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 #include <linux/iomap.h>
  35 #include "internal.h"
  36
  37 /*
  38  * We use lowest available bit in exceptional entry for locking, other two
  39  * bits to determine entry type. In total 3 special bits.
  40  */
  41 #define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 3)
  42 #define RADIX_DAX_PTE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
  43 #define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
  44 #define RADIX_DAX_TYPE_MASK (RADIX_DAX_PTE | RADIX_DAX_PMD)
  45 #define RADIX_DAX_TYPE(entry) ((unsigned long)entry & RADIX_DAX_TYPE_MASK)
  46 #define RADIX_DAX_SECTOR(entry) (((unsigned long)entry >> RADIX_DAX_SHIFT))
  47 #define RADIX_DAX_ENTRY(sector, pmd) ((void *)((unsigned long)sector << \
  48                 RADIX_DAX_SHIFT | (pmd ? RADIX_DAX_PMD : RADIX_DAX_PTE) | \
  49                 RADIX_TREE_EXCEPTIONAL_ENTRY))
  50
  51 /* We choose 4096 entries - same as per-zone page wait tables */
  52 #define DAX_WAIT_TABLE_BITS 12
  53 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  54
  55 wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  56
  57 static int __init init_dax_wait_table(void)
  58 {
  59         int i;
  60
  61         for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  62                 init_waitqueue_head(wait_table + i);
  63         return 0;
  64 }
  65 fs_initcall(init_dax_wait_table);
  66
  67 static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
  68                                               pgoff_t index)
  69 {
  70         unsigned long hash = hash_long((unsigned long)mapping ^ index,
  71                                        DAX_WAIT_TABLE_BITS);
  72         return wait_table + hash;
  73 }
  74
  75 static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  76 {
  77         struct request_queue *q = bdev->bd_queue;
  78         long rc = -EIO;
  79
  80         dax->addr = ERR_PTR(-EIO);
  81         if (blk_queue_enter(q, true) != 0)
  82                 return rc;
  83
  84         rc = bdev_direct_access(bdev, dax);
  85         if (rc < 0) {
  86                 dax->addr = ERR_PTR(rc);
  87                 blk_queue_exit(q);
  88                 return rc;
  89         }
  90         return rc;
  91 }
  92
  93 static void dax_unmap_atomic(struct block_device *bdev,
  94                 const struct blk_dax_ctl *dax)
  95 {
  96         if (IS_ERR(dax->addr))
  97                 return;
  98         blk_queue_exit(bdev->bd_queue);
  99 }
 100
 101 struct page *read_dax_sector(struct block_device *bdev, sector_t n)
 102 {
 103         struct page *page = alloc_pages(GFP_KERNEL, 0);
 104         struct blk_dax_ctl dax = {
 105                 .size = PAGE_SIZE,
 106                 .sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
 107         };
 108         long rc;
 109
 110         if (!page)
 111                 return ERR_PTR(-ENOMEM);
 112
 113         rc = dax_map_atomic(bdev, &dax);
 114         if (rc < 0)
 115                 return ERR_PTR(rc);
 116         memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
 117         dax_unmap_atomic(bdev, &dax);
 118         return page;
 119 }
 120
 121 static bool buffer_written(struct buffer_head *bh)
 122 {
 123         return buffer_mapped(bh) && !buffer_unwritten(bh);
 124 }
 125
 126 /*
 127  * When ext4 encounters a hole, it returns without modifying the buffer_head
 128  * which means that we can't trust b_size.  To cope with this, we set b_state
 129  * to 0 before calling get_block and, if any bit is set, we know we can trust
 130  * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 131  * and would save us time calling get_block repeatedly.
 132  */
 133 static bool buffer_size_valid(struct buffer_head *bh)
 134 {
 135         return bh->b_state != 0;
 136 }
 137
 138
 139 static sector_t to_sector(const struct buffer_head *bh,
 140                 const struct inode *inode)
 141 {
 142         sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 143
 144         return sector;
 145 }
 146
 147 static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 148                       loff_t start, loff_t end, get_block_t get_block,
 149                       struct buffer_head *bh)
 150 {
 151         loff_t pos = start, max = start, bh_max = start;
 152         bool hole = false;
 153         struct block_device *bdev = NULL;
 154         int rw = iov_iter_rw(iter), rc;
 155         long map_len = 0;
 156         struct blk_dax_ctl dax = {
 157                 .addr = ERR_PTR(-EIO),
 158         };
 159         unsigned blkbits = inode->i_blkbits;
 160         sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1)
 161                                                                 >> blkbits;
 162
 163         if (rw == READ)
 164                 end = min(end, i_size_read(inode));
 165
 166         while (pos < end) {
 167                 size_t len;
 168                 if (pos == max) {
 169                         long page = pos >> PAGE_SHIFT;
 170                         sector_t block = page << (PAGE_SHIFT - blkbits);
 171                         unsigned first = pos - (block << blkbits);
 172                         long size;
 173
 174                         if (pos == bh_max) {
 175                                 bh->b_size = PAGE_ALIGN(end - pos);
 176                                 bh->b_state = 0;
 177                                 rc = get_block(inode, block, bh, rw == WRITE);
 178                                 if (rc)
 179                                         break;
 180                                 if (!buffer_size_valid(bh))
 181                                         bh->b_size = 1 << blkbits;
 182                                 bh_max = pos - first + bh->b_size;
 183                                 bdev = bh->b_bdev;
 184                                 /*
 185                                  * We allow uninitialized buffers for writes
 186                                  * beyond EOF as those cannot race with faults
 187                                  */
 188                                 WARN_ON_ONCE(
 189                                         (buffer_new(bh) && block < file_blks) ||
 190                                         (rw == WRITE && buffer_unwritten(bh)));
 191                         } else {
 192                                 unsigned done = bh->b_size -
 193                                                 (bh_max - (pos - first));
 194                                 bh->b_blocknr += done >> blkbits;
 195                                 bh->b_size -= done;
 196                         }
 197
 198                         hole = rw == READ && !buffer_written(bh);
 199                         if (hole) {
 200                                 size = bh->b_size - first;
 201                         } else {
 202                                 dax_unmap_atomic(bdev, &dax);
 203                                 dax.sector = to_sector(bh, inode);
 204                                 dax.size = bh->b_size;
 205                                 map_len = dax_map_atomic(bdev, &dax);
 206                                 if (map_len < 0) {
 207                                         rc = map_len;
 208                                         break;
 209                                 }
 210                                 dax.addr += first;
 211                                 size = map_len - first;
 212                         }
 213                         /*
 214                          * pos + size is one past the last offset for IO,
 215                          * so pos + size can overflow loff_t at extreme offsets.
 216                          * Cast to u64 to catch this and get the true minimum.
 217                          */
 218                         max = min_t(u64, pos + size, end);
 219                 }
 220
 221                 if (iov_iter_rw(iter) == WRITE) {
 222                         len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 223                 } else if (!hole)
 224                         len = copy_to_iter((void __force *) dax.addr, max - pos,
 225                                         iter);
 226                 else
 227                         len = iov_iter_zero(max - pos, iter);
 228
 229                 if (!len) {
 230                         rc = -EFAULT;
 231                         break;
 232                 }
 233
 234                 pos += len;
 235                 if (!IS_ERR(dax.addr))
 236                         dax.addr += len;
 237         }
 238
 239         dax_unmap_atomic(bdev, &dax);
 240
 241         return (pos == start) ? rc : pos - start;
 242 }
 243
 244 /**
 245  * dax_do_io - Perform I/O to a DAX file
 246  * @iocb: The control block for this I/O
 247  * @inode: The file which the I/O is directed at
 248  * @iter: The addresses to do I/O from or to
 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, 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 pos = iocb->ki_pos;
 267         loff_t end = pos + iov_iter_count(iter);
 268
 269         memset(&bh, 0, sizeof(bh));
 270         bh.b_bdev = inode->i_sb->s_bdev;
 271
 272         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 273                 inode_lock(inode);
 274
 275         /* Protects against truncate */
 276         if (!(flags & DIO_SKIP_DIO_COUNT))
 277                 inode_dio_begin(inode);
 278
 279         retval = dax_io(inode, iter, pos, end, get_block, &bh);
 280
 281         if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 282                 inode_unlock(inode);
 283
 284         if (end_io) {
 285                 int err;
 286
 287                 err = end_io(iocb, pos, retval, bh.b_private);
 288                 if (err)
 289                         retval = err;
 290         }
 291
 292         if (!(flags & DIO_SKIP_DIO_COUNT))
 293                 inode_dio_end(inode);
 294         return retval;
 295 }
 296 EXPORT_SYMBOL_GPL(dax_do_io);
 297
 298 /*
 299  * DAX radix tree locking
 300  */
 301 struct exceptional_entry_key {
 302         struct address_space *mapping;
 303         unsigned long index;
 304 };
 305
 306 struct wait_exceptional_entry_queue {
 307         wait_queue_t wait;
 308         struct exceptional_entry_key key;
 309 };
 310
 311 static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode,
 312                                        int sync, void *keyp)
 313 {
 314         struct exceptional_entry_key *key = keyp;
 315         struct wait_exceptional_entry_queue *ewait =
 316                 container_of(wait, struct wait_exceptional_entry_queue, wait);
 317
 318         if (key->mapping != ewait->key.mapping ||
 319             key->index != ewait->key.index)
 320                 return 0;
 321         return autoremove_wake_function(wait, mode, sync, NULL);
 322 }
 323
 324 /*
 325  * Check whether the given slot is locked. The function must be called with
 326  * mapping->tree_lock held
 327  */
 328 static inline int slot_locked(struct address_space *mapping, void **slot)
 329 {
 330         unsigned long entry = (unsigned long)
 331                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 332         return entry & RADIX_DAX_ENTRY_LOCK;
 333 }
 334
 335 /*
 336  * Mark the given slot is locked. The function must be called with
 337  * mapping->tree_lock held
 338  */
 339 static inline void *lock_slot(struct address_space *mapping, void **slot)
 340 {
 341         unsigned long entry = (unsigned long)
 342                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 343
 344         entry |= RADIX_DAX_ENTRY_LOCK;
 345         radix_tree_replace_slot(slot, (void *)entry);
 346         return (void *)entry;
 347 }
 348
 349 /*
 350  * Mark the given slot is unlocked. The function must be called with
 351  * mapping->tree_lock held
 352  */
 353 static inline void *unlock_slot(struct address_space *mapping, void **slot)
 354 {
 355         unsigned long entry = (unsigned long)
 356                 radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
 357
 358         entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
 359         radix_tree_replace_slot(slot, (void *)entry);
 360         return (void *)entry;
 361 }
 362
 363 /*
 364  * Lookup entry in radix tree, wait for it to become unlocked if it is
 365  * exceptional entry and return it. The caller must call
 366  * put_unlocked_mapping_entry() when he decided not to lock the entry or
 367  * put_locked_mapping_entry() when he locked the entry and now wants to
 368  * unlock it.
 369  *
 370  * The function must be called with mapping->tree_lock held.
 371  */
 372 static void *get_unlocked_mapping_entry(struct address_space *mapping,
 373                                         pgoff_t index, void ***slotp)
 374 {
 375         void *ret, **slot;
 376         struct wait_exceptional_entry_queue ewait;
 377         wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
 378
 379         init_wait(&ewait.wait);
 380         ewait.wait.func = wake_exceptional_entry_func;
 381         ewait.key.mapping = mapping;
 382         ewait.key.index = index;
 383
 384         for (;;) {
 385                 ret = __radix_tree_lookup(&mapping->page_tree, index, NULL,
 386                                           &slot);
 387                 if (!ret || !radix_tree_exceptional_entry(ret) ||
 388                     !slot_locked(mapping, slot)) {
 389                         if (slotp)
 390                                 *slotp = slot;
 391                         return ret;
 392                 }
 393                 prepare_to_wait_exclusive(wq, &ewait.wait,
 394                                           TASK_UNINTERRUPTIBLE);
 395                 spin_unlock_irq(&mapping->tree_lock);
 396                 schedule();
 397                 finish_wait(wq, &ewait.wait);
 398                 spin_lock_irq(&mapping->tree_lock);
 399         }
 400 }
 401
 402 /*
 403  * Find radix tree entry at given index. If it points to a page, return with
 404  * the page locked. If it points to the exceptional entry, return with the
 405  * radix tree entry locked. If the radix tree doesn't contain given index,
 406  * create empty exceptional entry for the index and return with it locked.
 407  *
 408  * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 409  * persistent memory the benefit is doubtful. We can add that later if we can
 410  * show it helps.
 411  */
 412 static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index)
 413 {
 414         void *ret, **slot;
 415
 416 restart:
 417         spin_lock_irq(&mapping->tree_lock);
 418         ret = get_unlocked_mapping_entry(mapping, index, &slot);
 419         /* No entry for given index? Make sure radix tree is big enough. */
 420         if (!ret) {
 421                 int err;
 422
 423                 spin_unlock_irq(&mapping->tree_lock);
 424                 err = radix_tree_preload(
 425                                 mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
 426                 if (err)
 427                         return ERR_PTR(err);
 428                 ret = (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY |
 429                                RADIX_DAX_ENTRY_LOCK);
 430                 spin_lock_irq(&mapping->tree_lock);
 431                 err = radix_tree_insert(&mapping->page_tree, index, ret);
 432                 radix_tree_preload_end();
 433                 if (err) {
 434                         spin_unlock_irq(&mapping->tree_lock);
 435                         /* Someone already created the entry? */
 436                         if (err == -EEXIST)
 437                                 goto restart;
 438                         return ERR_PTR(err);
 439                 }
 440                 /* Good, we have inserted empty locked entry into the tree. */
 441                 mapping->nrexceptional++;
 442                 spin_unlock_irq(&mapping->tree_lock);
 443                 return ret;
 444         }
 445         /* Normal page in radix tree? */
 446         if (!radix_tree_exceptional_entry(ret)) {
 447                 struct page *page = ret;
 448
 449                 get_page(page);
 450                 spin_unlock_irq(&mapping->tree_lock);
 451                 lock_page(page);
 452                 /* Page got truncated? Retry... */
 453                 if (unlikely(page->mapping != mapping)) {
 454                         unlock_page(page);
 455                         put_page(page);
 456                         goto restart;
 457                 }
 458                 return page;
 459         }
 460         ret = lock_slot(mapping, slot);
 461         spin_unlock_irq(&mapping->tree_lock);
 462         return ret;
 463 }
 464
 465 void dax_wake_mapping_entry_waiter(struct address_space *mapping,
 466                                    pgoff_t index, bool wake_all)
 467 {
 468         wait_queue_head_t *wq = dax_entry_waitqueue(mapping, index);
 469
 470         /*
 471          * Checking for locked entry and prepare_to_wait_exclusive() happens
 472          * under mapping->tree_lock, ditto for entry handling in our callers.
 473          * So at this point all tasks that could have seen our entry locked
 474          * must be in the waitqueue and the following check will see them.
 475          */
 476         if (waitqueue_active(wq)) {
 477                 struct exceptional_entry_key key;
 478
 479                 key.mapping = mapping;
 480                 key.index = index;
 481                 __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
 482         }
 483 }
 484
 485 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index)
 486 {
 487         void *ret, **slot;
 488
 489         spin_lock_irq(&mapping->tree_lock);
 490         ret = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
 491         if (WARN_ON_ONCE(!ret || !radix_tree_exceptional_entry(ret) ||
 492                          !slot_locked(mapping, slot))) {
 493                 spin_unlock_irq(&mapping->tree_lock);
 494                 return;
 495         }
 496         unlock_slot(mapping, slot);
 497         spin_unlock_irq(&mapping->tree_lock);
 498         dax_wake_mapping_entry_waiter(mapping, index, false);
 499 }
 500
 501 static void put_locked_mapping_entry(struct address_space *mapping,
 502                                      pgoff_t index, void *entry)
 503 {
 504         if (!radix_tree_exceptional_entry(entry)) {
 505                 unlock_page(entry);
 506                 put_page(entry);
 507         } else {
 508                 dax_unlock_mapping_entry(mapping, index);
 509         }
 510 }
 511
 512 /*
 513  * Called when we are done with radix tree entry we looked up via
 514  * get_unlocked_mapping_entry() and which we didn't lock in the end.
 515  */
 516 static void put_unlocked_mapping_entry(struct address_space *mapping,
 517                                        pgoff_t index, void *entry)
 518 {
 519         if (!radix_tree_exceptional_entry(entry))
 520                 return;
 521
 522         /* We have to wake up next waiter for the radix tree entry lock */
 523         dax_wake_mapping_entry_waiter(mapping, index, false);
 524 }
 525
 526 /*
 527  * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
 528  * entry to get unlocked before deleting it.
 529  */
 530 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 531 {
 532         void *entry;
 533
 534         spin_lock_irq(&mapping->tree_lock);
 535         entry = get_unlocked_mapping_entry(mapping, index, NULL);
 536         /*
 537          * This gets called from truncate / punch_hole path. As such, the caller
 538          * must hold locks protecting against concurrent modifications of the
 539          * radix tree (usually fs-private i_mmap_sem for writing). Since the
 540          * caller has seen exceptional entry for this index, we better find it
 541          * at that index as well...
 542          */
 543         if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) {
 544                 spin_unlock_irq(&mapping->tree_lock);
 545                 return 0;
 546         }
 547         radix_tree_delete(&mapping->page_tree, index);
 548         mapping->nrexceptional--;
 549         spin_unlock_irq(&mapping->tree_lock);
 550         dax_wake_mapping_entry_waiter(mapping, index, true);
 551
 552         return 1;
 553 }
 554
 555 /*
 556  * The user has performed a load from a hole in the file.  Allocating
 557  * a new page in the file would cause excessive storage usage for
 558  * workloads with sparse files.  We allocate a page cache page instead.
 559  * We'll kick it out of the page cache if it's ever written to,
 560  * otherwise it will simply fall out of the page cache under memory
 561  * pressure without ever having been dirtied.
 562  */
 563 static int dax_load_hole(struct address_space *mapping, void *entry,
 564                          struct vm_fault *vmf)
 565 {
 566         struct page *page;
 567
 568         /* Hole page already exists? Return it...  */
 569         if (!radix_tree_exceptional_entry(entry)) {
 570                 vmf->page = entry;
 571                 return VM_FAULT_LOCKED;
 572         }
 573
 574         /* This will replace locked radix tree entry with a hole page */
 575         page = find_or_create_page(mapping, vmf->pgoff,
 576                                    vmf->gfp_mask | __GFP_ZERO);
 577         if (!page) {
 578                 put_locked_mapping_entry(mapping, vmf->pgoff, entry);
 579                 return VM_FAULT_OOM;
 580         }
 581         vmf->page = page;
 582         return VM_FAULT_LOCKED;
 583 }
 584
 585 static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size,
 586                 struct page *to, unsigned long vaddr)
 587 {
 588         struct blk_dax_ctl dax = {
 589                 .sector = sector,
 590                 .size = size,
 591         };
 592         void *vto;
 593
 594         if (dax_map_atomic(bdev, &dax) < 0)
 595                 return PTR_ERR(dax.addr);
 596         vto = kmap_atomic(to);
 597         copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 598         kunmap_atomic(vto);
 599         dax_unmap_atomic(bdev, &dax);
 600         return 0;
 601 }
 602
 603 #define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 604
 605 static void *dax_insert_mapping_entry(struct address_space *mapping,
 606                                       struct vm_fault *vmf,
 607                                       void *entry, sector_t sector)
 608 {
 609         struct radix_tree_root *page_tree = &mapping->page_tree;
 610         int error = 0;
 611         bool hole_fill = false;
 612         void *new_entry;
 613         pgoff_t index = vmf->pgoff;
 614
 615         if (vmf->flags & FAULT_FLAG_WRITE)
 616                 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 617
 618         /* Replacing hole page with block mapping? */
 619         if (!radix_tree_exceptional_entry(entry)) {
 620                 hole_fill = true;
 621                 /*
 622                  * Unmap the page now before we remove it from page cache below.
 623                  * The page is locked so it cannot be faulted in again.
 624                  */
 625                 unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 626                                     PAGE_SIZE, 0);
 627                 error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM);
 628                 if (error)
 629                         return ERR_PTR(error);
 630         }
 631
 632         spin_lock_irq(&mapping->tree_lock);
 633         new_entry = (void *)((unsigned long)RADIX_DAX_ENTRY(sector, false) |
 634                        RADIX_DAX_ENTRY_LOCK);
 635         if (hole_fill) {
 636                 __delete_from_page_cache(entry, NULL);
 637                 /* Drop pagecache reference */
 638                 put_page(entry);
 639                 error = radix_tree_insert(page_tree, index, new_entry);
 640                 if (error) {
 641                         new_entry = ERR_PTR(error);
 642                         goto unlock;
 643                 }
 644                 mapping->nrexceptional++;
 645         } else {
 646                 void **slot;
 647                 void *ret;
 648
 649                 ret = __radix_tree_lookup(page_tree, index, NULL, &slot);
 650                 WARN_ON_ONCE(ret != entry);
 651                 radix_tree_replace_slot(slot, new_entry);
 652         }
 653         if (vmf->flags & FAULT_FLAG_WRITE)
 654                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 655  unlock:
 656         spin_unlock_irq(&mapping->tree_lock);
 657         if (hole_fill) {
 658                 radix_tree_preload_end();
 659                 /*
 660                  * We don't need hole page anymore, it has been replaced with
 661                  * locked radix tree entry now.
 662                  */
 663                 if (mapping->a_ops->freepage)
 664                         mapping->a_ops->freepage(entry);
 665                 unlock_page(entry);
 666                 put_page(entry);
 667         }
 668         return new_entry;
 669 }
 670
 671 static int dax_writeback_one(struct block_device *bdev,
 672                 struct address_space *mapping, pgoff_t index, void *entry)
 673 {
 674         struct radix_tree_root *page_tree = &mapping->page_tree;
 675         int type = RADIX_DAX_TYPE(entry);
 676         struct radix_tree_node *node;
 677         struct blk_dax_ctl dax;
 678         void **slot;
 679         int ret = 0;
 680
 681         spin_lock_irq(&mapping->tree_lock);
 682         /*
 683          * Regular page slots are stabilized by the page lock even
 684          * without the tree itself locked.  These unlocked entries
 685          * need verification under the tree lock.
 686          */
 687         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 688                 goto unlock;
 689         if (*slot != entry)
 690                 goto unlock;
 691
 692         /* another fsync thread may have already written back this entry */
 693         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 694                 goto unlock;
 695
 696         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 697                 ret = -EIO;
 698                 goto unlock;
 699         }
 700
 701         dax.sector = RADIX_DAX_SECTOR(entry);
 702         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 703         spin_unlock_irq(&mapping->tree_lock);
 704
 705         /*
 706          * We cannot hold tree_lock while calling dax_map_atomic() because it
 707          * eventually calls cond_resched().
 708          */
 709         ret = dax_map_atomic(bdev, &dax);
 710         if (ret < 0)
 711                 return ret;
 712
 713         if (WARN_ON_ONCE(ret < dax.size)) {
 714                 ret = -EIO;
 715                 goto unmap;
 716         }
 717
 718         wb_cache_pmem(dax.addr, dax.size);
 719
 720         spin_lock_irq(&mapping->tree_lock);
 721         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 722         spin_unlock_irq(&mapping->tree_lock);
 723  unmap:
 724         dax_unmap_atomic(bdev, &dax);
 725         return ret;
 726
 727  unlock:
 728         spin_unlock_irq(&mapping->tree_lock);
 729         return ret;
 730 }
 731
 732 /*
 733  * Flush the mapping to the persistent domain within the byte range of [start,
 734  * end]. This is required by data integrity operations to ensure file data is
 735  * on persistent storage prior to completion of the operation.
 736  */
 737 int dax_writeback_mapping_range(struct address_space *mapping,
 738                 struct block_device *bdev, struct writeback_control *wbc)
 739 {
 740         struct inode *inode = mapping->host;
 741         pgoff_t start_index, end_index, pmd_index;
 742         pgoff_t indices[PAGEVEC_SIZE];
 743         struct pagevec pvec;
 744         bool done = false;
 745         int i, ret = 0;
 746         void *entry;
 747
 748         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 749                 return -EIO;
 750
 751         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 752                 return 0;
 753
 754         start_index = wbc->range_start >> PAGE_SHIFT;
 755         end_index = wbc->range_end >> PAGE_SHIFT;
 756         pmd_index = DAX_PMD_INDEX(start_index);
 757
 758         rcu_read_lock();
 759         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 760         rcu_read_unlock();
 761
 762         /* see if the start of our range is covered by a PMD entry */
 763         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 764                 start_index = pmd_index;
 765
 766         tag_pages_for_writeback(mapping, start_index, end_index);
 767
 768         pagevec_init(&pvec, 0);
 769         while (!done) {
 770                 pvec.nr = find_get_entries_tag(mapping, start_index,
 771                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 772                                 pvec.pages, indices);
 773
 774                 if (pvec.nr == 0)
 775                         break;
 776
 777                 for (i = 0; i < pvec.nr; i++) {
 778                         if (indices[i] > end_index) {
 779                                 done = true;
 780                                 break;
 781                         }
 782
 783                         ret = dax_writeback_one(bdev, mapping, indices[i],
 784                                         pvec.pages[i]);
 785                         if (ret < 0)
 786                                 return ret;
 787                 }
 788         }
 789         return 0;
 790 }
 791 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 792
 793 static int dax_insert_mapping(struct address_space *mapping,
 794                 struct block_device *bdev, sector_t sector, size_t size,
 795                 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
 796 {
 797         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 798         struct blk_dax_ctl dax = {
 799                 .sector = sector,
 800                 .size = size,
 801         };
 802         void *ret;
 803         void *entry = *entryp;
 804
 805         if (dax_map_atomic(bdev, &dax) < 0)
 806                 return PTR_ERR(dax.addr);
 807         dax_unmap_atomic(bdev, &dax);
 808
 809         ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector);
 810         if (IS_ERR(ret))
 811                 return PTR_ERR(ret);
 812         *entryp = ret;
 813
 814         return vm_insert_mixed(vma, vaddr, dax.pfn);
 815 }
 816
 817 /**
 818  * dax_fault - handle a page fault on a DAX file
 819  * @vma: The virtual memory area where the fault occurred
 820  * @vmf: The description of the fault
 821  * @get_block: The filesystem method used to translate file offsets to blocks
 822  *
 823  * When a page fault occurs, filesystems may call this helper in their
 824  * fault handler for DAX files. dax_fault() assumes the caller has done all
 825  * the necessary locking for the page fault to proceed successfully.
 826  */
 827 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 828                         get_block_t get_block)
 829 {
 830         struct file *file = vma->vm_file;
 831         struct address_space *mapping = file->f_mapping;
 832         struct inode *inode = mapping->host;
 833         void *entry;
 834         struct buffer_head bh;
 835         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 836         unsigned blkbits = inode->i_blkbits;
 837         sector_t block;
 838         pgoff_t size;
 839         int error;
 840         int major = 0;
 841
 842         /*
 843          * Check whether offset isn't beyond end of file now. Caller is supposed
 844          * to hold locks serializing us with truncate / punch hole so this is
 845          * a reliable test.
 846          */
 847         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 848         if (vmf->pgoff >= size)
 849                 return VM_FAULT_SIGBUS;
 850
 851         memset(&bh, 0, sizeof(bh));
 852         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 853         bh.b_bdev = inode->i_sb->s_bdev;
 854         bh.b_size = PAGE_SIZE;
 855
 856         entry = grab_mapping_entry(mapping, vmf->pgoff);
 857         if (IS_ERR(entry)) {
 858                 error = PTR_ERR(entry);
 859                 goto out;
 860         }
 861
 862         error = get_block(inode, block, &bh, 0);
 863         if (!error && (bh.b_size < PAGE_SIZE))
 864                 error = -EIO;           /* fs corruption? */
 865         if (error)
 866                 goto unlock_entry;
 867
 868         if (vmf->cow_page) {
 869                 struct page *new_page = vmf->cow_page;
 870                 if (buffer_written(&bh))
 871                         error = copy_user_dax(bh.b_bdev, to_sector(&bh, inode),
 872                                         bh.b_size, new_page, vaddr);
 873                 else
 874                         clear_user_highpage(new_page, vaddr);
 875                 if (error)
 876                         goto unlock_entry;
 877                 if (!radix_tree_exceptional_entry(entry)) {
 878                         vmf->page = entry;
 879                         return VM_FAULT_LOCKED;
 880                 }
 881                 vmf->entry = entry;
 882                 return VM_FAULT_DAX_LOCKED;
 883         }
 884
 885         if (!buffer_mapped(&bh)) {
 886                 if (vmf->flags & FAULT_FLAG_WRITE) {
 887                         error = get_block(inode, block, &bh, 1);
 888                         count_vm_event(PGMAJFAULT);
 889                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 890                         major = VM_FAULT_MAJOR;
 891                         if (!error && (bh.b_size < PAGE_SIZE))
 892                                 error = -EIO;
 893                         if (error)
 894                                 goto unlock_entry;
 895                 } else {
 896                         return dax_load_hole(mapping, entry, vmf);
 897                 }
 898         }
 899
 900         /* Filesystem should not return unwritten buffers to us! */
 901         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 902         error = dax_insert_mapping(mapping, bh.b_bdev, to_sector(&bh, inode),
 903                         bh.b_size, &entry, vma, vmf);
 904  unlock_entry:
 905         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
 906  out:
 907         if (error == -ENOMEM)
 908                 return VM_FAULT_OOM | major;
 909         /* -EBUSY is fine, somebody else faulted on the same PTE */
 910         if ((error < 0) && (error != -EBUSY))
 911                 return VM_FAULT_SIGBUS | major;
 912         return VM_FAULT_NOPAGE | major;
 913 }
 914 EXPORT_SYMBOL_GPL(dax_fault);
 915
 916 #if defined(CONFIG_TRANSPARENT_HUGEPAGE)
 917 /*
 918  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 919  * more often than one might expect in the below function.
 920  */
 921 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 922
 923 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 924                 const char *reason, const char *fn)
 925 {
 926         if (bh) {
 927                 char bname[BDEVNAME_SIZE];
 928                 bdevname(bh->b_bdev, bname);
 929                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 930                         "length %zd fallback: %s\n", fn, current->comm,
 931                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 932                         bh->b_size, reason);
 933         } else {
 934                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 935                         current->comm, address, reason);
 936         }
 937 }
 938
 939 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 940
 941 /**
 942  * dax_pmd_fault - handle a PMD fault on a DAX file
 943  * @vma: The virtual memory area where the fault occurred
 944  * @vmf: The description of the fault
 945  * @get_block: The filesystem method used to translate file offsets to blocks
 946  *
 947  * When a page fault occurs, filesystems may call this helper in their
 948  * pmd_fault handler for DAX files.
 949  */
 950 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 951                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 952 {
 953         struct file *file = vma->vm_file;
 954         struct address_space *mapping = file->f_mapping;
 955         struct inode *inode = mapping->host;
 956         struct buffer_head bh;
 957         unsigned blkbits = inode->i_blkbits;
 958         unsigned long pmd_addr = address & PMD_MASK;
 959         bool write = flags & FAULT_FLAG_WRITE;
 960         struct block_device *bdev;
 961         pgoff_t size, pgoff;
 962         sector_t block;
 963         int result = 0;
 964         bool alloc = false;
 965
 966         /* dax pmd mappings require pfn_t_devmap() */
 967         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 968                 return VM_FAULT_FALLBACK;
 969
 970         /* Fall back to PTEs if we're going to COW */
 971         if (write && !(vma->vm_flags & VM_SHARED)) {
 972                 split_huge_pmd(vma, pmd, address);
 973                 dax_pmd_dbg(NULL, address, "cow write");
 974                 return VM_FAULT_FALLBACK;
 975         }
 976         /* If the PMD would extend outside the VMA */
 977         if (pmd_addr < vma->vm_start) {
 978                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 979                 return VM_FAULT_FALLBACK;
 980         }
 981         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 982                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 983                 return VM_FAULT_FALLBACK;
 984         }
 985
 986         pgoff = linear_page_index(vma, pmd_addr);
 987         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 988         if (pgoff >= size)
 989                 return VM_FAULT_SIGBUS;
 990         /* If the PMD would cover blocks out of the file */
 991         if ((pgoff | PG_PMD_COLOUR) >= size) {
 992                 dax_pmd_dbg(NULL, address,
 993                                 "offset + huge page size > file size");
 994                 return VM_FAULT_FALLBACK;
 995         }
 996
 997         memset(&bh, 0, sizeof(bh));
 998         bh.b_bdev = inode->i_sb->s_bdev;
 999         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
1000
1001         bh.b_size = PMD_SIZE;
1002
1003         if (get_block(inode, block, &bh, 0) != 0)
1004                 return VM_FAULT_SIGBUS;
1005
1006         if (!buffer_mapped(&bh) && write) {
1007                 if (get_block(inode, block, &bh, 1) != 0)
1008                         return VM_FAULT_SIGBUS;
1009                 alloc = true;
1010                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
1011         }
1012
1013         bdev = bh.b_bdev;
1014
1015         /*
1016          * If the filesystem isn't willing to tell us the length of a hole,
1017          * just fall back to PTEs.  Calling get_block 512 times in a loop
1018          * would be silly.
1019          */
1020         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
1021                 dax_pmd_dbg(&bh, address, "allocated block too small");
1022                 return VM_FAULT_FALLBACK;
1023         }
1024
1025         /*
1026          * If we allocated new storage, make sure no process has any
1027          * zero pages covering this hole
1028          */
1029         if (alloc) {
1030                 loff_t lstart = pgoff << PAGE_SHIFT;
1031                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
1032
1033                 truncate_pagecache_range(inode, lstart, lend);
1034         }
1035
1036         if (!write && !buffer_mapped(&bh)) {
1037                 spinlock_t *ptl;
1038                 pmd_t entry;
1039                 struct page *zero_page = mm_get_huge_zero_page(vma->vm_mm);
1040
1041                 if (unlikely(!zero_page)) {
1042                         dax_pmd_dbg(&bh, address, "no zero page");
1043                         goto fallback;
1044                 }
1045
1046                 ptl = pmd_lock(vma->vm_mm, pmd);
1047                 if (!pmd_none(*pmd)) {
1048                         spin_unlock(ptl);
1049                         dax_pmd_dbg(&bh, address, "pmd already present");
1050                         goto fallback;
1051                 }
1052
1053                 dev_dbg(part_to_dev(bdev->bd_part),
1054                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
1055                                 __func__, current->comm, address,
1056                                 (unsigned long long) to_sector(&bh, inode));
1057
1058                 entry = mk_pmd(zero_page, vma->vm_page_prot);
1059                 entry = pmd_mkhuge(entry);
1060                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
1061                 result = VM_FAULT_NOPAGE;
1062                 spin_unlock(ptl);
1063         } else {
1064                 struct blk_dax_ctl dax = {
1065                         .sector = to_sector(&bh, inode),
1066                         .size = PMD_SIZE,
1067                 };
1068                 long length = dax_map_atomic(bdev, &dax);
1069
1070                 if (length < 0) {
1071                         dax_pmd_dbg(&bh, address, "dax-error fallback");
1072                         goto fallback;
1073                 }
1074                 if (length < PMD_SIZE) {
1075                         dax_pmd_dbg(&bh, address, "dax-length too small");
1076                         dax_unmap_atomic(bdev, &dax);
1077                         goto fallback;
1078                 }
1079                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
1080                         dax_pmd_dbg(&bh, address, "pfn unaligned");
1081                         dax_unmap_atomic(bdev, &dax);
1082                         goto fallback;
1083                 }
1084
1085                 if (!pfn_t_devmap(dax.pfn)) {
1086                         dax_unmap_atomic(bdev, &dax);
1087                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
1088                         goto fallback;
1089                 }
1090                 dax_unmap_atomic(bdev, &dax);
1091
1092                 /*
1093                  * For PTE faults we insert a radix tree entry for reads, and
1094                  * leave it clean.  Then on the first write we dirty the radix
1095                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
1096                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
1097                  * call into get_block() to translate the pgoff to a sector in
1098                  * order to be able to create a new radix tree entry.
1099                  *
1100                  * The PMD path doesn't have an equivalent to
1101                  * dax_pfn_mkwrite(), though, so for a read followed by a
1102                  * write we traverse all the way through dax_pmd_fault()
1103                  * twice.  This means we can just skip inserting a radix tree
1104                  * entry completely on the initial read and just wait until
1105                  * the write to insert a dirty entry.
1106                  */
1107                 if (write) {
1108                         /*
1109                          * We should insert radix-tree entry and dirty it here.
1110                          * For now this is broken...
1111                          */
1112                 }
1113
1114                 dev_dbg(part_to_dev(bdev->bd_part),
1115                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1116                                 __func__, current->comm, address,
1117                                 pfn_t_to_pfn(dax.pfn),
1118                                 (unsigned long long) dax.sector);
1119                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1120                                 dax.pfn, write);
1121         }
1122
1123  out:
1124         return result;
1125
1126  fallback:
1127         count_vm_event(THP_FAULT_FALLBACK);
1128         result = VM_FAULT_FALLBACK;
1129         goto out;
1130 }
1131 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1132 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1133
1134 /**
1135  * dax_pfn_mkwrite - handle first write to DAX page
1136  * @vma: The virtual memory area where the fault occurred
1137  * @vmf: The description of the fault
1138  */
1139 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1140 {
1141         struct file *file = vma->vm_file;
1142         struct address_space *mapping = file->f_mapping;
1143         void *entry;
1144         pgoff_t index = vmf->pgoff;
1145
1146         spin_lock_irq(&mapping->tree_lock);
1147         entry = get_unlocked_mapping_entry(mapping, index, NULL);
1148         if (!entry || !radix_tree_exceptional_entry(entry))
1149                 goto out;
1150         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1151         put_unlocked_mapping_entry(mapping, index, entry);
1152 out:
1153         spin_unlock_irq(&mapping->tree_lock);
1154         return VM_FAULT_NOPAGE;
1155 }
1156 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1157
1158 static bool dax_range_is_aligned(struct block_device *bdev,
1159                                  unsigned int offset, unsigned int length)
1160 {
1161         unsigned short sector_size = bdev_logical_block_size(bdev);
1162
1163         if (!IS_ALIGNED(offset, sector_size))
1164                 return false;
1165         if (!IS_ALIGNED(length, sector_size))
1166                 return false;
1167
1168         return true;
1169 }
1170
1171 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1172                 unsigned int offset, unsigned int length)
1173 {
1174         struct blk_dax_ctl dax = {
1175                 .sector         = sector,
1176                 .size           = PAGE_SIZE,
1177         };
1178
1179         if (dax_range_is_aligned(bdev, offset, length)) {
1180                 sector_t start_sector = dax.sector + (offset >> 9);
1181
1182                 return blkdev_issue_zeroout(bdev, start_sector,
1183                                 length >> 9, GFP_NOFS, true);
1184         } else {
1185                 if (dax_map_atomic(bdev, &dax) < 0)
1186                         return PTR_ERR(dax.addr);
1187                 clear_pmem(dax.addr + offset, length);
1188                 dax_unmap_atomic(bdev, &dax);
1189         }
1190         return 0;
1191 }
1192 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1193
1194 /**
1195  * dax_zero_page_range - zero a range within a page of a DAX file
1196  * @inode: The file being truncated
1197  * @from: The file offset that is being truncated to
1198  * @length: The number of bytes to zero
1199  * @get_block: The filesystem method used to translate file offsets to blocks
1200  *
1201  * This function can be called by a filesystem when it is zeroing part of a
1202  * page in a DAX file.  This is intended for hole-punch operations.  If
1203  * you are truncating a file, the helper function dax_truncate_page() may be
1204  * more convenient.
1205  */
1206 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1207                                                         get_block_t get_block)
1208 {
1209         struct buffer_head bh;
1210         pgoff_t index = from >> PAGE_SHIFT;
1211         unsigned offset = from & (PAGE_SIZE-1);
1212         int err;
1213
1214         /* Block boundary? Nothing to do */
1215         if (!length)
1216                 return 0;
1217         BUG_ON((offset + length) > PAGE_SIZE);
1218
1219         memset(&bh, 0, sizeof(bh));
1220         bh.b_bdev = inode->i_sb->s_bdev;
1221         bh.b_size = PAGE_SIZE;
1222         err = get_block(inode, index, &bh, 0);
1223         if (err < 0 || !buffer_written(&bh))
1224                 return err;
1225
1226         return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode),
1227                         offset, length);
1228 }
1229 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1230
1231 /**
1232  * dax_truncate_page - handle a partial page being truncated in a DAX file
1233  * @inode: The file being truncated
1234  * @from: The file offset that is being truncated to
1235  * @get_block: The filesystem method used to translate file offsets to blocks
1236  *
1237  * Similar to block_truncate_page(), this function can be called by a
1238  * filesystem when it is truncating a DAX file to handle the partial page.
1239  */
1240 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1241 {
1242         unsigned length = PAGE_ALIGN(from) - from;
1243         return dax_zero_page_range(inode, from, length, get_block);
1244 }
1245 EXPORT_SYMBOL_GPL(dax_truncate_page);
1246
1247 #ifdef CONFIG_FS_IOMAP
1248 static loff_t
1249 iomap_dax_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1250                 struct iomap *iomap)
1251 {
1252         struct iov_iter *iter = data;
1253         loff_t end = pos + length, done = 0;
1254         ssize_t ret = 0;
1255
1256         if (iov_iter_rw(iter) == READ) {
1257                 end = min(end, i_size_read(inode));
1258                 if (pos >= end)
1259                         return 0;
1260
1261                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1262                         return iov_iter_zero(min(length, end - pos), iter);
1263         }
1264
1265         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1266                 return -EIO;
1267
1268         while (pos < end) {
1269                 unsigned offset = pos & (PAGE_SIZE - 1);
1270                 struct blk_dax_ctl dax = { 0 };
1271                 ssize_t map_len;
1272
1273                 dax.sector = iomap->blkno +
1274                         (((pos & PAGE_MASK) - iomap->offset) >> 9);
1275                 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
1276                 map_len = dax_map_atomic(iomap->bdev, &dax);
1277                 if (map_len < 0) {
1278                         ret = map_len;
1279                         break;
1280                 }
1281
1282                 dax.addr += offset;
1283                 map_len -= offset;
1284                 if (map_len > end - pos)
1285                         map_len = end - pos;
1286
1287                 if (iov_iter_rw(iter) == WRITE)
1288                         map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1289                 else
1290                         map_len = copy_to_iter(dax.addr, map_len, iter);
1291                 dax_unmap_atomic(iomap->bdev, &dax);
1292                 if (map_len <= 0) {
1293                         ret = map_len ? map_len : -EFAULT;
1294                         break;
1295                 }
1296
1297                 pos += map_len;
1298                 length -= map_len;
1299                 done += map_len;
1300         }
1301
1302         return done ? done : ret;
1303 }
1304
1305 /**
1306  * iomap_dax_rw - Perform I/O to a DAX file
1307  * @iocb:       The control block for this I/O
1308  * @iter:       The addresses to do I/O from or to
1309  * @ops:        iomap ops passed from the file system
1310  *
1311  * This function performs read and write operations to directly mapped
1312  * persistent memory.  The callers needs to take care of read/write exclusion
1313  * and evicting any page cache pages in the region under I/O.
1314  */
1315 ssize_t
1316 iomap_dax_rw(struct kiocb *iocb, struct iov_iter *iter,
1317                 struct iomap_ops *ops)
1318 {
1319         struct address_space *mapping = iocb->ki_filp->f_mapping;
1320         struct inode *inode = mapping->host;
1321         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1322         unsigned flags = 0;
1323
1324         if (iov_iter_rw(iter) == WRITE)
1325                 flags |= IOMAP_WRITE;
1326
1327         /*
1328          * Yes, even DAX files can have page cache attached to them:  A zeroed
1329          * page is inserted into the pagecache when we have to serve a write
1330          * fault on a hole.  It should never be dirtied and can simply be
1331          * dropped from the pagecache once we get real data for the page.
1332          *
1333          * XXX: This is racy against mmap, and there's nothing we can do about
1334          * it. We'll eventually need to shift this down even further so that
1335          * we can check if we allocated blocks over a hole first.
1336          */
1337         if (mapping->nrpages) {
1338                 ret = invalidate_inode_pages2_range(mapping,
1339                                 pos >> PAGE_SHIFT,
1340                                 (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT);
1341                 WARN_ON_ONCE(ret);
1342         }
1343
1344         while (iov_iter_count(iter)) {
1345                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1346                                 iter, iomap_dax_actor);
1347                 if (ret <= 0)
1348                         break;
1349                 pos += ret;
1350                 done += ret;
1351         }
1352
1353         iocb->ki_pos += done;
1354         return done ? done : ret;
1355 }
1356 EXPORT_SYMBOL_GPL(iomap_dax_rw);
1357
1358 /**
1359  * iomap_dax_fault - handle a page fault on a DAX file
1360  * @vma: The virtual memory area where the fault occurred
1361  * @vmf: The description of the fault
1362  * @ops: iomap ops passed from the file system
1363  *
1364  * When a page fault occurs, filesystems may call this helper in their fault
1365  * or mkwrite handler for DAX files. Assumes the caller has done all the
1366  * necessary locking for the page fault to proceed successfully.
1367  */
1368 int iomap_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1369                         struct iomap_ops *ops)
1370 {
1371         struct address_space *mapping = vma->vm_file->f_mapping;
1372         struct inode *inode = mapping->host;
1373         unsigned long vaddr = (unsigned long)vmf->virtual_address;
1374         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1375         sector_t sector;
1376         struct iomap iomap = { 0 };
1377         unsigned flags = 0;
1378         int error, major = 0;
1379         void *entry;
1380
1381         /*
1382          * Check whether offset isn't beyond end of file now. Caller is supposed
1383          * to hold locks serializing us with truncate / punch hole so this is
1384          * a reliable test.
1385          */
1386         if (pos >= i_size_read(inode))
1387                 return VM_FAULT_SIGBUS;
1388
1389         entry = grab_mapping_entry(mapping, vmf->pgoff);
1390         if (IS_ERR(entry)) {
1391                 error = PTR_ERR(entry);
1392                 goto out;
1393         }
1394
1395         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1396                 flags |= IOMAP_WRITE;
1397
1398         /*
1399          * Note that we don't bother to use iomap_apply here: DAX required
1400          * the file system block size to be equal the page size, which means
1401          * that we never have to deal with more than a single extent here.
1402          */
1403         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1404         if (error)
1405                 goto unlock_entry;
1406         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1407                 error = -EIO;           /* fs corruption? */
1408                 goto unlock_entry;
1409         }
1410
1411         sector = iomap.blkno + (((pos & PAGE_MASK) - iomap.offset) >> 9);
1412
1413         if (vmf->cow_page) {
1414                 switch (iomap.type) {
1415                 case IOMAP_HOLE:
1416                 case IOMAP_UNWRITTEN:
1417                         clear_user_highpage(vmf->cow_page, vaddr);
1418                         break;
1419                 case IOMAP_MAPPED:
1420                         error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1421                                         vmf->cow_page, vaddr);
1422                         break;
1423                 default:
1424                         WARN_ON_ONCE(1);
1425                         error = -EIO;
1426                         break;
1427                 }
1428
1429                 if (error)
1430                         goto unlock_entry;
1431                 if (!radix_tree_exceptional_entry(entry)) {
1432                         vmf->page = entry;
1433                         return VM_FAULT_LOCKED;
1434                 }
1435                 vmf->entry = entry;
1436                 return VM_FAULT_DAX_LOCKED;
1437         }
1438
1439         switch (iomap.type) {
1440         case IOMAP_MAPPED:
1441                 if (iomap.flags & IOMAP_F_NEW) {
1442                         count_vm_event(PGMAJFAULT);
1443                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1444                         major = VM_FAULT_MAJOR;
1445                 }
1446                 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1447                                 PAGE_SIZE, &entry, vma, vmf);
1448                 break;
1449         case IOMAP_UNWRITTEN:
1450         case IOMAP_HOLE:
1451                 if (!(vmf->flags & FAULT_FLAG_WRITE))
1452                         return dax_load_hole(mapping, entry, vmf);
1453                 /*FALLTHRU*/
1454         default:
1455                 WARN_ON_ONCE(1);
1456                 error = -EIO;
1457                 break;
1458         }
1459
1460  unlock_entry:
1461         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1462  out:
1463         if (error == -ENOMEM)
1464                 return VM_FAULT_OOM | major;
1465         /* -EBUSY is fine, somebody else faulted on the same PTE */
1466         if (error < 0 && error != -EBUSY)
1467                 return VM_FAULT_SIGBUS | major;
1468         return VM_FAULT_NOPAGE | major;
1469 }
1470 EXPORT_SYMBOL_GPL(iomap_dax_fault);
1471 #endif /* CONFIG_FS_IOMAP */