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