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 (end_io) {
 271                 int err;
 272
 273                 err = end_io(iocb, pos, retval, bh.b_private);
 274                 if (err)
 275                         retval = err;
 276         }
 277
 278         if (!(flags & DIO_SKIP_DIO_COUNT))
 279                 inode_dio_end(inode);
 280  out:
 281         return retval;
 282 }
 283 EXPORT_SYMBOL_GPL(dax_do_io);
 284
 285 /*
 286  * The user has performed a load from a hole in the file.  Allocating
 287  * a new page in the file would cause excessive storage usage for
 288  * workloads with sparse files.  We allocate a page cache page instead.
 289  * We'll kick it out of the page cache if it's ever written to,
 290  * otherwise it will simply fall out of the page cache under memory
 291  * pressure without ever having been dirtied.
 292  */
 293 static int dax_load_hole(struct address_space *mapping, struct page *page,
 294                                                         struct vm_fault *vmf)
 295 {
 296         unsigned long size;
 297         struct inode *inode = mapping->host;
 298         if (!page)
 299                 page = find_or_create_page(mapping, vmf->pgoff,
 300                                                 GFP_KERNEL | __GFP_ZERO);
 301         if (!page)
 302                 return VM_FAULT_OOM;
 303         /* Recheck i_size under page lock to avoid truncate race */
 304         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 305         if (vmf->pgoff >= size) {
 306                 unlock_page(page);
 307                 page_cache_release(page);
 308                 return VM_FAULT_SIGBUS;
 309         }
 310
 311         vmf->page = page;
 312         return VM_FAULT_LOCKED;
 313 }
 314
 315 static int copy_user_bh(struct page *to, struct inode *inode,
 316                 struct buffer_head *bh, unsigned long vaddr)
 317 {
 318         struct blk_dax_ctl dax = {
 319                 .sector = to_sector(bh, inode),
 320                 .size = bh->b_size,
 321         };
 322         struct block_device *bdev = bh->b_bdev;
 323         void *vto;
 324
 325         if (dax_map_atomic(bdev, &dax) < 0)
 326                 return PTR_ERR(dax.addr);
 327         vto = kmap_atomic(to);
 328         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 329         kunmap_atomic(vto);
 330         dax_unmap_atomic(bdev, &dax);
 331         return 0;
 332 }
 333
 334 #define NO_SECTOR -1
 335 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_CACHE_SHIFT))
 336
 337 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 338                 sector_t sector, bool pmd_entry, bool dirty)
 339 {
 340         struct radix_tree_root *page_tree = &mapping->page_tree;
 341         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 342         int type, error = 0;
 343         void *entry;
 344
 345         WARN_ON_ONCE(pmd_entry && !dirty);
 346         __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 347
 348         spin_lock_irq(&mapping->tree_lock);
 349
 350         entry = radix_tree_lookup(page_tree, pmd_index);
 351         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 352                 index = pmd_index;
 353                 goto dirty;
 354         }
 355
 356         entry = radix_tree_lookup(page_tree, index);
 357         if (entry) {
 358                 type = RADIX_DAX_TYPE(entry);
 359                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 360                                         type != RADIX_DAX_PMD)) {
 361                         error = -EIO;
 362                         goto unlock;
 363                 }
 364
 365                 if (!pmd_entry || type == RADIX_DAX_PMD)
 366                         goto dirty;
 367
 368                 /*
 369                  * We only insert dirty PMD entries into the radix tree.  This
 370                  * means we don't need to worry about removing a dirty PTE
 371                  * entry and inserting a clean PMD entry, thus reducing the
 372                  * range we would flush with a follow-up fsync/msync call.
 373                  */
 374                 radix_tree_delete(&mapping->page_tree, index);
 375                 mapping->nrexceptional--;
 376         }
 377
 378         if (sector == NO_SECTOR) {
 379                 /*
 380                  * This can happen during correct operation if our pfn_mkwrite
 381                  * fault raced against a hole punch operation.  If this
 382                  * happens the pte that was hole punched will have been
 383                  * unmapped and the radix tree entry will have been removed by
 384                  * the time we are called, but the call will still happen.  We
 385                  * will return all the way up to wp_pfn_shared(), where the
 386                  * pte_same() check will fail, eventually causing page fault
 387                  * to be retried by the CPU.
 388                  */
 389                 goto unlock;
 390         }
 391
 392         error = radix_tree_insert(page_tree, index,
 393                         RADIX_DAX_ENTRY(sector, pmd_entry));
 394         if (error)
 395                 goto unlock;
 396
 397         mapping->nrexceptional++;
 398  dirty:
 399         if (dirty)
 400                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 401  unlock:
 402         spin_unlock_irq(&mapping->tree_lock);
 403         return error;
 404 }
 405
 406 static int dax_writeback_one(struct block_device *bdev,
 407                 struct address_space *mapping, pgoff_t index, void *entry)
 408 {
 409         struct radix_tree_root *page_tree = &mapping->page_tree;
 410         int type = RADIX_DAX_TYPE(entry);
 411         struct radix_tree_node *node;
 412         struct blk_dax_ctl dax;
 413         void **slot;
 414         int ret = 0;
 415
 416         spin_lock_irq(&mapping->tree_lock);
 417         /*
 418          * Regular page slots are stabilized by the page lock even
 419          * without the tree itself locked.  These unlocked entries
 420          * need verification under the tree lock.
 421          */
 422         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 423                 goto unlock;
 424         if (*slot != entry)
 425                 goto unlock;
 426
 427         /* another fsync thread may have already written back this entry */
 428         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 429                 goto unlock;
 430
 431         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 432                 ret = -EIO;
 433                 goto unlock;
 434         }
 435
 436         dax.sector = RADIX_DAX_SECTOR(entry);
 437         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 438         spin_unlock_irq(&mapping->tree_lock);
 439
 440         /*
 441          * We cannot hold tree_lock while calling dax_map_atomic() because it
 442          * eventually calls cond_resched().
 443          */
 444         ret = dax_map_atomic(bdev, &dax);
 445         if (ret < 0)
 446                 return ret;
 447
 448         if (WARN_ON_ONCE(ret < dax.size)) {
 449                 ret = -EIO;
 450                 goto unmap;
 451         }
 452
 453         wb_cache_pmem(dax.addr, dax.size);
 454
 455         spin_lock_irq(&mapping->tree_lock);
 456         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 457         spin_unlock_irq(&mapping->tree_lock);
 458  unmap:
 459         dax_unmap_atomic(bdev, &dax);
 460         return ret;
 461
 462  unlock:
 463         spin_unlock_irq(&mapping->tree_lock);
 464         return ret;
 465 }
 466
 467 /*
 468  * Flush the mapping to the persistent domain within the byte range of [start,
 469  * end]. This is required by data integrity operations to ensure file data is
 470  * on persistent storage prior to completion of the operation.
 471  */
 472 int dax_writeback_mapping_range(struct address_space *mapping, loff_t start,
 473                 loff_t end)
 474 {
 475         struct inode *inode = mapping->host;
 476         struct block_device *bdev = inode->i_sb->s_bdev;
 477         pgoff_t start_index, end_index, pmd_index;
 478         pgoff_t indices[PAGEVEC_SIZE];
 479         struct pagevec pvec;
 480         bool done = false;
 481         int i, ret = 0;
 482         void *entry;
 483
 484         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 485                 return -EIO;
 486
 487         start_index = start >> PAGE_CACHE_SHIFT;
 488         end_index = end >> PAGE_CACHE_SHIFT;
 489         pmd_index = DAX_PMD_INDEX(start_index);
 490
 491         rcu_read_lock();
 492         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 493         rcu_read_unlock();
 494
 495         /* see if the start of our range is covered by a PMD entry */
 496         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 497                 start_index = pmd_index;
 498
 499         tag_pages_for_writeback(mapping, start_index, end_index);
 500
 501         pagevec_init(&pvec, 0);
 502         while (!done) {
 503                 pvec.nr = find_get_entries_tag(mapping, start_index,
 504                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 505                                 pvec.pages, indices);
 506
 507                 if (pvec.nr == 0)
 508                         break;
 509
 510                 for (i = 0; i < pvec.nr; i++) {
 511                         if (indices[i] > end_index) {
 512                                 done = true;
 513                                 break;
 514                         }
 515
 516                         ret = dax_writeback_one(bdev, mapping, indices[i],
 517                                         pvec.pages[i]);
 518                         if (ret < 0)
 519                                 return ret;
 520                 }
 521         }
 522         wmb_pmem();
 523         return 0;
 524 }
 525 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 526
 527 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 528                         struct vm_area_struct *vma, struct vm_fault *vmf)
 529 {
 530         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 531         struct address_space *mapping = inode->i_mapping;
 532         struct block_device *bdev = bh->b_bdev;
 533         struct blk_dax_ctl dax = {
 534                 .sector = to_sector(bh, inode),
 535                 .size = bh->b_size,
 536         };
 537         pgoff_t size;
 538         int error;
 539
 540         i_mmap_lock_read(mapping);
 541
 542         /*
 543          * Check truncate didn't happen while we were allocating a block.
 544          * If it did, this block may or may not be still allocated to the
 545          * file.  We can't tell the filesystem to free it because we can't
 546          * take i_mutex here.  In the worst case, the file still has blocks
 547          * allocated past the end of the file.
 548          */
 549         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 550         if (unlikely(vmf->pgoff >= size)) {
 551                 error = -EIO;
 552                 goto out;
 553         }
 554
 555         if (dax_map_atomic(bdev, &dax) < 0) {
 556                 error = PTR_ERR(dax.addr);
 557                 goto out;
 558         }
 559
 560         if (buffer_unwritten(bh) || buffer_new(bh)) {
 561                 clear_pmem(dax.addr, PAGE_SIZE);
 562                 wmb_pmem();
 563         }
 564         dax_unmap_atomic(bdev, &dax);
 565
 566         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 567                         vmf->flags & FAULT_FLAG_WRITE);
 568         if (error)
 569                 goto out;
 570
 571         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 572
 573  out:
 574         i_mmap_unlock_read(mapping);
 575
 576         return error;
 577 }
 578
 579 /**
 580  * __dax_fault - handle a page fault on a DAX file
 581  * @vma: The virtual memory area where the fault occurred
 582  * @vmf: The description of the fault
 583  * @get_block: The filesystem method used to translate file offsets to blocks
 584  * @complete_unwritten: The filesystem method used to convert unwritten blocks
 585  *      to written so the data written to them is exposed. This is required for
 586  *      required by write faults for filesystems that will return unwritten
 587  *      extent mappings from @get_block, but it is optional for reads as
 588  *      dax_insert_mapping() will always zero unwritten blocks. If the fs does
 589  *      not support unwritten extents, the it should pass NULL.
 590  *
 591  * When a page fault occurs, filesystems may call this helper in their
 592  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 593  * the necessary locking for the page fault to proceed successfully.
 594  */
 595 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 596                         get_block_t get_block, dax_iodone_t complete_unwritten)
 597 {
 598         struct file *file = vma->vm_file;
 599         struct address_space *mapping = file->f_mapping;
 600         struct inode *inode = mapping->host;
 601         struct page *page;
 602         struct buffer_head bh;
 603         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 604         unsigned blkbits = inode->i_blkbits;
 605         sector_t block;
 606         pgoff_t size;
 607         int error;
 608         int major = 0;
 609
 610         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 611         if (vmf->pgoff >= size)
 612                 return VM_FAULT_SIGBUS;
 613
 614         memset(&bh, 0, sizeof(bh));
 615         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 616         bh.b_bdev = inode->i_sb->s_bdev;
 617         bh.b_size = PAGE_SIZE;
 618
 619  repeat:
 620         page = find_get_page(mapping, vmf->pgoff);
 621         if (page) {
 622                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 623                         page_cache_release(page);
 624                         return VM_FAULT_RETRY;
 625                 }
 626                 if (unlikely(page->mapping != mapping)) {
 627                         unlock_page(page);
 628                         page_cache_release(page);
 629                         goto repeat;
 630                 }
 631                 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 632                 if (unlikely(vmf->pgoff >= size)) {
 633                         /*
 634                          * We have a struct page covering a hole in the file
 635                          * from a read fault and we've raced with a truncate
 636                          */
 637                         error = -EIO;
 638                         goto unlock_page;
 639                 }
 640         }
 641
 642         error = get_block(inode, block, &bh, 0);
 643         if (!error && (bh.b_size < PAGE_SIZE))
 644                 error = -EIO;           /* fs corruption? */
 645         if (error)
 646                 goto unlock_page;
 647
 648         if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 649                 if (vmf->flags & FAULT_FLAG_WRITE) {
 650                         error = get_block(inode, block, &bh, 1);
 651                         count_vm_event(PGMAJFAULT);
 652                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 653                         major = VM_FAULT_MAJOR;
 654                         if (!error && (bh.b_size < PAGE_SIZE))
 655                                 error = -EIO;
 656                         if (error)
 657                                 goto unlock_page;
 658                 } else {
 659                         return dax_load_hole(mapping, page, vmf);
 660                 }
 661         }
 662
 663         if (vmf->cow_page) {
 664                 struct page *new_page = vmf->cow_page;
 665                 if (buffer_written(&bh))
 666                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 667                 else
 668                         clear_user_highpage(new_page, vaddr);
 669                 if (error)
 670                         goto unlock_page;
 671                 vmf->page = page;
 672                 if (!page) {
 673                         i_mmap_lock_read(mapping);
 674                         /* Check we didn't race with truncate */
 675                         size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 676                                                                 PAGE_SHIFT;
 677                         if (vmf->pgoff >= size) {
 678                                 i_mmap_unlock_read(mapping);
 679                                 error = -EIO;
 680                                 goto out;
 681                         }
 682                 }
 683                 return VM_FAULT_LOCKED;
 684         }
 685
 686         /* Check we didn't race with a read fault installing a new page */
 687         if (!page && major)
 688                 page = find_lock_page(mapping, vmf->pgoff);
 689
 690         if (page) {
 691                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 692                                                         PAGE_CACHE_SIZE, 0);
 693                 delete_from_page_cache(page);
 694                 unlock_page(page);
 695                 page_cache_release(page);
 696                 page = NULL;
 697         }
 698
 699         /*
 700          * If we successfully insert the new mapping over an unwritten extent,
 701          * we need to ensure we convert the unwritten extent. If there is an
 702          * error inserting the mapping, the filesystem needs to leave it as
 703          * unwritten to prevent exposure of the stale underlying data to
 704          * userspace, but we still need to call the completion function so
 705          * the private resources on the mapping buffer can be released. We
 706          * indicate what the callback should do via the uptodate variable, same
 707          * as for normal BH based IO completions.
 708          */
 709         error = dax_insert_mapping(inode, &bh, vma, vmf);
 710         if (buffer_unwritten(&bh)) {
 711                 if (complete_unwritten)
 712                         complete_unwritten(&bh, !error);
 713                 else
 714                         WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 715         }
 716
 717  out:
 718         if (error == -ENOMEM)
 719                 return VM_FAULT_OOM | major;
 720         /* -EBUSY is fine, somebody else faulted on the same PTE */
 721         if ((error < 0) && (error != -EBUSY))
 722                 return VM_FAULT_SIGBUS | major;
 723         return VM_FAULT_NOPAGE | major;
 724
 725  unlock_page:
 726         if (page) {
 727                 unlock_page(page);
 728                 page_cache_release(page);
 729         }
 730         goto out;
 731 }
 732 EXPORT_SYMBOL(__dax_fault);
 733
 734 /**
 735  * dax_fault - handle a page fault on a DAX file
 736  * @vma: The virtual memory area where the fault occurred
 737  * @vmf: The description of the fault
 738  * @get_block: The filesystem method used to translate file offsets to blocks
 739  *
 740  * When a page fault occurs, filesystems may call this helper in their
 741  * fault handler for DAX files.
 742  */
 743 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 744               get_block_t get_block, dax_iodone_t complete_unwritten)
 745 {
 746         int result;
 747         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 748
 749         if (vmf->flags & FAULT_FLAG_WRITE) {
 750                 sb_start_pagefault(sb);
 751                 file_update_time(vma->vm_file);
 752         }
 753         result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 754         if (vmf->flags & FAULT_FLAG_WRITE)
 755                 sb_end_pagefault(sb);
 756
 757         return result;
 758 }
 759 EXPORT_SYMBOL_GPL(dax_fault);
 760
 761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 762 /*
 763  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 764  * more often than one might expect in the below function.
 765  */
 766 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 767
 768 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 769                 const char *reason, const char *fn)
 770 {
 771         if (bh) {
 772                 char bname[BDEVNAME_SIZE];
 773                 bdevname(bh->b_bdev, bname);
 774                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 775                         "length %zd fallback: %s\n", fn, current->comm,
 776                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 777                         bh->b_size, reason);
 778         } else {
 779                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 780                         current->comm, address, reason);
 781         }
 782 }
 783
 784 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 785
 786 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 787                 pmd_t *pmd, unsigned int flags, get_block_t get_block,
 788                 dax_iodone_t complete_unwritten)
 789 {
 790         struct file *file = vma->vm_file;
 791         struct address_space *mapping = file->f_mapping;
 792         struct inode *inode = mapping->host;
 793         struct buffer_head bh;
 794         unsigned blkbits = inode->i_blkbits;
 795         unsigned long pmd_addr = address & PMD_MASK;
 796         bool write = flags & FAULT_FLAG_WRITE;
 797         struct block_device *bdev;
 798         pgoff_t size, pgoff;
 799         sector_t block;
 800         int error, result = 0;
 801         bool alloc = false;
 802
 803         /* dax pmd mappings require pfn_t_devmap() */
 804         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 805                 return VM_FAULT_FALLBACK;
 806
 807         /* Fall back to PTEs if we're going to COW */
 808         if (write && !(vma->vm_flags & VM_SHARED)) {
 809                 split_huge_pmd(vma, pmd, address);
 810                 dax_pmd_dbg(NULL, address, "cow write");
 811                 return VM_FAULT_FALLBACK;
 812         }
 813         /* If the PMD would extend outside the VMA */
 814         if (pmd_addr < vma->vm_start) {
 815                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 816                 return VM_FAULT_FALLBACK;
 817         }
 818         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 819                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 820                 return VM_FAULT_FALLBACK;
 821         }
 822
 823         pgoff = linear_page_index(vma, pmd_addr);
 824         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 825         if (pgoff >= size)
 826                 return VM_FAULT_SIGBUS;
 827         /* If the PMD would cover blocks out of the file */
 828         if ((pgoff | PG_PMD_COLOUR) >= size) {
 829                 dax_pmd_dbg(NULL, address,
 830                                 "offset + huge page size > file size");
 831                 return VM_FAULT_FALLBACK;
 832         }
 833
 834         memset(&bh, 0, sizeof(bh));
 835         bh.b_bdev = inode->i_sb->s_bdev;
 836         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 837
 838         bh.b_size = PMD_SIZE;
 839
 840         if (get_block(inode, block, &bh, 0) != 0)
 841                 return VM_FAULT_SIGBUS;
 842
 843         if (!buffer_mapped(&bh) && write) {
 844                 if (get_block(inode, block, &bh, 1) != 0)
 845                         return VM_FAULT_SIGBUS;
 846                 alloc = true;
 847         }
 848
 849         bdev = bh.b_bdev;
 850
 851         /*
 852          * If the filesystem isn't willing to tell us the length of a hole,
 853          * just fall back to PTEs.  Calling get_block 512 times in a loop
 854          * would be silly.
 855          */
 856         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 857                 dax_pmd_dbg(&bh, address, "allocated block too small");
 858                 return VM_FAULT_FALLBACK;
 859         }
 860
 861         /*
 862          * If we allocated new storage, make sure no process has any
 863          * zero pages covering this hole
 864          */
 865         if (alloc) {
 866                 loff_t lstart = pgoff << PAGE_SHIFT;
 867                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 868
 869                 truncate_pagecache_range(inode, lstart, lend);
 870         }
 871
 872         i_mmap_lock_read(mapping);
 873
 874         /*
 875          * If a truncate happened while we were allocating blocks, we may
 876          * leave blocks allocated to the file that are beyond EOF.  We can't
 877          * take i_mutex here, so just leave them hanging; they'll be freed
 878          * when the file is deleted.
 879          */
 880         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 881         if (pgoff >= size) {
 882                 result = VM_FAULT_SIGBUS;
 883                 goto out;
 884         }
 885         if ((pgoff | PG_PMD_COLOUR) >= size) {
 886                 dax_pmd_dbg(&bh, address,
 887                                 "offset + huge page size > file size");
 888                 goto fallback;
 889         }
 890
 891         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 892                 spinlock_t *ptl;
 893                 pmd_t entry;
 894                 struct page *zero_page = get_huge_zero_page();
 895
 896                 if (unlikely(!zero_page)) {
 897                         dax_pmd_dbg(&bh, address, "no zero page");
 898                         goto fallback;
 899                 }
 900
 901                 ptl = pmd_lock(vma->vm_mm, pmd);
 902                 if (!pmd_none(*pmd)) {
 903                         spin_unlock(ptl);
 904                         dax_pmd_dbg(&bh, address, "pmd already present");
 905                         goto fallback;
 906                 }
 907
 908                 dev_dbg(part_to_dev(bdev->bd_part),
 909                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 910                                 __func__, current->comm, address,
 911                                 (unsigned long long) to_sector(&bh, inode));
 912
 913                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 914                 entry = pmd_mkhuge(entry);
 915                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 916                 result = VM_FAULT_NOPAGE;
 917                 spin_unlock(ptl);
 918         } else {
 919                 struct blk_dax_ctl dax = {
 920                         .sector = to_sector(&bh, inode),
 921                         .size = PMD_SIZE,
 922                 };
 923                 long length = dax_map_atomic(bdev, &dax);
 924
 925                 if (length < 0) {
 926                         result = VM_FAULT_SIGBUS;
 927                         goto out;
 928                 }
 929                 if (length < PMD_SIZE) {
 930                         dax_pmd_dbg(&bh, address, "dax-length too small");
 931                         dax_unmap_atomic(bdev, &dax);
 932                         goto fallback;
 933                 }
 934                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 935                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 936                         dax_unmap_atomic(bdev, &dax);
 937                         goto fallback;
 938                 }
 939
 940                 if (!pfn_t_devmap(dax.pfn)) {
 941                         dax_unmap_atomic(bdev, &dax);
 942                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 943                         goto fallback;
 944                 }
 945
 946                 if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 947                         clear_pmem(dax.addr, PMD_SIZE);
 948                         wmb_pmem();
 949                         count_vm_event(PGMAJFAULT);
 950                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 951                         result |= VM_FAULT_MAJOR;
 952                 }
 953                 dax_unmap_atomic(bdev, &dax);
 954
 955                 /*
 956                  * For PTE faults we insert a radix tree entry for reads, and
 957                  * leave it clean.  Then on the first write we dirty the radix
 958                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 959                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 960                  * call into get_block() to translate the pgoff to a sector in
 961                  * order to be able to create a new radix tree entry.
 962                  *
 963                  * The PMD path doesn't have an equivalent to
 964                  * dax_pfn_mkwrite(), though, so for a read followed by a
 965                  * write we traverse all the way through __dax_pmd_fault()
 966                  * twice.  This means we can just skip inserting a radix tree
 967                  * entry completely on the initial read and just wait until
 968                  * the write to insert a dirty entry.
 969                  */
 970                 if (write) {
 971                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 972                                         true, true);
 973                         if (error) {
 974                                 dax_pmd_dbg(&bh, address,
 975                                                 "PMD radix insertion failed");
 976                                 goto fallback;
 977                         }
 978                 }
 979
 980                 dev_dbg(part_to_dev(bdev->bd_part),
 981                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 982                                 __func__, current->comm, address,
 983                                 pfn_t_to_pfn(dax.pfn),
 984                                 (unsigned long long) dax.sector);
 985                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
 986                                 dax.pfn, write);
 987         }
 988
 989  out:
 990         i_mmap_unlock_read(mapping);
 991
 992         if (buffer_unwritten(&bh))
 993                 complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
 994
 995         return result;
 996
 997  fallback:
 998         count_vm_event(THP_FAULT_FALLBACK);
 999         result = VM_FAULT_FALLBACK;
1000         goto out;
1001 }
1002 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1003
1004 /**
1005  * dax_pmd_fault - handle a PMD fault on a DAX file
1006  * @vma: The virtual memory area where the fault occurred
1007  * @vmf: The description of the fault
1008  * @get_block: The filesystem method used to translate file offsets to blocks
1009  *
1010  * When a page fault occurs, filesystems may call this helper in their
1011  * pmd_fault handler for DAX files.
1012  */
1013 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1014                         pmd_t *pmd, unsigned int flags, get_block_t get_block,
1015                         dax_iodone_t complete_unwritten)
1016 {
1017         int result;
1018         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
1019
1020         if (flags & FAULT_FLAG_WRITE) {
1021                 sb_start_pagefault(sb);
1022                 file_update_time(vma->vm_file);
1023         }
1024         result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1025                                 complete_unwritten);
1026         if (flags & FAULT_FLAG_WRITE)
1027                 sb_end_pagefault(sb);
1028
1029         return result;
1030 }
1031 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1032 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1033
1034 /**
1035  * dax_pfn_mkwrite - handle first write to DAX page
1036  * @vma: The virtual memory area where the fault occurred
1037  * @vmf: The description of the fault
1038  */
1039 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1040 {
1041         struct file *file = vma->vm_file;
1042
1043         /*
1044          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1045          * RADIX_DAX_PTE entry already exists in the radix tree from a
1046          * previous call to __dax_fault().  We just want to look up that PTE
1047          * entry using vmf->pgoff and make sure the dirty tag is set.  This
1048          * saves us from having to make a call to get_block() here to look
1049          * up the sector.
1050          */
1051         dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false, true);
1052         return VM_FAULT_NOPAGE;
1053 }
1054 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1055
1056 /**
1057  * dax_zero_page_range - zero a range within a page of a DAX file
1058  * @inode: The file being truncated
1059  * @from: The file offset that is being truncated to
1060  * @length: The number of bytes to zero
1061  * @get_block: The filesystem method used to translate file offsets to blocks
1062  *
1063  * This function can be called by a filesystem when it is zeroing part of a
1064  * page in a DAX file.  This is intended for hole-punch operations.  If
1065  * you are truncating a file, the helper function dax_truncate_page() may be
1066  * more convenient.
1067  *
1068  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1069  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1070  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1071  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1072  * since the file might be mmapped.
1073  */
1074 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1075                                                         get_block_t get_block)
1076 {
1077         struct buffer_head bh;
1078         pgoff_t index = from >> PAGE_CACHE_SHIFT;
1079         unsigned offset = from & (PAGE_CACHE_SIZE-1);
1080         int err;
1081
1082         /* Block boundary? Nothing to do */
1083         if (!length)
1084                 return 0;
1085         BUG_ON((offset + length) > PAGE_CACHE_SIZE);
1086
1087         memset(&bh, 0, sizeof(bh));
1088         bh.b_bdev = inode->i_sb->s_bdev;
1089         bh.b_size = PAGE_CACHE_SIZE;
1090         err = get_block(inode, index, &bh, 0);
1091         if (err < 0)
1092                 return err;
1093         if (buffer_written(&bh)) {
1094                 struct block_device *bdev = bh.b_bdev;
1095                 struct blk_dax_ctl dax = {
1096                         .sector = to_sector(&bh, inode),
1097                         .size = PAGE_CACHE_SIZE,
1098                 };
1099
1100                 if (dax_map_atomic(bdev, &dax) < 0)
1101                         return PTR_ERR(dax.addr);
1102                 clear_pmem(dax.addr + offset, length);
1103                 wmb_pmem();
1104                 dax_unmap_atomic(bdev, &dax);
1105         }
1106
1107         return 0;
1108 }
1109 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1110
1111 /**
1112  * dax_truncate_page - handle a partial page being truncated in a DAX file
1113  * @inode: The file being truncated
1114  * @from: The file offset that is being truncated to
1115  * @get_block: The filesystem method used to translate file offsets to blocks
1116  *
1117  * Similar to block_truncate_page(), this function can be called by a
1118  * filesystem when it is truncating a DAX file to handle the partial page.
1119  *
1120  * We work in terms of PAGE_CACHE_SIZE here for commonality with
1121  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1122  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1123  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1124  * since the file might be mmapped.
1125  */
1126 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1127 {
1128         unsigned length = PAGE_CACHE_ALIGN(from) - from;
1129         return dax_zero_page_range(inode, from, length, get_block);
1130 }
1131 EXPORT_SYMBOL_GPL(dax_truncate_page);