fs/dax.c

   1 /*
   2  * fs/dax.c - Direct Access filesystem code
   3  * Copyright (c) 2013-2014 Intel Corporation
   4  * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   5  * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  */
  16
  17 #include <linux/atomic.h>
  18 #include <linux/blkdev.h>
  19 #include <linux/buffer_head.h>
  20 #include <linux/dax.h>
  21 #include <linux/fs.h>
  22 #include <linux/genhd.h>
  23 #include <linux/highmem.h>
  24 #include <linux/memcontrol.h>
  25 #include <linux/mm.h>
  26 #include <linux/mutex.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/pmem.h>
  29 #include <linux/sched.h>
  30 #include <linux/uio.h>
  31 #include <linux/vmstat.h>
  32 #include <linux/pfn_t.h>
  33 #include <linux/sizes.h>
  34
  35 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  36 {
  37         struct request_queue *q = bdev->bd_queue;
  38         long rc = -EIO;
  39
  40         dax->addr = (void __pmem *) ERR_PTR(-EIO);
  41         if (blk_queue_enter(q, true) != 0)
  42                 return rc;
  43
  44         rc = bdev_direct_access(bdev, dax);
  45         if (rc < 0) {
  46                 dax->addr = (void __pmem *) ERR_PTR(rc);
  47                 blk_queue_exit(q);
  48                 return rc;
  49         }
  50         return rc;
  51 }
  52
  53 static void dax_unmap_atomic(struct block_device *bdev,
  54                 const struct blk_dax_ctl *dax)
  55 {
  56         if (IS_ERR(dax->addr))
  57                 return;
  58         blk_queue_exit(bdev->bd_queue);
  59 }
  60
  61 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
  62 {
  63         struct page *page = alloc_pages(GFP_KERNEL, 0);
  64         struct blk_dax_ctl dax = {
  65                 .size = PAGE_SIZE,
  66                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
  67         };
  68         long rc;
  69
  70         if (!page)
  71                 return ERR_PTR(-ENOMEM);
  72
  73         rc = dax_map_atomic(bdev, &dax);
  74         if (rc < 0)
  75                 return ERR_PTR(rc);
  76         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
  77         dax_unmap_atomic(bdev, &dax);
  78         return page;
  79 }
  80
  81 /*
  82  * dax_clear_sectors() is called from within transaction context from XFS,
  83  * and hence this means the stack from this point must follow GFP_NOFS
  84  * semantics for all operations.
  85  */
  86 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
  87 {
  88         struct blk_dax_ctl dax = {
  89                 .sector = _sector,
  90                 .size = _size,
  91         };
  92
  93         might_sleep();
  94         do {
  95                 long count, sz;
  96
  97                 count = dax_map_atomic(bdev, &dax);
  98                 if (count < 0)
  99                         return count;
 100                 sz = min_t(long, count, SZ_128K);
 101                 clear_pmem(dax.addr, sz);
 102                 dax.size -= sz;
 103                 dax.sector += sz / 512;
 104                 dax_unmap_atomic(bdev, &dax);
 105                 cond_resched();
 106         } while (dax.size);
 107
 108         wmb_pmem();
 109         return 0;
 110 }
 111 EXPORT_SYMBOL_GPL(dax_clear_sectors);
 112
 113 /* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
 114 static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
 115                 loff_t pos, loff_t end)
 116 {
 117         loff_t final = end - pos + first; /* The final byte of the buffer */
 118
 119         if (first > 0)
 120                 clear_pmem(addr, first);
 121         if (final < size)
 122                 clear_pmem(addr + final, size - final);
 123 }
 124
 125 static bool buffer_written(struct buffer_head *bh)
 126 {
 127         return buffer_mapped(bh) && !buffer_unwritten(bh);
 128 }
 129
 130 /*
 131  * When ext4 encounters a hole, it returns without modifying the buffer_head
 132  * which means that we can't trust b_size.  To cope with this, we set b_state
 133  * to 0 before calling get_block and, if any bit is set, we know we can trust
 134  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 135  * and would save us time calling get_block repeatedly.
 136  */
 137 static bool buffer_size_valid(struct buffer_head *bh)
 138 {
 139         return bh->b_state != 0;
 140 }
 141
 142
 143 static sector_t to_sector(const struct buffer_head *bh,
 144                 const struct inode *inode)
 145 {
 146         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 147
 148         return sector;
 149 }
 150
 151 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 152                       loff_t start, loff_t end, get_block_t get_block,
 153                       struct buffer_head *bh)
 154 {
 155         loff_t pos = start, max = start, bh_max = start;
 156         bool hole = false, need_wmb = false;
 157         struct block_device *bdev = NULL;
 158         int rw = iov_iter_rw(iter), rc;
 159         long map_len = 0;
 160         struct blk_dax_ctl dax = {
 161                 .addr = (void __pmem *) ERR_PTR(-EIO),
 162         };
 163
 164         if (rw == READ)
 165                 end = min(end, i_size_read(inode));
 166
 167         while (pos < end) {
 168                 size_t len;
 169                 if (pos == max) {
 170                         unsigned blkbits = inode->i_blkbits;
 171                         long page = pos >> PAGE_SHIFT;
 172                         sector_t block = page << (PAGE_SHIFT - blkbits);
 173                         unsigned first = pos - (block << blkbits);
 174                         long size;
 175
 176                         if (pos == bh_max) {
 177                                 bh->b_size = PAGE_ALIGN(end - pos);
 178                                 bh->b_state = 0;
 179                                 rc = get_block(inode, block, bh, rw == WRITE);
 180                                 if (rc)
 181                                         break;
 182                                 if (!buffer_size_valid(bh))
 183                                         bh->b_size = 1 << blkbits;
 184                                 bh_max = pos - first + bh->b_size;
 185                                 bdev = bh->b_bdev;
 186                         } else {
 187                                 unsigned done = bh->b_size -
 188                                                 (bh_max - (pos - first));
 189                                 bh->b_blocknr += done >> blkbits;
 190                                 bh->b_size -= done;
 191                         }
 192
 193                         hole = rw == READ && !buffer_written(bh);
 194                         if (hole) {
 195                                 size = bh->b_size - first;
 196                         } else {
 197                                 dax_unmap_atomic(bdev, &dax);
 198                                 dax.sector = to_sector(bh, inode);
 199                                 dax.size = bh->b_size;
 200                                 map_len = dax_map_atomic(bdev, &dax);
 201                                 if (map_len < 0) {
 202                                         rc = map_len;
 203                                         break;
 204                                 }
 205                                 if (buffer_unwritten(bh) || buffer_new(bh)) {
 206                                         dax_new_buf(dax.addr, map_len, first,
 207                                                         pos, end);
 208                                         need_wmb = true;
 209                                 }
 210                                 dax.addr += first;
 211                                 size = map_len - first;
 212                         }
 213                         max = min(pos + size, end);
 214                 }
 215
 216                 if (iov_iter_rw(iter) == WRITE) {
 217                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 218                         need_wmb = true;
 219                 } else if (!hole)
 220                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 221                                         iter);
 222                 else
 223                         len = iov_iter_zero(max - pos, iter);
 224
 225                 if (!len) {
 226                         rc = -EFAULT;
 227                         break;
 228                 }
 229
 230                 pos += len;
 231                 if (!IS_ERR(dax.addr))
 232                         dax.addr += len;
 233         }
 234
 235         if (need_wmb)
 236                 wmb_pmem();
 237         dax_unmap_atomic(bdev, &dax);
 238
 239         return (pos == start) ? rc : pos - start;
 240 }
 241
 242 /**
 243  * dax_do_io - Perform I/O to a DAX file
 244  * @iocb: The control block for this I/O
 245  * @inode: The file which the I/O is directed at
 246  * @iter: The addresses to do I/O from or to
 247  * @get_block: The filesystem method used to translate file offsets to blocks
 248  * @end_io: A filesystem callback for I/O completion
 249  * @flags: See below
 250  *
 251  * This function uses the same locking scheme as do_blockdev_direct_IO:
 252  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 253  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 254  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 255  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 256  * is in progress.
 257  */
 258 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 259                   struct iov_iter *iter, get_block_t get_block,
 260                   dio_iodone_t end_io, int flags)
 261 {
 262         struct buffer_head bh;
 263         ssize_t retval = -EINVAL;
 264         loff_t pos = iocb->ki_pos;
 265         loff_t end = pos + iov_iter_count(iter);
 266
 267         memset(&bh, 0, sizeof(bh));
 268         bh.b_bdev = inode->i_sb->s_bdev;
 269
 270         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
 271                 struct address_space *mapping = inode->i_mapping;
 272                 inode_lock(inode);
 273                 retval = filemap_write_and_wait_range(mapping, pos, end - 1);
 274                 if (retval) {
 275                         inode_unlock(inode);
 276                         goto out;
 277                 }
 278         }
 279
 280         /* Protects against truncate */
 281         if (!(flags & DIO_SKIP_DIO_COUNT))
 282                 inode_dio_begin(inode);
 283
 284         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 285
 286         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 287                 inode_unlock(inode);
 288
 289         if (end_io) {
 290                 int err;
 291
 292                 err = end_io(iocb, pos, retval, bh.b_private);
 293                 if (err)
 294                         retval = err;
 295         }
 296
 297         if (!(flags & DIO_SKIP_DIO_COUNT))
 298                 inode_dio_end(inode);
 299  out:
 300         return retval;
 301 }
 302 EXPORT_SYMBOL_GPL(dax_do_io);
 303
 304 /*
 305  * The user has performed a load from a hole in the file.  Allocating
 306  * a new page in the file would cause excessive storage usage for
 307  * workloads with sparse files.  We allocate a page cache page instead.
 308  * We'll kick it out of the page cache if it's ever written to,
 309  * otherwise it will simply fall out of the page cache under memory
 310  * pressure without ever having been dirtied.
 311  */
 312 static int dax_load_hole(struct address_space *mapping, struct page *page,
 313                                                         struct vm_fault *vmf)
 314 {
 315         unsigned long size;
 316         struct inode *inode = mapping->host;
 317         if (!page)
 318                 page = find_or_create_page(mapping, vmf->pgoff,
 319                                                 GFP_KERNEL | __GFP_ZERO);
 320         if (!page)
 321                 return VM_FAULT_OOM;
 322         /* Recheck i_size under page lock to avoid truncate race */
 323         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 324         if (vmf->pgoff >= size) {
 325                 unlock_page(page);
 326                 put_page(page);
 327                 return VM_FAULT_SIGBUS;
 328         }
 329
 330         vmf->page = page;
 331         return VM_FAULT_LOCKED;
 332 }
 333
 334 static int copy_user_bh(struct page *to, struct inode *inode,
 335                 struct buffer_head *bh, unsigned long vaddr)
 336 {
 337         struct blk_dax_ctl dax = {
 338                 .sector = to_sector(bh, inode),
 339                 .size = bh->b_size,
 340         };
 341         struct block_device *bdev = bh->b_bdev;
 342         void *vto;
 343
 344         if (dax_map_atomic(bdev, &dax) < 0)
 345                 return PTR_ERR(dax.addr);
 346         vto = kmap_atomic(to);
 347         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 348         kunmap_atomic(vto);
 349         dax_unmap_atomic(bdev, &dax);
 350         return 0;
 351 }
 352
 353 #define NO_SECTOR -1
 354 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 355
 356 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 357                 sector_t sector, bool pmd_entry, bool dirty)
 358 {
 359         struct radix_tree_root *page_tree = &mapping->page_tree;
 360         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 361         int type, error = 0;
 362         void *entry;
 363
 364         WARN_ON_ONCE(pmd_entry && !dirty);
 365         if (dirty)
 366                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 367
 368         spin_lock_irq(&mapping->tree_lock);
 369
 370         entry = radix_tree_lookup(page_tree, pmd_index);
 371         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 372                 index = pmd_index;
 373                 goto dirty;
 374         }
 375
 376         entry = radix_tree_lookup(page_tree, index);
 377         if (entry) {
 378                 type = RADIX_DAX_TYPE(entry);
 379                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 380                                         type != RADIX_DAX_PMD)) {
 381                         error = -EIO;
 382                         goto unlock;
 383                 }
 384
 385                 if (!pmd_entry || type == RADIX_DAX_PMD)
 386                         goto dirty;
 387
 388                 /*
 389                  * We only insert dirty PMD entries into the radix tree.  This
 390                  * means we don't need to worry about removing a dirty PTE
 391                  * entry and inserting a clean PMD entry, thus reducing the
 392                  * range we would flush with a follow-up fsync/msync call.
 393                  */
 394                 radix_tree_delete(&mapping->page_tree, index);
 395                 mapping->nrexceptional--;
 396         }
 397
 398         if (sector == NO_SECTOR) {
 399                 /*
 400                  * This can happen during correct operation if our pfn_mkwrite
 401                  * fault raced against a hole punch operation.  If this
 402                  * happens the pte that was hole punched will have been
 403                  * unmapped and the radix tree entry will have been removed by
 404                  * the time we are called, but the call will still happen.  We
 405                  * will return all the way up to wp_pfn_shared(), where the
 406                  * pte_same() check will fail, eventually causing page fault
 407                  * to be retried by the CPU.
 408                  */
 409                 goto unlock;
 410         }
 411
 412         error = radix_tree_insert(page_tree, index,
 413                         RADIX_DAX_ENTRY(sector, pmd_entry));
 414         if (error)
 415                 goto unlock;
 416
 417         mapping->nrexceptional++;
 418  dirty:
 419         if (dirty)
 420                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 421  unlock:
 422         spin_unlock_irq(&mapping->tree_lock);
 423         return error;
 424 }
 425
 426 static int dax_writeback_one(struct block_device *bdev,
 427                 struct address_space *mapping, pgoff_t index, void *entry)
 428 {
 429         struct radix_tree_root *page_tree = &mapping->page_tree;
 430         int type = RADIX_DAX_TYPE(entry);
 431         struct radix_tree_node *node;
 432         struct blk_dax_ctl dax;
 433         void **slot;
 434         int ret = 0;
 435
 436         spin_lock_irq(&mapping->tree_lock);
 437         /*
 438          * Regular page slots are stabilized by the page lock even
 439          * without the tree itself locked.  These unlocked entries
 440          * need verification under the tree lock.
 441          */
 442         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 443                 goto unlock;
 444         if (*slot != entry)
 445                 goto unlock;
 446
 447         /* another fsync thread may have already written back this entry */
 448         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 449                 goto unlock;
 450
 451         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 452                 ret = -EIO;
 453                 goto unlock;
 454         }
 455
 456         dax.sector = RADIX_DAX_SECTOR(entry);
 457         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 458         spin_unlock_irq(&mapping->tree_lock);
 459
 460         /*
 461          * We cannot hold tree_lock while calling dax_map_atomic() because it
 462          * eventually calls cond_resched().
 463          */
 464         ret = dax_map_atomic(bdev, &dax);
 465         if (ret < 0)
 466                 return ret;
 467
 468         if (WARN_ON_ONCE(ret < dax.size)) {
 469                 ret = -EIO;
 470                 goto unmap;
 471         }
 472
 473         wb_cache_pmem(dax.addr, dax.size);
 474
 475         spin_lock_irq(&mapping->tree_lock);
 476         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 477         spin_unlock_irq(&mapping->tree_lock);
 478  unmap:
 479         dax_unmap_atomic(bdev, &dax);
 480         return ret;
 481
 482  unlock:
 483         spin_unlock_irq(&mapping->tree_lock);
 484         return ret;
 485 }
 486
 487 /*
 488  * Flush the mapping to the persistent domain within the byte range of [start,
 489  * end]. This is required by data integrity operations to ensure file data is
 490  * on persistent storage prior to completion of the operation.
 491  */
 492 int dax_writeback_mapping_range(struct address_space *mapping,
 493                 struct block_device *bdev, struct writeback_control *wbc)
 494 {
 495         struct inode *inode = mapping->host;
 496         pgoff_t start_index, end_index, pmd_index;
 497         pgoff_t indices[PAGEVEC_SIZE];
 498         struct pagevec pvec;
 499         bool done = false;
 500         int i, ret = 0;
 501         void *entry;
 502
 503         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 504                 return -EIO;
 505
 506         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 507                 return 0;
 508
 509         start_index = wbc->range_start >> PAGE_SHIFT;
 510         end_index = wbc->range_end >> PAGE_SHIFT;
 511         pmd_index = DAX_PMD_INDEX(start_index);
 512
 513         rcu_read_lock();
 514         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 515         rcu_read_unlock();
 516
 517         /* see if the start of our range is covered by a PMD entry */
 518         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 519                 start_index = pmd_index;
 520
 521         tag_pages_for_writeback(mapping, start_index, end_index);
 522
 523         pagevec_init(&pvec, 0);
 524         while (!done) {
 525                 pvec.nr = find_get_entries_tag(mapping, start_index,
 526                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 527                                 pvec.pages, indices);
 528
 529                 if (pvec.nr == 0)
 530                         break;
 531
 532                 for (i = 0; i < pvec.nr; i++) {
 533                         if (indices[i] > end_index) {
 534                                 done = true;
 535                                 break;
 536                         }
 537
 538                         ret = dax_writeback_one(bdev, mapping, indices[i],
 539                                         pvec.pages[i]);
 540                         if (ret < 0)
 541                                 return ret;
 542                 }
 543         }
 544         wmb_pmem();
 545         return 0;
 546 }
 547 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 548
 549 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 550                         struct vm_area_struct *vma, struct vm_fault *vmf)
 551 {
 552         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 553         struct address_space *mapping = inode->i_mapping;
 554         struct block_device *bdev = bh->b_bdev;
 555         struct blk_dax_ctl dax = {
 556                 .sector = to_sector(bh, inode),
 557                 .size = bh->b_size,
 558         };
 559         pgoff_t size;
 560         int error;
 561
 562         i_mmap_lock_read(mapping);
 563
 564         /*
 565          * Check truncate didn't happen while we were allocating a block.
 566          * If it did, this block may or may not be still allocated to the
 567          * file.  We can't tell the filesystem to free it because we can't
 568          * take i_mutex here.  In the worst case, the file still has blocks
 569          * allocated past the end of the file.
 570          */
 571         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 572         if (unlikely(vmf->pgoff >= size)) {
 573                 error = -EIO;
 574                 goto out;
 575         }
 576
 577         if (dax_map_atomic(bdev, &dax) < 0) {
 578                 error = PTR_ERR(dax.addr);
 579                 goto out;
 580         }
 581
 582         if (buffer_unwritten(bh) || buffer_new(bh)) {
 583                 clear_pmem(dax.addr, PAGE_SIZE);
 584                 wmb_pmem();
 585         }
 586         dax_unmap_atomic(bdev, &dax);
 587
 588         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 589                         vmf->flags & FAULT_FLAG_WRITE);
 590         if (error)
 591                 goto out;
 592
 593         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 594
 595  out:
 596         i_mmap_unlock_read(mapping);
 597
 598         return error;
 599 }
 600
 601 /**
 602  * __dax_fault - handle a page fault on a DAX file
 603  * @vma: The virtual memory area where the fault occurred
 604  * @vmf: The description of the fault
 605  * @get_block: The filesystem method used to translate file offsets to blocks
 606  * @complete_unwritten: The filesystem method used to convert unwritten blocks
 607  *      to written so the data written to them is exposed. This is required for
 608  *      required by write faults for filesystems that will return unwritten
 609  *      extent mappings from @get_block, but it is optional for reads as
 610  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
 611  *      not support unwritten extents, the it should pass NULL.
 612  *
 613  * When a page fault occurs, filesystems may call this helper in their
 614  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 615  * the necessary locking for the page fault to proceed successfully.
 616  */
 617 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 618                         get_block_t get_block, dax_iodone_t complete_unwritten)
 619 {
 620         struct file *file = vma->vm_file;
 621         struct address_space *mapping = file->f_mapping;
 622         struct inode *inode = mapping->host;
 623         struct page *page;
 624         struct buffer_head bh;
 625         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 626         unsigned blkbits = inode->i_blkbits;
 627         sector_t block;
 628         pgoff_t size;
 629         int error;
 630         int major = 0;
 631
 632         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 633         if (vmf->pgoff >= size)
 634                 return VM_FAULT_SIGBUS;
 635
 636         memset(&bh, 0, sizeof(bh));
 637         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 638         bh.b_bdev = inode->i_sb->s_bdev;
 639         bh.b_size = PAGE_SIZE;
 640
 641  repeat:
 642         page = find_get_page(mapping, vmf->pgoff);
 643         if (page) {
 644                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 645                         put_page(page);
 646                         return VM_FAULT_RETRY;
 647                 }
 648                 if (unlikely(page->mapping != mapping)) {
 649                         unlock_page(page);
 650                         put_page(page);
 651                         goto repeat;
 652                 }
 653                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 654                 if (unlikely(vmf->pgoff >= size)) {
 655                         /*
 656                          * We have a struct page covering a hole in the file
 657                          * from a read fault and we've raced with a truncate
 658                          */
 659                         error = -EIO;
 660                         goto unlock_page;
 661                 }
 662         }
 663
 664         error = get_block(inode, block, &bh, 0);
 665         if (!error && (bh.b_size < PAGE_SIZE))
 666                 error = -EIO;           /* fs corruption? */
 667         if (error)
 668                 goto unlock_page;
 669
 670         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 671                 if (vmf->flags & FAULT_FLAG_WRITE) {
 672                         error = get_block(inode, block, &bh, 1);
 673                         count_vm_event(PGMAJFAULT);
 674                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 675                         major = VM_FAULT_MAJOR;
 676                         if (!error && (bh.b_size < PAGE_SIZE))
 677                                 error = -EIO;
 678                         if (error)
 679                                 goto unlock_page;
 680                 } else {
 681                         return dax_load_hole(mapping, page, vmf);
 682                 }
 683         }
 684
 685         if (vmf->cow_page) {
 686                 struct page *new_page = vmf->cow_page;
 687                 if (buffer_written(&bh))
 688                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 689                 else
 690                         clear_user_highpage(new_page, vaddr);
 691                 if (error)
 692                         goto unlock_page;
 693                 vmf->page = page;
 694                 if (!page) {
 695                         i_mmap_lock_read(mapping);
 696                         /* Check we didn't race with truncate */
 697                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 698                                                                 PAGE_SHIFT;
 699                         if (vmf->pgoff >= size) {
 700                                 i_mmap_unlock_read(mapping);
 701                                 error = -EIO;
 702                                 goto out;
 703                         }
 704                 }
 705                 return VM_FAULT_LOCKED;
 706         }
 707
 708         /* Check we didn't race with a read fault installing a new page */
 709         if (!page && major)
 710                 page = find_lock_page(mapping, vmf->pgoff);
 711
 712         if (page) {
 713                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 714                                                         PAGE_SIZE, 0);
 715                 delete_from_page_cache(page);
 716                 unlock_page(page);
 717                 put_page(page);
 718                 page = NULL;
 719         }
 720
 721         /*
 722          * If we successfully insert the new mapping over an unwritten extent,
 723          * we need to ensure we convert the unwritten extent. If there is an
 724          * error inserting the mapping, the filesystem needs to leave it as
 725          * unwritten to prevent exposure of the stale underlying data to
 726          * userspace, but we still need to call the completion function so
 727          * the private resources on the mapping buffer can be released. We
 728          * indicate what the callback should do via the uptodate variable, same
 729          * as for normal BH based IO completions.
 730          */
 731         error = dax_insert_mapping(inode, &bh, vma, vmf);
 732         if (buffer_unwritten(&bh)) {
 733                 if (complete_unwritten)
 734                         complete_unwritten(&bh, !error);
 735                 else
 736                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 737         }
 738
 739  out:
 740         if (error == -ENOMEM)
 741                 return VM_FAULT_OOM | major;
 742         /* -EBUSY is fine, somebody else faulted on the same PTE */
 743         if ((error < 0) && (error != -EBUSY))
 744                 return VM_FAULT_SIGBUS | major;
 745         return VM_FAULT_NOPAGE | major;
 746
 747  unlock_page:
 748         if (page) {
 749                 unlock_page(page);
 750                 put_page(page);
 751         }
 752         goto out;
 753 }
 754 EXPORT_SYMBOL(__dax_fault);
 755
 756 /**
 757  * dax_fault - handle a page fault on a DAX file
 758  * @vma: The virtual memory area where the fault occurred
 759  * @vmf: The description of the fault
 760  * @get_block: The filesystem method used to translate file offsets to blocks
 761  *
 762  * When a page fault occurs, filesystems may call this helper in their
 763  * fault handler for DAX files.
 764  */
 765 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 766               get_block_t get_block, dax_iodone_t complete_unwritten)
 767 {
 768         int result;
 769         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 770
 771         if (vmf->flags & FAULT_FLAG_WRITE) {
 772                 sb_start_pagefault(sb);
 773                 file_update_time(vma->vm_file);
 774         }
 775         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 776         if (vmf->flags & FAULT_FLAG_WRITE)
 777                 sb_end_pagefault(sb);
 778
 779         return result;
 780 }
 781 EXPORT_SYMBOL_GPL(dax_fault);
 782
 783 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 784 /*
 785  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 786  * more often than one might expect in the below function.
 787  */
 788 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 789
 790 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 791                 const char *reason, const char *fn)
 792 {
 793         if (bh) {
 794                 char bname[BDEVNAME_SIZE];
 795                 bdevname(bh->b_bdev, bname);
 796                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 797                         "length %zd fallback: %s\n", fn, current->comm,
 798                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 799                         bh->b_size, reason);
 800         } else {
 801                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 802                         current->comm, address, reason);
 803         }
 804 }
 805
 806 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 807
 808 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 809                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
 810                 dax_iodone_t complete_unwritten)
 811 {
 812         struct file *file = vma->vm_file;
 813         struct address_space *mapping = file->f_mapping;
 814         struct inode *inode = mapping->host;
 815         struct buffer_head bh;
 816         unsigned blkbits = inode->i_blkbits;
 817         unsigned long pmd_addr = address & PMD_MASK;
 818         bool write = flags & FAULT_FLAG_WRITE;
 819         struct block_device *bdev;
 820         pgoff_t size, pgoff;
 821         sector_t block;
 822         int error, result = 0;
 823         bool alloc = false;
 824
 825         /* dax pmd mappings require pfn_t_devmap() */
 826         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 827                 return VM_FAULT_FALLBACK;
 828
 829         /* Fall back to PTEs if we're going to COW */
 830         if (write && !(vma->vm_flags & VM_SHARED)) {
 831                 split_huge_pmd(vma, pmd, address);
 832                 dax_pmd_dbg(NULL, address, "cow write");
 833                 return VM_FAULT_FALLBACK;
 834         }
 835         /* If the PMD would extend outside the VMA */
 836         if (pmd_addr < vma->vm_start) {
 837                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 838                 return VM_FAULT_FALLBACK;
 839         }
 840         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 841                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 842                 return VM_FAULT_FALLBACK;
 843         }
 844
 845         pgoff = linear_page_index(vma, pmd_addr);
 846         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 847         if (pgoff >= size)
 848                 return VM_FAULT_SIGBUS;
 849         /* If the PMD would cover blocks out of the file */
 850         if ((pgoff | PG_PMD_COLOUR) >= size) {
 851                 dax_pmd_dbg(NULL, address,
 852                                 "offset + huge page size > file size");
 853                 return VM_FAULT_FALLBACK;
 854         }
 855
 856         memset(&bh, 0, sizeof(bh));
 857         bh.b_bdev = inode->i_sb->s_bdev;
 858         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 859
 860         bh.b_size = PMD_SIZE;
 861
 862         if (get_block(inode, block, &bh, 0) != 0)
 863                 return VM_FAULT_SIGBUS;
 864
 865         if (!buffer_mapped(&bh) && write) {
 866                 if (get_block(inode, block, &bh, 1) != 0)
 867                         return VM_FAULT_SIGBUS;
 868                 alloc = true;
 869         }
 870
 871         bdev = bh.b_bdev;
 872
 873         /*
 874          * If the filesystem isn't willing to tell us the length of a hole,
 875          * just fall back to PTEs.  Calling get_block 512 times in a loop
 876          * would be silly.
 877          */
 878         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 879                 dax_pmd_dbg(&bh, address, "allocated block too small");
 880                 return VM_FAULT_FALLBACK;
 881         }
 882
 883         /*
 884          * If we allocated new storage, make sure no process has any
 885          * zero pages covering this hole
 886          */
 887         if (alloc) {
 888                 loff_t lstart = pgoff << PAGE_SHIFT;
 889                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 890
 891                 truncate_pagecache_range(inode, lstart, lend);
 892         }
 893
 894         i_mmap_lock_read(mapping);
 895
 896         /*
 897          * If a truncate happened while we were allocating blocks, we may
 898          * leave blocks allocated to the file that are beyond EOF.  We can't
 899          * take i_mutex here, so just leave them hanging; they'll be freed
 900          * when the file is deleted.
 901          */
 902         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 903         if (pgoff >= size) {
 904                 result = VM_FAULT_SIGBUS;
 905                 goto out;
 906         }
 907         if ((pgoff | PG_PMD_COLOUR) >= size) {
 908                 dax_pmd_dbg(&bh, address,
 909                                 "offset + huge page size > file size");
 910                 goto fallback;
 911         }
 912
 913         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 914                 spinlock_t *ptl;
 915                 pmd_t entry;
 916                 struct page *zero_page = get_huge_zero_page();
 917
 918                 if (unlikely(!zero_page)) {
 919                         dax_pmd_dbg(&bh, address, "no zero page");
 920                         goto fallback;
 921                 }
 922
 923                 ptl = pmd_lock(vma->vm_mm, pmd);
 924                 if (!pmd_none(*pmd)) {
 925                         spin_unlock(ptl);
 926                         dax_pmd_dbg(&bh, address, "pmd already present");
 927                         goto fallback;
 928                 }
 929
 930                 dev_dbg(part_to_dev(bdev->bd_part),
 931                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 932                                 __func__, current->comm, address,
 933                                 (unsigned long long) to_sector(&bh, inode));
 934
 935                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 936                 entry = pmd_mkhuge(entry);
 937                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 938                 result = VM_FAULT_NOPAGE;
 939                 spin_unlock(ptl);
 940         } else {
 941                 struct blk_dax_ctl dax = {
 942                         .sector = to_sector(&bh, inode),
 943                         .size = PMD_SIZE,
 944                 };
 945                 long length = dax_map_atomic(bdev, &dax);
 946
 947                 if (length < 0) {
 948                         result = VM_FAULT_SIGBUS;
 949                         goto out;
 950                 }
 951                 if (length < PMD_SIZE) {
 952                         dax_pmd_dbg(&bh, address, "dax-length too small");
 953                         dax_unmap_atomic(bdev, &dax);
 954                         goto fallback;
 955                 }
 956                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 957                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 958                         dax_unmap_atomic(bdev, &dax);
 959                         goto fallback;
 960                 }
 961
 962                 if (!pfn_t_devmap(dax.pfn)) {
 963                         dax_unmap_atomic(bdev, &dax);
 964                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 965                         goto fallback;
 966                 }
 967
 968                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 969                         clear_pmem(dax.addr, PMD_SIZE);
 970                         wmb_pmem();
 971                         count_vm_event(PGMAJFAULT);
 972                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 973                         result |= VM_FAULT_MAJOR;
 974                 }
 975                 dax_unmap_atomic(bdev, &dax);
 976
 977                 /*
 978                  * For PTE faults we insert a radix tree entry for reads, and
 979                  * leave it clean.  Then on the first write we dirty the radix
 980                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 981                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 982                  * call into get_block() to translate the pgoff to a sector in
 983                  * order to be able to create a new radix tree entry.
 984                  *
 985                  * The PMD path doesn't have an equivalent to
 986                  * dax_pfn_mkwrite(), though, so for a read followed by a
 987                  * write we traverse all the way through __dax_pmd_fault()
 988                  * twice.  This means we can just skip inserting a radix tree
 989                  * entry completely on the initial read and just wait until
 990                  * the write to insert a dirty entry.
 991                  */
 992                 if (write) {
 993                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 994                                         true, true);
 995                         if (error) {
 996                                 dax_pmd_dbg(&bh, address,
 997                                                 "PMD radix insertion failed");
 998                                 goto fallback;
 999                         }
1000                 }
1001
1002                 dev_dbg(part_to_dev(bdev->bd_part),
1003                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1004                                 __func__, current->comm, address,
1005                                 pfn_t_to_pfn(dax.pfn),
1006                                 (unsigned long long) dax.sector);
1007                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1008                                 dax.pfn, write);
1009         }
1010
1011  out:
1012         i_mmap_unlock_read(mapping);
1013
1014         if (buffer_unwritten(&bh))
1015                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1016
1017         return result;
1018
1019  fallback:
1020         count_vm_event(THP_FAULT_FALLBACK);
1021         result = VM_FAULT_FALLBACK;
1022         goto out;
1023 }
1024 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1025
1026 /**
1027  * dax_pmd_fault - handle a PMD fault on a DAX file
1028  * @vma: The virtual memory area where the fault occurred
1029  * @vmf: The description of the fault
1030  * @get_block: The filesystem method used to translate file offsets to blocks
1031  *
1032  * When a page fault occurs, filesystems may call this helper in their
1033  * pmd_fault handler for DAX files.
1034  */
1035 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1036                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
1037                         dax_iodone_t complete_unwritten)
1038 {
1039         int result;
1040         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1041
1042         if (flags & FAULT_FLAG_WRITE) {
1043                 sb_start_pagefault(sb);
1044                 file_update_time(vma->vm_file);
1045         }
1046         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1047                                 complete_unwritten);
1048         if (flags & FAULT_FLAG_WRITE)
1049                 sb_end_pagefault(sb);
1050
1051         return result;
1052 }
1053 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1054 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1055
1056 /**
1057  * dax_pfn_mkwrite - handle first write to DAX page
1058  * @vma: The virtual memory area where the fault occurred
1059  * @vmf: The description of the fault
1060  */
1061 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1062 {
1063         struct file *file = vma->vm_file;
1064         int error;
1065
1066         /*
1067          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1068          * RADIX_DAX_PTE entry already exists in the radix tree from a
1069          * previous call to __dax_fault().  We just want to look up that PTE
1070          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1071          * saves us from having to make a call to get_block() here to look
1072          * up the sector.
1073          */
1074         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1075                         true);
1076
1077         if (error == -ENOMEM)
1078                 return VM_FAULT_OOM;
1079         if (error)
1080                 return VM_FAULT_SIGBUS;
1081         return VM_FAULT_NOPAGE;
1082 }
1083 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1084
1085 /**
1086  * dax_zero_page_range - zero a range within a page of a DAX file
1087  * @inode: The file being truncated
1088  * @from: The file offset that is being truncated to
1089  * @length: The number of bytes to zero
1090  * @get_block: The filesystem method used to translate file offsets to blocks
1091  *
1092  * This function can be called by a filesystem when it is zeroing part of a
1093  * page in a DAX file.  This is intended for hole-punch operations.  If
1094  * you are truncating a file, the helper function dax_truncate_page() may be
1095  * more convenient.
1096  *
1097  * We work in terms of PAGE_SIZE here for commonality with
1098  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1099  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1100  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1101  * since the file might be mmapped.
1102  */
1103 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1104                                                         get_block_t get_block)
1105 {
1106         struct buffer_head bh;
1107         pgoff_t index = from >> PAGE_SHIFT;
1108         unsigned offset = from & (PAGE_SIZE-1);
1109         int err;
1110
1111         /* Block boundary? Nothing to do */
1112         if (!length)
1113                 return 0;
1114         BUG_ON((offset + length) > PAGE_SIZE);
1115
1116         memset(&bh, 0, sizeof(bh));
1117         bh.b_bdev = inode->i_sb->s_bdev;
1118         bh.b_size = PAGE_SIZE;
1119         err = get_block(inode, index, &bh, 0);
1120         if (err < 0)
1121                 return err;
1122         if (buffer_written(&bh)) {
1123                 struct block_device *bdev = bh.b_bdev;
1124                 struct blk_dax_ctl dax = {
1125                         .sector = to_sector(&bh, inode),
1126                         .size = PAGE_SIZE,
1127                 };
1128
1129                 if (dax_map_atomic(bdev, &dax) < 0)
1130                         return PTR_ERR(dax.addr);
1131                 clear_pmem(dax.addr + offset, length);
1132                 wmb_pmem();
1133                 dax_unmap_atomic(bdev, &dax);
1134         }
1135
1136         return 0;
1137 }
1138 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1139
1140 /**
1141  * dax_truncate_page - handle a partial page being truncated in a DAX file
1142  * @inode: The file being truncated
1143  * @from: The file offset that is being truncated to
1144  * @get_block: The filesystem method used to translate file offsets to blocks
1145  *
1146  * Similar to block_truncate_page(), this function can be called by a
1147  * filesystem when it is truncating a DAX file to handle the partial page.
1148  *
1149  * We work in terms of PAGE_SIZE here for commonality with
1150  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1151  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1152  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1153  * since the file might be mmapped.
1154  */
1155 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1156 {
1157         unsigned length = PAGE_ALIGN(from) - from;
1158         return dax_zero_page_range(inode, from, length, get_block);
1159 }
1160 EXPORT_SYMBOL_GPL(dax_truncate_page);