1d475c634SMatthew Wilcox /* 2d475c634SMatthew Wilcox * fs/dax.c - Direct Access filesystem code 3d475c634SMatthew Wilcox * Copyright (c) 2013-2014 Intel Corporation 4d475c634SMatthew Wilcox * Author: Matthew Wilcox <matthew.r.wilcox@intel.com> 5d475c634SMatthew Wilcox * Author: Ross Zwisler <ross.zwisler@linux.intel.com> 6d475c634SMatthew Wilcox * 7d475c634SMatthew Wilcox * This program is free software; you can redistribute it and/or modify it 8d475c634SMatthew Wilcox * under the terms and conditions of the GNU General Public License, 9d475c634SMatthew Wilcox * version 2, as published by the Free Software Foundation. 10d475c634SMatthew Wilcox * 11d475c634SMatthew Wilcox * This program is distributed in the hope it will be useful, but WITHOUT 12d475c634SMatthew Wilcox * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13d475c634SMatthew Wilcox * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 14d475c634SMatthew Wilcox * more details. 15d475c634SMatthew Wilcox */ 16d475c634SMatthew Wilcox 17d475c634SMatthew Wilcox #include <linux/atomic.h> 18d475c634SMatthew Wilcox #include <linux/blkdev.h> 19d475c634SMatthew Wilcox #include <linux/buffer_head.h> 20d77e92e2SRoss Zwisler #include <linux/dax.h> 21d475c634SMatthew Wilcox #include <linux/fs.h> 22d475c634SMatthew Wilcox #include <linux/genhd.h> 23f7ca90b1SMatthew Wilcox #include <linux/highmem.h> 24f7ca90b1SMatthew Wilcox #include <linux/memcontrol.h> 25f7ca90b1SMatthew Wilcox #include <linux/mm.h> 26d475c634SMatthew Wilcox #include <linux/mutex.h> 279973c98eSRoss Zwisler #include <linux/pagevec.h> 282765cfbbSRoss Zwisler #include <linux/pmem.h> 29289c6aedSMatthew Wilcox #include <linux/sched.h> 30d475c634SMatthew Wilcox #include <linux/uio.h> 31f7ca90b1SMatthew Wilcox #include <linux/vmstat.h> 3234c0fd54SDan Williams #include <linux/pfn_t.h> 330e749e54SDan Williams #include <linux/sizes.h> 34a254e568SChristoph Hellwig #include <linux/iomap.h> 35a254e568SChristoph Hellwig #include "internal.h" 36d475c634SMatthew Wilcox 37ac401cc7SJan Kara /* We choose 4096 entries - same as per-zone page wait tables */ 38ac401cc7SJan Kara #define DAX_WAIT_TABLE_BITS 12 39ac401cc7SJan Kara #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) 40ac401cc7SJan Kara 41ce95ab0fSRoss Zwisler static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; 42ac401cc7SJan Kara 43ac401cc7SJan Kara static int __init init_dax_wait_table(void) 44ac401cc7SJan Kara { 45ac401cc7SJan Kara int i; 46ac401cc7SJan Kara 47ac401cc7SJan Kara for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) 48ac401cc7SJan Kara init_waitqueue_head(wait_table + i); 49ac401cc7SJan Kara return 0; 50ac401cc7SJan Kara } 51ac401cc7SJan Kara fs_initcall(init_dax_wait_table); 52ac401cc7SJan Kara 53b2e0d162SDan Williams static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax) 54b2e0d162SDan Williams { 55b2e0d162SDan Williams struct request_queue *q = bdev->bd_queue; 56b2e0d162SDan Williams long rc = -EIO; 57b2e0d162SDan Williams 587a9eb206SDan Williams dax->addr = ERR_PTR(-EIO); 59b2e0d162SDan Williams if (blk_queue_enter(q, true) != 0) 60b2e0d162SDan Williams return rc; 61b2e0d162SDan Williams 62b2e0d162SDan Williams rc = bdev_direct_access(bdev, dax); 63b2e0d162SDan Williams if (rc < 0) { 647a9eb206SDan Williams dax->addr = ERR_PTR(rc); 65b2e0d162SDan Williams blk_queue_exit(q); 66b2e0d162SDan Williams return rc; 67b2e0d162SDan Williams } 68b2e0d162SDan Williams return rc; 69b2e0d162SDan Williams } 70b2e0d162SDan Williams 71b2e0d162SDan Williams static void dax_unmap_atomic(struct block_device *bdev, 72b2e0d162SDan Williams const struct blk_dax_ctl *dax) 73b2e0d162SDan Williams { 74b2e0d162SDan Williams if (IS_ERR(dax->addr)) 75b2e0d162SDan Williams return; 76b2e0d162SDan Williams blk_queue_exit(bdev->bd_queue); 77b2e0d162SDan Williams } 78b2e0d162SDan Williams 79642261acSRoss Zwisler static int dax_is_pmd_entry(void *entry) 80642261acSRoss Zwisler { 81642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_PMD; 82642261acSRoss Zwisler } 83642261acSRoss Zwisler 84642261acSRoss Zwisler static int dax_is_pte_entry(void *entry) 85642261acSRoss Zwisler { 86642261acSRoss Zwisler return !((unsigned long)entry & RADIX_DAX_PMD); 87642261acSRoss Zwisler } 88642261acSRoss Zwisler 89642261acSRoss Zwisler static int dax_is_zero_entry(void *entry) 90642261acSRoss Zwisler { 91642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_HZP; 92642261acSRoss Zwisler } 93642261acSRoss Zwisler 94642261acSRoss Zwisler static int dax_is_empty_entry(void *entry) 95642261acSRoss Zwisler { 96642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_EMPTY; 97642261acSRoss Zwisler } 98642261acSRoss Zwisler 99d1a5f2b4SDan Williams struct page *read_dax_sector(struct block_device *bdev, sector_t n) 100d1a5f2b4SDan Williams { 101d1a5f2b4SDan Williams struct page *page = alloc_pages(GFP_KERNEL, 0); 102d1a5f2b4SDan Williams struct blk_dax_ctl dax = { 103d1a5f2b4SDan Williams .size = PAGE_SIZE, 104d1a5f2b4SDan Williams .sector = n & ~((((int) PAGE_SIZE) / 512) - 1), 105d1a5f2b4SDan Williams }; 106d1a5f2b4SDan Williams long rc; 107d1a5f2b4SDan Williams 108d1a5f2b4SDan Williams if (!page) 109d1a5f2b4SDan Williams return ERR_PTR(-ENOMEM); 110d1a5f2b4SDan Williams 111d1a5f2b4SDan Williams rc = dax_map_atomic(bdev, &dax); 112d1a5f2b4SDan Williams if (rc < 0) 113d1a5f2b4SDan Williams return ERR_PTR(rc); 114d1a5f2b4SDan Williams memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE); 115d1a5f2b4SDan Williams dax_unmap_atomic(bdev, &dax); 116d1a5f2b4SDan Williams return page; 117d1a5f2b4SDan Williams } 118d1a5f2b4SDan Williams 119d475c634SMatthew Wilcox static bool buffer_written(struct buffer_head *bh) 120d475c634SMatthew Wilcox { 121d475c634SMatthew Wilcox return buffer_mapped(bh) && !buffer_unwritten(bh); 122d475c634SMatthew Wilcox } 123d475c634SMatthew Wilcox 124b2e0d162SDan Williams static sector_t to_sector(const struct buffer_head *bh, 125b2e0d162SDan Williams const struct inode *inode) 126b2e0d162SDan Williams { 127b2e0d162SDan Williams sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9); 128b2e0d162SDan Williams 129b2e0d162SDan Williams return sector; 130b2e0d162SDan Williams } 131b2e0d162SDan Williams 132a95cd631SOmar Sandoval static ssize_t dax_io(struct inode *inode, struct iov_iter *iter, 133d475c634SMatthew Wilcox loff_t start, loff_t end, get_block_t get_block, 134d475c634SMatthew Wilcox struct buffer_head *bh) 135d475c634SMatthew Wilcox { 136b2e0d162SDan Williams loff_t pos = start, max = start, bh_max = start; 13714df6a4eSDan Williams bool hole = false; 138b2e0d162SDan Williams struct block_device *bdev = NULL; 139b2e0d162SDan Williams int rw = iov_iter_rw(iter), rc; 140b2e0d162SDan Williams long map_len = 0; 141b2e0d162SDan Williams struct blk_dax_ctl dax = { 1427a9eb206SDan Williams .addr = ERR_PTR(-EIO), 143b2e0d162SDan Williams }; 144069c77bcSJan Kara unsigned blkbits = inode->i_blkbits; 145069c77bcSJan Kara sector_t file_blks = (i_size_read(inode) + (1 << blkbits) - 1) 146069c77bcSJan Kara >> blkbits; 147d475c634SMatthew Wilcox 148b2e0d162SDan Williams if (rw == READ) 149d475c634SMatthew Wilcox end = min(end, i_size_read(inode)); 150d475c634SMatthew Wilcox 151d475c634SMatthew Wilcox while (pos < end) { 1522765cfbbSRoss Zwisler size_t len; 153d475c634SMatthew Wilcox if (pos == max) { 154e94f5a22SJeff Moyer long page = pos >> PAGE_SHIFT; 155e94f5a22SJeff Moyer sector_t block = page << (PAGE_SHIFT - blkbits); 156d475c634SMatthew Wilcox unsigned first = pos - (block << blkbits); 157d475c634SMatthew Wilcox long size; 158d475c634SMatthew Wilcox 159d475c634SMatthew Wilcox if (pos == bh_max) { 160d475c634SMatthew Wilcox bh->b_size = PAGE_ALIGN(end - pos); 161d475c634SMatthew Wilcox bh->b_state = 0; 162b2e0d162SDan Williams rc = get_block(inode, block, bh, rw == WRITE); 163b2e0d162SDan Williams if (rc) 164d475c634SMatthew Wilcox break; 165d475c634SMatthew Wilcox bh_max = pos - first + bh->b_size; 166b2e0d162SDan Williams bdev = bh->b_bdev; 167069c77bcSJan Kara /* 168069c77bcSJan Kara * We allow uninitialized buffers for writes 169069c77bcSJan Kara * beyond EOF as those cannot race with faults 170069c77bcSJan Kara */ 171069c77bcSJan Kara WARN_ON_ONCE( 172069c77bcSJan Kara (buffer_new(bh) && block < file_blks) || 173069c77bcSJan Kara (rw == WRITE && buffer_unwritten(bh))); 174d475c634SMatthew Wilcox } else { 175d475c634SMatthew Wilcox unsigned done = bh->b_size - 176d475c634SMatthew Wilcox (bh_max - (pos - first)); 177d475c634SMatthew Wilcox bh->b_blocknr += done >> blkbits; 178d475c634SMatthew Wilcox bh->b_size -= done; 179d475c634SMatthew Wilcox } 180d475c634SMatthew Wilcox 181b2e0d162SDan Williams hole = rw == READ && !buffer_written(bh); 182d475c634SMatthew Wilcox if (hole) { 183d475c634SMatthew Wilcox size = bh->b_size - first; 184d475c634SMatthew Wilcox } else { 185b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 186b2e0d162SDan Williams dax.sector = to_sector(bh, inode); 187b2e0d162SDan Williams dax.size = bh->b_size; 188b2e0d162SDan Williams map_len = dax_map_atomic(bdev, &dax); 189b2e0d162SDan Williams if (map_len < 0) { 190b2e0d162SDan Williams rc = map_len; 191d475c634SMatthew Wilcox break; 192b2e0d162SDan Williams } 193b2e0d162SDan Williams dax.addr += first; 194b2e0d162SDan Williams size = map_len - first; 195d475c634SMatthew Wilcox } 19602395435SEric Sandeen /* 19702395435SEric Sandeen * pos + size is one past the last offset for IO, 19802395435SEric Sandeen * so pos + size can overflow loff_t at extreme offsets. 19902395435SEric Sandeen * Cast to u64 to catch this and get the true minimum. 20002395435SEric Sandeen */ 20102395435SEric Sandeen max = min_t(u64, pos + size, end); 202d475c634SMatthew Wilcox } 203d475c634SMatthew Wilcox 2042765cfbbSRoss Zwisler if (iov_iter_rw(iter) == WRITE) { 205b2e0d162SDan Williams len = copy_from_iter_pmem(dax.addr, max - pos, iter); 2062765cfbbSRoss Zwisler } else if (!hole) 207b2e0d162SDan Williams len = copy_to_iter((void __force *) dax.addr, max - pos, 208e2e05394SRoss Zwisler iter); 209d475c634SMatthew Wilcox else 210d475c634SMatthew Wilcox len = iov_iter_zero(max - pos, iter); 211d475c634SMatthew Wilcox 212cadfbb6eSAl Viro if (!len) { 213b2e0d162SDan Williams rc = -EFAULT; 214d475c634SMatthew Wilcox break; 215cadfbb6eSAl Viro } 216d475c634SMatthew Wilcox 217d475c634SMatthew Wilcox pos += len; 218b2e0d162SDan Williams if (!IS_ERR(dax.addr)) 219b2e0d162SDan Williams dax.addr += len; 220d475c634SMatthew Wilcox } 221d475c634SMatthew Wilcox 222b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 2232765cfbbSRoss Zwisler 224b2e0d162SDan Williams return (pos == start) ? rc : pos - start; 225d475c634SMatthew Wilcox } 226d475c634SMatthew Wilcox 227d475c634SMatthew Wilcox /** 228d475c634SMatthew Wilcox * dax_do_io - Perform I/O to a DAX file 229d475c634SMatthew Wilcox * @iocb: The control block for this I/O 230d475c634SMatthew Wilcox * @inode: The file which the I/O is directed at 231d475c634SMatthew Wilcox * @iter: The addresses to do I/O from or to 232d475c634SMatthew Wilcox * @get_block: The filesystem method used to translate file offsets to blocks 233d475c634SMatthew Wilcox * @end_io: A filesystem callback for I/O completion 234d475c634SMatthew Wilcox * @flags: See below 235d475c634SMatthew Wilcox * 236d475c634SMatthew Wilcox * This function uses the same locking scheme as do_blockdev_direct_IO: 237d475c634SMatthew Wilcox * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the 238d475c634SMatthew Wilcox * caller for writes. For reads, we take and release the i_mutex ourselves. 239d475c634SMatthew Wilcox * If DIO_LOCKING is not set, the filesystem takes care of its own locking. 240d475c634SMatthew Wilcox * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O 241d475c634SMatthew Wilcox * is in progress. 242d475c634SMatthew Wilcox */ 243a95cd631SOmar Sandoval ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode, 244c8b8e32dSChristoph Hellwig struct iov_iter *iter, get_block_t get_block, 245a95cd631SOmar Sandoval dio_iodone_t end_io, int flags) 246d475c634SMatthew Wilcox { 247d475c634SMatthew Wilcox struct buffer_head bh; 248d475c634SMatthew Wilcox ssize_t retval = -EINVAL; 249c8b8e32dSChristoph Hellwig loff_t pos = iocb->ki_pos; 250d475c634SMatthew Wilcox loff_t end = pos + iov_iter_count(iter); 251d475c634SMatthew Wilcox 252d475c634SMatthew Wilcox memset(&bh, 0, sizeof(bh)); 253eab95db6SRoss Zwisler bh.b_bdev = inode->i_sb->s_bdev; 254d475c634SMatthew Wilcox 255c3d98e39SJan Kara if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) 2565955102cSAl Viro inode_lock(inode); 257d475c634SMatthew Wilcox 258d475c634SMatthew Wilcox /* Protects against truncate */ 259bbab37ddSMatthew Wilcox if (!(flags & DIO_SKIP_DIO_COUNT)) 260fe0f07d0SJens Axboe inode_dio_begin(inode); 261d475c634SMatthew Wilcox 262a95cd631SOmar Sandoval retval = dax_io(inode, iter, pos, end, get_block, &bh); 263d475c634SMatthew Wilcox 264a95cd631SOmar Sandoval if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) 2655955102cSAl Viro inode_unlock(inode); 266d475c634SMatthew Wilcox 267187372a3SChristoph Hellwig if (end_io) { 268187372a3SChristoph Hellwig int err; 269187372a3SChristoph Hellwig 270187372a3SChristoph Hellwig err = end_io(iocb, pos, retval, bh.b_private); 271187372a3SChristoph Hellwig if (err) 272187372a3SChristoph Hellwig retval = err; 273187372a3SChristoph Hellwig } 274d475c634SMatthew Wilcox 275bbab37ddSMatthew Wilcox if (!(flags & DIO_SKIP_DIO_COUNT)) 276fe0f07d0SJens Axboe inode_dio_end(inode); 277d475c634SMatthew Wilcox return retval; 278d475c634SMatthew Wilcox } 279d475c634SMatthew Wilcox EXPORT_SYMBOL_GPL(dax_do_io); 280f7ca90b1SMatthew Wilcox 281f7ca90b1SMatthew Wilcox /* 282ac401cc7SJan Kara * DAX radix tree locking 283ac401cc7SJan Kara */ 284ac401cc7SJan Kara struct exceptional_entry_key { 285ac401cc7SJan Kara struct address_space *mapping; 28663e95b5cSRoss Zwisler pgoff_t entry_start; 287ac401cc7SJan Kara }; 288ac401cc7SJan Kara 289ac401cc7SJan Kara struct wait_exceptional_entry_queue { 290ac401cc7SJan Kara wait_queue_t wait; 291ac401cc7SJan Kara struct exceptional_entry_key key; 292ac401cc7SJan Kara }; 293ac401cc7SJan Kara 29463e95b5cSRoss Zwisler static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, 29563e95b5cSRoss Zwisler pgoff_t index, void *entry, struct exceptional_entry_key *key) 29663e95b5cSRoss Zwisler { 29763e95b5cSRoss Zwisler unsigned long hash; 29863e95b5cSRoss Zwisler 29963e95b5cSRoss Zwisler /* 30063e95b5cSRoss Zwisler * If 'entry' is a PMD, align the 'index' that we use for the wait 30163e95b5cSRoss Zwisler * queue to the start of that PMD. This ensures that all offsets in 30263e95b5cSRoss Zwisler * the range covered by the PMD map to the same bit lock. 30363e95b5cSRoss Zwisler */ 304642261acSRoss Zwisler if (dax_is_pmd_entry(entry)) 30563e95b5cSRoss Zwisler index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); 30663e95b5cSRoss Zwisler 30763e95b5cSRoss Zwisler key->mapping = mapping; 30863e95b5cSRoss Zwisler key->entry_start = index; 30963e95b5cSRoss Zwisler 31063e95b5cSRoss Zwisler hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); 31163e95b5cSRoss Zwisler return wait_table + hash; 31263e95b5cSRoss Zwisler } 31363e95b5cSRoss Zwisler 314ac401cc7SJan Kara static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, 315ac401cc7SJan Kara int sync, void *keyp) 316ac401cc7SJan Kara { 317ac401cc7SJan Kara struct exceptional_entry_key *key = keyp; 318ac401cc7SJan Kara struct wait_exceptional_entry_queue *ewait = 319ac401cc7SJan Kara container_of(wait, struct wait_exceptional_entry_queue, wait); 320ac401cc7SJan Kara 321ac401cc7SJan Kara if (key->mapping != ewait->key.mapping || 32263e95b5cSRoss Zwisler key->entry_start != ewait->key.entry_start) 323ac401cc7SJan Kara return 0; 324ac401cc7SJan Kara return autoremove_wake_function(wait, mode, sync, NULL); 325ac401cc7SJan Kara } 326ac401cc7SJan Kara 327ac401cc7SJan Kara /* 328ac401cc7SJan Kara * Check whether the given slot is locked. The function must be called with 329ac401cc7SJan Kara * mapping->tree_lock held 330ac401cc7SJan Kara */ 331ac401cc7SJan Kara static inline int slot_locked(struct address_space *mapping, void **slot) 332ac401cc7SJan Kara { 333ac401cc7SJan Kara unsigned long entry = (unsigned long) 334ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 335ac401cc7SJan Kara return entry & RADIX_DAX_ENTRY_LOCK; 336ac401cc7SJan Kara } 337ac401cc7SJan Kara 338ac401cc7SJan Kara /* 339ac401cc7SJan Kara * Mark the given slot is locked. The function must be called with 340ac401cc7SJan Kara * mapping->tree_lock held 341ac401cc7SJan Kara */ 342ac401cc7SJan Kara static inline void *lock_slot(struct address_space *mapping, void **slot) 343ac401cc7SJan Kara { 344ac401cc7SJan Kara unsigned long entry = (unsigned long) 345ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 346ac401cc7SJan Kara 347ac401cc7SJan Kara entry |= RADIX_DAX_ENTRY_LOCK; 348ac401cc7SJan Kara radix_tree_replace_slot(slot, (void *)entry); 349ac401cc7SJan Kara return (void *)entry; 350ac401cc7SJan Kara } 351ac401cc7SJan Kara 352ac401cc7SJan Kara /* 353ac401cc7SJan Kara * Mark the given slot is unlocked. The function must be called with 354ac401cc7SJan Kara * mapping->tree_lock held 355ac401cc7SJan Kara */ 356ac401cc7SJan Kara static inline void *unlock_slot(struct address_space *mapping, void **slot) 357ac401cc7SJan Kara { 358ac401cc7SJan Kara unsigned long entry = (unsigned long) 359ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 360ac401cc7SJan Kara 361ac401cc7SJan Kara entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; 362ac401cc7SJan Kara radix_tree_replace_slot(slot, (void *)entry); 363ac401cc7SJan Kara return (void *)entry; 364ac401cc7SJan Kara } 365ac401cc7SJan Kara 366ac401cc7SJan Kara /* 367ac401cc7SJan Kara * Lookup entry in radix tree, wait for it to become unlocked if it is 368ac401cc7SJan Kara * exceptional entry and return it. The caller must call 369ac401cc7SJan Kara * put_unlocked_mapping_entry() when he decided not to lock the entry or 370ac401cc7SJan Kara * put_locked_mapping_entry() when he locked the entry and now wants to 371ac401cc7SJan Kara * unlock it. 372ac401cc7SJan Kara * 373ac401cc7SJan Kara * The function must be called with mapping->tree_lock held. 374ac401cc7SJan Kara */ 375ac401cc7SJan Kara static void *get_unlocked_mapping_entry(struct address_space *mapping, 376ac401cc7SJan Kara pgoff_t index, void ***slotp) 377ac401cc7SJan Kara { 378e3ad61c6SRoss Zwisler void *entry, **slot; 379ac401cc7SJan Kara struct wait_exceptional_entry_queue ewait; 38063e95b5cSRoss Zwisler wait_queue_head_t *wq; 381ac401cc7SJan Kara 382ac401cc7SJan Kara init_wait(&ewait.wait); 383ac401cc7SJan Kara ewait.wait.func = wake_exceptional_entry_func; 384ac401cc7SJan Kara 385ac401cc7SJan Kara for (;;) { 386e3ad61c6SRoss Zwisler entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, 387ac401cc7SJan Kara &slot); 388e3ad61c6SRoss Zwisler if (!entry || !radix_tree_exceptional_entry(entry) || 389ac401cc7SJan Kara !slot_locked(mapping, slot)) { 390ac401cc7SJan Kara if (slotp) 391ac401cc7SJan Kara *slotp = slot; 392e3ad61c6SRoss Zwisler return entry; 393ac401cc7SJan Kara } 39463e95b5cSRoss Zwisler 39563e95b5cSRoss Zwisler wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); 396ac401cc7SJan Kara prepare_to_wait_exclusive(wq, &ewait.wait, 397ac401cc7SJan Kara TASK_UNINTERRUPTIBLE); 398ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 399ac401cc7SJan Kara schedule(); 400ac401cc7SJan Kara finish_wait(wq, &ewait.wait); 401ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 402ac401cc7SJan Kara } 403ac401cc7SJan Kara } 404ac401cc7SJan Kara 405422476c4SRoss Zwisler static void put_locked_mapping_entry(struct address_space *mapping, 406422476c4SRoss Zwisler pgoff_t index, void *entry) 407422476c4SRoss Zwisler { 408422476c4SRoss Zwisler if (!radix_tree_exceptional_entry(entry)) { 409422476c4SRoss Zwisler unlock_page(entry); 410422476c4SRoss Zwisler put_page(entry); 411422476c4SRoss Zwisler } else { 412422476c4SRoss Zwisler dax_unlock_mapping_entry(mapping, index); 413422476c4SRoss Zwisler } 414422476c4SRoss Zwisler } 415422476c4SRoss Zwisler 416422476c4SRoss Zwisler /* 417422476c4SRoss Zwisler * Called when we are done with radix tree entry we looked up via 418422476c4SRoss Zwisler * get_unlocked_mapping_entry() and which we didn't lock in the end. 419422476c4SRoss Zwisler */ 420422476c4SRoss Zwisler static void put_unlocked_mapping_entry(struct address_space *mapping, 421422476c4SRoss Zwisler pgoff_t index, void *entry) 422422476c4SRoss Zwisler { 423422476c4SRoss Zwisler if (!radix_tree_exceptional_entry(entry)) 424422476c4SRoss Zwisler return; 425422476c4SRoss Zwisler 426422476c4SRoss Zwisler /* We have to wake up next waiter for the radix tree entry lock */ 427422476c4SRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, false); 428422476c4SRoss Zwisler } 429422476c4SRoss Zwisler 430ac401cc7SJan Kara /* 431ac401cc7SJan Kara * Find radix tree entry at given index. If it points to a page, return with 432ac401cc7SJan Kara * the page locked. If it points to the exceptional entry, return with the 433ac401cc7SJan Kara * radix tree entry locked. If the radix tree doesn't contain given index, 434ac401cc7SJan Kara * create empty exceptional entry for the index and return with it locked. 435ac401cc7SJan Kara * 436642261acSRoss Zwisler * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will 437642261acSRoss Zwisler * either return that locked entry or will return an error. This error will 438642261acSRoss Zwisler * happen if there are any 4k entries (either zero pages or DAX entries) 439642261acSRoss Zwisler * within the 2MiB range that we are requesting. 440642261acSRoss Zwisler * 441642261acSRoss Zwisler * We always favor 4k entries over 2MiB entries. There isn't a flow where we 442642261acSRoss Zwisler * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB 443642261acSRoss Zwisler * insertion will fail if it finds any 4k entries already in the tree, and a 444642261acSRoss Zwisler * 4k insertion will cause an existing 2MiB entry to be unmapped and 445642261acSRoss Zwisler * downgraded to 4k entries. This happens for both 2MiB huge zero pages as 446642261acSRoss Zwisler * well as 2MiB empty entries. 447642261acSRoss Zwisler * 448642261acSRoss Zwisler * The exception to this downgrade path is for 2MiB DAX PMD entries that have 449642261acSRoss Zwisler * real storage backing them. We will leave these real 2MiB DAX entries in 450642261acSRoss Zwisler * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. 451642261acSRoss Zwisler * 452ac401cc7SJan Kara * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 453ac401cc7SJan Kara * persistent memory the benefit is doubtful. We can add that later if we can 454ac401cc7SJan Kara * show it helps. 455ac401cc7SJan Kara */ 456642261acSRoss Zwisler static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, 457642261acSRoss Zwisler unsigned long size_flag) 458ac401cc7SJan Kara { 459642261acSRoss Zwisler bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ 460e3ad61c6SRoss Zwisler void *entry, **slot; 461ac401cc7SJan Kara 462ac401cc7SJan Kara restart: 463ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 464e3ad61c6SRoss Zwisler entry = get_unlocked_mapping_entry(mapping, index, &slot); 465642261acSRoss Zwisler 466642261acSRoss Zwisler if (entry) { 467642261acSRoss Zwisler if (size_flag & RADIX_DAX_PMD) { 468642261acSRoss Zwisler if (!radix_tree_exceptional_entry(entry) || 469642261acSRoss Zwisler dax_is_pte_entry(entry)) { 470642261acSRoss Zwisler put_unlocked_mapping_entry(mapping, index, 471642261acSRoss Zwisler entry); 472642261acSRoss Zwisler entry = ERR_PTR(-EEXIST); 473642261acSRoss Zwisler goto out_unlock; 474642261acSRoss Zwisler } 475642261acSRoss Zwisler } else { /* trying to grab a PTE entry */ 476642261acSRoss Zwisler if (radix_tree_exceptional_entry(entry) && 477642261acSRoss Zwisler dax_is_pmd_entry(entry) && 478642261acSRoss Zwisler (dax_is_zero_entry(entry) || 479642261acSRoss Zwisler dax_is_empty_entry(entry))) { 480642261acSRoss Zwisler pmd_downgrade = true; 481642261acSRoss Zwisler } 482642261acSRoss Zwisler } 483642261acSRoss Zwisler } 484642261acSRoss Zwisler 485ac401cc7SJan Kara /* No entry for given index? Make sure radix tree is big enough. */ 486642261acSRoss Zwisler if (!entry || pmd_downgrade) { 487ac401cc7SJan Kara int err; 488ac401cc7SJan Kara 489642261acSRoss Zwisler if (pmd_downgrade) { 490642261acSRoss Zwisler /* 491642261acSRoss Zwisler * Make sure 'entry' remains valid while we drop 492642261acSRoss Zwisler * mapping->tree_lock. 493642261acSRoss Zwisler */ 494642261acSRoss Zwisler entry = lock_slot(mapping, slot); 495642261acSRoss Zwisler } 496642261acSRoss Zwisler 497ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 498ac401cc7SJan Kara err = radix_tree_preload( 499ac401cc7SJan Kara mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); 500642261acSRoss Zwisler if (err) { 501642261acSRoss Zwisler if (pmd_downgrade) 502642261acSRoss Zwisler put_locked_mapping_entry(mapping, index, entry); 503ac401cc7SJan Kara return ERR_PTR(err); 504642261acSRoss Zwisler } 505642261acSRoss Zwisler 506642261acSRoss Zwisler /* 507642261acSRoss Zwisler * Besides huge zero pages the only other thing that gets 508642261acSRoss Zwisler * downgraded are empty entries which don't need to be 509642261acSRoss Zwisler * unmapped. 510642261acSRoss Zwisler */ 511642261acSRoss Zwisler if (pmd_downgrade && dax_is_zero_entry(entry)) 512642261acSRoss Zwisler unmap_mapping_range(mapping, 513642261acSRoss Zwisler (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); 514642261acSRoss Zwisler 515ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 516642261acSRoss Zwisler 517642261acSRoss Zwisler if (pmd_downgrade) { 518642261acSRoss Zwisler radix_tree_delete(&mapping->page_tree, index); 519642261acSRoss Zwisler mapping->nrexceptional--; 520642261acSRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, 521642261acSRoss Zwisler true); 522642261acSRoss Zwisler } 523642261acSRoss Zwisler 524642261acSRoss Zwisler entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); 525642261acSRoss Zwisler 526642261acSRoss Zwisler err = __radix_tree_insert(&mapping->page_tree, index, 527642261acSRoss Zwisler dax_radix_order(entry), entry); 528ac401cc7SJan Kara radix_tree_preload_end(); 529ac401cc7SJan Kara if (err) { 530ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 531642261acSRoss Zwisler /* 532642261acSRoss Zwisler * Someone already created the entry? This is a 533642261acSRoss Zwisler * normal failure when inserting PMDs in a range 534642261acSRoss Zwisler * that already contains PTEs. In that case we want 535642261acSRoss Zwisler * to return -EEXIST immediately. 536642261acSRoss Zwisler */ 537642261acSRoss Zwisler if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD)) 538ac401cc7SJan Kara goto restart; 539642261acSRoss Zwisler /* 540642261acSRoss Zwisler * Our insertion of a DAX PMD entry failed, most 541642261acSRoss Zwisler * likely because it collided with a PTE sized entry 542642261acSRoss Zwisler * at a different index in the PMD range. We haven't 543642261acSRoss Zwisler * inserted anything into the radix tree and have no 544642261acSRoss Zwisler * waiters to wake. 545642261acSRoss Zwisler */ 546ac401cc7SJan Kara return ERR_PTR(err); 547ac401cc7SJan Kara } 548ac401cc7SJan Kara /* Good, we have inserted empty locked entry into the tree. */ 549ac401cc7SJan Kara mapping->nrexceptional++; 550ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 551e3ad61c6SRoss Zwisler return entry; 552ac401cc7SJan Kara } 553ac401cc7SJan Kara /* Normal page in radix tree? */ 554e3ad61c6SRoss Zwisler if (!radix_tree_exceptional_entry(entry)) { 555e3ad61c6SRoss Zwisler struct page *page = entry; 556ac401cc7SJan Kara 557ac401cc7SJan Kara get_page(page); 558ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 559ac401cc7SJan Kara lock_page(page); 560ac401cc7SJan Kara /* Page got truncated? Retry... */ 561ac401cc7SJan Kara if (unlikely(page->mapping != mapping)) { 562ac401cc7SJan Kara unlock_page(page); 563ac401cc7SJan Kara put_page(page); 564ac401cc7SJan Kara goto restart; 565ac401cc7SJan Kara } 566ac401cc7SJan Kara return page; 567ac401cc7SJan Kara } 568e3ad61c6SRoss Zwisler entry = lock_slot(mapping, slot); 569642261acSRoss Zwisler out_unlock: 570ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 571e3ad61c6SRoss Zwisler return entry; 572ac401cc7SJan Kara } 573ac401cc7SJan Kara 57463e95b5cSRoss Zwisler /* 57563e95b5cSRoss Zwisler * We do not necessarily hold the mapping->tree_lock when we call this 57663e95b5cSRoss Zwisler * function so it is possible that 'entry' is no longer a valid item in the 577642261acSRoss Zwisler * radix tree. This is okay because all we really need to do is to find the 578642261acSRoss Zwisler * correct waitqueue where tasks might be waiting for that old 'entry' and 579642261acSRoss Zwisler * wake them. 58063e95b5cSRoss Zwisler */ 581ac401cc7SJan Kara void dax_wake_mapping_entry_waiter(struct address_space *mapping, 58263e95b5cSRoss Zwisler pgoff_t index, void *entry, bool wake_all) 583ac401cc7SJan Kara { 58463e95b5cSRoss Zwisler struct exceptional_entry_key key; 58563e95b5cSRoss Zwisler wait_queue_head_t *wq; 58663e95b5cSRoss Zwisler 58763e95b5cSRoss Zwisler wq = dax_entry_waitqueue(mapping, index, entry, &key); 588ac401cc7SJan Kara 589ac401cc7SJan Kara /* 590ac401cc7SJan Kara * Checking for locked entry and prepare_to_wait_exclusive() happens 591ac401cc7SJan Kara * under mapping->tree_lock, ditto for entry handling in our callers. 592ac401cc7SJan Kara * So at this point all tasks that could have seen our entry locked 593ac401cc7SJan Kara * must be in the waitqueue and the following check will see them. 594ac401cc7SJan Kara */ 59563e95b5cSRoss Zwisler if (waitqueue_active(wq)) 596ac401cc7SJan Kara __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); 597ac401cc7SJan Kara } 598ac401cc7SJan Kara 599bc2466e4SJan Kara void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index) 600ac401cc7SJan Kara { 601e3ad61c6SRoss Zwisler void *entry, **slot; 602ac401cc7SJan Kara 603ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 604e3ad61c6SRoss Zwisler entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); 605e3ad61c6SRoss Zwisler if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || 606ac401cc7SJan Kara !slot_locked(mapping, slot))) { 607ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 608ac401cc7SJan Kara return; 609ac401cc7SJan Kara } 610ac401cc7SJan Kara unlock_slot(mapping, slot); 611ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 61263e95b5cSRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, false); 613ac401cc7SJan Kara } 614ac401cc7SJan Kara 615ac401cc7SJan Kara /* 616ac401cc7SJan Kara * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree 617ac401cc7SJan Kara * entry to get unlocked before deleting it. 618ac401cc7SJan Kara */ 619ac401cc7SJan Kara int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 620ac401cc7SJan Kara { 621ac401cc7SJan Kara void *entry; 622ac401cc7SJan Kara 623ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 624ac401cc7SJan Kara entry = get_unlocked_mapping_entry(mapping, index, NULL); 625ac401cc7SJan Kara /* 626ac401cc7SJan Kara * This gets called from truncate / punch_hole path. As such, the caller 627ac401cc7SJan Kara * must hold locks protecting against concurrent modifications of the 628ac401cc7SJan Kara * radix tree (usually fs-private i_mmap_sem for writing). Since the 629ac401cc7SJan Kara * caller has seen exceptional entry for this index, we better find it 630ac401cc7SJan Kara * at that index as well... 631ac401cc7SJan Kara */ 632ac401cc7SJan Kara if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry))) { 633ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 634ac401cc7SJan Kara return 0; 635ac401cc7SJan Kara } 636ac401cc7SJan Kara radix_tree_delete(&mapping->page_tree, index); 637ac401cc7SJan Kara mapping->nrexceptional--; 638ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 63963e95b5cSRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, true); 640ac401cc7SJan Kara 641ac401cc7SJan Kara return 1; 642ac401cc7SJan Kara } 643ac401cc7SJan Kara 644ac401cc7SJan Kara /* 645f7ca90b1SMatthew Wilcox * The user has performed a load from a hole in the file. Allocating 646f7ca90b1SMatthew Wilcox * a new page in the file would cause excessive storage usage for 647f7ca90b1SMatthew Wilcox * workloads with sparse files. We allocate a page cache page instead. 648f7ca90b1SMatthew Wilcox * We'll kick it out of the page cache if it's ever written to, 649f7ca90b1SMatthew Wilcox * otherwise it will simply fall out of the page cache under memory 650f7ca90b1SMatthew Wilcox * pressure without ever having been dirtied. 651f7ca90b1SMatthew Wilcox */ 652ac401cc7SJan Kara static int dax_load_hole(struct address_space *mapping, void *entry, 653f7ca90b1SMatthew Wilcox struct vm_fault *vmf) 654f7ca90b1SMatthew Wilcox { 655ac401cc7SJan Kara struct page *page; 656f7ca90b1SMatthew Wilcox 657ac401cc7SJan Kara /* Hole page already exists? Return it... */ 658ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) { 659ac401cc7SJan Kara vmf->page = entry; 660ac401cc7SJan Kara return VM_FAULT_LOCKED; 661ac401cc7SJan Kara } 662ac401cc7SJan Kara 663ac401cc7SJan Kara /* This will replace locked radix tree entry with a hole page */ 664ac401cc7SJan Kara page = find_or_create_page(mapping, vmf->pgoff, 665ac401cc7SJan Kara vmf->gfp_mask | __GFP_ZERO); 666ac401cc7SJan Kara if (!page) { 667ac401cc7SJan Kara put_locked_mapping_entry(mapping, vmf->pgoff, entry); 668ac401cc7SJan Kara return VM_FAULT_OOM; 669ac401cc7SJan Kara } 670f7ca90b1SMatthew Wilcox vmf->page = page; 671f7ca90b1SMatthew Wilcox return VM_FAULT_LOCKED; 672f7ca90b1SMatthew Wilcox } 673f7ca90b1SMatthew Wilcox 674b0d5e82fSChristoph Hellwig static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size, 675b0d5e82fSChristoph Hellwig struct page *to, unsigned long vaddr) 676f7ca90b1SMatthew Wilcox { 677b2e0d162SDan Williams struct blk_dax_ctl dax = { 678b0d5e82fSChristoph Hellwig .sector = sector, 679b0d5e82fSChristoph Hellwig .size = size, 680b2e0d162SDan Williams }; 681e2e05394SRoss Zwisler void *vto; 682e2e05394SRoss Zwisler 683b2e0d162SDan Williams if (dax_map_atomic(bdev, &dax) < 0) 684b2e0d162SDan Williams return PTR_ERR(dax.addr); 685f7ca90b1SMatthew Wilcox vto = kmap_atomic(to); 686b2e0d162SDan Williams copy_user_page(vto, (void __force *)dax.addr, vaddr, to); 687f7ca90b1SMatthew Wilcox kunmap_atomic(vto); 688b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 689f7ca90b1SMatthew Wilcox return 0; 690f7ca90b1SMatthew Wilcox } 691f7ca90b1SMatthew Wilcox 692642261acSRoss Zwisler /* 693642261acSRoss Zwisler * By this point grab_mapping_entry() has ensured that we have a locked entry 694642261acSRoss Zwisler * of the appropriate size so we don't have to worry about downgrading PMDs to 695642261acSRoss Zwisler * PTEs. If we happen to be trying to insert a PTE and there is a PMD 696642261acSRoss Zwisler * already in the tree, we will skip the insertion and just dirty the PMD as 697642261acSRoss Zwisler * appropriate. 698642261acSRoss Zwisler */ 699ac401cc7SJan Kara static void *dax_insert_mapping_entry(struct address_space *mapping, 700ac401cc7SJan Kara struct vm_fault *vmf, 701642261acSRoss Zwisler void *entry, sector_t sector, 702642261acSRoss Zwisler unsigned long flags) 7039973c98eSRoss Zwisler { 7049973c98eSRoss Zwisler struct radix_tree_root *page_tree = &mapping->page_tree; 705ac401cc7SJan Kara int error = 0; 706ac401cc7SJan Kara bool hole_fill = false; 707ac401cc7SJan Kara void *new_entry; 708ac401cc7SJan Kara pgoff_t index = vmf->pgoff; 7099973c98eSRoss Zwisler 710ac401cc7SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) 7119973c98eSRoss Zwisler __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 7129973c98eSRoss Zwisler 713ac401cc7SJan Kara /* Replacing hole page with block mapping? */ 714ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) { 715ac401cc7SJan Kara hole_fill = true; 7169973c98eSRoss Zwisler /* 717ac401cc7SJan Kara * Unmap the page now before we remove it from page cache below. 718ac401cc7SJan Kara * The page is locked so it cannot be faulted in again. 7199973c98eSRoss Zwisler */ 720ac401cc7SJan Kara unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, 721ac401cc7SJan Kara PAGE_SIZE, 0); 722ac401cc7SJan Kara error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); 7239973c98eSRoss Zwisler if (error) 724ac401cc7SJan Kara return ERR_PTR(error); 725642261acSRoss Zwisler } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) { 726642261acSRoss Zwisler /* replacing huge zero page with PMD block mapping */ 727642261acSRoss Zwisler unmap_mapping_range(mapping, 728642261acSRoss Zwisler (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); 729ac401cc7SJan Kara } 7309973c98eSRoss Zwisler 731ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 732642261acSRoss Zwisler new_entry = dax_radix_locked_entry(sector, flags); 733642261acSRoss Zwisler 734ac401cc7SJan Kara if (hole_fill) { 735ac401cc7SJan Kara __delete_from_page_cache(entry, NULL); 736ac401cc7SJan Kara /* Drop pagecache reference */ 737ac401cc7SJan Kara put_page(entry); 738642261acSRoss Zwisler error = __radix_tree_insert(page_tree, index, 739642261acSRoss Zwisler dax_radix_order(new_entry), new_entry); 740ac401cc7SJan Kara if (error) { 741ac401cc7SJan Kara new_entry = ERR_PTR(error); 742ac401cc7SJan Kara goto unlock; 743ac401cc7SJan Kara } 7449973c98eSRoss Zwisler mapping->nrexceptional++; 745642261acSRoss Zwisler } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 746642261acSRoss Zwisler /* 747642261acSRoss Zwisler * Only swap our new entry into the radix tree if the current 748642261acSRoss Zwisler * entry is a zero page or an empty entry. If a normal PTE or 749642261acSRoss Zwisler * PMD entry is already in the tree, we leave it alone. This 750642261acSRoss Zwisler * means that if we are trying to insert a PTE and the 751642261acSRoss Zwisler * existing entry is a PMD, we will just leave the PMD in the 752642261acSRoss Zwisler * tree and dirty it if necessary. 753642261acSRoss Zwisler */ 754ac401cc7SJan Kara void **slot; 755ac401cc7SJan Kara void *ret; 756ac401cc7SJan Kara 757ac401cc7SJan Kara ret = __radix_tree_lookup(page_tree, index, NULL, &slot); 758ac401cc7SJan Kara WARN_ON_ONCE(ret != entry); 759ac401cc7SJan Kara radix_tree_replace_slot(slot, new_entry); 760ac401cc7SJan Kara } 761ac401cc7SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) 7629973c98eSRoss Zwisler radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); 7639973c98eSRoss Zwisler unlock: 7649973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 765ac401cc7SJan Kara if (hole_fill) { 766ac401cc7SJan Kara radix_tree_preload_end(); 767ac401cc7SJan Kara /* 768ac401cc7SJan Kara * We don't need hole page anymore, it has been replaced with 769ac401cc7SJan Kara * locked radix tree entry now. 770ac401cc7SJan Kara */ 771ac401cc7SJan Kara if (mapping->a_ops->freepage) 772ac401cc7SJan Kara mapping->a_ops->freepage(entry); 773ac401cc7SJan Kara unlock_page(entry); 774ac401cc7SJan Kara put_page(entry); 775ac401cc7SJan Kara } 776ac401cc7SJan Kara return new_entry; 7779973c98eSRoss Zwisler } 7789973c98eSRoss Zwisler 7799973c98eSRoss Zwisler static int dax_writeback_one(struct block_device *bdev, 7809973c98eSRoss Zwisler struct address_space *mapping, pgoff_t index, void *entry) 7819973c98eSRoss Zwisler { 7829973c98eSRoss Zwisler struct radix_tree_root *page_tree = &mapping->page_tree; 7839973c98eSRoss Zwisler struct radix_tree_node *node; 7849973c98eSRoss Zwisler struct blk_dax_ctl dax; 7859973c98eSRoss Zwisler void **slot; 7869973c98eSRoss Zwisler int ret = 0; 7879973c98eSRoss Zwisler 7889973c98eSRoss Zwisler spin_lock_irq(&mapping->tree_lock); 7899973c98eSRoss Zwisler /* 7909973c98eSRoss Zwisler * Regular page slots are stabilized by the page lock even 7919973c98eSRoss Zwisler * without the tree itself locked. These unlocked entries 7929973c98eSRoss Zwisler * need verification under the tree lock. 7939973c98eSRoss Zwisler */ 7949973c98eSRoss Zwisler if (!__radix_tree_lookup(page_tree, index, &node, &slot)) 7959973c98eSRoss Zwisler goto unlock; 7969973c98eSRoss Zwisler if (*slot != entry) 7979973c98eSRoss Zwisler goto unlock; 7989973c98eSRoss Zwisler 7999973c98eSRoss Zwisler /* another fsync thread may have already written back this entry */ 8009973c98eSRoss Zwisler if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) 8019973c98eSRoss Zwisler goto unlock; 8029973c98eSRoss Zwisler 803642261acSRoss Zwisler if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 804642261acSRoss Zwisler dax_is_zero_entry(entry))) { 8059973c98eSRoss Zwisler ret = -EIO; 8069973c98eSRoss Zwisler goto unlock; 8079973c98eSRoss Zwisler } 8089973c98eSRoss Zwisler 809642261acSRoss Zwisler /* 810642261acSRoss Zwisler * Even if dax_writeback_mapping_range() was given a wbc->range_start 811642261acSRoss Zwisler * in the middle of a PMD, the 'index' we are given will be aligned to 812642261acSRoss Zwisler * the start index of the PMD, as will the sector we pull from 813642261acSRoss Zwisler * 'entry'. This allows us to flush for PMD_SIZE and not have to 814642261acSRoss Zwisler * worry about partial PMD writebacks. 815642261acSRoss Zwisler */ 816642261acSRoss Zwisler dax.sector = dax_radix_sector(entry); 817642261acSRoss Zwisler dax.size = PAGE_SIZE << dax_radix_order(entry); 8189973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 8199973c98eSRoss Zwisler 8209973c98eSRoss Zwisler /* 8219973c98eSRoss Zwisler * We cannot hold tree_lock while calling dax_map_atomic() because it 8229973c98eSRoss Zwisler * eventually calls cond_resched(). 8239973c98eSRoss Zwisler */ 8249973c98eSRoss Zwisler ret = dax_map_atomic(bdev, &dax); 8259973c98eSRoss Zwisler if (ret < 0) 8269973c98eSRoss Zwisler return ret; 8279973c98eSRoss Zwisler 8289973c98eSRoss Zwisler if (WARN_ON_ONCE(ret < dax.size)) { 8299973c98eSRoss Zwisler ret = -EIO; 8309973c98eSRoss Zwisler goto unmap; 8319973c98eSRoss Zwisler } 8329973c98eSRoss Zwisler 8339973c98eSRoss Zwisler wb_cache_pmem(dax.addr, dax.size); 8349973c98eSRoss Zwisler 8359973c98eSRoss Zwisler spin_lock_irq(&mapping->tree_lock); 8369973c98eSRoss Zwisler radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); 8379973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 8389973c98eSRoss Zwisler unmap: 8399973c98eSRoss Zwisler dax_unmap_atomic(bdev, &dax); 8409973c98eSRoss Zwisler return ret; 8419973c98eSRoss Zwisler 8429973c98eSRoss Zwisler unlock: 8439973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 8449973c98eSRoss Zwisler return ret; 8459973c98eSRoss Zwisler } 8469973c98eSRoss Zwisler 8479973c98eSRoss Zwisler /* 8489973c98eSRoss Zwisler * Flush the mapping to the persistent domain within the byte range of [start, 8499973c98eSRoss Zwisler * end]. This is required by data integrity operations to ensure file data is 8509973c98eSRoss Zwisler * on persistent storage prior to completion of the operation. 8519973c98eSRoss Zwisler */ 8527f6d5b52SRoss Zwisler int dax_writeback_mapping_range(struct address_space *mapping, 8537f6d5b52SRoss Zwisler struct block_device *bdev, struct writeback_control *wbc) 8549973c98eSRoss Zwisler { 8559973c98eSRoss Zwisler struct inode *inode = mapping->host; 856642261acSRoss Zwisler pgoff_t start_index, end_index; 8579973c98eSRoss Zwisler pgoff_t indices[PAGEVEC_SIZE]; 8589973c98eSRoss Zwisler struct pagevec pvec; 8599973c98eSRoss Zwisler bool done = false; 8609973c98eSRoss Zwisler int i, ret = 0; 8619973c98eSRoss Zwisler 8629973c98eSRoss Zwisler if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 8639973c98eSRoss Zwisler return -EIO; 8649973c98eSRoss Zwisler 8657f6d5b52SRoss Zwisler if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) 8667f6d5b52SRoss Zwisler return 0; 8677f6d5b52SRoss Zwisler 86809cbfeafSKirill A. Shutemov start_index = wbc->range_start >> PAGE_SHIFT; 86909cbfeafSKirill A. Shutemov end_index = wbc->range_end >> PAGE_SHIFT; 8709973c98eSRoss Zwisler 8719973c98eSRoss Zwisler tag_pages_for_writeback(mapping, start_index, end_index); 8729973c98eSRoss Zwisler 8739973c98eSRoss Zwisler pagevec_init(&pvec, 0); 8749973c98eSRoss Zwisler while (!done) { 8759973c98eSRoss Zwisler pvec.nr = find_get_entries_tag(mapping, start_index, 8769973c98eSRoss Zwisler PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, 8779973c98eSRoss Zwisler pvec.pages, indices); 8789973c98eSRoss Zwisler 8799973c98eSRoss Zwisler if (pvec.nr == 0) 8809973c98eSRoss Zwisler break; 8819973c98eSRoss Zwisler 8829973c98eSRoss Zwisler for (i = 0; i < pvec.nr; i++) { 8839973c98eSRoss Zwisler if (indices[i] > end_index) { 8849973c98eSRoss Zwisler done = true; 8859973c98eSRoss Zwisler break; 8869973c98eSRoss Zwisler } 8879973c98eSRoss Zwisler 8889973c98eSRoss Zwisler ret = dax_writeback_one(bdev, mapping, indices[i], 8899973c98eSRoss Zwisler pvec.pages[i]); 8909973c98eSRoss Zwisler if (ret < 0) 8919973c98eSRoss Zwisler return ret; 8929973c98eSRoss Zwisler } 8939973c98eSRoss Zwisler } 8949973c98eSRoss Zwisler return 0; 8959973c98eSRoss Zwisler } 8969973c98eSRoss Zwisler EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 8979973c98eSRoss Zwisler 898ac401cc7SJan Kara static int dax_insert_mapping(struct address_space *mapping, 8991aaba095SChristoph Hellwig struct block_device *bdev, sector_t sector, size_t size, 9001aaba095SChristoph Hellwig void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf) 901f7ca90b1SMatthew Wilcox { 902f7ca90b1SMatthew Wilcox unsigned long vaddr = (unsigned long)vmf->virtual_address; 903b2e0d162SDan Williams struct blk_dax_ctl dax = { 9041aaba095SChristoph Hellwig .sector = sector, 9051aaba095SChristoph Hellwig .size = size, 906b2e0d162SDan Williams }; 907ac401cc7SJan Kara void *ret; 908ac401cc7SJan Kara void *entry = *entryp; 909f7ca90b1SMatthew Wilcox 9104d9a2c87SJan Kara if (dax_map_atomic(bdev, &dax) < 0) 9114d9a2c87SJan Kara return PTR_ERR(dax.addr); 912b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 913f7ca90b1SMatthew Wilcox 914642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0); 9154d9a2c87SJan Kara if (IS_ERR(ret)) 9164d9a2c87SJan Kara return PTR_ERR(ret); 917ac401cc7SJan Kara *entryp = ret; 9189973c98eSRoss Zwisler 9194d9a2c87SJan Kara return vm_insert_mixed(vma, vaddr, dax.pfn); 920f7ca90b1SMatthew Wilcox } 921f7ca90b1SMatthew Wilcox 922ce5c5d55SDave Chinner /** 9236b524995SRoss Zwisler * dax_fault - handle a page fault on a DAX file 924ce5c5d55SDave Chinner * @vma: The virtual memory area where the fault occurred 925ce5c5d55SDave Chinner * @vmf: The description of the fault 926ce5c5d55SDave Chinner * @get_block: The filesystem method used to translate file offsets to blocks 927ce5c5d55SDave Chinner * 928ce5c5d55SDave Chinner * When a page fault occurs, filesystems may call this helper in their 9296b524995SRoss Zwisler * fault handler for DAX files. dax_fault() assumes the caller has done all 930ce5c5d55SDave Chinner * the necessary locking for the page fault to proceed successfully. 931ce5c5d55SDave Chinner */ 9326b524995SRoss Zwisler int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf, 93302fbd139SJan Kara get_block_t get_block) 934f7ca90b1SMatthew Wilcox { 935f7ca90b1SMatthew Wilcox struct file *file = vma->vm_file; 936f7ca90b1SMatthew Wilcox struct address_space *mapping = file->f_mapping; 937f7ca90b1SMatthew Wilcox struct inode *inode = mapping->host; 938ac401cc7SJan Kara void *entry; 939f7ca90b1SMatthew Wilcox struct buffer_head bh; 940f7ca90b1SMatthew Wilcox unsigned long vaddr = (unsigned long)vmf->virtual_address; 941f7ca90b1SMatthew Wilcox unsigned blkbits = inode->i_blkbits; 942f7ca90b1SMatthew Wilcox sector_t block; 943f7ca90b1SMatthew Wilcox pgoff_t size; 944f7ca90b1SMatthew Wilcox int error; 945f7ca90b1SMatthew Wilcox int major = 0; 946f7ca90b1SMatthew Wilcox 947ac401cc7SJan Kara /* 948ac401cc7SJan Kara * Check whether offset isn't beyond end of file now. Caller is supposed 949ac401cc7SJan Kara * to hold locks serializing us with truncate / punch hole so this is 950ac401cc7SJan Kara * a reliable test. 951ac401cc7SJan Kara */ 952f7ca90b1SMatthew Wilcox size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT; 953f7ca90b1SMatthew Wilcox if (vmf->pgoff >= size) 954f7ca90b1SMatthew Wilcox return VM_FAULT_SIGBUS; 955f7ca90b1SMatthew Wilcox 956f7ca90b1SMatthew Wilcox memset(&bh, 0, sizeof(bh)); 957f7ca90b1SMatthew Wilcox block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits); 958eab95db6SRoss Zwisler bh.b_bdev = inode->i_sb->s_bdev; 959f7ca90b1SMatthew Wilcox bh.b_size = PAGE_SIZE; 960f7ca90b1SMatthew Wilcox 961642261acSRoss Zwisler entry = grab_mapping_entry(mapping, vmf->pgoff, 0); 962ac401cc7SJan Kara if (IS_ERR(entry)) { 963ac401cc7SJan Kara error = PTR_ERR(entry); 964ac401cc7SJan Kara goto out; 965f7ca90b1SMatthew Wilcox } 966f7ca90b1SMatthew Wilcox 967f7ca90b1SMatthew Wilcox error = get_block(inode, block, &bh, 0); 968f7ca90b1SMatthew Wilcox if (!error && (bh.b_size < PAGE_SIZE)) 969f7ca90b1SMatthew Wilcox error = -EIO; /* fs corruption? */ 970f7ca90b1SMatthew Wilcox if (error) 971ac401cc7SJan Kara goto unlock_entry; 972f7ca90b1SMatthew Wilcox 973f7ca90b1SMatthew Wilcox if (vmf->cow_page) { 974f7ca90b1SMatthew Wilcox struct page *new_page = vmf->cow_page; 975f7ca90b1SMatthew Wilcox if (buffer_written(&bh)) 976b0d5e82fSChristoph Hellwig error = copy_user_dax(bh.b_bdev, to_sector(&bh, inode), 977b0d5e82fSChristoph Hellwig bh.b_size, new_page, vaddr); 978f7ca90b1SMatthew Wilcox else 979f7ca90b1SMatthew Wilcox clear_user_highpage(new_page, vaddr); 980f7ca90b1SMatthew Wilcox if (error) 981ac401cc7SJan Kara goto unlock_entry; 982ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) { 983ac401cc7SJan Kara vmf->page = entry; 984f7ca90b1SMatthew Wilcox return VM_FAULT_LOCKED; 985f7ca90b1SMatthew Wilcox } 986bc2466e4SJan Kara vmf->entry = entry; 987bc2466e4SJan Kara return VM_FAULT_DAX_LOCKED; 988bc2466e4SJan Kara } 989f7ca90b1SMatthew Wilcox 990ac401cc7SJan Kara if (!buffer_mapped(&bh)) { 991ac401cc7SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) { 992ac401cc7SJan Kara error = get_block(inode, block, &bh, 1); 993ac401cc7SJan Kara count_vm_event(PGMAJFAULT); 994ac401cc7SJan Kara mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); 995ac401cc7SJan Kara major = VM_FAULT_MAJOR; 996ac401cc7SJan Kara if (!error && (bh.b_size < PAGE_SIZE)) 997ac401cc7SJan Kara error = -EIO; 998ac401cc7SJan Kara if (error) 999ac401cc7SJan Kara goto unlock_entry; 1000ac401cc7SJan Kara } else { 1001ac401cc7SJan Kara return dax_load_hole(mapping, entry, vmf); 1002ac401cc7SJan Kara } 1003f7ca90b1SMatthew Wilcox } 1004f7ca90b1SMatthew Wilcox 100502fbd139SJan Kara /* Filesystem should not return unwritten buffers to us! */ 10062b10945cSJan Kara WARN_ON_ONCE(buffer_unwritten(&bh) || buffer_new(&bh)); 10071aaba095SChristoph Hellwig error = dax_insert_mapping(mapping, bh.b_bdev, to_sector(&bh, inode), 10081aaba095SChristoph Hellwig bh.b_size, &entry, vma, vmf); 1009ac401cc7SJan Kara unlock_entry: 1010ac401cc7SJan Kara put_locked_mapping_entry(mapping, vmf->pgoff, entry); 1011f7ca90b1SMatthew Wilcox out: 1012f7ca90b1SMatthew Wilcox if (error == -ENOMEM) 1013f7ca90b1SMatthew Wilcox return VM_FAULT_OOM | major; 1014f7ca90b1SMatthew Wilcox /* -EBUSY is fine, somebody else faulted on the same PTE */ 1015f7ca90b1SMatthew Wilcox if ((error < 0) && (error != -EBUSY)) 1016f7ca90b1SMatthew Wilcox return VM_FAULT_SIGBUS | major; 1017f7ca90b1SMatthew Wilcox return VM_FAULT_NOPAGE | major; 1018f7ca90b1SMatthew Wilcox } 1019f7ca90b1SMatthew Wilcox EXPORT_SYMBOL_GPL(dax_fault); 10204c0ccfefSMatthew Wilcox 10214c0ccfefSMatthew Wilcox /** 10220e3b210cSBoaz Harrosh * dax_pfn_mkwrite - handle first write to DAX page 10230e3b210cSBoaz Harrosh * @vma: The virtual memory area where the fault occurred 10240e3b210cSBoaz Harrosh * @vmf: The description of the fault 10250e3b210cSBoaz Harrosh */ 10260e3b210cSBoaz Harrosh int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 10270e3b210cSBoaz Harrosh { 10289973c98eSRoss Zwisler struct file *file = vma->vm_file; 1029ac401cc7SJan Kara struct address_space *mapping = file->f_mapping; 1030ac401cc7SJan Kara void *entry; 1031ac401cc7SJan Kara pgoff_t index = vmf->pgoff; 10320e3b210cSBoaz Harrosh 1033ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 1034ac401cc7SJan Kara entry = get_unlocked_mapping_entry(mapping, index, NULL); 1035ac401cc7SJan Kara if (!entry || !radix_tree_exceptional_entry(entry)) 1036ac401cc7SJan Kara goto out; 1037ac401cc7SJan Kara radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); 1038ac401cc7SJan Kara put_unlocked_mapping_entry(mapping, index, entry); 1039ac401cc7SJan Kara out: 1040ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 10410e3b210cSBoaz Harrosh return VM_FAULT_NOPAGE; 10420e3b210cSBoaz Harrosh } 10430e3b210cSBoaz Harrosh EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); 10440e3b210cSBoaz Harrosh 10454b0228faSVishal Verma static bool dax_range_is_aligned(struct block_device *bdev, 10464b0228faSVishal Verma unsigned int offset, unsigned int length) 10474b0228faSVishal Verma { 10484b0228faSVishal Verma unsigned short sector_size = bdev_logical_block_size(bdev); 10494b0228faSVishal Verma 10504b0228faSVishal Verma if (!IS_ALIGNED(offset, sector_size)) 10514b0228faSVishal Verma return false; 10524b0228faSVishal Verma if (!IS_ALIGNED(length, sector_size)) 10534b0228faSVishal Verma return false; 10544b0228faSVishal Verma 10554b0228faSVishal Verma return true; 10564b0228faSVishal Verma } 10574b0228faSVishal Verma 1058679c8bd3SChristoph Hellwig int __dax_zero_page_range(struct block_device *bdev, sector_t sector, 1059679c8bd3SChristoph Hellwig unsigned int offset, unsigned int length) 1060679c8bd3SChristoph Hellwig { 1061679c8bd3SChristoph Hellwig struct blk_dax_ctl dax = { 1062679c8bd3SChristoph Hellwig .sector = sector, 1063679c8bd3SChristoph Hellwig .size = PAGE_SIZE, 1064679c8bd3SChristoph Hellwig }; 1065679c8bd3SChristoph Hellwig 10664b0228faSVishal Verma if (dax_range_is_aligned(bdev, offset, length)) { 10674b0228faSVishal Verma sector_t start_sector = dax.sector + (offset >> 9); 10684b0228faSVishal Verma 10694b0228faSVishal Verma return blkdev_issue_zeroout(bdev, start_sector, 10704b0228faSVishal Verma length >> 9, GFP_NOFS, true); 10714b0228faSVishal Verma } else { 1072679c8bd3SChristoph Hellwig if (dax_map_atomic(bdev, &dax) < 0) 1073679c8bd3SChristoph Hellwig return PTR_ERR(dax.addr); 1074679c8bd3SChristoph Hellwig clear_pmem(dax.addr + offset, length); 1075679c8bd3SChristoph Hellwig dax_unmap_atomic(bdev, &dax); 10764b0228faSVishal Verma } 1077679c8bd3SChristoph Hellwig return 0; 1078679c8bd3SChristoph Hellwig } 1079679c8bd3SChristoph Hellwig EXPORT_SYMBOL_GPL(__dax_zero_page_range); 1080679c8bd3SChristoph Hellwig 10810e3b210cSBoaz Harrosh /** 108225726bc1SMatthew Wilcox * dax_zero_page_range - zero a range within a page of a DAX file 10834c0ccfefSMatthew Wilcox * @inode: The file being truncated 10844c0ccfefSMatthew Wilcox * @from: The file offset that is being truncated to 108525726bc1SMatthew Wilcox * @length: The number of bytes to zero 10864c0ccfefSMatthew Wilcox * @get_block: The filesystem method used to translate file offsets to blocks 10874c0ccfefSMatthew Wilcox * 108825726bc1SMatthew Wilcox * This function can be called by a filesystem when it is zeroing part of a 108925726bc1SMatthew Wilcox * page in a DAX file. This is intended for hole-punch operations. If 109025726bc1SMatthew Wilcox * you are truncating a file, the helper function dax_truncate_page() may be 109125726bc1SMatthew Wilcox * more convenient. 10924c0ccfefSMatthew Wilcox */ 109325726bc1SMatthew Wilcox int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length, 109425726bc1SMatthew Wilcox get_block_t get_block) 10954c0ccfefSMatthew Wilcox { 10964c0ccfefSMatthew Wilcox struct buffer_head bh; 109709cbfeafSKirill A. Shutemov pgoff_t index = from >> PAGE_SHIFT; 109809cbfeafSKirill A. Shutemov unsigned offset = from & (PAGE_SIZE-1); 10994c0ccfefSMatthew Wilcox int err; 11004c0ccfefSMatthew Wilcox 11014c0ccfefSMatthew Wilcox /* Block boundary? Nothing to do */ 11024c0ccfefSMatthew Wilcox if (!length) 11034c0ccfefSMatthew Wilcox return 0; 1104aada54f9SRoss Zwisler if (WARN_ON_ONCE((offset + length) > PAGE_SIZE)) 1105aada54f9SRoss Zwisler return -EINVAL; 11064c0ccfefSMatthew Wilcox 11074c0ccfefSMatthew Wilcox memset(&bh, 0, sizeof(bh)); 1108eab95db6SRoss Zwisler bh.b_bdev = inode->i_sb->s_bdev; 110909cbfeafSKirill A. Shutemov bh.b_size = PAGE_SIZE; 11104c0ccfefSMatthew Wilcox err = get_block(inode, index, &bh, 0); 1111679c8bd3SChristoph Hellwig if (err < 0 || !buffer_written(&bh)) 11124c0ccfefSMatthew Wilcox return err; 1113b2e0d162SDan Williams 1114679c8bd3SChristoph Hellwig return __dax_zero_page_range(bh.b_bdev, to_sector(&bh, inode), 1115679c8bd3SChristoph Hellwig offset, length); 11164c0ccfefSMatthew Wilcox } 111725726bc1SMatthew Wilcox EXPORT_SYMBOL_GPL(dax_zero_page_range); 111825726bc1SMatthew Wilcox 111925726bc1SMatthew Wilcox /** 112025726bc1SMatthew Wilcox * dax_truncate_page - handle a partial page being truncated in a DAX file 112125726bc1SMatthew Wilcox * @inode: The file being truncated 112225726bc1SMatthew Wilcox * @from: The file offset that is being truncated to 112325726bc1SMatthew Wilcox * @get_block: The filesystem method used to translate file offsets to blocks 112425726bc1SMatthew Wilcox * 112525726bc1SMatthew Wilcox * Similar to block_truncate_page(), this function can be called by a 112625726bc1SMatthew Wilcox * filesystem when it is truncating a DAX file to handle the partial page. 112725726bc1SMatthew Wilcox */ 112825726bc1SMatthew Wilcox int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block) 112925726bc1SMatthew Wilcox { 113009cbfeafSKirill A. Shutemov unsigned length = PAGE_ALIGN(from) - from; 113125726bc1SMatthew Wilcox return dax_zero_page_range(inode, from, length, get_block); 113225726bc1SMatthew Wilcox } 11334c0ccfefSMatthew Wilcox EXPORT_SYMBOL_GPL(dax_truncate_page); 1134a254e568SChristoph Hellwig 1135a254e568SChristoph Hellwig #ifdef CONFIG_FS_IOMAP 1136333ccc97SRoss Zwisler static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) 1137333ccc97SRoss Zwisler { 1138333ccc97SRoss Zwisler return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); 1139333ccc97SRoss Zwisler } 1140333ccc97SRoss Zwisler 1141a254e568SChristoph Hellwig static loff_t 114211c59c92SRoss Zwisler dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, 1143a254e568SChristoph Hellwig struct iomap *iomap) 1144a254e568SChristoph Hellwig { 1145a254e568SChristoph Hellwig struct iov_iter *iter = data; 1146a254e568SChristoph Hellwig loff_t end = pos + length, done = 0; 1147a254e568SChristoph Hellwig ssize_t ret = 0; 1148a254e568SChristoph Hellwig 1149a254e568SChristoph Hellwig if (iov_iter_rw(iter) == READ) { 1150a254e568SChristoph Hellwig end = min(end, i_size_read(inode)); 1151a254e568SChristoph Hellwig if (pos >= end) 1152a254e568SChristoph Hellwig return 0; 1153a254e568SChristoph Hellwig 1154a254e568SChristoph Hellwig if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1155a254e568SChristoph Hellwig return iov_iter_zero(min(length, end - pos), iter); 1156a254e568SChristoph Hellwig } 1157a254e568SChristoph Hellwig 1158a254e568SChristoph Hellwig if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) 1159a254e568SChristoph Hellwig return -EIO; 1160a254e568SChristoph Hellwig 1161a254e568SChristoph Hellwig while (pos < end) { 1162a254e568SChristoph Hellwig unsigned offset = pos & (PAGE_SIZE - 1); 1163a254e568SChristoph Hellwig struct blk_dax_ctl dax = { 0 }; 1164a254e568SChristoph Hellwig ssize_t map_len; 1165a254e568SChristoph Hellwig 1166333ccc97SRoss Zwisler dax.sector = dax_iomap_sector(iomap, pos); 1167a254e568SChristoph Hellwig dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK; 1168a254e568SChristoph Hellwig map_len = dax_map_atomic(iomap->bdev, &dax); 1169a254e568SChristoph Hellwig if (map_len < 0) { 1170a254e568SChristoph Hellwig ret = map_len; 1171a254e568SChristoph Hellwig break; 1172a254e568SChristoph Hellwig } 1173a254e568SChristoph Hellwig 1174a254e568SChristoph Hellwig dax.addr += offset; 1175a254e568SChristoph Hellwig map_len -= offset; 1176a254e568SChristoph Hellwig if (map_len > end - pos) 1177a254e568SChristoph Hellwig map_len = end - pos; 1178a254e568SChristoph Hellwig 1179a254e568SChristoph Hellwig if (iov_iter_rw(iter) == WRITE) 1180a254e568SChristoph Hellwig map_len = copy_from_iter_pmem(dax.addr, map_len, iter); 1181a254e568SChristoph Hellwig else 1182a254e568SChristoph Hellwig map_len = copy_to_iter(dax.addr, map_len, iter); 1183a254e568SChristoph Hellwig dax_unmap_atomic(iomap->bdev, &dax); 1184a254e568SChristoph Hellwig if (map_len <= 0) { 1185a254e568SChristoph Hellwig ret = map_len ? map_len : -EFAULT; 1186a254e568SChristoph Hellwig break; 1187a254e568SChristoph Hellwig } 1188a254e568SChristoph Hellwig 1189a254e568SChristoph Hellwig pos += map_len; 1190a254e568SChristoph Hellwig length -= map_len; 1191a254e568SChristoph Hellwig done += map_len; 1192a254e568SChristoph Hellwig } 1193a254e568SChristoph Hellwig 1194a254e568SChristoph Hellwig return done ? done : ret; 1195a254e568SChristoph Hellwig } 1196a254e568SChristoph Hellwig 1197a254e568SChristoph Hellwig /** 119811c59c92SRoss Zwisler * dax_iomap_rw - Perform I/O to a DAX file 1199a254e568SChristoph Hellwig * @iocb: The control block for this I/O 1200a254e568SChristoph Hellwig * @iter: The addresses to do I/O from or to 1201a254e568SChristoph Hellwig * @ops: iomap ops passed from the file system 1202a254e568SChristoph Hellwig * 1203a254e568SChristoph Hellwig * This function performs read and write operations to directly mapped 1204a254e568SChristoph Hellwig * persistent memory. The callers needs to take care of read/write exclusion 1205a254e568SChristoph Hellwig * and evicting any page cache pages in the region under I/O. 1206a254e568SChristoph Hellwig */ 1207a254e568SChristoph Hellwig ssize_t 120811c59c92SRoss Zwisler dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 1209a254e568SChristoph Hellwig struct iomap_ops *ops) 1210a254e568SChristoph Hellwig { 1211a254e568SChristoph Hellwig struct address_space *mapping = iocb->ki_filp->f_mapping; 1212a254e568SChristoph Hellwig struct inode *inode = mapping->host; 1213a254e568SChristoph Hellwig loff_t pos = iocb->ki_pos, ret = 0, done = 0; 1214a254e568SChristoph Hellwig unsigned flags = 0; 1215a254e568SChristoph Hellwig 1216a254e568SChristoph Hellwig if (iov_iter_rw(iter) == WRITE) 1217a254e568SChristoph Hellwig flags |= IOMAP_WRITE; 1218a254e568SChristoph Hellwig 1219a254e568SChristoph Hellwig /* 1220a254e568SChristoph Hellwig * Yes, even DAX files can have page cache attached to them: A zeroed 1221a254e568SChristoph Hellwig * page is inserted into the pagecache when we have to serve a write 1222a254e568SChristoph Hellwig * fault on a hole. It should never be dirtied and can simply be 1223a254e568SChristoph Hellwig * dropped from the pagecache once we get real data for the page. 1224a254e568SChristoph Hellwig * 1225a254e568SChristoph Hellwig * XXX: This is racy against mmap, and there's nothing we can do about 1226a254e568SChristoph Hellwig * it. We'll eventually need to shift this down even further so that 1227a254e568SChristoph Hellwig * we can check if we allocated blocks over a hole first. 1228a254e568SChristoph Hellwig */ 1229a254e568SChristoph Hellwig if (mapping->nrpages) { 1230a254e568SChristoph Hellwig ret = invalidate_inode_pages2_range(mapping, 1231a254e568SChristoph Hellwig pos >> PAGE_SHIFT, 1232a254e568SChristoph Hellwig (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT); 1233a254e568SChristoph Hellwig WARN_ON_ONCE(ret); 1234a254e568SChristoph Hellwig } 1235a254e568SChristoph Hellwig 1236a254e568SChristoph Hellwig while (iov_iter_count(iter)) { 1237a254e568SChristoph Hellwig ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, 123811c59c92SRoss Zwisler iter, dax_iomap_actor); 1239a254e568SChristoph Hellwig if (ret <= 0) 1240a254e568SChristoph Hellwig break; 1241a254e568SChristoph Hellwig pos += ret; 1242a254e568SChristoph Hellwig done += ret; 1243a254e568SChristoph Hellwig } 1244a254e568SChristoph Hellwig 1245a254e568SChristoph Hellwig iocb->ki_pos += done; 1246a254e568SChristoph Hellwig return done ? done : ret; 1247a254e568SChristoph Hellwig } 124811c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_rw); 1249a7d73fe6SChristoph Hellwig 1250a7d73fe6SChristoph Hellwig /** 125111c59c92SRoss Zwisler * dax_iomap_fault - handle a page fault on a DAX file 1252a7d73fe6SChristoph Hellwig * @vma: The virtual memory area where the fault occurred 1253a7d73fe6SChristoph Hellwig * @vmf: The description of the fault 1254a7d73fe6SChristoph Hellwig * @ops: iomap ops passed from the file system 1255a7d73fe6SChristoph Hellwig * 1256a7d73fe6SChristoph Hellwig * When a page fault occurs, filesystems may call this helper in their fault 1257a7d73fe6SChristoph Hellwig * or mkwrite handler for DAX files. Assumes the caller has done all the 1258a7d73fe6SChristoph Hellwig * necessary locking for the page fault to proceed successfully. 1259a7d73fe6SChristoph Hellwig */ 126011c59c92SRoss Zwisler int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf, 1261a7d73fe6SChristoph Hellwig struct iomap_ops *ops) 1262a7d73fe6SChristoph Hellwig { 1263a7d73fe6SChristoph Hellwig struct address_space *mapping = vma->vm_file->f_mapping; 1264a7d73fe6SChristoph Hellwig struct inode *inode = mapping->host; 1265a7d73fe6SChristoph Hellwig unsigned long vaddr = (unsigned long)vmf->virtual_address; 1266a7d73fe6SChristoph Hellwig loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; 1267a7d73fe6SChristoph Hellwig sector_t sector; 1268a7d73fe6SChristoph Hellwig struct iomap iomap = { 0 }; 1269a7d73fe6SChristoph Hellwig unsigned flags = 0; 1270a7d73fe6SChristoph Hellwig int error, major = 0; 12711550290bSRoss Zwisler int locked_status = 0; 1272a7d73fe6SChristoph Hellwig void *entry; 1273a7d73fe6SChristoph Hellwig 1274a7d73fe6SChristoph Hellwig /* 1275a7d73fe6SChristoph Hellwig * Check whether offset isn't beyond end of file now. Caller is supposed 1276a7d73fe6SChristoph Hellwig * to hold locks serializing us with truncate / punch hole so this is 1277a7d73fe6SChristoph Hellwig * a reliable test. 1278a7d73fe6SChristoph Hellwig */ 1279a7d73fe6SChristoph Hellwig if (pos >= i_size_read(inode)) 1280a7d73fe6SChristoph Hellwig return VM_FAULT_SIGBUS; 1281a7d73fe6SChristoph Hellwig 1282642261acSRoss Zwisler entry = grab_mapping_entry(mapping, vmf->pgoff, 0); 1283a7d73fe6SChristoph Hellwig if (IS_ERR(entry)) { 1284a7d73fe6SChristoph Hellwig error = PTR_ERR(entry); 1285a7d73fe6SChristoph Hellwig goto out; 1286a7d73fe6SChristoph Hellwig } 1287a7d73fe6SChristoph Hellwig 1288a7d73fe6SChristoph Hellwig if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) 1289a7d73fe6SChristoph Hellwig flags |= IOMAP_WRITE; 1290a7d73fe6SChristoph Hellwig 1291a7d73fe6SChristoph Hellwig /* 1292a7d73fe6SChristoph Hellwig * Note that we don't bother to use iomap_apply here: DAX required 1293a7d73fe6SChristoph Hellwig * the file system block size to be equal the page size, which means 1294a7d73fe6SChristoph Hellwig * that we never have to deal with more than a single extent here. 1295a7d73fe6SChristoph Hellwig */ 1296a7d73fe6SChristoph Hellwig error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); 1297a7d73fe6SChristoph Hellwig if (error) 1298a7d73fe6SChristoph Hellwig goto unlock_entry; 1299a7d73fe6SChristoph Hellwig if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { 1300a7d73fe6SChristoph Hellwig error = -EIO; /* fs corruption? */ 13011550290bSRoss Zwisler goto finish_iomap; 1302a7d73fe6SChristoph Hellwig } 1303a7d73fe6SChristoph Hellwig 1304333ccc97SRoss Zwisler sector = dax_iomap_sector(&iomap, pos); 1305a7d73fe6SChristoph Hellwig 1306a7d73fe6SChristoph Hellwig if (vmf->cow_page) { 1307a7d73fe6SChristoph Hellwig switch (iomap.type) { 1308a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 1309a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1310a7d73fe6SChristoph Hellwig clear_user_highpage(vmf->cow_page, vaddr); 1311a7d73fe6SChristoph Hellwig break; 1312a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1313a7d73fe6SChristoph Hellwig error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE, 1314a7d73fe6SChristoph Hellwig vmf->cow_page, vaddr); 1315a7d73fe6SChristoph Hellwig break; 1316a7d73fe6SChristoph Hellwig default: 1317a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1318a7d73fe6SChristoph Hellwig error = -EIO; 1319a7d73fe6SChristoph Hellwig break; 1320a7d73fe6SChristoph Hellwig } 1321a7d73fe6SChristoph Hellwig 1322a7d73fe6SChristoph Hellwig if (error) 13231550290bSRoss Zwisler goto finish_iomap; 1324a7d73fe6SChristoph Hellwig if (!radix_tree_exceptional_entry(entry)) { 1325a7d73fe6SChristoph Hellwig vmf->page = entry; 13261550290bSRoss Zwisler locked_status = VM_FAULT_LOCKED; 13271550290bSRoss Zwisler } else { 1328a7d73fe6SChristoph Hellwig vmf->entry = entry; 13291550290bSRoss Zwisler locked_status = VM_FAULT_DAX_LOCKED; 13301550290bSRoss Zwisler } 13311550290bSRoss Zwisler goto finish_iomap; 1332a7d73fe6SChristoph Hellwig } 1333a7d73fe6SChristoph Hellwig 1334a7d73fe6SChristoph Hellwig switch (iomap.type) { 1335a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1336a7d73fe6SChristoph Hellwig if (iomap.flags & IOMAP_F_NEW) { 1337a7d73fe6SChristoph Hellwig count_vm_event(PGMAJFAULT); 1338a7d73fe6SChristoph Hellwig mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); 1339a7d73fe6SChristoph Hellwig major = VM_FAULT_MAJOR; 1340a7d73fe6SChristoph Hellwig } 1341a7d73fe6SChristoph Hellwig error = dax_insert_mapping(mapping, iomap.bdev, sector, 1342a7d73fe6SChristoph Hellwig PAGE_SIZE, &entry, vma, vmf); 1343a7d73fe6SChristoph Hellwig break; 1344a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1345a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 13461550290bSRoss Zwisler if (!(vmf->flags & FAULT_FLAG_WRITE)) { 13471550290bSRoss Zwisler locked_status = dax_load_hole(mapping, entry, vmf); 13481550290bSRoss Zwisler break; 13491550290bSRoss Zwisler } 1350a7d73fe6SChristoph Hellwig /*FALLTHRU*/ 1351a7d73fe6SChristoph Hellwig default: 1352a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1353a7d73fe6SChristoph Hellwig error = -EIO; 1354a7d73fe6SChristoph Hellwig break; 1355a7d73fe6SChristoph Hellwig } 1356a7d73fe6SChristoph Hellwig 13571550290bSRoss Zwisler finish_iomap: 13581550290bSRoss Zwisler if (ops->iomap_end) { 13591550290bSRoss Zwisler if (error) { 13601550290bSRoss Zwisler /* keep previous error */ 13611550290bSRoss Zwisler ops->iomap_end(inode, pos, PAGE_SIZE, 0, flags, 13621550290bSRoss Zwisler &iomap); 13631550290bSRoss Zwisler } else { 13641550290bSRoss Zwisler error = ops->iomap_end(inode, pos, PAGE_SIZE, 13651550290bSRoss Zwisler PAGE_SIZE, flags, &iomap); 13661550290bSRoss Zwisler } 13671550290bSRoss Zwisler } 1368a7d73fe6SChristoph Hellwig unlock_entry: 13691550290bSRoss Zwisler if (!locked_status || error) 1370a7d73fe6SChristoph Hellwig put_locked_mapping_entry(mapping, vmf->pgoff, entry); 1371a7d73fe6SChristoph Hellwig out: 1372a7d73fe6SChristoph Hellwig if (error == -ENOMEM) 1373a7d73fe6SChristoph Hellwig return VM_FAULT_OOM | major; 1374a7d73fe6SChristoph Hellwig /* -EBUSY is fine, somebody else faulted on the same PTE */ 1375a7d73fe6SChristoph Hellwig if (error < 0 && error != -EBUSY) 1376a7d73fe6SChristoph Hellwig return VM_FAULT_SIGBUS | major; 13771550290bSRoss Zwisler if (locked_status) { 13781550290bSRoss Zwisler WARN_ON_ONCE(error); /* -EBUSY from ops->iomap_end? */ 13791550290bSRoss Zwisler return locked_status; 13801550290bSRoss Zwisler } 1381a7d73fe6SChristoph Hellwig return VM_FAULT_NOPAGE | major; 1382a7d73fe6SChristoph Hellwig } 138311c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_fault); 1384642261acSRoss Zwisler 1385642261acSRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 1386642261acSRoss Zwisler /* 1387642261acSRoss Zwisler * The 'colour' (ie low bits) within a PMD of a page offset. This comes up 1388642261acSRoss Zwisler * more often than one might expect in the below functions. 1389642261acSRoss Zwisler */ 1390642261acSRoss Zwisler #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) 1391642261acSRoss Zwisler 1392642261acSRoss Zwisler static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd, 1393642261acSRoss Zwisler struct vm_fault *vmf, unsigned long address, 1394642261acSRoss Zwisler struct iomap *iomap, loff_t pos, bool write, void **entryp) 1395642261acSRoss Zwisler { 1396642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1397642261acSRoss Zwisler struct block_device *bdev = iomap->bdev; 1398642261acSRoss Zwisler struct blk_dax_ctl dax = { 1399642261acSRoss Zwisler .sector = dax_iomap_sector(iomap, pos), 1400642261acSRoss Zwisler .size = PMD_SIZE, 1401642261acSRoss Zwisler }; 1402642261acSRoss Zwisler long length = dax_map_atomic(bdev, &dax); 1403642261acSRoss Zwisler void *ret; 1404642261acSRoss Zwisler 1405642261acSRoss Zwisler if (length < 0) /* dax_map_atomic() failed */ 1406642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1407642261acSRoss Zwisler if (length < PMD_SIZE) 1408642261acSRoss Zwisler goto unmap_fallback; 1409642261acSRoss Zwisler if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) 1410642261acSRoss Zwisler goto unmap_fallback; 1411642261acSRoss Zwisler if (!pfn_t_devmap(dax.pfn)) 1412642261acSRoss Zwisler goto unmap_fallback; 1413642261acSRoss Zwisler 1414642261acSRoss Zwisler dax_unmap_atomic(bdev, &dax); 1415642261acSRoss Zwisler 1416642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector, 1417642261acSRoss Zwisler RADIX_DAX_PMD); 1418642261acSRoss Zwisler if (IS_ERR(ret)) 1419642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1420642261acSRoss Zwisler *entryp = ret; 1421642261acSRoss Zwisler 1422642261acSRoss Zwisler return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write); 1423642261acSRoss Zwisler 1424642261acSRoss Zwisler unmap_fallback: 1425642261acSRoss Zwisler dax_unmap_atomic(bdev, &dax); 1426642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1427642261acSRoss Zwisler } 1428642261acSRoss Zwisler 1429642261acSRoss Zwisler static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd, 1430642261acSRoss Zwisler struct vm_fault *vmf, unsigned long address, 1431642261acSRoss Zwisler struct iomap *iomap, void **entryp) 1432642261acSRoss Zwisler { 1433642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1434642261acSRoss Zwisler unsigned long pmd_addr = address & PMD_MASK; 1435642261acSRoss Zwisler struct page *zero_page; 1436642261acSRoss Zwisler spinlock_t *ptl; 1437642261acSRoss Zwisler pmd_t pmd_entry; 1438642261acSRoss Zwisler void *ret; 1439642261acSRoss Zwisler 1440642261acSRoss Zwisler zero_page = mm_get_huge_zero_page(vma->vm_mm); 1441642261acSRoss Zwisler 1442642261acSRoss Zwisler if (unlikely(!zero_page)) 1443642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1444642261acSRoss Zwisler 1445642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0, 1446642261acSRoss Zwisler RADIX_DAX_PMD | RADIX_DAX_HZP); 1447642261acSRoss Zwisler if (IS_ERR(ret)) 1448642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1449642261acSRoss Zwisler *entryp = ret; 1450642261acSRoss Zwisler 1451642261acSRoss Zwisler ptl = pmd_lock(vma->vm_mm, pmd); 1452642261acSRoss Zwisler if (!pmd_none(*pmd)) { 1453642261acSRoss Zwisler spin_unlock(ptl); 1454642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1455642261acSRoss Zwisler } 1456642261acSRoss Zwisler 1457642261acSRoss Zwisler pmd_entry = mk_pmd(zero_page, vma->vm_page_prot); 1458642261acSRoss Zwisler pmd_entry = pmd_mkhuge(pmd_entry); 1459642261acSRoss Zwisler set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry); 1460642261acSRoss Zwisler spin_unlock(ptl); 1461642261acSRoss Zwisler return VM_FAULT_NOPAGE; 1462642261acSRoss Zwisler } 1463642261acSRoss Zwisler 1464642261acSRoss Zwisler int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address, 1465642261acSRoss Zwisler pmd_t *pmd, unsigned int flags, struct iomap_ops *ops) 1466642261acSRoss Zwisler { 1467642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1468642261acSRoss Zwisler unsigned long pmd_addr = address & PMD_MASK; 1469642261acSRoss Zwisler bool write = flags & FAULT_FLAG_WRITE; 1470642261acSRoss Zwisler unsigned int iomap_flags = write ? IOMAP_WRITE : 0; 1471642261acSRoss Zwisler struct inode *inode = mapping->host; 1472642261acSRoss Zwisler int result = VM_FAULT_FALLBACK; 1473642261acSRoss Zwisler struct iomap iomap = { 0 }; 1474642261acSRoss Zwisler pgoff_t max_pgoff, pgoff; 1475642261acSRoss Zwisler struct vm_fault vmf; 1476642261acSRoss Zwisler void *entry; 1477642261acSRoss Zwisler loff_t pos; 1478642261acSRoss Zwisler int error; 1479642261acSRoss Zwisler 1480642261acSRoss Zwisler /* Fall back to PTEs if we're going to COW */ 1481642261acSRoss Zwisler if (write && !(vma->vm_flags & VM_SHARED)) 1482642261acSRoss Zwisler goto fallback; 1483642261acSRoss Zwisler 1484642261acSRoss Zwisler /* If the PMD would extend outside the VMA */ 1485642261acSRoss Zwisler if (pmd_addr < vma->vm_start) 1486642261acSRoss Zwisler goto fallback; 1487642261acSRoss Zwisler if ((pmd_addr + PMD_SIZE) > vma->vm_end) 1488642261acSRoss Zwisler goto fallback; 1489642261acSRoss Zwisler 1490642261acSRoss Zwisler /* 1491642261acSRoss Zwisler * Check whether offset isn't beyond end of file now. Caller is 1492642261acSRoss Zwisler * supposed to hold locks serializing us with truncate / punch hole so 1493642261acSRoss Zwisler * this is a reliable test. 1494642261acSRoss Zwisler */ 1495642261acSRoss Zwisler pgoff = linear_page_index(vma, pmd_addr); 1496642261acSRoss Zwisler max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; 1497642261acSRoss Zwisler 1498642261acSRoss Zwisler if (pgoff > max_pgoff) 1499642261acSRoss Zwisler return VM_FAULT_SIGBUS; 1500642261acSRoss Zwisler 1501642261acSRoss Zwisler /* If the PMD would extend beyond the file size */ 1502642261acSRoss Zwisler if ((pgoff | PG_PMD_COLOUR) > max_pgoff) 1503642261acSRoss Zwisler goto fallback; 1504642261acSRoss Zwisler 1505642261acSRoss Zwisler /* 1506642261acSRoss Zwisler * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX 1507642261acSRoss Zwisler * PMD or a HZP entry. If it can't (because a 4k page is already in 1508642261acSRoss Zwisler * the tree, for instance), it will return -EEXIST and we just fall 1509642261acSRoss Zwisler * back to 4k entries. 1510642261acSRoss Zwisler */ 1511642261acSRoss Zwisler entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); 1512642261acSRoss Zwisler if (IS_ERR(entry)) 1513642261acSRoss Zwisler goto fallback; 1514642261acSRoss Zwisler 1515642261acSRoss Zwisler /* 1516642261acSRoss Zwisler * Note that we don't use iomap_apply here. We aren't doing I/O, only 1517642261acSRoss Zwisler * setting up a mapping, so really we're using iomap_begin() as a way 1518642261acSRoss Zwisler * to look up our filesystem block. 1519642261acSRoss Zwisler */ 1520642261acSRoss Zwisler pos = (loff_t)pgoff << PAGE_SHIFT; 1521642261acSRoss Zwisler error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); 1522642261acSRoss Zwisler if (error) 1523642261acSRoss Zwisler goto unlock_entry; 1524642261acSRoss Zwisler if (iomap.offset + iomap.length < pos + PMD_SIZE) 1525642261acSRoss Zwisler goto finish_iomap; 1526642261acSRoss Zwisler 1527642261acSRoss Zwisler vmf.pgoff = pgoff; 1528642261acSRoss Zwisler vmf.flags = flags; 1529642261acSRoss Zwisler vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO; 1530642261acSRoss Zwisler 1531642261acSRoss Zwisler switch (iomap.type) { 1532642261acSRoss Zwisler case IOMAP_MAPPED: 1533642261acSRoss Zwisler result = dax_pmd_insert_mapping(vma, pmd, &vmf, address, 1534642261acSRoss Zwisler &iomap, pos, write, &entry); 1535642261acSRoss Zwisler break; 1536642261acSRoss Zwisler case IOMAP_UNWRITTEN: 1537642261acSRoss Zwisler case IOMAP_HOLE: 1538642261acSRoss Zwisler if (WARN_ON_ONCE(write)) 1539642261acSRoss Zwisler goto finish_iomap; 1540642261acSRoss Zwisler result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap, 1541642261acSRoss Zwisler &entry); 1542642261acSRoss Zwisler break; 1543642261acSRoss Zwisler default: 1544642261acSRoss Zwisler WARN_ON_ONCE(1); 1545642261acSRoss Zwisler break; 1546642261acSRoss Zwisler } 1547642261acSRoss Zwisler 1548642261acSRoss Zwisler finish_iomap: 1549642261acSRoss Zwisler if (ops->iomap_end) { 1550642261acSRoss Zwisler if (result == VM_FAULT_FALLBACK) { 1551642261acSRoss Zwisler ops->iomap_end(inode, pos, PMD_SIZE, 0, iomap_flags, 1552642261acSRoss Zwisler &iomap); 1553642261acSRoss Zwisler } else { 1554642261acSRoss Zwisler error = ops->iomap_end(inode, pos, PMD_SIZE, PMD_SIZE, 1555642261acSRoss Zwisler iomap_flags, &iomap); 1556642261acSRoss Zwisler if (error) 1557642261acSRoss Zwisler result = VM_FAULT_FALLBACK; 1558642261acSRoss Zwisler } 1559642261acSRoss Zwisler } 1560642261acSRoss Zwisler unlock_entry: 1561642261acSRoss Zwisler put_locked_mapping_entry(mapping, pgoff, entry); 1562642261acSRoss Zwisler fallback: 1563642261acSRoss Zwisler if (result == VM_FAULT_FALLBACK) { 1564642261acSRoss Zwisler split_huge_pmd(vma, pmd, address); 1565642261acSRoss Zwisler count_vm_event(THP_FAULT_FALLBACK); 1566642261acSRoss Zwisler } 1567642261acSRoss Zwisler return result; 1568642261acSRoss Zwisler } 1569642261acSRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault); 1570642261acSRoss Zwisler #endif /* CONFIG_FS_DAX_PMD */ 1571a254e568SChristoph Hellwig #endif /* CONFIG_FS_IOMAP */ 1572