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> 28289c6aedSMatthew Wilcox #include <linux/sched.h> 29f361bf4aSIngo Molnar #include <linux/sched/signal.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 38282a8e03SRoss Zwisler #define CREATE_TRACE_POINTS 39282a8e03SRoss Zwisler #include <trace/events/fs_dax.h> 40282a8e03SRoss Zwisler 41cfc93c6cSMatthew Wilcox static inline unsigned int pe_order(enum page_entry_size pe_size) 42cfc93c6cSMatthew Wilcox { 43cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PTE) 44cfc93c6cSMatthew Wilcox return PAGE_SHIFT - PAGE_SHIFT; 45cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PMD) 46cfc93c6cSMatthew Wilcox return PMD_SHIFT - PAGE_SHIFT; 47cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PUD) 48cfc93c6cSMatthew Wilcox return PUD_SHIFT - PAGE_SHIFT; 49cfc93c6cSMatthew Wilcox return ~0; 50cfc93c6cSMatthew Wilcox } 51cfc93c6cSMatthew Wilcox 52ac401cc7SJan Kara /* We choose 4096 entries - same as per-zone page wait tables */ 53ac401cc7SJan Kara #define DAX_WAIT_TABLE_BITS 12 54ac401cc7SJan Kara #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) 55ac401cc7SJan Kara 56917f3452SRoss Zwisler /* The 'colour' (ie low bits) within a PMD of a page offset. */ 57917f3452SRoss Zwisler #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) 58977fbdcdSMatthew Wilcox #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) 59917f3452SRoss Zwisler 60cfc93c6cSMatthew Wilcox /* The order of a PMD entry */ 61cfc93c6cSMatthew Wilcox #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT) 62cfc93c6cSMatthew Wilcox 63ce95ab0fSRoss Zwisler static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; 64ac401cc7SJan Kara 65ac401cc7SJan Kara static int __init init_dax_wait_table(void) 66ac401cc7SJan Kara { 67ac401cc7SJan Kara int i; 68ac401cc7SJan Kara 69ac401cc7SJan Kara for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) 70ac401cc7SJan Kara init_waitqueue_head(wait_table + i); 71ac401cc7SJan Kara return 0; 72ac401cc7SJan Kara } 73ac401cc7SJan Kara fs_initcall(init_dax_wait_table); 74ac401cc7SJan Kara 75527b19d0SRoss Zwisler /* 763159f943SMatthew Wilcox * DAX pagecache entries use XArray value entries so they can't be mistaken 773159f943SMatthew Wilcox * for pages. We use one bit for locking, one bit for the entry size (PMD) 783159f943SMatthew Wilcox * and two more to tell us if the entry is a zero page or an empty entry that 793159f943SMatthew Wilcox * is just used for locking. In total four special bits. 80527b19d0SRoss Zwisler * 81527b19d0SRoss Zwisler * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE 82527b19d0SRoss Zwisler * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem 83527b19d0SRoss Zwisler * block allocation. 84527b19d0SRoss Zwisler */ 853159f943SMatthew Wilcox #define DAX_SHIFT (4) 863159f943SMatthew Wilcox #define DAX_LOCKED (1UL << 0) 873159f943SMatthew Wilcox #define DAX_PMD (1UL << 1) 883159f943SMatthew Wilcox #define DAX_ZERO_PAGE (1UL << 2) 893159f943SMatthew Wilcox #define DAX_EMPTY (1UL << 3) 90527b19d0SRoss Zwisler 91a77d19f4SMatthew Wilcox static unsigned long dax_to_pfn(void *entry) 92527b19d0SRoss Zwisler { 933159f943SMatthew Wilcox return xa_to_value(entry) >> DAX_SHIFT; 94527b19d0SRoss Zwisler } 95527b19d0SRoss Zwisler 969f32d221SMatthew Wilcox static void *dax_make_entry(pfn_t pfn, unsigned long flags) 979f32d221SMatthew Wilcox { 989f32d221SMatthew Wilcox return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT)); 999f32d221SMatthew Wilcox } 1009f32d221SMatthew Wilcox 101cfc93c6cSMatthew Wilcox static bool dax_is_locked(void *entry) 102cfc93c6cSMatthew Wilcox { 103cfc93c6cSMatthew Wilcox return xa_to_value(entry) & DAX_LOCKED; 104cfc93c6cSMatthew Wilcox } 105cfc93c6cSMatthew Wilcox 106a77d19f4SMatthew Wilcox static unsigned int dax_entry_order(void *entry) 107527b19d0SRoss Zwisler { 1083159f943SMatthew Wilcox if (xa_to_value(entry) & DAX_PMD) 109cfc93c6cSMatthew Wilcox return PMD_ORDER; 110527b19d0SRoss Zwisler return 0; 111527b19d0SRoss Zwisler } 112527b19d0SRoss Zwisler 113fda490d3SMatthew Wilcox static unsigned long dax_is_pmd_entry(void *entry) 114642261acSRoss Zwisler { 1153159f943SMatthew Wilcox return xa_to_value(entry) & DAX_PMD; 116642261acSRoss Zwisler } 117642261acSRoss Zwisler 118fda490d3SMatthew Wilcox static bool dax_is_pte_entry(void *entry) 119642261acSRoss Zwisler { 1203159f943SMatthew Wilcox return !(xa_to_value(entry) & DAX_PMD); 121642261acSRoss Zwisler } 122642261acSRoss Zwisler 123642261acSRoss Zwisler static int dax_is_zero_entry(void *entry) 124642261acSRoss Zwisler { 1253159f943SMatthew Wilcox return xa_to_value(entry) & DAX_ZERO_PAGE; 126642261acSRoss Zwisler } 127642261acSRoss Zwisler 128642261acSRoss Zwisler static int dax_is_empty_entry(void *entry) 129642261acSRoss Zwisler { 1303159f943SMatthew Wilcox return xa_to_value(entry) & DAX_EMPTY; 131642261acSRoss Zwisler } 132642261acSRoss Zwisler 133f7ca90b1SMatthew Wilcox /* 134a77d19f4SMatthew Wilcox * DAX page cache entry locking 135ac401cc7SJan Kara */ 136ac401cc7SJan Kara struct exceptional_entry_key { 137ec4907ffSMatthew Wilcox struct xarray *xa; 13863e95b5cSRoss Zwisler pgoff_t entry_start; 139ac401cc7SJan Kara }; 140ac401cc7SJan Kara 141ac401cc7SJan Kara struct wait_exceptional_entry_queue { 142ac6424b9SIngo Molnar wait_queue_entry_t wait; 143ac401cc7SJan Kara struct exceptional_entry_key key; 144ac401cc7SJan Kara }; 145ac401cc7SJan Kara 146b15cd800SMatthew Wilcox static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas, 147b15cd800SMatthew Wilcox void *entry, struct exceptional_entry_key *key) 14863e95b5cSRoss Zwisler { 14963e95b5cSRoss Zwisler unsigned long hash; 150b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 15163e95b5cSRoss Zwisler 15263e95b5cSRoss Zwisler /* 15363e95b5cSRoss Zwisler * If 'entry' is a PMD, align the 'index' that we use for the wait 15463e95b5cSRoss Zwisler * queue to the start of that PMD. This ensures that all offsets in 15563e95b5cSRoss Zwisler * the range covered by the PMD map to the same bit lock. 15663e95b5cSRoss Zwisler */ 157642261acSRoss Zwisler if (dax_is_pmd_entry(entry)) 158917f3452SRoss Zwisler index &= ~PG_PMD_COLOUR; 159b15cd800SMatthew Wilcox key->xa = xas->xa; 16063e95b5cSRoss Zwisler key->entry_start = index; 16163e95b5cSRoss Zwisler 162b15cd800SMatthew Wilcox hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS); 16363e95b5cSRoss Zwisler return wait_table + hash; 16463e95b5cSRoss Zwisler } 16563e95b5cSRoss Zwisler 166ec4907ffSMatthew Wilcox static int wake_exceptional_entry_func(wait_queue_entry_t *wait, 167ec4907ffSMatthew Wilcox unsigned int mode, int sync, void *keyp) 168ac401cc7SJan Kara { 169ac401cc7SJan Kara struct exceptional_entry_key *key = keyp; 170ac401cc7SJan Kara struct wait_exceptional_entry_queue *ewait = 171ac401cc7SJan Kara container_of(wait, struct wait_exceptional_entry_queue, wait); 172ac401cc7SJan Kara 173ec4907ffSMatthew Wilcox if (key->xa != ewait->key.xa || 17463e95b5cSRoss Zwisler key->entry_start != ewait->key.entry_start) 175ac401cc7SJan Kara return 0; 176ac401cc7SJan Kara return autoremove_wake_function(wait, mode, sync, NULL); 177ac401cc7SJan Kara } 178ac401cc7SJan Kara 179ac401cc7SJan Kara /* 180b93b0163SMatthew Wilcox * @entry may no longer be the entry at the index in the mapping. 181b93b0163SMatthew Wilcox * The important information it's conveying is whether the entry at 182b93b0163SMatthew Wilcox * this index used to be a PMD entry. 183e30331ffSRoss Zwisler */ 184b15cd800SMatthew Wilcox static void dax_wake_entry(struct xa_state *xas, void *entry, bool wake_all) 185e30331ffSRoss Zwisler { 186e30331ffSRoss Zwisler struct exceptional_entry_key key; 187e30331ffSRoss Zwisler wait_queue_head_t *wq; 188e30331ffSRoss Zwisler 189b15cd800SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &key); 190e30331ffSRoss Zwisler 191e30331ffSRoss Zwisler /* 192e30331ffSRoss Zwisler * Checking for locked entry and prepare_to_wait_exclusive() happens 193b93b0163SMatthew Wilcox * under the i_pages lock, ditto for entry handling in our callers. 194e30331ffSRoss Zwisler * So at this point all tasks that could have seen our entry locked 195e30331ffSRoss Zwisler * must be in the waitqueue and the following check will see them. 196e30331ffSRoss Zwisler */ 197e30331ffSRoss Zwisler if (waitqueue_active(wq)) 198e30331ffSRoss Zwisler __wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key); 199e30331ffSRoss Zwisler } 200e30331ffSRoss Zwisler 201cfc93c6cSMatthew Wilcox /* 202cfc93c6cSMatthew Wilcox * Look up entry in page cache, wait for it to become unlocked if it 203cfc93c6cSMatthew Wilcox * is a DAX entry and return it. The caller must subsequently call 204cfc93c6cSMatthew Wilcox * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry() 205cfc93c6cSMatthew Wilcox * if it did. 206cfc93c6cSMatthew Wilcox * 207cfc93c6cSMatthew Wilcox * Must be called with the i_pages lock held. 208cfc93c6cSMatthew Wilcox */ 209cfc93c6cSMatthew Wilcox static void *get_unlocked_entry(struct xa_state *xas) 210cfc93c6cSMatthew Wilcox { 211cfc93c6cSMatthew Wilcox void *entry; 212cfc93c6cSMatthew Wilcox struct wait_exceptional_entry_queue ewait; 213cfc93c6cSMatthew Wilcox wait_queue_head_t *wq; 214cfc93c6cSMatthew Wilcox 215cfc93c6cSMatthew Wilcox init_wait(&ewait.wait); 216cfc93c6cSMatthew Wilcox ewait.wait.func = wake_exceptional_entry_func; 217cfc93c6cSMatthew Wilcox 218cfc93c6cSMatthew Wilcox for (;;) { 2190e40de03SMatthew Wilcox entry = xas_find_conflict(xas); 2200e40de03SMatthew Wilcox if (!entry || WARN_ON_ONCE(!xa_is_value(entry)) || 221cfc93c6cSMatthew Wilcox !dax_is_locked(entry)) 222cfc93c6cSMatthew Wilcox return entry; 223cfc93c6cSMatthew Wilcox 224b15cd800SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &ewait.key); 225cfc93c6cSMatthew Wilcox prepare_to_wait_exclusive(wq, &ewait.wait, 226cfc93c6cSMatthew Wilcox TASK_UNINTERRUPTIBLE); 227cfc93c6cSMatthew Wilcox xas_unlock_irq(xas); 228cfc93c6cSMatthew Wilcox xas_reset(xas); 229cfc93c6cSMatthew Wilcox schedule(); 230cfc93c6cSMatthew Wilcox finish_wait(wq, &ewait.wait); 231cfc93c6cSMatthew Wilcox xas_lock_irq(xas); 232cfc93c6cSMatthew Wilcox } 233cfc93c6cSMatthew Wilcox } 234cfc93c6cSMatthew Wilcox 23555e56f06SMatthew Wilcox /* 23655e56f06SMatthew Wilcox * The only thing keeping the address space around is the i_pages lock 23755e56f06SMatthew Wilcox * (it's cycled in clear_inode() after removing the entries from i_pages) 23855e56f06SMatthew Wilcox * After we call xas_unlock_irq(), we cannot touch xas->xa. 23955e56f06SMatthew Wilcox */ 24055e56f06SMatthew Wilcox static void wait_entry_unlocked(struct xa_state *xas, void *entry) 24155e56f06SMatthew Wilcox { 24255e56f06SMatthew Wilcox struct wait_exceptional_entry_queue ewait; 24355e56f06SMatthew Wilcox wait_queue_head_t *wq; 24455e56f06SMatthew Wilcox 24555e56f06SMatthew Wilcox init_wait(&ewait.wait); 24655e56f06SMatthew Wilcox ewait.wait.func = wake_exceptional_entry_func; 24755e56f06SMatthew Wilcox 24855e56f06SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &ewait.key); 24955e56f06SMatthew Wilcox prepare_to_wait_exclusive(wq, &ewait.wait, TASK_UNINTERRUPTIBLE); 25055e56f06SMatthew Wilcox xas_unlock_irq(xas); 25155e56f06SMatthew Wilcox schedule(); 25255e56f06SMatthew Wilcox finish_wait(wq, &ewait.wait); 25355e56f06SMatthew Wilcox 25455e56f06SMatthew Wilcox /* 25555e56f06SMatthew Wilcox * Entry lock waits are exclusive. Wake up the next waiter since 25655e56f06SMatthew Wilcox * we aren't sure we will acquire the entry lock and thus wake 25755e56f06SMatthew Wilcox * the next waiter up on unlock. 25855e56f06SMatthew Wilcox */ 25955e56f06SMatthew Wilcox if (waitqueue_active(wq)) 26055e56f06SMatthew Wilcox __wake_up(wq, TASK_NORMAL, 1, &ewait.key); 26155e56f06SMatthew Wilcox } 26255e56f06SMatthew Wilcox 263cfc93c6cSMatthew Wilcox static void put_unlocked_entry(struct xa_state *xas, void *entry) 264cfc93c6cSMatthew Wilcox { 265cfc93c6cSMatthew Wilcox /* If we were the only waiter woken, wake the next one */ 266cfc93c6cSMatthew Wilcox if (entry) 267cfc93c6cSMatthew Wilcox dax_wake_entry(xas, entry, false); 268cfc93c6cSMatthew Wilcox } 269cfc93c6cSMatthew Wilcox 270cfc93c6cSMatthew Wilcox /* 271cfc93c6cSMatthew Wilcox * We used the xa_state to get the entry, but then we locked the entry and 272cfc93c6cSMatthew Wilcox * dropped the xa_lock, so we know the xa_state is stale and must be reset 273cfc93c6cSMatthew Wilcox * before use. 274cfc93c6cSMatthew Wilcox */ 275cfc93c6cSMatthew Wilcox static void dax_unlock_entry(struct xa_state *xas, void *entry) 276cfc93c6cSMatthew Wilcox { 277cfc93c6cSMatthew Wilcox void *old; 278cfc93c6cSMatthew Wilcox 2797ae2ea7dSMatthew Wilcox BUG_ON(dax_is_locked(entry)); 280cfc93c6cSMatthew Wilcox xas_reset(xas); 281cfc93c6cSMatthew Wilcox xas_lock_irq(xas); 282cfc93c6cSMatthew Wilcox old = xas_store(xas, entry); 283cfc93c6cSMatthew Wilcox xas_unlock_irq(xas); 284cfc93c6cSMatthew Wilcox BUG_ON(!dax_is_locked(old)); 285cfc93c6cSMatthew Wilcox dax_wake_entry(xas, entry, false); 286cfc93c6cSMatthew Wilcox } 287cfc93c6cSMatthew Wilcox 288cfc93c6cSMatthew Wilcox /* 289cfc93c6cSMatthew Wilcox * Return: The entry stored at this location before it was locked. 290cfc93c6cSMatthew Wilcox */ 291cfc93c6cSMatthew Wilcox static void *dax_lock_entry(struct xa_state *xas, void *entry) 292cfc93c6cSMatthew Wilcox { 293cfc93c6cSMatthew Wilcox unsigned long v = xa_to_value(entry); 294cfc93c6cSMatthew Wilcox return xas_store(xas, xa_mk_value(v | DAX_LOCKED)); 295cfc93c6cSMatthew Wilcox } 296cfc93c6cSMatthew Wilcox 297d2c997c0SDan Williams static unsigned long dax_entry_size(void *entry) 298d2c997c0SDan Williams { 299d2c997c0SDan Williams if (dax_is_zero_entry(entry)) 300d2c997c0SDan Williams return 0; 301d2c997c0SDan Williams else if (dax_is_empty_entry(entry)) 302d2c997c0SDan Williams return 0; 303d2c997c0SDan Williams else if (dax_is_pmd_entry(entry)) 304d2c997c0SDan Williams return PMD_SIZE; 305d2c997c0SDan Williams else 306d2c997c0SDan Williams return PAGE_SIZE; 307d2c997c0SDan Williams } 308d2c997c0SDan Williams 309a77d19f4SMatthew Wilcox static unsigned long dax_end_pfn(void *entry) 310d2c997c0SDan Williams { 311a77d19f4SMatthew Wilcox return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; 312d2c997c0SDan Williams } 313d2c997c0SDan Williams 314d2c997c0SDan Williams /* 315d2c997c0SDan Williams * Iterate through all mapped pfns represented by an entry, i.e. skip 316d2c997c0SDan Williams * 'empty' and 'zero' entries. 317d2c997c0SDan Williams */ 318d2c997c0SDan Williams #define for_each_mapped_pfn(entry, pfn) \ 319a77d19f4SMatthew Wilcox for (pfn = dax_to_pfn(entry); \ 320a77d19f4SMatthew Wilcox pfn < dax_end_pfn(entry); pfn++) 321d2c997c0SDan Williams 32273449dafSDan Williams /* 32373449dafSDan Williams * TODO: for reflink+dax we need a way to associate a single page with 32473449dafSDan Williams * multiple address_space instances at different linear_page_index() 32573449dafSDan Williams * offsets. 32673449dafSDan Williams */ 32773449dafSDan Williams static void dax_associate_entry(void *entry, struct address_space *mapping, 32873449dafSDan Williams struct vm_area_struct *vma, unsigned long address) 329d2c997c0SDan Williams { 33073449dafSDan Williams unsigned long size = dax_entry_size(entry), pfn, index; 33173449dafSDan Williams int i = 0; 332d2c997c0SDan Williams 333d2c997c0SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 334d2c997c0SDan Williams return; 335d2c997c0SDan Williams 33673449dafSDan Williams index = linear_page_index(vma, address & ~(size - 1)); 337d2c997c0SDan Williams for_each_mapped_pfn(entry, pfn) { 338d2c997c0SDan Williams struct page *page = pfn_to_page(pfn); 339d2c997c0SDan Williams 340d2c997c0SDan Williams WARN_ON_ONCE(page->mapping); 341d2c997c0SDan Williams page->mapping = mapping; 34273449dafSDan Williams page->index = index + i++; 343d2c997c0SDan Williams } 344d2c997c0SDan Williams } 345d2c997c0SDan Williams 346d2c997c0SDan Williams static void dax_disassociate_entry(void *entry, struct address_space *mapping, 347d2c997c0SDan Williams bool trunc) 348d2c997c0SDan Williams { 349d2c997c0SDan Williams unsigned long pfn; 350d2c997c0SDan Williams 351d2c997c0SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 352d2c997c0SDan Williams return; 353d2c997c0SDan Williams 354d2c997c0SDan Williams for_each_mapped_pfn(entry, pfn) { 355d2c997c0SDan Williams struct page *page = pfn_to_page(pfn); 356d2c997c0SDan Williams 357d2c997c0SDan Williams WARN_ON_ONCE(trunc && page_ref_count(page) > 1); 358d2c997c0SDan Williams WARN_ON_ONCE(page->mapping && page->mapping != mapping); 359d2c997c0SDan Williams page->mapping = NULL; 36073449dafSDan Williams page->index = 0; 361d2c997c0SDan Williams } 362d2c997c0SDan Williams } 363d2c997c0SDan Williams 3645fac7408SDan Williams static struct page *dax_busy_page(void *entry) 3655fac7408SDan Williams { 3665fac7408SDan Williams unsigned long pfn; 3675fac7408SDan Williams 3685fac7408SDan Williams for_each_mapped_pfn(entry, pfn) { 3695fac7408SDan Williams struct page *page = pfn_to_page(pfn); 3705fac7408SDan Williams 3715fac7408SDan Williams if (page_ref_count(page) > 1) 3725fac7408SDan Williams return page; 3735fac7408SDan Williams } 3745fac7408SDan Williams return NULL; 3755fac7408SDan Williams } 3765fac7408SDan Williams 377c5bbd451SMatthew Wilcox /* 378c5bbd451SMatthew Wilcox * dax_lock_mapping_entry - Lock the DAX entry corresponding to a page 379c5bbd451SMatthew Wilcox * @page: The page whose entry we want to lock 380c5bbd451SMatthew Wilcox * 381c5bbd451SMatthew Wilcox * Context: Process context. 382c5bbd451SMatthew Wilcox * Return: %true if the entry was locked or does not need to be locked. 383c5bbd451SMatthew Wilcox */ 384c2a7d2a1SDan Williams bool dax_lock_mapping_entry(struct page *page) 385c2a7d2a1SDan Williams { 3869f32d221SMatthew Wilcox XA_STATE(xas, NULL, 0); 3879f32d221SMatthew Wilcox void *entry; 388c5bbd451SMatthew Wilcox bool locked; 389c2a7d2a1SDan Williams 390c5bbd451SMatthew Wilcox /* Ensure page->mapping isn't freed while we look at it */ 391c5bbd451SMatthew Wilcox rcu_read_lock(); 392c2a7d2a1SDan Williams for (;;) { 3939f32d221SMatthew Wilcox struct address_space *mapping = READ_ONCE(page->mapping); 394c2a7d2a1SDan Williams 395c5bbd451SMatthew Wilcox locked = false; 396c93db7bbSMatthew Wilcox if (!mapping || !dax_mapping(mapping)) 397c5bbd451SMatthew Wilcox break; 398c2a7d2a1SDan Williams 399c2a7d2a1SDan Williams /* 400c2a7d2a1SDan Williams * In the device-dax case there's no need to lock, a 401c2a7d2a1SDan Williams * struct dev_pagemap pin is sufficient to keep the 402c2a7d2a1SDan Williams * inode alive, and we assume we have dev_pagemap pin 403c2a7d2a1SDan Williams * otherwise we would not have a valid pfn_to_page() 404c2a7d2a1SDan Williams * translation. 405c2a7d2a1SDan Williams */ 406c5bbd451SMatthew Wilcox locked = true; 4079f32d221SMatthew Wilcox if (S_ISCHR(mapping->host->i_mode)) 408c5bbd451SMatthew Wilcox break; 409c2a7d2a1SDan Williams 4109f32d221SMatthew Wilcox xas.xa = &mapping->i_pages; 4119f32d221SMatthew Wilcox xas_lock_irq(&xas); 412c2a7d2a1SDan Williams if (mapping != page->mapping) { 4139f32d221SMatthew Wilcox xas_unlock_irq(&xas); 414c2a7d2a1SDan Williams continue; 415c2a7d2a1SDan Williams } 4169f32d221SMatthew Wilcox xas_set(&xas, page->index); 4179f32d221SMatthew Wilcox entry = xas_load(&xas); 4189f32d221SMatthew Wilcox if (dax_is_locked(entry)) { 419c5bbd451SMatthew Wilcox rcu_read_unlock(); 42055e56f06SMatthew Wilcox wait_entry_unlocked(&xas, entry); 421c5bbd451SMatthew Wilcox rcu_read_lock(); 422c2a7d2a1SDan Williams continue; 423c2a7d2a1SDan Williams } 4249f32d221SMatthew Wilcox dax_lock_entry(&xas, entry); 4259f32d221SMatthew Wilcox xas_unlock_irq(&xas); 426c5bbd451SMatthew Wilcox break; 4279f32d221SMatthew Wilcox } 428c5bbd451SMatthew Wilcox rcu_read_unlock(); 429c5bbd451SMatthew Wilcox return locked; 430c2a7d2a1SDan Williams } 431c2a7d2a1SDan Williams 432c2a7d2a1SDan Williams void dax_unlock_mapping_entry(struct page *page) 433c2a7d2a1SDan Williams { 434c2a7d2a1SDan Williams struct address_space *mapping = page->mapping; 4359f32d221SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, page->index); 436fda490d3SMatthew Wilcox void *entry; 437c2a7d2a1SDan Williams 4389f32d221SMatthew Wilcox if (S_ISCHR(mapping->host->i_mode)) 439c2a7d2a1SDan Williams return; 440c2a7d2a1SDan Williams 441fda490d3SMatthew Wilcox rcu_read_lock(); 442fda490d3SMatthew Wilcox entry = xas_load(&xas); 443fda490d3SMatthew Wilcox rcu_read_unlock(); 444fda490d3SMatthew Wilcox entry = dax_make_entry(page_to_pfn_t(page), dax_is_pmd_entry(entry)); 445fda490d3SMatthew Wilcox dax_unlock_entry(&xas, entry); 446c2a7d2a1SDan Williams } 447c2a7d2a1SDan Williams 448ac401cc7SJan Kara /* 449a77d19f4SMatthew Wilcox * Find page cache entry at given index. If it is a DAX entry, return it 450a77d19f4SMatthew Wilcox * with the entry locked. If the page cache doesn't contain an entry at 451a77d19f4SMatthew Wilcox * that index, add a locked empty entry. 452ac401cc7SJan Kara * 4533159f943SMatthew Wilcox * When requesting an entry with size DAX_PMD, grab_mapping_entry() will 454b15cd800SMatthew Wilcox * either return that locked entry or will return VM_FAULT_FALLBACK. 455b15cd800SMatthew Wilcox * This will happen if there are any PTE entries within the PMD range 456b15cd800SMatthew Wilcox * that we are requesting. 457642261acSRoss Zwisler * 458b15cd800SMatthew Wilcox * We always favor PTE entries over PMD entries. There isn't a flow where we 459b15cd800SMatthew Wilcox * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD 460b15cd800SMatthew Wilcox * insertion will fail if it finds any PTE entries already in the tree, and a 461b15cd800SMatthew Wilcox * PTE insertion will cause an existing PMD entry to be unmapped and 462b15cd800SMatthew Wilcox * downgraded to PTE entries. This happens for both PMD zero pages as 463b15cd800SMatthew Wilcox * well as PMD empty entries. 464642261acSRoss Zwisler * 465b15cd800SMatthew Wilcox * The exception to this downgrade path is for PMD entries that have 466b15cd800SMatthew Wilcox * real storage backing them. We will leave these real PMD entries in 467b15cd800SMatthew Wilcox * the tree, and PTE writes will simply dirty the entire PMD entry. 468642261acSRoss Zwisler * 469ac401cc7SJan Kara * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 470ac401cc7SJan Kara * persistent memory the benefit is doubtful. We can add that later if we can 471ac401cc7SJan Kara * show it helps. 472b15cd800SMatthew Wilcox * 473b15cd800SMatthew Wilcox * On error, this function does not return an ERR_PTR. Instead it returns 474b15cd800SMatthew Wilcox * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values 475b15cd800SMatthew Wilcox * overlap with xarray value entries. 476ac401cc7SJan Kara */ 477b15cd800SMatthew Wilcox static void *grab_mapping_entry(struct xa_state *xas, 478b15cd800SMatthew Wilcox struct address_space *mapping, unsigned long size_flag) 479ac401cc7SJan Kara { 480b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 481b15cd800SMatthew Wilcox bool pmd_downgrade = false; /* splitting PMD entry into PTE entries? */ 482b15cd800SMatthew Wilcox void *entry; 483ac401cc7SJan Kara 484b15cd800SMatthew Wilcox retry: 485b15cd800SMatthew Wilcox xas_lock_irq(xas); 486b15cd800SMatthew Wilcox entry = get_unlocked_entry(xas); 487642261acSRoss Zwisler 488b15cd800SMatthew Wilcox if (entry) { 4890e40de03SMatthew Wilcox if (!xa_is_value(entry)) { 490b15cd800SMatthew Wilcox xas_set_err(xas, EIO); 49191d25ba8SRoss Zwisler goto out_unlock; 49291d25ba8SRoss Zwisler } 49391d25ba8SRoss Zwisler 4943159f943SMatthew Wilcox if (size_flag & DAX_PMD) { 49591d25ba8SRoss Zwisler if (dax_is_pte_entry(entry)) { 496b15cd800SMatthew Wilcox put_unlocked_entry(xas, entry); 497b15cd800SMatthew Wilcox goto fallback; 498642261acSRoss Zwisler } 499642261acSRoss Zwisler } else { /* trying to grab a PTE entry */ 50091d25ba8SRoss Zwisler if (dax_is_pmd_entry(entry) && 501642261acSRoss Zwisler (dax_is_zero_entry(entry) || 502642261acSRoss Zwisler dax_is_empty_entry(entry))) { 503642261acSRoss Zwisler pmd_downgrade = true; 504642261acSRoss Zwisler } 505642261acSRoss Zwisler } 506642261acSRoss Zwisler } 507642261acSRoss Zwisler 508642261acSRoss Zwisler if (pmd_downgrade) { 509642261acSRoss Zwisler /* 510642261acSRoss Zwisler * Make sure 'entry' remains valid while we drop 511b93b0163SMatthew Wilcox * the i_pages lock. 512642261acSRoss Zwisler */ 513b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 514642261acSRoss Zwisler 515642261acSRoss Zwisler /* 516642261acSRoss Zwisler * Besides huge zero pages the only other thing that gets 517642261acSRoss Zwisler * downgraded are empty entries which don't need to be 518642261acSRoss Zwisler * unmapped. 519642261acSRoss Zwisler */ 520b15cd800SMatthew Wilcox if (dax_is_zero_entry(entry)) { 521b15cd800SMatthew Wilcox xas_unlock_irq(xas); 522b15cd800SMatthew Wilcox unmap_mapping_pages(mapping, 523b15cd800SMatthew Wilcox xas->xa_index & ~PG_PMD_COLOUR, 524977fbdcdSMatthew Wilcox PG_PMD_NR, false); 525b15cd800SMatthew Wilcox xas_reset(xas); 526b15cd800SMatthew Wilcox xas_lock_irq(xas); 527e11f8b7bSRoss Zwisler } 528e11f8b7bSRoss Zwisler 529d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, false); 530b15cd800SMatthew Wilcox xas_store(xas, NULL); /* undo the PMD join */ 531b15cd800SMatthew Wilcox dax_wake_entry(xas, entry, true); 532642261acSRoss Zwisler mapping->nrexceptional--; 533b15cd800SMatthew Wilcox entry = NULL; 534b15cd800SMatthew Wilcox xas_set(xas, index); 535642261acSRoss Zwisler } 536642261acSRoss Zwisler 537b15cd800SMatthew Wilcox if (entry) { 538b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 539b15cd800SMatthew Wilcox } else { 540b15cd800SMatthew Wilcox entry = dax_make_entry(pfn_to_pfn_t(0), size_flag | DAX_EMPTY); 541b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 542b15cd800SMatthew Wilcox if (xas_error(xas)) 543b15cd800SMatthew Wilcox goto out_unlock; 544ac401cc7SJan Kara mapping->nrexceptional++; 545ac401cc7SJan Kara } 546b15cd800SMatthew Wilcox 547642261acSRoss Zwisler out_unlock: 548b15cd800SMatthew Wilcox xas_unlock_irq(xas); 549b15cd800SMatthew Wilcox if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM)) 550b15cd800SMatthew Wilcox goto retry; 551b15cd800SMatthew Wilcox if (xas->xa_node == XA_ERROR(-ENOMEM)) 552b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_OOM); 553b15cd800SMatthew Wilcox if (xas_error(xas)) 554b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_SIGBUS); 555e3ad61c6SRoss Zwisler return entry; 556b15cd800SMatthew Wilcox fallback: 557b15cd800SMatthew Wilcox xas_unlock_irq(xas); 558b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_FALLBACK); 559ac401cc7SJan Kara } 560ac401cc7SJan Kara 5615fac7408SDan Williams /** 5625fac7408SDan Williams * dax_layout_busy_page - find first pinned page in @mapping 5635fac7408SDan Williams * @mapping: address space to scan for a page with ref count > 1 5645fac7408SDan Williams * 5655fac7408SDan Williams * DAX requires ZONE_DEVICE mapped pages. These pages are never 5665fac7408SDan Williams * 'onlined' to the page allocator so they are considered idle when 5675fac7408SDan Williams * page->count == 1. A filesystem uses this interface to determine if 5685fac7408SDan Williams * any page in the mapping is busy, i.e. for DMA, or other 5695fac7408SDan Williams * get_user_pages() usages. 5705fac7408SDan Williams * 5715fac7408SDan Williams * It is expected that the filesystem is holding locks to block the 5725fac7408SDan Williams * establishment of new mappings in this address_space. I.e. it expects 5735fac7408SDan Williams * to be able to run unmap_mapping_range() and subsequently not race 5745fac7408SDan Williams * mapping_mapped() becoming true. 5755fac7408SDan Williams */ 5765fac7408SDan Williams struct page *dax_layout_busy_page(struct address_space *mapping) 5775fac7408SDan Williams { 578084a8990SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, 0); 579084a8990SMatthew Wilcox void *entry; 580084a8990SMatthew Wilcox unsigned int scanned = 0; 5815fac7408SDan Williams struct page *page = NULL; 5825fac7408SDan Williams 5835fac7408SDan Williams /* 5845fac7408SDan Williams * In the 'limited' case get_user_pages() for dax is disabled. 5855fac7408SDan Williams */ 5865fac7408SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 5875fac7408SDan Williams return NULL; 5885fac7408SDan Williams 5895fac7408SDan Williams if (!dax_mapping(mapping) || !mapping_mapped(mapping)) 5905fac7408SDan Williams return NULL; 5915fac7408SDan Williams 5925fac7408SDan Williams /* 5935fac7408SDan Williams * If we race get_user_pages_fast() here either we'll see the 594084a8990SMatthew Wilcox * elevated page count in the iteration and wait, or 5955fac7408SDan Williams * get_user_pages_fast() will see that the page it took a reference 5965fac7408SDan Williams * against is no longer mapped in the page tables and bail to the 5975fac7408SDan Williams * get_user_pages() slow path. The slow path is protected by 5985fac7408SDan Williams * pte_lock() and pmd_lock(). New references are not taken without 5995fac7408SDan Williams * holding those locks, and unmap_mapping_range() will not zero the 6005fac7408SDan Williams * pte or pmd without holding the respective lock, so we are 6015fac7408SDan Williams * guaranteed to either see new references or prevent new 6025fac7408SDan Williams * references from being established. 6035fac7408SDan Williams */ 6045fac7408SDan Williams unmap_mapping_range(mapping, 0, 0, 1); 6055fac7408SDan Williams 606084a8990SMatthew Wilcox xas_lock_irq(&xas); 607084a8990SMatthew Wilcox xas_for_each(&xas, entry, ULONG_MAX) { 608084a8990SMatthew Wilcox if (WARN_ON_ONCE(!xa_is_value(entry))) 6095fac7408SDan Williams continue; 610084a8990SMatthew Wilcox if (unlikely(dax_is_locked(entry))) 611084a8990SMatthew Wilcox entry = get_unlocked_entry(&xas); 6125fac7408SDan Williams if (entry) 6135fac7408SDan Williams page = dax_busy_page(entry); 614084a8990SMatthew Wilcox put_unlocked_entry(&xas, entry); 6155fac7408SDan Williams if (page) 6165fac7408SDan Williams break; 617084a8990SMatthew Wilcox if (++scanned % XA_CHECK_SCHED) 618084a8990SMatthew Wilcox continue; 619cdbf8897SRoss Zwisler 620084a8990SMatthew Wilcox xas_pause(&xas); 621084a8990SMatthew Wilcox xas_unlock_irq(&xas); 622084a8990SMatthew Wilcox cond_resched(); 623084a8990SMatthew Wilcox xas_lock_irq(&xas); 6245fac7408SDan Williams } 625084a8990SMatthew Wilcox xas_unlock_irq(&xas); 6265fac7408SDan Williams return page; 6275fac7408SDan Williams } 6285fac7408SDan Williams EXPORT_SYMBOL_GPL(dax_layout_busy_page); 6295fac7408SDan Williams 630a77d19f4SMatthew Wilcox static int __dax_invalidate_entry(struct address_space *mapping, 631c6dcf52cSJan Kara pgoff_t index, bool trunc) 632c6dcf52cSJan Kara { 63307f2d89cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, index); 634c6dcf52cSJan Kara int ret = 0; 635c6dcf52cSJan Kara void *entry; 636c6dcf52cSJan Kara 63707f2d89cSMatthew Wilcox xas_lock_irq(&xas); 63807f2d89cSMatthew Wilcox entry = get_unlocked_entry(&xas); 6393159f943SMatthew Wilcox if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 640c6dcf52cSJan Kara goto out; 641c6dcf52cSJan Kara if (!trunc && 64207f2d89cSMatthew Wilcox (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) || 64307f2d89cSMatthew Wilcox xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE))) 644c6dcf52cSJan Kara goto out; 645d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, trunc); 64607f2d89cSMatthew Wilcox xas_store(&xas, NULL); 647c6dcf52cSJan Kara mapping->nrexceptional--; 648c6dcf52cSJan Kara ret = 1; 649c6dcf52cSJan Kara out: 65007f2d89cSMatthew Wilcox put_unlocked_entry(&xas, entry); 65107f2d89cSMatthew Wilcox xas_unlock_irq(&xas); 652c6dcf52cSJan Kara return ret; 653c6dcf52cSJan Kara } 65407f2d89cSMatthew Wilcox 655ac401cc7SJan Kara /* 6563159f943SMatthew Wilcox * Delete DAX entry at @index from @mapping. Wait for it 6573159f943SMatthew Wilcox * to be unlocked before deleting it. 658ac401cc7SJan Kara */ 659ac401cc7SJan Kara int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 660ac401cc7SJan Kara { 661a77d19f4SMatthew Wilcox int ret = __dax_invalidate_entry(mapping, index, true); 662ac401cc7SJan Kara 663ac401cc7SJan Kara /* 664ac401cc7SJan Kara * This gets called from truncate / punch_hole path. As such, the caller 665ac401cc7SJan Kara * must hold locks protecting against concurrent modifications of the 666a77d19f4SMatthew Wilcox * page cache (usually fs-private i_mmap_sem for writing). Since the 6673159f943SMatthew Wilcox * caller has seen a DAX entry for this index, we better find it 668ac401cc7SJan Kara * at that index as well... 669ac401cc7SJan Kara */ 670c6dcf52cSJan Kara WARN_ON_ONCE(!ret); 671c6dcf52cSJan Kara return ret; 672ac401cc7SJan Kara } 673ac401cc7SJan Kara 674c6dcf52cSJan Kara /* 6753159f943SMatthew Wilcox * Invalidate DAX entry if it is clean. 676c6dcf52cSJan Kara */ 677c6dcf52cSJan Kara int dax_invalidate_mapping_entry_sync(struct address_space *mapping, 678c6dcf52cSJan Kara pgoff_t index) 679c6dcf52cSJan Kara { 680a77d19f4SMatthew Wilcox return __dax_invalidate_entry(mapping, index, false); 681ac401cc7SJan Kara } 682ac401cc7SJan Kara 683cccbce67SDan Williams static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev, 684cccbce67SDan Williams sector_t sector, size_t size, struct page *to, 685cccbce67SDan Williams unsigned long vaddr) 686f7ca90b1SMatthew Wilcox { 687cccbce67SDan Williams void *vto, *kaddr; 688cccbce67SDan Williams pgoff_t pgoff; 689cccbce67SDan Williams long rc; 690cccbce67SDan Williams int id; 691e2e05394SRoss Zwisler 692cccbce67SDan Williams rc = bdev_dax_pgoff(bdev, sector, size, &pgoff); 693cccbce67SDan Williams if (rc) 694cccbce67SDan Williams return rc; 695cccbce67SDan Williams 696cccbce67SDan Williams id = dax_read_lock(); 69786ed913bSHuaisheng Ye rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, NULL); 698cccbce67SDan Williams if (rc < 0) { 699cccbce67SDan Williams dax_read_unlock(id); 700cccbce67SDan Williams return rc; 701cccbce67SDan Williams } 702f7ca90b1SMatthew Wilcox vto = kmap_atomic(to); 703cccbce67SDan Williams copy_user_page(vto, (void __force *)kaddr, vaddr, to); 704f7ca90b1SMatthew Wilcox kunmap_atomic(vto); 705cccbce67SDan Williams dax_read_unlock(id); 706f7ca90b1SMatthew Wilcox return 0; 707f7ca90b1SMatthew Wilcox } 708f7ca90b1SMatthew Wilcox 709642261acSRoss Zwisler /* 710642261acSRoss Zwisler * By this point grab_mapping_entry() has ensured that we have a locked entry 711642261acSRoss Zwisler * of the appropriate size so we don't have to worry about downgrading PMDs to 712642261acSRoss Zwisler * PTEs. If we happen to be trying to insert a PTE and there is a PMD 713642261acSRoss Zwisler * already in the tree, we will skip the insertion and just dirty the PMD as 714642261acSRoss Zwisler * appropriate. 715642261acSRoss Zwisler */ 716b15cd800SMatthew Wilcox static void *dax_insert_entry(struct xa_state *xas, 717b15cd800SMatthew Wilcox struct address_space *mapping, struct vm_fault *vmf, 718b15cd800SMatthew Wilcox void *entry, pfn_t pfn, unsigned long flags, bool dirty) 7199973c98eSRoss Zwisler { 720b15cd800SMatthew Wilcox void *new_entry = dax_make_entry(pfn, flags); 7219973c98eSRoss Zwisler 722f5b7b748SJan Kara if (dirty) 7239973c98eSRoss Zwisler __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 7249973c98eSRoss Zwisler 7253159f943SMatthew Wilcox if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) { 726b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 72791d25ba8SRoss Zwisler /* we are replacing a zero page with block mapping */ 72891d25ba8SRoss Zwisler if (dax_is_pmd_entry(entry)) 729977fbdcdSMatthew Wilcox unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, 730977fbdcdSMatthew Wilcox PG_PMD_NR, false); 73191d25ba8SRoss Zwisler else /* pte entry */ 732b15cd800SMatthew Wilcox unmap_mapping_pages(mapping, index, 1, false); 733ac401cc7SJan Kara } 7349973c98eSRoss Zwisler 735b15cd800SMatthew Wilcox xas_reset(xas); 736b15cd800SMatthew Wilcox xas_lock_irq(xas); 737d2c997c0SDan Williams if (dax_entry_size(entry) != dax_entry_size(new_entry)) { 738d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, false); 73973449dafSDan Williams dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address); 740d2c997c0SDan Williams } 741642261acSRoss Zwisler 74291d25ba8SRoss Zwisler if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 743642261acSRoss Zwisler /* 744a77d19f4SMatthew Wilcox * Only swap our new entry into the page cache if the current 745642261acSRoss Zwisler * entry is a zero page or an empty entry. If a normal PTE or 746a77d19f4SMatthew Wilcox * PMD entry is already in the cache, we leave it alone. This 747642261acSRoss Zwisler * means that if we are trying to insert a PTE and the 748642261acSRoss Zwisler * existing entry is a PMD, we will just leave the PMD in the 749642261acSRoss Zwisler * tree and dirty it if necessary. 750642261acSRoss Zwisler */ 751b15cd800SMatthew Wilcox void *old = dax_lock_entry(xas, new_entry); 752b15cd800SMatthew Wilcox WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | 753b15cd800SMatthew Wilcox DAX_LOCKED)); 75491d25ba8SRoss Zwisler entry = new_entry; 755b15cd800SMatthew Wilcox } else { 756b15cd800SMatthew Wilcox xas_load(xas); /* Walk the xa_state */ 757ac401cc7SJan Kara } 75891d25ba8SRoss Zwisler 759f5b7b748SJan Kara if (dirty) 760b15cd800SMatthew Wilcox xas_set_mark(xas, PAGECACHE_TAG_DIRTY); 76191d25ba8SRoss Zwisler 762b15cd800SMatthew Wilcox xas_unlock_irq(xas); 76391d25ba8SRoss Zwisler return entry; 7649973c98eSRoss Zwisler } 7659973c98eSRoss Zwisler 766a77d19f4SMatthew Wilcox static inline 767a77d19f4SMatthew Wilcox unsigned long pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma) 7684b4bb46dSJan Kara { 7694b4bb46dSJan Kara unsigned long address; 7704b4bb46dSJan Kara 7714b4bb46dSJan Kara address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 7724b4bb46dSJan Kara VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 7734b4bb46dSJan Kara return address; 7744b4bb46dSJan Kara } 7754b4bb46dSJan Kara 7764b4bb46dSJan Kara /* Walk all mappings of a given index of a file and writeprotect them */ 777a77d19f4SMatthew Wilcox static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index, 778a77d19f4SMatthew Wilcox unsigned long pfn) 7794b4bb46dSJan Kara { 7804b4bb46dSJan Kara struct vm_area_struct *vma; 781f729c8c9SRoss Zwisler pte_t pte, *ptep = NULL; 782f729c8c9SRoss Zwisler pmd_t *pmdp = NULL; 7834b4bb46dSJan Kara spinlock_t *ptl; 7844b4bb46dSJan Kara 7854b4bb46dSJan Kara i_mmap_lock_read(mapping); 7864b4bb46dSJan Kara vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) { 787a4d1a885SJérôme Glisse unsigned long address, start, end; 7884b4bb46dSJan Kara 7894b4bb46dSJan Kara cond_resched(); 7904b4bb46dSJan Kara 7914b4bb46dSJan Kara if (!(vma->vm_flags & VM_SHARED)) 7924b4bb46dSJan Kara continue; 7934b4bb46dSJan Kara 7944b4bb46dSJan Kara address = pgoff_address(index, vma); 795a4d1a885SJérôme Glisse 796a4d1a885SJérôme Glisse /* 797a4d1a885SJérôme Glisse * Note because we provide start/end to follow_pte_pmd it will 798a4d1a885SJérôme Glisse * call mmu_notifier_invalidate_range_start() on our behalf 799a4d1a885SJérôme Glisse * before taking any lock. 800a4d1a885SJérôme Glisse */ 801a4d1a885SJérôme Glisse if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl)) 8024b4bb46dSJan Kara continue; 803f729c8c9SRoss Zwisler 8040f10851eSJérôme Glisse /* 8050f10851eSJérôme Glisse * No need to call mmu_notifier_invalidate_range() as we are 8060f10851eSJérôme Glisse * downgrading page table protection not changing it to point 8070f10851eSJérôme Glisse * to a new page. 8080f10851eSJérôme Glisse * 809ad56b738SMike Rapoport * See Documentation/vm/mmu_notifier.rst 8100f10851eSJérôme Glisse */ 811f729c8c9SRoss Zwisler if (pmdp) { 812f729c8c9SRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 813f729c8c9SRoss Zwisler pmd_t pmd; 814f729c8c9SRoss Zwisler 815f729c8c9SRoss Zwisler if (pfn != pmd_pfn(*pmdp)) 816f729c8c9SRoss Zwisler goto unlock_pmd; 817f6f37321SLinus Torvalds if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp)) 818f729c8c9SRoss Zwisler goto unlock_pmd; 819f729c8c9SRoss Zwisler 820f729c8c9SRoss Zwisler flush_cache_page(vma, address, pfn); 821f729c8c9SRoss Zwisler pmd = pmdp_huge_clear_flush(vma, address, pmdp); 822f729c8c9SRoss Zwisler pmd = pmd_wrprotect(pmd); 823f729c8c9SRoss Zwisler pmd = pmd_mkclean(pmd); 824f729c8c9SRoss Zwisler set_pmd_at(vma->vm_mm, address, pmdp, pmd); 825f729c8c9SRoss Zwisler unlock_pmd: 826f729c8c9SRoss Zwisler #endif 827ee190ca6SJan H. Schönherr spin_unlock(ptl); 828f729c8c9SRoss Zwisler } else { 8294b4bb46dSJan Kara if (pfn != pte_pfn(*ptep)) 830f729c8c9SRoss Zwisler goto unlock_pte; 8314b4bb46dSJan Kara if (!pte_dirty(*ptep) && !pte_write(*ptep)) 832f729c8c9SRoss Zwisler goto unlock_pte; 8334b4bb46dSJan Kara 8344b4bb46dSJan Kara flush_cache_page(vma, address, pfn); 8354b4bb46dSJan Kara pte = ptep_clear_flush(vma, address, ptep); 8364b4bb46dSJan Kara pte = pte_wrprotect(pte); 8374b4bb46dSJan Kara pte = pte_mkclean(pte); 8384b4bb46dSJan Kara set_pte_at(vma->vm_mm, address, ptep, pte); 839f729c8c9SRoss Zwisler unlock_pte: 8404b4bb46dSJan Kara pte_unmap_unlock(ptep, ptl); 841f729c8c9SRoss Zwisler } 8424b4bb46dSJan Kara 843a4d1a885SJérôme Glisse mmu_notifier_invalidate_range_end(vma->vm_mm, start, end); 8444b4bb46dSJan Kara } 8454b4bb46dSJan Kara i_mmap_unlock_read(mapping); 8464b4bb46dSJan Kara } 8474b4bb46dSJan Kara 8489fc747f6SMatthew Wilcox static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, 8499fc747f6SMatthew Wilcox struct address_space *mapping, void *entry) 8509973c98eSRoss Zwisler { 8513fe0791cSDan Williams unsigned long pfn; 8523fe0791cSDan Williams long ret = 0; 853cccbce67SDan Williams size_t size; 8549973c98eSRoss Zwisler 8559973c98eSRoss Zwisler /* 856a6abc2c0SJan Kara * A page got tagged dirty in DAX mapping? Something is seriously 857a6abc2c0SJan Kara * wrong. 8589973c98eSRoss Zwisler */ 8593159f943SMatthew Wilcox if (WARN_ON(!xa_is_value(entry))) 860a6abc2c0SJan Kara return -EIO; 8619973c98eSRoss Zwisler 8629fc747f6SMatthew Wilcox if (unlikely(dax_is_locked(entry))) { 8639fc747f6SMatthew Wilcox void *old_entry = entry; 8649fc747f6SMatthew Wilcox 8659fc747f6SMatthew Wilcox entry = get_unlocked_entry(xas); 8669fc747f6SMatthew Wilcox 867a6abc2c0SJan Kara /* Entry got punched out / reallocated? */ 8689fc747f6SMatthew Wilcox if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 869a6abc2c0SJan Kara goto put_unlocked; 870a6abc2c0SJan Kara /* 8719fc747f6SMatthew Wilcox * Entry got reallocated elsewhere? No need to writeback. 8729fc747f6SMatthew Wilcox * We have to compare pfns as we must not bail out due to 8739fc747f6SMatthew Wilcox * difference in lockbit or entry type. 874a6abc2c0SJan Kara */ 8759fc747f6SMatthew Wilcox if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) 876a6abc2c0SJan Kara goto put_unlocked; 877642261acSRoss Zwisler if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 878642261acSRoss Zwisler dax_is_zero_entry(entry))) { 8799973c98eSRoss Zwisler ret = -EIO; 880a6abc2c0SJan Kara goto put_unlocked; 8819973c98eSRoss Zwisler } 8829973c98eSRoss Zwisler 8839fc747f6SMatthew Wilcox /* Another fsync thread may have already done this entry */ 8849fc747f6SMatthew Wilcox if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) 885a6abc2c0SJan Kara goto put_unlocked; 8869fc747f6SMatthew Wilcox } 8879fc747f6SMatthew Wilcox 888a6abc2c0SJan Kara /* Lock the entry to serialize with page faults */ 8899fc747f6SMatthew Wilcox dax_lock_entry(xas, entry); 8909fc747f6SMatthew Wilcox 891a6abc2c0SJan Kara /* 892a6abc2c0SJan Kara * We can clear the tag now but we have to be careful so that concurrent 893a6abc2c0SJan Kara * dax_writeback_one() calls for the same index cannot finish before we 894a6abc2c0SJan Kara * actually flush the caches. This is achieved as the calls will look 895b93b0163SMatthew Wilcox * at the entry only under the i_pages lock and once they do that 896b93b0163SMatthew Wilcox * they will see the entry locked and wait for it to unlock. 897a6abc2c0SJan Kara */ 8989fc747f6SMatthew Wilcox xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); 8999fc747f6SMatthew Wilcox xas_unlock_irq(xas); 900a6abc2c0SJan Kara 901642261acSRoss Zwisler /* 902642261acSRoss Zwisler * Even if dax_writeback_mapping_range() was given a wbc->range_start 903642261acSRoss Zwisler * in the middle of a PMD, the 'index' we are given will be aligned to 9043fe0791cSDan Williams * the start index of the PMD, as will the pfn we pull from 'entry'. 9053fe0791cSDan Williams * This allows us to flush for PMD_SIZE and not have to worry about 9063fe0791cSDan Williams * partial PMD writebacks. 907642261acSRoss Zwisler */ 908a77d19f4SMatthew Wilcox pfn = dax_to_pfn(entry); 909a77d19f4SMatthew Wilcox size = PAGE_SIZE << dax_entry_order(entry); 910cccbce67SDan Williams 9119fc747f6SMatthew Wilcox dax_entry_mkclean(mapping, xas->xa_index, pfn); 9123fe0791cSDan Williams dax_flush(dax_dev, page_address(pfn_to_page(pfn)), size); 9134b4bb46dSJan Kara /* 9144b4bb46dSJan Kara * After we have flushed the cache, we can clear the dirty tag. There 9154b4bb46dSJan Kara * cannot be new dirty data in the pfn after the flush has completed as 9164b4bb46dSJan Kara * the pfn mappings are writeprotected and fault waits for mapping 9174b4bb46dSJan Kara * entry lock. 9184b4bb46dSJan Kara */ 9199fc747f6SMatthew Wilcox xas_reset(xas); 9209fc747f6SMatthew Wilcox xas_lock_irq(xas); 9219fc747f6SMatthew Wilcox xas_store(xas, entry); 9229fc747f6SMatthew Wilcox xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); 9239fc747f6SMatthew Wilcox dax_wake_entry(xas, entry, false); 9249fc747f6SMatthew Wilcox 9259fc747f6SMatthew Wilcox trace_dax_writeback_one(mapping->host, xas->xa_index, 9269fc747f6SMatthew Wilcox size >> PAGE_SHIFT); 9279973c98eSRoss Zwisler return ret; 9289973c98eSRoss Zwisler 929a6abc2c0SJan Kara put_unlocked: 9309fc747f6SMatthew Wilcox put_unlocked_entry(xas, entry); 9319973c98eSRoss Zwisler return ret; 9329973c98eSRoss Zwisler } 9339973c98eSRoss Zwisler 9349973c98eSRoss Zwisler /* 9359973c98eSRoss Zwisler * Flush the mapping to the persistent domain within the byte range of [start, 9369973c98eSRoss Zwisler * end]. This is required by data integrity operations to ensure file data is 9379973c98eSRoss Zwisler * on persistent storage prior to completion of the operation. 9389973c98eSRoss Zwisler */ 9397f6d5b52SRoss Zwisler int dax_writeback_mapping_range(struct address_space *mapping, 9407f6d5b52SRoss Zwisler struct block_device *bdev, struct writeback_control *wbc) 9419973c98eSRoss Zwisler { 9429fc747f6SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); 9439973c98eSRoss Zwisler struct inode *inode = mapping->host; 9449fc747f6SMatthew Wilcox pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; 945cccbce67SDan Williams struct dax_device *dax_dev; 9469fc747f6SMatthew Wilcox void *entry; 9479fc747f6SMatthew Wilcox int ret = 0; 9489fc747f6SMatthew Wilcox unsigned int scanned = 0; 9499973c98eSRoss Zwisler 9509973c98eSRoss Zwisler if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 9519973c98eSRoss Zwisler return -EIO; 9529973c98eSRoss Zwisler 9537f6d5b52SRoss Zwisler if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL) 9547f6d5b52SRoss Zwisler return 0; 9557f6d5b52SRoss Zwisler 956cccbce67SDan Williams dax_dev = dax_get_by_host(bdev->bd_disk->disk_name); 957cccbce67SDan Williams if (!dax_dev) 958cccbce67SDan Williams return -EIO; 959cccbce67SDan Williams 9609fc747f6SMatthew Wilcox trace_dax_writeback_range(inode, xas.xa_index, end_index); 9619973c98eSRoss Zwisler 9629fc747f6SMatthew Wilcox tag_pages_for_writeback(mapping, xas.xa_index, end_index); 963d14a3f48SRoss Zwisler 9649fc747f6SMatthew Wilcox xas_lock_irq(&xas); 9659fc747f6SMatthew Wilcox xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { 9669fc747f6SMatthew Wilcox ret = dax_writeback_one(&xas, dax_dev, mapping, entry); 967819ec6b9SJeff Layton if (ret < 0) { 968819ec6b9SJeff Layton mapping_set_error(mapping, ret); 9699fc747f6SMatthew Wilcox break; 970d14a3f48SRoss Zwisler } 9719fc747f6SMatthew Wilcox if (++scanned % XA_CHECK_SCHED) 9729fc747f6SMatthew Wilcox continue; 9739fc747f6SMatthew Wilcox 9749fc747f6SMatthew Wilcox xas_pause(&xas); 9759fc747f6SMatthew Wilcox xas_unlock_irq(&xas); 9769fc747f6SMatthew Wilcox cond_resched(); 9779fc747f6SMatthew Wilcox xas_lock_irq(&xas); 978d14a3f48SRoss Zwisler } 9799fc747f6SMatthew Wilcox xas_unlock_irq(&xas); 980cccbce67SDan Williams put_dax(dax_dev); 9819fc747f6SMatthew Wilcox trace_dax_writeback_range_done(inode, xas.xa_index, end_index); 9829fc747f6SMatthew Wilcox return ret; 9839973c98eSRoss Zwisler } 9849973c98eSRoss Zwisler EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 9859973c98eSRoss Zwisler 98631a6f1a6SJan Kara static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos) 987f7ca90b1SMatthew Wilcox { 988a3841f94SLinus Torvalds return (iomap->addr + (pos & PAGE_MASK) - iomap->offset) >> 9; 98931a6f1a6SJan Kara } 990f7ca90b1SMatthew Wilcox 9915e161e40SJan Kara static int dax_iomap_pfn(struct iomap *iomap, loff_t pos, size_t size, 9925e161e40SJan Kara pfn_t *pfnp) 9935e161e40SJan Kara { 9945e161e40SJan Kara const sector_t sector = dax_iomap_sector(iomap, pos); 9955e161e40SJan Kara pgoff_t pgoff; 9965e161e40SJan Kara int id, rc; 9975e161e40SJan Kara long length; 9985e161e40SJan Kara 9995e161e40SJan Kara rc = bdev_dax_pgoff(iomap->bdev, sector, size, &pgoff); 1000cccbce67SDan Williams if (rc) 1001cccbce67SDan Williams return rc; 1002cccbce67SDan Williams id = dax_read_lock(); 10035e161e40SJan Kara length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), 100486ed913bSHuaisheng Ye NULL, pfnp); 10055e161e40SJan Kara if (length < 0) { 10065e161e40SJan Kara rc = length; 10075e161e40SJan Kara goto out; 10085e161e40SJan Kara } 10095e161e40SJan Kara rc = -EINVAL; 10105e161e40SJan Kara if (PFN_PHYS(length) < size) 10115e161e40SJan Kara goto out; 10125e161e40SJan Kara if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) 10135e161e40SJan Kara goto out; 10145e161e40SJan Kara /* For larger pages we need devmap */ 10155e161e40SJan Kara if (length > 1 && !pfn_t_devmap(*pfnp)) 10165e161e40SJan Kara goto out; 10175e161e40SJan Kara rc = 0; 10185e161e40SJan Kara out: 1019cccbce67SDan Williams dax_read_unlock(id); 1020cccbce67SDan Williams return rc; 1021cccbce67SDan Williams } 1022f7ca90b1SMatthew Wilcox 10232f89dc12SJan Kara /* 102491d25ba8SRoss Zwisler * The user has performed a load from a hole in the file. Allocating a new 102591d25ba8SRoss Zwisler * page in the file would cause excessive storage usage for workloads with 102691d25ba8SRoss Zwisler * sparse files. Instead we insert a read-only mapping of the 4k zero page. 102791d25ba8SRoss Zwisler * If this page is ever written to we will re-fault and change the mapping to 102891d25ba8SRoss Zwisler * point to real DAX storage instead. 10292f89dc12SJan Kara */ 1030b15cd800SMatthew Wilcox static vm_fault_t dax_load_hole(struct xa_state *xas, 1031b15cd800SMatthew Wilcox struct address_space *mapping, void **entry, 1032e30331ffSRoss Zwisler struct vm_fault *vmf) 1033e30331ffSRoss Zwisler { 1034e30331ffSRoss Zwisler struct inode *inode = mapping->host; 103591d25ba8SRoss Zwisler unsigned long vaddr = vmf->address; 1036b90ca5ccSMatthew Wilcox pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); 1037b90ca5ccSMatthew Wilcox vm_fault_t ret; 1038e30331ffSRoss Zwisler 1039b15cd800SMatthew Wilcox *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 10403159f943SMatthew Wilcox DAX_ZERO_PAGE, false); 10413159f943SMatthew Wilcox 1042ab77dab4SSouptick Joarder ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); 1043e30331ffSRoss Zwisler trace_dax_load_hole(inode, vmf, ret); 1044e30331ffSRoss Zwisler return ret; 1045e30331ffSRoss Zwisler } 1046e30331ffSRoss Zwisler 10474b0228faSVishal Verma static bool dax_range_is_aligned(struct block_device *bdev, 10484b0228faSVishal Verma unsigned int offset, unsigned int length) 10494b0228faSVishal Verma { 10504b0228faSVishal Verma unsigned short sector_size = bdev_logical_block_size(bdev); 10514b0228faSVishal Verma 10524b0228faSVishal Verma if (!IS_ALIGNED(offset, sector_size)) 10534b0228faSVishal Verma return false; 10544b0228faSVishal Verma if (!IS_ALIGNED(length, sector_size)) 10554b0228faSVishal Verma return false; 10564b0228faSVishal Verma 10574b0228faSVishal Verma return true; 10584b0228faSVishal Verma } 10594b0228faSVishal Verma 1060cccbce67SDan Williams int __dax_zero_page_range(struct block_device *bdev, 1061cccbce67SDan Williams struct dax_device *dax_dev, sector_t sector, 1062cccbce67SDan Williams unsigned int offset, unsigned int size) 1063679c8bd3SChristoph Hellwig { 1064cccbce67SDan Williams if (dax_range_is_aligned(bdev, offset, size)) { 1065cccbce67SDan Williams sector_t start_sector = sector + (offset >> 9); 10664b0228faSVishal Verma 10674b0228faSVishal Verma return blkdev_issue_zeroout(bdev, start_sector, 106853ef7d0eSLinus Torvalds size >> 9, GFP_NOFS, 0); 10694b0228faSVishal Verma } else { 1070cccbce67SDan Williams pgoff_t pgoff; 1071cccbce67SDan Williams long rc, id; 1072cccbce67SDan Williams void *kaddr; 1073cccbce67SDan Williams 1074e84b83b9SDan Williams rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff); 1075cccbce67SDan Williams if (rc) 1076cccbce67SDan Williams return rc; 1077cccbce67SDan Williams 1078cccbce67SDan Williams id = dax_read_lock(); 107986ed913bSHuaisheng Ye rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr, NULL); 1080cccbce67SDan Williams if (rc < 0) { 1081cccbce67SDan Williams dax_read_unlock(id); 1082cccbce67SDan Williams return rc; 1083cccbce67SDan Williams } 108481f55870SDan Williams memset(kaddr + offset, 0, size); 1085c3ca015fSMikulas Patocka dax_flush(dax_dev, kaddr + offset, size); 1086cccbce67SDan Williams dax_read_unlock(id); 10874b0228faSVishal Verma } 1088679c8bd3SChristoph Hellwig return 0; 1089679c8bd3SChristoph Hellwig } 1090679c8bd3SChristoph Hellwig EXPORT_SYMBOL_GPL(__dax_zero_page_range); 1091679c8bd3SChristoph Hellwig 1092a254e568SChristoph Hellwig static loff_t 109311c59c92SRoss Zwisler dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data, 1094a254e568SChristoph Hellwig struct iomap *iomap) 1095a254e568SChristoph Hellwig { 1096cccbce67SDan Williams struct block_device *bdev = iomap->bdev; 1097cccbce67SDan Williams struct dax_device *dax_dev = iomap->dax_dev; 1098a254e568SChristoph Hellwig struct iov_iter *iter = data; 1099a254e568SChristoph Hellwig loff_t end = pos + length, done = 0; 1100a254e568SChristoph Hellwig ssize_t ret = 0; 1101a77d4786SDan Williams size_t xfer; 1102cccbce67SDan Williams int id; 1103a254e568SChristoph Hellwig 1104a254e568SChristoph Hellwig if (iov_iter_rw(iter) == READ) { 1105a254e568SChristoph Hellwig end = min(end, i_size_read(inode)); 1106a254e568SChristoph Hellwig if (pos >= end) 1107a254e568SChristoph Hellwig return 0; 1108a254e568SChristoph Hellwig 1109a254e568SChristoph Hellwig if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1110a254e568SChristoph Hellwig return iov_iter_zero(min(length, end - pos), iter); 1111a254e568SChristoph Hellwig } 1112a254e568SChristoph Hellwig 1113a254e568SChristoph Hellwig if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) 1114a254e568SChristoph Hellwig return -EIO; 1115a254e568SChristoph Hellwig 1116e3fce68cSJan Kara /* 1117e3fce68cSJan Kara * Write can allocate block for an area which has a hole page mapped 1118e3fce68cSJan Kara * into page tables. We have to tear down these mappings so that data 1119e3fce68cSJan Kara * written by write(2) is visible in mmap. 1120e3fce68cSJan Kara */ 1121cd656375SJan Kara if (iomap->flags & IOMAP_F_NEW) { 1122e3fce68cSJan Kara invalidate_inode_pages2_range(inode->i_mapping, 1123e3fce68cSJan Kara pos >> PAGE_SHIFT, 1124e3fce68cSJan Kara (end - 1) >> PAGE_SHIFT); 1125e3fce68cSJan Kara } 1126e3fce68cSJan Kara 1127cccbce67SDan Williams id = dax_read_lock(); 1128a254e568SChristoph Hellwig while (pos < end) { 1129a254e568SChristoph Hellwig unsigned offset = pos & (PAGE_SIZE - 1); 1130cccbce67SDan Williams const size_t size = ALIGN(length + offset, PAGE_SIZE); 1131cccbce67SDan Williams const sector_t sector = dax_iomap_sector(iomap, pos); 1132a254e568SChristoph Hellwig ssize_t map_len; 1133cccbce67SDan Williams pgoff_t pgoff; 1134cccbce67SDan Williams void *kaddr; 1135a254e568SChristoph Hellwig 1136d1908f52SMichal Hocko if (fatal_signal_pending(current)) { 1137d1908f52SMichal Hocko ret = -EINTR; 1138d1908f52SMichal Hocko break; 1139d1908f52SMichal Hocko } 1140d1908f52SMichal Hocko 1141cccbce67SDan Williams ret = bdev_dax_pgoff(bdev, sector, size, &pgoff); 1142cccbce67SDan Williams if (ret) 1143cccbce67SDan Williams break; 1144cccbce67SDan Williams 1145cccbce67SDan Williams map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), 114686ed913bSHuaisheng Ye &kaddr, NULL); 1147a254e568SChristoph Hellwig if (map_len < 0) { 1148a254e568SChristoph Hellwig ret = map_len; 1149a254e568SChristoph Hellwig break; 1150a254e568SChristoph Hellwig } 1151a254e568SChristoph Hellwig 1152cccbce67SDan Williams map_len = PFN_PHYS(map_len); 1153cccbce67SDan Williams kaddr += offset; 1154a254e568SChristoph Hellwig map_len -= offset; 1155a254e568SChristoph Hellwig if (map_len > end - pos) 1156a254e568SChristoph Hellwig map_len = end - pos; 1157a254e568SChristoph Hellwig 1158a2e050f5SRoss Zwisler /* 1159a2e050f5SRoss Zwisler * The userspace address for the memory copy has already been 1160a2e050f5SRoss Zwisler * validated via access_ok() in either vfs_read() or 1161a2e050f5SRoss Zwisler * vfs_write(), depending on which operation we are doing. 1162a2e050f5SRoss Zwisler */ 1163a254e568SChristoph Hellwig if (iov_iter_rw(iter) == WRITE) 1164a77d4786SDan Williams xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, 1165fec53774SDan Williams map_len, iter); 1166a254e568SChristoph Hellwig else 1167a77d4786SDan Williams xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, 1168b3a9a0c3SDan Williams map_len, iter); 1169a254e568SChristoph Hellwig 1170a77d4786SDan Williams pos += xfer; 1171a77d4786SDan Williams length -= xfer; 1172a77d4786SDan Williams done += xfer; 1173a77d4786SDan Williams 1174a77d4786SDan Williams if (xfer == 0) 1175a77d4786SDan Williams ret = -EFAULT; 1176a77d4786SDan Williams if (xfer < map_len) 1177a77d4786SDan Williams break; 1178a254e568SChristoph Hellwig } 1179cccbce67SDan Williams dax_read_unlock(id); 1180a254e568SChristoph Hellwig 1181a254e568SChristoph Hellwig return done ? done : ret; 1182a254e568SChristoph Hellwig } 1183a254e568SChristoph Hellwig 1184a254e568SChristoph Hellwig /** 118511c59c92SRoss Zwisler * dax_iomap_rw - Perform I/O to a DAX file 1186a254e568SChristoph Hellwig * @iocb: The control block for this I/O 1187a254e568SChristoph Hellwig * @iter: The addresses to do I/O from or to 1188a254e568SChristoph Hellwig * @ops: iomap ops passed from the file system 1189a254e568SChristoph Hellwig * 1190a254e568SChristoph Hellwig * This function performs read and write operations to directly mapped 1191a254e568SChristoph Hellwig * persistent memory. The callers needs to take care of read/write exclusion 1192a254e568SChristoph Hellwig * and evicting any page cache pages in the region under I/O. 1193a254e568SChristoph Hellwig */ 1194a254e568SChristoph Hellwig ssize_t 119511c59c92SRoss Zwisler dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 11968ff6daa1SChristoph Hellwig const struct iomap_ops *ops) 1197a254e568SChristoph Hellwig { 1198a254e568SChristoph Hellwig struct address_space *mapping = iocb->ki_filp->f_mapping; 1199a254e568SChristoph Hellwig struct inode *inode = mapping->host; 1200a254e568SChristoph Hellwig loff_t pos = iocb->ki_pos, ret = 0, done = 0; 1201a254e568SChristoph Hellwig unsigned flags = 0; 1202a254e568SChristoph Hellwig 1203168316dbSChristoph Hellwig if (iov_iter_rw(iter) == WRITE) { 1204168316dbSChristoph Hellwig lockdep_assert_held_exclusive(&inode->i_rwsem); 1205a254e568SChristoph Hellwig flags |= IOMAP_WRITE; 1206168316dbSChristoph Hellwig } else { 1207168316dbSChristoph Hellwig lockdep_assert_held(&inode->i_rwsem); 1208168316dbSChristoph Hellwig } 1209a254e568SChristoph Hellwig 1210a254e568SChristoph Hellwig while (iov_iter_count(iter)) { 1211a254e568SChristoph Hellwig ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops, 121211c59c92SRoss Zwisler iter, dax_iomap_actor); 1213a254e568SChristoph Hellwig if (ret <= 0) 1214a254e568SChristoph Hellwig break; 1215a254e568SChristoph Hellwig pos += ret; 1216a254e568SChristoph Hellwig done += ret; 1217a254e568SChristoph Hellwig } 1218a254e568SChristoph Hellwig 1219a254e568SChristoph Hellwig iocb->ki_pos += done; 1220a254e568SChristoph Hellwig return done ? done : ret; 1221a254e568SChristoph Hellwig } 122211c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_rw); 1223a7d73fe6SChristoph Hellwig 1224ab77dab4SSouptick Joarder static vm_fault_t dax_fault_return(int error) 12259f141d6eSJan Kara { 12269f141d6eSJan Kara if (error == 0) 12279f141d6eSJan Kara return VM_FAULT_NOPAGE; 12289f141d6eSJan Kara if (error == -ENOMEM) 12299f141d6eSJan Kara return VM_FAULT_OOM; 12309f141d6eSJan Kara return VM_FAULT_SIGBUS; 12319f141d6eSJan Kara } 12329f141d6eSJan Kara 1233aaa422c4SDan Williams /* 1234aaa422c4SDan Williams * MAP_SYNC on a dax mapping guarantees dirty metadata is 1235aaa422c4SDan Williams * flushed on write-faults (non-cow), but not read-faults. 1236aaa422c4SDan Williams */ 1237aaa422c4SDan Williams static bool dax_fault_is_synchronous(unsigned long flags, 1238aaa422c4SDan Williams struct vm_area_struct *vma, struct iomap *iomap) 1239aaa422c4SDan Williams { 1240aaa422c4SDan Williams return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) 1241aaa422c4SDan Williams && (iomap->flags & IOMAP_F_DIRTY); 1242aaa422c4SDan Williams } 1243aaa422c4SDan Williams 1244ab77dab4SSouptick Joarder static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, 1245c0b24625SJan Kara int *iomap_errp, const struct iomap_ops *ops) 1246a7d73fe6SChristoph Hellwig { 1247a0987ad5SJan Kara struct vm_area_struct *vma = vmf->vma; 1248a0987ad5SJan Kara struct address_space *mapping = vma->vm_file->f_mapping; 1249b15cd800SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, vmf->pgoff); 1250a7d73fe6SChristoph Hellwig struct inode *inode = mapping->host; 12511a29d85eSJan Kara unsigned long vaddr = vmf->address; 1252a7d73fe6SChristoph Hellwig loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT; 1253a7d73fe6SChristoph Hellwig struct iomap iomap = { 0 }; 12549484ab1bSJan Kara unsigned flags = IOMAP_FAULT; 1255a7d73fe6SChristoph Hellwig int error, major = 0; 1256d2c43ef1SJan Kara bool write = vmf->flags & FAULT_FLAG_WRITE; 1257caa51d26SJan Kara bool sync; 1258ab77dab4SSouptick Joarder vm_fault_t ret = 0; 1259a7d73fe6SChristoph Hellwig void *entry; 12601b5a1cb2SJan Kara pfn_t pfn; 1261a7d73fe6SChristoph Hellwig 1262ab77dab4SSouptick Joarder trace_dax_pte_fault(inode, vmf, ret); 1263a7d73fe6SChristoph Hellwig /* 1264a7d73fe6SChristoph Hellwig * Check whether offset isn't beyond end of file now. Caller is supposed 1265a7d73fe6SChristoph Hellwig * to hold locks serializing us with truncate / punch hole so this is 1266a7d73fe6SChristoph Hellwig * a reliable test. 1267a7d73fe6SChristoph Hellwig */ 1268a9c42b33SRoss Zwisler if (pos >= i_size_read(inode)) { 1269ab77dab4SSouptick Joarder ret = VM_FAULT_SIGBUS; 1270a9c42b33SRoss Zwisler goto out; 1271a9c42b33SRoss Zwisler } 1272a7d73fe6SChristoph Hellwig 1273d2c43ef1SJan Kara if (write && !vmf->cow_page) 1274a7d73fe6SChristoph Hellwig flags |= IOMAP_WRITE; 1275a7d73fe6SChristoph Hellwig 1276b15cd800SMatthew Wilcox entry = grab_mapping_entry(&xas, mapping, 0); 1277b15cd800SMatthew Wilcox if (xa_is_internal(entry)) { 1278b15cd800SMatthew Wilcox ret = xa_to_internal(entry); 127913e451fdSJan Kara goto out; 128013e451fdSJan Kara } 128113e451fdSJan Kara 1282a7d73fe6SChristoph Hellwig /* 1283e2093926SRoss Zwisler * It is possible, particularly with mixed reads & writes to private 1284e2093926SRoss Zwisler * mappings, that we have raced with a PMD fault that overlaps with 1285e2093926SRoss Zwisler * the PTE we need to set up. If so just return and the fault will be 1286e2093926SRoss Zwisler * retried. 1287e2093926SRoss Zwisler */ 1288e2093926SRoss Zwisler if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { 1289ab77dab4SSouptick Joarder ret = VM_FAULT_NOPAGE; 1290e2093926SRoss Zwisler goto unlock_entry; 1291e2093926SRoss Zwisler } 1292e2093926SRoss Zwisler 1293e2093926SRoss Zwisler /* 1294a7d73fe6SChristoph Hellwig * Note that we don't bother to use iomap_apply here: DAX required 1295a7d73fe6SChristoph Hellwig * the file system block size to be equal the page size, which means 1296a7d73fe6SChristoph Hellwig * that we never have to deal with more than a single extent here. 1297a7d73fe6SChristoph Hellwig */ 1298a7d73fe6SChristoph Hellwig error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap); 1299c0b24625SJan Kara if (iomap_errp) 1300c0b24625SJan Kara *iomap_errp = error; 1301a9c42b33SRoss Zwisler if (error) { 1302ab77dab4SSouptick Joarder ret = dax_fault_return(error); 130313e451fdSJan Kara goto unlock_entry; 1304a9c42b33SRoss Zwisler } 1305a7d73fe6SChristoph Hellwig if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) { 130613e451fdSJan Kara error = -EIO; /* fs corruption? */ 130713e451fdSJan Kara goto error_finish_iomap; 1308a7d73fe6SChristoph Hellwig } 1309a7d73fe6SChristoph Hellwig 1310a7d73fe6SChristoph Hellwig if (vmf->cow_page) { 131131a6f1a6SJan Kara sector_t sector = dax_iomap_sector(&iomap, pos); 131231a6f1a6SJan Kara 1313a7d73fe6SChristoph Hellwig switch (iomap.type) { 1314a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 1315a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1316a7d73fe6SChristoph Hellwig clear_user_highpage(vmf->cow_page, vaddr); 1317a7d73fe6SChristoph Hellwig break; 1318a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1319cccbce67SDan Williams error = copy_user_dax(iomap.bdev, iomap.dax_dev, 1320cccbce67SDan Williams sector, PAGE_SIZE, vmf->cow_page, vaddr); 1321a7d73fe6SChristoph Hellwig break; 1322a7d73fe6SChristoph Hellwig default: 1323a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1324a7d73fe6SChristoph Hellwig error = -EIO; 1325a7d73fe6SChristoph Hellwig break; 1326a7d73fe6SChristoph Hellwig } 1327a7d73fe6SChristoph Hellwig 1328a7d73fe6SChristoph Hellwig if (error) 132913e451fdSJan Kara goto error_finish_iomap; 1330b1aa812bSJan Kara 1331b1aa812bSJan Kara __SetPageUptodate(vmf->cow_page); 1332ab77dab4SSouptick Joarder ret = finish_fault(vmf); 1333ab77dab4SSouptick Joarder if (!ret) 1334ab77dab4SSouptick Joarder ret = VM_FAULT_DONE_COW; 133513e451fdSJan Kara goto finish_iomap; 1336a7d73fe6SChristoph Hellwig } 1337a7d73fe6SChristoph Hellwig 1338aaa422c4SDan Williams sync = dax_fault_is_synchronous(flags, vma, &iomap); 1339caa51d26SJan Kara 1340a7d73fe6SChristoph Hellwig switch (iomap.type) { 1341a7d73fe6SChristoph Hellwig case IOMAP_MAPPED: 1342a7d73fe6SChristoph Hellwig if (iomap.flags & IOMAP_F_NEW) { 1343a7d73fe6SChristoph Hellwig count_vm_event(PGMAJFAULT); 1344a0987ad5SJan Kara count_memcg_event_mm(vma->vm_mm, PGMAJFAULT); 1345a7d73fe6SChristoph Hellwig major = VM_FAULT_MAJOR; 1346a7d73fe6SChristoph Hellwig } 13471b5a1cb2SJan Kara error = dax_iomap_pfn(&iomap, pos, PAGE_SIZE, &pfn); 13481b5a1cb2SJan Kara if (error < 0) 13491b5a1cb2SJan Kara goto error_finish_iomap; 13501b5a1cb2SJan Kara 1351b15cd800SMatthew Wilcox entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, 1352caa51d26SJan Kara 0, write && !sync); 13531b5a1cb2SJan Kara 1354caa51d26SJan Kara /* 1355caa51d26SJan Kara * If we are doing synchronous page fault and inode needs fsync, 1356caa51d26SJan Kara * we can insert PTE into page tables only after that happens. 1357caa51d26SJan Kara * Skip insertion for now and return the pfn so that caller can 1358caa51d26SJan Kara * insert it after fsync is done. 1359caa51d26SJan Kara */ 1360caa51d26SJan Kara if (sync) { 1361caa51d26SJan Kara if (WARN_ON_ONCE(!pfnp)) { 1362caa51d26SJan Kara error = -EIO; 1363caa51d26SJan Kara goto error_finish_iomap; 1364caa51d26SJan Kara } 1365caa51d26SJan Kara *pfnp = pfn; 1366ab77dab4SSouptick Joarder ret = VM_FAULT_NEEDDSYNC | major; 1367caa51d26SJan Kara goto finish_iomap; 1368caa51d26SJan Kara } 13691b5a1cb2SJan Kara trace_dax_insert_mapping(inode, vmf, entry); 13701b5a1cb2SJan Kara if (write) 1371ab77dab4SSouptick Joarder ret = vmf_insert_mixed_mkwrite(vma, vaddr, pfn); 13721b5a1cb2SJan Kara else 1373ab77dab4SSouptick Joarder ret = vmf_insert_mixed(vma, vaddr, pfn); 13741b5a1cb2SJan Kara 1375ab77dab4SSouptick Joarder goto finish_iomap; 1376a7d73fe6SChristoph Hellwig case IOMAP_UNWRITTEN: 1377a7d73fe6SChristoph Hellwig case IOMAP_HOLE: 1378d2c43ef1SJan Kara if (!write) { 1379b15cd800SMatthew Wilcox ret = dax_load_hole(&xas, mapping, &entry, vmf); 138013e451fdSJan Kara goto finish_iomap; 13811550290bSRoss Zwisler } 1382a7d73fe6SChristoph Hellwig /*FALLTHRU*/ 1383a7d73fe6SChristoph Hellwig default: 1384a7d73fe6SChristoph Hellwig WARN_ON_ONCE(1); 1385a7d73fe6SChristoph Hellwig error = -EIO; 1386a7d73fe6SChristoph Hellwig break; 1387a7d73fe6SChristoph Hellwig } 1388a7d73fe6SChristoph Hellwig 138913e451fdSJan Kara error_finish_iomap: 1390ab77dab4SSouptick Joarder ret = dax_fault_return(error); 13919f141d6eSJan Kara finish_iomap: 13929f141d6eSJan Kara if (ops->iomap_end) { 13939f141d6eSJan Kara int copied = PAGE_SIZE; 13949f141d6eSJan Kara 1395ab77dab4SSouptick Joarder if (ret & VM_FAULT_ERROR) 13969f141d6eSJan Kara copied = 0; 13979f141d6eSJan Kara /* 13989f141d6eSJan Kara * The fault is done by now and there's no way back (other 13999f141d6eSJan Kara * thread may be already happily using PTE we have installed). 14009f141d6eSJan Kara * Just ignore error from ->iomap_end since we cannot do much 14019f141d6eSJan Kara * with it. 14029f141d6eSJan Kara */ 14039f141d6eSJan Kara ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap); 14041550290bSRoss Zwisler } 140513e451fdSJan Kara unlock_entry: 1406b15cd800SMatthew Wilcox dax_unlock_entry(&xas, entry); 1407a9c42b33SRoss Zwisler out: 1408ab77dab4SSouptick Joarder trace_dax_pte_fault_done(inode, vmf, ret); 1409ab77dab4SSouptick Joarder return ret | major; 1410a7d73fe6SChristoph Hellwig } 1411642261acSRoss Zwisler 1412642261acSRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 1413b15cd800SMatthew Wilcox static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1414b15cd800SMatthew Wilcox struct iomap *iomap, void **entry) 1415642261acSRoss Zwisler { 1416f4200391SDave Jiang struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1417f4200391SDave Jiang unsigned long pmd_addr = vmf->address & PMD_MASK; 1418653b2ea3SRoss Zwisler struct inode *inode = mapping->host; 1419642261acSRoss Zwisler struct page *zero_page; 1420642261acSRoss Zwisler spinlock_t *ptl; 1421642261acSRoss Zwisler pmd_t pmd_entry; 14223fe0791cSDan Williams pfn_t pfn; 1423642261acSRoss Zwisler 1424f4200391SDave Jiang zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); 1425642261acSRoss Zwisler 1426642261acSRoss Zwisler if (unlikely(!zero_page)) 1427653b2ea3SRoss Zwisler goto fallback; 1428642261acSRoss Zwisler 14293fe0791cSDan Williams pfn = page_to_pfn_t(zero_page); 1430b15cd800SMatthew Wilcox *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 14313159f943SMatthew Wilcox DAX_PMD | DAX_ZERO_PAGE, false); 1432642261acSRoss Zwisler 1433f4200391SDave Jiang ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1434f4200391SDave Jiang if (!pmd_none(*(vmf->pmd))) { 1435642261acSRoss Zwisler spin_unlock(ptl); 1436653b2ea3SRoss Zwisler goto fallback; 1437642261acSRoss Zwisler } 1438642261acSRoss Zwisler 1439f4200391SDave Jiang pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); 1440642261acSRoss Zwisler pmd_entry = pmd_mkhuge(pmd_entry); 1441f4200391SDave Jiang set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); 1442642261acSRoss Zwisler spin_unlock(ptl); 1443b15cd800SMatthew Wilcox trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); 1444642261acSRoss Zwisler return VM_FAULT_NOPAGE; 1445653b2ea3SRoss Zwisler 1446653b2ea3SRoss Zwisler fallback: 1447b15cd800SMatthew Wilcox trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); 1448642261acSRoss Zwisler return VM_FAULT_FALLBACK; 1449642261acSRoss Zwisler } 1450642261acSRoss Zwisler 1451ab77dab4SSouptick Joarder static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1452a2d58167SDave Jiang const struct iomap_ops *ops) 1453642261acSRoss Zwisler { 1454f4200391SDave Jiang struct vm_area_struct *vma = vmf->vma; 1455642261acSRoss Zwisler struct address_space *mapping = vma->vm_file->f_mapping; 1456b15cd800SMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); 1457d8a849e1SDave Jiang unsigned long pmd_addr = vmf->address & PMD_MASK; 1458d8a849e1SDave Jiang bool write = vmf->flags & FAULT_FLAG_WRITE; 1459caa51d26SJan Kara bool sync; 14609484ab1bSJan Kara unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT; 1461642261acSRoss Zwisler struct inode *inode = mapping->host; 1462ab77dab4SSouptick Joarder vm_fault_t result = VM_FAULT_FALLBACK; 1463642261acSRoss Zwisler struct iomap iomap = { 0 }; 1464b15cd800SMatthew Wilcox pgoff_t max_pgoff; 1465642261acSRoss Zwisler void *entry; 1466642261acSRoss Zwisler loff_t pos; 1467642261acSRoss Zwisler int error; 1468302a5e31SJan Kara pfn_t pfn; 1469642261acSRoss Zwisler 1470282a8e03SRoss Zwisler /* 1471282a8e03SRoss Zwisler * Check whether offset isn't beyond end of file now. Caller is 1472282a8e03SRoss Zwisler * supposed to hold locks serializing us with truncate / punch hole so 1473282a8e03SRoss Zwisler * this is a reliable test. 1474282a8e03SRoss Zwisler */ 1475957ac8c4SJeff Moyer max_pgoff = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 1476282a8e03SRoss Zwisler 1477f4200391SDave Jiang trace_dax_pmd_fault(inode, vmf, max_pgoff, 0); 1478282a8e03SRoss Zwisler 1479fffa281bSRoss Zwisler /* 1480fffa281bSRoss Zwisler * Make sure that the faulting address's PMD offset (color) matches 1481fffa281bSRoss Zwisler * the PMD offset from the start of the file. This is necessary so 1482fffa281bSRoss Zwisler * that a PMD range in the page table overlaps exactly with a PMD 1483a77d19f4SMatthew Wilcox * range in the page cache. 1484fffa281bSRoss Zwisler */ 1485fffa281bSRoss Zwisler if ((vmf->pgoff & PG_PMD_COLOUR) != 1486fffa281bSRoss Zwisler ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) 1487fffa281bSRoss Zwisler goto fallback; 1488fffa281bSRoss Zwisler 1489642261acSRoss Zwisler /* Fall back to PTEs if we're going to COW */ 1490642261acSRoss Zwisler if (write && !(vma->vm_flags & VM_SHARED)) 1491642261acSRoss Zwisler goto fallback; 1492642261acSRoss Zwisler 1493642261acSRoss Zwisler /* If the PMD would extend outside the VMA */ 1494642261acSRoss Zwisler if (pmd_addr < vma->vm_start) 1495642261acSRoss Zwisler goto fallback; 1496642261acSRoss Zwisler if ((pmd_addr + PMD_SIZE) > vma->vm_end) 1497642261acSRoss Zwisler goto fallback; 1498642261acSRoss Zwisler 1499b15cd800SMatthew Wilcox if (xas.xa_index >= max_pgoff) { 1500282a8e03SRoss Zwisler result = VM_FAULT_SIGBUS; 1501282a8e03SRoss Zwisler goto out; 1502282a8e03SRoss Zwisler } 1503642261acSRoss Zwisler 1504642261acSRoss Zwisler /* If the PMD would extend beyond the file size */ 1505b15cd800SMatthew Wilcox if ((xas.xa_index | PG_PMD_COLOUR) >= max_pgoff) 1506642261acSRoss Zwisler goto fallback; 1507642261acSRoss Zwisler 1508642261acSRoss Zwisler /* 1509b15cd800SMatthew Wilcox * grab_mapping_entry() will make sure we get an empty PMD entry, 1510b15cd800SMatthew Wilcox * a zero PMD entry or a DAX PMD. If it can't (because a PTE 1511b15cd800SMatthew Wilcox * entry is already in the array, for instance), it will return 1512b15cd800SMatthew Wilcox * VM_FAULT_FALLBACK. 15139f141d6eSJan Kara */ 1514b15cd800SMatthew Wilcox entry = grab_mapping_entry(&xas, mapping, DAX_PMD); 1515b15cd800SMatthew Wilcox if (xa_is_internal(entry)) { 1516b15cd800SMatthew Wilcox result = xa_to_internal(entry); 1517876f2946SRoss Zwisler goto fallback; 1518b15cd800SMatthew Wilcox } 1519876f2946SRoss Zwisler 1520876f2946SRoss Zwisler /* 1521e2093926SRoss Zwisler * It is possible, particularly with mixed reads & writes to private 1522e2093926SRoss Zwisler * mappings, that we have raced with a PTE fault that overlaps with 1523e2093926SRoss Zwisler * the PMD we need to set up. If so just return and the fault will be 1524e2093926SRoss Zwisler * retried. 1525e2093926SRoss Zwisler */ 1526e2093926SRoss Zwisler if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && 1527e2093926SRoss Zwisler !pmd_devmap(*vmf->pmd)) { 1528e2093926SRoss Zwisler result = 0; 1529e2093926SRoss Zwisler goto unlock_entry; 1530e2093926SRoss Zwisler } 1531e2093926SRoss Zwisler 1532e2093926SRoss Zwisler /* 1533876f2946SRoss Zwisler * Note that we don't use iomap_apply here. We aren't doing I/O, only 1534876f2946SRoss Zwisler * setting up a mapping, so really we're using iomap_begin() as a way 1535876f2946SRoss Zwisler * to look up our filesystem block. 1536876f2946SRoss Zwisler */ 1537b15cd800SMatthew Wilcox pos = (loff_t)xas.xa_index << PAGE_SHIFT; 1538876f2946SRoss Zwisler error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap); 1539876f2946SRoss Zwisler if (error) 1540876f2946SRoss Zwisler goto unlock_entry; 1541876f2946SRoss Zwisler 1542876f2946SRoss Zwisler if (iomap.offset + iomap.length < pos + PMD_SIZE) 15439f141d6eSJan Kara goto finish_iomap; 15449f141d6eSJan Kara 1545aaa422c4SDan Williams sync = dax_fault_is_synchronous(iomap_flags, vma, &iomap); 1546caa51d26SJan Kara 1547642261acSRoss Zwisler switch (iomap.type) { 1548642261acSRoss Zwisler case IOMAP_MAPPED: 1549302a5e31SJan Kara error = dax_iomap_pfn(&iomap, pos, PMD_SIZE, &pfn); 1550302a5e31SJan Kara if (error < 0) 1551302a5e31SJan Kara goto finish_iomap; 1552302a5e31SJan Kara 1553b15cd800SMatthew Wilcox entry = dax_insert_entry(&xas, mapping, vmf, entry, pfn, 15543159f943SMatthew Wilcox DAX_PMD, write && !sync); 1555302a5e31SJan Kara 1556caa51d26SJan Kara /* 1557caa51d26SJan Kara * If we are doing synchronous page fault and inode needs fsync, 1558caa51d26SJan Kara * we can insert PMD into page tables only after that happens. 1559caa51d26SJan Kara * Skip insertion for now and return the pfn so that caller can 1560caa51d26SJan Kara * insert it after fsync is done. 1561caa51d26SJan Kara */ 1562caa51d26SJan Kara if (sync) { 1563caa51d26SJan Kara if (WARN_ON_ONCE(!pfnp)) 1564caa51d26SJan Kara goto finish_iomap; 1565caa51d26SJan Kara *pfnp = pfn; 1566caa51d26SJan Kara result = VM_FAULT_NEEDDSYNC; 1567caa51d26SJan Kara goto finish_iomap; 1568caa51d26SJan Kara } 1569caa51d26SJan Kara 1570302a5e31SJan Kara trace_dax_pmd_insert_mapping(inode, vmf, PMD_SIZE, pfn, entry); 1571302a5e31SJan Kara result = vmf_insert_pfn_pmd(vma, vmf->address, vmf->pmd, pfn, 1572302a5e31SJan Kara write); 1573642261acSRoss Zwisler break; 1574642261acSRoss Zwisler case IOMAP_UNWRITTEN: 1575642261acSRoss Zwisler case IOMAP_HOLE: 1576642261acSRoss Zwisler if (WARN_ON_ONCE(write)) 1577876f2946SRoss Zwisler break; 1578b15cd800SMatthew Wilcox result = dax_pmd_load_hole(&xas, vmf, &iomap, &entry); 1579642261acSRoss Zwisler break; 1580642261acSRoss Zwisler default: 1581642261acSRoss Zwisler WARN_ON_ONCE(1); 1582642261acSRoss Zwisler break; 1583642261acSRoss Zwisler } 1584642261acSRoss Zwisler 15859f141d6eSJan Kara finish_iomap: 15869f141d6eSJan Kara if (ops->iomap_end) { 15879f141d6eSJan Kara int copied = PMD_SIZE; 15889f141d6eSJan Kara 15899f141d6eSJan Kara if (result == VM_FAULT_FALLBACK) 15909f141d6eSJan Kara copied = 0; 15919f141d6eSJan Kara /* 15929f141d6eSJan Kara * The fault is done by now and there's no way back (other 15939f141d6eSJan Kara * thread may be already happily using PMD we have installed). 15949f141d6eSJan Kara * Just ignore error from ->iomap_end since we cannot do much 15959f141d6eSJan Kara * with it. 15969f141d6eSJan Kara */ 15979f141d6eSJan Kara ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags, 15989f141d6eSJan Kara &iomap); 15999f141d6eSJan Kara } 1600876f2946SRoss Zwisler unlock_entry: 1601b15cd800SMatthew Wilcox dax_unlock_entry(&xas, entry); 1602642261acSRoss Zwisler fallback: 1603642261acSRoss Zwisler if (result == VM_FAULT_FALLBACK) { 1604d8a849e1SDave Jiang split_huge_pmd(vma, vmf->pmd, vmf->address); 1605642261acSRoss Zwisler count_vm_event(THP_FAULT_FALLBACK); 1606642261acSRoss Zwisler } 1607282a8e03SRoss Zwisler out: 1608f4200391SDave Jiang trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result); 1609642261acSRoss Zwisler return result; 1610642261acSRoss Zwisler } 1611a2d58167SDave Jiang #else 1612ab77dab4SSouptick Joarder static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 161301cddfe9SArnd Bergmann const struct iomap_ops *ops) 1614a2d58167SDave Jiang { 1615a2d58167SDave Jiang return VM_FAULT_FALLBACK; 1616a2d58167SDave Jiang } 1617642261acSRoss Zwisler #endif /* CONFIG_FS_DAX_PMD */ 1618a2d58167SDave Jiang 1619a2d58167SDave Jiang /** 1620a2d58167SDave Jiang * dax_iomap_fault - handle a page fault on a DAX file 1621a2d58167SDave Jiang * @vmf: The description of the fault 1622cec04e8cSJan Kara * @pe_size: Size of the page to fault in 16239a0dd422SJan Kara * @pfnp: PFN to insert for synchronous faults if fsync is required 1624c0b24625SJan Kara * @iomap_errp: Storage for detailed error code in case of error 1625cec04e8cSJan Kara * @ops: Iomap ops passed from the file system 1626a2d58167SDave Jiang * 1627a2d58167SDave Jiang * When a page fault occurs, filesystems may call this helper in 1628a2d58167SDave Jiang * their fault handler for DAX files. dax_iomap_fault() assumes the caller 1629a2d58167SDave Jiang * has done all the necessary locking for page fault to proceed 1630a2d58167SDave Jiang * successfully. 1631a2d58167SDave Jiang */ 1632ab77dab4SSouptick Joarder vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, 1633c0b24625SJan Kara pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) 1634a2d58167SDave Jiang { 1635c791ace1SDave Jiang switch (pe_size) { 1636c791ace1SDave Jiang case PE_SIZE_PTE: 1637c0b24625SJan Kara return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); 1638c791ace1SDave Jiang case PE_SIZE_PMD: 16399a0dd422SJan Kara return dax_iomap_pmd_fault(vmf, pfnp, ops); 1640a2d58167SDave Jiang default: 1641a2d58167SDave Jiang return VM_FAULT_FALLBACK; 1642a2d58167SDave Jiang } 1643a2d58167SDave Jiang } 1644a2d58167SDave Jiang EXPORT_SYMBOL_GPL(dax_iomap_fault); 164571eab6dfSJan Kara 1646a77d19f4SMatthew Wilcox /* 164771eab6dfSJan Kara * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables 164871eab6dfSJan Kara * @vmf: The description of the fault 164971eab6dfSJan Kara * @pfn: PFN to insert 1650cfc93c6cSMatthew Wilcox * @order: Order of entry to insert. 165171eab6dfSJan Kara * 1652a77d19f4SMatthew Wilcox * This function inserts a writeable PTE or PMD entry into the page tables 1653a77d19f4SMatthew Wilcox * for an mmaped DAX file. It also marks the page cache entry as dirty. 165471eab6dfSJan Kara */ 1655cfc93c6cSMatthew Wilcox static vm_fault_t 1656cfc93c6cSMatthew Wilcox dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) 165771eab6dfSJan Kara { 165871eab6dfSJan Kara struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1659cfc93c6cSMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); 1660cfc93c6cSMatthew Wilcox void *entry; 1661ab77dab4SSouptick Joarder vm_fault_t ret; 166271eab6dfSJan Kara 1663cfc93c6cSMatthew Wilcox xas_lock_irq(&xas); 1664cfc93c6cSMatthew Wilcox entry = get_unlocked_entry(&xas); 166571eab6dfSJan Kara /* Did we race with someone splitting entry or so? */ 166671eab6dfSJan Kara if (!entry || 1667cfc93c6cSMatthew Wilcox (order == 0 && !dax_is_pte_entry(entry)) || 16680e40de03SMatthew Wilcox (order == PMD_ORDER && !dax_is_pmd_entry(entry))) { 1669cfc93c6cSMatthew Wilcox put_unlocked_entry(&xas, entry); 1670cfc93c6cSMatthew Wilcox xas_unlock_irq(&xas); 167171eab6dfSJan Kara trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, 167271eab6dfSJan Kara VM_FAULT_NOPAGE); 167371eab6dfSJan Kara return VM_FAULT_NOPAGE; 167471eab6dfSJan Kara } 1675cfc93c6cSMatthew Wilcox xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); 1676cfc93c6cSMatthew Wilcox dax_lock_entry(&xas, entry); 1677cfc93c6cSMatthew Wilcox xas_unlock_irq(&xas); 1678cfc93c6cSMatthew Wilcox if (order == 0) 1679ab77dab4SSouptick Joarder ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 168071eab6dfSJan Kara #ifdef CONFIG_FS_DAX_PMD 1681cfc93c6cSMatthew Wilcox else if (order == PMD_ORDER) 1682ab77dab4SSouptick Joarder ret = vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, 168371eab6dfSJan Kara pfn, true); 168471eab6dfSJan Kara #endif 1685cfc93c6cSMatthew Wilcox else 1686ab77dab4SSouptick Joarder ret = VM_FAULT_FALLBACK; 1687cfc93c6cSMatthew Wilcox dax_unlock_entry(&xas, entry); 1688ab77dab4SSouptick Joarder trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); 1689ab77dab4SSouptick Joarder return ret; 169071eab6dfSJan Kara } 169171eab6dfSJan Kara 169271eab6dfSJan Kara /** 169371eab6dfSJan Kara * dax_finish_sync_fault - finish synchronous page fault 169471eab6dfSJan Kara * @vmf: The description of the fault 169571eab6dfSJan Kara * @pe_size: Size of entry to be inserted 169671eab6dfSJan Kara * @pfn: PFN to insert 169771eab6dfSJan Kara * 169871eab6dfSJan Kara * This function ensures that the file range touched by the page fault is 169971eab6dfSJan Kara * stored persistently on the media and handles inserting of appropriate page 170071eab6dfSJan Kara * table entry. 170171eab6dfSJan Kara */ 1702ab77dab4SSouptick Joarder vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, 1703ab77dab4SSouptick Joarder enum page_entry_size pe_size, pfn_t pfn) 170471eab6dfSJan Kara { 170571eab6dfSJan Kara int err; 170671eab6dfSJan Kara loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; 1707cfc93c6cSMatthew Wilcox unsigned int order = pe_order(pe_size); 1708cfc93c6cSMatthew Wilcox size_t len = PAGE_SIZE << order; 170971eab6dfSJan Kara 171071eab6dfSJan Kara err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); 171171eab6dfSJan Kara if (err) 171271eab6dfSJan Kara return VM_FAULT_SIGBUS; 1713cfc93c6cSMatthew Wilcox return dax_insert_pfn_mkwrite(vmf, pfn, order); 171471eab6dfSJan Kara } 171571eab6dfSJan Kara EXPORT_SYMBOL_GPL(dax_finish_sync_fault); 1716