12025cf9eSThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 2d475c634SMatthew Wilcox /* 3d475c634SMatthew Wilcox * fs/dax.c - Direct Access filesystem code 4d475c634SMatthew Wilcox * Copyright (c) 2013-2014 Intel Corporation 5d475c634SMatthew Wilcox * Author: Matthew Wilcox <matthew.r.wilcox@intel.com> 6d475c634SMatthew Wilcox * Author: Ross Zwisler <ross.zwisler@linux.intel.com> 7d475c634SMatthew Wilcox */ 8d475c634SMatthew Wilcox 9d475c634SMatthew Wilcox #include <linux/atomic.h> 10d475c634SMatthew Wilcox #include <linux/blkdev.h> 11d475c634SMatthew Wilcox #include <linux/buffer_head.h> 12d77e92e2SRoss Zwisler #include <linux/dax.h> 13d475c634SMatthew Wilcox #include <linux/fs.h> 14f7ca90b1SMatthew Wilcox #include <linux/highmem.h> 15f7ca90b1SMatthew Wilcox #include <linux/memcontrol.h> 16f7ca90b1SMatthew Wilcox #include <linux/mm.h> 17d475c634SMatthew Wilcox #include <linux/mutex.h> 189973c98eSRoss Zwisler #include <linux/pagevec.h> 19289c6aedSMatthew Wilcox #include <linux/sched.h> 20f361bf4aSIngo Molnar #include <linux/sched/signal.h> 21d475c634SMatthew Wilcox #include <linux/uio.h> 22f7ca90b1SMatthew Wilcox #include <linux/vmstat.h> 2334c0fd54SDan Williams #include <linux/pfn_t.h> 240e749e54SDan Williams #include <linux/sizes.h> 254b4bb46dSJan Kara #include <linux/mmu_notifier.h> 26a254e568SChristoph Hellwig #include <linux/iomap.h> 2706083a09SMuchun Song #include <linux/rmap.h> 2811cf9d86SAneesh Kumar K.V #include <asm/pgalloc.h> 29d475c634SMatthew Wilcox 30282a8e03SRoss Zwisler #define CREATE_TRACE_POINTS 31282a8e03SRoss Zwisler #include <trace/events/fs_dax.h> 32282a8e03SRoss Zwisler 33cfc93c6cSMatthew Wilcox static inline unsigned int pe_order(enum page_entry_size pe_size) 34cfc93c6cSMatthew Wilcox { 35cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PTE) 36cfc93c6cSMatthew Wilcox return PAGE_SHIFT - PAGE_SHIFT; 37cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PMD) 38cfc93c6cSMatthew Wilcox return PMD_SHIFT - PAGE_SHIFT; 39cfc93c6cSMatthew Wilcox if (pe_size == PE_SIZE_PUD) 40cfc93c6cSMatthew Wilcox return PUD_SHIFT - PAGE_SHIFT; 41cfc93c6cSMatthew Wilcox return ~0; 42cfc93c6cSMatthew Wilcox } 43cfc93c6cSMatthew Wilcox 44ac401cc7SJan Kara /* We choose 4096 entries - same as per-zone page wait tables */ 45ac401cc7SJan Kara #define DAX_WAIT_TABLE_BITS 12 46ac401cc7SJan Kara #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) 47ac401cc7SJan Kara 48917f3452SRoss Zwisler /* The 'colour' (ie low bits) within a PMD of a page offset. */ 49917f3452SRoss Zwisler #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) 50977fbdcdSMatthew Wilcox #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) 51917f3452SRoss Zwisler 52cfc93c6cSMatthew Wilcox /* The order of a PMD entry */ 53cfc93c6cSMatthew Wilcox #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT) 54cfc93c6cSMatthew Wilcox 55ce95ab0fSRoss Zwisler static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; 56ac401cc7SJan Kara 57ac401cc7SJan Kara static int __init init_dax_wait_table(void) 58ac401cc7SJan Kara { 59ac401cc7SJan Kara int i; 60ac401cc7SJan Kara 61ac401cc7SJan Kara for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) 62ac401cc7SJan Kara init_waitqueue_head(wait_table + i); 63ac401cc7SJan Kara return 0; 64ac401cc7SJan Kara } 65ac401cc7SJan Kara fs_initcall(init_dax_wait_table); 66ac401cc7SJan Kara 67527b19d0SRoss Zwisler /* 683159f943SMatthew Wilcox * DAX pagecache entries use XArray value entries so they can't be mistaken 693159f943SMatthew Wilcox * for pages. We use one bit for locking, one bit for the entry size (PMD) 703159f943SMatthew Wilcox * and two more to tell us if the entry is a zero page or an empty entry that 713159f943SMatthew Wilcox * is just used for locking. In total four special bits. 72527b19d0SRoss Zwisler * 73527b19d0SRoss Zwisler * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE 74527b19d0SRoss Zwisler * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem 75527b19d0SRoss Zwisler * block allocation. 76527b19d0SRoss Zwisler */ 773159f943SMatthew Wilcox #define DAX_SHIFT (4) 783159f943SMatthew Wilcox #define DAX_LOCKED (1UL << 0) 793159f943SMatthew Wilcox #define DAX_PMD (1UL << 1) 803159f943SMatthew Wilcox #define DAX_ZERO_PAGE (1UL << 2) 813159f943SMatthew Wilcox #define DAX_EMPTY (1UL << 3) 82527b19d0SRoss Zwisler 83a77d19f4SMatthew Wilcox static unsigned long dax_to_pfn(void *entry) 84527b19d0SRoss Zwisler { 853159f943SMatthew Wilcox return xa_to_value(entry) >> DAX_SHIFT; 86527b19d0SRoss Zwisler } 87527b19d0SRoss Zwisler 889f32d221SMatthew Wilcox static void *dax_make_entry(pfn_t pfn, unsigned long flags) 899f32d221SMatthew Wilcox { 909f32d221SMatthew Wilcox return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT)); 919f32d221SMatthew Wilcox } 929f32d221SMatthew Wilcox 93cfc93c6cSMatthew Wilcox static bool dax_is_locked(void *entry) 94cfc93c6cSMatthew Wilcox { 95cfc93c6cSMatthew Wilcox return xa_to_value(entry) & DAX_LOCKED; 96cfc93c6cSMatthew Wilcox } 97cfc93c6cSMatthew Wilcox 98a77d19f4SMatthew Wilcox static unsigned int dax_entry_order(void *entry) 99527b19d0SRoss Zwisler { 1003159f943SMatthew Wilcox if (xa_to_value(entry) & DAX_PMD) 101cfc93c6cSMatthew Wilcox return PMD_ORDER; 102527b19d0SRoss Zwisler return 0; 103527b19d0SRoss Zwisler } 104527b19d0SRoss Zwisler 105fda490d3SMatthew Wilcox static unsigned long dax_is_pmd_entry(void *entry) 106642261acSRoss Zwisler { 1073159f943SMatthew Wilcox return xa_to_value(entry) & DAX_PMD; 108642261acSRoss Zwisler } 109642261acSRoss Zwisler 110fda490d3SMatthew Wilcox static bool dax_is_pte_entry(void *entry) 111642261acSRoss Zwisler { 1123159f943SMatthew Wilcox return !(xa_to_value(entry) & DAX_PMD); 113642261acSRoss Zwisler } 114642261acSRoss Zwisler 115642261acSRoss Zwisler static int dax_is_zero_entry(void *entry) 116642261acSRoss Zwisler { 1173159f943SMatthew Wilcox return xa_to_value(entry) & DAX_ZERO_PAGE; 118642261acSRoss Zwisler } 119642261acSRoss Zwisler 120642261acSRoss Zwisler static int dax_is_empty_entry(void *entry) 121642261acSRoss Zwisler { 1223159f943SMatthew Wilcox return xa_to_value(entry) & DAX_EMPTY; 123642261acSRoss Zwisler } 124642261acSRoss Zwisler 125f7ca90b1SMatthew Wilcox /* 12623c84eb7SMatthew Wilcox (Oracle) * true if the entry that was found is of a smaller order than the entry 12723c84eb7SMatthew Wilcox (Oracle) * we were looking for 12823c84eb7SMatthew Wilcox (Oracle) */ 12923c84eb7SMatthew Wilcox (Oracle) static bool dax_is_conflict(void *entry) 13023c84eb7SMatthew Wilcox (Oracle) { 13123c84eb7SMatthew Wilcox (Oracle) return entry == XA_RETRY_ENTRY; 13223c84eb7SMatthew Wilcox (Oracle) } 13323c84eb7SMatthew Wilcox (Oracle) 13423c84eb7SMatthew Wilcox (Oracle) /* 135a77d19f4SMatthew Wilcox * DAX page cache entry locking 136ac401cc7SJan Kara */ 137ac401cc7SJan Kara struct exceptional_entry_key { 138ec4907ffSMatthew Wilcox struct xarray *xa; 13963e95b5cSRoss Zwisler pgoff_t entry_start; 140ac401cc7SJan Kara }; 141ac401cc7SJan Kara 142ac401cc7SJan Kara struct wait_exceptional_entry_queue { 143ac6424b9SIngo Molnar wait_queue_entry_t wait; 144ac401cc7SJan Kara struct exceptional_entry_key key; 145ac401cc7SJan Kara }; 146ac401cc7SJan Kara 147698ab77aSVivek Goyal /** 148698ab77aSVivek Goyal * enum dax_wake_mode: waitqueue wakeup behaviour 149698ab77aSVivek Goyal * @WAKE_ALL: wake all waiters in the waitqueue 150698ab77aSVivek Goyal * @WAKE_NEXT: wake only the first waiter in the waitqueue 151698ab77aSVivek Goyal */ 152698ab77aSVivek Goyal enum dax_wake_mode { 153698ab77aSVivek Goyal WAKE_ALL, 154698ab77aSVivek Goyal WAKE_NEXT, 155698ab77aSVivek Goyal }; 156698ab77aSVivek Goyal 157b15cd800SMatthew Wilcox static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas, 158b15cd800SMatthew Wilcox void *entry, struct exceptional_entry_key *key) 15963e95b5cSRoss Zwisler { 16063e95b5cSRoss Zwisler unsigned long hash; 161b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 16263e95b5cSRoss Zwisler 16363e95b5cSRoss Zwisler /* 16463e95b5cSRoss Zwisler * If 'entry' is a PMD, align the 'index' that we use for the wait 16563e95b5cSRoss Zwisler * queue to the start of that PMD. This ensures that all offsets in 16663e95b5cSRoss Zwisler * the range covered by the PMD map to the same bit lock. 16763e95b5cSRoss Zwisler */ 168642261acSRoss Zwisler if (dax_is_pmd_entry(entry)) 169917f3452SRoss Zwisler index &= ~PG_PMD_COLOUR; 170b15cd800SMatthew Wilcox key->xa = xas->xa; 17163e95b5cSRoss Zwisler key->entry_start = index; 17263e95b5cSRoss Zwisler 173b15cd800SMatthew Wilcox hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS); 17463e95b5cSRoss Zwisler return wait_table + hash; 17563e95b5cSRoss Zwisler } 17663e95b5cSRoss Zwisler 177ec4907ffSMatthew Wilcox static int wake_exceptional_entry_func(wait_queue_entry_t *wait, 178ec4907ffSMatthew Wilcox unsigned int mode, int sync, void *keyp) 179ac401cc7SJan Kara { 180ac401cc7SJan Kara struct exceptional_entry_key *key = keyp; 181ac401cc7SJan Kara struct wait_exceptional_entry_queue *ewait = 182ac401cc7SJan Kara container_of(wait, struct wait_exceptional_entry_queue, wait); 183ac401cc7SJan Kara 184ec4907ffSMatthew Wilcox if (key->xa != ewait->key.xa || 18563e95b5cSRoss Zwisler key->entry_start != ewait->key.entry_start) 186ac401cc7SJan Kara return 0; 187ac401cc7SJan Kara return autoremove_wake_function(wait, mode, sync, NULL); 188ac401cc7SJan Kara } 189ac401cc7SJan Kara 190ac401cc7SJan Kara /* 191b93b0163SMatthew Wilcox * @entry may no longer be the entry at the index in the mapping. 192b93b0163SMatthew Wilcox * The important information it's conveying is whether the entry at 193b93b0163SMatthew Wilcox * this index used to be a PMD entry. 194e30331ffSRoss Zwisler */ 195698ab77aSVivek Goyal static void dax_wake_entry(struct xa_state *xas, void *entry, 196698ab77aSVivek Goyal enum dax_wake_mode mode) 197e30331ffSRoss Zwisler { 198e30331ffSRoss Zwisler struct exceptional_entry_key key; 199e30331ffSRoss Zwisler wait_queue_head_t *wq; 200e30331ffSRoss Zwisler 201b15cd800SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &key); 202e30331ffSRoss Zwisler 203e30331ffSRoss Zwisler /* 204e30331ffSRoss Zwisler * Checking for locked entry and prepare_to_wait_exclusive() happens 205b93b0163SMatthew Wilcox * under the i_pages lock, ditto for entry handling in our callers. 206e30331ffSRoss Zwisler * So at this point all tasks that could have seen our entry locked 207e30331ffSRoss Zwisler * must be in the waitqueue and the following check will see them. 208e30331ffSRoss Zwisler */ 209e30331ffSRoss Zwisler if (waitqueue_active(wq)) 210698ab77aSVivek Goyal __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key); 211e30331ffSRoss Zwisler } 212e30331ffSRoss Zwisler 213cfc93c6cSMatthew Wilcox /* 214cfc93c6cSMatthew Wilcox * Look up entry in page cache, wait for it to become unlocked if it 215cfc93c6cSMatthew Wilcox * is a DAX entry and return it. The caller must subsequently call 216cfc93c6cSMatthew Wilcox * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry() 21723c84eb7SMatthew Wilcox (Oracle) * if it did. The entry returned may have a larger order than @order. 21823c84eb7SMatthew Wilcox (Oracle) * If @order is larger than the order of the entry found in i_pages, this 21923c84eb7SMatthew Wilcox (Oracle) * function returns a dax_is_conflict entry. 220cfc93c6cSMatthew Wilcox * 221cfc93c6cSMatthew Wilcox * Must be called with the i_pages lock held. 222cfc93c6cSMatthew Wilcox */ 22323c84eb7SMatthew Wilcox (Oracle) static void *get_unlocked_entry(struct xa_state *xas, unsigned int order) 224cfc93c6cSMatthew Wilcox { 225cfc93c6cSMatthew Wilcox void *entry; 226cfc93c6cSMatthew Wilcox struct wait_exceptional_entry_queue ewait; 227cfc93c6cSMatthew Wilcox wait_queue_head_t *wq; 228cfc93c6cSMatthew Wilcox 229cfc93c6cSMatthew Wilcox init_wait(&ewait.wait); 230cfc93c6cSMatthew Wilcox ewait.wait.func = wake_exceptional_entry_func; 231cfc93c6cSMatthew Wilcox 232cfc93c6cSMatthew Wilcox for (;;) { 2330e40de03SMatthew Wilcox entry = xas_find_conflict(xas); 2346370740eSDan Williams if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 2356370740eSDan Williams return entry; 23623c84eb7SMatthew Wilcox (Oracle) if (dax_entry_order(entry) < order) 23723c84eb7SMatthew Wilcox (Oracle) return XA_RETRY_ENTRY; 2386370740eSDan Williams if (!dax_is_locked(entry)) 239cfc93c6cSMatthew Wilcox return entry; 240cfc93c6cSMatthew Wilcox 241b15cd800SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &ewait.key); 242cfc93c6cSMatthew Wilcox prepare_to_wait_exclusive(wq, &ewait.wait, 243cfc93c6cSMatthew Wilcox TASK_UNINTERRUPTIBLE); 244cfc93c6cSMatthew Wilcox xas_unlock_irq(xas); 245cfc93c6cSMatthew Wilcox xas_reset(xas); 246cfc93c6cSMatthew Wilcox schedule(); 247cfc93c6cSMatthew Wilcox finish_wait(wq, &ewait.wait); 248cfc93c6cSMatthew Wilcox xas_lock_irq(xas); 249cfc93c6cSMatthew Wilcox } 250cfc93c6cSMatthew Wilcox } 251cfc93c6cSMatthew Wilcox 25255e56f06SMatthew Wilcox /* 25355e56f06SMatthew Wilcox * The only thing keeping the address space around is the i_pages lock 25455e56f06SMatthew Wilcox * (it's cycled in clear_inode() after removing the entries from i_pages) 25555e56f06SMatthew Wilcox * After we call xas_unlock_irq(), we cannot touch xas->xa. 25655e56f06SMatthew Wilcox */ 25755e56f06SMatthew Wilcox static void wait_entry_unlocked(struct xa_state *xas, void *entry) 25855e56f06SMatthew Wilcox { 25955e56f06SMatthew Wilcox struct wait_exceptional_entry_queue ewait; 26055e56f06SMatthew Wilcox wait_queue_head_t *wq; 26155e56f06SMatthew Wilcox 26255e56f06SMatthew Wilcox init_wait(&ewait.wait); 26355e56f06SMatthew Wilcox ewait.wait.func = wake_exceptional_entry_func; 26455e56f06SMatthew Wilcox 26555e56f06SMatthew Wilcox wq = dax_entry_waitqueue(xas, entry, &ewait.key); 266d8a70641SDan Williams /* 267d8a70641SDan Williams * Unlike get_unlocked_entry() there is no guarantee that this 268d8a70641SDan Williams * path ever successfully retrieves an unlocked entry before an 269d8a70641SDan Williams * inode dies. Perform a non-exclusive wait in case this path 270d8a70641SDan Williams * never successfully performs its own wake up. 271d8a70641SDan Williams */ 272d8a70641SDan Williams prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE); 27355e56f06SMatthew Wilcox xas_unlock_irq(xas); 27455e56f06SMatthew Wilcox schedule(); 27555e56f06SMatthew Wilcox finish_wait(wq, &ewait.wait); 27655e56f06SMatthew Wilcox } 27755e56f06SMatthew Wilcox 2784c3d043dSVivek Goyal static void put_unlocked_entry(struct xa_state *xas, void *entry, 2794c3d043dSVivek Goyal enum dax_wake_mode mode) 280cfc93c6cSMatthew Wilcox { 28161c30c98SJan Kara if (entry && !dax_is_conflict(entry)) 2824c3d043dSVivek Goyal dax_wake_entry(xas, entry, mode); 283cfc93c6cSMatthew Wilcox } 284cfc93c6cSMatthew Wilcox 285cfc93c6cSMatthew Wilcox /* 286cfc93c6cSMatthew Wilcox * We used the xa_state to get the entry, but then we locked the entry and 287cfc93c6cSMatthew Wilcox * dropped the xa_lock, so we know the xa_state is stale and must be reset 288cfc93c6cSMatthew Wilcox * before use. 289cfc93c6cSMatthew Wilcox */ 290cfc93c6cSMatthew Wilcox static void dax_unlock_entry(struct xa_state *xas, void *entry) 291cfc93c6cSMatthew Wilcox { 292cfc93c6cSMatthew Wilcox void *old; 293cfc93c6cSMatthew Wilcox 2947ae2ea7dSMatthew Wilcox BUG_ON(dax_is_locked(entry)); 295cfc93c6cSMatthew Wilcox xas_reset(xas); 296cfc93c6cSMatthew Wilcox xas_lock_irq(xas); 297cfc93c6cSMatthew Wilcox old = xas_store(xas, entry); 298cfc93c6cSMatthew Wilcox xas_unlock_irq(xas); 299cfc93c6cSMatthew Wilcox BUG_ON(!dax_is_locked(old)); 300698ab77aSVivek Goyal dax_wake_entry(xas, entry, WAKE_NEXT); 301cfc93c6cSMatthew Wilcox } 302cfc93c6cSMatthew Wilcox 303cfc93c6cSMatthew Wilcox /* 304cfc93c6cSMatthew Wilcox * Return: The entry stored at this location before it was locked. 305cfc93c6cSMatthew Wilcox */ 306cfc93c6cSMatthew Wilcox static void *dax_lock_entry(struct xa_state *xas, void *entry) 307cfc93c6cSMatthew Wilcox { 308cfc93c6cSMatthew Wilcox unsigned long v = xa_to_value(entry); 309cfc93c6cSMatthew Wilcox return xas_store(xas, xa_mk_value(v | DAX_LOCKED)); 310cfc93c6cSMatthew Wilcox } 311cfc93c6cSMatthew Wilcox 312d2c997c0SDan Williams static unsigned long dax_entry_size(void *entry) 313d2c997c0SDan Williams { 314d2c997c0SDan Williams if (dax_is_zero_entry(entry)) 315d2c997c0SDan Williams return 0; 316d2c997c0SDan Williams else if (dax_is_empty_entry(entry)) 317d2c997c0SDan Williams return 0; 318d2c997c0SDan Williams else if (dax_is_pmd_entry(entry)) 319d2c997c0SDan Williams return PMD_SIZE; 320d2c997c0SDan Williams else 321d2c997c0SDan Williams return PAGE_SIZE; 322d2c997c0SDan Williams } 323d2c997c0SDan Williams 324a77d19f4SMatthew Wilcox static unsigned long dax_end_pfn(void *entry) 325d2c997c0SDan Williams { 326a77d19f4SMatthew Wilcox return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; 327d2c997c0SDan Williams } 328d2c997c0SDan Williams 329d2c997c0SDan Williams /* 330d2c997c0SDan Williams * Iterate through all mapped pfns represented by an entry, i.e. skip 331d2c997c0SDan Williams * 'empty' and 'zero' entries. 332d2c997c0SDan Williams */ 333d2c997c0SDan Williams #define for_each_mapped_pfn(entry, pfn) \ 334a77d19f4SMatthew Wilcox for (pfn = dax_to_pfn(entry); \ 335a77d19f4SMatthew Wilcox pfn < dax_end_pfn(entry); pfn++) 336d2c997c0SDan Williams 33773449dafSDan Williams /* 33873449dafSDan Williams * TODO: for reflink+dax we need a way to associate a single page with 33973449dafSDan Williams * multiple address_space instances at different linear_page_index() 34073449dafSDan Williams * offsets. 34173449dafSDan Williams */ 34273449dafSDan Williams static void dax_associate_entry(void *entry, struct address_space *mapping, 34373449dafSDan Williams struct vm_area_struct *vma, unsigned long address) 344d2c997c0SDan Williams { 34573449dafSDan Williams unsigned long size = dax_entry_size(entry), pfn, index; 34673449dafSDan Williams int i = 0; 347d2c997c0SDan Williams 348d2c997c0SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 349d2c997c0SDan Williams return; 350d2c997c0SDan Williams 35173449dafSDan Williams index = linear_page_index(vma, address & ~(size - 1)); 352d2c997c0SDan Williams for_each_mapped_pfn(entry, pfn) { 353d2c997c0SDan Williams struct page *page = pfn_to_page(pfn); 354d2c997c0SDan Williams 355d2c997c0SDan Williams WARN_ON_ONCE(page->mapping); 356d2c997c0SDan Williams page->mapping = mapping; 35773449dafSDan Williams page->index = index + i++; 358d2c997c0SDan Williams } 359d2c997c0SDan Williams } 360d2c997c0SDan Williams 361d2c997c0SDan Williams static void dax_disassociate_entry(void *entry, struct address_space *mapping, 362d2c997c0SDan Williams bool trunc) 363d2c997c0SDan Williams { 364d2c997c0SDan Williams unsigned long pfn; 365d2c997c0SDan Williams 366d2c997c0SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 367d2c997c0SDan Williams return; 368d2c997c0SDan Williams 369d2c997c0SDan Williams for_each_mapped_pfn(entry, pfn) { 370d2c997c0SDan Williams struct page *page = pfn_to_page(pfn); 371d2c997c0SDan Williams 372d2c997c0SDan Williams WARN_ON_ONCE(trunc && page_ref_count(page) > 1); 373d2c997c0SDan Williams WARN_ON_ONCE(page->mapping && page->mapping != mapping); 374d2c997c0SDan Williams page->mapping = NULL; 37573449dafSDan Williams page->index = 0; 376d2c997c0SDan Williams } 377d2c997c0SDan Williams } 378d2c997c0SDan Williams 3795fac7408SDan Williams static struct page *dax_busy_page(void *entry) 3805fac7408SDan Williams { 3815fac7408SDan Williams unsigned long pfn; 3825fac7408SDan Williams 3835fac7408SDan Williams for_each_mapped_pfn(entry, pfn) { 3845fac7408SDan Williams struct page *page = pfn_to_page(pfn); 3855fac7408SDan Williams 3865fac7408SDan Williams if (page_ref_count(page) > 1) 3875fac7408SDan Williams return page; 3885fac7408SDan Williams } 3895fac7408SDan Williams return NULL; 3905fac7408SDan Williams } 3915fac7408SDan Williams 392c5bbd451SMatthew Wilcox /* 393c2e8021aSShiyang Ruan * dax_lock_page - Lock the DAX entry corresponding to a page 394c5bbd451SMatthew Wilcox * @page: The page whose entry we want to lock 395c5bbd451SMatthew Wilcox * 396c5bbd451SMatthew Wilcox * Context: Process context. 39727359fd6SMatthew Wilcox * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could 39827359fd6SMatthew Wilcox * not be locked. 399c5bbd451SMatthew Wilcox */ 40027359fd6SMatthew Wilcox dax_entry_t dax_lock_page(struct page *page) 401c2a7d2a1SDan Williams { 4029f32d221SMatthew Wilcox XA_STATE(xas, NULL, 0); 4039f32d221SMatthew Wilcox void *entry; 404c2a7d2a1SDan Williams 405c5bbd451SMatthew Wilcox /* Ensure page->mapping isn't freed while we look at it */ 406c5bbd451SMatthew Wilcox rcu_read_lock(); 407c2a7d2a1SDan Williams for (;;) { 4089f32d221SMatthew Wilcox struct address_space *mapping = READ_ONCE(page->mapping); 409c2a7d2a1SDan Williams 41027359fd6SMatthew Wilcox entry = NULL; 411c93db7bbSMatthew Wilcox if (!mapping || !dax_mapping(mapping)) 412c5bbd451SMatthew Wilcox break; 413c2a7d2a1SDan Williams 414c2a7d2a1SDan Williams /* 415c2a7d2a1SDan Williams * In the device-dax case there's no need to lock, a 416c2a7d2a1SDan Williams * struct dev_pagemap pin is sufficient to keep the 417c2a7d2a1SDan Williams * inode alive, and we assume we have dev_pagemap pin 418c2a7d2a1SDan Williams * otherwise we would not have a valid pfn_to_page() 419c2a7d2a1SDan Williams * translation. 420c2a7d2a1SDan Williams */ 42127359fd6SMatthew Wilcox entry = (void *)~0UL; 4229f32d221SMatthew Wilcox if (S_ISCHR(mapping->host->i_mode)) 423c5bbd451SMatthew Wilcox break; 424c2a7d2a1SDan Williams 4259f32d221SMatthew Wilcox xas.xa = &mapping->i_pages; 4269f32d221SMatthew Wilcox xas_lock_irq(&xas); 427c2a7d2a1SDan Williams if (mapping != page->mapping) { 4289f32d221SMatthew Wilcox xas_unlock_irq(&xas); 429c2a7d2a1SDan Williams continue; 430c2a7d2a1SDan Williams } 4319f32d221SMatthew Wilcox xas_set(&xas, page->index); 4329f32d221SMatthew Wilcox entry = xas_load(&xas); 4339f32d221SMatthew Wilcox if (dax_is_locked(entry)) { 434c5bbd451SMatthew Wilcox rcu_read_unlock(); 43555e56f06SMatthew Wilcox wait_entry_unlocked(&xas, entry); 436c5bbd451SMatthew Wilcox rcu_read_lock(); 437c2a7d2a1SDan Williams continue; 438c2a7d2a1SDan Williams } 4399f32d221SMatthew Wilcox dax_lock_entry(&xas, entry); 4409f32d221SMatthew Wilcox xas_unlock_irq(&xas); 441c5bbd451SMatthew Wilcox break; 4429f32d221SMatthew Wilcox } 443c5bbd451SMatthew Wilcox rcu_read_unlock(); 44427359fd6SMatthew Wilcox return (dax_entry_t)entry; 445c2a7d2a1SDan Williams } 446c2a7d2a1SDan Williams 44727359fd6SMatthew Wilcox void dax_unlock_page(struct page *page, dax_entry_t cookie) 448c2a7d2a1SDan Williams { 449c2a7d2a1SDan Williams struct address_space *mapping = page->mapping; 4509f32d221SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, page->index); 451c2a7d2a1SDan Williams 4529f32d221SMatthew Wilcox if (S_ISCHR(mapping->host->i_mode)) 453c2a7d2a1SDan Williams return; 454c2a7d2a1SDan Williams 45527359fd6SMatthew Wilcox dax_unlock_entry(&xas, (void *)cookie); 456c2a7d2a1SDan Williams } 457c2a7d2a1SDan Williams 458ac401cc7SJan Kara /* 459*2f437effSShiyang Ruan * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping 460*2f437effSShiyang Ruan * @mapping: the file's mapping whose entry we want to lock 461*2f437effSShiyang Ruan * @index: the offset within this file 462*2f437effSShiyang Ruan * @page: output the dax page corresponding to this dax entry 463*2f437effSShiyang Ruan * 464*2f437effSShiyang Ruan * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry 465*2f437effSShiyang Ruan * could not be locked. 466*2f437effSShiyang Ruan */ 467*2f437effSShiyang Ruan dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index, 468*2f437effSShiyang Ruan struct page **page) 469*2f437effSShiyang Ruan { 470*2f437effSShiyang Ruan XA_STATE(xas, NULL, 0); 471*2f437effSShiyang Ruan void *entry; 472*2f437effSShiyang Ruan 473*2f437effSShiyang Ruan rcu_read_lock(); 474*2f437effSShiyang Ruan for (;;) { 475*2f437effSShiyang Ruan entry = NULL; 476*2f437effSShiyang Ruan if (!dax_mapping(mapping)) 477*2f437effSShiyang Ruan break; 478*2f437effSShiyang Ruan 479*2f437effSShiyang Ruan xas.xa = &mapping->i_pages; 480*2f437effSShiyang Ruan xas_lock_irq(&xas); 481*2f437effSShiyang Ruan xas_set(&xas, index); 482*2f437effSShiyang Ruan entry = xas_load(&xas); 483*2f437effSShiyang Ruan if (dax_is_locked(entry)) { 484*2f437effSShiyang Ruan rcu_read_unlock(); 485*2f437effSShiyang Ruan wait_entry_unlocked(&xas, entry); 486*2f437effSShiyang Ruan rcu_read_lock(); 487*2f437effSShiyang Ruan continue; 488*2f437effSShiyang Ruan } 489*2f437effSShiyang Ruan if (!entry || 490*2f437effSShiyang Ruan dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 491*2f437effSShiyang Ruan /* 492*2f437effSShiyang Ruan * Because we are looking for entry from file's mapping 493*2f437effSShiyang Ruan * and index, so the entry may not be inserted for now, 494*2f437effSShiyang Ruan * or even a zero/empty entry. We don't think this is 495*2f437effSShiyang Ruan * an error case. So, return a special value and do 496*2f437effSShiyang Ruan * not output @page. 497*2f437effSShiyang Ruan */ 498*2f437effSShiyang Ruan entry = (void *)~0UL; 499*2f437effSShiyang Ruan } else { 500*2f437effSShiyang Ruan *page = pfn_to_page(dax_to_pfn(entry)); 501*2f437effSShiyang Ruan dax_lock_entry(&xas, entry); 502*2f437effSShiyang Ruan } 503*2f437effSShiyang Ruan xas_unlock_irq(&xas); 504*2f437effSShiyang Ruan break; 505*2f437effSShiyang Ruan } 506*2f437effSShiyang Ruan rcu_read_unlock(); 507*2f437effSShiyang Ruan return (dax_entry_t)entry; 508*2f437effSShiyang Ruan } 509*2f437effSShiyang Ruan 510*2f437effSShiyang Ruan void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index, 511*2f437effSShiyang Ruan dax_entry_t cookie) 512*2f437effSShiyang Ruan { 513*2f437effSShiyang Ruan XA_STATE(xas, &mapping->i_pages, index); 514*2f437effSShiyang Ruan 515*2f437effSShiyang Ruan if (cookie == ~0UL) 516*2f437effSShiyang Ruan return; 517*2f437effSShiyang Ruan 518*2f437effSShiyang Ruan dax_unlock_entry(&xas, (void *)cookie); 519*2f437effSShiyang Ruan } 520*2f437effSShiyang Ruan 521*2f437effSShiyang Ruan /* 522a77d19f4SMatthew Wilcox * Find page cache entry at given index. If it is a DAX entry, return it 523a77d19f4SMatthew Wilcox * with the entry locked. If the page cache doesn't contain an entry at 524a77d19f4SMatthew Wilcox * that index, add a locked empty entry. 525ac401cc7SJan Kara * 5263159f943SMatthew Wilcox * When requesting an entry with size DAX_PMD, grab_mapping_entry() will 527b15cd800SMatthew Wilcox * either return that locked entry or will return VM_FAULT_FALLBACK. 528b15cd800SMatthew Wilcox * This will happen if there are any PTE entries within the PMD range 529b15cd800SMatthew Wilcox * that we are requesting. 530642261acSRoss Zwisler * 531b15cd800SMatthew Wilcox * We always favor PTE entries over PMD entries. There isn't a flow where we 532b15cd800SMatthew Wilcox * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD 533b15cd800SMatthew Wilcox * insertion will fail if it finds any PTE entries already in the tree, and a 534b15cd800SMatthew Wilcox * PTE insertion will cause an existing PMD entry to be unmapped and 535b15cd800SMatthew Wilcox * downgraded to PTE entries. This happens for both PMD zero pages as 536b15cd800SMatthew Wilcox * well as PMD empty entries. 537642261acSRoss Zwisler * 538b15cd800SMatthew Wilcox * The exception to this downgrade path is for PMD entries that have 539b15cd800SMatthew Wilcox * real storage backing them. We will leave these real PMD entries in 540b15cd800SMatthew Wilcox * the tree, and PTE writes will simply dirty the entire PMD entry. 541642261acSRoss Zwisler * 542ac401cc7SJan Kara * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 543ac401cc7SJan Kara * persistent memory the benefit is doubtful. We can add that later if we can 544ac401cc7SJan Kara * show it helps. 545b15cd800SMatthew Wilcox * 546b15cd800SMatthew Wilcox * On error, this function does not return an ERR_PTR. Instead it returns 547b15cd800SMatthew Wilcox * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values 548b15cd800SMatthew Wilcox * overlap with xarray value entries. 549ac401cc7SJan Kara */ 550b15cd800SMatthew Wilcox static void *grab_mapping_entry(struct xa_state *xas, 55123c84eb7SMatthew Wilcox (Oracle) struct address_space *mapping, unsigned int order) 552ac401cc7SJan Kara { 553b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 5541a14e377SJan Kara bool pmd_downgrade; /* splitting PMD entry into PTE entries? */ 555b15cd800SMatthew Wilcox void *entry; 556ac401cc7SJan Kara 557b15cd800SMatthew Wilcox retry: 5581a14e377SJan Kara pmd_downgrade = false; 559b15cd800SMatthew Wilcox xas_lock_irq(xas); 56023c84eb7SMatthew Wilcox (Oracle) entry = get_unlocked_entry(xas, order); 561642261acSRoss Zwisler 562b15cd800SMatthew Wilcox if (entry) { 56323c84eb7SMatthew Wilcox (Oracle) if (dax_is_conflict(entry)) 56423c84eb7SMatthew Wilcox (Oracle) goto fallback; 5650e40de03SMatthew Wilcox if (!xa_is_value(entry)) { 56649688e65SHao Li xas_set_err(xas, -EIO); 56791d25ba8SRoss Zwisler goto out_unlock; 56891d25ba8SRoss Zwisler } 56991d25ba8SRoss Zwisler 57023c84eb7SMatthew Wilcox (Oracle) if (order == 0) { 57191d25ba8SRoss Zwisler if (dax_is_pmd_entry(entry) && 572642261acSRoss Zwisler (dax_is_zero_entry(entry) || 573642261acSRoss Zwisler dax_is_empty_entry(entry))) { 574642261acSRoss Zwisler pmd_downgrade = true; 575642261acSRoss Zwisler } 576642261acSRoss Zwisler } 577642261acSRoss Zwisler } 578642261acSRoss Zwisler 579642261acSRoss Zwisler if (pmd_downgrade) { 580642261acSRoss Zwisler /* 581642261acSRoss Zwisler * Make sure 'entry' remains valid while we drop 582b93b0163SMatthew Wilcox * the i_pages lock. 583642261acSRoss Zwisler */ 584b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 585642261acSRoss Zwisler 586642261acSRoss Zwisler /* 587642261acSRoss Zwisler * Besides huge zero pages the only other thing that gets 588642261acSRoss Zwisler * downgraded are empty entries which don't need to be 589642261acSRoss Zwisler * unmapped. 590642261acSRoss Zwisler */ 591b15cd800SMatthew Wilcox if (dax_is_zero_entry(entry)) { 592b15cd800SMatthew Wilcox xas_unlock_irq(xas); 593b15cd800SMatthew Wilcox unmap_mapping_pages(mapping, 594b15cd800SMatthew Wilcox xas->xa_index & ~PG_PMD_COLOUR, 595977fbdcdSMatthew Wilcox PG_PMD_NR, false); 596b15cd800SMatthew Wilcox xas_reset(xas); 597b15cd800SMatthew Wilcox xas_lock_irq(xas); 598e11f8b7bSRoss Zwisler } 599e11f8b7bSRoss Zwisler 600d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, false); 601b15cd800SMatthew Wilcox xas_store(xas, NULL); /* undo the PMD join */ 602698ab77aSVivek Goyal dax_wake_entry(xas, entry, WAKE_ALL); 6037f0e07fbSMatthew Wilcox (Oracle) mapping->nrpages -= PG_PMD_NR; 604b15cd800SMatthew Wilcox entry = NULL; 605b15cd800SMatthew Wilcox xas_set(xas, index); 606642261acSRoss Zwisler } 607642261acSRoss Zwisler 608b15cd800SMatthew Wilcox if (entry) { 609b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 610b15cd800SMatthew Wilcox } else { 61123c84eb7SMatthew Wilcox (Oracle) unsigned long flags = DAX_EMPTY; 61223c84eb7SMatthew Wilcox (Oracle) 61323c84eb7SMatthew Wilcox (Oracle) if (order > 0) 61423c84eb7SMatthew Wilcox (Oracle) flags |= DAX_PMD; 61523c84eb7SMatthew Wilcox (Oracle) entry = dax_make_entry(pfn_to_pfn_t(0), flags); 616b15cd800SMatthew Wilcox dax_lock_entry(xas, entry); 617b15cd800SMatthew Wilcox if (xas_error(xas)) 618b15cd800SMatthew Wilcox goto out_unlock; 6197f0e07fbSMatthew Wilcox (Oracle) mapping->nrpages += 1UL << order; 620ac401cc7SJan Kara } 621b15cd800SMatthew Wilcox 622642261acSRoss Zwisler out_unlock: 623b15cd800SMatthew Wilcox xas_unlock_irq(xas); 624b15cd800SMatthew Wilcox if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM)) 625b15cd800SMatthew Wilcox goto retry; 626b15cd800SMatthew Wilcox if (xas->xa_node == XA_ERROR(-ENOMEM)) 627b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_OOM); 628b15cd800SMatthew Wilcox if (xas_error(xas)) 629b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_SIGBUS); 630e3ad61c6SRoss Zwisler return entry; 631b15cd800SMatthew Wilcox fallback: 632b15cd800SMatthew Wilcox xas_unlock_irq(xas); 633b15cd800SMatthew Wilcox return xa_mk_internal(VM_FAULT_FALLBACK); 634ac401cc7SJan Kara } 635ac401cc7SJan Kara 6365fac7408SDan Williams /** 6376bbdd563SVivek Goyal * dax_layout_busy_page_range - find first pinned page in @mapping 6385fac7408SDan Williams * @mapping: address space to scan for a page with ref count > 1 6396bbdd563SVivek Goyal * @start: Starting offset. Page containing 'start' is included. 6406bbdd563SVivek Goyal * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX, 6416bbdd563SVivek Goyal * pages from 'start' till the end of file are included. 6425fac7408SDan Williams * 6435fac7408SDan Williams * DAX requires ZONE_DEVICE mapped pages. These pages are never 6445fac7408SDan Williams * 'onlined' to the page allocator so they are considered idle when 6455fac7408SDan Williams * page->count == 1. A filesystem uses this interface to determine if 6465fac7408SDan Williams * any page in the mapping is busy, i.e. for DMA, or other 6475fac7408SDan Williams * get_user_pages() usages. 6485fac7408SDan Williams * 6495fac7408SDan Williams * It is expected that the filesystem is holding locks to block the 6505fac7408SDan Williams * establishment of new mappings in this address_space. I.e. it expects 6515fac7408SDan Williams * to be able to run unmap_mapping_range() and subsequently not race 6525fac7408SDan Williams * mapping_mapped() becoming true. 6535fac7408SDan Williams */ 6546bbdd563SVivek Goyal struct page *dax_layout_busy_page_range(struct address_space *mapping, 6556bbdd563SVivek Goyal loff_t start, loff_t end) 6565fac7408SDan Williams { 657084a8990SMatthew Wilcox void *entry; 658084a8990SMatthew Wilcox unsigned int scanned = 0; 6595fac7408SDan Williams struct page *page = NULL; 6606bbdd563SVivek Goyal pgoff_t start_idx = start >> PAGE_SHIFT; 6616bbdd563SVivek Goyal pgoff_t end_idx; 6626bbdd563SVivek Goyal XA_STATE(xas, &mapping->i_pages, start_idx); 6635fac7408SDan Williams 6645fac7408SDan Williams /* 6655fac7408SDan Williams * In the 'limited' case get_user_pages() for dax is disabled. 6665fac7408SDan Williams */ 6675fac7408SDan Williams if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 6685fac7408SDan Williams return NULL; 6695fac7408SDan Williams 6705fac7408SDan Williams if (!dax_mapping(mapping) || !mapping_mapped(mapping)) 6715fac7408SDan Williams return NULL; 6725fac7408SDan Williams 6736bbdd563SVivek Goyal /* If end == LLONG_MAX, all pages from start to till end of file */ 6746bbdd563SVivek Goyal if (end == LLONG_MAX) 6756bbdd563SVivek Goyal end_idx = ULONG_MAX; 6766bbdd563SVivek Goyal else 6776bbdd563SVivek Goyal end_idx = end >> PAGE_SHIFT; 6785fac7408SDan Williams /* 6795fac7408SDan Williams * If we race get_user_pages_fast() here either we'll see the 680084a8990SMatthew Wilcox * elevated page count in the iteration and wait, or 6815fac7408SDan Williams * get_user_pages_fast() will see that the page it took a reference 6825fac7408SDan Williams * against is no longer mapped in the page tables and bail to the 6835fac7408SDan Williams * get_user_pages() slow path. The slow path is protected by 6845fac7408SDan Williams * pte_lock() and pmd_lock(). New references are not taken without 6856bbdd563SVivek Goyal * holding those locks, and unmap_mapping_pages() will not zero the 6865fac7408SDan Williams * pte or pmd without holding the respective lock, so we are 6875fac7408SDan Williams * guaranteed to either see new references or prevent new 6885fac7408SDan Williams * references from being established. 6895fac7408SDan Williams */ 6906bbdd563SVivek Goyal unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0); 6915fac7408SDan Williams 692084a8990SMatthew Wilcox xas_lock_irq(&xas); 6936bbdd563SVivek Goyal xas_for_each(&xas, entry, end_idx) { 694084a8990SMatthew Wilcox if (WARN_ON_ONCE(!xa_is_value(entry))) 6955fac7408SDan Williams continue; 696084a8990SMatthew Wilcox if (unlikely(dax_is_locked(entry))) 69723c84eb7SMatthew Wilcox (Oracle) entry = get_unlocked_entry(&xas, 0); 6985fac7408SDan Williams if (entry) 6995fac7408SDan Williams page = dax_busy_page(entry); 7004c3d043dSVivek Goyal put_unlocked_entry(&xas, entry, WAKE_NEXT); 7015fac7408SDan Williams if (page) 7025fac7408SDan Williams break; 703084a8990SMatthew Wilcox if (++scanned % XA_CHECK_SCHED) 704084a8990SMatthew Wilcox continue; 705cdbf8897SRoss Zwisler 706084a8990SMatthew Wilcox xas_pause(&xas); 707084a8990SMatthew Wilcox xas_unlock_irq(&xas); 708084a8990SMatthew Wilcox cond_resched(); 709084a8990SMatthew Wilcox xas_lock_irq(&xas); 7105fac7408SDan Williams } 711084a8990SMatthew Wilcox xas_unlock_irq(&xas); 7125fac7408SDan Williams return page; 7135fac7408SDan Williams } 7146bbdd563SVivek Goyal EXPORT_SYMBOL_GPL(dax_layout_busy_page_range); 7156bbdd563SVivek Goyal 7166bbdd563SVivek Goyal struct page *dax_layout_busy_page(struct address_space *mapping) 7176bbdd563SVivek Goyal { 7186bbdd563SVivek Goyal return dax_layout_busy_page_range(mapping, 0, LLONG_MAX); 7196bbdd563SVivek Goyal } 7205fac7408SDan Williams EXPORT_SYMBOL_GPL(dax_layout_busy_page); 7215fac7408SDan Williams 722a77d19f4SMatthew Wilcox static int __dax_invalidate_entry(struct address_space *mapping, 723c6dcf52cSJan Kara pgoff_t index, bool trunc) 724c6dcf52cSJan Kara { 72507f2d89cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, index); 726c6dcf52cSJan Kara int ret = 0; 727c6dcf52cSJan Kara void *entry; 728c6dcf52cSJan Kara 72907f2d89cSMatthew Wilcox xas_lock_irq(&xas); 73023c84eb7SMatthew Wilcox (Oracle) entry = get_unlocked_entry(&xas, 0); 7313159f943SMatthew Wilcox if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 732c6dcf52cSJan Kara goto out; 733c6dcf52cSJan Kara if (!trunc && 73407f2d89cSMatthew Wilcox (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) || 73507f2d89cSMatthew Wilcox xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE))) 736c6dcf52cSJan Kara goto out; 737d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, trunc); 73807f2d89cSMatthew Wilcox xas_store(&xas, NULL); 7397f0e07fbSMatthew Wilcox (Oracle) mapping->nrpages -= 1UL << dax_entry_order(entry); 740c6dcf52cSJan Kara ret = 1; 741c6dcf52cSJan Kara out: 74223738832SVivek Goyal put_unlocked_entry(&xas, entry, WAKE_ALL); 74307f2d89cSMatthew Wilcox xas_unlock_irq(&xas); 744c6dcf52cSJan Kara return ret; 745c6dcf52cSJan Kara } 74607f2d89cSMatthew Wilcox 747ac401cc7SJan Kara /* 7483159f943SMatthew Wilcox * Delete DAX entry at @index from @mapping. Wait for it 7493159f943SMatthew Wilcox * to be unlocked before deleting it. 750ac401cc7SJan Kara */ 751ac401cc7SJan Kara int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 752ac401cc7SJan Kara { 753a77d19f4SMatthew Wilcox int ret = __dax_invalidate_entry(mapping, index, true); 754ac401cc7SJan Kara 755ac401cc7SJan Kara /* 756ac401cc7SJan Kara * This gets called from truncate / punch_hole path. As such, the caller 757ac401cc7SJan Kara * must hold locks protecting against concurrent modifications of the 758a77d19f4SMatthew Wilcox * page cache (usually fs-private i_mmap_sem for writing). Since the 7593159f943SMatthew Wilcox * caller has seen a DAX entry for this index, we better find it 760ac401cc7SJan Kara * at that index as well... 761ac401cc7SJan Kara */ 762c6dcf52cSJan Kara WARN_ON_ONCE(!ret); 763c6dcf52cSJan Kara return ret; 764ac401cc7SJan Kara } 765ac401cc7SJan Kara 766c6dcf52cSJan Kara /* 7673159f943SMatthew Wilcox * Invalidate DAX entry if it is clean. 768c6dcf52cSJan Kara */ 769c6dcf52cSJan Kara int dax_invalidate_mapping_entry_sync(struct address_space *mapping, 770c6dcf52cSJan Kara pgoff_t index) 771c6dcf52cSJan Kara { 772a77d19f4SMatthew Wilcox return __dax_invalidate_entry(mapping, index, false); 773ac401cc7SJan Kara } 774ac401cc7SJan Kara 77560696eb2SChristoph Hellwig static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos) 776f7ca90b1SMatthew Wilcox { 777de205114SChristoph Hellwig return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset); 778429f8de7SChristoph Hellwig } 779429f8de7SChristoph Hellwig 780429f8de7SChristoph Hellwig static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter) 781429f8de7SChristoph Hellwig { 78260696eb2SChristoph Hellwig pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos); 783cccbce67SDan Williams void *vto, *kaddr; 784cccbce67SDan Williams long rc; 785cccbce67SDan Williams int id; 786e2e05394SRoss Zwisler 787cccbce67SDan Williams id = dax_read_lock(); 788e511c4a3SJane Chu rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS, 789e511c4a3SJane Chu &kaddr, NULL); 790cccbce67SDan Williams if (rc < 0) { 791cccbce67SDan Williams dax_read_unlock(id); 792cccbce67SDan Williams return rc; 793cccbce67SDan Williams } 794429f8de7SChristoph Hellwig vto = kmap_atomic(vmf->cow_page); 795429f8de7SChristoph Hellwig copy_user_page(vto, kaddr, vmf->address, vmf->cow_page); 796f7ca90b1SMatthew Wilcox kunmap_atomic(vto); 797cccbce67SDan Williams dax_read_unlock(id); 798f7ca90b1SMatthew Wilcox return 0; 799f7ca90b1SMatthew Wilcox } 800f7ca90b1SMatthew Wilcox 801642261acSRoss Zwisler /* 802642261acSRoss Zwisler * By this point grab_mapping_entry() has ensured that we have a locked entry 803642261acSRoss Zwisler * of the appropriate size so we don't have to worry about downgrading PMDs to 804642261acSRoss Zwisler * PTEs. If we happen to be trying to insert a PTE and there is a PMD 805642261acSRoss Zwisler * already in the tree, we will skip the insertion and just dirty the PMD as 806642261acSRoss Zwisler * appropriate. 807642261acSRoss Zwisler */ 808b15cd800SMatthew Wilcox static void *dax_insert_entry(struct xa_state *xas, 809b15cd800SMatthew Wilcox struct address_space *mapping, struct vm_fault *vmf, 810b15cd800SMatthew Wilcox void *entry, pfn_t pfn, unsigned long flags, bool dirty) 8119973c98eSRoss Zwisler { 812b15cd800SMatthew Wilcox void *new_entry = dax_make_entry(pfn, flags); 8139973c98eSRoss Zwisler 814f5b7b748SJan Kara if (dirty) 8159973c98eSRoss Zwisler __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 8169973c98eSRoss Zwisler 8173159f943SMatthew Wilcox if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) { 818b15cd800SMatthew Wilcox unsigned long index = xas->xa_index; 81991d25ba8SRoss Zwisler /* we are replacing a zero page with block mapping */ 82091d25ba8SRoss Zwisler if (dax_is_pmd_entry(entry)) 821977fbdcdSMatthew Wilcox unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, 822977fbdcdSMatthew Wilcox PG_PMD_NR, false); 82391d25ba8SRoss Zwisler else /* pte entry */ 824b15cd800SMatthew Wilcox unmap_mapping_pages(mapping, index, 1, false); 825ac401cc7SJan Kara } 8269973c98eSRoss Zwisler 827b15cd800SMatthew Wilcox xas_reset(xas); 828b15cd800SMatthew Wilcox xas_lock_irq(xas); 8291571c029SJan Kara if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 8301571c029SJan Kara void *old; 8311571c029SJan Kara 832d2c997c0SDan Williams dax_disassociate_entry(entry, mapping, false); 83373449dafSDan Williams dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address); 834642261acSRoss Zwisler /* 835a77d19f4SMatthew Wilcox * Only swap our new entry into the page cache if the current 836642261acSRoss Zwisler * entry is a zero page or an empty entry. If a normal PTE or 837a77d19f4SMatthew Wilcox * PMD entry is already in the cache, we leave it alone. This 838642261acSRoss Zwisler * means that if we are trying to insert a PTE and the 839642261acSRoss Zwisler * existing entry is a PMD, we will just leave the PMD in the 840642261acSRoss Zwisler * tree and dirty it if necessary. 841642261acSRoss Zwisler */ 8421571c029SJan Kara old = dax_lock_entry(xas, new_entry); 843b15cd800SMatthew Wilcox WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | 844b15cd800SMatthew Wilcox DAX_LOCKED)); 84591d25ba8SRoss Zwisler entry = new_entry; 846b15cd800SMatthew Wilcox } else { 847b15cd800SMatthew Wilcox xas_load(xas); /* Walk the xa_state */ 848ac401cc7SJan Kara } 84991d25ba8SRoss Zwisler 850f5b7b748SJan Kara if (dirty) 851b15cd800SMatthew Wilcox xas_set_mark(xas, PAGECACHE_TAG_DIRTY); 85291d25ba8SRoss Zwisler 853b15cd800SMatthew Wilcox xas_unlock_irq(xas); 85491d25ba8SRoss Zwisler return entry; 8559973c98eSRoss Zwisler } 8569973c98eSRoss Zwisler 8579fc747f6SMatthew Wilcox static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, 8589fc747f6SMatthew Wilcox struct address_space *mapping, void *entry) 8599973c98eSRoss Zwisler { 86006083a09SMuchun Song unsigned long pfn, index, count, end; 8613fe0791cSDan Williams long ret = 0; 86206083a09SMuchun Song struct vm_area_struct *vma; 8639973c98eSRoss Zwisler 8649973c98eSRoss Zwisler /* 865a6abc2c0SJan Kara * A page got tagged dirty in DAX mapping? Something is seriously 866a6abc2c0SJan Kara * wrong. 8679973c98eSRoss Zwisler */ 8683159f943SMatthew Wilcox if (WARN_ON(!xa_is_value(entry))) 869a6abc2c0SJan Kara return -EIO; 8709973c98eSRoss Zwisler 8719fc747f6SMatthew Wilcox if (unlikely(dax_is_locked(entry))) { 8729fc747f6SMatthew Wilcox void *old_entry = entry; 8739fc747f6SMatthew Wilcox 87423c84eb7SMatthew Wilcox (Oracle) entry = get_unlocked_entry(xas, 0); 8759fc747f6SMatthew Wilcox 876a6abc2c0SJan Kara /* Entry got punched out / reallocated? */ 8779fc747f6SMatthew Wilcox if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 878a6abc2c0SJan Kara goto put_unlocked; 879a6abc2c0SJan Kara /* 8809fc747f6SMatthew Wilcox * Entry got reallocated elsewhere? No need to writeback. 8819fc747f6SMatthew Wilcox * We have to compare pfns as we must not bail out due to 8829fc747f6SMatthew Wilcox * difference in lockbit or entry type. 883a6abc2c0SJan Kara */ 8849fc747f6SMatthew Wilcox if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) 885a6abc2c0SJan Kara goto put_unlocked; 886642261acSRoss Zwisler if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 887642261acSRoss Zwisler dax_is_zero_entry(entry))) { 8889973c98eSRoss Zwisler ret = -EIO; 889a6abc2c0SJan Kara goto put_unlocked; 8909973c98eSRoss Zwisler } 8919973c98eSRoss Zwisler 8929fc747f6SMatthew Wilcox /* Another fsync thread may have already done this entry */ 8939fc747f6SMatthew Wilcox if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) 894a6abc2c0SJan Kara goto put_unlocked; 8959fc747f6SMatthew Wilcox } 8969fc747f6SMatthew Wilcox 897a6abc2c0SJan Kara /* Lock the entry to serialize with page faults */ 8989fc747f6SMatthew Wilcox dax_lock_entry(xas, entry); 8999fc747f6SMatthew Wilcox 900a6abc2c0SJan Kara /* 901a6abc2c0SJan Kara * We can clear the tag now but we have to be careful so that concurrent 902a6abc2c0SJan Kara * dax_writeback_one() calls for the same index cannot finish before we 903a6abc2c0SJan Kara * actually flush the caches. This is achieved as the calls will look 904b93b0163SMatthew Wilcox * at the entry only under the i_pages lock and once they do that 905b93b0163SMatthew Wilcox * they will see the entry locked and wait for it to unlock. 906a6abc2c0SJan Kara */ 9079fc747f6SMatthew Wilcox xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); 9089fc747f6SMatthew Wilcox xas_unlock_irq(xas); 909a6abc2c0SJan Kara 910642261acSRoss Zwisler /* 911e4b3448bSMatthew Wilcox * If dax_writeback_mapping_range() was given a wbc->range_start 912e4b3448bSMatthew Wilcox * in the middle of a PMD, the 'index' we use needs to be 913e4b3448bSMatthew Wilcox * aligned to the start of the PMD. 9143fe0791cSDan Williams * This allows us to flush for PMD_SIZE and not have to worry about 9153fe0791cSDan Williams * partial PMD writebacks. 916642261acSRoss Zwisler */ 917a77d19f4SMatthew Wilcox pfn = dax_to_pfn(entry); 918e4b3448bSMatthew Wilcox count = 1UL << dax_entry_order(entry); 919e4b3448bSMatthew Wilcox index = xas->xa_index & ~(count - 1); 92006083a09SMuchun Song end = index + count - 1; 921cccbce67SDan Williams 92206083a09SMuchun Song /* Walk all mappings of a given index of a file and writeprotect them */ 92306083a09SMuchun Song i_mmap_lock_read(mapping); 92406083a09SMuchun Song vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) { 92506083a09SMuchun Song pfn_mkclean_range(pfn, count, index, vma); 92606083a09SMuchun Song cond_resched(); 92706083a09SMuchun Song } 92806083a09SMuchun Song i_mmap_unlock_read(mapping); 92906083a09SMuchun Song 930e4b3448bSMatthew Wilcox dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE); 9314b4bb46dSJan Kara /* 9324b4bb46dSJan Kara * After we have flushed the cache, we can clear the dirty tag. There 9334b4bb46dSJan Kara * cannot be new dirty data in the pfn after the flush has completed as 9344b4bb46dSJan Kara * the pfn mappings are writeprotected and fault waits for mapping 9354b4bb46dSJan Kara * entry lock. 9364b4bb46dSJan Kara */ 9379fc747f6SMatthew Wilcox xas_reset(xas); 9389fc747f6SMatthew Wilcox xas_lock_irq(xas); 9399fc747f6SMatthew Wilcox xas_store(xas, entry); 9409fc747f6SMatthew Wilcox xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); 941698ab77aSVivek Goyal dax_wake_entry(xas, entry, WAKE_NEXT); 9429fc747f6SMatthew Wilcox 943e4b3448bSMatthew Wilcox trace_dax_writeback_one(mapping->host, index, count); 9449973c98eSRoss Zwisler return ret; 9459973c98eSRoss Zwisler 946a6abc2c0SJan Kara put_unlocked: 9474c3d043dSVivek Goyal put_unlocked_entry(xas, entry, WAKE_NEXT); 9489973c98eSRoss Zwisler return ret; 9499973c98eSRoss Zwisler } 9509973c98eSRoss Zwisler 9519973c98eSRoss Zwisler /* 9529973c98eSRoss Zwisler * Flush the mapping to the persistent domain within the byte range of [start, 9539973c98eSRoss Zwisler * end]. This is required by data integrity operations to ensure file data is 9549973c98eSRoss Zwisler * on persistent storage prior to completion of the operation. 9559973c98eSRoss Zwisler */ 9567f6d5b52SRoss Zwisler int dax_writeback_mapping_range(struct address_space *mapping, 9573f666c56SVivek Goyal struct dax_device *dax_dev, struct writeback_control *wbc) 9589973c98eSRoss Zwisler { 9599fc747f6SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); 9609973c98eSRoss Zwisler struct inode *inode = mapping->host; 9619fc747f6SMatthew Wilcox pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; 9629fc747f6SMatthew Wilcox void *entry; 9639fc747f6SMatthew Wilcox int ret = 0; 9649fc747f6SMatthew Wilcox unsigned int scanned = 0; 9659973c98eSRoss Zwisler 9669973c98eSRoss Zwisler if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 9679973c98eSRoss Zwisler return -EIO; 9689973c98eSRoss Zwisler 9697716506aSMatthew Wilcox (Oracle) if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL) 9707f6d5b52SRoss Zwisler return 0; 9717f6d5b52SRoss Zwisler 9729fc747f6SMatthew Wilcox trace_dax_writeback_range(inode, xas.xa_index, end_index); 9739973c98eSRoss Zwisler 9749fc747f6SMatthew Wilcox tag_pages_for_writeback(mapping, xas.xa_index, end_index); 975d14a3f48SRoss Zwisler 9769fc747f6SMatthew Wilcox xas_lock_irq(&xas); 9779fc747f6SMatthew Wilcox xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { 9789fc747f6SMatthew Wilcox ret = dax_writeback_one(&xas, dax_dev, mapping, entry); 979819ec6b9SJeff Layton if (ret < 0) { 980819ec6b9SJeff Layton mapping_set_error(mapping, ret); 9819fc747f6SMatthew Wilcox break; 982d14a3f48SRoss Zwisler } 9839fc747f6SMatthew Wilcox if (++scanned % XA_CHECK_SCHED) 9849fc747f6SMatthew Wilcox continue; 9859fc747f6SMatthew Wilcox 9869fc747f6SMatthew Wilcox xas_pause(&xas); 9879fc747f6SMatthew Wilcox xas_unlock_irq(&xas); 9889fc747f6SMatthew Wilcox cond_resched(); 9899fc747f6SMatthew Wilcox xas_lock_irq(&xas); 990d14a3f48SRoss Zwisler } 9919fc747f6SMatthew Wilcox xas_unlock_irq(&xas); 9929fc747f6SMatthew Wilcox trace_dax_writeback_range_done(inode, xas.xa_index, end_index); 9939fc747f6SMatthew Wilcox return ret; 9949973c98eSRoss Zwisler } 9959973c98eSRoss Zwisler EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 9969973c98eSRoss Zwisler 99765dd814aSChristoph Hellwig static int dax_iomap_pfn(const struct iomap *iomap, loff_t pos, size_t size, 9985e161e40SJan Kara pfn_t *pfnp) 9995e161e40SJan Kara { 100060696eb2SChristoph Hellwig pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 10015e161e40SJan Kara int id, rc; 10025e161e40SJan Kara long length; 10035e161e40SJan Kara 1004cccbce67SDan Williams id = dax_read_lock(); 10055e161e40SJan Kara length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), 1006e511c4a3SJane Chu DAX_ACCESS, NULL, pfnp); 10075e161e40SJan Kara if (length < 0) { 10085e161e40SJan Kara rc = length; 10095e161e40SJan Kara goto out; 10105e161e40SJan Kara } 10115e161e40SJan Kara rc = -EINVAL; 10125e161e40SJan Kara if (PFN_PHYS(length) < size) 10135e161e40SJan Kara goto out; 10145e161e40SJan Kara if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) 10155e161e40SJan Kara goto out; 10165e161e40SJan Kara /* For larger pages we need devmap */ 10175e161e40SJan Kara if (length > 1 && !pfn_t_devmap(*pfnp)) 10185e161e40SJan Kara goto out; 10195e161e40SJan Kara rc = 0; 10205e161e40SJan Kara out: 1021cccbce67SDan Williams dax_read_unlock(id); 1022cccbce67SDan Williams return rc; 1023cccbce67SDan Williams } 1024f7ca90b1SMatthew Wilcox 10252f89dc12SJan Kara /* 102691d25ba8SRoss Zwisler * The user has performed a load from a hole in the file. Allocating a new 102791d25ba8SRoss Zwisler * page in the file would cause excessive storage usage for workloads with 102891d25ba8SRoss Zwisler * sparse files. Instead we insert a read-only mapping of the 4k zero page. 102991d25ba8SRoss Zwisler * If this page is ever written to we will re-fault and change the mapping to 103091d25ba8SRoss Zwisler * point to real DAX storage instead. 10312f89dc12SJan Kara */ 1032b15cd800SMatthew Wilcox static vm_fault_t dax_load_hole(struct xa_state *xas, 1033b15cd800SMatthew Wilcox struct address_space *mapping, void **entry, 1034e30331ffSRoss Zwisler struct vm_fault *vmf) 1035e30331ffSRoss Zwisler { 1036e30331ffSRoss Zwisler struct inode *inode = mapping->host; 103791d25ba8SRoss Zwisler unsigned long vaddr = vmf->address; 1038b90ca5ccSMatthew Wilcox pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); 1039b90ca5ccSMatthew Wilcox vm_fault_t ret; 1040e30331ffSRoss Zwisler 1041b15cd800SMatthew Wilcox *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 10423159f943SMatthew Wilcox DAX_ZERO_PAGE, false); 10433159f943SMatthew Wilcox 1044ab77dab4SSouptick Joarder ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); 1045e30331ffSRoss Zwisler trace_dax_load_hole(inode, vmf, ret); 1046e30331ffSRoss Zwisler return ret; 1047e30331ffSRoss Zwisler } 1048e30331ffSRoss Zwisler 1049c2436190SShiyang Ruan #ifdef CONFIG_FS_DAX_PMD 1050c2436190SShiyang Ruan static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 105165dd814aSChristoph Hellwig const struct iomap *iomap, void **entry) 1052c2436190SShiyang Ruan { 1053c2436190SShiyang Ruan struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1054c2436190SShiyang Ruan unsigned long pmd_addr = vmf->address & PMD_MASK; 1055c2436190SShiyang Ruan struct vm_area_struct *vma = vmf->vma; 1056c2436190SShiyang Ruan struct inode *inode = mapping->host; 1057c2436190SShiyang Ruan pgtable_t pgtable = NULL; 1058c2436190SShiyang Ruan struct page *zero_page; 1059c2436190SShiyang Ruan spinlock_t *ptl; 1060c2436190SShiyang Ruan pmd_t pmd_entry; 1061c2436190SShiyang Ruan pfn_t pfn; 1062c2436190SShiyang Ruan 1063c2436190SShiyang Ruan zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); 1064c2436190SShiyang Ruan 1065c2436190SShiyang Ruan if (unlikely(!zero_page)) 1066c2436190SShiyang Ruan goto fallback; 1067c2436190SShiyang Ruan 1068c2436190SShiyang Ruan pfn = page_to_pfn_t(zero_page); 1069c2436190SShiyang Ruan *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 1070c2436190SShiyang Ruan DAX_PMD | DAX_ZERO_PAGE, false); 1071c2436190SShiyang Ruan 1072c2436190SShiyang Ruan if (arch_needs_pgtable_deposit()) { 1073c2436190SShiyang Ruan pgtable = pte_alloc_one(vma->vm_mm); 1074c2436190SShiyang Ruan if (!pgtable) 1075c2436190SShiyang Ruan return VM_FAULT_OOM; 1076c2436190SShiyang Ruan } 1077c2436190SShiyang Ruan 1078c2436190SShiyang Ruan ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1079c2436190SShiyang Ruan if (!pmd_none(*(vmf->pmd))) { 1080c2436190SShiyang Ruan spin_unlock(ptl); 1081c2436190SShiyang Ruan goto fallback; 1082c2436190SShiyang Ruan } 1083c2436190SShiyang Ruan 1084c2436190SShiyang Ruan if (pgtable) { 1085c2436190SShiyang Ruan pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 1086c2436190SShiyang Ruan mm_inc_nr_ptes(vma->vm_mm); 1087c2436190SShiyang Ruan } 1088c2436190SShiyang Ruan pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); 1089c2436190SShiyang Ruan pmd_entry = pmd_mkhuge(pmd_entry); 1090c2436190SShiyang Ruan set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); 1091c2436190SShiyang Ruan spin_unlock(ptl); 1092c2436190SShiyang Ruan trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); 1093c2436190SShiyang Ruan return VM_FAULT_NOPAGE; 1094c2436190SShiyang Ruan 1095c2436190SShiyang Ruan fallback: 1096c2436190SShiyang Ruan if (pgtable) 1097c2436190SShiyang Ruan pte_free(vma->vm_mm, pgtable); 1098c2436190SShiyang Ruan trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); 1099c2436190SShiyang Ruan return VM_FAULT_FALLBACK; 1100c2436190SShiyang Ruan } 1101c2436190SShiyang Ruan #else 1102c2436190SShiyang Ruan static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 110365dd814aSChristoph Hellwig const struct iomap *iomap, void **entry) 1104c2436190SShiyang Ruan { 1105c2436190SShiyang Ruan return VM_FAULT_FALLBACK; 1106c2436190SShiyang Ruan } 1107c2436190SShiyang Ruan #endif /* CONFIG_FS_DAX_PMD */ 1108c2436190SShiyang Ruan 1109e5c71954SChristoph Hellwig static int dax_memzero(struct dax_device *dax_dev, pgoff_t pgoff, 1110e5c71954SChristoph Hellwig unsigned int offset, size_t size) 1111e5c71954SChristoph Hellwig { 1112e5c71954SChristoph Hellwig void *kaddr; 1113e5c71954SChristoph Hellwig long ret; 1114e5c71954SChristoph Hellwig 1115e511c4a3SJane Chu ret = dax_direct_access(dax_dev, pgoff, 1, DAX_ACCESS, &kaddr, NULL); 1116e5c71954SChristoph Hellwig if (ret > 0) { 1117e5c71954SChristoph Hellwig memset(kaddr + offset, 0, size); 1118e5c71954SChristoph Hellwig dax_flush(dax_dev, kaddr + offset, size); 1119e5c71954SChristoph Hellwig } 1120e5c71954SChristoph Hellwig return ret; 1121e5c71954SChristoph Hellwig } 1122e5c71954SChristoph Hellwig 1123c6f40468SChristoph Hellwig static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero) 1124679c8bd3SChristoph Hellwig { 1125c6f40468SChristoph Hellwig const struct iomap *iomap = &iter->iomap; 1126c6f40468SChristoph Hellwig const struct iomap *srcmap = iomap_iter_srcmap(iter); 1127c6f40468SChristoph Hellwig loff_t pos = iter->pos; 1128c6f40468SChristoph Hellwig u64 length = iomap_length(iter); 1129c6f40468SChristoph Hellwig s64 written = 0; 1130c6f40468SChristoph Hellwig 1131c6f40468SChristoph Hellwig /* already zeroed? we're done. */ 1132c6f40468SChristoph Hellwig if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN) 1133c6f40468SChristoph Hellwig return length; 1134c6f40468SChristoph Hellwig 1135c6f40468SChristoph Hellwig do { 113681ee8e52SMatthew Wilcox (Oracle) unsigned offset = offset_in_page(pos); 113781ee8e52SMatthew Wilcox (Oracle) unsigned size = min_t(u64, PAGE_SIZE - offset, length); 1138c6f40468SChristoph Hellwig pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1139c6f40468SChristoph Hellwig long rc; 1140c6f40468SChristoph Hellwig int id; 11410a23f9ffSVivek Goyal 1142cccbce67SDan Williams id = dax_read_lock(); 1143e5c71954SChristoph Hellwig if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE) 114481ee8e52SMatthew Wilcox (Oracle) rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1); 11450a23f9ffSVivek Goyal else 1146e5c71954SChristoph Hellwig rc = dax_memzero(iomap->dax_dev, pgoff, offset, size); 1147cccbce67SDan Williams dax_read_unlock(id); 11480a23f9ffSVivek Goyal 1149e5c71954SChristoph Hellwig if (rc < 0) 1150e5c71954SChristoph Hellwig return rc; 1151c6f40468SChristoph Hellwig pos += size; 1152c6f40468SChristoph Hellwig length -= size; 1153c6f40468SChristoph Hellwig written += size; 1154c6f40468SChristoph Hellwig if (did_zero) 1155c6f40468SChristoph Hellwig *did_zero = true; 1156c6f40468SChristoph Hellwig } while (length > 0); 1157c6f40468SChristoph Hellwig 1158c6f40468SChristoph Hellwig return written; 1159679c8bd3SChristoph Hellwig } 1160679c8bd3SChristoph Hellwig 1161c6f40468SChristoph Hellwig int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero, 1162c6f40468SChristoph Hellwig const struct iomap_ops *ops) 1163c6f40468SChristoph Hellwig { 1164c6f40468SChristoph Hellwig struct iomap_iter iter = { 1165c6f40468SChristoph Hellwig .inode = inode, 1166c6f40468SChristoph Hellwig .pos = pos, 1167c6f40468SChristoph Hellwig .len = len, 1168952da063SChristoph Hellwig .flags = IOMAP_DAX | IOMAP_ZERO, 1169c6f40468SChristoph Hellwig }; 1170c6f40468SChristoph Hellwig int ret; 1171c6f40468SChristoph Hellwig 1172c6f40468SChristoph Hellwig while ((ret = iomap_iter(&iter, ops)) > 0) 1173c6f40468SChristoph Hellwig iter.processed = dax_zero_iter(&iter, did_zero); 1174c6f40468SChristoph Hellwig return ret; 1175c6f40468SChristoph Hellwig } 1176c6f40468SChristoph Hellwig EXPORT_SYMBOL_GPL(dax_zero_range); 1177c6f40468SChristoph Hellwig 1178c6f40468SChristoph Hellwig int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero, 1179c6f40468SChristoph Hellwig const struct iomap_ops *ops) 1180c6f40468SChristoph Hellwig { 1181c6f40468SChristoph Hellwig unsigned int blocksize = i_blocksize(inode); 1182c6f40468SChristoph Hellwig unsigned int off = pos & (blocksize - 1); 1183c6f40468SChristoph Hellwig 1184c6f40468SChristoph Hellwig /* Block boundary? Nothing to do */ 1185c6f40468SChristoph Hellwig if (!off) 1186c6f40468SChristoph Hellwig return 0; 1187c6f40468SChristoph Hellwig return dax_zero_range(inode, pos, blocksize - off, did_zero, ops); 1188c6f40468SChristoph Hellwig } 1189c6f40468SChristoph Hellwig EXPORT_SYMBOL_GPL(dax_truncate_page); 1190c6f40468SChristoph Hellwig 1191ca289e0bSChristoph Hellwig static loff_t dax_iomap_iter(const struct iomap_iter *iomi, 1192ca289e0bSChristoph Hellwig struct iov_iter *iter) 1193a254e568SChristoph Hellwig { 1194ca289e0bSChristoph Hellwig const struct iomap *iomap = &iomi->iomap; 1195ca289e0bSChristoph Hellwig loff_t length = iomap_length(iomi); 1196ca289e0bSChristoph Hellwig loff_t pos = iomi->pos; 1197cccbce67SDan Williams struct dax_device *dax_dev = iomap->dax_dev; 1198a254e568SChristoph Hellwig loff_t end = pos + length, done = 0; 1199a254e568SChristoph Hellwig ssize_t ret = 0; 1200a77d4786SDan Williams size_t xfer; 1201cccbce67SDan Williams int id; 1202a254e568SChristoph Hellwig 1203a254e568SChristoph Hellwig if (iov_iter_rw(iter) == READ) { 1204ca289e0bSChristoph Hellwig end = min(end, i_size_read(iomi->inode)); 1205a254e568SChristoph Hellwig if (pos >= end) 1206a254e568SChristoph Hellwig return 0; 1207a254e568SChristoph Hellwig 1208a254e568SChristoph Hellwig if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1209a254e568SChristoph Hellwig return iov_iter_zero(min(length, end - pos), iter); 1210a254e568SChristoph Hellwig } 1211a254e568SChristoph Hellwig 1212a254e568SChristoph Hellwig if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) 1213a254e568SChristoph Hellwig return -EIO; 1214a254e568SChristoph Hellwig 1215e3fce68cSJan Kara /* 1216e3fce68cSJan Kara * Write can allocate block for an area which has a hole page mapped 1217e3fce68cSJan Kara * into page tables. We have to tear down these mappings so that data 1218e3fce68cSJan Kara * written by write(2) is visible in mmap. 1219e3fce68cSJan Kara */ 1220cd656375SJan Kara if (iomap->flags & IOMAP_F_NEW) { 1221ca289e0bSChristoph Hellwig invalidate_inode_pages2_range(iomi->inode->i_mapping, 1222e3fce68cSJan Kara pos >> PAGE_SHIFT, 1223e3fce68cSJan Kara (end - 1) >> PAGE_SHIFT); 1224e3fce68cSJan Kara } 1225e3fce68cSJan Kara 1226cccbce67SDan Williams id = dax_read_lock(); 1227a254e568SChristoph Hellwig while (pos < end) { 1228a254e568SChristoph Hellwig unsigned offset = pos & (PAGE_SIZE - 1); 1229cccbce67SDan Williams const size_t size = ALIGN(length + offset, PAGE_SIZE); 123060696eb2SChristoph Hellwig pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1231a254e568SChristoph Hellwig ssize_t map_len; 1232047218ecSJane Chu bool recovery = false; 1233cccbce67SDan Williams void *kaddr; 1234a254e568SChristoph Hellwig 1235d1908f52SMichal Hocko if (fatal_signal_pending(current)) { 1236d1908f52SMichal Hocko ret = -EINTR; 1237d1908f52SMichal Hocko break; 1238d1908f52SMichal Hocko } 1239d1908f52SMichal Hocko 1240cccbce67SDan Williams map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), 1241e511c4a3SJane Chu DAX_ACCESS, &kaddr, NULL); 1242047218ecSJane Chu if (map_len == -EIO && iov_iter_rw(iter) == WRITE) { 1243047218ecSJane Chu map_len = dax_direct_access(dax_dev, pgoff, 1244047218ecSJane Chu PHYS_PFN(size), DAX_RECOVERY_WRITE, 124586ed913bSHuaisheng Ye &kaddr, NULL); 1246047218ecSJane Chu if (map_len > 0) 1247047218ecSJane Chu recovery = true; 1248047218ecSJane Chu } 1249a254e568SChristoph Hellwig if (map_len < 0) { 1250a254e568SChristoph Hellwig ret = map_len; 1251a254e568SChristoph Hellwig break; 1252a254e568SChristoph Hellwig } 1253a254e568SChristoph Hellwig 1254cccbce67SDan Williams map_len = PFN_PHYS(map_len); 1255cccbce67SDan Williams kaddr += offset; 1256a254e568SChristoph Hellwig map_len -= offset; 1257a254e568SChristoph Hellwig if (map_len > end - pos) 1258a254e568SChristoph Hellwig map_len = end - pos; 1259a254e568SChristoph Hellwig 1260047218ecSJane Chu if (recovery) 1261047218ecSJane Chu xfer = dax_recovery_write(dax_dev, pgoff, kaddr, 1262047218ecSJane Chu map_len, iter); 1263047218ecSJane Chu else if (iov_iter_rw(iter) == WRITE) 1264a77d4786SDan Williams xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, 1265fec53774SDan Williams map_len, iter); 1266a254e568SChristoph Hellwig else 1267a77d4786SDan Williams xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, 1268b3a9a0c3SDan Williams map_len, iter); 1269a254e568SChristoph Hellwig 1270a77d4786SDan Williams pos += xfer; 1271a77d4786SDan Williams length -= xfer; 1272a77d4786SDan Williams done += xfer; 1273a77d4786SDan Williams 1274a77d4786SDan Williams if (xfer == 0) 1275a77d4786SDan Williams ret = -EFAULT; 1276a77d4786SDan Williams if (xfer < map_len) 1277a77d4786SDan Williams break; 1278a254e568SChristoph Hellwig } 1279cccbce67SDan Williams dax_read_unlock(id); 1280a254e568SChristoph Hellwig 1281a254e568SChristoph Hellwig return done ? done : ret; 1282a254e568SChristoph Hellwig } 1283a254e568SChristoph Hellwig 1284a254e568SChristoph Hellwig /** 128511c59c92SRoss Zwisler * dax_iomap_rw - Perform I/O to a DAX file 1286a254e568SChristoph Hellwig * @iocb: The control block for this I/O 1287a254e568SChristoph Hellwig * @iter: The addresses to do I/O from or to 1288a254e568SChristoph Hellwig * @ops: iomap ops passed from the file system 1289a254e568SChristoph Hellwig * 1290a254e568SChristoph Hellwig * This function performs read and write operations to directly mapped 1291a254e568SChristoph Hellwig * persistent memory. The callers needs to take care of read/write exclusion 1292a254e568SChristoph Hellwig * and evicting any page cache pages in the region under I/O. 1293a254e568SChristoph Hellwig */ 1294a254e568SChristoph Hellwig ssize_t 129511c59c92SRoss Zwisler dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 12968ff6daa1SChristoph Hellwig const struct iomap_ops *ops) 1297a254e568SChristoph Hellwig { 1298ca289e0bSChristoph Hellwig struct iomap_iter iomi = { 1299ca289e0bSChristoph Hellwig .inode = iocb->ki_filp->f_mapping->host, 1300ca289e0bSChristoph Hellwig .pos = iocb->ki_pos, 1301ca289e0bSChristoph Hellwig .len = iov_iter_count(iter), 1302952da063SChristoph Hellwig .flags = IOMAP_DAX, 1303ca289e0bSChristoph Hellwig }; 1304ca289e0bSChristoph Hellwig loff_t done = 0; 1305ca289e0bSChristoph Hellwig int ret; 1306a254e568SChristoph Hellwig 1307168316dbSChristoph Hellwig if (iov_iter_rw(iter) == WRITE) { 1308ca289e0bSChristoph Hellwig lockdep_assert_held_write(&iomi.inode->i_rwsem); 1309ca289e0bSChristoph Hellwig iomi.flags |= IOMAP_WRITE; 1310168316dbSChristoph Hellwig } else { 1311ca289e0bSChristoph Hellwig lockdep_assert_held(&iomi.inode->i_rwsem); 1312168316dbSChristoph Hellwig } 1313a254e568SChristoph Hellwig 131496222d53SJeff Moyer if (iocb->ki_flags & IOCB_NOWAIT) 1315ca289e0bSChristoph Hellwig iomi.flags |= IOMAP_NOWAIT; 131696222d53SJeff Moyer 1317ca289e0bSChristoph Hellwig while ((ret = iomap_iter(&iomi, ops)) > 0) 1318ca289e0bSChristoph Hellwig iomi.processed = dax_iomap_iter(&iomi, iter); 1319a254e568SChristoph Hellwig 1320ca289e0bSChristoph Hellwig done = iomi.pos - iocb->ki_pos; 1321ca289e0bSChristoph Hellwig iocb->ki_pos = iomi.pos; 1322a254e568SChristoph Hellwig return done ? done : ret; 1323a254e568SChristoph Hellwig } 132411c59c92SRoss Zwisler EXPORT_SYMBOL_GPL(dax_iomap_rw); 1325a7d73fe6SChristoph Hellwig 1326ab77dab4SSouptick Joarder static vm_fault_t dax_fault_return(int error) 13279f141d6eSJan Kara { 13289f141d6eSJan Kara if (error == 0) 13299f141d6eSJan Kara return VM_FAULT_NOPAGE; 1330c9aed74eSSouptick Joarder return vmf_error(error); 13319f141d6eSJan Kara } 13329f141d6eSJan Kara 1333aaa422c4SDan Williams /* 1334aaa422c4SDan Williams * MAP_SYNC on a dax mapping guarantees dirty metadata is 1335aaa422c4SDan Williams * flushed on write-faults (non-cow), but not read-faults. 1336aaa422c4SDan Williams */ 1337aaa422c4SDan Williams static bool dax_fault_is_synchronous(unsigned long flags, 133865dd814aSChristoph Hellwig struct vm_area_struct *vma, const struct iomap *iomap) 1339aaa422c4SDan Williams { 1340aaa422c4SDan Williams return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) 1341aaa422c4SDan Williams && (iomap->flags & IOMAP_F_DIRTY); 1342aaa422c4SDan Williams } 1343aaa422c4SDan Williams 134455f81639SShiyang Ruan /* 134555f81639SShiyang Ruan * When handling a synchronous page fault and the inode need a fsync, we can 134655f81639SShiyang Ruan * insert the PTE/PMD into page tables only after that fsync happened. Skip 134755f81639SShiyang Ruan * insertion for now and return the pfn so that caller can insert it after the 134855f81639SShiyang Ruan * fsync is done. 134955f81639SShiyang Ruan */ 135055f81639SShiyang Ruan static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn) 135155f81639SShiyang Ruan { 135255f81639SShiyang Ruan if (WARN_ON_ONCE(!pfnp)) 135355f81639SShiyang Ruan return VM_FAULT_SIGBUS; 135455f81639SShiyang Ruan *pfnp = pfn; 135555f81639SShiyang Ruan return VM_FAULT_NEEDDSYNC; 135655f81639SShiyang Ruan } 135755f81639SShiyang Ruan 135865dd814aSChristoph Hellwig static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf, 135965dd814aSChristoph Hellwig const struct iomap_iter *iter) 136055f81639SShiyang Ruan { 136155f81639SShiyang Ruan vm_fault_t ret; 136255f81639SShiyang Ruan int error = 0; 136355f81639SShiyang Ruan 136465dd814aSChristoph Hellwig switch (iter->iomap.type) { 136555f81639SShiyang Ruan case IOMAP_HOLE: 136655f81639SShiyang Ruan case IOMAP_UNWRITTEN: 1367429f8de7SChristoph Hellwig clear_user_highpage(vmf->cow_page, vmf->address); 136855f81639SShiyang Ruan break; 136955f81639SShiyang Ruan case IOMAP_MAPPED: 1370429f8de7SChristoph Hellwig error = copy_cow_page_dax(vmf, iter); 137155f81639SShiyang Ruan break; 137255f81639SShiyang Ruan default: 137355f81639SShiyang Ruan WARN_ON_ONCE(1); 137455f81639SShiyang Ruan error = -EIO; 137555f81639SShiyang Ruan break; 137655f81639SShiyang Ruan } 137755f81639SShiyang Ruan 137855f81639SShiyang Ruan if (error) 137955f81639SShiyang Ruan return dax_fault_return(error); 138055f81639SShiyang Ruan 138155f81639SShiyang Ruan __SetPageUptodate(vmf->cow_page); 138255f81639SShiyang Ruan ret = finish_fault(vmf); 138355f81639SShiyang Ruan if (!ret) 138455f81639SShiyang Ruan return VM_FAULT_DONE_COW; 138555f81639SShiyang Ruan return ret; 138655f81639SShiyang Ruan } 138755f81639SShiyang Ruan 1388c2436190SShiyang Ruan /** 138965dd814aSChristoph Hellwig * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault. 1390c2436190SShiyang Ruan * @vmf: vm fault instance 139165dd814aSChristoph Hellwig * @iter: iomap iter 1392c2436190SShiyang Ruan * @pfnp: pfn to be returned 1393c2436190SShiyang Ruan * @xas: the dax mapping tree of a file 1394c2436190SShiyang Ruan * @entry: an unlocked dax entry to be inserted 1395c2436190SShiyang Ruan * @pmd: distinguish whether it is a pmd fault 1396c2436190SShiyang Ruan */ 139765dd814aSChristoph Hellwig static vm_fault_t dax_fault_iter(struct vm_fault *vmf, 139865dd814aSChristoph Hellwig const struct iomap_iter *iter, pfn_t *pfnp, 139965dd814aSChristoph Hellwig struct xa_state *xas, void **entry, bool pmd) 1400c2436190SShiyang Ruan { 1401c2436190SShiyang Ruan struct address_space *mapping = vmf->vma->vm_file->f_mapping; 140265dd814aSChristoph Hellwig const struct iomap *iomap = &iter->iomap; 1403c2436190SShiyang Ruan size_t size = pmd ? PMD_SIZE : PAGE_SIZE; 1404c2436190SShiyang Ruan loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT; 1405c2436190SShiyang Ruan bool write = vmf->flags & FAULT_FLAG_WRITE; 140665dd814aSChristoph Hellwig bool sync = dax_fault_is_synchronous(iter->flags, vmf->vma, iomap); 1407c2436190SShiyang Ruan unsigned long entry_flags = pmd ? DAX_PMD : 0; 1408c2436190SShiyang Ruan int err = 0; 1409c2436190SShiyang Ruan pfn_t pfn; 1410c2436190SShiyang Ruan 141165dd814aSChristoph Hellwig if (!pmd && vmf->cow_page) 141265dd814aSChristoph Hellwig return dax_fault_cow_page(vmf, iter); 141365dd814aSChristoph Hellwig 1414c2436190SShiyang Ruan /* if we are reading UNWRITTEN and HOLE, return a hole. */ 1415c2436190SShiyang Ruan if (!write && 1416c2436190SShiyang Ruan (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) { 1417c2436190SShiyang Ruan if (!pmd) 1418c2436190SShiyang Ruan return dax_load_hole(xas, mapping, entry, vmf); 1419c2436190SShiyang Ruan return dax_pmd_load_hole(xas, vmf, iomap, entry); 1420c2436190SShiyang Ruan } 1421c2436190SShiyang Ruan 1422c2436190SShiyang Ruan if (iomap->type != IOMAP_MAPPED) { 1423c2436190SShiyang Ruan WARN_ON_ONCE(1); 1424c2436190SShiyang Ruan return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS; 1425c2436190SShiyang Ruan } 1426c2436190SShiyang Ruan 142765dd814aSChristoph Hellwig err = dax_iomap_pfn(&iter->iomap, pos, size, &pfn); 1428c2436190SShiyang Ruan if (err) 1429c2436190SShiyang Ruan return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err); 1430c2436190SShiyang Ruan 1431c2436190SShiyang Ruan *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, entry_flags, 1432c2436190SShiyang Ruan write && !sync); 1433c2436190SShiyang Ruan 1434c2436190SShiyang Ruan if (sync) 1435c2436190SShiyang Ruan return dax_fault_synchronous_pfnp(pfnp, pfn); 1436c2436190SShiyang Ruan 1437c2436190SShiyang Ruan /* insert PMD pfn */ 1438c2436190SShiyang Ruan if (pmd) 1439c2436190SShiyang Ruan return vmf_insert_pfn_pmd(vmf, pfn, write); 1440c2436190SShiyang Ruan 1441c2436190SShiyang Ruan /* insert PTE pfn */ 1442c2436190SShiyang Ruan if (write) 1443c2436190SShiyang Ruan return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 1444c2436190SShiyang Ruan return vmf_insert_mixed(vmf->vma, vmf->address, pfn); 1445c2436190SShiyang Ruan } 1446c2436190SShiyang Ruan 1447ab77dab4SSouptick Joarder static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, 1448c0b24625SJan Kara int *iomap_errp, const struct iomap_ops *ops) 1449a7d73fe6SChristoph Hellwig { 145065dd814aSChristoph Hellwig struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1451b15cd800SMatthew Wilcox XA_STATE(xas, &mapping->i_pages, vmf->pgoff); 145265dd814aSChristoph Hellwig struct iomap_iter iter = { 145365dd814aSChristoph Hellwig .inode = mapping->host, 145465dd814aSChristoph Hellwig .pos = (loff_t)vmf->pgoff << PAGE_SHIFT, 145565dd814aSChristoph Hellwig .len = PAGE_SIZE, 1456952da063SChristoph Hellwig .flags = IOMAP_DAX | IOMAP_FAULT, 145765dd814aSChristoph Hellwig }; 1458ab77dab4SSouptick Joarder vm_fault_t ret = 0; 1459a7d73fe6SChristoph Hellwig void *entry; 146065dd814aSChristoph Hellwig int error; 1461a7d73fe6SChristoph Hellwig 146265dd814aSChristoph Hellwig trace_dax_pte_fault(iter.inode, vmf, ret); 1463a7d73fe6SChristoph Hellwig /* 1464a7d73fe6SChristoph Hellwig * Check whether offset isn't beyond end of file now. Caller is supposed 1465a7d73fe6SChristoph Hellwig * to hold locks serializing us with truncate / punch hole so this is 1466a7d73fe6SChristoph Hellwig * a reliable test. 1467a7d73fe6SChristoph Hellwig */ 146865dd814aSChristoph Hellwig if (iter.pos >= i_size_read(iter.inode)) { 1469ab77dab4SSouptick Joarder ret = VM_FAULT_SIGBUS; 1470a9c42b33SRoss Zwisler goto out; 1471a9c42b33SRoss Zwisler } 1472a7d73fe6SChristoph Hellwig 147365dd814aSChristoph Hellwig if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) 147465dd814aSChristoph Hellwig iter.flags |= IOMAP_WRITE; 1475a7d73fe6SChristoph Hellwig 1476b15cd800SMatthew Wilcox entry = grab_mapping_entry(&xas, mapping, 0); 1477b15cd800SMatthew Wilcox if (xa_is_internal(entry)) { 1478b15cd800SMatthew Wilcox ret = xa_to_internal(entry); 147913e451fdSJan Kara goto out; 148013e451fdSJan Kara } 148113e451fdSJan Kara 1482a7d73fe6SChristoph Hellwig /* 1483e2093926SRoss Zwisler * It is possible, particularly with mixed reads & writes to private 1484e2093926SRoss Zwisler * mappings, that we have raced with a PMD fault that overlaps with 1485e2093926SRoss Zwisler * the PTE we need to set up. If so just return and the fault will be 1486e2093926SRoss Zwisler * retried. 1487e2093926SRoss Zwisler */ 1488e2093926SRoss Zwisler if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { 1489ab77dab4SSouptick Joarder ret = VM_FAULT_NOPAGE; 1490e2093926SRoss Zwisler goto unlock_entry; 1491e2093926SRoss Zwisler } 1492e2093926SRoss Zwisler 149365dd814aSChristoph Hellwig while ((error = iomap_iter(&iter, ops)) > 0) { 149465dd814aSChristoph Hellwig if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) { 149565dd814aSChristoph Hellwig iter.processed = -EIO; /* fs corruption? */ 149665dd814aSChristoph Hellwig continue; 149765dd814aSChristoph Hellwig } 149865dd814aSChristoph Hellwig 149965dd814aSChristoph Hellwig ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false); 150065dd814aSChristoph Hellwig if (ret != VM_FAULT_SIGBUS && 150165dd814aSChristoph Hellwig (iter.iomap.flags & IOMAP_F_NEW)) { 150265dd814aSChristoph Hellwig count_vm_event(PGMAJFAULT); 150365dd814aSChristoph Hellwig count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); 150465dd814aSChristoph Hellwig ret |= VM_FAULT_MAJOR; 150565dd814aSChristoph Hellwig } 150665dd814aSChristoph Hellwig 150765dd814aSChristoph Hellwig if (!(ret & VM_FAULT_ERROR)) 150865dd814aSChristoph Hellwig iter.processed = PAGE_SIZE; 150965dd814aSChristoph Hellwig } 151065dd814aSChristoph Hellwig 1511c0b24625SJan Kara if (iomap_errp) 1512c0b24625SJan Kara *iomap_errp = error; 151365dd814aSChristoph Hellwig if (!ret && error) 1514ab77dab4SSouptick Joarder ret = dax_fault_return(error); 1515a7d73fe6SChristoph Hellwig 151613e451fdSJan Kara unlock_entry: 1517b15cd800SMatthew Wilcox dax_unlock_entry(&xas, entry); 1518a9c42b33SRoss Zwisler out: 151965dd814aSChristoph Hellwig trace_dax_pte_fault_done(iter.inode, vmf, ret); 152065dd814aSChristoph Hellwig return ret; 1521a7d73fe6SChristoph Hellwig } 1522642261acSRoss Zwisler 1523642261acSRoss Zwisler #ifdef CONFIG_FS_DAX_PMD 152455f81639SShiyang Ruan static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas, 152555f81639SShiyang Ruan pgoff_t max_pgoff) 1526642261acSRoss Zwisler { 1527d8a849e1SDave Jiang unsigned long pmd_addr = vmf->address & PMD_MASK; 1528d8a849e1SDave Jiang bool write = vmf->flags & FAULT_FLAG_WRITE; 1529282a8e03SRoss Zwisler 1530fffa281bSRoss Zwisler /* 1531fffa281bSRoss Zwisler * Make sure that the faulting address's PMD offset (color) matches 1532fffa281bSRoss Zwisler * the PMD offset from the start of the file. This is necessary so 1533fffa281bSRoss Zwisler * that a PMD range in the page table overlaps exactly with a PMD 1534a77d19f4SMatthew Wilcox * range in the page cache. 1535fffa281bSRoss Zwisler */ 1536fffa281bSRoss Zwisler if ((vmf->pgoff & PG_PMD_COLOUR) != 1537fffa281bSRoss Zwisler ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) 153855f81639SShiyang Ruan return true; 1539fffa281bSRoss Zwisler 1540642261acSRoss Zwisler /* Fall back to PTEs if we're going to COW */ 154155f81639SShiyang Ruan if (write && !(vmf->vma->vm_flags & VM_SHARED)) 154255f81639SShiyang Ruan return true; 1543642261acSRoss Zwisler 1544642261acSRoss Zwisler /* If the PMD would extend outside the VMA */ 154555f81639SShiyang Ruan if (pmd_addr < vmf->vma->vm_start) 154655f81639SShiyang Ruan return true; 154755f81639SShiyang Ruan if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end) 154855f81639SShiyang Ruan return true; 154955f81639SShiyang Ruan 155055f81639SShiyang Ruan /* If the PMD would extend beyond the file size */ 155155f81639SShiyang Ruan if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff) 155255f81639SShiyang Ruan return true; 155355f81639SShiyang Ruan 155455f81639SShiyang Ruan return false; 155555f81639SShiyang Ruan } 155655f81639SShiyang Ruan 1557642261acSRoss Zwisler static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1558642261acSRoss Zwisler const struct iomap_ops *ops) 1559642261acSRoss Zwisler { 156065dd814aSChristoph Hellwig struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1561642261acSRoss Zwisler XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); 156265dd814aSChristoph Hellwig struct iomap_iter iter = { 156365dd814aSChristoph Hellwig .inode = mapping->host, 156465dd814aSChristoph Hellwig .len = PMD_SIZE, 1565952da063SChristoph Hellwig .flags = IOMAP_DAX | IOMAP_FAULT, 156665dd814aSChristoph Hellwig }; 1567c2436190SShiyang Ruan vm_fault_t ret = VM_FAULT_FALLBACK; 1568642261acSRoss Zwisler pgoff_t max_pgoff; 1569642261acSRoss Zwisler void *entry; 1570642261acSRoss Zwisler int error; 1571642261acSRoss Zwisler 157265dd814aSChristoph Hellwig if (vmf->flags & FAULT_FLAG_WRITE) 157365dd814aSChristoph Hellwig iter.flags |= IOMAP_WRITE; 157465dd814aSChristoph Hellwig 1575642261acSRoss Zwisler /* 1576642261acSRoss Zwisler * Check whether offset isn't beyond end of file now. Caller is 1577642261acSRoss Zwisler * supposed to hold locks serializing us with truncate / punch hole so 1578642261acSRoss Zwisler * this is a reliable test. 1579642261acSRoss Zwisler */ 158065dd814aSChristoph Hellwig max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE); 1581642261acSRoss Zwisler 158265dd814aSChristoph Hellwig trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0); 1583642261acSRoss Zwisler 1584b15cd800SMatthew Wilcox if (xas.xa_index >= max_pgoff) { 1585c2436190SShiyang Ruan ret = VM_FAULT_SIGBUS; 1586282a8e03SRoss Zwisler goto out; 1587282a8e03SRoss Zwisler } 1588642261acSRoss Zwisler 158955f81639SShiyang Ruan if (dax_fault_check_fallback(vmf, &xas, max_pgoff)) 1590642261acSRoss Zwisler goto fallback; 1591642261acSRoss Zwisler 1592642261acSRoss Zwisler /* 1593b15cd800SMatthew Wilcox * grab_mapping_entry() will make sure we get an empty PMD entry, 1594b15cd800SMatthew Wilcox * a zero PMD entry or a DAX PMD. If it can't (because a PTE 1595b15cd800SMatthew Wilcox * entry is already in the array, for instance), it will return 1596b15cd800SMatthew Wilcox * VM_FAULT_FALLBACK. 15979f141d6eSJan Kara */ 159823c84eb7SMatthew Wilcox (Oracle) entry = grab_mapping_entry(&xas, mapping, PMD_ORDER); 1599b15cd800SMatthew Wilcox if (xa_is_internal(entry)) { 1600c2436190SShiyang Ruan ret = xa_to_internal(entry); 1601876f2946SRoss Zwisler goto fallback; 1602b15cd800SMatthew Wilcox } 1603876f2946SRoss Zwisler 1604876f2946SRoss Zwisler /* 1605e2093926SRoss Zwisler * It is possible, particularly with mixed reads & writes to private 1606e2093926SRoss Zwisler * mappings, that we have raced with a PTE fault that overlaps with 1607e2093926SRoss Zwisler * the PMD we need to set up. If so just return and the fault will be 1608e2093926SRoss Zwisler * retried. 1609e2093926SRoss Zwisler */ 1610e2093926SRoss Zwisler if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && 1611e2093926SRoss Zwisler !pmd_devmap(*vmf->pmd)) { 1612c2436190SShiyang Ruan ret = 0; 1613e2093926SRoss Zwisler goto unlock_entry; 1614e2093926SRoss Zwisler } 1615e2093926SRoss Zwisler 161665dd814aSChristoph Hellwig iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT; 161765dd814aSChristoph Hellwig while ((error = iomap_iter(&iter, ops)) > 0) { 161865dd814aSChristoph Hellwig if (iomap_length(&iter) < PMD_SIZE) 161965dd814aSChristoph Hellwig continue; /* actually breaks out of the loop */ 1620876f2946SRoss Zwisler 162165dd814aSChristoph Hellwig ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true); 162265dd814aSChristoph Hellwig if (ret != VM_FAULT_FALLBACK) 162365dd814aSChristoph Hellwig iter.processed = PMD_SIZE; 1624caa51d26SJan Kara } 1625caa51d26SJan Kara 1626876f2946SRoss Zwisler unlock_entry: 1627b15cd800SMatthew Wilcox dax_unlock_entry(&xas, entry); 1628642261acSRoss Zwisler fallback: 1629c2436190SShiyang Ruan if (ret == VM_FAULT_FALLBACK) { 163065dd814aSChristoph Hellwig split_huge_pmd(vmf->vma, vmf->pmd, vmf->address); 1631642261acSRoss Zwisler count_vm_event(THP_FAULT_FALLBACK); 1632642261acSRoss Zwisler } 1633282a8e03SRoss Zwisler out: 163465dd814aSChristoph Hellwig trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret); 1635c2436190SShiyang Ruan return ret; 1636642261acSRoss Zwisler } 1637a2d58167SDave Jiang #else 1638ab77dab4SSouptick Joarder static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 163901cddfe9SArnd Bergmann const struct iomap_ops *ops) 1640a2d58167SDave Jiang { 1641a2d58167SDave Jiang return VM_FAULT_FALLBACK; 1642a2d58167SDave Jiang } 1643642261acSRoss Zwisler #endif /* CONFIG_FS_DAX_PMD */ 1644a2d58167SDave Jiang 1645a2d58167SDave Jiang /** 1646a2d58167SDave Jiang * dax_iomap_fault - handle a page fault on a DAX file 1647a2d58167SDave Jiang * @vmf: The description of the fault 1648cec04e8cSJan Kara * @pe_size: Size of the page to fault in 16499a0dd422SJan Kara * @pfnp: PFN to insert for synchronous faults if fsync is required 1650c0b24625SJan Kara * @iomap_errp: Storage for detailed error code in case of error 1651cec04e8cSJan Kara * @ops: Iomap ops passed from the file system 1652a2d58167SDave Jiang * 1653a2d58167SDave Jiang * When a page fault occurs, filesystems may call this helper in 1654a2d58167SDave Jiang * their fault handler for DAX files. dax_iomap_fault() assumes the caller 1655a2d58167SDave Jiang * has done all the necessary locking for page fault to proceed 1656a2d58167SDave Jiang * successfully. 1657a2d58167SDave Jiang */ 1658ab77dab4SSouptick Joarder vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, 1659c0b24625SJan Kara pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) 1660a2d58167SDave Jiang { 1661c791ace1SDave Jiang switch (pe_size) { 1662c791ace1SDave Jiang case PE_SIZE_PTE: 1663c0b24625SJan Kara return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); 1664c791ace1SDave Jiang case PE_SIZE_PMD: 16659a0dd422SJan Kara return dax_iomap_pmd_fault(vmf, pfnp, ops); 1666a2d58167SDave Jiang default: 1667a2d58167SDave Jiang return VM_FAULT_FALLBACK; 1668a2d58167SDave Jiang } 1669a2d58167SDave Jiang } 1670a2d58167SDave Jiang EXPORT_SYMBOL_GPL(dax_iomap_fault); 167171eab6dfSJan Kara 1672a77d19f4SMatthew Wilcox /* 167371eab6dfSJan Kara * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables 167471eab6dfSJan Kara * @vmf: The description of the fault 167571eab6dfSJan Kara * @pfn: PFN to insert 1676cfc93c6cSMatthew Wilcox * @order: Order of entry to insert. 167771eab6dfSJan Kara * 1678a77d19f4SMatthew Wilcox * This function inserts a writeable PTE or PMD entry into the page tables 1679a77d19f4SMatthew Wilcox * for an mmaped DAX file. It also marks the page cache entry as dirty. 168071eab6dfSJan Kara */ 1681cfc93c6cSMatthew Wilcox static vm_fault_t 1682cfc93c6cSMatthew Wilcox dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) 168371eab6dfSJan Kara { 168471eab6dfSJan Kara struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1685cfc93c6cSMatthew Wilcox XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); 1686cfc93c6cSMatthew Wilcox void *entry; 1687ab77dab4SSouptick Joarder vm_fault_t ret; 168871eab6dfSJan Kara 1689cfc93c6cSMatthew Wilcox xas_lock_irq(&xas); 169023c84eb7SMatthew Wilcox (Oracle) entry = get_unlocked_entry(&xas, order); 169171eab6dfSJan Kara /* Did we race with someone splitting entry or so? */ 169223c84eb7SMatthew Wilcox (Oracle) if (!entry || dax_is_conflict(entry) || 169323c84eb7SMatthew Wilcox (Oracle) (order == 0 && !dax_is_pte_entry(entry))) { 16944c3d043dSVivek Goyal put_unlocked_entry(&xas, entry, WAKE_NEXT); 1695cfc93c6cSMatthew Wilcox xas_unlock_irq(&xas); 169671eab6dfSJan Kara trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, 169771eab6dfSJan Kara VM_FAULT_NOPAGE); 169871eab6dfSJan Kara return VM_FAULT_NOPAGE; 169971eab6dfSJan Kara } 1700cfc93c6cSMatthew Wilcox xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); 1701cfc93c6cSMatthew Wilcox dax_lock_entry(&xas, entry); 1702cfc93c6cSMatthew Wilcox xas_unlock_irq(&xas); 1703cfc93c6cSMatthew Wilcox if (order == 0) 1704ab77dab4SSouptick Joarder ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 170571eab6dfSJan Kara #ifdef CONFIG_FS_DAX_PMD 1706cfc93c6cSMatthew Wilcox else if (order == PMD_ORDER) 1707fce86ff5SDan Williams ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE); 170871eab6dfSJan Kara #endif 1709cfc93c6cSMatthew Wilcox else 1710ab77dab4SSouptick Joarder ret = VM_FAULT_FALLBACK; 1711cfc93c6cSMatthew Wilcox dax_unlock_entry(&xas, entry); 1712ab77dab4SSouptick Joarder trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); 1713ab77dab4SSouptick Joarder return ret; 171471eab6dfSJan Kara } 171571eab6dfSJan Kara 171671eab6dfSJan Kara /** 171771eab6dfSJan Kara * dax_finish_sync_fault - finish synchronous page fault 171871eab6dfSJan Kara * @vmf: The description of the fault 171971eab6dfSJan Kara * @pe_size: Size of entry to be inserted 172071eab6dfSJan Kara * @pfn: PFN to insert 172171eab6dfSJan Kara * 172271eab6dfSJan Kara * This function ensures that the file range touched by the page fault is 172371eab6dfSJan Kara * stored persistently on the media and handles inserting of appropriate page 172471eab6dfSJan Kara * table entry. 172571eab6dfSJan Kara */ 1726ab77dab4SSouptick Joarder vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, 1727ab77dab4SSouptick Joarder enum page_entry_size pe_size, pfn_t pfn) 172871eab6dfSJan Kara { 172971eab6dfSJan Kara int err; 173071eab6dfSJan Kara loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; 1731cfc93c6cSMatthew Wilcox unsigned int order = pe_order(pe_size); 1732cfc93c6cSMatthew Wilcox size_t len = PAGE_SIZE << order; 173371eab6dfSJan Kara 173471eab6dfSJan Kara err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); 173571eab6dfSJan Kara if (err) 173671eab6dfSJan Kara return VM_FAULT_SIGBUS; 1737cfc93c6cSMatthew Wilcox return dax_insert_pfn_mkwrite(vmf, pfn, order); 173871eab6dfSJan Kara } 173971eab6dfSJan Kara EXPORT_SYMBOL_GPL(dax_finish_sync_fault); 1740