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