 349 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 350                 sector_t sector, bool pmd_entry, bool dirty)
 351 {
 352         struct radix_tree_root *page_tree = &mapping->page_tree;
 353         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 354         int type, error = 0;
 355         void *entry;
 356
 357         WARN_ON_ONCE(pmd_entry && !dirty);
 358         if (dirty)
 359                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 360
 361         spin_lock_irq(&mapping->tree_lock);
 362
 363         entry = radix_tree_lookup(page_tree, pmd_index);
 364         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 365                 index = pmd_index;
 366                 goto dirty;
 367         }
 368
 369         entry = radix_tree_lookup(page_tree, index);
 370         if (entry) {
 371                 type = RADIX_DAX_TYPE(entry);
 372                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 373                                         type != RADIX_DAX_PMD)) {
 374                         error = -EIO;
 375                         goto unlock;
 376                 }
 377
 378                 if (!pmd_entry || type == RADIX_DAX_PMD)
 379                         goto dirty;
 380
 381                 /*
 382                  * We only insert dirty PMD entries into the radix tree.  This
 383                  * means we don't need to worry about removing a dirty PTE
 384                  * entry and inserting a clean PMD entry, thus reducing the
 385                  * range we would flush with a follow-up fsync/msync call.
 386                  */
 387                 radix_tree_delete(&mapping->page_tree, index);
 388                 mapping->nrexceptional--;
 389         }
 390
 391         if (sector == NO_SECTOR) {
 392                 /*
 393                  * This can happen during correct operation if our pfn_mkwrite
 394                  * fault raced against a hole punch operation.  If this
 395                  * happens the pte that was hole punched will have been
 396                  * unmapped and the radix tree entry will have been removed by
 397                  * the time we are called, but the call will still happen.  We
 398                  * will return all the way up to wp_pfn_shared(), where the
 399                  * pte_same() check will fail, eventually causing page fault
 400                  * to be retried by the CPU.
 401                  */
 402                 goto unlock;
 403         }
 404
 405         error = radix_tree_insert(page_tree, index,
 406                         RADIX_DAX_ENTRY(sector, pmd_entry));
 407         if (error)
 408                 goto unlock;
 409
 410         mapping->nrexceptional++;
 411  dirty:
 412         if (dirty)
 413                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 414  unlock:
 415         spin_unlock_irq(&mapping->tree_lock);
 416         return error;
 417 }
 418
 419 static int dax_writeback_one(struct block_device *bdev,
 420                 struct address_space *mapping, pgoff_t index, void *entry)
 421 {
 422         struct radix_tree_root *page_tree = &mapping->page_tree;
 423         int type = RADIX_DAX_TYPE(entry);
 424         struct radix_tree_node *node;
 425         struct blk_dax_ctl dax;
 426         void **slot;
 427         int ret = 0;
 428
 429         spin_lock_irq(&mapping->tree_lock);
 430         /*
 431          * Regular page slots are stabilized by the page lock even
 432          * without the tree itself locked.  These unlocked entries
 433          * need verification under the tree lock.
 434          */
 435         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 436                 goto unlock;
 437         if (*slot != entry)
 438                 goto unlock;
 439
 440         /* another fsync thread may have already written back this entry */
 441         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 442                 goto unlock;
 443
 444         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 445                 ret = -EIO;
 446                 goto unlock;
 447         }
 448
 449         dax.sector = RADIX_DAX_SECTOR(entry);
 450         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 451         spin_unlock_irq(&mapping->tree_lock);
 452
 453         /*
 454          * We cannot hold tree_lock while calling dax_map_atomic() because it
 455          * eventually calls cond_resched().
 456          */
 457         ret = dax_map_atomic(bdev, &dax);
 458         if (ret < 0)
 459                 return ret;
 460
 461         if (WARN_ON_ONCE(ret < dax.size)) {
 462                 ret = -EIO;
 463                 goto unmap;
 464         }
 465
 466         wb_cache_pmem(dax.addr, dax.size);
 467
 468         spin_lock_irq(&mapping->tree_lock);
 469         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 470         spin_unlock_irq(&mapping->tree_lock);
 471  unmap:
 472         dax_unmap_atomic(bdev, &dax);
 473         return ret;
 474
 475  unlock:
 476         spin_unlock_irq(&mapping->tree_lock);
 477         return ret;
 478 }
 479
 480 /*
 481  * Flush the mapping to the persistent domain within the byte range of [start,
 482  * end]. This is required by data integrity operations to ensure file data is
 483  * on persistent storage prior to completion of the operation.
 484  */
 485 int dax_writeback_mapping_range(struct address_space *mapping,
 486                 struct block_device *bdev, struct writeback_control *wbc)
 487 {
 488         struct inode *inode = mapping->host;
 489         pgoff_t start_index, end_index, pmd_index;
 490         pgoff_t indices[PAGEVEC_SIZE];
 491         struct pagevec pvec;
 492         bool done = false;
 493         int i, ret = 0;
 494         void *entry;
 495
 496         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 497                 return -EIO;
 498
 499         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 500                 return 0;
 501
 502         start_index = wbc->range_start >> PAGE_SHIFT;
 503         end_index = wbc->range_end >> PAGE_SHIFT;
 504         pmd_index = DAX_PMD_INDEX(start_index);
 505
 506         rcu_read_lock();
 507         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 508         rcu_read_unlock();
 509
 510         /* see if the start of our range is covered by a PMD entry */
 511         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 512                 start_index = pmd_index;
 513
 514         tag_pages_for_writeback(mapping, start_index, end_index);
 515
 516         pagevec_init(&pvec, 0);
 517         while (!done) {
 518                 pvec.nr = find_get_entries_tag(mapping, start_index,
 519                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 520                                 pvec.pages, indices);
 521
 522                 if (pvec.nr == 0)
 523                         break;
 524
 525                 for (i = 0; i < pvec.nr; i++) {
 526                         if (indices[i] > end_index) {
 527                                 done = true;
 528                                 break;
 529                         }
 530
 531                         ret = dax_writeback_one(bdev, mapping, indices[i],
 532                                         pvec.pages[i]);
 533                         if (ret < 0)
 534                                 return ret;
 535                 }
 536         }
 537         wmb_pmem();
 538         return 0;
 539 }
 540 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 541
 542 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 543                         struct vm_area_struct *vma, struct vm_fault *vmf)
 544 {
 545         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 546         struct address_space *mapping = inode->i_mapping;
 547         struct block_device *bdev = bh->b_bdev;
 548         struct blk_dax_ctl dax = {
 549                 .sector = to_sector(bh, inode),
 550                 .size = bh->b_size,
 551         };
 552         pgoff_t size;
 553         int error;
 554
 555         i_mmap_lock_read(mapping);
 556
 557         /*
 558          * Check truncate didn't happen while we were allocating a block.
 559          * If it did, this block may or may not be still allocated to the
 560          * file.  We can't tell the filesystem to free it because we can't
 561          * take i_mutex here.  In the worst case, the file still has blocks
 562          * allocated past the end of the file.
 563          */
 564         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 565         if (unlikely(vmf->pgoff >= size)) {
 566                 error = -EIO;
 567                 goto out;
 568         }
 569
 570         if (dax_map_atomic(bdev, &dax) < 0) {
 571                 error = PTR_ERR(dax.addr);
 572                 goto out;
 573         }
 574         dax_unmap_atomic(bdev, &dax);
 575
 576         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 577                         vmf->flags & FAULT_FLAG_WRITE);
 578         if (error)
 579                 goto out;
 580
 581         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 582
 583  out:
 584         i_mmap_unlock_read(mapping);
 585
 586         return error;
 587 }
 588
 589 /**
 590  * __dax_fault - handle a page fault on a DAX file
 591  * @vma: The virtual memory area where the fault occurred
 592  * @vmf: The description of the fault
 593  * @get_block: The filesystem method used to translate file offsets to blocks
 594  *
 595  * When a page fault occurs, filesystems may call this helper in their
 596  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 597  * the necessary locking for the page fault to proceed successfully.
 598  */
 599 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 600                         get_block_t get_block)
 601 {
 602         struct file *file = vma->vm_file;
 603         struct address_space *mapping = file->f_mapping;
 604         struct inode *inode = mapping->host;
 605         struct page *page;
 606         struct buffer_head bh;
 607         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 608         unsigned blkbits = inode->i_blkbits;
 609         sector_t block;
 610         pgoff_t size;
 611         int error;
 612         int major = 0;
 613
 614         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 615         if (vmf->pgoff >= size)
 616                 return VM_FAULT_SIGBUS;
 617
 618         memset(&bh, 0, sizeof(bh));
 619         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 620         bh.b_bdev = inode->i_sb->s_bdev;
 621         bh.b_size = PAGE_SIZE;
 622
 623  repeat:
 624         page = find_get_page(mapping, vmf->pgoff);
 625         if (page) {
 626                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 627                         put_page(page);
 628                         return VM_FAULT_RETRY;
 629                 }
 630                 if (unlikely(page->mapping != mapping)) {
 631                         unlock_page(page);
 632                         put_page(page);
 633                         goto repeat;
 634                 }
 635                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 636                 if (unlikely(vmf->pgoff >= size)) {
 637                         /*
 638                          * We have a struct page covering a hole in the file
 639                          * from a read fault and we've raced with a truncate
 640                          */
 641                         error = -EIO;
 642                         goto unlock_page;
 643                 }
 644         }
 645
 646         error = get_block(inode, block, &bh, 0);
 647         if (!error && (bh.b_size < PAGE_SIZE))
 648                 error = -EIO;           /* fs corruption? */
 649         if (error)
 650                 goto unlock_page;
 651
 652         if (!buffer_mapped(&bh) && !vmf->cow_page) {
 653                 if (vmf->flags & FAULT_FLAG_WRITE) {
 654                         error = get_block(inode, block, &bh, 1);
 655                         count_vm_event(PGMAJFAULT);
 656                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 657                         major = VM_FAULT_MAJOR;
 658                         if (!error && (bh.b_size < PAGE_SIZE))
 659                                 error = -EIO;
 660                         if (error)
 661                                 goto unlock_page;
 662                 } else {
 663                         return dax_load_hole(mapping, page, vmf);
 664                 }
 665         }
 666
 667         if (vmf->cow_page) {
 668                 struct page *new_page = vmf->cow_page;
 669                 if (buffer_written(&bh))
 670                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 671                 else
 672                         clear_user_highpage(new_page, vaddr);
 673                 if (error)
 674                         goto unlock_page;
 675                 vmf->page = page;
 676                 if (!page) {
 677                         i_mmap_lock_read(mapping);
 678                         /* Check we didn't race with truncate */
 679                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 680                                                                 PAGE_SHIFT;
 681                         if (vmf->pgoff >= size) {
 682                                 i_mmap_unlock_read(mapping);
 683                                 error = -EIO;
 684                                 goto out;
 685                         }
 686                 }
 687                 return VM_FAULT_LOCKED;
 688         }
 689
 690         /* Check we didn't race with a read fault installing a new page */
 691         if (!page && major)
 692                 page = find_lock_page(mapping, vmf->pgoff);
 693
 694         if (page) {
 695                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 696                                                         PAGE_SIZE, 0);
 697                 delete_from_page_cache(page);
 698                 unlock_page(page);
 699                 put_page(page);
 700                 page = NULL;
 701         }
 702
 703         /* Filesystem should not return unwritten buffers to us! */
 704         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 705         error = dax_insert_mapping(inode, &bh, vma, vmf);
 706
 707  out:
 708         if (error == -ENOMEM)
 709                 return VM_FAULT_OOM | major;
 710         /* -EBUSY is fine, somebody else faulted on the same PTE */
 711         if ((error < 0) && (error != -EBUSY))
 712                 return VM_FAULT_SIGBUS | major;
 713         return VM_FAULT_NOPAGE | major;
 714
 715  unlock_page:
 716         if (page) {
 717                 unlock_page(page);
 718                 put_page(page);
 719         }
 720         goto out;
 721 }
 722 EXPORT_SYMBOL(__dax_fault);
 723
 724 /**
 725  * dax_fault - handle a page fault on a DAX file
 726  * @vma: The virtual memory area where the fault occurred
 727  * @vmf: The description of the fault
 728  * @get_block: The filesystem method used to translate file offsets to blocks
 729  *
 730  * When a page fault occurs, filesystems may call this helper in their
 731  * fault handler for DAX files.
 732  */
 733 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 734               get_block_t get_block)
 735 {
 736         int result;
 737         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 738
 739         if (vmf->flags & FAULT_FLAG_WRITE) {
 740                 sb_start_pagefault(sb);
 741                 file_update_time(vma->vm_file);
 742         }
 743         result = __dax_fault(vma, vmf, get_block);
 744         if (vmf->flags & FAULT_FLAG_WRITE)
 745                 sb_end_pagefault(sb);
 746
 747         return result;
 748 }
 749 EXPORT_SYMBOL_GPL(dax_fault);
 750
 751 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 752 /*
 753  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 754  * more often than one might expect in the below function.
 755  */
 756 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 757
 758 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 759                 const char *reason, const char *fn)
 760 {
 761         if (bh) {
 762                 char bname[BDEVNAME_SIZE];
 763                 bdevname(bh->b_bdev, bname);
 764                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 765                         "length %zd fallback: %s\n", fn, current->comm,
 766                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 767                         bh->b_size, reason);
 768         } else {
 769                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 770                         current->comm, address, reason);
 771         }
 772 }
 773
 774 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 775
 776 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 777                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 778 {
 779         struct file *file = vma->vm_file;
 780         struct address_space *mapping = file->f_mapping;
 781         struct inode *inode = mapping->host;
 782         struct buffer_head bh;
 783         unsigned blkbits = inode->i_blkbits;
 784         unsigned long pmd_addr = address & PMD_MASK;
 785         bool write = flags & FAULT_FLAG_WRITE;
 786         struct block_device *bdev;
 787         pgoff_t size, pgoff;
 788         sector_t block;
 789         int error, result = 0;
 790         bool alloc = false;
 791
 792         /* dax pmd mappings require pfn_t_devmap() */
 793         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 794                 return VM_FAULT_FALLBACK;
 795
 796         /* Fall back to PTEs if we're going to COW */
 797         if (write && !(vma->vm_flags & VM_SHARED)) {
 798                 split_huge_pmd(vma, pmd, address);
 799                 dax_pmd_dbg(NULL, address, "cow write");
 800                 return VM_FAULT_FALLBACK;
 801         }
 802         /* If the PMD would extend outside the VMA */
 803         if (pmd_addr < vma->vm_start) {
 804                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 805                 return VM_FAULT_FALLBACK;
 806         }
 807         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 808                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 809                 return VM_FAULT_FALLBACK;
 810         }
 811
 812         pgoff = linear_page_index(vma, pmd_addr);
 813         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 814         if (pgoff >= size)
 815                 return VM_FAULT_SIGBUS;
 816         /* If the PMD would cover blocks out of the file */
 817         if ((pgoff | PG_PMD_COLOUR) >= size) {
 818                 dax_pmd_dbg(NULL, address,
 819                                 "offset + huge page size > file size");
 820                 return VM_FAULT_FALLBACK;
 821         }
 822
 823         memset(&bh, 0, sizeof(bh));
 824         bh.b_bdev = inode->i_sb->s_bdev;
 825         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 826
 827         bh.b_size = PMD_SIZE;
 828
 829         if (get_block(inode, block, &bh, 0) != 0)
 830                 return VM_FAULT_SIGBUS;
 831
 832         if (!buffer_mapped(&bh) && write) {
 833                 if (get_block(inode, block, &bh, 1) != 0)
 834                         return VM_FAULT_SIGBUS;
 835                 alloc = true;
 836                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 837         }
 838
 839         bdev = bh.b_bdev;
 840
 841         /*
 842          * If the filesystem isn't willing to tell us the length of a hole,
 843          * just fall back to PTEs.  Calling get_block 512 times in a loop
 844          * would be silly.
 845          */
 846         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 847                 dax_pmd_dbg(&bh, address, "allocated block too small");
 848                 return VM_FAULT_FALLBACK;
 849         }
 850
 851         /*
 852          * If we allocated new storage, make sure no process has any
 853          * zero pages covering this hole
 854          */
 855         if (alloc) {
 856                 loff_t lstart = pgoff << PAGE_SHIFT;
 857                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 858
 859                 truncate_pagecache_range(inode, lstart, lend);
 860         }
 861
 862         i_mmap_lock_read(mapping);
 863
 864         /*
 865          * If a truncate happened while we were allocating blocks, we may
 866          * leave blocks allocated to the file that are beyond EOF.  We can't
 867          * take i_mutex here, so just leave them hanging; they'll be freed
 868          * when the file is deleted.
 869          */
 870         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 871         if (pgoff >= size) {
 872                 result = VM_FAULT_SIGBUS;
 873                 goto out;
 874         }
 875         if ((pgoff | PG_PMD_COLOUR) >= size) {
 876                 dax_pmd_dbg(&bh, address,
 877                                 "offset + huge page size > file size");
 878                 goto fallback;
 879         }
 880
 881         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 882                 spinlock_t *ptl;
 883                 pmd_t entry;
 884                 struct page *zero_page = get_huge_zero_page();
 885
 886                 if (unlikely(!zero_page)) {
 887                         dax_pmd_dbg(&bh, address, "no zero page");
 888                         goto fallback;
 889                 }
 890
 891                 ptl = pmd_lock(vma->vm_mm, pmd);
 892                 if (!pmd_none(*pmd)) {
 893                         spin_unlock(ptl);
 894                         dax_pmd_dbg(&bh, address, "pmd already present");
 895                         goto fallback;
 896                 }
 897
 898                 dev_dbg(part_to_dev(bdev->bd_part),
 899                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 900                                 __func__, current->comm, address,
 901                                 (unsigned long long) to_sector(&bh, inode));
 902
 903                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 904                 entry = pmd_mkhuge(entry);
 905                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 906                 result = VM_FAULT_NOPAGE;
 907                 spin_unlock(ptl);
 908         } else {
 909                 struct blk_dax_ctl dax = {
 910                         .sector = to_sector(&bh, inode),
 911                         .size = PMD_SIZE,
 912                 };
 913                 long length = dax_map_atomic(bdev, &dax);
 914
 915                 if (length < 0) {
 916                         result = VM_FAULT_SIGBUS;
 917                         goto out;
 918                 }
 919                 if (length < PMD_SIZE) {
 920                         dax_pmd_dbg(&bh, address, "dax-length too small");
 921                         dax_unmap_atomic(bdev, &dax);
 922                         goto fallback;
 923                 }
 924                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 925                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 926                         dax_unmap_atomic(bdev, &dax);
 927                         goto fallback;
 928                 }
 929
 930                 if (!pfn_t_devmap(dax.pfn)) {
 931                         dax_unmap_atomic(bdev, &dax);
 932                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 933                         goto fallback;
 934                 }
 935                 dax_unmap_atomic(bdev, &dax);
 936
 937                 /*
 938                  * For PTE faults we insert a radix tree entry for reads, and
 939                  * leave it clean.  Then on the first write we dirty the radix
 940                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 941                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 942                  * call into get_block() to translate the pgoff to a sector in
 943                  * order to be able to create a new radix tree entry.
 944                  *
 945                  * The PMD path doesn't have an equivalent to
 946                  * dax_pfn_mkwrite(), though, so for a read followed by a
 947                  * write we traverse all the way through __dax_pmd_fault()
 948                  * twice.  This means we can just skip inserting a radix tree
 949                  * entry completely on the initial read and just wait until
 950                  * the write to insert a dirty entry.
 951                  */
 952                 if (write) {
 953                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 954                                         true, true);
 955                         if (error) {
 956                                 dax_pmd_dbg(&bh, address,
 957                                                 "PMD radix insertion failed");
 958                                 goto fallback;
 959                         }
 960                 }
 961
 962                 dev_dbg(part_to_dev(bdev->bd_part),
 963                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 964                                 __func__, current->comm, address,
 965                                 pfn_t_to_pfn(dax.pfn),
 966                                 (unsigned long long) dax.sector);
 967                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
 968                                 dax.pfn, write);
 969         }
 970
 971  out:
 972         i_mmap_unlock_read(mapping);
 973
 974         return result;
 975
 976  fallback:
 977         count_vm_event(THP_FAULT_FALLBACK);
 978         result = VM_FAULT_FALLBACK;
 979         goto out;
 980 }
 981 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
 982
 983 /**
 984  * dax_pmd_fault - handle a PMD fault on a DAX file
 985  * @vma: The virtual memory area where the fault occurred
 986  * @vmf: The description of the fault
 987  * @get_block: The filesystem method used to translate file offsets to blocks
 988  *
 989  * When a page fault occurs, filesystems may call this helper in their
 990  * pmd_fault handler for DAX files.
 991  */
 992 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 993                         pmd_t *pmd, unsigned int flags, get_block_t get_block)
 994 {
 995         int result;
 996         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 997
 998         if (flags & FAULT_FLAG_WRITE) {
 999                 sb_start_pagefault(sb);
1000                 file_update_time(vma->vm_file);
1001         }
1002         result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
1003         if (flags & FAULT_FLAG_WRITE)
1004                 sb_end_pagefault(sb);
1005
1006         return result;
1007 }
1008 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1009 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1010
1011 /**
1012  * dax_pfn_mkwrite - handle first write to DAX page
1013  * @vma: The virtual memory area where the fault occurred
1014  * @vmf: The description of the fault
1015  */
1016 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1017 {
1018         struct file *file = vma->vm_file;
1019         int error;
1020
1021         /*
1022          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1023          * RADIX_DAX_PTE entry already exists in the radix tree from a
1024          * previous call to __dax_fault().  We just want to look up that PTE
1025          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1026          * saves us from having to make a call to get_block() here to look
1027          * up the sector.
1028          */
1029         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1030                         true);
1031
1032         if (error == -ENOMEM)
1033                 return VM_FAULT_OOM;
1034         if (error)
1035                 return VM_FAULT_SIGBUS;
1036         return VM_FAULT_NOPAGE;
1037 }
1038 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1039
1040 /**
1041  * dax_zero_page_range - zero a range within a page of a DAX file
1042  * @inode: The file being truncated
1043  * @from: The file offset that is being truncated to
1044  * @length: The number of bytes to zero
1045  * @get_block: The filesystem method used to translate file offsets to blocks
1046  *
1047  * This function can be called by a filesystem when it is zeroing part of a
1048  * page in a DAX file.  This is intended for hole-punch operations.  If
1049  * you are truncating a file, the helper function dax_truncate_page() may be
1050  * more convenient.
1051  *
1052  * We work in terms of PAGE_SIZE here for commonality with
1053  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1054  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1055  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1056  * since the file might be mmapped.
1057  */
1058 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1059                                                         get_block_t get_block)
1060 {
1061         struct buffer_head bh;
1062         pgoff_t index = from >> PAGE_SHIFT;
1063         unsigned offset = from & (PAGE_SIZE-1);
1064         int err;
1065
1066         /* Block boundary? Nothing to do */
1067         if (!length)
1068                 return 0;
1069         BUG_ON((offset + length) > PAGE_SIZE);
1070
1071         memset(&bh, 0, sizeof(bh));
1072         bh.b_bdev = inode->i_sb->s_bdev;
1073         bh.b_size = PAGE_SIZE;
1074         err = get_block(inode, index, &bh, 0);
1075         if (err < 0)
1076                 return err;
1077         if (buffer_written(&bh)) {
1078                 struct block_device *bdev = bh.b_bdev;
1079                 struct blk_dax_ctl dax = {
1080                         .sector = to_sector(&bh, inode),
1081                         .size = PAGE_SIZE,
1082                 };
1083
1084                 if (dax_map_atomic(bdev, &dax) < 0)
1085                         return PTR_ERR(dax.addr);
1086                 clear_pmem(dax.addr + offset, length);
1087                 wmb_pmem();
1088                 dax_unmap_atomic(bdev, &dax);
1089         }
1090
1091         return 0;
1092 }
1093 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1094
1095 /**
1096  * dax_truncate_page - handle a partial page being truncated in a DAX file
1097  * @inode: The file being truncated
1098  * @from: The file offset that is being truncated to
1099  * @get_block: The filesystem method used to translate file offsets to blocks
1100  *
1101  * Similar to block_truncate_page(), this function can be called by a
1102  * filesystem when it is truncating a DAX file to handle the partial page.
1103  *
1104  * We work in terms of PAGE_SIZE here for commonality with
1105  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1106  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1107  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1108  * since the file might be mmapped.
1109  */
1110 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1111 {
1112         unsigned length = PAGE_ALIGN(from) - from;
1113         return dax_zero_page_range(inode, from, length, get_block);
1114 }
1115 EXPORT_SYMBOL_GPL(dax_truncate_page);