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(&mapping->page_tree, 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(&mapping->page_tree, 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                 struct radix_tree_node *node;
 647                 void **slot;
 648                 void *ret;
 649
 650                 ret = __radix_tree_lookup(page_tree, index, &node, &slot);
 651                 WARN_ON_ONCE(ret != entry);
 652                 __radix_tree_replace(page_tree, node, slot,
 653                                      new_entry, NULL, NULL);
 654         }
 655         if (vmf->flags & FAULT_FLAG_WRITE)
 656                 radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 657  unlock:
 658         spin_unlock_irq(&mapping->tree_lock);
 659         if (hole_fill) {
 660                 radix_tree_preload_end();
 661                 /*
 662                  * We don't need hole page anymore, it has been replaced with
 663                  * locked radix tree entry now.
 664                  */
 665                 if (mapping->a_ops->freepage)
 666                         mapping->a_ops->freepage(entry);
 667                 unlock_page(entry);
 668                 put_page(entry);
 669         }
 670         return new_entry;
 671 }
 672
 673 static int dax_writeback_one(struct block_device *bdev,
 674                 struct address_space *mapping, pgoff_t index, void *entry)
 675 {
 676         struct radix_tree_root *page_tree = &mapping->page_tree;
 677         int type = RADIX_DAX_TYPE(entry);
 678         struct radix_tree_node *node;
 679         struct blk_dax_ctl dax;
 680         void **slot;
 681         int ret = 0;
 682
 683         spin_lock_irq(&mapping->tree_lock);
 684         /*
 685          * Regular page slots are stabilized by the page lock even
 686          * without the tree itself locked.  These unlocked entries
 687          * need verification under the tree lock.
 688          */
 689         if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 690                 goto unlock;
 691         if (*slot != entry)
 692                 goto unlock;
 693
 694         /* another fsync thread may have already written back this entry */
 695         if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 696                 goto unlock;
 697
 698         if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 699                 ret = -EIO;
 700                 goto unlock;
 701         }
 702
 703         dax.sector = RADIX_DAX_SECTOR(entry);
 704         dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 705         spin_unlock_irq(&mapping->tree_lock);
 706
 707         /*
 708          * We cannot hold tree_lock while calling dax_map_atomic() because it
 709          * eventually calls cond_resched().
 710          */
 711         ret = dax_map_atomic(bdev, &dax);
 712         if (ret < 0)
 713                 return ret;
 714
 715         if (WARN_ON_ONCE(ret < dax.size)) {
 716                 ret = -EIO;
 717                 goto unmap;
 718         }
 719
 720         wb_cache_pmem(dax.addr, dax.size);
 721
 722         spin_lock_irq(&mapping->tree_lock);
 723         radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 724         spin_unlock_irq(&mapping->tree_lock);
 725  unmap:
 726         dax_unmap_atomic(bdev, &dax);
 727         return ret;
 728
 729  unlock:
 730         spin_unlock_irq(&mapping->tree_lock);
 731         return ret;
 732 }
 733
 734 /*
 735  * Flush the mapping to the persistent domain within the byte range of [start,
 736  * end]. This is required by data integrity operations to ensure file data is
 737  * on persistent storage prior to completion of the operation.
 738  */
 739 int dax_writeback_mapping_range(struct address_space *mapping,
 740                 struct block_device *bdev, struct writeback_control *wbc)
 741 {
 742         struct inode *inode = mapping->host;
 743         pgoff_t start_index, end_index, pmd_index;
 744         pgoff_t indices[PAGEVEC_SIZE];
 745         struct pagevec pvec;
 746         bool done = false;
 747         int i, ret = 0;
 748         void *entry;
 749
 750         if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 751                 return -EIO;
 752
 753         if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 754                 return 0;
 755
 756         start_index = wbc->range_start >> PAGE_SHIFT;
 757         end_index = wbc->range_end >> PAGE_SHIFT;
 758         pmd_index = DAX_PMD_INDEX(start_index);
 759
 760         rcu_read_lock();
 761         entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 762         rcu_read_unlock();
 763
 764         /* see if the start of our range is covered by a PMD entry */
 765         if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 766                 start_index = pmd_index;
 767
 768         tag_pages_for_writeback(mapping, start_index, end_index);
 769
 770         pagevec_init(&pvec, 0);
 771         while (!done) {
 772                 pvec.nr = find_get_entries_tag(mapping, start_index,
 773                                 PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 774                                 pvec.pages, indices);
 775
 776                 if (pvec.nr == 0)
 777                         break;
 778
 779                 for (i = 0; i < pvec.nr; i++) {
 780                         if (indices[i] > end_index) {
 781                                 done = true;
 782                                 break;
 783                         }
 784
 785                         ret = dax_writeback_one(bdev, mapping, indices[i],
 786                                         pvec.pages[i]);
 787                         if (ret < 0)
 788                                 return ret;
 789                 }
 790         }
 791         return 0;
 792 }
 793 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 794
 795 static int dax_insert_mapping(struct address_space *mapping,
 796                 struct block_device *bdev, sector_t sector, size_t size,
 797                 void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf)
 798 {
 799         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 800         struct blk_dax_ctl dax = {
 801                 .sector = sector,
 802                 .size = size,
 803         };
 804         void *ret;
 805         void *entry = *entryp;
 806
 807         if (dax_map_atomic(bdev, &dax) < 0)
 808                 return PTR_ERR(dax.addr);
 809         dax_unmap_atomic(bdev, &dax);
 810
 811         ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector);
 812         if (IS_ERR(ret))
 813                 return PTR_ERR(ret);
 814         *entryp = ret;
 815
 816         return vm_insert_mixed(vma, vaddr, dax.pfn);
 817 }
 818
 819 /**
 820  * dax_fault - handle a page fault on a DAX file
 821  * @vma: The virtual memory area where the fault occurred
 822  * @vmf: The description of the fault
 823  * @get_block: The filesystem method used to translate file offsets to blocks
 824  *
 825  * When a page fault occurs, filesystems may call this helper in their
 826  * fault handler for DAX files. dax_fault() assumes the caller has done all
 827  * the necessary locking for the page fault to proceed successfully.
 828  */
 829 int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 830                         get_block_t get_block)
 831 {
 832         struct file *file = vma->vm_file;
 833         struct address_space *mapping = file->f_mapping;
 834         struct inode *inode = mapping->host;
 835         void *entry;
 836         struct buffer_head bh;
 837         unsigned long vaddr = (unsigned long)vmf->virtual_address;
 838         unsigned blkbits = inode->i_blkbits;
 839         sector_t block;
 840         pgoff_t size;
 841         int error;
 842         int major = 0;
 843
 844         /*
 845          * Check whether offset isn't beyond end of file now. Caller is supposed
 846          * to hold locks serializing us with truncate / punch hole so this is
 847          * a reliable test.
 848          */
 849         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 850         if (vmf->pgoff >= size)
 851                 return VM_FAULT_SIGBUS;
 852
 853         memset(&bh, 0, sizeof(bh));
 854         block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 855         bh.b_bdev = inode->i_sb->s_bdev;
 856         bh.b_size = PAGE_SIZE;
 857
 858         entry = grab_mapping_entry(mapping, vmf->pgoff);
 859         if (IS_ERR(entry)) {
 860                 error = PTR_ERR(entry);
 861                 goto out;
 862         }
 863
 864         error = get_block(inode, block, &bh, 0);
 865         if (!error && (bh.b_size < PAGE_SIZE))
 866                 error = -EIO;           /* fs corruption? */
 867         if (error)
 868                 goto unlock_entry;
 869
 870         if (vmf->cow_page) {
 871                 struct page *new_page = vmf->cow_page;
 872                 if (buffer_written(&bh))
 873                         error = copy_user_dax(bh.b_bdev, to_sector(&bh, inode),
 874                                         bh.b_size, new_page, vaddr);
 875                 else
 876                         clear_user_highpage(new_page, vaddr);
 877                 if (error)
 878                         goto unlock_entry;
 879                 if (!radix_tree_exceptional_entry(entry)) {
 880                         vmf->page = entry;
 881                         return VM_FAULT_LOCKED;
 882                 }
 883                 vmf->entry = entry;
 884                 return VM_FAULT_DAX_LOCKED;
 885         }
 886
 887         if (!buffer_mapped(&bh)) {
 888                 if (vmf->flags & FAULT_FLAG_WRITE) {
 889                         error = get_block(inode, block, &bh, 1);
 890                         count_vm_event(PGMAJFAULT);
 891                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 892                         major = VM_FAULT_MAJOR;
 893                         if (!error && (bh.b_size < PAGE_SIZE))
 894                                 error = -EIO;
 895                         if (error)
 896                                 goto unlock_entry;
 897                 } else {
 898                         return dax_load_hole(mapping, entry, vmf);
 899                 }
 900         }
 901
 902         /* Filesystem should not return unwritten buffers to us! */
 903         WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
 904         error = dax_insert_mapping(mapping, bh.b_bdev, to_sector(&bh, inode),
 905                         bh.b_size, &entry, vma, vmf);
 906  unlock_entry:
 907         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
 908  out:
 909         if (error == -ENOMEM)
 910                 return VM_FAULT_OOM | major;
 911         /* -EBUSY is fine, somebody else faulted on the same PTE */
 912         if ((error < 0) && (error != -EBUSY))
 913                 return VM_FAULT_SIGBUS | major;
 914         return VM_FAULT_NOPAGE | major;
 915 }
 916 EXPORT_SYMBOL_GPL(dax_fault);
 917
 918 #if defined(CONFIG_TRANSPARENT_HUGEPAGE)
 919 /*
 920  * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 921  * more often than one might expect in the below function.
 922  */
 923 #define PG_PMD_COLOUR   ((PMD_SIZE >> PAGE_SHIFT) - 1)
 924
 925 static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 926                 const char *reason, const char *fn)
 927 {
 928         if (bh) {
 929                 char bname[BDEVNAME_SIZE];
 930                 bdevname(bh->b_bdev, bname);
 931                 pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 932                         "length %zd fallback: %s\n", fn, current->comm,
 933                         address, bname, bh->b_state, (u64)bh->b_blocknr,
 934                         bh->b_size, reason);
 935         } else {
 936                 pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 937                         current->comm, address, reason);
 938         }
 939 }
 940
 941 #define dax_pmd_dbg(bh, address, reason)        __dax_dbg(bh, address, reason, "dax_pmd")
 942
 943 /**
 944  * dax_pmd_fault - handle a PMD fault on a DAX file
 945  * @vma: The virtual memory area where the fault occurred
 946  * @vmf: The description of the fault
 947  * @get_block: The filesystem method used to translate file offsets to blocks
 948  *
 949  * When a page fault occurs, filesystems may call this helper in their
 950  * pmd_fault handler for DAX files.
 951  */
 952 int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 953                 pmd_t *pmd, unsigned int flags, get_block_t get_block)
 954 {
 955         struct file *file = vma->vm_file;
 956         struct address_space *mapping = file->f_mapping;
 957         struct inode *inode = mapping->host;
 958         struct buffer_head bh;
 959         unsigned blkbits = inode->i_blkbits;
 960         unsigned long pmd_addr = address & PMD_MASK;
 961         bool write = flags & FAULT_FLAG_WRITE;
 962         struct block_device *bdev;
 963         pgoff_t size, pgoff;
 964         sector_t block;
 965         int result = 0;
 966         bool alloc = false;
 967
 968         /* dax pmd mappings require pfn_t_devmap() */
 969         if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 970                 return VM_FAULT_FALLBACK;
 971
 972         /* Fall back to PTEs if we're going to COW */
 973         if (write && !(vma->vm_flags & VM_SHARED)) {
 974                 split_huge_pmd(vma, pmd, address);
 975                 dax_pmd_dbg(NULL, address, "cow write");
 976                 return VM_FAULT_FALLBACK;
 977         }
 978         /* If the PMD would extend outside the VMA */
 979         if (pmd_addr < vma->vm_start) {
 980                 dax_pmd_dbg(NULL, address, "vma start unaligned");
 981                 return VM_FAULT_FALLBACK;
 982         }
 983         if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 984                 dax_pmd_dbg(NULL, address, "vma end unaligned");
 985                 return VM_FAULT_FALLBACK;
 986         }
 987
 988         pgoff = linear_page_index(vma, pmd_addr);
 989         size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 990         if (pgoff >= size)
 991                 return VM_FAULT_SIGBUS;
 992         /* If the PMD would cover blocks out of the file */
 993         if ((pgoff | PG_PMD_COLOUR) >= size) {
 994                 dax_pmd_dbg(NULL, address,
 995                                 "offset + huge page size > file size");
 996                 return VM_FAULT_FALLBACK;
 997         }
 998
 999         memset(&bh, 0, sizeof(bh));
1000         bh.b_bdev = inode->i_sb->s_bdev;
1001         block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
1002
1003         bh.b_size = PMD_SIZE;
1004
1005         if (get_block(inode, block, &bh, 0) != 0)
1006                 return VM_FAULT_SIGBUS;
1007
1008         if (!buffer_mapped(&bh) && write) {
1009                 if (get_block(inode, block, &bh, 1) != 0)
1010                         return VM_FAULT_SIGBUS;
1011                 alloc = true;
1012                 WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh));
1013         }
1014
1015         bdev = bh.b_bdev;
1016
1017         /*
1018          * If the filesystem isn't willing to tell us the length of a hole,
1019          * just fall back to PTEs.  Calling get_block 512 times in a loop
1020          * would be silly.
1021          */
1022         if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
1023                 dax_pmd_dbg(&bh, address, "allocated block too small");
1024                 return VM_FAULT_FALLBACK;
1025         }
1026
1027         /*
1028          * If we allocated new storage, make sure no process has any
1029          * zero pages covering this hole
1030          */
1031         if (alloc) {
1032                 loff_t lstart = pgoff << PAGE_SHIFT;
1033                 loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
1034
1035                 truncate_pagecache_range(inode, lstart, lend);
1036         }
1037
1038         if (!write && !buffer_mapped(&bh)) {
1039                 spinlock_t *ptl;
1040                 pmd_t entry;
1041                 struct page *zero_page = mm_get_huge_zero_page(vma->vm_mm);
1042
1043                 if (unlikely(!zero_page)) {
1044                         dax_pmd_dbg(&bh, address, "no zero page");
1045                         goto fallback;
1046                 }
1047
1048                 ptl = pmd_lock(vma->vm_mm, pmd);
1049                 if (!pmd_none(*pmd)) {
1050                         spin_unlock(ptl);
1051                         dax_pmd_dbg(&bh, address, "pmd already present");
1052                         goto fallback;
1053                 }
1054
1055                 dev_dbg(part_to_dev(bdev->bd_part),
1056                                 "%s: %s addr: %lx pfn: <zero> sect: %llx\n",
1057                                 __func__, current->comm, address,
1058                                 (unsigned long long) to_sector(&bh, inode));
1059
1060                 entry = mk_pmd(zero_page, vma->vm_page_prot);
1061                 entry = pmd_mkhuge(entry);
1062                 set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
1063                 result = VM_FAULT_NOPAGE;
1064                 spin_unlock(ptl);
1065         } else {
1066                 struct blk_dax_ctl dax = {
1067                         .sector = to_sector(&bh, inode),
1068                         .size = PMD_SIZE,
1069                 };
1070                 long length = dax_map_atomic(bdev, &dax);
1071
1072                 if (length < 0) {
1073                         dax_pmd_dbg(&bh, address, "dax-error fallback");
1074                         goto fallback;
1075                 }
1076                 if (length < PMD_SIZE) {
1077                         dax_pmd_dbg(&bh, address, "dax-length too small");
1078                         dax_unmap_atomic(bdev, &dax);
1079                         goto fallback;
1080                 }
1081                 if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
1082                         dax_pmd_dbg(&bh, address, "pfn unaligned");
1083                         dax_unmap_atomic(bdev, &dax);
1084                         goto fallback;
1085                 }
1086
1087                 if (!pfn_t_devmap(dax.pfn)) {
1088                         dax_unmap_atomic(bdev, &dax);
1089                         dax_pmd_dbg(&bh, address, "pfn not in memmap");
1090                         goto fallback;
1091                 }
1092                 dax_unmap_atomic(bdev, &dax);
1093
1094                 /*
1095                  * For PTE faults we insert a radix tree entry for reads, and
1096                  * leave it clean.  Then on the first write we dirty the radix
1097                  * tree entry via the dax_pfn_mkwrite() path.  This sequence
1098                  * allows the dax_pfn_mkwrite() call to be simpler and avoid a
1099                  * call into get_block() to translate the pgoff to a sector in
1100                  * order to be able to create a new radix tree entry.
1101                  *
1102                  * The PMD path doesn't have an equivalent to
1103                  * dax_pfn_mkwrite(), though, so for a read followed by a
1104                  * write we traverse all the way through dax_pmd_fault()
1105                  * twice.  This means we can just skip inserting a radix tree
1106                  * entry completely on the initial read and just wait until
1107                  * the write to insert a dirty entry.
1108                  */
1109                 if (write) {
1110                         /*
1111                          * We should insert radix-tree entry and dirty it here.
1112                          * For now this is broken...
1113                          */
1114                 }
1115
1116                 dev_dbg(part_to_dev(bdev->bd_part),
1117                                 "%s: %s addr: %lx pfn: %lx sect: %llx\n",
1118                                 __func__, current->comm, address,
1119                                 pfn_t_to_pfn(dax.pfn),
1120                                 (unsigned long long) dax.sector);
1121                 result |= vmf_insert_pfn_pmd(vma, address, pmd,
1122                                 dax.pfn, write);
1123         }
1124
1125  out:
1126         return result;
1127
1128  fallback:
1129         count_vm_event(THP_FAULT_FALLBACK);
1130         result = VM_FAULT_FALLBACK;
1131         goto out;
1132 }
1133 EXPORT_SYMBOL_GPL(dax_pmd_fault);
1134 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1135
1136 /**
1137  * dax_pfn_mkwrite - handle first write to DAX page
1138  * @vma: The virtual memory area where the fault occurred
1139  * @vmf: The description of the fault
1140  */
1141 int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1142 {
1143         struct file *file = vma->vm_file;
1144         struct address_space *mapping = file->f_mapping;
1145         void *entry;
1146         pgoff_t index = vmf->pgoff;
1147
1148         spin_lock_irq(&mapping->tree_lock);
1149         entry = get_unlocked_mapping_entry(mapping, index, NULL);
1150         if (!entry || !radix_tree_exceptional_entry(entry))
1151                 goto out;
1152         radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY);
1153         put_unlocked_mapping_entry(mapping, index, entry);
1154 out:
1155         spin_unlock_irq(&mapping->tree_lock);
1156         return VM_FAULT_NOPAGE;
1157 }
1158 EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
1159
1160 static bool dax_range_is_aligned(struct block_device *bdev,
1161                                  unsigned int offset, unsigned int length)
1162 {
1163         unsigned short sector_size = bdev_logical_block_size(bdev);
1164
1165         if (!IS_ALIGNED(offset, sector_size))
1166                 return false;
1167         if (!IS_ALIGNED(length, sector_size))
1168                 return false;
1169
1170         return true;
1171 }
1172
1173 int __dax_zero_page_range(struct block_device *bdev, sector_t sector,
1174                 unsigned int offset, unsigned int length)
1175 {
1176         struct blk_dax_ctl dax = {
1177                 .sector         = sector,
1178                 .size           = PAGE_SIZE,
1179         };
1180
1181         if (dax_range_is_aligned(bdev, offset, length)) {
1182                 sector_t start_sector = dax.sector + (offset >> 9);
1183
1184                 return blkdev_issue_zeroout(bdev, start_sector,
1185                                 length >> 9, GFP_NOFS, true);
1186         } else {
1187                 if (dax_map_atomic(bdev, &dax) < 0)
1188                         return PTR_ERR(dax.addr);
1189                 clear_pmem(dax.addr + offset, length);
1190                 dax_unmap_atomic(bdev, &dax);
1191         }
1192         return 0;
1193 }
1194 EXPORT_SYMBOL_GPL(__dax_zero_page_range);
1195
1196 /**
1197  * dax_zero_page_range - zero a range within a page of a DAX file
1198  * @inode: The file being truncated
1199  * @from: The file offset that is being truncated to
1200  * @length: The number of bytes to zero
1201  * @get_block: The filesystem method used to translate file offsets to blocks
1202  *
1203  * This function can be called by a filesystem when it is zeroing part of a
1204  * page in a DAX file.  This is intended for hole-punch operations.  If
1205  * you are truncating a file, the helper function dax_truncate_page() may be
1206  * more convenient.
1207  */
1208 int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1209                                                         get_block_t get_block)
1210 {
1211         struct buffer_head bh;
1212         pgoff_t index = from >> PAGE_SHIFT;
1213         unsigned offset = from & (PAGE_SIZE-1);
1214         int err;
1215
1216         /* Block boundary? Nothing to do */
1217         if (!length)
1218                 return 0;
1219         BUG_ON((offset + length) > PAGE_SIZE);
1220
1221         memset(&bh, 0, sizeof(bh));
1222         bh.b_bdev = inode->i_sb->s_bdev;
1223         bh.b_size = PAGE_SIZE;
1224         err = get_block(inode, index, &bh, 0);
1225         if (err < 0 || !buffer_written(&bh))
1226                 return err;
1227
1228         return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode),
1229                         offset, length);
1230 }
1231 EXPORT_SYMBOL_GPL(dax_zero_page_range);
1232
1233 /**
1234  * dax_truncate_page - handle a partial page being truncated in a DAX file
1235  * @inode: The file being truncated
1236  * @from: The file offset that is being truncated to
1237  * @get_block: The filesystem method used to translate file offsets to blocks
1238  *
1239  * Similar to block_truncate_page(), this function can be called by a
1240  * filesystem when it is truncating a DAX file to handle the partial page.
1241  */
1242 int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1243 {
1244         unsigned length = PAGE_ALIGN(from) - from;
1245         return dax_zero_page_range(inode, from, length, get_block);
1246 }
1247 EXPORT_SYMBOL_GPL(dax_truncate_page);
1248
1249 #ifdef CONFIG_FS_IOMAP
1250 static loff_t
1251 iomap_dax_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
1252                 struct iomap *iomap)
1253 {
1254         struct iov_iter *iter = data;
1255         loff_t end = pos + length, done = 0;
1256         ssize_t ret = 0;
1257
1258         if (iov_iter_rw(iter) == READ) {
1259                 end = min(end, i_size_read(inode));
1260                 if (pos >= end)
1261                         return 0;
1262
1263                 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1264                         return iov_iter_zero(min(length, end - pos), iter);
1265         }
1266
1267         if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
1268                 return -EIO;
1269
1270         while (pos < end) {
1271                 unsigned offset = pos & (PAGE_SIZE - 1);
1272                 struct blk_dax_ctl dax = { 0 };
1273                 ssize_t map_len;
1274
1275                 dax.sector = iomap->blkno +
1276                         (((pos & PAGE_MASK) - iomap->offset) >> 9);
1277                 dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK;
1278                 map_len = dax_map_atomic(iomap->bdev, &dax);
1279                 if (map_len < 0) {
1280                         ret = map_len;
1281                         break;
1282                 }
1283
1284                 dax.addr += offset;
1285                 map_len -= offset;
1286                 if (map_len > end - pos)
1287                         map_len = end - pos;
1288
1289                 if (iov_iter_rw(iter) == WRITE)
1290                         map_len = copy_from_iter_pmem(dax.addr, map_len, iter);
1291                 else
1292                         map_len = copy_to_iter(dax.addr, map_len, iter);
1293                 dax_unmap_atomic(iomap->bdev, &dax);
1294                 if (map_len <= 0) {
1295                         ret = map_len ? map_len : -EFAULT;
1296                         break;
1297                 }
1298
1299                 pos += map_len;
1300                 length -= map_len;
1301                 done += map_len;
1302         }
1303
1304         return done ? done : ret;
1305 }
1306
1307 /**
1308  * iomap_dax_rw - Perform I/O to a DAX file
1309  * @iocb:       The control block for this I/O
1310  * @iter:       The addresses to do I/O from or to
1311  * @ops:        iomap ops passed from the file system
1312  *
1313  * This function performs read and write operations to directly mapped
1314  * persistent memory.  The callers needs to take care of read/write exclusion
1315  * and evicting any page cache pages in the region under I/O.
1316  */
1317 ssize_t
1318 iomap_dax_rw(struct kiocb *iocb, struct iov_iter *iter,
1319                 struct iomap_ops *ops)
1320 {
1321         struct address_space *mapping = iocb->ki_filp->f_mapping;
1322         struct inode *inode = mapping->host;
1323         loff_t pos = iocb->ki_pos, ret = 0, done = 0;
1324         unsigned flags = 0;
1325
1326         if (iov_iter_rw(iter) == WRITE)
1327                 flags |= IOMAP_WRITE;
1328
1329         /*
1330          * Yes, even DAX files can have page cache attached to them:  A zeroed
1331          * page is inserted into the pagecache when we have to serve a write
1332          * fault on a hole.  It should never be dirtied and can simply be
1333          * dropped from the pagecache once we get real data for the page.
1334          *
1335          * XXX: This is racy against mmap, and there's nothing we can do about
1336          * it. We'll eventually need to shift this down even further so that
1337          * we can check if we allocated blocks over a hole first.
1338          */
1339         if (mapping->nrpages) {
1340                 ret = invalidate_inode_pages2_range(mapping,
1341                                 pos >> PAGE_SHIFT,
1342                                 (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT);
1343                 WARN_ON_ONCE(ret);
1344         }
1345
1346         while (iov_iter_count(iter)) {
1347                 ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
1348                                 iter, iomap_dax_actor);
1349                 if (ret <= 0)
1350                         break;
1351                 pos += ret;
1352                 done += ret;
1353         }
1354
1355         iocb->ki_pos += done;
1356         return done ? done : ret;
1357 }
1358 EXPORT_SYMBOL_GPL(iomap_dax_rw);
1359
1360 /**
1361  * iomap_dax_fault - handle a page fault on a DAX file
1362  * @vma: The virtual memory area where the fault occurred
1363  * @vmf: The description of the fault
1364  * @ops: iomap ops passed from the file system
1365  *
1366  * When a page fault occurs, filesystems may call this helper in their fault
1367  * or mkwrite handler for DAX files. Assumes the caller has done all the
1368  * necessary locking for the page fault to proceed successfully.
1369  */
1370 int iomap_dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
1371                         struct iomap_ops *ops)
1372 {
1373         struct address_space *mapping = vma->vm_file->f_mapping;
1374         struct inode *inode = mapping->host;
1375         unsigned long vaddr = (unsigned long)vmf->virtual_address;
1376         loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
1377         sector_t sector;
1378         struct iomap iomap = { 0 };
1379         unsigned flags = 0;
1380         int error, major = 0;
1381         void *entry;
1382
1383         /*
1384          * Check whether offset isn't beyond end of file now. Caller is supposed
1385          * to hold locks serializing us with truncate / punch hole so this is
1386          * a reliable test.
1387          */
1388         if (pos >= i_size_read(inode))
1389                 return VM_FAULT_SIGBUS;
1390
1391         entry = grab_mapping_entry(mapping, vmf->pgoff);
1392         if (IS_ERR(entry)) {
1393                 error = PTR_ERR(entry);
1394                 goto out;
1395         }
1396
1397         if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1398                 flags |= IOMAP_WRITE;
1399
1400         /*
1401          * Note that we don't bother to use iomap_apply here: DAX required
1402          * the file system block size to be equal the page size, which means
1403          * that we never have to deal with more than a single extent here.
1404          */
1405         error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
1406         if (error)
1407                 goto unlock_entry;
1408         if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
1409                 error = -EIO;           /* fs corruption? */
1410                 goto unlock_entry;
1411         }
1412
1413         sector = iomap.blkno + (((pos & PAGE_MASK) - iomap.offset) >> 9);
1414
1415         if (vmf->cow_page) {
1416                 switch (iomap.type) {
1417                 case IOMAP_HOLE:
1418                 case IOMAP_UNWRITTEN:
1419                         clear_user_highpage(vmf->cow_page, vaddr);
1420                         break;
1421                 case IOMAP_MAPPED:
1422                         error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE,
1423                                         vmf->cow_page, vaddr);
1424                         break;
1425                 default:
1426                         WARN_ON_ONCE(1);
1427                         error = -EIO;
1428                         break;
1429                 }
1430
1431                 if (error)
1432                         goto unlock_entry;
1433                 if (!radix_tree_exceptional_entry(entry)) {
1434                         vmf->page = entry;
1435                         return VM_FAULT_LOCKED;
1436                 }
1437                 vmf->entry = entry;
1438                 return VM_FAULT_DAX_LOCKED;
1439         }
1440
1441         switch (iomap.type) {
1442         case IOMAP_MAPPED:
1443                 if (iomap.flags & IOMAP_F_NEW) {
1444                         count_vm_event(PGMAJFAULT);
1445                         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1446                         major = VM_FAULT_MAJOR;
1447                 }
1448                 error = dax_insert_mapping(mapping, iomap.bdev, sector,
1449                                 PAGE_SIZE, &entry, vma, vmf);
1450                 break;
1451         case IOMAP_UNWRITTEN:
1452         case IOMAP_HOLE:
1453                 if (!(vmf->flags & FAULT_FLAG_WRITE))
1454                         return dax_load_hole(mapping, entry, vmf);
1455                 /*FALLTHRU*/
1456         default:
1457                 WARN_ON_ONCE(1);
1458                 error = -EIO;
1459                 break;
1460         }
1461
1462  unlock_entry:
1463         put_locked_mapping_entry(mapping, vmf->pgoff, entry);
1464  out:
1465         if (error == -ENOMEM)
1466                 return VM_FAULT_OOM | major;
1467         /* -EBUSY is fine, somebody else faulted on the same PTE */
1468         if (error < 0 && error != -EBUSY)
1469                 return VM_FAULT_SIGBUS | major;
1470         return VM_FAULT_NOPAGE | major;
1471 }
1472 EXPORT_SYMBOL_GPL(iomap_dax_fault);
1473 #endif /* CONFIG_FS_IOMAP */