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                 struct address_space *mapping = inode->i_mapping;
 273                 inode_lock(inode);
 274                 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
 275                 if (retval) {
 276                         inode_unlock(inode);
 277                         goto out;
 278                 }
 279         }
 280
 281         /* Protects against truncate */
 282         if (!(flags & DIO_SKIP_DIO_COUNT))
 283                 inode_dio_begin(inode);
 284
 285         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 286
 287         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 288                 inode_unlock(inode);
 289
 290         if (end_io) {
 291                 int err;
 292
 293                 err = end_io(iocb, pos, retval, bh.b_private);
 294                 if (err)
 295                         retval = err;
 296         }
 297
 298         if (!(flags & DIO_SKIP_DIO_COUNT))
 299                 inode_dio_end(inode);
 300  out:
 301         return retval;
 302 }
 303 EXPORT_SYMBOL_GPL(dax_do_io);
 304
 305 /*
 306  * The user has performed a load from a hole in the file.  Allocating
 307  * a new page in the file would cause excessive storage usage for
 308  * workloads with sparse files.  We allocate a page cache page instead.
 309  * We'll kick it out of the page cache if it's ever written to,
 310  * otherwise it will simply fall out of the page cache under memory
 311  * pressure without ever having been dirtied.
 312  */
 313 static int dax_load_hole(struct address_space *mapping, struct page *page,
 314                                                         struct vm_fault *vmf)
 315 {
 316         unsigned long size;
 317         struct inode *inode = mapping->host;
 318         if (!page)
 319                 page = find_or_create_page(mapping, vmf->pgoff,
 320                                                 GFP_KERNEL | __GFP_ZERO);
 321         if (!page)
 322                 return VM_FAULT_OOM;
 323         /* Recheck i_size under page lock to avoid truncate race */
 324         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 325         if (vmf->pgoff >= size) {
 326                 unlock_page(page);
 327                 put_page(page);
 328                 return VM_FAULT_SIGBUS;
 329         }
 330
 331         vmf->page = page;
 332         return VM_FAULT_LOCKED;
 333 }
 334
 335 static int copy_user_bh(struct page *to, struct inode *inode,
 336                 struct buffer_head *bh, unsigned long vaddr)
 337 {
 338         struct blk_dax_ctl dax = {
 339                 .sector = to_sector(bh, inode),
 340                 .size = bh->b_size,
 341         };
 342         struct block_device *bdev = bh->b_bdev;
 343         void *vto;
 344
 345         if (dax_map_atomic(bdev, &dax) < 0)
 346                 return PTR_ERR(dax.addr);
 347         vto = kmap_atomic(to);
 348         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 349         kunmap_atomic(vto);
 350         dax_unmap_atomic(bdev, &dax);
 351         return 0;
 352 }
 353
 354 #define NO_SECTOR -1
 355 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 356
 357 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 358                 sector_t sector, bool pmd_entry, bool dirty)
 359 {
 360         struct radix_tree_root *page_tree = &mapping->page_tree;
 361         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 362         int type, error = 0;
 363         void *entry;
 364
 365         WARN_ON_ONCE(pmd_entry && !dirty);
 366         if (dirty)
 367                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 368
 369         spin_lock_irq(&mapping->tree_lock);
 370
 371         entry = radix_tree_lookup(page_tree, pmd_index);
 372         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 373                 index = pmd_index;
 374                 goto dirty;
 375         }
 376
 377         entry = radix_tree_lookup(page_tree, index);
 378         if (entry) {
 379                 type = RADIX_DAX_TYPE(entry);
 380                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 381                                         type != RADIX_DAX_PMD)) {
 382                         error = -EIO;
 383                         goto unlock;
 384                 }
 385
 386                 if (!pmd_entry || type == RADIX_DAX_PMD)
 387                         goto dirty;
 388
 389                 /*
 390                  * We only insert dirty PMD entries into the radix tree.  This
 391                  * means we don't need to worry about removing a dirty PTE
 392                  * entry and inserting a clean PMD entry, thus reducing the
 393                  * range we would flush with a follow-up fsync/msync call.
 394                  */
 395                 radix_tree_delete(&mapping->page_tree, index);
 396                 mapping->nrexceptional--;
 397         }
 398
 399         if (sector == NO_SECTOR) {
 400                 /*
 401                  * This can happen during correct operation if our pfn_mkwrite
 402                  * fault raced against a hole punch operation.  If this
 403                  * happens the pte that was hole punched will have been
 404                  * unmapped and the radix tree entry will have been removed by
 405                  * the time we are called, but the call will still happen.  We
 406                  * will return all the way up to wp_pfn_shared(), where the
 407                  * pte_same() check will fail, eventually causing page fault
 408                  * to be retried by the CPU.
 409                  */
 410                 goto unlock;
 411         }
 412
 413         error = radix_tree_insert(page_tree, index,
 414                         RADIX_DAX_ENTRY(sector, pmd_entry));
 415         if (error)
 416                 goto unlock;
 417
 418         mapping->nrexceptional++;
 419  dirty:
 420         if (dirty)
 421                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 422  unlock:
 423         spin_unlock_irq(&mapping->tree_lock);
 424         return error;
 425 }
 426
 427 static int dax_writeback_one(struct block_device *bdev,
 428                 struct address_space *mapping, pgoff_t index, void *entry)
 429 {
 430         struct radix_tree_root *page_tree = &mapping->page_tree;
 431         int type = RADIX_DAX_TYPE(entry);
 432         struct radix_tree_node *node;
 433         struct blk_dax_ctl dax;
 434         void **slot;
 435         int ret = 0;
 436
 437         spin_lock_irq(&mapping->tree_lock);
 438         /*
 439          * Regular page slots are stabilized by the page lock even
 440          * without the tree itself locked.  These unlocked entries
 441          * need verification under the tree lock.
 442          */
 443         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 444                 goto unlock;
 445         if (*slot != entry)
 446                 goto unlock;
 447
 448         /* another fsync thread may have already written back this entry */
 449         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 450                 goto unlock;
 451
 452         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 453                 ret = -EIO;
 454                 goto unlock;
 455         }
 456
 457         dax.sector = RADIX_DAX_SECTOR(entry);
 458         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 459         spin_unlock_irq(&mapping->tree_lock);
 460
 461         /*
 462          * We cannot hold tree_lock while calling dax_map_atomic() because it
 463          * eventually calls cond_resched().
 464          */
 465         ret = dax_map_atomic(bdev, &dax);
 466         if (ret < 0)
 467                 return ret;
 468
 469         if (WARN_ON_ONCE(ret < dax.size)) {
 470                 ret = -EIO;
 471                 goto unmap;
 472         }
 473
 474         wb_cache_pmem(dax.addr, dax.size);
 475
 476         spin_lock_irq(&mapping->tree_lock);
 477         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 478         spin_unlock_irq(&mapping->tree_lock);
 479  unmap:
 480         dax_unmap_atomic(bdev, &dax);
 481         return ret;
 482
 483  unlock:
 484         spin_unlock_irq(&mapping->tree_lock);
 485         return ret;
 486 }
 487
 488 /*
 489  * Flush the mapping to the persistent domain within the byte range of [start,
 490  * end]. This is required by data integrity operations to ensure file data is
 491  * on persistent storage prior to completion of the operation.
 492  */
 493 int dax_writeback_mapping_range(struct address_space *mapping,
 494                 struct block_device *bdev, struct writeback_control *wbc)
 495 {
 496         struct inode *inode = mapping->host;
 497         pgoff_t start_index, end_index, pmd_index;
 498         pgoff_t indices[PAGEVEC_SIZE];
 499         struct pagevec pvec;
 500         bool done = false;
 501         int i, ret = 0;
 502         void *entry;
 503
 504         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 505                 return -EIO;
 506
 507         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 508                 return 0;
 509
 510         start_index = wbc->range_start >> PAGE_SHIFT;
 511         end_index = wbc->range_end >> PAGE_SHIFT;
 512         pmd_index = DAX_PMD_INDEX(start_index);
 513
 514         rcu_read_lock();
 515         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 516         rcu_read_unlock();
 517
 518         /* see if the start of our range is covered by a PMD entry */
 519         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 520                 start_index = pmd_index;
 521
 522         tag_pages_for_writeback(mapping, start_index, end_index);
 523
 524         pagevec_init(&pvec, 0);
 525         while (!done) {
 526                 pvec.nr = find_get_entries_tag(mapping, start_index,
 527                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 528                                 pvec.pages, indices);
 529
 530                 if (pvec.nr == 0)
 531                         break;
 532
 533                 for (i = 0; i < pvec.nr; i++) {
 534                         if (indices[i] > end_index) {
 535                                 done = true;
 536                                 break;
 537                         }
 538
 539                         ret = dax_writeback_one(bdev, mapping, indices[i],
 540                                         pvec.pages[i]);
 541                         if (ret < 0)
 542                                 return ret;
 543                 }
 544         }
 545         wmb_pmem();
 546         return 0;
 547 }
 548 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 549
 550 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 551                         struct vm_area_struct *vma, struct vm_fault *vmf)
 552 {
 553         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 554         struct address_space *mapping = inode->i_mapping;
 555         struct block_device *bdev = bh->b_bdev;
 556         struct blk_dax_ctl dax = {
 557                 .sector = to_sector(bh, inode),
 558                 .size = bh->b_size,
 559         };
 560         pgoff_t size;
 561         int error;
 562
 563         i_mmap_lock_read(mapping);
 564
 565         /*
 566          * Check truncate didn't happen while we were allocating a block.
 567          * If it did, this block may or may not be still allocated to the
 568          * file.  We can't tell the filesystem to free it because we can't
 569          * take i_mutex here.  In the worst case, the file still has blocks
 570          * allocated past the end of the file.
 571          */
 572         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 573         if (unlikely(vmf->pgoff >= size)) {
 574                 error = -EIO;
 575                 goto out;
 576         }
 577
 578         if (dax_map_atomic(bdev, &dax) < 0) {
 579                 error = PTR_ERR(dax.addr);
 580                 goto out;
 581         }
 582         dax_unmap_atomic(bdev, &dax);
 583
 584         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 585                         vmf->flags & FAULT_FLAG_WRITE);
 586         if (error)
 587                 goto out;
 588
 589         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 590
 591  out:
 592         i_mmap_unlock_read(mapping);
 593
 594         return error;
 595 }
 596
 597 /**
 598  * __dax_fault - handle a page fault on a DAX file
 599  * @vma: The virtual memory area where the fault occurred
 600  * @vmf: The description of the fault
 601  * @get_block: The filesystem method used to translate file offsets to blocks
 602  *
 603  * When a page fault occurs, filesystems may call this helper in their
 604  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 605  * the necessary locking for the page fault to proceed successfully.
 606  */
 607 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 608                         get_block_t get_block)
 609 {
 610         struct file *file = vma->vm_file;
 611         struct address_space *mapping = file->f_mapping;
 612         struct inode *inode = mapping->host;
 613         struct page *page;
 614         struct buffer_head bh;
 615         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 616         unsigned blkbits = inode->i_blkbits;
 617         sector_t block;
 618         pgoff_t size;
 619         int error;
 620         int major = 0;
 621
 622         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 623         if (vmf->pgoff >= size)
 624                 return VM_FAULT_SIGBUS;
 625
 626         memset(&bh, 0, sizeof(bh));
 627         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 628         bh.b_bdev = inode->i_sb->s_bdev;
 629         bh.b_size = PAGE_SIZE;
 630
 631  repeat:
 632         page = find_get_page(mapping, vmf->pgoff);
 633         if (page) {
 634                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 635                         put_page(page);
 636                         return VM_FAULT_RETRY;
 637                 }
 638                 if (unlikely(page->mapping != mapping)) {
 639                         unlock_page(page);
 640                         put_page(page);
 641                         goto repeat;
 642                 }
 643                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 644                 if (unlikely(vmf->pgoff >= size)) {
 645                         /*
 646                          * We have a struct page covering a hole in the file
 647                          * from a read fault and we've raced with a truncate
 648                          */
 649                         error = -EIO;
 650                         goto unlock_page;
 651                 }
 652         }
 653
 654         error = get_block(inode, block, &bh, 0);
 655         if (!error && (bh.b_size < PAGE_SIZE))
 656                 error = -EIO;           /* fs corruption? */
 657         if (error)
 658                 goto unlock_page;
 659
 660         if (!buffer_mapped(&bh) && !vmf->cow_page) {
 661                 if (vmf->flags & FAULT_FLAG_WRITE) {
 662                         error = get_block(inode, block, &bh, 1);
 663                         count_vm_event(PGMAJFAULT);
 664                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 665                         major = VM_FAULT_MAJOR;
 666                         if (!error && (bh.b_size < PAGE_SIZE))
 667                                 error = -EIO;
 668                         if (error)
 669                                 goto unlock_page;
 670                 } else {
 671                         return dax_load_hole(mapping, page, vmf);
 672                 }
 673         }
 674
 675         if (vmf->cow_page) {
 676                 struct page *new_page = vmf->cow_page;
 677                 if (buffer_written(&bh))
 678                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 679                 else
 680                         clear_user_highpage(new_page, vaddr);
 681                 if (error)
 682                         goto unlock_page;
 683                 vmf->page = page;
 684                 if (!page) {
 685                         i_mmap_lock_read(mapping);
 686                         /* Check we didn't race with truncate */
 687                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 688                                                                 PAGE_SHIFT;
 689                         if (vmf->pgoff >= size) {
 690                                 i_mmap_unlock_read(mapping);
 691                                 error = -EIO;
 692                                 goto out;
 693                         }
 694                 }
 695                 return VM_FAULT_LOCKED;
 696         }
 697
 698         /* Check we didn't race with a read fault installing a new page */
 699         if (!page && major)
 700                 page = find_lock_page(mapping, vmf->pgoff);
 701
 702         if (page) {
 703                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 704                                                         PAGE_SIZE, 0);
 705                 delete_from_page_cache(page);
 706                 unlock_page(page);
 707                 put_page(page);
 708                 page = NULL;
 709         }
 710
 711         /* Filesystem should not return unwritten buffers to us! */
 712         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 713         error = dax_insert_mapping(inode, &bh, vma, vmf);
 714
 715  out:
 716         if (error == -ENOMEM)
 717                 return VM_FAULT_OOM | major;
 718         /* -EBUSY is fine, somebody else faulted on the same PTE */
 719         if ((error < 0) && (error != -EBUSY))
 720                 return VM_FAULT_SIGBUS | major;
 721         return VM_FAULT_NOPAGE | major;
 722
 723  unlock_page:
 724         if (page) {
 725                 unlock_page(page);
 726                 put_page(page);
 727         }
 728         goto out;
 729 }
 730 EXPORT_SYMBOL(__dax_fault);
 731
 732 /**
 733  * dax_fault - handle a page fault on a DAX file
 734  * @vma: The virtual memory area where the fault occurred
 735  * @vmf: The description of the fault
 736  * @get_block: The filesystem method used to translate file offsets to blocks
 737  *
 738  * When a page fault occurs, filesystems may call this helper in their
 739  * fault handler for DAX files.
 740  */
 741 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 742               get_block_t get_block)
 743 {
 744         int result;
 745         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 746
 747         if (vmf->flags & FAULT_FLAG_WRITE) {
 748                 sb_start_pagefault(sb);
 749                 file_update_time(vma->vm_file);
 750         }
 751         result = __dax_fault(vma, vmf, get_block);
 752         if (vmf->flags & FAULT_FLAG_WRITE)
 753                 sb_end_pagefault(sb);
 754
 755         return result;
 756 }
 757 EXPORT_SYMBOL_GPL(dax_fault);
 758
 759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 760 /*
 761  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 762  * more often than one might expect in the below function.
 763  */
 764 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 765
 766 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 767                 const char *reason, const char *fn)
 768 {
 769         if (bh) {
 770                 char bname[BDEVNAME_SIZE];
 771                 bdevname(bh->b_bdev, bname);
 772                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 773                         "length %zd fallback: %s\n", fn, current->comm,
 774                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 775                         bh->b_size, reason);
 776         } else {
 777                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 778                         current->comm, address, reason);
 779         }
 780 }
 781
 782 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 783
 784 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 785                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 786 {
 787         struct file *file = vma->vm_file;
 788         struct address_space *mapping = file->f_mapping;
 789         struct inode *inode = mapping->host;
 790         struct buffer_head bh;
 791         unsigned blkbits = inode->i_blkbits;
 792         unsigned long pmd_addr = address & PMD_MASK;
 793         bool write = flags & FAULT_FLAG_WRITE;
 794         struct block_device *bdev;
 795         pgoff_t size, pgoff;
 796         sector_t block;
 797         int error, result = 0;
 798         bool alloc = false;
 799
 800         /* dax pmd mappings require pfn_t_devmap() */
 801         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 802                 return VM_FAULT_FALLBACK;
 803
 804         /* Fall back to PTEs if we're going to COW */
 805         if (write && !(vma->vm_flags & VM_SHARED)) {
 806                 split_huge_pmd(vma, pmd, address);
 807                 dax_pmd_dbg(NULL, address, "cow write");
 808                 return VM_FAULT_FALLBACK;
 809         }
 810         /* If the PMD would extend outside the VMA */
 811         if (pmd_addr < vma->vm_start) {
 812                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 813                 return VM_FAULT_FALLBACK;
 814         }
 815         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 816                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 817                 return VM_FAULT_FALLBACK;
 818         }
 819
 820         pgoff = linear_page_index(vma, pmd_addr);
 821         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 822         if (pgoff >= size)
 823                 return VM_FAULT_SIGBUS;
 824         /* If the PMD would cover blocks out of the file */
 825         if ((pgoff | PG_PMD_COLOUR) >= size) {
 826                 dax_pmd_dbg(NULL, address,
 827                                 "offset + huge page size > file size");
 828                 return VM_FAULT_FALLBACK;
 829         }
 830
 831         memset(&bh, 0, sizeof(bh));
 832         bh.b_bdev = inode->i_sb->s_bdev;
 833         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 834
 835         bh.b_size = PMD_SIZE;
 836
 837         if (get_block(inode, block, &bh, 0) != 0)
 838                 return VM_FAULT_SIGBUS;
 839
 840         if (!buffer_mapped(&bh) && write) {
 841                 if (get_block(inode, block, &bh, 1) != 0)
 842                         return VM_FAULT_SIGBUS;
 843                 alloc = true;
 844                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 845         }
 846
 847         bdev = bh.b_bdev;
 848
 849         /*
 850          * If the filesystem isn't willing to tell us the length of a hole,
 851          * just fall back to PTEs.  Calling get_block 512 times in a loop
 852          * would be silly.
 853          */
 854         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 855                 dax_pmd_dbg(&bh, address, "allocated block too small");
 856                 return VM_FAULT_FALLBACK;
 857         }
 858
 859         /*
 860          * If we allocated new storage, make sure no process has any
 861          * zero pages covering this hole
 862          */
 863         if (alloc) {
 864                 loff_t lstart = pgoff << PAGE_SHIFT;
 865                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 866
 867                 truncate_pagecache_range(inode, lstart, lend);
 868         }
 869
 870         i_mmap_lock_read(mapping);
 871
 872         /*
 873          * If a truncate happened while we were allocating blocks, we may
 874          * leave blocks allocated to the file that are beyond EOF.  We can't
 875          * take i_mutex here, so just leave them hanging; they'll be freed
 876          * when the file is deleted.
 877          */
 878         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 879         if (pgoff >= size) {
 880                 result = VM_FAULT_SIGBUS;
 881                 goto out;
 882         }
 883         if ((pgoff | PG_PMD_COLOUR) >= size) {
 884                 dax_pmd_dbg(&bh, address,
 885                                 "offset + huge page size > file size");
 886                 goto fallback;
 887         }
 888
 889         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 890                 spinlock_t *ptl;
 891                 pmd_t entry;
 892                 struct page *zero_page = get_huge_zero_page();
 893
 894                 if (unlikely(!zero_page)) {
 895                         dax_pmd_dbg(&bh, address, "no zero page");
 896                         goto fallback;
 897                 }
 898
 899                 ptl = pmd_lock(vma->vm_mm, pmd);
 900                 if (!pmd_none(*pmd)) {
 901                         spin_unlock(ptl);
 902                         dax_pmd_dbg(&bh, address, "pmd already present");
 903                         goto fallback;
 904                 }
 905
 906                 dev_dbg(part_to_dev(bdev->bd_part),
 907                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 908                                 __func__, current->comm, address,
 909                                 (unsigned long long) to_sector(&bh, inode));
 910
 911                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 912                 entry = pmd_mkhuge(entry);
 913                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 914                 result = VM_FAULT_NOPAGE;
 915                 spin_unlock(ptl);
 916         } else {
 917                 struct blk_dax_ctl dax = {
 918                         .sector = to_sector(&bh, inode),
 919                         .size = PMD_SIZE,
 920                 };
 921                 long length = dax_map_atomic(bdev, &dax);
 922
 923                 if (length < 0) {
 924                         result = VM_FAULT_SIGBUS;
 925                         goto out;
 926                 }
 927                 if (length < PMD_SIZE) {
 928                         dax_pmd_dbg(&bh, address, "dax-length too small");
 929                         dax_unmap_atomic(bdev, &dax);
 930                         goto fallback;
 931                 }
 932                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 933                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 934                         dax_unmap_atomic(bdev, &dax);
 935                         goto fallback;
 936                 }
 937
 938                 if (!pfn_t_devmap(dax.pfn)) {
 939                         dax_unmap_atomic(bdev, &dax);
 940                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 941                         goto fallback;
 942                 }
 943                 dax_unmap_atomic(bdev, &dax);
 944
 945                 /*
 946                  * For PTE faults we insert a radix tree entry for reads, and
 947                  * leave it clean.  Then on the first write we dirty the radix
 948                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 949                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 950                  * call into get_block() to translate the pgoff to a sector in
 951                  * order to be able to create a new radix tree entry.
 952                  *
 953                  * The PMD path doesn't have an equivalent to
 954                  * dax_pfn_mkwrite(), though, so for a read followed by a
 955                  * write we traverse all the way through __dax_pmd_fault()
 956                  * twice.  This means we can just skip inserting a radix tree
 957                  * entry completely on the initial read and just wait until
 958                  * the write to insert a dirty entry.
 959                  */
 960                 if (write) {
 961                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 962                                         true, true);
 963                         if (error) {
 964                                 dax_pmd_dbg(&bh, address,
 965                                                 "PMD radix insertion failed");
 966                                 goto fallback;
 967                         }
 968                 }
 969
 970                 dev_dbg(part_to_dev(bdev->bd_part),
 971                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 972                                 __func__, current->comm, address,
 973                                 pfn_t_to_pfn(dax.pfn),
 974                                 (unsigned long long) dax.sector);
 975                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
 976                                 dax.pfn, write);
 977         }
 978
 979  out:
 980         i_mmap_unlock_read(mapping);
 981
 982         return result;
 983
 984  fallback:
 985         count_vm_event(THP_FAULT_FALLBACK);
 986         result = VM_FAULT_FALLBACK;
 987         goto out;
 988 }
 989 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
 990
 991 /**
 992  * dax_pmd_fault - handle a PMD fault on a DAX file
 993  * @vma: The virtual memory area where the fault occurred
 994  * @vmf: The description of the fault
 995  * @get_block: The filesystem method used to translate file offsets to blocks
 996  *
 997  * When a page fault occurs, filesystems may call this helper in their
 998  * pmd_fault handler for DAX files.
 999  */
1000 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1001                         pmd_t *pmd, unsigned int flags, get_block_t get_block)
1002 {
1003         int result;
1004         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1005
1006         if (flags & FAULT_FLAG_WRITE) {
1007                 sb_start_pagefault(sb);
1008                 file_update_time(vma->vm_file);
1009         }
1010         result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
1011         if (flags & FAULT_FLAG_WRITE)
1012                 sb_end_pagefault(sb);
1013
1014         return result;
1015 }
1016 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1017 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1018
1019 /**
1020  * dax_pfn_mkwrite - handle first write to DAX page
1021  * @vma: The virtual memory area where the fault occurred
1022  * @vmf: The description of the fault
1023  */
1024 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1025 {
1026         struct file *file = vma->vm_file;
1027         int error;
1028
1029         /*
1030          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1031          * RADIX_DAX_PTE entry already exists in the radix tree from a
1032          * previous call to __dax_fault().  We just want to look up that PTE
1033          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1034          * saves us from having to make a call to get_block() here to look
1035          * up the sector.
1036          */
1037         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1038                         true);
1039
1040         if (error == -ENOMEM)
1041                 return VM_FAULT_OOM;
1042         if (error)
1043                 return VM_FAULT_SIGBUS;
1044         return VM_FAULT_NOPAGE;
1045 }
1046 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1047
1048 /**
1049  * dax_zero_page_range - zero a range within a page of a DAX file
1050  * @inode: The file being truncated
1051  * @from: The file offset that is being truncated to
1052  * @length: The number of bytes to zero
1053  * @get_block: The filesystem method used to translate file offsets to blocks
1054  *
1055  * This function can be called by a filesystem when it is zeroing part of a
1056  * page in a DAX file.  This is intended for hole-punch operations.  If
1057  * you are truncating a file, the helper function dax_truncate_page() may be
1058  * more convenient.
1059  *
1060  * We work in terms of PAGE_SIZE here for commonality with
1061  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1062  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1063  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1064  * since the file might be mmapped.
1065  */
1066 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1067                                                         get_block_t get_block)
1068 {
1069         struct buffer_head bh;
1070         pgoff_t index = from >> PAGE_SHIFT;
1071         unsigned offset = from & (PAGE_SIZE-1);
1072         int err;
1073
1074         /* Block boundary? Nothing to do */
1075         if (!length)
1076                 return 0;
1077         BUG_ON((offset + length) > PAGE_SIZE);
1078
1079         memset(&bh, 0, sizeof(bh));
1080         bh.b_bdev = inode->i_sb->s_bdev;
1081         bh.b_size = PAGE_SIZE;
1082         err = get_block(inode, index, &bh, 0);
1083         if (err < 0)
1084                 return err;
1085         if (buffer_written(&bh)) {
1086                 struct block_device *bdev = bh.b_bdev;
1087                 struct blk_dax_ctl dax = {
1088                         .sector = to_sector(&bh, inode),
1089                         .size = PAGE_SIZE,
1090                 };
1091
1092                 if (dax_map_atomic(bdev, &dax) < 0)
1093                         return PTR_ERR(dax.addr);
1094                 clear_pmem(dax.addr + offset, length);
1095                 wmb_pmem();
1096                 dax_unmap_atomic(bdev, &dax);
1097         }
1098
1099         return 0;
1100 }
1101 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1102
1103 /**
1104  * dax_truncate_page - handle a partial page being truncated in a DAX file
1105  * @inode: The file being truncated
1106  * @from: The file offset that is being truncated to
1107  * @get_block: The filesystem method used to translate file offsets to blocks
1108  *
1109  * Similar to block_truncate_page(), this function can be called by a
1110  * filesystem when it is truncating a DAX file to handle the partial page.
1111  *
1112  * We work in terms of PAGE_SIZE here for commonality with
1113  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1114  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1115  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1116  * since the file might be mmapped.
1117  */
1118 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1119 {
1120         unsigned length = PAGE_ALIGN(from) - from;
1121         return dax_zero_page_range(inode, from, length, get_block);
1122 }
1123 EXPORT_SYMBOL_GPL(dax_truncate_page);