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 (end_io) {
 291                 int err;
 292
 293                 err = end_io(iocb, pos, retval, bh.b_private);
 294                 if (err)
 295                         retval = err;
 296         }
 297
 298         if (!(flags & DIO_SKIP_DIO_COUNT))
 299                 inode_dio_end(inode);
 300  out:
 301         return retval;
 302 }
 303 EXPORT_SYMBOL_GPL(dax_do_io);
 304
 305 /*
 306  * The user has performed a load from a hole in the file.  Allocating
 307  * a new page in the file would cause excessive storage usage for
 308  * workloads with sparse files.  We allocate a page cache page instead.
 309  * We'll kick it out of the page cache if it's ever written to,
 310  * otherwise it will simply fall out of the page cache under memory
 311  * pressure without ever having been dirtied.
 312  */
 313 static int dax_load_hole(struct address_space *mapping, struct page *page,
 314                                                         struct vm_fault *vmf)
 315 {
 316         unsigned long size;
 317         struct inode *inode = mapping->host;
 318         if (!page)
 319                 page = find_or_create_page(mapping, vmf->pgoff,
 320                                                 GFP_KERNEL | __GFP_ZERO);
 321         if (!page)
 322                 return VM_FAULT_OOM;
 323         /* Recheck i_size under page lock to avoid truncate race */
 324         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 325         if (vmf->pgoff >= size) {
 326                 unlock_page(page);
 327                 page_cache_release(page);
 328                 return VM_FAULT_SIGBUS;
 329         }
 330
 331         vmf->page = page;
 332         return VM_FAULT_LOCKED;
 333 }
 334
 335 static int copy_user_bh(struct page *to, struct inode *inode,
 336                 struct buffer_head *bh, unsigned long vaddr)
 337 {
 338         struct blk_dax_ctl dax = {
 339                 .sector = to_sector(bh, inode),
 340                 .size = bh->b_size,
 341         };
 342         struct block_device *bdev = bh->b_bdev;
 343         void *vto;
 344
 345         if (dax_map_atomic(bdev, &dax) < 0)
 346                 return PTR_ERR(dax.addr);
 347         vto = kmap_atomic(to);
 348         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 349         kunmap_atomic(vto);
 350         dax_unmap_atomic(bdev, &dax);
 351         return 0;
 352 }
 353
 354 #define NO_SECTOR -1
 355 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_CACHE_SHIFT))
 356
 357 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 358                 sector_t sector, bool pmd_entry, bool dirty)
 359 {
 360         struct radix_tree_root *page_tree = &mapping->page_tree;
 361         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 362         int type, error = 0;
 363         void *entry;
 364
 365         WARN_ON_ONCE(pmd_entry && !dirty);
 366         if (dirty)
 367                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 368
 369         spin_lock_irq(&mapping->tree_lock);
 370
 371         entry = radix_tree_lookup(page_tree, pmd_index);
 372         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 373                 index = pmd_index;
 374                 goto dirty;
 375         }
 376
 377         entry = radix_tree_lookup(page_tree, index);
 378         if (entry) {
 379                 type = RADIX_DAX_TYPE(entry);
 380                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 381                                         type != RADIX_DAX_PMD)) {
 382                         error = -EIO;
 383                         goto unlock;
 384                 }
 385
 386                 if (!pmd_entry || type == RADIX_DAX_PMD)
 387                         goto dirty;
 388
 389                 /*
 390                  * We only insert dirty PMD entries into the radix tree.  This
 391                  * means we don't need to worry about removing a dirty PTE
 392                  * entry and inserting a clean PMD entry, thus reducing the
 393                  * range we would flush with a follow-up fsync/msync call.
 394                  */
 395                 radix_tree_delete(&mapping->page_tree, index);
 396                 mapping->nrexceptional--;
 397         }
 398
 399         if (sector == NO_SECTOR) {
 400                 /*
 401                  * This can happen during correct operation if our pfn_mkwrite
 402                  * fault raced against a hole punch operation.  If this
 403                  * happens the pte that was hole punched will have been
 404                  * unmapped and the radix tree entry will have been removed by
 405                  * the time we are called, but the call will still happen.  We
 406                  * will return all the way up to wp_pfn_shared(), where the
 407                  * pte_same() check will fail, eventually causing page fault
 408                  * to be retried by the CPU.
 409                  */
 410                 goto unlock;
 411         }
 412
 413         error = radix_tree_insert(page_tree, index,
 414                         RADIX_DAX_ENTRY(sector, pmd_entry));
 415         if (error)
 416                 goto unlock;
 417
 418         mapping->nrexceptional++;
 419  dirty:
 420         if (dirty)
 421                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 422  unlock:
 423         spin_unlock_irq(&mapping->tree_lock);
 424         return error;
 425 }
 426
 427 static int dax_writeback_one(struct block_device *bdev,
 428                 struct address_space *mapping, pgoff_t index, void *entry)
 429 {
 430         struct radix_tree_root *page_tree = &mapping->page_tree;
 431         int type = RADIX_DAX_TYPE(entry);
 432         struct radix_tree_node *node;
 433         struct blk_dax_ctl dax;
 434         void **slot;
 435         int ret = 0;
 436
 437         spin_lock_irq(&mapping->tree_lock);
 438         /*
 439          * Regular page slots are stabilized by the page lock even
 440          * without the tree itself locked.  These unlocked entries
 441          * need verification under the tree lock.
 442          */
 443         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 444                 goto unlock;
 445         if (*slot != entry)
 446                 goto unlock;
 447
 448         /* another fsync thread may have already written back this entry */
 449         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 450                 goto unlock;
 451
 452         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 453                 ret = -EIO;
 454                 goto unlock;
 455         }
 456
 457         dax.sector = RADIX_DAX_SECTOR(entry);
 458         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 459         spin_unlock_irq(&mapping->tree_lock);
 460
 461         /*
 462          * We cannot hold tree_lock while calling dax_map_atomic() because it
 463          * eventually calls cond_resched().
 464          */
 465         ret = dax_map_atomic(bdev, &dax);
 466         if (ret < 0)
 467                 return ret;
 468
 469         if (WARN_ON_ONCE(ret < dax.size)) {
 470                 ret = -EIO;
 471                 goto unmap;
 472         }
 473
 474         wb_cache_pmem(dax.addr, dax.size);
 475
 476         spin_lock_irq(&mapping->tree_lock);
 477         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 478         spin_unlock_irq(&mapping->tree_lock);
 479  unmap:
 480         dax_unmap_atomic(bdev, &dax);
 481         return ret;
 482
 483  unlock:
 484         spin_unlock_irq(&mapping->tree_lock);
 485         return ret;
 486 }
 487
 488 /*
 489  * Flush the mapping to the persistent domain within the byte range of [start,
 490  * end]. This is required by data integrity operations to ensure file data is
 491  * on persistent storage prior to completion of the operation.
 492  */
 493 int dax_writeback_mapping_range(struct address_space *mapping, loff_t start,
 494                 loff_t end)
 495 {
 496         struct inode *inode = mapping->host;
 497         struct block_device *bdev = inode->i_sb->s_bdev;
 498         pgoff_t start_index, end_index, pmd_index;
 499         pgoff_t indices[PAGEVEC_SIZE];
 500         struct pagevec pvec;
 501         bool done = false;
 502         int i, ret = 0;
 503         void *entry;
 504
 505         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 506                 return -EIO;
 507
 508         start_index = start >> PAGE_CACHE_SHIFT;
 509         end_index = end >> PAGE_CACHE_SHIFT;
 510         pmd_index = DAX_PMD_INDEX(start_index);
 511
 512         rcu_read_lock();
 513         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 514         rcu_read_unlock();
 515
 516         /* see if the start of our range is covered by a PMD entry */
 517         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 518                 start_index = pmd_index;
 519
 520         tag_pages_for_writeback(mapping, start_index, end_index);
 521
 522         pagevec_init(&pvec, 0);
 523         while (!done) {
 524                 pvec.nr = find_get_entries_tag(mapping, start_index,
 525                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 526                                 pvec.pages, indices);
 527
 528                 if (pvec.nr == 0)
 529                         break;
 530
 531                 for (i = 0; i < pvec.nr; i++) {
 532                         if (indices[i] > end_index) {
 533                                 done = true;
 534                                 break;
 535                         }
 536
 537                         ret = dax_writeback_one(bdev, mapping, indices[i],
 538                                         pvec.pages[i]);
 539                         if (ret < 0)
 540                                 return ret;
 541                 }
 542         }
 543         wmb_pmem();
 544         return 0;
 545 }
 546 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 547
 548 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 549                         struct vm_area_struct *vma, struct vm_fault *vmf)
 550 {
 551         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 552         struct address_space *mapping = inode->i_mapping;
 553         struct block_device *bdev = bh->b_bdev;
 554         struct blk_dax_ctl dax = {
 555                 .sector = to_sector(bh, inode),
 556                 .size = bh->b_size,
 557         };
 558         pgoff_t size;
 559         int error;
 560
 561         i_mmap_lock_read(mapping);
 562
 563         /*
 564          * Check truncate didn't happen while we were allocating a block.
 565          * If it did, this block may or may not be still allocated to the
 566          * file.  We can't tell the filesystem to free it because we can't
 567          * take i_mutex here.  In the worst case, the file still has blocks
 568          * allocated past the end of the file.
 569          */
 570         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 571         if (unlikely(vmf->pgoff >= size)) {
 572                 error = -EIO;
 573                 goto out;
 574         }
 575
 576         if (dax_map_atomic(bdev, &dax) < 0) {
 577                 error = PTR_ERR(dax.addr);
 578                 goto out;
 579         }
 580
 581         if (buffer_unwritten(bh) || buffer_new(bh)) {
 582                 clear_pmem(dax.addr, PAGE_SIZE);
 583                 wmb_pmem();
 584         }
 585         dax_unmap_atomic(bdev, &dax);
 586
 587         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 588                         vmf->flags & FAULT_FLAG_WRITE);
 589         if (error)
 590                 goto out;
 591
 592         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 593
 594  out:
 595         i_mmap_unlock_read(mapping);
 596
 597         return error;
 598 }
 599
 600 /**
 601  * __dax_fault - handle a page fault on a DAX file
 602  * @vma: The virtual memory area where the fault occurred
 603  * @vmf: The description of the fault
 604  * @get_block: The filesystem method used to translate file offsets to blocks
 605  * @complete_unwritten: The filesystem method used to convert unwritten blocks
 606  *      to written so the data written to them is exposed. This is required for
 607  *      required by write faults for filesystems that will return unwritten
 608  *      extent mappings from @get_block, but it is optional for reads as
 609  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
 610  *      not support unwritten extents, the it should pass NULL.
 611  *
 612  * When a page fault occurs, filesystems may call this helper in their
 613  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 614  * the necessary locking for the page fault to proceed successfully.
 615  */
 616 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 617                         get_block_t get_block, dax_iodone_t complete_unwritten)
 618 {
 619         struct file *file = vma->vm_file;
 620         struct address_space *mapping = file->f_mapping;
 621         struct inode *inode = mapping->host;
 622         struct page *page;
 623         struct buffer_head bh;
 624         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 625         unsigned blkbits = inode->i_blkbits;
 626         sector_t block;
 627         pgoff_t size;
 628         int error;
 629         int major = 0;
 630
 631         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 632         if (vmf->pgoff >= size)
 633                 return VM_FAULT_SIGBUS;
 634
 635         memset(&bh, 0, sizeof(bh));
 636         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 637         bh.b_bdev = inode->i_sb->s_bdev;
 638         bh.b_size = PAGE_SIZE;
 639
 640  repeat:
 641         page = find_get_page(mapping, vmf->pgoff);
 642         if (page) {
 643                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 644                         page_cache_release(page);
 645                         return VM_FAULT_RETRY;
 646                 }
 647                 if (unlikely(page->mapping != mapping)) {
 648                         unlock_page(page);
 649                         page_cache_release(page);
 650                         goto repeat;
 651                 }
 652                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 653                 if (unlikely(vmf->pgoff >= size)) {
 654                         /*
 655                          * We have a struct page covering a hole in the file
 656                          * from a read fault and we've raced with a truncate
 657                          */
 658                         error = -EIO;
 659                         goto unlock_page;
 660                 }
 661         }
 662
 663         error = get_block(inode, block, &bh, 0);
 664         if (!error && (bh.b_size < PAGE_SIZE))
 665                 error = -EIO;           /* fs corruption? */
 666         if (error)
 667                 goto unlock_page;
 668
 669         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 670                 if (vmf->flags & FAULT_FLAG_WRITE) {
 671                         error = get_block(inode, block, &bh, 1);
 672                         count_vm_event(PGMAJFAULT);
 673                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 674                         major = VM_FAULT_MAJOR;
 675                         if (!error && (bh.b_size < PAGE_SIZE))
 676                                 error = -EIO;
 677                         if (error)
 678                                 goto unlock_page;
 679                 } else {
 680                         return dax_load_hole(mapping, page, vmf);
 681                 }
 682         }
 683
 684         if (vmf->cow_page) {
 685                 struct page *new_page = vmf->cow_page;
 686                 if (buffer_written(&bh))
 687                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 688                 else
 689                         clear_user_highpage(new_page, vaddr);
 690                 if (error)
 691                         goto unlock_page;
 692                 vmf->page = page;
 693                 if (!page) {
 694                         i_mmap_lock_read(mapping);
 695                         /* Check we didn't race with truncate */
 696                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 697                                                                 PAGE_SHIFT;
 698                         if (vmf->pgoff >= size) {
 699                                 i_mmap_unlock_read(mapping);
 700                                 error = -EIO;
 701                                 goto out;
 702                         }
 703                 }
 704                 return VM_FAULT_LOCKED;
 705         }
 706
 707         /* Check we didn't race with a read fault installing a new page */
 708         if (!page && major)
 709                 page = find_lock_page(mapping, vmf->pgoff);
 710
 711         if (page) {
 712                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 713                                                         PAGE_CACHE_SIZE, 0);
 714                 delete_from_page_cache(page);
 715                 unlock_page(page);
 716                 page_cache_release(page);
 717                 page = NULL;
 718         }
 719
 720         /*
 721          * If we successfully insert the new mapping over an unwritten extent,
 722          * we need to ensure we convert the unwritten extent. If there is an
 723          * error inserting the mapping, the filesystem needs to leave it as
 724          * unwritten to prevent exposure of the stale underlying data to
 725          * userspace, but we still need to call the completion function so
 726          * the private resources on the mapping buffer can be released. We
 727          * indicate what the callback should do via the uptodate variable, same
 728          * as for normal BH based IO completions.
 729          */
 730         error = dax_insert_mapping(inode, &bh, vma, vmf);
 731         if (buffer_unwritten(&bh)) {
 732                 if (complete_unwritten)
 733                         complete_unwritten(&bh, !error);
 734                 else
 735                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 736         }
 737
 738  out:
 739         if (error == -ENOMEM)
 740                 return VM_FAULT_OOM | major;
 741         /* -EBUSY is fine, somebody else faulted on the same PTE */
 742         if ((error < 0) && (error != -EBUSY))
 743                 return VM_FAULT_SIGBUS | major;
 744         return VM_FAULT_NOPAGE | major;
 745
 746  unlock_page:
 747         if (page) {
 748                 unlock_page(page);
 749                 page_cache_release(page);
 750         }
 751         goto out;
 752 }
 753 EXPORT_SYMBOL(__dax_fault);
 754
 755 /**
 756  * dax_fault - handle a page fault on a DAX file
 757  * @vma: The virtual memory area where the fault occurred
 758  * @vmf: The description of the fault
 759  * @get_block: The filesystem method used to translate file offsets to blocks
 760  *
 761  * When a page fault occurs, filesystems may call this helper in their
 762  * fault handler for DAX files.
 763  */
 764 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 765               get_block_t get_block, dax_iodone_t complete_unwritten)
 766 {
 767         int result;
 768         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 769
 770         if (vmf->flags & FAULT_FLAG_WRITE) {
 771                 sb_start_pagefault(sb);
 772                 file_update_time(vma->vm_file);
 773         }
 774         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 775         if (vmf->flags & FAULT_FLAG_WRITE)
 776                 sb_end_pagefault(sb);
 777
 778         return result;
 779 }
 780 EXPORT_SYMBOL_GPL(dax_fault);
 781
 782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 783 /*
 784  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 785  * more often than one might expect in the below function.
 786  */
 787 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 788
 789 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 790                 const char *reason, const char *fn)
 791 {
 792         if (bh) {
 793                 char bname[BDEVNAME_SIZE];
 794                 bdevname(bh->b_bdev, bname);
 795                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 796                         "length %zd fallback: %s\n", fn, current->comm,
 797                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 798                         bh->b_size, reason);
 799         } else {
 800                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 801                         current->comm, address, reason);
 802         }
 803 }
 804
 805 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 806
 807 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 808                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
 809                 dax_iodone_t complete_unwritten)
 810 {
 811         struct file *file = vma->vm_file;
 812         struct address_space *mapping = file->f_mapping;
 813         struct inode *inode = mapping->host;
 814         struct buffer_head bh;
 815         unsigned blkbits = inode->i_blkbits;
 816         unsigned long pmd_addr = address & PMD_MASK;
 817         bool write = flags & FAULT_FLAG_WRITE;
 818         struct block_device *bdev;
 819         pgoff_t size, pgoff;
 820         sector_t block;
 821         int error, result = 0;
 822         bool alloc = false;
 823
 824         /* dax pmd mappings require pfn_t_devmap() */
 825         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 826                 return VM_FAULT_FALLBACK;
 827
 828         /* Fall back to PTEs if we're going to COW */
 829         if (write && !(vma->vm_flags & VM_SHARED)) {
 830                 split_huge_pmd(vma, pmd, address);
 831                 dax_pmd_dbg(NULL, address, "cow write");
 832                 return VM_FAULT_FALLBACK;
 833         }
 834         /* If the PMD would extend outside the VMA */
 835         if (pmd_addr < vma->vm_start) {
 836                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 837                 return VM_FAULT_FALLBACK;
 838         }
 839         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 840                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 841                 return VM_FAULT_FALLBACK;
 842         }
 843
 844         pgoff = linear_page_index(vma, pmd_addr);
 845         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 846         if (pgoff >= size)
 847                 return VM_FAULT_SIGBUS;
 848         /* If the PMD would cover blocks out of the file */
 849         if ((pgoff | PG_PMD_COLOUR) >= size) {
 850                 dax_pmd_dbg(NULL, address,
 851                                 "offset + huge page size > file size");
 852                 return VM_FAULT_FALLBACK;
 853         }
 854
 855         memset(&bh, 0, sizeof(bh));
 856         bh.b_bdev = inode->i_sb->s_bdev;
 857         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 858
 859         bh.b_size = PMD_SIZE;
 860
 861         if (get_block(inode, block, &bh, 0) != 0)
 862                 return VM_FAULT_SIGBUS;
 863
 864         if (!buffer_mapped(&bh) && write) {
 865                 if (get_block(inode, block, &bh, 1) != 0)
 866                         return VM_FAULT_SIGBUS;
 867                 alloc = true;
 868         }
 869
 870         bdev = bh.b_bdev;
 871
 872         /*
 873          * If the filesystem isn't willing to tell us the length of a hole,
 874          * just fall back to PTEs.  Calling get_block 512 times in a loop
 875          * would be silly.
 876          */
 877         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 878                 dax_pmd_dbg(&bh, address, "allocated block too small");
 879                 return VM_FAULT_FALLBACK;
 880         }
 881
 882         /*
 883          * If we allocated new storage, make sure no process has any
 884          * zero pages covering this hole
 885          */
 886         if (alloc) {
 887                 loff_t lstart = pgoff << PAGE_SHIFT;
 888                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 889
 890                 truncate_pagecache_range(inode, lstart, lend);
 891         }
 892
 893         i_mmap_lock_read(mapping);
 894
 895         /*
 896          * If a truncate happened while we were allocating blocks, we may
 897          * leave blocks allocated to the file that are beyond EOF.  We can't
 898          * take i_mutex here, so just leave them hanging; they'll be freed
 899          * when the file is deleted.
 900          */
 901         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 902         if (pgoff >= size) {
 903                 result = VM_FAULT_SIGBUS;
 904                 goto out;
 905         }
 906         if ((pgoff | PG_PMD_COLOUR) >= size) {
 907                 dax_pmd_dbg(&bh, address,
 908                                 "offset + huge page size > file size");
 909                 goto fallback;
 910         }
 911
 912         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 913                 spinlock_t *ptl;
 914                 pmd_t entry;
 915                 struct page *zero_page = get_huge_zero_page();
 916
 917                 if (unlikely(!zero_page)) {
 918                         dax_pmd_dbg(&bh, address, "no zero page");
 919                         goto fallback;
 920                 }
 921
 922                 ptl = pmd_lock(vma->vm_mm, pmd);
 923                 if (!pmd_none(*pmd)) {
 924                         spin_unlock(ptl);
 925                         dax_pmd_dbg(&bh, address, "pmd already present");
 926                         goto fallback;
 927                 }
 928
 929                 dev_dbg(part_to_dev(bdev->bd_part),
 930                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 931                                 __func__, current->comm, address,
 932                                 (unsigned long long) to_sector(&bh, inode));
 933
 934                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 935                 entry = pmd_mkhuge(entry);
 936                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 937                 result = VM_FAULT_NOPAGE;
 938                 spin_unlock(ptl);
 939         } else {
 940                 struct blk_dax_ctl dax = {
 941                         .sector = to_sector(&bh, inode),
 942                         .size = PMD_SIZE,
 943                 };
 944                 long length = dax_map_atomic(bdev, &dax);
 945
 946                 if (length < 0) {
 947                         result = VM_FAULT_SIGBUS;
 948                         goto out;
 949                 }
 950                 if (length < PMD_SIZE) {
 951                         dax_pmd_dbg(&bh, address, "dax-length too small");
 952                         dax_unmap_atomic(bdev, &dax);
 953                         goto fallback;
 954                 }
 955                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 956                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 957                         dax_unmap_atomic(bdev, &dax);
 958                         goto fallback;
 959                 }
 960
 961                 if (!pfn_t_devmap(dax.pfn)) {
 962                         dax_unmap_atomic(bdev, &dax);
 963                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 964                         goto fallback;
 965                 }
 966
 967                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 968                         clear_pmem(dax.addr, PMD_SIZE);
 969                         wmb_pmem();
 970                         count_vm_event(PGMAJFAULT);
 971                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 972                         result |= VM_FAULT_MAJOR;
 973                 }
 974                 dax_unmap_atomic(bdev, &dax);
 975
 976                 /*
 977                  * For PTE faults we insert a radix tree entry for reads, and
 978                  * leave it clean.  Then on the first write we dirty the radix
 979                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 980                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 981                  * call into get_block() to translate the pgoff to a sector in
 982                  * order to be able to create a new radix tree entry.
 983                  *
 984                  * The PMD path doesn't have an equivalent to
 985                  * dax_pfn_mkwrite(), though, so for a read followed by a
 986                  * write we traverse all the way through __dax_pmd_fault()
 987                  * twice.  This means we can just skip inserting a radix tree
 988                  * entry completely on the initial read and just wait until
 989                  * the write to insert a dirty entry.
 990                  */
 991                 if (write) {
 992                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 993                                         true, true);
 994                         if (error) {
 995                                 dax_pmd_dbg(&bh, address,
 996                                                 "PMD radix insertion failed");
 997                                 goto fallback;
 998                         }
 999                 }
1000
1001                 dev_dbg(part_to_dev(bdev->bd_part),
1002                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1003                                 __func__, current->comm, address,
1004                                 pfn_t_to_pfn(dax.pfn),
1005                                 (unsigned long long) dax.sector);
1006                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1007                                 dax.pfn, write);
1008         }
1009
1010  out:
1011         i_mmap_unlock_read(mapping);
1012
1013         if (buffer_unwritten(&bh))
1014                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
1015
1016         return result;
1017
1018  fallback:
1019         count_vm_event(THP_FAULT_FALLBACK);
1020         result = VM_FAULT_FALLBACK;
1021         goto out;
1022 }
1023 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1024
1025 /**
1026  * dax_pmd_fault - handle a PMD fault on a DAX file
1027  * @vma: The virtual memory area where the fault occurred
1028  * @vmf: The description of the fault
1029  * @get_block: The filesystem method used to translate file offsets to blocks
1030  *
1031  * When a page fault occurs, filesystems may call this helper in their
1032  * pmd_fault handler for DAX files.
1033  */
1034 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1035                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
1036                         dax_iodone_t complete_unwritten)
1037 {
1038         int result;
1039         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1040
1041         if (flags & FAULT_FLAG_WRITE) {
1042                 sb_start_pagefault(sb);
1043                 file_update_time(vma->vm_file);
1044         }
1045         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1046                                 complete_unwritten);
1047         if (flags & FAULT_FLAG_WRITE)
1048                 sb_end_pagefault(sb);
1049
1050         return result;
1051 }
1052 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1053 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1054
1055 /**
1056  * dax_pfn_mkwrite - handle first write to DAX page
1057  * @vma: The virtual memory area where the fault occurred
1058  * @vmf: The description of the fault
1059  */
1060 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1061 {
1062         struct file *file = vma->vm_file;
1063
1064         /*
1065          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1066          * RADIX_DAX_PTE entry already exists in the radix tree from a
1067          * previous call to __dax_fault().  We just want to look up that PTE
1068          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1069          * saves us from having to make a call to get_block() here to look
1070          * up the sector.
1071          */
1072         dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false, true);
1073         return VM_FAULT_NOPAGE;
1074 }
1075 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1076
1077 /**
1078  * dax_zero_page_range - zero a range within a page of a DAX file
1079  * @inode: The file being truncated
1080  * @from: The file offset that is being truncated to
1081  * @length: The number of bytes to zero
1082  * @get_block: The filesystem method used to translate file offsets to blocks
1083  *
1084  * This function can be called by a filesystem when it is zeroing part of a
1085  * page in a DAX file.  This is intended for hole-punch operations.  If
1086  * you are truncating a file, the helper function dax_truncate_page() may be
1087  * more convenient.
1088  *
1089  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1090  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1091  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1092  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1093  * since the file might be mmapped.
1094  */
1095 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1096                                                         get_block_t get_block)
1097 {
1098         struct buffer_head bh;
1099         pgoff_t index = from >> PAGE_CACHE_SHIFT;
1100         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1101         int err;
1102
1103         /* Block boundary? Nothing to do */
1104         if (!length)
1105                 return 0;
1106         BUG_ON((offset + length) > PAGE_CACHE_SIZE);
1107
1108         memset(&bh, 0, sizeof(bh));
1109         bh.b_bdev = inode->i_sb->s_bdev;
1110         bh.b_size = PAGE_CACHE_SIZE;
1111         err = get_block(inode, index, &bh, 0);
1112         if (err < 0)
1113                 return err;
1114         if (buffer_written(&bh)) {
1115                 struct block_device *bdev = bh.b_bdev;
1116                 struct blk_dax_ctl dax = {
1117                         .sector = to_sector(&bh, inode),
1118                         .size = PAGE_CACHE_SIZE,
1119                 };
1120
1121                 if (dax_map_atomic(bdev, &dax) < 0)
1122                         return PTR_ERR(dax.addr);
1123                 clear_pmem(dax.addr + offset, length);
1124                 wmb_pmem();
1125                 dax_unmap_atomic(bdev, &dax);
1126         }
1127
1128         return 0;
1129 }
1130 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1131
1132 /**
1133  * dax_truncate_page - handle a partial page being truncated in a DAX file
1134  * @inode: The file being truncated
1135  * @from: The file offset that is being truncated to
1136  * @get_block: The filesystem method used to translate file offsets to blocks
1137  *
1138  * Similar to block_truncate_page(), this function can be called by a
1139  * filesystem when it is truncating a DAX file to handle the partial page.
1140  *
1141  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1142  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1143  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1144  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1145  * since the file might be mmapped.
1146  */
1147 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1148 {
1149         unsigned length = PAGE_CACHE_ALIGN(from) - from;
1150         return dax_zero_page_range(inode, from, length, get_block);
1151 }
1152 EXPORT_SYMBOL_GPL(dax_truncate_page);