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> 344b4bb46dSJan Kara #include <linux/mmu_notifier.h> 35a254e568SChristoph Hellwig #include <linux/iomap.h> 36a254e568SChristoph Hellwig #include "internal.h" 37d475c634SMatthew Wilcox 38ac401cc7SJan Kara /* We choose 4096 entries - same as per-zone page wait tables */ 39ac401cc7SJan Kara #define DAX_WAIT_TABLE_BITS 12 40ac401cc7SJan Kara #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) 41ac401cc7SJan Kara 42ce95ab0fSRoss Zwisler static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; 43ac401cc7SJan Kara 44ac401cc7SJan Kara static int __init init_dax_wait_table(void) 45ac401cc7SJan Kara { 46ac401cc7SJan Kara int i; 47ac401cc7SJan Kara 48ac401cc7SJan Kara for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) 49ac401cc7SJan Kara init_waitqueue_head(wait_table + i); 50ac401cc7SJan Kara return 0; 51ac401cc7SJan Kara } 52ac401cc7SJan Kara fs_initcall(init_dax_wait_table); 53ac401cc7SJan Kara 54b2e0d162SDan Williams static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax) 55b2e0d162SDan Williams { 56b2e0d162SDan Williams struct request_queue *q = bdev->bd_queue; 57b2e0d162SDan Williams long rc = -EIO; 58b2e0d162SDan Williams 597a9eb206SDan Williams dax->addr = ERR_PTR(-EIO); 60b2e0d162SDan Williams if (blk_queue_enter(q, true) != 0) 61b2e0d162SDan Williams return rc; 62b2e0d162SDan Williams 63b2e0d162SDan Williams rc = bdev_direct_access(bdev, dax); 64b2e0d162SDan Williams if (rc < 0) { 657a9eb206SDan Williams dax->addr = ERR_PTR(rc); 66b2e0d162SDan Williams blk_queue_exit(q); 67b2e0d162SDan Williams return rc; 68b2e0d162SDan Williams } 69b2e0d162SDan Williams return rc; 70b2e0d162SDan Williams } 71b2e0d162SDan Williams 72b2e0d162SDan Williams static void dax_unmap_atomic(struct block_device *bdev, 73b2e0d162SDan Williams const struct blk_dax_ctl *dax) 74b2e0d162SDan Williams { 75b2e0d162SDan Williams if (IS_ERR(dax->addr)) 76b2e0d162SDan Williams return; 77b2e0d162SDan Williams blk_queue_exit(bdev->bd_queue); 78b2e0d162SDan Williams } 79b2e0d162SDan Williams 80642261acSRoss Zwisler static int dax_is_pmd_entry(void *entry) 81642261acSRoss Zwisler { 82642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_PMD; 83642261acSRoss Zwisler } 84642261acSRoss Zwisler 85642261acSRoss Zwisler static int dax_is_pte_entry(void *entry) 86642261acSRoss Zwisler { 87642261acSRoss Zwisler return !((unsigned long)entry & RADIX_DAX_PMD); 88642261acSRoss Zwisler } 89642261acSRoss Zwisler 90642261acSRoss Zwisler static int dax_is_zero_entry(void *entry) 91642261acSRoss Zwisler { 92642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_HZP; 93642261acSRoss Zwisler } 94642261acSRoss Zwisler 95642261acSRoss Zwisler static int dax_is_empty_entry(void *entry) 96642261acSRoss Zwisler { 97642261acSRoss Zwisler return (unsigned long)entry & RADIX_DAX_EMPTY; 98642261acSRoss Zwisler } 99642261acSRoss Zwisler 100d1a5f2b4SDan Williams struct page *read_dax_sector(struct block_device *bdev, sector_t n) 101d1a5f2b4SDan Williams { 102d1a5f2b4SDan Williams struct page *page = alloc_pages(GFP_KERNEL, 0); 103d1a5f2b4SDan Williams struct blk_dax_ctl dax = { 104d1a5f2b4SDan Williams .size = PAGE_SIZE, 105d1a5f2b4SDan Williams .sector = n & ~((((int) PAGE_SIZE) / 512) - 1), 106d1a5f2b4SDan Williams }; 107d1a5f2b4SDan Williams long rc; 108d1a5f2b4SDan Williams 109d1a5f2b4SDan Williams if (!page) 110d1a5f2b4SDan Williams return ERR_PTR(-ENOMEM); 111d1a5f2b4SDan Williams 112d1a5f2b4SDan Williams rc = dax_map_atomic(bdev, &dax); 113d1a5f2b4SDan Williams if (rc < 0) 114d1a5f2b4SDan Williams return ERR_PTR(rc); 115d1a5f2b4SDan Williams memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE); 116d1a5f2b4SDan Williams dax_unmap_atomic(bdev, &dax); 117d1a5f2b4SDan Williams return page; 118d1a5f2b4SDan Williams } 119d1a5f2b4SDan Williams 120f7ca90b1SMatthew Wilcox /* 121ac401cc7SJan Kara * DAX radix tree locking 122ac401cc7SJan Kara */ 123ac401cc7SJan Kara struct exceptional_entry_key { 124ac401cc7SJan Kara struct address_space *mapping; 12563e95b5cSRoss Zwisler pgoff_t entry_start; 126ac401cc7SJan Kara }; 127ac401cc7SJan Kara 128ac401cc7SJan Kara struct wait_exceptional_entry_queue { 129ac401cc7SJan Kara wait_queue_t wait; 130ac401cc7SJan Kara struct exceptional_entry_key key; 131ac401cc7SJan Kara }; 132ac401cc7SJan Kara 13363e95b5cSRoss Zwisler static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping, 13463e95b5cSRoss Zwisler pgoff_t index, void *entry, struct exceptional_entry_key *key) 13563e95b5cSRoss Zwisler { 13663e95b5cSRoss Zwisler unsigned long hash; 13763e95b5cSRoss Zwisler 13863e95b5cSRoss Zwisler /* 13963e95b5cSRoss Zwisler * If 'entry' is a PMD, align the 'index' that we use for the wait 14063e95b5cSRoss Zwisler * queue to the start of that PMD. This ensures that all offsets in 14163e95b5cSRoss Zwisler * the range covered by the PMD map to the same bit lock. 14263e95b5cSRoss Zwisler */ 143642261acSRoss Zwisler if (dax_is_pmd_entry(entry)) 14463e95b5cSRoss Zwisler index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1); 14563e95b5cSRoss Zwisler 14663e95b5cSRoss Zwisler key->mapping = mapping; 14763e95b5cSRoss Zwisler key->entry_start = index; 14863e95b5cSRoss Zwisler 14963e95b5cSRoss Zwisler hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS); 15063e95b5cSRoss Zwisler return wait_table + hash; 15163e95b5cSRoss Zwisler } 15263e95b5cSRoss Zwisler 153ac401cc7SJan Kara static int wake_exceptional_entry_func(wait_queue_t *wait, unsigned int mode, 154ac401cc7SJan Kara int sync, void *keyp) 155ac401cc7SJan Kara { 156ac401cc7SJan Kara struct exceptional_entry_key *key = keyp; 157ac401cc7SJan Kara struct wait_exceptional_entry_queue *ewait = 158ac401cc7SJan Kara container_of(wait, struct wait_exceptional_entry_queue, wait); 159ac401cc7SJan Kara 160ac401cc7SJan Kara if (key->mapping != ewait->key.mapping || 16163e95b5cSRoss Zwisler key->entry_start != ewait->key.entry_start) 162ac401cc7SJan Kara return 0; 163ac401cc7SJan Kara return autoremove_wake_function(wait, mode, sync, NULL); 164ac401cc7SJan Kara } 165ac401cc7SJan Kara 166ac401cc7SJan Kara /* 167ac401cc7SJan Kara * Check whether the given slot is locked. The function must be called with 168ac401cc7SJan Kara * mapping->tree_lock held 169ac401cc7SJan Kara */ 170ac401cc7SJan Kara static inline int slot_locked(struct address_space *mapping, void **slot) 171ac401cc7SJan Kara { 172ac401cc7SJan Kara unsigned long entry = (unsigned long) 173ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 174ac401cc7SJan Kara return entry & RADIX_DAX_ENTRY_LOCK; 175ac401cc7SJan Kara } 176ac401cc7SJan Kara 177ac401cc7SJan Kara /* 178ac401cc7SJan Kara * Mark the given slot is locked. The function must be called with 179ac401cc7SJan Kara * mapping->tree_lock held 180ac401cc7SJan Kara */ 181ac401cc7SJan Kara static inline void *lock_slot(struct address_space *mapping, void **slot) 182ac401cc7SJan Kara { 183ac401cc7SJan Kara unsigned long entry = (unsigned long) 184ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 185ac401cc7SJan Kara 186ac401cc7SJan Kara entry |= RADIX_DAX_ENTRY_LOCK; 1876d75f366SJohannes Weiner radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); 188ac401cc7SJan Kara return (void *)entry; 189ac401cc7SJan Kara } 190ac401cc7SJan Kara 191ac401cc7SJan Kara /* 192ac401cc7SJan Kara * Mark the given slot is unlocked. The function must be called with 193ac401cc7SJan Kara * mapping->tree_lock held 194ac401cc7SJan Kara */ 195ac401cc7SJan Kara static inline void *unlock_slot(struct address_space *mapping, void **slot) 196ac401cc7SJan Kara { 197ac401cc7SJan Kara unsigned long entry = (unsigned long) 198ac401cc7SJan Kara radix_tree_deref_slot_protected(slot, &mapping->tree_lock); 199ac401cc7SJan Kara 200ac401cc7SJan Kara entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK; 2016d75f366SJohannes Weiner radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry); 202ac401cc7SJan Kara return (void *)entry; 203ac401cc7SJan Kara } 204ac401cc7SJan Kara 205ac401cc7SJan Kara /* 206ac401cc7SJan Kara * Lookup entry in radix tree, wait for it to become unlocked if it is 207ac401cc7SJan Kara * exceptional entry and return it. The caller must call 208ac401cc7SJan Kara * put_unlocked_mapping_entry() when he decided not to lock the entry or 209ac401cc7SJan Kara * put_locked_mapping_entry() when he locked the entry and now wants to 210ac401cc7SJan Kara * unlock it. 211ac401cc7SJan Kara * 212ac401cc7SJan Kara * The function must be called with mapping->tree_lock held. 213ac401cc7SJan Kara */ 214ac401cc7SJan Kara static void *get_unlocked_mapping_entry(struct address_space *mapping, 215ac401cc7SJan Kara pgoff_t index, void ***slotp) 216ac401cc7SJan Kara { 217e3ad61c6SRoss Zwisler void *entry, **slot; 218ac401cc7SJan Kara struct wait_exceptional_entry_queue ewait; 21963e95b5cSRoss Zwisler wait_queue_head_t *wq; 220ac401cc7SJan Kara 221ac401cc7SJan Kara init_wait(&ewait.wait); 222ac401cc7SJan Kara ewait.wait.func = wake_exceptional_entry_func; 223ac401cc7SJan Kara 224ac401cc7SJan Kara for (;;) { 225e3ad61c6SRoss Zwisler entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, 226ac401cc7SJan Kara &slot); 227e3ad61c6SRoss Zwisler if (!entry || !radix_tree_exceptional_entry(entry) || 228ac401cc7SJan Kara !slot_locked(mapping, slot)) { 229ac401cc7SJan Kara if (slotp) 230ac401cc7SJan Kara *slotp = slot; 231e3ad61c6SRoss Zwisler return entry; 232ac401cc7SJan Kara } 23363e95b5cSRoss Zwisler 23463e95b5cSRoss Zwisler wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key); 235ac401cc7SJan Kara prepare_to_wait_exclusive(wq, &ewait.wait, 236ac401cc7SJan Kara TASK_UNINTERRUPTIBLE); 237ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 238ac401cc7SJan Kara schedule(); 239ac401cc7SJan Kara finish_wait(wq, &ewait.wait); 240ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 241ac401cc7SJan Kara } 242ac401cc7SJan Kara } 243ac401cc7SJan Kara 244b1aa812bSJan Kara static void dax_unlock_mapping_entry(struct address_space *mapping, 245b1aa812bSJan Kara pgoff_t index) 246b1aa812bSJan Kara { 247b1aa812bSJan Kara void *entry, **slot; 248b1aa812bSJan Kara 249b1aa812bSJan Kara spin_lock_irq(&mapping->tree_lock); 250b1aa812bSJan Kara entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot); 251b1aa812bSJan Kara if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) || 252b1aa812bSJan Kara !slot_locked(mapping, slot))) { 253b1aa812bSJan Kara spin_unlock_irq(&mapping->tree_lock); 254b1aa812bSJan Kara return; 255b1aa812bSJan Kara } 256b1aa812bSJan Kara unlock_slot(mapping, slot); 257b1aa812bSJan Kara spin_unlock_irq(&mapping->tree_lock); 258b1aa812bSJan Kara dax_wake_mapping_entry_waiter(mapping, index, entry, false); 259b1aa812bSJan Kara } 260b1aa812bSJan Kara 261ac401cc7SJan Kara static void put_locked_mapping_entry(struct address_space *mapping, 262ac401cc7SJan Kara pgoff_t index, void *entry) 263ac401cc7SJan Kara { 264ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) { 265ac401cc7SJan Kara unlock_page(entry); 266ac401cc7SJan Kara put_page(entry); 267ac401cc7SJan Kara } else { 268bc2466e4SJan Kara dax_unlock_mapping_entry(mapping, index); 269ac401cc7SJan Kara } 270ac401cc7SJan Kara } 271ac401cc7SJan Kara 272ac401cc7SJan Kara /* 273ac401cc7SJan Kara * Called when we are done with radix tree entry we looked up via 274ac401cc7SJan Kara * get_unlocked_mapping_entry() and which we didn't lock in the end. 275ac401cc7SJan Kara */ 276ac401cc7SJan Kara static void put_unlocked_mapping_entry(struct address_space *mapping, 277ac401cc7SJan Kara pgoff_t index, void *entry) 278ac401cc7SJan Kara { 279ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) 280ac401cc7SJan Kara return; 281ac401cc7SJan Kara 282ac401cc7SJan Kara /* We have to wake up next waiter for the radix tree entry lock */ 283422476c4SRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, false); 284422476c4SRoss Zwisler } 285422476c4SRoss Zwisler 286ac401cc7SJan Kara /* 287ac401cc7SJan Kara * Find radix tree entry at given index. If it points to a page, return with 288ac401cc7SJan Kara * the page locked. If it points to the exceptional entry, return with the 289ac401cc7SJan Kara * radix tree entry locked. If the radix tree doesn't contain given index, 290ac401cc7SJan Kara * create empty exceptional entry for the index and return with it locked. 291ac401cc7SJan Kara * 292642261acSRoss Zwisler * When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will 293642261acSRoss Zwisler * either return that locked entry or will return an error. This error will 294642261acSRoss Zwisler * happen if there are any 4k entries (either zero pages or DAX entries) 295642261acSRoss Zwisler * within the 2MiB range that we are requesting. 296642261acSRoss Zwisler * 297642261acSRoss Zwisler * We always favor 4k entries over 2MiB entries. There isn't a flow where we 298642261acSRoss Zwisler * evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB 299642261acSRoss Zwisler * insertion will fail if it finds any 4k entries already in the tree, and a 300642261acSRoss Zwisler * 4k insertion will cause an existing 2MiB entry to be unmapped and 301642261acSRoss Zwisler * downgraded to 4k entries. This happens for both 2MiB huge zero pages as 302642261acSRoss Zwisler * well as 2MiB empty entries. 303642261acSRoss Zwisler * 304642261acSRoss Zwisler * The exception to this downgrade path is for 2MiB DAX PMD entries that have 305642261acSRoss Zwisler * real storage backing them. We will leave these real 2MiB DAX entries in 306642261acSRoss Zwisler * the tree, and PTE writes will simply dirty the entire 2MiB DAX entry. 307642261acSRoss Zwisler * 308ac401cc7SJan Kara * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 309ac401cc7SJan Kara * persistent memory the benefit is doubtful. We can add that later if we can 310ac401cc7SJan Kara * show it helps. 311ac401cc7SJan Kara */ 312642261acSRoss Zwisler static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index, 313642261acSRoss Zwisler unsigned long size_flag) 314ac401cc7SJan Kara { 315642261acSRoss Zwisler bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */ 316e3ad61c6SRoss Zwisler void *entry, **slot; 317ac401cc7SJan Kara 318ac401cc7SJan Kara restart: 319ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 320e3ad61c6SRoss Zwisler entry = get_unlocked_mapping_entry(mapping, index, &slot); 321642261acSRoss Zwisler 322642261acSRoss Zwisler if (entry) { 323642261acSRoss Zwisler if (size_flag & RADIX_DAX_PMD) { 324642261acSRoss Zwisler if (!radix_tree_exceptional_entry(entry) || 325642261acSRoss Zwisler dax_is_pte_entry(entry)) { 326642261acSRoss Zwisler put_unlocked_mapping_entry(mapping, index, 327642261acSRoss Zwisler entry); 328642261acSRoss Zwisler entry = ERR_PTR(-EEXIST); 329642261acSRoss Zwisler goto out_unlock; 330642261acSRoss Zwisler } 331642261acSRoss Zwisler } else { /* trying to grab a PTE entry */ 332642261acSRoss Zwisler if (radix_tree_exceptional_entry(entry) && 333642261acSRoss Zwisler dax_is_pmd_entry(entry) && 334642261acSRoss Zwisler (dax_is_zero_entry(entry) || 335642261acSRoss Zwisler dax_is_empty_entry(entry))) { 336642261acSRoss Zwisler pmd_downgrade = true; 337642261acSRoss Zwisler } 338642261acSRoss Zwisler } 339642261acSRoss Zwisler } 340642261acSRoss Zwisler 341ac401cc7SJan Kara /* No entry for given index? Make sure radix tree is big enough. */ 342642261acSRoss Zwisler if (!entry || pmd_downgrade) { 343ac401cc7SJan Kara int err; 344ac401cc7SJan Kara 345642261acSRoss Zwisler if (pmd_downgrade) { 346642261acSRoss Zwisler /* 347642261acSRoss Zwisler * Make sure 'entry' remains valid while we drop 348642261acSRoss Zwisler * mapping->tree_lock. 349642261acSRoss Zwisler */ 350642261acSRoss Zwisler entry = lock_slot(mapping, slot); 351642261acSRoss Zwisler } 352642261acSRoss Zwisler 353ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 354642261acSRoss Zwisler /* 355642261acSRoss Zwisler * Besides huge zero pages the only other thing that gets 356642261acSRoss Zwisler * downgraded are empty entries which don't need to be 357642261acSRoss Zwisler * unmapped. 358642261acSRoss Zwisler */ 359642261acSRoss Zwisler if (pmd_downgrade && dax_is_zero_entry(entry)) 360642261acSRoss Zwisler unmap_mapping_range(mapping, 361642261acSRoss Zwisler (index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); 362642261acSRoss Zwisler 3630cb80b48SJan Kara err = radix_tree_preload( 3640cb80b48SJan Kara mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM); 3650cb80b48SJan Kara if (err) { 3660cb80b48SJan Kara if (pmd_downgrade) 3670cb80b48SJan Kara put_locked_mapping_entry(mapping, index, entry); 3680cb80b48SJan Kara return ERR_PTR(err); 3690cb80b48SJan Kara } 370ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 371642261acSRoss Zwisler 372642261acSRoss Zwisler if (pmd_downgrade) { 373642261acSRoss Zwisler radix_tree_delete(&mapping->page_tree, index); 374642261acSRoss Zwisler mapping->nrexceptional--; 375642261acSRoss Zwisler dax_wake_mapping_entry_waiter(mapping, index, entry, 376642261acSRoss Zwisler true); 377642261acSRoss Zwisler } 378642261acSRoss Zwisler 379642261acSRoss Zwisler entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY); 380642261acSRoss Zwisler 381642261acSRoss Zwisler err = __radix_tree_insert(&mapping->page_tree, index, 382642261acSRoss Zwisler dax_radix_order(entry), entry); 383ac401cc7SJan Kara radix_tree_preload_end(); 384ac401cc7SJan Kara if (err) { 385ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 386642261acSRoss Zwisler /* 387642261acSRoss Zwisler * Someone already created the entry? This is a 388642261acSRoss Zwisler * normal failure when inserting PMDs in a range 389642261acSRoss Zwisler * that already contains PTEs. In that case we want 390642261acSRoss Zwisler * to return -EEXIST immediately. 391642261acSRoss Zwisler */ 392642261acSRoss Zwisler if (err == -EEXIST && !(size_flag & RADIX_DAX_PMD)) 393ac401cc7SJan Kara goto restart; 394642261acSRoss Zwisler /* 395642261acSRoss Zwisler * Our insertion of a DAX PMD entry failed, most 396642261acSRoss Zwisler * likely because it collided with a PTE sized entry 397642261acSRoss Zwisler * at a different index in the PMD range. We haven't 398642261acSRoss Zwisler * inserted anything into the radix tree and have no 399642261acSRoss Zwisler * waiters to wake. 400642261acSRoss Zwisler */ 401ac401cc7SJan Kara return ERR_PTR(err); 402ac401cc7SJan Kara } 403ac401cc7SJan Kara /* Good, we have inserted empty locked entry into the tree. */ 404ac401cc7SJan Kara mapping->nrexceptional++; 405ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 406e3ad61c6SRoss Zwisler return entry; 407ac401cc7SJan Kara } 408ac401cc7SJan Kara /* Normal page in radix tree? */ 409e3ad61c6SRoss Zwisler if (!radix_tree_exceptional_entry(entry)) { 410e3ad61c6SRoss Zwisler struct page *page = entry; 411ac401cc7SJan Kara 412ac401cc7SJan Kara get_page(page); 413ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 414ac401cc7SJan Kara lock_page(page); 415ac401cc7SJan Kara /* Page got truncated? Retry... */ 416ac401cc7SJan Kara if (unlikely(page->mapping != mapping)) { 417ac401cc7SJan Kara unlock_page(page); 418ac401cc7SJan Kara put_page(page); 419ac401cc7SJan Kara goto restart; 420ac401cc7SJan Kara } 421ac401cc7SJan Kara return page; 422ac401cc7SJan Kara } 423e3ad61c6SRoss Zwisler entry = lock_slot(mapping, slot); 424642261acSRoss Zwisler out_unlock: 425ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 426e3ad61c6SRoss Zwisler return entry; 427ac401cc7SJan Kara } 428ac401cc7SJan Kara 42963e95b5cSRoss Zwisler /* 43063e95b5cSRoss Zwisler * We do not necessarily hold the mapping->tree_lock when we call this 43163e95b5cSRoss Zwisler * function so it is possible that 'entry' is no longer a valid item in the 432642261acSRoss Zwisler * radix tree. This is okay because all we really need to do is to find the 433642261acSRoss Zwisler * correct waitqueue where tasks might be waiting for that old 'entry' and 434642261acSRoss Zwisler * wake them. 43563e95b5cSRoss Zwisler */ 436ac401cc7SJan Kara void dax_wake_mapping_entry_waiter(struct address_space *mapping, 43763e95b5cSRoss Zwisler pgoff_t index, void *entry, bool wake_all) 438ac401cc7SJan Kara { 43963e95b5cSRoss Zwisler struct exceptional_entry_key key; 44063e95b5cSRoss Zwisler wait_queue_head_t *wq; 44163e95b5cSRoss Zwisler 44263e95b5cSRoss Zwisler wq = dax_entry_waitqueue(mapping, index, entry, &key); 443ac401cc7SJan Kara 444ac401cc7SJan Kara /* 445ac401cc7SJan Kara * Checking for locked entry and prepare_to_wait_exclusive() happens 446ac401cc7SJan Kara * under mapping->tree_lock, ditto for entry handling in our callers. 447ac401cc7SJan Kara * So at this point all tasks that could have seen our entry locked 448ac401cc7SJan Kara * must be in the waitqueue and the following check will see them. 449ac401cc7SJan Kara */ 45063e95b5cSRoss Zwisler if (waitqueue_active(wq)) 451ac401cc7SJan Kara __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); 452ac401cc7SJan Kara } 453ac401cc7SJan Kara 454c6dcf52cSJan Kara static int __dax_invalidate_mapping_entry(struct address_space *mapping, 455c6dcf52cSJan Kara pgoff_t index, bool trunc) 456c6dcf52cSJan Kara { 457c6dcf52cSJan Kara int ret = 0; 458c6dcf52cSJan Kara void *entry; 459c6dcf52cSJan Kara struct radix_tree_root *page_tree = &mapping->page_tree; 460c6dcf52cSJan Kara 461c6dcf52cSJan Kara spin_lock_irq(&mapping->tree_lock); 462c6dcf52cSJan Kara entry = get_unlocked_mapping_entry(mapping, index, NULL); 463c6dcf52cSJan Kara if (!entry || !radix_tree_exceptional_entry(entry)) 464c6dcf52cSJan Kara goto out; 465c6dcf52cSJan Kara if (!trunc && 466c6dcf52cSJan Kara (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || 467c6dcf52cSJan Kara radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))) 468c6dcf52cSJan Kara goto out; 469c6dcf52cSJan Kara radix_tree_delete(page_tree, index); 470c6dcf52cSJan Kara mapping->nrexceptional--; 471c6dcf52cSJan Kara ret = 1; 472c6dcf52cSJan Kara out: 473c6dcf52cSJan Kara put_unlocked_mapping_entry(mapping, index, entry); 474c6dcf52cSJan Kara spin_unlock_irq(&mapping->tree_lock); 475c6dcf52cSJan Kara return ret; 476c6dcf52cSJan Kara } 477ac401cc7SJan Kara /* 478ac401cc7SJan Kara * Delete exceptional DAX entry at @index from @mapping. Wait for radix tree 479ac401cc7SJan Kara * entry to get unlocked before deleting it. 480ac401cc7SJan Kara */ 481ac401cc7SJan Kara int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 482ac401cc7SJan Kara { 483c6dcf52cSJan Kara int ret = __dax_invalidate_mapping_entry(mapping, index, true); 484ac401cc7SJan Kara 485ac401cc7SJan Kara /* 486ac401cc7SJan Kara * This gets called from truncate / punch_hole path. As such, the caller 487ac401cc7SJan Kara * must hold locks protecting against concurrent modifications of the 488ac401cc7SJan Kara * radix tree (usually fs-private i_mmap_sem for writing). Since the 489ac401cc7SJan Kara * caller has seen exceptional entry for this index, we better find it 490ac401cc7SJan Kara * at that index as well... 491ac401cc7SJan Kara */ 492c6dcf52cSJan Kara WARN_ON_ONCE(!ret); 493c6dcf52cSJan Kara return ret; 494ac401cc7SJan Kara } 495ac401cc7SJan Kara 496c6dcf52cSJan Kara /* 497c6dcf52cSJan Kara * Invalidate exceptional DAX entry if easily possible. This handles DAX 498c6dcf52cSJan Kara * entries for invalidate_inode_pages() so we evict the entry only if we can 499c6dcf52cSJan Kara * do so without blocking. 500c6dcf52cSJan Kara */ 501c6dcf52cSJan Kara int dax_invalidate_mapping_entry(struct address_space *mapping, pgoff_t index) 502c6dcf52cSJan Kara { 503c6dcf52cSJan Kara int ret = 0; 504c6dcf52cSJan Kara void *entry, **slot; 505c6dcf52cSJan Kara struct radix_tree_root *page_tree = &mapping->page_tree; 506c6dcf52cSJan Kara 507c6dcf52cSJan Kara spin_lock_irq(&mapping->tree_lock); 508c6dcf52cSJan Kara entry = __radix_tree_lookup(page_tree, index, NULL, &slot); 509c6dcf52cSJan Kara if (!entry || !radix_tree_exceptional_entry(entry) || 510c6dcf52cSJan Kara slot_locked(mapping, slot)) 511c6dcf52cSJan Kara goto out; 512c6dcf52cSJan Kara if (radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) || 513c6dcf52cSJan Kara radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) 514c6dcf52cSJan Kara goto out; 515c6dcf52cSJan Kara radix_tree_delete(page_tree, index); 516c6dcf52cSJan Kara mapping->nrexceptional--; 517c6dcf52cSJan Kara ret = 1; 518c6dcf52cSJan Kara out: 519c6dcf52cSJan Kara spin_unlock_irq(&mapping->tree_lock); 520c6dcf52cSJan Kara if (ret) 521c6dcf52cSJan Kara dax_wake_mapping_entry_waiter(mapping, index, entry, true); 522c6dcf52cSJan Kara return ret; 523c6dcf52cSJan Kara } 524c6dcf52cSJan Kara 525c6dcf52cSJan Kara /* 526c6dcf52cSJan Kara * Invalidate exceptional DAX entry if it is clean. 527c6dcf52cSJan Kara */ 528c6dcf52cSJan Kara int dax_invalidate_mapping_entry_sync(struct address_space *mapping, 529c6dcf52cSJan Kara pgoff_t index) 530c6dcf52cSJan Kara { 531c6dcf52cSJan Kara return __dax_invalidate_mapping_entry(mapping, index, false); 532ac401cc7SJan Kara } 533ac401cc7SJan Kara 534ac401cc7SJan Kara /* 535f7ca90b1SMatthew Wilcox * The user has performed a load from a hole in the file. Allocating 536f7ca90b1SMatthew Wilcox * a new page in the file would cause excessive storage usage for 537f7ca90b1SMatthew Wilcox * workloads with sparse files. We allocate a page cache page instead. 538f7ca90b1SMatthew Wilcox * We'll kick it out of the page cache if it's ever written to, 539f7ca90b1SMatthew Wilcox * otherwise it will simply fall out of the page cache under memory 540f7ca90b1SMatthew Wilcox * pressure without ever having been dirtied. 541f7ca90b1SMatthew Wilcox */ 542f449b936SJan Kara static int dax_load_hole(struct address_space *mapping, void **entry, 543f7ca90b1SMatthew Wilcox struct vm_fault *vmf) 544f7ca90b1SMatthew Wilcox { 545ac401cc7SJan Kara struct page *page; 546f449b936SJan Kara int ret; 547f7ca90b1SMatthew Wilcox 548ac401cc7SJan Kara /* Hole page already exists? Return it... */ 549f449b936SJan Kara if (!radix_tree_exceptional_entry(*entry)) { 550f449b936SJan Kara page = *entry; 551f449b936SJan Kara goto out; 552ac401cc7SJan Kara } 553ac401cc7SJan Kara 554ac401cc7SJan Kara /* This will replace locked radix tree entry with a hole page */ 555ac401cc7SJan Kara page = find_or_create_page(mapping, vmf->pgoff, 556ac401cc7SJan Kara vmf->gfp_mask | __GFP_ZERO); 557b1aa812bSJan Kara if (!page) 558ac401cc7SJan Kara return VM_FAULT_OOM; 559f449b936SJan Kara out: 560f7ca90b1SMatthew Wilcox vmf->page = page; 561f449b936SJan Kara ret = finish_fault(vmf); 562f449b936SJan Kara vmf->page = NULL; 563f449b936SJan Kara *entry = page; 564f449b936SJan Kara if (!ret) { 565f449b936SJan Kara /* Grab reference for PTE that is now referencing the page */ 566f449b936SJan Kara get_page(page); 567f449b936SJan Kara return VM_FAULT_NOPAGE; 568f449b936SJan Kara } 569f449b936SJan Kara return ret; 570f7ca90b1SMatthew Wilcox } 571f7ca90b1SMatthew Wilcox 572b0d5e82fSChristoph Hellwig static int copy_user_dax(struct block_device *bdev, sector_t sector, size_t size, 573b0d5e82fSChristoph Hellwig struct page *to, unsigned long vaddr) 574f7ca90b1SMatthew Wilcox { 575b2e0d162SDan Williams struct blk_dax_ctl dax = { 576b0d5e82fSChristoph Hellwig .sector = sector, 577b0d5e82fSChristoph Hellwig .size = size, 578b2e0d162SDan Williams }; 579e2e05394SRoss Zwisler void *vto; 580e2e05394SRoss Zwisler 581b2e0d162SDan Williams if (dax_map_atomic(bdev, &dax) < 0) 582b2e0d162SDan Williams return PTR_ERR(dax.addr); 583f7ca90b1SMatthew Wilcox vto = kmap_atomic(to); 584b2e0d162SDan Williams copy_user_page(vto, (void __force *)dax.addr, vaddr, to); 585f7ca90b1SMatthew Wilcox kunmap_atomic(vto); 586b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 587f7ca90b1SMatthew Wilcox return 0; 588f7ca90b1SMatthew Wilcox } 589f7ca90b1SMatthew Wilcox 590642261acSRoss Zwisler /* 591642261acSRoss Zwisler * By this point grab_mapping_entry() has ensured that we have a locked entry 592642261acSRoss Zwisler * of the appropriate size so we don't have to worry about downgrading PMDs to 593642261acSRoss Zwisler * PTEs. If we happen to be trying to insert a PTE and there is a PMD 594642261acSRoss Zwisler * already in the tree, we will skip the insertion and just dirty the PMD as 595642261acSRoss Zwisler * appropriate. 596642261acSRoss Zwisler */ 597ac401cc7SJan Kara static void *dax_insert_mapping_entry(struct address_space *mapping, 598ac401cc7SJan Kara struct vm_fault *vmf, 599642261acSRoss Zwisler void *entry, sector_t sector, 600642261acSRoss Zwisler unsigned long flags) 6019973c98eSRoss Zwisler { 6029973c98eSRoss Zwisler struct radix_tree_root *page_tree = &mapping->page_tree; 603ac401cc7SJan Kara int error = 0; 604ac401cc7SJan Kara bool hole_fill = false; 605ac401cc7SJan Kara void *new_entry; 606ac401cc7SJan Kara pgoff_t index = vmf->pgoff; 6079973c98eSRoss Zwisler 608ac401cc7SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) 6099973c98eSRoss Zwisler __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 6109973c98eSRoss Zwisler 611ac401cc7SJan Kara /* Replacing hole page with block mapping? */ 612ac401cc7SJan Kara if (!radix_tree_exceptional_entry(entry)) { 613ac401cc7SJan Kara hole_fill = true; 6149973c98eSRoss Zwisler /* 615ac401cc7SJan Kara * Unmap the page now before we remove it from page cache below. 616ac401cc7SJan Kara * The page is locked so it cannot be faulted in again. 6179973c98eSRoss Zwisler */ 618ac401cc7SJan Kara unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT, 619ac401cc7SJan Kara PAGE_SIZE, 0); 620ac401cc7SJan Kara error = radix_tree_preload(vmf->gfp_mask & ~__GFP_HIGHMEM); 6219973c98eSRoss Zwisler if (error) 622ac401cc7SJan Kara return ERR_PTR(error); 623642261acSRoss Zwisler } else if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_HZP)) { 624642261acSRoss Zwisler /* replacing huge zero page with PMD block mapping */ 625642261acSRoss Zwisler unmap_mapping_range(mapping, 626642261acSRoss Zwisler (vmf->pgoff << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0); 627ac401cc7SJan Kara } 6289973c98eSRoss Zwisler 629ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 630642261acSRoss Zwisler new_entry = dax_radix_locked_entry(sector, flags); 631642261acSRoss Zwisler 632ac401cc7SJan Kara if (hole_fill) { 633ac401cc7SJan Kara __delete_from_page_cache(entry, NULL); 634ac401cc7SJan Kara /* Drop pagecache reference */ 635ac401cc7SJan Kara put_page(entry); 636642261acSRoss Zwisler error = __radix_tree_insert(page_tree, index, 637642261acSRoss Zwisler dax_radix_order(new_entry), new_entry); 638ac401cc7SJan Kara if (error) { 639ac401cc7SJan Kara new_entry = ERR_PTR(error); 640ac401cc7SJan Kara goto unlock; 641ac401cc7SJan Kara } 6429973c98eSRoss Zwisler mapping->nrexceptional++; 643642261acSRoss Zwisler } else if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 644642261acSRoss Zwisler /* 645642261acSRoss Zwisler * Only swap our new entry into the radix tree if the current 646642261acSRoss Zwisler * entry is a zero page or an empty entry. If a normal PTE or 647642261acSRoss Zwisler * PMD entry is already in the tree, we leave it alone. This 648642261acSRoss Zwisler * means that if we are trying to insert a PTE and the 649642261acSRoss Zwisler * existing entry is a PMD, we will just leave the PMD in the 650642261acSRoss Zwisler * tree and dirty it if necessary. 651642261acSRoss Zwisler */ 652f7942430SJohannes Weiner struct radix_tree_node *node; 653ac401cc7SJan Kara void **slot; 654ac401cc7SJan Kara void *ret; 655ac401cc7SJan Kara 656f7942430SJohannes Weiner ret = __radix_tree_lookup(page_tree, index, &node, &slot); 657ac401cc7SJan Kara WARN_ON_ONCE(ret != entry); 6584d693d08SJohannes Weiner __radix_tree_replace(page_tree, node, slot, 6594d693d08SJohannes Weiner new_entry, NULL, NULL); 660ac401cc7SJan Kara } 661ac401cc7SJan Kara if (vmf->flags & FAULT_FLAG_WRITE) 6629973c98eSRoss Zwisler radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY); 6639973c98eSRoss Zwisler unlock: 6649973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 665ac401cc7SJan Kara if (hole_fill) { 666ac401cc7SJan Kara radix_tree_preload_end(); 667ac401cc7SJan Kara /* 668ac401cc7SJan Kara * We don't need hole page anymore, it has been replaced with 669ac401cc7SJan Kara * locked radix tree entry now. 670ac401cc7SJan Kara */ 671ac401cc7SJan Kara if (mapping->a_ops->freepage) 672ac401cc7SJan Kara mapping->a_ops->freepage(entry); 673ac401cc7SJan Kara unlock_page(entry); 674ac401cc7SJan Kara put_page(entry); 675ac401cc7SJan Kara } 676ac401cc7SJan Kara return new_entry; 6779973c98eSRoss Zwisler } 6789973c98eSRoss Zwisler 6794b4bb46dSJan Kara static inline unsigned long 6804b4bb46dSJan Kara pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) 6814b4bb46dSJan Kara { 6824b4bb46dSJan Kara unsigned long address; 6834b4bb46dSJan Kara 6844b4bb46dSJan Kara address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 6854b4bb46dSJan Kara VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 6864b4bb46dSJan Kara return address; 6874b4bb46dSJan Kara } 6884b4bb46dSJan Kara 6894b4bb46dSJan Kara /* Walk all mappings of a given index of a file and writeprotect them */ 6904b4bb46dSJan Kara static void dax_mapping_entry_mkclean(struct address_space *mapping, 6914b4bb46dSJan Kara pgoff_t index, unsigned long pfn) 6924b4bb46dSJan Kara { 6934b4bb46dSJan Kara struct vm_area_struct *vma; 694f729c8c9SRoss Zwisler pte_t pte, *ptep = NULL; 695f729c8c9SRoss Zwisler pmd_t *pmdp = NULL; 6964b4bb46dSJan Kara spinlock_t *ptl; 6974b4bb46dSJan Kara bool changed; 6984b4bb46dSJan Kara 6994b4bb46dSJan Kara i_mmap_lock_read(mapping); 7004b4bb46dSJan Kara vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { 7014b4bb46dSJan Kara unsigned long address; 7024b4bb46dSJan Kara 7034b4bb46dSJan Kara cond_resched(); 7044b4bb46dSJan Kara 7054b4bb46dSJan Kara if (!(vma->vm_flags & VM_SHARED)) 7064b4bb46dSJan Kara continue; 7074b4bb46dSJan Kara 7084b4bb46dSJan Kara address = pgoff_address(index, vma); 7094b4bb46dSJan Kara changed = false; 710f729c8c9SRoss Zwisler if (follow_pte_pmd(vma->vm_mm, address, &ptep, &pmdp, &ptl)) 7114b4bb46dSJan Kara continue; 712f729c8c9SRoss Zwisler 713f729c8c9SRoss Zwisler if (pmdp) { 714f729c8c9SRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 715f729c8c9SRoss Zwisler pmd_t pmd; 716f729c8c9SRoss Zwisler 717f729c8c9SRoss Zwisler if (pfn != pmd_pfn(*pmdp)) 718f729c8c9SRoss Zwisler goto unlock_pmd; 719f729c8c9SRoss Zwisler if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) 720f729c8c9SRoss Zwisler goto unlock_pmd; 721f729c8c9SRoss Zwisler 722f729c8c9SRoss Zwisler flush_cache_page(vma, address, pfn); 723f729c8c9SRoss Zwisler pmd = pmdp_huge_clear_flush(vma, address, pmdp); 724f729c8c9SRoss Zwisler pmd = pmd_wrprotect(pmd); 725f729c8c9SRoss Zwisler pmd = pmd_mkclean(pmd); 726f729c8c9SRoss Zwisler set_pmd_at(vma->vm_mm, address, pmdp, pmd); 727f729c8c9SRoss Zwisler changed = true; 728f729c8c9SRoss Zwisler unlock_pmd: 729f729c8c9SRoss Zwisler spin_unlock(ptl); 730f729c8c9SRoss Zwisler #endif 731f729c8c9SRoss Zwisler } else { 7324b4bb46dSJan Kara if (pfn != pte_pfn(*ptep)) 733f729c8c9SRoss Zwisler goto unlock_pte; 7344b4bb46dSJan Kara if (!pte_dirty(*ptep) && !pte_write(*ptep)) 735f729c8c9SRoss Zwisler goto unlock_pte; 7364b4bb46dSJan Kara 7374b4bb46dSJan Kara flush_cache_page(vma, address, pfn); 7384b4bb46dSJan Kara pte = ptep_clear_flush(vma, address, ptep); 7394b4bb46dSJan Kara pte = pte_wrprotect(pte); 7404b4bb46dSJan Kara pte = pte_mkclean(pte); 7414b4bb46dSJan Kara set_pte_at(vma->vm_mm, address, ptep, pte); 7424b4bb46dSJan Kara changed = true; 743f729c8c9SRoss Zwisler unlock_pte: 7444b4bb46dSJan Kara pte_unmap_unlock(ptep, ptl); 745f729c8c9SRoss Zwisler } 7464b4bb46dSJan Kara 7474b4bb46dSJan Kara if (changed) 7484b4bb46dSJan Kara mmu_notifier_invalidate_page(vma->vm_mm, address); 7494b4bb46dSJan Kara } 7504b4bb46dSJan Kara i_mmap_unlock_read(mapping); 7514b4bb46dSJan Kara } 7524b4bb46dSJan Kara 7539973c98eSRoss Zwisler static int dax_writeback_one(struct block_device *bdev, 7549973c98eSRoss Zwisler struct address_space *mapping, pgoff_t index, void *entry) 7559973c98eSRoss Zwisler { 7569973c98eSRoss Zwisler struct radix_tree_root *page_tree = &mapping->page_tree; 7579973c98eSRoss Zwisler struct blk_dax_ctl dax; 758a6abc2c0SJan Kara void *entry2, **slot; 7599973c98eSRoss Zwisler int ret = 0; 7609973c98eSRoss Zwisler 7619973c98eSRoss Zwisler /* 762a6abc2c0SJan Kara * A page got tagged dirty in DAX mapping? Something is seriously 763a6abc2c0SJan Kara * wrong. 7649973c98eSRoss Zwisler */ 765a6abc2c0SJan Kara if (WARN_ON(!radix_tree_exceptional_entry(entry))) 766a6abc2c0SJan Kara return -EIO; 7679973c98eSRoss Zwisler 768a6abc2c0SJan Kara spin_lock_irq(&mapping->tree_lock); 769a6abc2c0SJan Kara entry2 = get_unlocked_mapping_entry(mapping, index, &slot); 770a6abc2c0SJan Kara /* Entry got punched out / reallocated? */ 771a6abc2c0SJan Kara if (!entry2 || !radix_tree_exceptional_entry(entry2)) 772a6abc2c0SJan Kara goto put_unlocked; 773a6abc2c0SJan Kara /* 774a6abc2c0SJan Kara * Entry got reallocated elsewhere? No need to writeback. We have to 775a6abc2c0SJan Kara * compare sectors as we must not bail out due to difference in lockbit 776a6abc2c0SJan Kara * or entry type. 777a6abc2c0SJan Kara */ 778a6abc2c0SJan Kara if (dax_radix_sector(entry2) != dax_radix_sector(entry)) 779a6abc2c0SJan Kara goto put_unlocked; 780642261acSRoss Zwisler if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 781642261acSRoss Zwisler dax_is_zero_entry(entry))) { 7829973c98eSRoss Zwisler ret = -EIO; 783a6abc2c0SJan Kara goto put_unlocked; 7849973c98eSRoss Zwisler } 7859973c98eSRoss Zwisler 786a6abc2c0SJan Kara /* Another fsync thread may have already written back this entry */ 787a6abc2c0SJan Kara if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)) 788a6abc2c0SJan Kara goto put_unlocked; 789a6abc2c0SJan Kara /* Lock the entry to serialize with page faults */ 790a6abc2c0SJan Kara entry = lock_slot(mapping, slot); 791a6abc2c0SJan Kara /* 792a6abc2c0SJan Kara * We can clear the tag now but we have to be careful so that concurrent 793a6abc2c0SJan Kara * dax_writeback_one() calls for the same index cannot finish before we 794a6abc2c0SJan Kara * actually flush the caches. This is achieved as the calls will look 795a6abc2c0SJan Kara * at the entry only under tree_lock and once they do that they will 796a6abc2c0SJan Kara * see the entry locked and wait for it to unlock. 797a6abc2c0SJan Kara */ 798a6abc2c0SJan Kara radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE); 799a6abc2c0SJan Kara spin_unlock_irq(&mapping->tree_lock); 800a6abc2c0SJan Kara 801642261acSRoss Zwisler /* 802642261acSRoss Zwisler * Even if dax_writeback_mapping_range() was given a wbc->range_start 803642261acSRoss Zwisler * in the middle of a PMD, the 'index' we are given will be aligned to 804642261acSRoss Zwisler * the start index of the PMD, as will the sector we pull from 805642261acSRoss Zwisler * 'entry'. This allows us to flush for PMD_SIZE and not have to 806642261acSRoss Zwisler * worry about partial PMD writebacks. 807642261acSRoss Zwisler */ 808642261acSRoss Zwisler dax.sector = dax_radix_sector(entry); 809642261acSRoss Zwisler dax.size = PAGE_SIZE << dax_radix_order(entry); 8109973c98eSRoss Zwisler 8119973c98eSRoss Zwisler /* 8129973c98eSRoss Zwisler * We cannot hold tree_lock while calling dax_map_atomic() because it 8139973c98eSRoss Zwisler * eventually calls cond_resched(). 8149973c98eSRoss Zwisler */ 8159973c98eSRoss Zwisler ret = dax_map_atomic(bdev, &dax); 816a6abc2c0SJan Kara if (ret < 0) { 817a6abc2c0SJan Kara put_locked_mapping_entry(mapping, index, entry); 8189973c98eSRoss Zwisler return ret; 819a6abc2c0SJan Kara } 8209973c98eSRoss Zwisler 8219973c98eSRoss Zwisler if (WARN_ON_ONCE(ret < dax.size)) { 8229973c98eSRoss Zwisler ret = -EIO; 8239973c98eSRoss Zwisler goto unmap; 8249973c98eSRoss Zwisler } 8259973c98eSRoss Zwisler 8264b4bb46dSJan Kara dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(dax.pfn)); 8279973c98eSRoss Zwisler wb_cache_pmem(dax.addr, dax.size); 8284b4bb46dSJan Kara /* 8294b4bb46dSJan Kara * After we have flushed the cache, we can clear the dirty tag. There 8304b4bb46dSJan Kara * cannot be new dirty data in the pfn after the flush has completed as 8314b4bb46dSJan Kara * the pfn mappings are writeprotected and fault waits for mapping 8324b4bb46dSJan Kara * entry lock. 8334b4bb46dSJan Kara */ 8344b4bb46dSJan Kara spin_lock_irq(&mapping->tree_lock); 8354b4bb46dSJan Kara radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY); 8364b4bb46dSJan Kara spin_unlock_irq(&mapping->tree_lock); 8379973c98eSRoss Zwisler unmap: 8389973c98eSRoss Zwisler dax_unmap_atomic(bdev, &dax); 839a6abc2c0SJan Kara put_locked_mapping_entry(mapping, index, entry); 8409973c98eSRoss Zwisler return ret; 8419973c98eSRoss Zwisler 842a6abc2c0SJan Kara put_unlocked: 843a6abc2c0SJan Kara put_unlocked_mapping_entry(mapping, index, entry2); 8449973c98eSRoss Zwisler spin_unlock_irq(&mapping->tree_lock); 8459973c98eSRoss Zwisler return ret; 8469973c98eSRoss Zwisler } 8479973c98eSRoss Zwisler 8489973c98eSRoss Zwisler /* 8499973c98eSRoss Zwisler * Flush the mapping to the persistent domain within the byte range of [start, 8509973c98eSRoss Zwisler * end]. This is required by data integrity operations to ensure file data is 8519973c98eSRoss Zwisler * on persistent storage prior to completion of the operation. 8529973c98eSRoss Zwisler */ 8537f6d5b52SRoss Zwisler int dax_writeback_mapping_range(struct address_space *mapping, 8547f6d5b52SRoss Zwisler struct block_device *bdev, struct writeback_control *wbc) 8559973c98eSRoss Zwisler { 8569973c98eSRoss Zwisler struct inode *inode = mapping->host; 857642261acSRoss Zwisler pgoff_t start_index, end_index; 8589973c98eSRoss Zwisler pgoff_t indices[PAGEVEC_SIZE]; 8599973c98eSRoss Zwisler struct pagevec pvec; 8609973c98eSRoss Zwisler bool done = false; 8619973c98eSRoss Zwisler int i, ret = 0; 8629973c98eSRoss Zwisler 8639973c98eSRoss Zwisler if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 8649973c98eSRoss Zwisler return -EIO; 8659973c98eSRoss Zwisler 8667f6d5b52SRoss Zwisler if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) 8677f6d5b52SRoss Zwisler return 0; 8687f6d5b52SRoss Zwisler 86909cbfeafSKirill A. Shutemov start_index = wbc->range_start >> PAGE_SHIFT; 87009cbfeafSKirill A. Shutemov end_index = wbc->range_end >> PAGE_SHIFT; 8719973c98eSRoss Zwisler 8729973c98eSRoss Zwisler tag_pages_for_writeback(mapping, start_index, end_index); 8739973c98eSRoss Zwisler 8749973c98eSRoss Zwisler pagevec_init(&pvec, 0); 8759973c98eSRoss Zwisler while (!done) { 8769973c98eSRoss Zwisler pvec.nr = find_get_entries_tag(mapping, start_index, 8779973c98eSRoss Zwisler PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE, 8789973c98eSRoss Zwisler pvec.pages, indices); 8799973c98eSRoss Zwisler 8809973c98eSRoss Zwisler if (pvec.nr == 0) 8819973c98eSRoss Zwisler break; 8829973c98eSRoss Zwisler 8839973c98eSRoss Zwisler for (i = 0; i < pvec.nr; i++) { 8849973c98eSRoss Zwisler if (indices[i] > end_index) { 8859973c98eSRoss Zwisler done = true; 8869973c98eSRoss Zwisler break; 8879973c98eSRoss Zwisler } 8889973c98eSRoss Zwisler 8899973c98eSRoss Zwisler ret = dax_writeback_one(bdev, mapping, indices[i], 8909973c98eSRoss Zwisler pvec.pages[i]); 8919973c98eSRoss Zwisler if (ret < 0) 8929973c98eSRoss Zwisler return ret; 8939973c98eSRoss Zwisler } 8949973c98eSRoss Zwisler } 8959973c98eSRoss Zwisler return 0; 8969973c98eSRoss Zwisler } 8979973c98eSRoss Zwisler EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 8989973c98eSRoss Zwisler 899ac401cc7SJan Kara static int dax_insert_mapping(struct address_space *mapping, 9001aaba095SChristoph Hellwig struct block_device *bdev, sector_t sector, size_t size, 9011aaba095SChristoph Hellwig void **entryp, struct vm_area_struct *vma, struct vm_fault *vmf) 902f7ca90b1SMatthew Wilcox { 9031a29d85eSJan Kara unsigned long vaddr = vmf->address; 904b2e0d162SDan Williams struct blk_dax_ctl dax = { 9051aaba095SChristoph Hellwig .sector = sector, 9061aaba095SChristoph Hellwig .size = size, 907b2e0d162SDan Williams }; 908ac401cc7SJan Kara void *ret; 909ac401cc7SJan Kara void *entry = *entryp; 910f7ca90b1SMatthew Wilcox 9114d9a2c87SJan Kara if (dax_map_atomic(bdev, &dax) < 0) 9124d9a2c87SJan Kara return PTR_ERR(dax.addr); 913b2e0d162SDan Williams dax_unmap_atomic(bdev, &dax); 914f7ca90b1SMatthew Wilcox 915642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, entry, dax.sector, 0); 9164d9a2c87SJan Kara if (IS_ERR(ret)) 9174d9a2c87SJan Kara return PTR_ERR(ret); 918ac401cc7SJan Kara *entryp = ret; 9199973c98eSRoss Zwisler 9204d9a2c87SJan Kara return vm_insert_mixed(vma, vaddr, dax.pfn); 921f7ca90b1SMatthew Wilcox } 922f7ca90b1SMatthew Wilcox 923ce5c5d55SDave Chinner /** 9240e3b210cSBoaz Harrosh * dax_pfn_mkwrite - handle first write to DAX page 9250e3b210cSBoaz Harrosh * @vma: The virtual memory area where the fault occurred 9260e3b210cSBoaz Harrosh * @vmf: The description of the fault 9270e3b210cSBoaz Harrosh */ 9280e3b210cSBoaz Harrosh int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf) 9290e3b210cSBoaz Harrosh { 9309973c98eSRoss Zwisler struct file *file = vma->vm_file; 931ac401cc7SJan Kara struct address_space *mapping = file->f_mapping; 9322f89dc12SJan Kara void *entry, **slot; 933ac401cc7SJan Kara pgoff_t index = vmf->pgoff; 9340e3b210cSBoaz Harrosh 935ac401cc7SJan Kara spin_lock_irq(&mapping->tree_lock); 9362f89dc12SJan Kara entry = get_unlocked_mapping_entry(mapping, index, &slot); 9372f89dc12SJan Kara if (!entry || !radix_tree_exceptional_entry(entry)) { 9382f89dc12SJan Kara if (entry) 939ac401cc7SJan Kara put_unlocked_mapping_entry(mapping, index, entry); 940ac401cc7SJan Kara spin_unlock_irq(&mapping->tree_lock); 9410e3b210cSBoaz Harrosh return VM_FAULT_NOPAGE; 9420e3b210cSBoaz Harrosh } 9432f89dc12SJan Kara radix_tree_tag_set(&mapping->page_tree, index, PAGECACHE_TAG_DIRTY); 9442f89dc12SJan Kara entry = lock_slot(mapping, slot); 9452f89dc12SJan Kara spin_unlock_irq(&mapping->tree_lock); 9462f89dc12SJan Kara /* 9472f89dc12SJan Kara * If we race with somebody updating the PTE and finish_mkwrite_fault() 9482f89dc12SJan Kara * fails, we don't care. We need to return VM_FAULT_NOPAGE and retry 9492f89dc12SJan Kara * the fault in either case. 9502f89dc12SJan Kara */ 9512f89dc12SJan Kara finish_mkwrite_fault(vmf); 9522f89dc12SJan Kara put_locked_mapping_entry(mapping, index, entry); 9532f89dc12SJan Kara return VM_FAULT_NOPAGE; 9542f89dc12SJan Kara } 9550e3b210cSBoaz Harrosh EXPORT_SYMBOL_GPL(dax_pfn_mkwrite); 9560e3b210cSBoaz Harrosh 9574b0228faSVishal Verma static bool dax_range_is_aligned(struct block_device *bdev, 9584b0228faSVishal Verma unsigned int offset, unsigned int length) 9594b0228faSVishal Verma { 9604b0228faSVishal Verma unsigned short sector_size = bdev_logical_block_size(bdev); 9614b0228faSVishal Verma 9624b0228faSVishal Verma if (!IS_ALIGNED(offset, sector_size)) 9634b0228faSVishal Verma return false; 9644b0228faSVishal Verma if (!IS_ALIGNED(length, sector_size)) 9654b0228faSVishal Verma return false; 9664b0228faSVishal Verma 9674b0228faSVishal Verma return true; 9684b0228faSVishal Verma } 9694b0228faSVishal Verma 970679c8bd3SChristoph Hellwig int __dax_zero_page_range(struct block_device *bdev, sector_t sector, 971679c8bd3SChristoph Hellwig unsigned int offset, unsigned int length) 972679c8bd3SChristoph Hellwig { 973679c8bd3SChristoph Hellwig struct blk_dax_ctl dax = { 974679c8bd3SChristoph Hellwig .sector = sector, 975679c8bd3SChristoph Hellwig .size = PAGE_SIZE, 976679c8bd3SChristoph Hellwig }; 977679c8bd3SChristoph Hellwig 9784b0228faSVishal Verma if (dax_range_is_aligned(bdev, offset, length)) { 9794b0228faSVishal Verma sector_t start_sector = dax.sector + (offset >> 9); 9804b0228faSVishal Verma 9814b0228faSVishal Verma return blkdev_issue_zeroout(bdev, start_sector, 9824b0228faSVishal Verma length >> 9, GFP_NOFS, true); 9834b0228faSVishal Verma } else { 984679c8bd3SChristoph Hellwig if (dax_map_atomic(bdev, &dax) < 0) 985679c8bd3SChristoph Hellwig return PTR_ERR(dax.addr); 986679c8bd3SChristoph Hellwig clear_pmem(dax.addr + offset, length); 987679c8bd3SChristoph Hellwig dax_unmap_atomic(bdev, &dax); 9884b0228faSVishal Verma } 989679c8bd3SChristoph Hellwig return 0; 990679c8bd3SChristoph Hellwig } 991679c8bd3SChristoph Hellwig EXPORT_SYMBOL_GPL(__dax_zero_page_range); 992679c8bd3SChristoph Hellwig 993333ccc97SRoss Zwisler static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) 994333ccc97SRoss Zwisler { 995333ccc97SRoss Zwisler return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9); 996333ccc97SRoss Zwisler } 997333ccc97SRoss Zwisler 998a254e568SChristoph Hellwig static loff_t 99911c59c92SRoss Zwisler dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, 1000a254e568SChristoph Hellwig struct iomap *iomap) 1001a254e568SChristoph Hellwig { 1002a254e568SChristoph Hellwig struct iov_iter *iter = data; 1003a254e568SChristoph Hellwig loff_t end = pos + length, done = 0; 1004a254e568SChristoph Hellwig ssize_t ret = 0; 1005a254e568SChristoph Hellwig 1006a254e568SChristoph Hellwig if (iov_iter_rw(iter) == READ) { 1007a254e568SChristoph Hellwig end = min(end, i_size_read(inode)); 1008a254e568SChristoph Hellwig if (pos >= end) 1009a254e568SChristoph Hellwig return 0; 1010a254e568SChristoph Hellwig 1011a254e568SChristoph Hellwig if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1012a254e568SChristoph Hellwig return iov_iter_zero(min(length, end - pos), iter); 1013a254e568SChristoph Hellwig } 1014a254e568SChristoph Hellwig 1015a254e568SChristoph Hellwig if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) 1016a254e568SChristoph Hellwig return -EIO; 1017a254e568SChristoph Hellwig 1018e3fce68cSJan Kara /* 1019e3fce68cSJan Kara * Write can allocate block for an area which has a hole page mapped 1020e3fce68cSJan Kara * into page tables. We have to tear down these mappings so that data 1021e3fce68cSJan Kara * written by write(2) is visible in mmap. 1022e3fce68cSJan Kara */ 1023e3fce68cSJan Kara if ((iomap->flags & IOMAP_F_NEW) && inode->i_mapping->nrpages) { 1024e3fce68cSJan Kara invalidate_inode_pages2_range(inode->i_mapping, 1025e3fce68cSJan Kara pos >> PAGE_SHIFT, 1026e3fce68cSJan Kara (end - 1) >> PAGE_SHIFT); 1027e3fce68cSJan Kara } 1028e3fce68cSJan Kara 1029a254e568SChristoph Hellwig while (pos < end) { 1030a254e568SChristoph Hellwig unsigned offset = pos & (PAGE_SIZE - 1); 1031a254e568SChristoph Hellwig struct blk_dax_ctl dax = { 0 }; 1032a254e568SChristoph Hellwig ssize_t map_len; 1033a254e568SChristoph Hellwig 1034333ccc97SRoss Zwisler dax.sector = dax_iomap_sector(iomap, pos); 1035a254e568SChristoph Hellwig dax.size = (length + offset + PAGE_SIZE - 1) & PAGE_MASK; 1036a254e568SChristoph Hellwig map_len = dax_map_atomic(iomap->bdev, &dax); 1037a254e568SChristoph Hellwig if (map_len < 0) { 1038a254e568SChristoph Hellwig ret = map_len; 1039a254e568SChristoph Hellwig break; 1040a254e568SChristoph Hellwig } 1041a254e568SChristoph Hellwig 1042a254e568SChristoph Hellwig dax.addr += offset; 1043a254e568SChristoph Hellwig map_len -= offset; 1044a254e568SChristoph Hellwig if (map_len > end - pos) 1045a254e568SChristoph Hellwig map_len = end - pos; 1046a254e568SChristoph Hellwig 1047a254e568SChristoph Hellwig if (iov_iter_rw(iter) == WRITE) 1048a254e568SChristoph Hellwig map_len = copy_from_iter_pmem(dax.addr, map_len, iter); 1049a254e568SChristoph Hellwig else 1050a254e568SChristoph Hellwig map_len = copy_to_iter(dax.addr, map_len, iter); 1051a254e568SChristoph Hellwig dax_unmap_atomic(iomap->bdev, &dax); 1052a254e568SChristoph Hellwig if (map_len <= 0) { 1053a254e568SChristoph Hellwig ret = map_len ? map_len : -EFAULT; 1054a254e568SChristoph Hellwig break; 1055a254e568SChristoph Hellwig } 1056a254e568SChristoph Hellwig 1057a254e568SChristoph Hellwig pos += map_len; 1058a254e568SChristoph Hellwig length -= map_len; 1059a254e568SChristoph Hellwig done += map_len; 1060a254e568SChristoph Hellwig } 1061a254e568SChristoph Hellwig 1062a254e568SChristoph Hellwig return done ? done : ret; 1063a254e568SChristoph Hellwig } 1064a254e568SChristoph Hellwig 1065a254e568SChristoph Hellwig /** 106611c59c92SRoss Zwisler * dax_iomap_rw - Perform I/O to a DAX file 1067a254e568SChristoph Hellwig * @iocb: The control block for this I/O 1068a254e568SChristoph Hellwig * @iter: The addresses to do I/O from or to 1069a254e568SChristoph Hellwig * @ops: iomap ops passed from the file system 1070a254e568SChristoph Hellwig * 1071a254e568SChristoph Hellwig * This function performs read and write operations to directly mapped 1072a254e568SChristoph Hellwig * persistent memory. The callers needs to take care of read/write exclusion 1073a254e568SChristoph Hellwig * and evicting any page cache pages in the region under I/O. 1074a254e568SChristoph Hellwig */ 1075a254e568SChristoph Hellwig ssize_t 107611c59c92SRoss Zwisler dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 10778ff6daa1SChristoph Hellwig const struct iomap_ops *ops) 1078a254e568SChristoph Hellwig { 1079a254e568SChristoph Hellwig struct address_space *mapping = iocb->ki_filp->f_mapping; 1080a254e568SChristoph Hellwig struct inode *inode = mapping->host; 1081a254e568SChristoph Hellwig loff_t pos = iocb->ki_pos, ret = 0, done = 0; 1082a254e568SChristoph Hellwig unsigned flags = 0; 1083a254e568SChristoph Hellwig 1084a254e568SChristoph Hellwig if (iov_iter_rw(iter) == WRITE) 1085a254e568SChristoph Hellwig flags |= IOMAP_WRITE; 1086a254e568SChristoph Hellwig 1087a254e568SChristoph Hellwig while (iov_iter_count(iter)) { 1088a254e568SChristoph Hellwig ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, 108911c59c92SRoss Zwisler iter, dax_iomap_actor); 1090a254e568SChristoph Hellwig if (ret <= 0) 1091a254e568SChristoph Hellwig break; 1092a254e568SChristoph Hellwig pos += ret; 1093a254e568SChristoph Hellwig done += ret; 1094a254e568SChristoph Hellwig } 1095a254e568SChristoph Hellwig 1096a254e568SChristoph Hellwig iocb->ki_pos += done; 1097a254e568SChristoph Hellwig return done ? done : ret; 1098a254e568SChristoph Hellwig } 109911c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_rw); 1100a7d73fe6SChristoph Hellwig 11019f141d6eSJan Kara static int dax_fault_return(int error) 11029f141d6eSJan Kara { 11039f141d6eSJan Kara if (error == 0) 11049f141d6eSJan Kara return VM_FAULT_NOPAGE; 11059f141d6eSJan Kara if (error == -ENOMEM) 11069f141d6eSJan Kara return VM_FAULT_OOM; 11079f141d6eSJan Kara return VM_FAULT_SIGBUS; 11089f141d6eSJan Kara } 11099f141d6eSJan Kara 1110a7d73fe6SChristoph Hellwig /** 111111c59c92SRoss Zwisler * dax_iomap_fault - handle a page fault on a DAX file 1112a7d73fe6SChristoph Hellwig * @vma: The virtual memory area where the fault occurred 1113a7d73fe6SChristoph Hellwig * @vmf: The description of the fault 1114a7d73fe6SChristoph Hellwig * @ops: iomap ops passed from the file system 1115a7d73fe6SChristoph Hellwig * 1116a7d73fe6SChristoph Hellwig * When a page fault occurs, filesystems may call this helper in their fault 1117a7d73fe6SChristoph Hellwig * or mkwrite handler for DAX files. Assumes the caller has done all the 1118a7d73fe6SChristoph Hellwig * necessary locking for the page fault to proceed successfully. 1119a7d73fe6SChristoph Hellwig */ 112011c59c92SRoss Zwisler int dax_iomap_fault(struct vm_area_struct *vma, struct vm_fault *vmf, 11218ff6daa1SChristoph Hellwig const struct iomap_ops *ops) 1122a7d73fe6SChristoph Hellwig { 1123a7d73fe6SChristoph Hellwig struct address_space *mapping = vma->vm_file->f_mapping; 1124a7d73fe6SChristoph Hellwig struct inode *inode = mapping->host; 11251a29d85eSJan Kara unsigned long vaddr = vmf->address; 1126a7d73fe6SChristoph Hellwig loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; 1127a7d73fe6SChristoph Hellwig sector_t sector; 1128a7d73fe6SChristoph Hellwig struct iomap iomap = { 0 }; 11299484ab1bSJan Kara unsigned flags = IOMAP_FAULT; 1130a7d73fe6SChristoph Hellwig int error, major = 0; 1131b1aa812bSJan Kara int vmf_ret = 0; 1132a7d73fe6SChristoph Hellwig void *entry; 1133a7d73fe6SChristoph Hellwig 1134a7d73fe6SChristoph Hellwig /* 1135a7d73fe6SChristoph Hellwig * Check whether offset isn't beyond end of file now. Caller is supposed 1136a7d73fe6SChristoph Hellwig * to hold locks serializing us with truncate / punch hole so this is 1137a7d73fe6SChristoph Hellwig * a reliable test. 1138a7d73fe6SChristoph Hellwig */ 1139a7d73fe6SChristoph Hellwig if (pos >= i_size_read(inode)) 1140a7d73fe6SChristoph Hellwig return VM_FAULT_SIGBUS; 1141a7d73fe6SChristoph Hellwig 1142a7d73fe6SChristoph Hellwig if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) 1143a7d73fe6SChristoph Hellwig flags |= IOMAP_WRITE; 1144a7d73fe6SChristoph Hellwig 1145a7d73fe6SChristoph Hellwig /* 1146a7d73fe6SChristoph Hellwig * Note that we don't bother to use iomap_apply here: DAX required 1147a7d73fe6SChristoph Hellwig * the file system block size to be equal the page size, which means 1148a7d73fe6SChristoph Hellwig * that we never have to deal with more than a single extent here. 1149a7d73fe6SChristoph Hellwig */ 1150a7d73fe6SChristoph Hellwig error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); 1151a7d73fe6SChristoph Hellwig if (error) 11529f141d6eSJan Kara return dax_fault_return(error); 1153a7d73fe6SChristoph Hellwig if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { 11549f141d6eSJan Kara vmf_ret = dax_fault_return(-EIO); /* fs corruption? */ 11559f141d6eSJan Kara goto finish_iomap; 11569f141d6eSJan Kara } 11579f141d6eSJan Kara 11589f141d6eSJan Kara entry = grab_mapping_entry(mapping, vmf->pgoff, 0); 11599f141d6eSJan Kara if (IS_ERR(entry)) { 11609f141d6eSJan Kara vmf_ret = dax_fault_return(PTR_ERR(entry)); 11611550290bSRoss Zwisler goto finish_iomap; 1162a7d73fe6SChristoph Hellwig } 1163a7d73fe6SChristoph Hellwig 1164333ccc97SRoss Zwisler sector = dax_iomap_sector(&iomap, pos); 1165a7d73fe6SChristoph Hellwig 1166a7d73fe6SChristoph Hellwig if (vmf->cow_page) { 1167a7d73fe6SChristoph Hellwig switch (iomap.type) { 1168a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 1169a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1170a7d73fe6SChristoph Hellwig clear_user_highpage(vmf->cow_page, vaddr); 1171a7d73fe6SChristoph Hellwig break; 1172a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1173a7d73fe6SChristoph Hellwig error = copy_user_dax(iomap.bdev, sector, PAGE_SIZE, 1174a7d73fe6SChristoph Hellwig vmf->cow_page, vaddr); 1175a7d73fe6SChristoph Hellwig break; 1176a7d73fe6SChristoph Hellwig default: 1177a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1178a7d73fe6SChristoph Hellwig error = -EIO; 1179a7d73fe6SChristoph Hellwig break; 1180a7d73fe6SChristoph Hellwig } 1181a7d73fe6SChristoph Hellwig 1182a7d73fe6SChristoph Hellwig if (error) 11839f141d6eSJan Kara goto error_unlock_entry; 1184b1aa812bSJan Kara 1185b1aa812bSJan Kara __SetPageUptodate(vmf->cow_page); 1186b1aa812bSJan Kara vmf_ret = finish_fault(vmf); 1187b1aa812bSJan Kara if (!vmf_ret) 1188b1aa812bSJan Kara vmf_ret = VM_FAULT_DONE_COW; 11899f141d6eSJan Kara goto unlock_entry; 1190a7d73fe6SChristoph Hellwig } 1191a7d73fe6SChristoph Hellwig 1192a7d73fe6SChristoph Hellwig switch (iomap.type) { 1193a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1194a7d73fe6SChristoph Hellwig if (iomap.flags & IOMAP_F_NEW) { 1195a7d73fe6SChristoph Hellwig count_vm_event(PGMAJFAULT); 1196a7d73fe6SChristoph Hellwig mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT); 1197a7d73fe6SChristoph Hellwig major = VM_FAULT_MAJOR; 1198a7d73fe6SChristoph Hellwig } 1199a7d73fe6SChristoph Hellwig error = dax_insert_mapping(mapping, iomap.bdev, sector, 1200a7d73fe6SChristoph Hellwig PAGE_SIZE, &entry, vma, vmf); 12019f141d6eSJan Kara /* -EBUSY is fine, somebody else faulted on the same PTE */ 12029f141d6eSJan Kara if (error == -EBUSY) 12039f141d6eSJan Kara error = 0; 1204a7d73fe6SChristoph Hellwig break; 1205a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1206a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 12071550290bSRoss Zwisler if (!(vmf->flags & FAULT_FLAG_WRITE)) { 1208f449b936SJan Kara vmf_ret = dax_load_hole(mapping, &entry, vmf); 12099f141d6eSJan Kara goto unlock_entry; 12101550290bSRoss Zwisler } 1211a7d73fe6SChristoph Hellwig /*FALLTHRU*/ 1212a7d73fe6SChristoph Hellwig default: 1213a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1214a7d73fe6SChristoph Hellwig error = -EIO; 1215a7d73fe6SChristoph Hellwig break; 1216a7d73fe6SChristoph Hellwig } 1217a7d73fe6SChristoph Hellwig 12189f141d6eSJan Kara error_unlock_entry: 12199f141d6eSJan Kara vmf_ret = dax_fault_return(error) | major; 1220a7d73fe6SChristoph Hellwig unlock_entry: 1221a7d73fe6SChristoph Hellwig put_locked_mapping_entry(mapping, vmf->pgoff, entry); 12229f141d6eSJan Kara finish_iomap: 12239f141d6eSJan Kara if (ops->iomap_end) { 12249f141d6eSJan Kara int copied = PAGE_SIZE; 12259f141d6eSJan Kara 12269f141d6eSJan Kara if (vmf_ret & VM_FAULT_ERROR) 12279f141d6eSJan Kara copied = 0; 12289f141d6eSJan Kara /* 12299f141d6eSJan Kara * The fault is done by now and there's no way back (other 12309f141d6eSJan Kara * thread may be already happily using PTE we have installed). 12319f141d6eSJan Kara * Just ignore error from ->iomap_end since we cannot do much 12329f141d6eSJan Kara * with it. 12339f141d6eSJan Kara */ 12349f141d6eSJan Kara ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); 12351550290bSRoss Zwisler } 12369f141d6eSJan Kara return vmf_ret; 1237a7d73fe6SChristoph Hellwig } 123811c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_fault); 1239642261acSRoss Zwisler 1240642261acSRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 1241642261acSRoss Zwisler /* 1242642261acSRoss Zwisler * The 'colour' (ie low bits) within a PMD of a page offset. This comes up 1243642261acSRoss Zwisler * more often than one might expect in the below functions. 1244642261acSRoss Zwisler */ 1245642261acSRoss Zwisler #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) 1246642261acSRoss Zwisler 1247642261acSRoss Zwisler static int dax_pmd_insert_mapping(struct vm_area_struct *vma, pmd_t *pmd, 1248642261acSRoss Zwisler struct vm_fault *vmf, unsigned long address, 1249642261acSRoss Zwisler struct iomap *iomap, loff_t pos, bool write, void **entryp) 1250642261acSRoss Zwisler { 1251642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1252642261acSRoss Zwisler struct block_device *bdev = iomap->bdev; 1253642261acSRoss Zwisler struct blk_dax_ctl dax = { 1254642261acSRoss Zwisler .sector = dax_iomap_sector(iomap, pos), 1255642261acSRoss Zwisler .size = PMD_SIZE, 1256642261acSRoss Zwisler }; 1257642261acSRoss Zwisler long length = dax_map_atomic(bdev, &dax); 1258642261acSRoss Zwisler void *ret; 1259642261acSRoss Zwisler 1260642261acSRoss Zwisler if (length < 0) /* dax_map_atomic() failed */ 1261642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1262642261acSRoss Zwisler if (length < PMD_SIZE) 1263642261acSRoss Zwisler goto unmap_fallback; 1264642261acSRoss Zwisler if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) 1265642261acSRoss Zwisler goto unmap_fallback; 1266642261acSRoss Zwisler if (!pfn_t_devmap(dax.pfn)) 1267642261acSRoss Zwisler goto unmap_fallback; 1268642261acSRoss Zwisler 1269642261acSRoss Zwisler dax_unmap_atomic(bdev, &dax); 1270642261acSRoss Zwisler 1271642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, *entryp, dax.sector, 1272642261acSRoss Zwisler RADIX_DAX_PMD); 1273642261acSRoss Zwisler if (IS_ERR(ret)) 1274642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1275642261acSRoss Zwisler *entryp = ret; 1276642261acSRoss Zwisler 1277642261acSRoss Zwisler return vmf_insert_pfn_pmd(vma, address, pmd, dax.pfn, write); 1278642261acSRoss Zwisler 1279642261acSRoss Zwisler unmap_fallback: 1280642261acSRoss Zwisler dax_unmap_atomic(bdev, &dax); 1281642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1282642261acSRoss Zwisler } 1283642261acSRoss Zwisler 1284642261acSRoss Zwisler static int dax_pmd_load_hole(struct vm_area_struct *vma, pmd_t *pmd, 1285642261acSRoss Zwisler struct vm_fault *vmf, unsigned long address, 1286642261acSRoss Zwisler struct iomap *iomap, void **entryp) 1287642261acSRoss Zwisler { 1288642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1289642261acSRoss Zwisler unsigned long pmd_addr = address & PMD_MASK; 1290642261acSRoss Zwisler struct page *zero_page; 1291642261acSRoss Zwisler spinlock_t *ptl; 1292642261acSRoss Zwisler pmd_t pmd_entry; 1293642261acSRoss Zwisler void *ret; 1294642261acSRoss Zwisler 1295642261acSRoss Zwisler zero_page = mm_get_huge_zero_page(vma->vm_mm); 1296642261acSRoss Zwisler 1297642261acSRoss Zwisler if (unlikely(!zero_page)) 1298642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1299642261acSRoss Zwisler 1300642261acSRoss Zwisler ret = dax_insert_mapping_entry(mapping, vmf, *entryp, 0, 1301642261acSRoss Zwisler RADIX_DAX_PMD | RADIX_DAX_HZP); 1302642261acSRoss Zwisler if (IS_ERR(ret)) 1303642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1304642261acSRoss Zwisler *entryp = ret; 1305642261acSRoss Zwisler 1306642261acSRoss Zwisler ptl = pmd_lock(vma->vm_mm, pmd); 1307642261acSRoss Zwisler if (!pmd_none(*pmd)) { 1308642261acSRoss Zwisler spin_unlock(ptl); 1309642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1310642261acSRoss Zwisler } 1311642261acSRoss Zwisler 1312642261acSRoss Zwisler pmd_entry = mk_pmd(zero_page, vma->vm_page_prot); 1313642261acSRoss Zwisler pmd_entry = pmd_mkhuge(pmd_entry); 1314642261acSRoss Zwisler set_pmd_at(vma->vm_mm, pmd_addr, pmd, pmd_entry); 1315642261acSRoss Zwisler spin_unlock(ptl); 1316642261acSRoss Zwisler return VM_FAULT_NOPAGE; 1317642261acSRoss Zwisler } 1318642261acSRoss Zwisler 1319642261acSRoss Zwisler int dax_iomap_pmd_fault(struct vm_area_struct *vma, unsigned long address, 13208ff6daa1SChristoph Hellwig pmd_t *pmd, unsigned int flags, const struct iomap_ops *ops) 1321642261acSRoss Zwisler { 1322642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1323642261acSRoss Zwisler unsigned long pmd_addr = address & PMD_MASK; 1324642261acSRoss Zwisler bool write = flags & FAULT_FLAG_WRITE; 13259484ab1bSJan Kara unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; 1326642261acSRoss Zwisler struct inode *inode = mapping->host; 1327642261acSRoss Zwisler int result = VM_FAULT_FALLBACK; 1328642261acSRoss Zwisler struct iomap iomap = { 0 }; 1329642261acSRoss Zwisler pgoff_t max_pgoff, pgoff; 1330642261acSRoss Zwisler struct vm_fault vmf; 1331642261acSRoss Zwisler void *entry; 1332642261acSRoss Zwisler loff_t pos; 1333642261acSRoss Zwisler int error; 1334642261acSRoss Zwisler 1335642261acSRoss Zwisler /* Fall back to PTEs if we're going to COW */ 1336642261acSRoss Zwisler if (write && !(vma->vm_flags & VM_SHARED)) 1337642261acSRoss Zwisler goto fallback; 1338642261acSRoss Zwisler 1339642261acSRoss Zwisler /* If the PMD would extend outside the VMA */ 1340642261acSRoss Zwisler if (pmd_addr < vma->vm_start) 1341642261acSRoss Zwisler goto fallback; 1342642261acSRoss Zwisler if ((pmd_addr + PMD_SIZE) > vma->vm_end) 1343642261acSRoss Zwisler goto fallback; 1344642261acSRoss Zwisler 1345642261acSRoss Zwisler /* 1346642261acSRoss Zwisler * Check whether offset isn't beyond end of file now. Caller is 1347642261acSRoss Zwisler * supposed to hold locks serializing us with truncate / punch hole so 1348642261acSRoss Zwisler * this is a reliable test. 1349642261acSRoss Zwisler */ 1350642261acSRoss Zwisler pgoff = linear_page_index(vma, pmd_addr); 1351642261acSRoss Zwisler max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT; 1352642261acSRoss Zwisler 1353642261acSRoss Zwisler if (pgoff > max_pgoff) 1354642261acSRoss Zwisler return VM_FAULT_SIGBUS; 1355642261acSRoss Zwisler 1356642261acSRoss Zwisler /* If the PMD would extend beyond the file size */ 1357642261acSRoss Zwisler if ((pgoff | PG_PMD_COLOUR) > max_pgoff) 1358642261acSRoss Zwisler goto fallback; 1359642261acSRoss Zwisler 1360642261acSRoss Zwisler /* 1361642261acSRoss Zwisler * Note that we don't use iomap_apply here. We aren't doing I/O, only 1362642261acSRoss Zwisler * setting up a mapping, so really we're using iomap_begin() as a way 1363642261acSRoss Zwisler * to look up our filesystem block. 1364642261acSRoss Zwisler */ 1365642261acSRoss Zwisler pos = (loff_t)pgoff << PAGE_SHIFT; 1366642261acSRoss Zwisler error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); 1367642261acSRoss Zwisler if (error) 13689f141d6eSJan Kara goto fallback; 13699f141d6eSJan Kara 1370642261acSRoss Zwisler if (iomap.offset + iomap.length < pos + PMD_SIZE) 1371642261acSRoss Zwisler goto finish_iomap; 1372642261acSRoss Zwisler 13739f141d6eSJan Kara /* 13749f141d6eSJan Kara * grab_mapping_entry() will make sure we get a 2M empty entry, a DAX 13759f141d6eSJan Kara * PMD or a HZP entry. If it can't (because a 4k page is already in 13769f141d6eSJan Kara * the tree, for instance), it will return -EEXIST and we just fall 13779f141d6eSJan Kara * back to 4k entries. 13789f141d6eSJan Kara */ 13799f141d6eSJan Kara entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD); 13809f141d6eSJan Kara if (IS_ERR(entry)) 13819f141d6eSJan Kara goto finish_iomap; 13829f141d6eSJan Kara 1383642261acSRoss Zwisler vmf.pgoff = pgoff; 1384642261acSRoss Zwisler vmf.flags = flags; 1385642261acSRoss Zwisler vmf.gfp_mask = mapping_gfp_mask(mapping) | __GFP_IO; 1386642261acSRoss Zwisler 1387642261acSRoss Zwisler switch (iomap.type) { 1388642261acSRoss Zwisler case IOMAP_MAPPED: 1389642261acSRoss Zwisler result = dax_pmd_insert_mapping(vma, pmd, &vmf, address, 1390642261acSRoss Zwisler &iomap, pos, write, &entry); 1391642261acSRoss Zwisler break; 1392642261acSRoss Zwisler case IOMAP_UNWRITTEN: 1393642261acSRoss Zwisler case IOMAP_HOLE: 1394642261acSRoss Zwisler if (WARN_ON_ONCE(write)) 13959f141d6eSJan Kara goto unlock_entry; 1396642261acSRoss Zwisler result = dax_pmd_load_hole(vma, pmd, &vmf, address, &iomap, 1397642261acSRoss Zwisler &entry); 1398642261acSRoss Zwisler break; 1399642261acSRoss Zwisler default: 1400642261acSRoss Zwisler WARN_ON_ONCE(1); 1401642261acSRoss Zwisler break; 1402642261acSRoss Zwisler } 1403642261acSRoss Zwisler 1404642261acSRoss Zwisler unlock_entry: 1405642261acSRoss Zwisler put_locked_mapping_entry(mapping, pgoff, entry); 14069f141d6eSJan Kara finish_iomap: 14079f141d6eSJan Kara if (ops->iomap_end) { 14089f141d6eSJan Kara int copied = PMD_SIZE; 14099f141d6eSJan Kara 14109f141d6eSJan Kara if (result == VM_FAULT_FALLBACK) 14119f141d6eSJan Kara copied = 0; 14129f141d6eSJan Kara /* 14139f141d6eSJan Kara * The fault is done by now and there's no way back (other 14149f141d6eSJan Kara * thread may be already happily using PMD we have installed). 14159f141d6eSJan Kara * Just ignore error from ->iomap_end since we cannot do much 14169f141d6eSJan Kara * with it. 14179f141d6eSJan Kara */ 14189f141d6eSJan Kara ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, 14199f141d6eSJan Kara &iomap); 14209f141d6eSJan Kara } 1421642261acSRoss Zwisler fallback: 1422642261acSRoss Zwisler if (result == VM_FAULT_FALLBACK) { 1423642261acSRoss Zwisler split_huge_pmd(vma, pmd, address); 1424642261acSRoss Zwisler count_vm_event(THP_FAULT_FALLBACK); 1425642261acSRoss Zwisler } 1426642261acSRoss Zwisler return result; 1427642261acSRoss Zwisler } 1428642261acSRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_pmd_fault); 1429642261acSRoss Zwisler #endif /* CONFIG_FS_DAX_PMD */ 1430