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