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 #define RADIX_DAX_MASK  0xf
  36 #define RADIX_DAX_SHIFT 4
  37 #define RADIX_DAX_PTE  (0x4 | RADIX_TREE_EXCEPTIONAL_ENTRY)
  38 #define RADIX_DAX_PMD  (0x8 | RADIX_TREE_EXCEPTIONAL_ENTRY)
  39 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_MASK)
  40 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
  41 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
  42                 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE)))
  43
  44 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  45 {
  46         struct request_queue *q = bdev->bd_queue;
  47         long rc = -EIO;
  48
  49         dax->addr = (void __pmem *) ERR_PTR(-EIO);
  50         if (blk_queue_enter(q, true) != 0)
  51                 return rc;
  52
  53         rc = bdev_direct_access(bdev, dax);
  54         if (rc < 0) {
  55                 dax->addr = (void __pmem *) ERR_PTR(rc);
  56                 blk_queue_exit(q);
  57                 return rc;
  58         }
  59         return rc;
  60 }
  61
  62 static void dax_unmap_atomic(struct block_device *bdev,
  63                 const struct blk_dax_ctl *dax)
  64 {
  65         if (IS_ERR(dax->addr))
  66                 return;
  67         blk_queue_exit(bdev->bd_queue);
  68 }
  69
  70 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
  71 {
  72         struct page *page = alloc_pages(GFP_KERNEL, 0);
  73         struct blk_dax_ctl dax = {
  74                 .size = PAGE_SIZE,
  75                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
  76         };
  77         long rc;
  78
  79         if (!page)
  80                 return ERR_PTR(-ENOMEM);
  81
  82         rc = dax_map_atomic(bdev, &dax);
  83         if (rc < 0)
  84                 return ERR_PTR(rc);
  85         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
  86         dax_unmap_atomic(bdev, &dax);
  87         return page;
  88 }
  89
  90 /*
  91  * dax_clear_sectors() is called from within transaction context from XFS,
  92  * and hence this means the stack from this point must follow GFP_NOFS
  93  * semantics for all operations.
  94  */
  95 int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
  96 {
  97         struct blk_dax_ctl dax = {
  98                 .sector = _sector,
  99                 .size = _size,
 100         };
 101
 102         might_sleep();
 103         do {
 104                 long count, sz;
 105
 106                 count = dax_map_atomic(bdev, &dax);
 107                 if (count < 0)
 108                         return count;
 109                 sz = min_t(long, count, SZ_128K);
 110                 clear_pmem(dax.addr, sz);
 111                 dax.size -= sz;
 112                 dax.sector += sz / 512;
 113                 dax_unmap_atomic(bdev, &dax);
 114                 cond_resched();
 115         } while (dax.size);
 116
 117         wmb_pmem();
 118         return 0;
 119 }
 120 EXPORT_SYMBOL_GPL(dax_clear_sectors);
 121
 122 static bool buffer_written(struct buffer_head *bh)
 123 {
 124         return buffer_mapped(bh) && !buffer_unwritten(bh);
 125 }
 126
 127 /*
 128  * When ext4 encounters a hole, it returns without modifying the buffer_head
 129  * which means that we can't trust b_size.  To cope with this, we set b_state
 130  * to 0 before calling get_block and, if any bit is set, we know we can trust
 131  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 132  * and would save us time calling get_block repeatedly.
 133  */
 134 static bool buffer_size_valid(struct buffer_head *bh)
 135 {
 136         return bh->b_state != 0;
 137 }
 138
 139
 140 static sector_t to_sector(const struct buffer_head *bh,
 141                 const struct inode *inode)
 142 {
 143         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 144
 145         return sector;
 146 }
 147
 148 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 149                       loff_t start, loff_t end, get_block_t get_block,
 150                       struct buffer_head *bh)
 151 {
 152         loff_t pos = start, max = start, bh_max = start;
 153         bool hole = false, need_wmb = false;
 154         struct block_device *bdev = NULL;
 155         int rw = iov_iter_rw(iter), rc;
 156         long map_len = 0;
 157         struct blk_dax_ctl dax = {
 158                 .addr = (void __pmem *) ERR_PTR(-EIO),
 159         };
 160         unsigned blkbits = inode->i_blkbits;
 161         sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
 162                                                                 >> blkbits;
 163
 164         if (rw == READ)
 165                 end = min(end, i_size_read(inode));
 166
 167         while (pos < end) {
 168                 size_t len;
 169                 if (pos == max) {
 170                         long page = pos >> PAGE_SHIFT;
 171                         sector_t block = page << (PAGE_SHIFT - blkbits);
 172                         unsigned first = pos - (block << blkbits);
 173                         long size;
 174
 175                         if (pos == bh_max) {
 176                                 bh->b_size = PAGE_ALIGN(end - pos);
 177                                 bh->b_state = 0;
 178                                 rc = get_block(inode, block, bh, rw == WRITE);
 179                                 if (rc)
 180                                         break;
 181                                 if (!buffer_size_valid(bh))
 182                                         bh->b_size = 1 << blkbits;
 183                                 bh_max = pos - first + bh->b_size;
 184                                 bdev = bh->b_bdev;
 185                                 /*
 186                                  * We allow uninitialized buffers for writes
 187                                  * beyond EOF as those cannot race with faults
 188                                  */
 189                                 WARN_ON_ONCE(
 190                                         (buffer_new(bh) && block < file_blks) ||
 191                                         (rw == WRITE && buffer_unwritten(bh)));
 192                         } else {
 193                                 unsigned done = bh->b_size -
 194                                                 (bh_max - (pos - first));
 195                                 bh->b_blocknr += done >> blkbits;
 196                                 bh->b_size -= done;
 197                         }
 198
 199                         hole = rw == READ && !buffer_written(bh);
 200                         if (hole) {
 201                                 size = bh->b_size - first;
 202                         } else {
 203                                 dax_unmap_atomic(bdev, &dax);
 204                                 dax.sector = to_sector(bh, inode);
 205                                 dax.size = bh->b_size;
 206                                 map_len = dax_map_atomic(bdev, &dax);
 207                                 if (map_len < 0) {
 208                                         rc = map_len;
 209                                         break;
 210                                 }
 211                                 dax.addr += first;
 212                                 size = map_len - first;
 213                         }
 214                         max = min(pos + size, end);
 215                 }
 216
 217                 if (iov_iter_rw(iter) == WRITE) {
 218                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 219                         need_wmb = true;
 220                 } else if (!hole)
 221                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 222                                         iter);
 223                 else
 224                         len = iov_iter_zero(max - pos, iter);
 225
 226                 if (!len) {
 227                         rc = -EFAULT;
 228                         break;
 229                 }
 230
 231                 pos += len;
 232                 if (!IS_ERR(dax.addr))
 233                         dax.addr += len;
 234         }
 235
 236         if (need_wmb)
 237                 wmb_pmem();
 238         dax_unmap_atomic(bdev, &dax);
 239
 240         return (pos == start) ? rc : pos - start;
 241 }
 242
 243 /**
 244  * dax_do_io - Perform I/O to a DAX file
 245  * @iocb: The control block for this I/O
 246  * @inode: The file which the I/O is directed at
 247  * @iter: The addresses to do I/O from or to
 248  * @pos: The file offset where the I/O starts
 249  * @get_block: The filesystem method used to translate file offsets to blocks
 250  * @end_io: A filesystem callback for I/O completion
 251  * @flags: See below
 252  *
 253  * This function uses the same locking scheme as do_blockdev_direct_IO:
 254  * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 255  * caller for writes.  For reads, we take and release the i_mutex ourselves.
 256  * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 257  * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 258  * is in progress.
 259  */
 260 ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 261                   struct iov_iter *iter, loff_t pos, get_block_t get_block,
 262                   dio_iodone_t end_io, int flags)
 263 {
 264         struct buffer_head bh;
 265         ssize_t retval = -EINVAL;
 266         loff_t end = pos + iov_iter_count(iter);
 267
 268         memset(&bh, 0, sizeof(bh));
 269         bh.b_bdev = inode->i_sb->s_bdev;
 270
 271         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 272                 inode_lock(inode);
 273
 274         /* Protects against truncate */
 275         if (!(flags & DIO_SKIP_DIO_COUNT))
 276                 inode_dio_begin(inode);
 277
 278         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 279
 280         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 281                 inode_unlock(inode);
 282
 283         if (end_io) {
 284                 int err;
 285
 286                 err = end_io(iocb, pos, retval, bh.b_private);
 287                 if (err)
 288                         retval = err;
 289         }
 290
 291         if (!(flags & DIO_SKIP_DIO_COUNT))
 292                 inode_dio_end(inode);
 293         return retval;
 294 }
 295 EXPORT_SYMBOL_GPL(dax_do_io);
 296
 297 /*
 298  * The user has performed a load from a hole in the file.  Allocating
 299  * a new page in the file would cause excessive storage usage for
 300  * workloads with sparse files.  We allocate a page cache page instead.
 301  * We'll kick it out of the page cache if it's ever written to,
 302  * otherwise it will simply fall out of the page cache under memory
 303  * pressure without ever having been dirtied.
 304  */
 305 static int dax_load_hole(struct address_space *mapping, struct page *page,
 306                                                         struct vm_fault *vmf)
 307 {
 308         if (!page)
 309                 page = find_or_create_page(mapping, vmf->pgoff,
 310                                                 GFP_KERNEL | __GFP_ZERO);
 311         if (!page)
 312                 return VM_FAULT_OOM;
 313
 314         vmf->page = page;
 315         return VM_FAULT_LOCKED;
 316 }
 317
 318 static int copy_user_bh(struct page *to, struct inode *inode,
 319                 struct buffer_head *bh, unsigned long vaddr)
 320 {
 321         struct blk_dax_ctl dax = {
 322                 .sector = to_sector(bh, inode),
 323                 .size = bh->b_size,
 324         };
 325         struct block_device *bdev = bh->b_bdev;
 326         void *vto;
 327
 328         if (dax_map_atomic(bdev, &dax) < 0)
 329                 return PTR_ERR(dax.addr);
 330         vto = kmap_atomic(to);
 331         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 332         kunmap_atomic(vto);
 333         dax_unmap_atomic(bdev, &dax);
 334         return 0;
 335 }
 336
 337 #define NO_SECTOR -1
 338 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 339
 340 static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 341                 sector_t sector, bool pmd_entry, bool dirty)
 342 {
 343         struct radix_tree_root *page_tree = &mapping->page_tree;
 344         pgoff_t pmd_index = DAX_PMD_INDEX(index);
 345         int type, error = 0;
 346         void *entry;
 347
 348         WARN_ON_ONCE(pmd_entry && !dirty);
 349         if (dirty)
 350                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 351
 352         spin_lock_irq(&mapping->tree_lock);
 353
 354         entry = radix_tree_lookup(page_tree, pmd_index);
 355         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 356                 index = pmd_index;
 357                 goto dirty;
 358         }
 359
 360         entry = radix_tree_lookup(page_tree, index);
 361         if (entry) {
 362                 type = RADIX_DAX_TYPE(entry);
 363                 if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 364                                         type != RADIX_DAX_PMD)) {
 365                         error = -EIO;
 366                         goto unlock;
 367                 }
 368
 369                 if (!pmd_entry || type == RADIX_DAX_PMD)
 370                         goto dirty;
 371
 372                 /*
 373                  * We only insert dirty PMD entries into the radix tree.  This
 374                  * means we don't need to worry about removing a dirty PTE
 375                  * entry and inserting a clean PMD entry, thus reducing the
 376                  * range we would flush with a follow-up fsync/msync call.
 377                  */
 378                 radix_tree_delete(&mapping->page_tree, index);
 379                 mapping->nrexceptional--;
 380         }
 381
 382         if (sector == NO_SECTOR) {
 383                 /*
 384                  * This can happen during correct operation if our pfn_mkwrite
 385                  * fault raced against a hole punch operation.  If this
 386                  * happens the pte that was hole punched will have been
 387                  * unmapped and the radix tree entry will have been removed by
 388                  * the time we are called, but the call will still happen.  We
 389                  * will return all the way up to wp_pfn_shared(), where the
 390                  * pte_same() check will fail, eventually causing page fault
 391                  * to be retried by the CPU.
 392                  */
 393                 goto unlock;
 394         }
 395
 396         error = radix_tree_insert(page_tree, index,
 397                         RADIX_DAX_ENTRY(sector, pmd_entry));
 398         if (error)
 399                 goto unlock;
 400
 401         mapping->nrexceptional++;
 402  dirty:
 403         if (dirty)
 404                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 405  unlock:
 406         spin_unlock_irq(&mapping->tree_lock);
 407         return error;
 408 }
 409
 410 static int dax_writeback_one(struct block_device *bdev,
 411                 struct address_space *mapping, pgoff_t index, void *entry)
 412 {
 413         struct radix_tree_root *page_tree = &mapping->page_tree;
 414         int type = RADIX_DAX_TYPE(entry);
 415         struct radix_tree_node *node;
 416         struct blk_dax_ctl dax;
 417         void **slot;
 418         int ret = 0;
 419
 420         spin_lock_irq(&mapping->tree_lock);
 421         /*
 422          * Regular page slots are stabilized by the page lock even
 423          * without the tree itself locked.  These unlocked entries
 424          * need verification under the tree lock.
 425          */
 426         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 427                 goto unlock;
 428         if (*slot != entry)
 429                 goto unlock;
 430
 431         /* another fsync thread may have already written back this entry */
 432         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 433                 goto unlock;
 434
 435         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 436                 ret = -EIO;
 437                 goto unlock;
 438         }
 439
 440         dax.sector = RADIX_DAX_SECTOR(entry);
 441         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 442         spin_unlock_irq(&mapping->tree_lock);
 443
 444         /*
 445          * We cannot hold tree_lock while calling dax_map_atomic() because it
 446          * eventually calls cond_resched().
 447          */
 448         ret = dax_map_atomic(bdev, &dax);
 449         if (ret < 0)
 450                 return ret;
 451
 452         if (WARN_ON_ONCE(ret < dax.size)) {
 453                 ret = -EIO;
 454                 goto unmap;
 455         }
 456
 457         wb_cache_pmem(dax.addr, dax.size);
 458
 459         spin_lock_irq(&mapping->tree_lock);
 460         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 461         spin_unlock_irq(&mapping->tree_lock);
 462  unmap:
 463         dax_unmap_atomic(bdev, &dax);
 464         return ret;
 465
 466  unlock:
 467         spin_unlock_irq(&mapping->tree_lock);
 468         return ret;
 469 }
 470
 471 /*
 472  * Flush the mapping to the persistent domain within the byte range of [start,
 473  * end]. This is required by data integrity operations to ensure file data is
 474  * on persistent storage prior to completion of the operation.
 475  */
 476 int dax_writeback_mapping_range(struct address_space *mapping,
 477                 struct block_device *bdev, struct writeback_control *wbc)
 478 {
 479         struct inode *inode = mapping->host;
 480         pgoff_t start_index, end_index, pmd_index;
 481         pgoff_t indices[PAGEVEC_SIZE];
 482         struct pagevec pvec;
 483         bool done = false;
 484         int i, ret = 0;
 485         void *entry;
 486
 487         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 488                 return -EIO;
 489
 490         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 491                 return 0;
 492
 493         start_index = wbc->range_start >> PAGE_SHIFT;
 494         end_index = wbc->range_end >> PAGE_SHIFT;
 495         pmd_index = DAX_PMD_INDEX(start_index);
 496
 497         rcu_read_lock();
 498         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 499         rcu_read_unlock();
 500
 501         /* see if the start of our range is covered by a PMD entry */
 502         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 503                 start_index = pmd_index;
 504
 505         tag_pages_for_writeback(mapping, start_index, end_index);
 506
 507         pagevec_init(&pvec, 0);
 508         while (!done) {
 509                 pvec.nr = find_get_entries_tag(mapping, start_index,
 510                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 511                                 pvec.pages, indices);
 512
 513                 if (pvec.nr == 0)
 514                         break;
 515
 516                 for (i = 0; i < pvec.nr; i++) {
 517                         if (indices[i] > end_index) {
 518                                 done = true;
 519                                 break;
 520                         }
 521
 522                         ret = dax_writeback_one(bdev, mapping, indices[i],
 523                                         pvec.pages[i]);
 524                         if (ret < 0)
 525                                 return ret;
 526                 }
 527         }
 528         wmb_pmem();
 529         return 0;
 530 }
 531 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 532
 533 static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 534                         struct vm_area_struct *vma, struct vm_fault *vmf)
 535 {
 536         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 537         struct address_space *mapping = inode->i_mapping;
 538         struct block_device *bdev = bh->b_bdev;
 539         struct blk_dax_ctl dax = {
 540                 .sector = to_sector(bh, inode),
 541                 .size = bh->b_size,
 542         };
 543         int error;
 544
 545         i_mmap_lock_read(mapping);
 546
 547         if (dax_map_atomic(bdev, &dax) < 0) {
 548                 error = PTR_ERR(dax.addr);
 549                 goto out;
 550         }
 551         dax_unmap_atomic(bdev, &dax);
 552
 553         error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 554                         vmf->flags & FAULT_FLAG_WRITE);
 555         if (error)
 556                 goto out;
 557
 558         error = vm_insert_mixed(vma, vaddr, dax.pfn);
 559
 560  out:
 561         i_mmap_unlock_read(mapping);
 562
 563         return error;
 564 }
 565
 566 /**
 567  * __dax_fault - handle a page fault on a DAX file
 568  * @vma: The virtual memory area where the fault occurred
 569  * @vmf: The description of the fault
 570  * @get_block: The filesystem method used to translate file offsets to blocks
 571  *
 572  * When a page fault occurs, filesystems may call this helper in their
 573  * fault handler for DAX files. __dax_fault() assumes the caller has done all
 574  * the necessary locking for the page fault to proceed successfully.
 575  */
 576 int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 577                         get_block_t get_block)
 578 {
 579         struct file *file = vma->vm_file;
 580         struct address_space *mapping = file->f_mapping;
 581         struct inode *inode = mapping->host;
 582         struct page *page;
 583         struct buffer_head bh;
 584         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 585         unsigned blkbits = inode->i_blkbits;
 586         sector_t block;
 587         pgoff_t size;
 588         int error;
 589         int major = 0;
 590
 591         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 592         if (vmf->pgoff >= size)
 593                 return VM_FAULT_SIGBUS;
 594
 595         memset(&bh, 0, sizeof(bh));
 596         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 597         bh.b_bdev = inode->i_sb->s_bdev;
 598         bh.b_size = PAGE_SIZE;
 599
 600  repeat:
 601         page = find_get_page(mapping, vmf->pgoff);
 602         if (page) {
 603                 if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 604                         put_page(page);
 605                         return VM_FAULT_RETRY;
 606                 }
 607                 if (unlikely(page->mapping != mapping)) {
 608                         unlock_page(page);
 609                         put_page(page);
 610                         goto repeat;
 611                 }
 612         }
 613
 614         error = get_block(inode, block, &bh, 0);
 615         if (!error && (bh.b_size < PAGE_SIZE))
 616                 error = -EIO;           /* fs corruption? */
 617         if (error)
 618                 goto unlock_page;
 619
 620         if (!buffer_mapped(&bh) && !vmf->cow_page) {
 621                 if (vmf->flags & FAULT_FLAG_WRITE) {
 622                         error = get_block(inode, block, &bh, 1);
 623                         count_vm_event(PGMAJFAULT);
 624                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 625                         major = VM_FAULT_MAJOR;
 626                         if (!error && (bh.b_size < PAGE_SIZE))
 627                                 error = -EIO;
 628                         if (error)
 629                                 goto unlock_page;
 630                 } else {
 631                         return dax_load_hole(mapping, page, vmf);
 632                 }
 633         }
 634
 635         if (vmf->cow_page) {
 636                 struct page *new_page = vmf->cow_page;
 637                 if (buffer_written(&bh))
 638                         error = copy_user_bh(new_page, inode, &bh, vaddr);
 639                 else
 640                         clear_user_highpage(new_page, vaddr);
 641                 if (error)
 642                         goto unlock_page;
 643                 vmf->page = page;
 644                 if (!page)
 645                         i_mmap_lock_read(mapping);
 646                 return VM_FAULT_LOCKED;
 647         }
 648
 649         /* Check we didn't race with a read fault installing a new page */
 650         if (!page && major)
 651                 page = find_lock_page(mapping, vmf->pgoff);
 652
 653         if (page) {
 654                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 655                                                         PAGE_SIZE, 0);
 656                 delete_from_page_cache(page);
 657                 unlock_page(page);
 658                 put_page(page);
 659                 page = NULL;
 660         }
 661
 662         /* Filesystem should not return unwritten buffers to us! */
 663         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 664         error = dax_insert_mapping(inode, &bh, vma, vmf);
 665
 666  out:
 667         if (error == -ENOMEM)
 668                 return VM_FAULT_OOM | major;
 669         /* -EBUSY is fine, somebody else faulted on the same PTE */
 670         if ((error < 0) && (error != -EBUSY))
 671                 return VM_FAULT_SIGBUS | major;
 672         return VM_FAULT_NOPAGE | major;
 673
 674  unlock_page:
 675         if (page) {
 676                 unlock_page(page);
 677                 put_page(page);
 678         }
 679         goto out;
 680 }
 681 EXPORT_SYMBOL(__dax_fault);
 682
 683 /**
 684  * dax_fault - handle a page fault on a DAX file
 685  * @vma: The virtual memory area where the fault occurred
 686  * @vmf: The description of the fault
 687  * @get_block: The filesystem method used to translate file offsets to blocks
 688  *
 689  * When a page fault occurs, filesystems may call this helper in their
 690  * fault handler for DAX files.
 691  */
 692 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 693               get_block_t get_block)
 694 {
 695         int result;
 696         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 697
 698         if (vmf->flags & FAULT_FLAG_WRITE) {
 699                 sb_start_pagefault(sb);
 700                 file_update_time(vma->vm_file);
 701         }
 702         result = __dax_fault(vma, vmf, get_block);
 703         if (vmf->flags & FAULT_FLAG_WRITE)
 704                 sb_end_pagefault(sb);
 705
 706         return result;
 707 }
 708 EXPORT_SYMBOL_GPL(dax_fault);
 709
 710 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 711 /*
 712  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 713  * more often than one might expect in the below function.
 714  */
 715 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 716
 717 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 718                 const char *reason, const char *fn)
 719 {
 720         if (bh) {
 721                 char bname[BDEVNAME_SIZE];
 722                 bdevname(bh->b_bdev, bname);
 723                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 724                         "length %zd fallback: %s\n", fn, current->comm,
 725                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 726                         bh->b_size, reason);
 727         } else {
 728                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 729                         current->comm, address, reason);
 730         }
 731 }
 732
 733 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 734
 735 int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 736                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 737 {
 738         struct file *file = vma->vm_file;
 739         struct address_space *mapping = file->f_mapping;
 740         struct inode *inode = mapping->host;
 741         struct buffer_head bh;
 742         unsigned blkbits = inode->i_blkbits;
 743         unsigned long pmd_addr = address & PMD_MASK;
 744         bool write = flags & FAULT_FLAG_WRITE;
 745         struct block_device *bdev;
 746         pgoff_t size, pgoff;
 747         sector_t block;
 748         int error, result = 0;
 749         bool alloc = false;
 750
 751         /* dax pmd mappings require pfn_t_devmap() */
 752         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 753                 return VM_FAULT_FALLBACK;
 754
 755         /* Fall back to PTEs if we're going to COW */
 756         if (write && !(vma->vm_flags & VM_SHARED)) {
 757                 split_huge_pmd(vma, pmd, address);
 758                 dax_pmd_dbg(NULL, address, "cow write");
 759                 return VM_FAULT_FALLBACK;
 760         }
 761         /* If the PMD would extend outside the VMA */
 762         if (pmd_addr < vma->vm_start) {
 763                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 764                 return VM_FAULT_FALLBACK;
 765         }
 766         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 767                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 768                 return VM_FAULT_FALLBACK;
 769         }
 770
 771         pgoff = linear_page_index(vma, pmd_addr);
 772         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 773         if (pgoff >= size)
 774                 return VM_FAULT_SIGBUS;
 775         /* If the PMD would cover blocks out of the file */
 776         if ((pgoff | PG_PMD_COLOUR) >= size) {
 777                 dax_pmd_dbg(NULL, address,
 778                                 "offset + huge page size > file size");
 779                 return VM_FAULT_FALLBACK;
 780         }
 781
 782         memset(&bh, 0, sizeof(bh));
 783         bh.b_bdev = inode->i_sb->s_bdev;
 784         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 785
 786         bh.b_size = PMD_SIZE;
 787
 788         if (get_block(inode, block, &bh, 0) != 0)
 789                 return VM_FAULT_SIGBUS;
 790
 791         if (!buffer_mapped(&bh) && write) {
 792                 if (get_block(inode, block, &bh, 1) != 0)
 793                         return VM_FAULT_SIGBUS;
 794                 alloc = true;
 795                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 796         }
 797
 798         bdev = bh.b_bdev;
 799
 800         /*
 801          * If the filesystem isn't willing to tell us the length of a hole,
 802          * just fall back to PTEs.  Calling get_block 512 times in a loop
 803          * would be silly.
 804          */
 805         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 806                 dax_pmd_dbg(&bh, address, "allocated block too small");
 807                 return VM_FAULT_FALLBACK;
 808         }
 809
 810         /*
 811          * If we allocated new storage, make sure no process has any
 812          * zero pages covering this hole
 813          */
 814         if (alloc) {
 815                 loff_t lstart = pgoff << PAGE_SHIFT;
 816                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 817
 818                 truncate_pagecache_range(inode, lstart, lend);
 819         }
 820
 821         i_mmap_lock_read(mapping);
 822
 823         if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 824                 spinlock_t *ptl;
 825                 pmd_t entry;
 826                 struct page *zero_page = get_huge_zero_page();
 827
 828                 if (unlikely(!zero_page)) {
 829                         dax_pmd_dbg(&bh, address, "no zero page");
 830                         goto fallback;
 831                 }
 832
 833                 ptl = pmd_lock(vma->vm_mm, pmd);
 834                 if (!pmd_none(*pmd)) {
 835                         spin_unlock(ptl);
 836                         dax_pmd_dbg(&bh, address, "pmd already present");
 837                         goto fallback;
 838                 }
 839
 840                 dev_dbg(part_to_dev(bdev->bd_part),
 841                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 842                                 __func__, current->comm, address,
 843                                 (unsigned long long) to_sector(&bh, inode));
 844
 845                 entry = mk_pmd(zero_page, vma->vm_page_prot);
 846                 entry = pmd_mkhuge(entry);
 847                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 848                 result = VM_FAULT_NOPAGE;
 849                 spin_unlock(ptl);
 850         } else {
 851                 struct blk_dax_ctl dax = {
 852                         .sector = to_sector(&bh, inode),
 853                         .size = PMD_SIZE,
 854                 };
 855                 long length = dax_map_atomic(bdev, &dax);
 856
 857                 if (length < 0) {
 858                         result = VM_FAULT_SIGBUS;
 859                         goto out;
 860                 }
 861                 if (length < PMD_SIZE) {
 862                         dax_pmd_dbg(&bh, address, "dax-length too small");
 863                         dax_unmap_atomic(bdev, &dax);
 864                         goto fallback;
 865                 }
 866                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 867                         dax_pmd_dbg(&bh, address, "pfn unaligned");
 868                         dax_unmap_atomic(bdev, &dax);
 869                         goto fallback;
 870                 }
 871
 872                 if (!pfn_t_devmap(dax.pfn)) {
 873                         dax_unmap_atomic(bdev, &dax);
 874                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
 875                         goto fallback;
 876                 }
 877                 dax_unmap_atomic(bdev, &dax);
 878
 879                 /*
 880                  * For PTE faults we insert a radix tree entry for reads, and
 881                  * leave it clean.  Then on the first write we dirty the radix
 882                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
 883                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 884                  * call into get_block() to translate the pgoff to a sector in
 885                  * order to be able to create a new radix tree entry.
 886                  *
 887                  * The PMD path doesn't have an equivalent to
 888                  * dax_pfn_mkwrite(), though, so for a read followed by a
 889                  * write we traverse all the way through __dax_pmd_fault()
 890                  * twice.  This means we can just skip inserting a radix tree
 891                  * entry completely on the initial read and just wait until
 892                  * the write to insert a dirty entry.
 893                  */
 894                 if (write) {
 895                         error = dax_radix_entry(mapping, pgoff, dax.sector,
 896                                         true, true);
 897                         if (error) {
 898                                 dax_pmd_dbg(&bh, address,
 899                                                 "PMD radix insertion failed");
 900                                 goto fallback;
 901                         }
 902                 }
 903
 904                 dev_dbg(part_to_dev(bdev->bd_part),
 905                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
 906                                 __func__, current->comm, address,
 907                                 pfn_t_to_pfn(dax.pfn),
 908                                 (unsigned long long) dax.sector);
 909                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
 910                                 dax.pfn, write);
 911         }
 912
 913  out:
 914         i_mmap_unlock_read(mapping);
 915
 916         return result;
 917
 918  fallback:
 919         count_vm_event(THP_FAULT_FALLBACK);
 920         result = VM_FAULT_FALLBACK;
 921         goto out;
 922 }
 923 EXPORT_SYMBOL_GPL(__dax_pmd_fault);
 924
 925 /**
 926  * dax_pmd_fault - handle a PMD fault on a DAX file
 927  * @vma: The virtual memory area where the fault occurred
 928  * @vmf: The description of the fault
 929  * @get_block: The filesystem method used to translate file offsets to blocks
 930  *
 931  * When a page fault occurs, filesystems may call this helper in their
 932  * pmd_fault handler for DAX files.
 933  */
 934 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 935                         pmd_t *pmd, unsigned int flags, get_block_t get_block)
 936 {
 937         int result;
 938         struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 939
 940         if (flags & FAULT_FLAG_WRITE) {
 941                 sb_start_pagefault(sb);
 942                 file_update_time(vma->vm_file);
 943         }
 944         result = __dax_pmd_fault(vma, address, pmd, flags, get_block);
 945         if (flags & FAULT_FLAG_WRITE)
 946                 sb_end_pagefault(sb);
 947
 948         return result;
 949 }
 950 EXPORT_SYMBOL_GPL(dax_pmd_fault);
 951 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 952
 953 /**
 954  * dax_pfn_mkwrite - handle first write to DAX page
 955  * @vma: The virtual memory area where the fault occurred
 956  * @vmf: The description of the fault
 957  */
 958 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 959 {
 960         struct file *file = vma->vm_file;
 961         int error;
 962
 963         /*
 964          * We pass NO_SECTOR to dax_radix_entry() because we expect that a
 965          * RADIX_DAX_PTE entry already exists in the radix tree from a
 966          * previous call to __dax_fault().  We just want to look up that PTE
 967          * entry using vmf->pgoff and make sure the dirty tag is set.  This
 968          * saves us from having to make a call to get_block() here to look
 969          * up the sector.
 970          */
 971         error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
 972                         true);
 973
 974         if (error == -ENOMEM)
 975                 return VM_FAULT_OOM;
 976         if (error)
 977                 return VM_FAULT_SIGBUS;
 978         return VM_FAULT_NOPAGE;
 979 }
 980 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 981
 982 /**
 983  * dax_zero_page_range - zero a range within a page of a DAX file
 984  * @inode: The file being truncated
 985  * @from: The file offset that is being truncated to
 986  * @length: The number of bytes to zero
 987  * @get_block: The filesystem method used to translate file offsets to blocks
 988  *
 989  * This function can be called by a filesystem when it is zeroing part of a
 990  * page in a DAX file.  This is intended for hole-punch operations.  If
 991  * you are truncating a file, the helper function dax_truncate_page() may be
 992  * more convenient.
 993  *
 994  * We work in terms of PAGE_SIZE here for commonality with
 995  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
 996  * took care of disposing of the unnecessary blocks.  Even if the filesystem
 997  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
 998  * since the file might be mmapped.
 999  */
1000 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1001                                                         get_block_t get_block)
1002 {
1003         struct buffer_head bh;
1004         pgoff_t index = from >> PAGE_SHIFT;
1005         unsigned offset = from & (PAGE_SIZE-1);
1006         int err;
1007
1008         /* Block boundary? Nothing to do */
1009         if (!length)
1010                 return 0;
1011         BUG_ON((offset + length) > PAGE_SIZE);
1012
1013         memset(&bh, 0, sizeof(bh));
1014         bh.b_bdev = inode->i_sb->s_bdev;
1015         bh.b_size = PAGE_SIZE;
1016         err = get_block(inode, index, &bh, 0);
1017         if (err < 0)
1018                 return err;
1019         if (buffer_written(&bh)) {
1020                 struct block_device *bdev = bh.b_bdev;
1021                 struct blk_dax_ctl dax = {
1022                         .sector = to_sector(&bh, inode),
1023                         .size = PAGE_SIZE,
1024                 };
1025
1026                 if (dax_map_atomic(bdev, &dax) < 0)
1027                         return PTR_ERR(dax.addr);
1028                 clear_pmem(dax.addr + offset, length);
1029                 wmb_pmem();
1030                 dax_unmap_atomic(bdev, &dax);
1031         }
1032
1033         return 0;
1034 }
1035 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1036
1037 /**
1038  * dax_truncate_page - handle a partial page being truncated in a DAX file
1039  * @inode: The file being truncated
1040  * @from: The file offset that is being truncated to
1041  * @get_block: The filesystem method used to translate file offsets to blocks
1042  *
1043  * Similar to block_truncate_page(), this function can be called by a
1044  * filesystem when it is truncating a DAX file to handle the partial page.
1045  *
1046  * We work in terms of PAGE_SIZE here for commonality with
1047  * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1048  * took care of disposing of the unnecessary blocks.  Even if the filesystem
1049  * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1050  * since the file might be mmapped.
1051  */
1052 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1053 {
1054         unsigned length = PAGE_ALIGN(from) - from;
1055         return dax_zero_page_range(inode, from, length, get_block);
1056 }
1057 EXPORT_SYMBOL_GPL(dax_truncate_page);