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