Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * fs/dax.c - Direct Access filesystem code
   4 * Copyright (c) 2013-2014 Intel Corporation
   5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   7 */
   8
   9#include <linux/atomic.h>
  10#include <linux/blkdev.h>
  11#include <linux/buffer_head.h>
  12#include <linux/dax.h>
  13#include <linux/fs.h>
  14#include <linux/highmem.h>
  15#include <linux/memcontrol.h>
  16#include <linux/mm.h>
  17#include <linux/mutex.h>
  18#include <linux/pagevec.h>
  19#include <linux/sched.h>
  20#include <linux/sched/signal.h>
  21#include <linux/uio.h>
  22#include <linux/vmstat.h>
  23#include <linux/pfn_t.h>
  24#include <linux/sizes.h>
  25#include <linux/mmu_notifier.h>
  26#include <linux/iomap.h>
  27#include <linux/rmap.h>
  28#include <asm/pgalloc.h>
  29
  30#define CREATE_TRACE_POINTS
  31#include <trace/events/fs_dax.h>
  32
  33static inline unsigned int pe_order(enum page_entry_size pe_size)
  34{
  35	if (pe_size == PE_SIZE_PTE)
  36		return PAGE_SHIFT - PAGE_SHIFT;
  37	if (pe_size == PE_SIZE_PMD)
  38		return PMD_SHIFT - PAGE_SHIFT;
  39	if (pe_size == PE_SIZE_PUD)
  40		return PUD_SHIFT - PAGE_SHIFT;
  41	return ~0;
  42}
  43
  44/* We choose 4096 entries - same as per-zone page wait tables */
  45#define DAX_WAIT_TABLE_BITS 12
  46#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  47
  48/* The 'colour' (ie low bits) within a PMD of a page offset.  */
  49#define PG_PMD_COLOUR	((PMD_SIZE >> PAGE_SHIFT) - 1)
  50#define PG_PMD_NR	(PMD_SIZE >> PAGE_SHIFT)
  51
  52/* The order of a PMD entry */
  53#define PMD_ORDER	(PMD_SHIFT - PAGE_SHIFT)
  54
  55static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  56
  57static int __init init_dax_wait_table(void)
  58{
  59	int i;
  60
  61	for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  62		init_waitqueue_head(wait_table + i);
  63	return 0;
  64}
  65fs_initcall(init_dax_wait_table);
  66
  67/*
  68 * DAX pagecache entries use XArray value entries so they can't be mistaken
  69 * for pages.  We use one bit for locking, one bit for the entry size (PMD)
  70 * and two more to tell us if the entry is a zero page or an empty entry that
  71 * is just used for locking.  In total four special bits.
  72 *
  73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
  74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
  75 * block allocation.
  76 */
  77#define DAX_SHIFT	(4)
  78#define DAX_LOCKED	(1UL << 0)
  79#define DAX_PMD		(1UL << 1)
  80#define DAX_ZERO_PAGE	(1UL << 2)
  81#define DAX_EMPTY	(1UL << 3)
  82
  83static unsigned long dax_to_pfn(void *entry)
  84{
  85	return xa_to_value(entry) >> DAX_SHIFT;
  86}
  87
  88static void *dax_make_entry(pfn_t pfn, unsigned long flags)
  89{
  90	return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
  91}
  92
  93static bool dax_is_locked(void *entry)
  94{
  95	return xa_to_value(entry) & DAX_LOCKED;
  96}
  97
  98static unsigned int dax_entry_order(void *entry)
  99{
 100	if (xa_to_value(entry) & DAX_PMD)
 101		return PMD_ORDER;
 102	return 0;
 103}
 104
 105static unsigned long dax_is_pmd_entry(void *entry)
 106{
 107	return xa_to_value(entry) & DAX_PMD;
 108}
 109
 110static bool dax_is_pte_entry(void *entry)
 111{
 112	return !(xa_to_value(entry) & DAX_PMD);
 113}
 114
 115static int dax_is_zero_entry(void *entry)
 116{
 117	return xa_to_value(entry) & DAX_ZERO_PAGE;
 118}
 119
 120static int dax_is_empty_entry(void *entry)
 121{
 122	return xa_to_value(entry) & DAX_EMPTY;
 123}
 124
 125/*
 126 * true if the entry that was found is of a smaller order than the entry
 127 * we were looking for
 128 */
 129static bool dax_is_conflict(void *entry)
 130{
 131	return entry == XA_RETRY_ENTRY;
 132}
 133
 134/*
 135 * DAX page cache entry locking
 136 */
 137struct exceptional_entry_key {
 138	struct xarray *xa;
 139	pgoff_t entry_start;
 140};
 141
 142struct wait_exceptional_entry_queue {
 143	wait_queue_entry_t wait;
 144	struct exceptional_entry_key key;
 145};
 146
 147/**
 148 * enum dax_wake_mode: waitqueue wakeup behaviour
 149 * @WAKE_ALL: wake all waiters in the waitqueue
 150 * @WAKE_NEXT: wake only the first waiter in the waitqueue
 151 */
 152enum dax_wake_mode {
 153	WAKE_ALL,
 154	WAKE_NEXT,
 155};
 156
 157static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
 158		void *entry, struct exceptional_entry_key *key)
 159{
 160	unsigned long hash;
 161	unsigned long index = xas->xa_index;
 162
 163	/*
 164	 * If 'entry' is a PMD, align the 'index' that we use for the wait
 165	 * queue to the start of that PMD.  This ensures that all offsets in
 166	 * the range covered by the PMD map to the same bit lock.
 167	 */
 168	if (dax_is_pmd_entry(entry))
 169		index &= ~PG_PMD_COLOUR;
 170	key->xa = xas->xa;
 171	key->entry_start = index;
 172
 173	hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
 174	return wait_table + hash;
 175}
 176
 177static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
 178		unsigned int mode, int sync, void *keyp)
 179{
 180	struct exceptional_entry_key *key = keyp;
 181	struct wait_exceptional_entry_queue *ewait =
 182		container_of(wait, struct wait_exceptional_entry_queue, wait);
 183
 184	if (key->xa != ewait->key.xa ||
 185	    key->entry_start != ewait->key.entry_start)
 186		return 0;
 187	return autoremove_wake_function(wait, mode, sync, NULL);
 188}
 189
 190/*
 191 * @entry may no longer be the entry at the index in the mapping.
 192 * The important information it's conveying is whether the entry at
 193 * this index used to be a PMD entry.
 194 */
 195static void dax_wake_entry(struct xa_state *xas, void *entry,
 196			   enum dax_wake_mode mode)
 197{
 198	struct exceptional_entry_key key;
 199	wait_queue_head_t *wq;
 200
 201	wq = dax_entry_waitqueue(xas, entry, &key);
 202
 203	/*
 204	 * Checking for locked entry and prepare_to_wait_exclusive() happens
 205	 * under the i_pages lock, ditto for entry handling in our callers.
 206	 * So at this point all tasks that could have seen our entry locked
 207	 * must be in the waitqueue and the following check will see them.
 208	 */
 209	if (waitqueue_active(wq))
 210		__wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
 211}
 212
 213/*
 214 * Look up entry in page cache, wait for it to become unlocked if it
 215 * is a DAX entry and return it.  The caller must subsequently call
 216 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
 217 * if it did.  The entry returned may have a larger order than @order.
 218 * If @order is larger than the order of the entry found in i_pages, this
 219 * function returns a dax_is_conflict entry.
 220 *
 221 * Must be called with the i_pages lock held.
 222 */
 223static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
 224{
 225	void *entry;
 226	struct wait_exceptional_entry_queue ewait;
 227	wait_queue_head_t *wq;
 228
 229	init_wait(&ewait.wait);
 230	ewait.wait.func = wake_exceptional_entry_func;
 231
 232	for (;;) {
 233		entry = xas_find_conflict(xas);
 234		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 235			return entry;
 236		if (dax_entry_order(entry) < order)
 237			return XA_RETRY_ENTRY;
 238		if (!dax_is_locked(entry))
 239			return entry;
 240
 241		wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 242		prepare_to_wait_exclusive(wq, &ewait.wait,
 243					  TASK_UNINTERRUPTIBLE);
 244		xas_unlock_irq(xas);
 245		xas_reset(xas);
 246		schedule();
 247		finish_wait(wq, &ewait.wait);
 248		xas_lock_irq(xas);
 249	}
 250}
 251
 252/*
 253 * The only thing keeping the address space around is the i_pages lock
 254 * (it's cycled in clear_inode() after removing the entries from i_pages)
 255 * After we call xas_unlock_irq(), we cannot touch xas->xa.
 256 */
 257static void wait_entry_unlocked(struct xa_state *xas, void *entry)
 258{
 259	struct wait_exceptional_entry_queue ewait;
 260	wait_queue_head_t *wq;
 261
 262	init_wait(&ewait.wait);
 263	ewait.wait.func = wake_exceptional_entry_func;
 264
 265	wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 266	/*
 267	 * Unlike get_unlocked_entry() there is no guarantee that this
 268	 * path ever successfully retrieves an unlocked entry before an
 269	 * inode dies. Perform a non-exclusive wait in case this path
 270	 * never successfully performs its own wake up.
 271	 */
 272	prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
 273	xas_unlock_irq(xas);
 274	schedule();
 275	finish_wait(wq, &ewait.wait);
 276}
 277
 278static void put_unlocked_entry(struct xa_state *xas, void *entry,
 279			       enum dax_wake_mode mode)
 280{
 281	if (entry && !dax_is_conflict(entry))
 282		dax_wake_entry(xas, entry, mode);
 283}
 284
 285/*
 286 * We used the xa_state to get the entry, but then we locked the entry and
 287 * dropped the xa_lock, so we know the xa_state is stale and must be reset
 288 * before use.
 289 */
 290static void dax_unlock_entry(struct xa_state *xas, void *entry)
 291{
 292	void *old;
 293
 294	BUG_ON(dax_is_locked(entry));
 295	xas_reset(xas);
 296	xas_lock_irq(xas);
 297	old = xas_store(xas, entry);
 298	xas_unlock_irq(xas);
 299	BUG_ON(!dax_is_locked(old));
 300	dax_wake_entry(xas, entry, WAKE_NEXT);
 301}
 302
 303/*
 304 * Return: The entry stored at this location before it was locked.
 305 */
 306static void *dax_lock_entry(struct xa_state *xas, void *entry)
 307{
 308	unsigned long v = xa_to_value(entry);
 309	return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
 310}
 311
 312static unsigned long dax_entry_size(void *entry)
 313{
 314	if (dax_is_zero_entry(entry))
 315		return 0;
 316	else if (dax_is_empty_entry(entry))
 317		return 0;
 318	else if (dax_is_pmd_entry(entry))
 319		return PMD_SIZE;
 320	else
 321		return PAGE_SIZE;
 322}
 323
 324static unsigned long dax_end_pfn(void *entry)
 325{
 326	return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
 327}
 328
 329/*
 330 * Iterate through all mapped pfns represented by an entry, i.e. skip
 331 * 'empty' and 'zero' entries.
 332 */
 333#define for_each_mapped_pfn(entry, pfn) \
 334	for (pfn = dax_to_pfn(entry); \
 335			pfn < dax_end_pfn(entry); pfn++)
 336
 337static inline bool dax_page_is_shared(struct page *page)
 338{
 339	return page->mapping == PAGE_MAPPING_DAX_SHARED;
 340}
 341
 342/*
 343 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
 344 * refcount.
 345 */
 346static inline void dax_page_share_get(struct page *page)
 347{
 348	if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
 349		/*
 350		 * Reset the index if the page was already mapped
 351		 * regularly before.
 352		 */
 353		if (page->mapping)
 354			page->share = 1;
 355		page->mapping = PAGE_MAPPING_DAX_SHARED;
 356	}
 357	page->share++;
 358}
 359
 360static inline unsigned long dax_page_share_put(struct page *page)
 361{
 362	return --page->share;
 363}
 364
 365/*
 366 * When it is called in dax_insert_entry(), the shared flag will indicate that
 367 * whether this entry is shared by multiple files.  If so, set the page->mapping
 368 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
 369 */
 370static void dax_associate_entry(void *entry, struct address_space *mapping,
 371		struct vm_area_struct *vma, unsigned long address, bool shared)
 372{
 373	unsigned long size = dax_entry_size(entry), pfn, index;
 374	int i = 0;
 375
 376	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 377		return;
 378
 379	index = linear_page_index(vma, address & ~(size - 1));
 380	for_each_mapped_pfn(entry, pfn) {
 381		struct page *page = pfn_to_page(pfn);
 382
 383		if (shared) {
 384			dax_page_share_get(page);
 385		} else {
 386			WARN_ON_ONCE(page->mapping);
 387			page->mapping = mapping;
 388			page->index = index + i++;
 389		}
 390	}
 391}
 392
 393static void dax_disassociate_entry(void *entry, struct address_space *mapping,
 394		bool trunc)
 395{
 396	unsigned long pfn;
 397
 398	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 399		return;
 400
 401	for_each_mapped_pfn(entry, pfn) {
 402		struct page *page = pfn_to_page(pfn);
 403
 404		WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
 405		if (dax_page_is_shared(page)) {
 406			/* keep the shared flag if this page is still shared */
 407			if (dax_page_share_put(page) > 0)
 408				continue;
 409		} else
 410			WARN_ON_ONCE(page->mapping && page->mapping != mapping);
 411		page->mapping = NULL;
 412		page->index = 0;
 413	}
 414}
 415
 416static struct page *dax_busy_page(void *entry)
 417{
 418	unsigned long pfn;
 419
 420	for_each_mapped_pfn(entry, pfn) {
 421		struct page *page = pfn_to_page(pfn);
 422
 423		if (page_ref_count(page) > 1)
 424			return page;
 425	}
 426	return NULL;
 427}
 428
 429/*
 430 * dax_lock_page - Lock the DAX entry corresponding to a page
 431 * @page: The page whose entry we want to lock
 432 *
 433 * Context: Process context.
 434 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could
 435 * not be locked.
 436 */
 437dax_entry_t dax_lock_page(struct page *page)
 438{
 439	XA_STATE(xas, NULL, 0);
 440	void *entry;
 441
 442	/* Ensure page->mapping isn't freed while we look at it */
 443	rcu_read_lock();
 444	for (;;) {
 445		struct address_space *mapping = READ_ONCE(page->mapping);
 446
 447		entry = NULL;
 448		if (!mapping || !dax_mapping(mapping))
 449			break;
 450
 451		/*
 452		 * In the device-dax case there's no need to lock, a
 453		 * struct dev_pagemap pin is sufficient to keep the
 454		 * inode alive, and we assume we have dev_pagemap pin
 455		 * otherwise we would not have a valid pfn_to_page()
 456		 * translation.
 457		 */
 458		entry = (void *)~0UL;
 459		if (S_ISCHR(mapping->host->i_mode))
 460			break;
 461
 462		xas.xa = &mapping->i_pages;
 463		xas_lock_irq(&xas);
 464		if (mapping != page->mapping) {
 465			xas_unlock_irq(&xas);
 466			continue;
 467		}
 468		xas_set(&xas, page->index);
 469		entry = xas_load(&xas);
 470		if (dax_is_locked(entry)) {
 471			rcu_read_unlock();
 472			wait_entry_unlocked(&xas, entry);
 473			rcu_read_lock();
 474			continue;
 475		}
 476		dax_lock_entry(&xas, entry);
 477		xas_unlock_irq(&xas);
 478		break;
 479	}
 480	rcu_read_unlock();
 481	return (dax_entry_t)entry;
 482}
 483
 484void dax_unlock_page(struct page *page, dax_entry_t cookie)
 485{
 486	struct address_space *mapping = page->mapping;
 487	XA_STATE(xas, &mapping->i_pages, page->index);
 488
 489	if (S_ISCHR(mapping->host->i_mode))
 490		return;
 491
 492	dax_unlock_entry(&xas, (void *)cookie);
 493}
 494
 495/*
 496 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
 497 * @mapping: the file's mapping whose entry we want to lock
 498 * @index: the offset within this file
 499 * @page: output the dax page corresponding to this dax entry
 500 *
 501 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
 502 * could not be locked.
 503 */
 504dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
 505		struct page **page)
 506{
 507	XA_STATE(xas, NULL, 0);
 508	void *entry;
 509
 510	rcu_read_lock();
 511	for (;;) {
 512		entry = NULL;
 513		if (!dax_mapping(mapping))
 514			break;
 515
 516		xas.xa = &mapping->i_pages;
 517		xas_lock_irq(&xas);
 518		xas_set(&xas, index);
 519		entry = xas_load(&xas);
 520		if (dax_is_locked(entry)) {
 521			rcu_read_unlock();
 522			wait_entry_unlocked(&xas, entry);
 523			rcu_read_lock();
 524			continue;
 525		}
 526		if (!entry ||
 527		    dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 528			/*
 529			 * Because we are looking for entry from file's mapping
 530			 * and index, so the entry may not be inserted for now,
 531			 * or even a zero/empty entry.  We don't think this is
 532			 * an error case.  So, return a special value and do
 533			 * not output @page.
 534			 */
 535			entry = (void *)~0UL;
 536		} else {
 537			*page = pfn_to_page(dax_to_pfn(entry));
 538			dax_lock_entry(&xas, entry);
 539		}
 540		xas_unlock_irq(&xas);
 541		break;
 542	}
 543	rcu_read_unlock();
 544	return (dax_entry_t)entry;
 545}
 546
 547void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
 548		dax_entry_t cookie)
 549{
 550	XA_STATE(xas, &mapping->i_pages, index);
 551
 552	if (cookie == ~0UL)
 553		return;
 554
 555	dax_unlock_entry(&xas, (void *)cookie);
 556}
 557
 558/*
 559 * Find page cache entry at given index. If it is a DAX entry, return it
 560 * with the entry locked. If the page cache doesn't contain an entry at
 561 * that index, add a locked empty entry.
 562 *
 563 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
 564 * either return that locked entry or will return VM_FAULT_FALLBACK.
 565 * This will happen if there are any PTE entries within the PMD range
 566 * that we are requesting.
 567 *
 568 * We always favor PTE entries over PMD entries. There isn't a flow where we
 569 * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
 570 * insertion will fail if it finds any PTE entries already in the tree, and a
 571 * PTE insertion will cause an existing PMD entry to be unmapped and
 572 * downgraded to PTE entries.  This happens for both PMD zero pages as
 573 * well as PMD empty entries.
 574 *
 575 * The exception to this downgrade path is for PMD entries that have
 576 * real storage backing them.  We will leave these real PMD entries in
 577 * the tree, and PTE writes will simply dirty the entire PMD entry.
 578 *
 579 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 580 * persistent memory the benefit is doubtful. We can add that later if we can
 581 * show it helps.
 582 *
 583 * On error, this function does not return an ERR_PTR.  Instead it returns
 584 * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
 585 * overlap with xarray value entries.
 586 */
 587static void *grab_mapping_entry(struct xa_state *xas,
 588		struct address_space *mapping, unsigned int order)
 589{
 590	unsigned long index = xas->xa_index;
 591	bool pmd_downgrade;	/* splitting PMD entry into PTE entries? */
 592	void *entry;
 593
 594retry:
 595	pmd_downgrade = false;
 596	xas_lock_irq(xas);
 597	entry = get_unlocked_entry(xas, order);
 598
 599	if (entry) {
 600		if (dax_is_conflict(entry))
 601			goto fallback;
 602		if (!xa_is_value(entry)) {
 603			xas_set_err(xas, -EIO);
 604			goto out_unlock;
 605		}
 606
 607		if (order == 0) {
 608			if (dax_is_pmd_entry(entry) &&
 609			    (dax_is_zero_entry(entry) ||
 610			     dax_is_empty_entry(entry))) {
 611				pmd_downgrade = true;
 612			}
 613		}
 614	}
 615
 616	if (pmd_downgrade) {
 617		/*
 618		 * Make sure 'entry' remains valid while we drop
 619		 * the i_pages lock.
 620		 */
 621		dax_lock_entry(xas, entry);
 622
 623		/*
 624		 * Besides huge zero pages the only other thing that gets
 625		 * downgraded are empty entries which don't need to be
 626		 * unmapped.
 627		 */
 628		if (dax_is_zero_entry(entry)) {
 629			xas_unlock_irq(xas);
 630			unmap_mapping_pages(mapping,
 631					xas->xa_index & ~PG_PMD_COLOUR,
 632					PG_PMD_NR, false);
 633			xas_reset(xas);
 634			xas_lock_irq(xas);
 635		}
 636
 637		dax_disassociate_entry(entry, mapping, false);
 638		xas_store(xas, NULL);	/* undo the PMD join */
 639		dax_wake_entry(xas, entry, WAKE_ALL);
 640		mapping->nrpages -= PG_PMD_NR;
 641		entry = NULL;
 642		xas_set(xas, index);
 643	}
 644
 645	if (entry) {
 646		dax_lock_entry(xas, entry);
 647	} else {
 648		unsigned long flags = DAX_EMPTY;
 649
 650		if (order > 0)
 651			flags |= DAX_PMD;
 652		entry = dax_make_entry(pfn_to_pfn_t(0), flags);
 653		dax_lock_entry(xas, entry);
 654		if (xas_error(xas))
 655			goto out_unlock;
 656		mapping->nrpages += 1UL << order;
 657	}
 658
 659out_unlock:
 660	xas_unlock_irq(xas);
 661	if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
 662		goto retry;
 663	if (xas->xa_node == XA_ERROR(-ENOMEM))
 664		return xa_mk_internal(VM_FAULT_OOM);
 665	if (xas_error(xas))
 666		return xa_mk_internal(VM_FAULT_SIGBUS);
 667	return entry;
 668fallback:
 669	xas_unlock_irq(xas);
 670	return xa_mk_internal(VM_FAULT_FALLBACK);
 671}
 672
 673/**
 674 * dax_layout_busy_page_range - find first pinned page in @mapping
 675 * @mapping: address space to scan for a page with ref count > 1
 676 * @start: Starting offset. Page containing 'start' is included.
 677 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
 678 *       pages from 'start' till the end of file are included.
 679 *
 680 * DAX requires ZONE_DEVICE mapped pages. These pages are never
 681 * 'onlined' to the page allocator so they are considered idle when
 682 * page->count == 1. A filesystem uses this interface to determine if
 683 * any page in the mapping is busy, i.e. for DMA, or other
 684 * get_user_pages() usages.
 685 *
 686 * It is expected that the filesystem is holding locks to block the
 687 * establishment of new mappings in this address_space. I.e. it expects
 688 * to be able to run unmap_mapping_range() and subsequently not race
 689 * mapping_mapped() becoming true.
 690 */
 691struct page *dax_layout_busy_page_range(struct address_space *mapping,
 692					loff_t start, loff_t end)
 693{
 694	void *entry;
 695	unsigned int scanned = 0;
 696	struct page *page = NULL;
 697	pgoff_t start_idx = start >> PAGE_SHIFT;
 698	pgoff_t end_idx;
 699	XA_STATE(xas, &mapping->i_pages, start_idx);
 700
 701	/*
 702	 * In the 'limited' case get_user_pages() for dax is disabled.
 703	 */
 704	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 705		return NULL;
 706
 707	if (!dax_mapping(mapping) || !mapping_mapped(mapping))
 708		return NULL;
 709
 710	/* If end == LLONG_MAX, all pages from start to till end of file */
 711	if (end == LLONG_MAX)
 712		end_idx = ULONG_MAX;
 713	else
 714		end_idx = end >> PAGE_SHIFT;
 715	/*
 716	 * If we race get_user_pages_fast() here either we'll see the
 717	 * elevated page count in the iteration and wait, or
 718	 * get_user_pages_fast() will see that the page it took a reference
 719	 * against is no longer mapped in the page tables and bail to the
 720	 * get_user_pages() slow path.  The slow path is protected by
 721	 * pte_lock() and pmd_lock(). New references are not taken without
 722	 * holding those locks, and unmap_mapping_pages() will not zero the
 723	 * pte or pmd without holding the respective lock, so we are
 724	 * guaranteed to either see new references or prevent new
 725	 * references from being established.
 726	 */
 727	unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
 728
 729	xas_lock_irq(&xas);
 730	xas_for_each(&xas, entry, end_idx) {
 731		if (WARN_ON_ONCE(!xa_is_value(entry)))
 732			continue;
 733		if (unlikely(dax_is_locked(entry)))
 734			entry = get_unlocked_entry(&xas, 0);
 735		if (entry)
 736			page = dax_busy_page(entry);
 737		put_unlocked_entry(&xas, entry, WAKE_NEXT);
 738		if (page)
 739			break;
 740		if (++scanned % XA_CHECK_SCHED)
 741			continue;
 742
 743		xas_pause(&xas);
 744		xas_unlock_irq(&xas);
 745		cond_resched();
 746		xas_lock_irq(&xas);
 747	}
 748	xas_unlock_irq(&xas);
 749	return page;
 750}
 751EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
 752
 753struct page *dax_layout_busy_page(struct address_space *mapping)
 754{
 755	return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
 756}
 757EXPORT_SYMBOL_GPL(dax_layout_busy_page);
 758
 759static int __dax_invalidate_entry(struct address_space *mapping,
 760					  pgoff_t index, bool trunc)
 761{
 762	XA_STATE(xas, &mapping->i_pages, index);
 763	int ret = 0;
 764	void *entry;
 765
 766	xas_lock_irq(&xas);
 767	entry = get_unlocked_entry(&xas, 0);
 768	if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 769		goto out;
 770	if (!trunc &&
 771	    (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
 772	     xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
 773		goto out;
 774	dax_disassociate_entry(entry, mapping, trunc);
 775	xas_store(&xas, NULL);
 776	mapping->nrpages -= 1UL << dax_entry_order(entry);
 777	ret = 1;
 778out:
 779	put_unlocked_entry(&xas, entry, WAKE_ALL);
 780	xas_unlock_irq(&xas);
 781	return ret;
 782}
 783
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 784/*
 785 * Delete DAX entry at @index from @mapping.  Wait for it
 786 * to be unlocked before deleting it.
 787 */
 788int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 789{
 790	int ret = __dax_invalidate_entry(mapping, index, true);
 791
 792	/*
 793	 * This gets called from truncate / punch_hole path. As such, the caller
 794	 * must hold locks protecting against concurrent modifications of the
 795	 * page cache (usually fs-private i_mmap_sem for writing). Since the
 796	 * caller has seen a DAX entry for this index, we better find it
 797	 * at that index as well...
 798	 */
 799	WARN_ON_ONCE(!ret);
 800	return ret;
 801}
 802
 803/*
 804 * Invalidate DAX entry if it is clean.
 805 */
 806int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 807				      pgoff_t index)
 808{
 809	return __dax_invalidate_entry(mapping, index, false);
 810}
 811
 812static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
 813{
 814	return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
 815}
 816
 817static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
 818{
 819	pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
 820	void *vto, *kaddr;
 821	long rc;
 822	int id;
 823
 824	id = dax_read_lock();
 825	rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
 826				&kaddr, NULL);
 827	if (rc < 0) {
 828		dax_read_unlock(id);
 829		return rc;
 830	}
 831	vto = kmap_atomic(vmf->cow_page);
 832	copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
 833	kunmap_atomic(vto);
 834	dax_read_unlock(id);
 835	return 0;
 836}
 837
 838/*
 839 * MAP_SYNC on a dax mapping guarantees dirty metadata is
 840 * flushed on write-faults (non-cow), but not read-faults.
 841 */
 842static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
 843		struct vm_area_struct *vma)
 844{
 845	return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
 846		(iter->iomap.flags & IOMAP_F_DIRTY);
 847}
 848
 849/*
 850 * By this point grab_mapping_entry() has ensured that we have a locked entry
 851 * of the appropriate size so we don't have to worry about downgrading PMDs to
 852 * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 853 * already in the tree, we will skip the insertion and just dirty the PMD as
 854 * appropriate.
 855 */
 856static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
 857		const struct iomap_iter *iter, void *entry, pfn_t pfn,
 858		unsigned long flags)
 859{
 860	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 861	void *new_entry = dax_make_entry(pfn, flags);
 862	bool write = iter->flags & IOMAP_WRITE;
 863	bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
 864	bool shared = iter->iomap.flags & IOMAP_F_SHARED;
 865
 866	if (dirty)
 867		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 868
 869	if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
 870		unsigned long index = xas->xa_index;
 871		/* we are replacing a zero page with block mapping */
 872		if (dax_is_pmd_entry(entry))
 873			unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
 874					PG_PMD_NR, false);
 875		else /* pte entry */
 876			unmap_mapping_pages(mapping, index, 1, false);
 877	}
 878
 879	xas_reset(xas);
 880	xas_lock_irq(xas);
 881	if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 882		void *old;
 883
 884		dax_disassociate_entry(entry, mapping, false);
 885		dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
 886				shared);
 887		/*
 888		 * Only swap our new entry into the page cache if the current
 889		 * entry is a zero page or an empty entry.  If a normal PTE or
 890		 * PMD entry is already in the cache, we leave it alone.  This
 891		 * means that if we are trying to insert a PTE and the
 892		 * existing entry is a PMD, we will just leave the PMD in the
 893		 * tree and dirty it if necessary.
 894		 */
 895		old = dax_lock_entry(xas, new_entry);
 896		WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
 897					DAX_LOCKED));
 898		entry = new_entry;
 899	} else {
 900		xas_load(xas);	/* Walk the xa_state */
 901	}
 902
 903	if (dirty)
 904		xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
 905
 906	if (write && shared)
 907		xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
 908
 909	xas_unlock_irq(xas);
 910	return entry;
 911}
 912
 913static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
 914		struct address_space *mapping, void *entry)
 915{
 916	unsigned long pfn, index, count, end;
 917	long ret = 0;
 918	struct vm_area_struct *vma;
 919
 920	/*
 921	 * A page got tagged dirty in DAX mapping? Something is seriously
 922	 * wrong.
 923	 */
 924	if (WARN_ON(!xa_is_value(entry)))
 925		return -EIO;
 926
 927	if (unlikely(dax_is_locked(entry))) {
 928		void *old_entry = entry;
 929
 930		entry = get_unlocked_entry(xas, 0);
 931
 932		/* Entry got punched out / reallocated? */
 933		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 934			goto put_unlocked;
 935		/*
 936		 * Entry got reallocated elsewhere? No need to writeback.
 937		 * We have to compare pfns as we must not bail out due to
 938		 * difference in lockbit or entry type.
 939		 */
 940		if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
 941			goto put_unlocked;
 942		if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 943					dax_is_zero_entry(entry))) {
 944			ret = -EIO;
 945			goto put_unlocked;
 946		}
 947
 948		/* Another fsync thread may have already done this entry */
 949		if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
 950			goto put_unlocked;
 951	}
 952
 953	/* Lock the entry to serialize with page faults */
 954	dax_lock_entry(xas, entry);
 955
 956	/*
 957	 * We can clear the tag now but we have to be careful so that concurrent
 958	 * dax_writeback_one() calls for the same index cannot finish before we
 959	 * actually flush the caches. This is achieved as the calls will look
 960	 * at the entry only under the i_pages lock and once they do that
 961	 * they will see the entry locked and wait for it to unlock.
 962	 */
 963	xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
 964	xas_unlock_irq(xas);
 965
 966	/*
 967	 * If dax_writeback_mapping_range() was given a wbc->range_start
 968	 * in the middle of a PMD, the 'index' we use needs to be
 969	 * aligned to the start of the PMD.
 970	 * This allows us to flush for PMD_SIZE and not have to worry about
 971	 * partial PMD writebacks.
 972	 */
 973	pfn = dax_to_pfn(entry);
 974	count = 1UL << dax_entry_order(entry);
 975	index = xas->xa_index & ~(count - 1);
 976	end = index + count - 1;
 977
 978	/* Walk all mappings of a given index of a file and writeprotect them */
 979	i_mmap_lock_read(mapping);
 980	vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
 981		pfn_mkclean_range(pfn, count, index, vma);
 982		cond_resched();
 983	}
 984	i_mmap_unlock_read(mapping);
 985
 986	dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
 987	/*
 988	 * After we have flushed the cache, we can clear the dirty tag. There
 989	 * cannot be new dirty data in the pfn after the flush has completed as
 990	 * the pfn mappings are writeprotected and fault waits for mapping
 991	 * entry lock.
 992	 */
 993	xas_reset(xas);
 994	xas_lock_irq(xas);
 995	xas_store(xas, entry);
 996	xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
 997	dax_wake_entry(xas, entry, WAKE_NEXT);
 998
 999	trace_dax_writeback_one(mapping->host, index, count);
1000	return ret;
1001
1002 put_unlocked:
1003	put_unlocked_entry(xas, entry, WAKE_NEXT);
1004	return ret;
1005}
1006
1007/*
1008 * Flush the mapping to the persistent domain within the byte range of [start,
1009 * end]. This is required by data integrity operations to ensure file data is
1010 * on persistent storage prior to completion of the operation.
1011 */
1012int dax_writeback_mapping_range(struct address_space *mapping,
1013		struct dax_device *dax_dev, struct writeback_control *wbc)
1014{
1015	XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
1016	struct inode *inode = mapping->host;
1017	pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
1018	void *entry;
1019	int ret = 0;
1020	unsigned int scanned = 0;
1021
1022	if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1023		return -EIO;
1024
1025	if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
1026		return 0;
1027
1028	trace_dax_writeback_range(inode, xas.xa_index, end_index);
1029
1030	tag_pages_for_writeback(mapping, xas.xa_index, end_index);
1031
1032	xas_lock_irq(&xas);
1033	xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1034		ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1035		if (ret < 0) {
1036			mapping_set_error(mapping, ret);
1037			break;
1038		}
1039		if (++scanned % XA_CHECK_SCHED)
1040			continue;
1041
1042		xas_pause(&xas);
1043		xas_unlock_irq(&xas);
1044		cond_resched();
1045		xas_lock_irq(&xas);
1046	}
1047	xas_unlock_irq(&xas);
1048	trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1049	return ret;
1050}
1051EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1052
1053static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1054		size_t size, void **kaddr, pfn_t *pfnp)
1055{
1056	pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1057	int id, rc = 0;
1058	long length;
1059
1060	id = dax_read_lock();
1061	length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1062				   DAX_ACCESS, kaddr, pfnp);
1063	if (length < 0) {
1064		rc = length;
1065		goto out;
1066	}
1067	if (!pfnp)
1068		goto out_check_addr;
1069	rc = -EINVAL;
1070	if (PFN_PHYS(length) < size)
1071		goto out;
1072	if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1073		goto out;
1074	/* For larger pages we need devmap */
1075	if (length > 1 && !pfn_t_devmap(*pfnp))
1076		goto out;
1077	rc = 0;
1078
1079out_check_addr:
1080	if (!kaddr)
1081		goto out;
1082	if (!*kaddr)
1083		rc = -EFAULT;
1084out:
1085	dax_read_unlock(id);
1086	return rc;
1087}
1088
1089/**
1090 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1091 * by copying the data before and after the range to be written.
1092 * @pos:	address to do copy from.
1093 * @length:	size of copy operation.
1094 * @align_size:	aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1095 * @srcmap:	iomap srcmap
1096 * @daddr:	destination address to copy to.
1097 *
1098 * This can be called from two places. Either during DAX write fault (page
1099 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1100 * write operation, dax_iomap_iter() might call this to do the copy of either
1101 * start or end unaligned address. In the latter case the rest of the copy of
1102 * aligned ranges is taken care by dax_iomap_iter() itself.
1103 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1104 * area to make sure no old data remains.
1105 */
1106static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
1107		const struct iomap *srcmap, void *daddr)
1108{
1109	loff_t head_off = pos & (align_size - 1);
1110	size_t size = ALIGN(head_off + length, align_size);
1111	loff_t end = pos + length;
1112	loff_t pg_end = round_up(end, align_size);
1113	/* copy_all is usually in page fault case */
1114	bool copy_all = head_off == 0 && end == pg_end;
1115	/* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1116	bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1117			 srcmap->type == IOMAP_UNWRITTEN;
1118	void *saddr = 0;
1119	int ret = 0;
1120
1121	if (!zero_edge) {
1122		ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1123		if (ret)
1124			return ret;
1125	}
1126
1127	if (copy_all) {
1128		if (zero_edge)
1129			memset(daddr, 0, size);
1130		else
1131			ret = copy_mc_to_kernel(daddr, saddr, length);
1132		goto out;
1133	}
1134
1135	/* Copy the head part of the range */
1136	if (head_off) {
1137		if (zero_edge)
1138			memset(daddr, 0, head_off);
1139		else {
1140			ret = copy_mc_to_kernel(daddr, saddr, head_off);
1141			if (ret)
1142				return -EIO;
1143		}
1144	}
1145
1146	/* Copy the tail part of the range */
1147	if (end < pg_end) {
1148		loff_t tail_off = head_off + length;
1149		loff_t tail_len = pg_end - end;
1150
1151		if (zero_edge)
1152			memset(daddr + tail_off, 0, tail_len);
1153		else {
1154			ret = copy_mc_to_kernel(daddr + tail_off,
1155						saddr + tail_off, tail_len);
1156			if (ret)
1157				return -EIO;
1158		}
1159	}
1160out:
1161	if (zero_edge)
1162		dax_flush(srcmap->dax_dev, daddr, size);
1163	return ret ? -EIO : 0;
1164}
1165
1166/*
1167 * The user has performed a load from a hole in the file.  Allocating a new
1168 * page in the file would cause excessive storage usage for workloads with
1169 * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1170 * If this page is ever written to we will re-fault and change the mapping to
1171 * point to real DAX storage instead.
1172 */
1173static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1174		const struct iomap_iter *iter, void **entry)
1175{
1176	struct inode *inode = iter->inode;
1177	unsigned long vaddr = vmf->address;
1178	pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1179	vm_fault_t ret;
1180
1181	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
1182
1183	ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1184	trace_dax_load_hole(inode, vmf, ret);
1185	return ret;
1186}
1187
1188#ifdef CONFIG_FS_DAX_PMD
1189static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1190		const struct iomap_iter *iter, void **entry)
1191{
1192	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1193	unsigned long pmd_addr = vmf->address & PMD_MASK;
1194	struct vm_area_struct *vma = vmf->vma;
1195	struct inode *inode = mapping->host;
1196	pgtable_t pgtable = NULL;
1197	struct page *zero_page;
1198	spinlock_t *ptl;
1199	pmd_t pmd_entry;
1200	pfn_t pfn;
1201
1202	zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1203
1204	if (unlikely(!zero_page))
1205		goto fallback;
1206
1207	pfn = page_to_pfn_t(zero_page);
1208	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1209				  DAX_PMD | DAX_ZERO_PAGE);
1210
1211	if (arch_needs_pgtable_deposit()) {
1212		pgtable = pte_alloc_one(vma->vm_mm);
1213		if (!pgtable)
1214			return VM_FAULT_OOM;
1215	}
1216
1217	ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1218	if (!pmd_none(*(vmf->pmd))) {
1219		spin_unlock(ptl);
1220		goto fallback;
1221	}
1222
1223	if (pgtable) {
1224		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1225		mm_inc_nr_ptes(vma->vm_mm);
1226	}
1227	pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1228	pmd_entry = pmd_mkhuge(pmd_entry);
1229	set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1230	spin_unlock(ptl);
1231	trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1232	return VM_FAULT_NOPAGE;
1233
1234fallback:
1235	if (pgtable)
1236		pte_free(vma->vm_mm, pgtable);
1237	trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1238	return VM_FAULT_FALLBACK;
1239}
1240#else
1241static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1242		const struct iomap_iter *iter, void **entry)
1243{
1244	return VM_FAULT_FALLBACK;
1245}
1246#endif /* CONFIG_FS_DAX_PMD */
1247
1248static s64 dax_unshare_iter(struct iomap_iter *iter)
1249{
1250	struct iomap *iomap = &iter->iomap;
1251	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1252	loff_t pos = iter->pos;
1253	loff_t length = iomap_length(iter);
1254	int id = 0;
1255	s64 ret = 0;
1256	void *daddr = NULL, *saddr = NULL;
1257
1258	/* don't bother with blocks that are not shared to start with */
1259	if (!(iomap->flags & IOMAP_F_SHARED))
1260		return length;
1261	/* don't bother with holes or unwritten extents */
1262	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1263		return length;
1264
1265	id = dax_read_lock();
1266	ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1267	if (ret < 0)
1268		goto out_unlock;
1269
 
 
 
 
 
 
 
 
1270	ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1271	if (ret < 0)
1272		goto out_unlock;
1273
1274	if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1275		ret = length;
1276	else
1277		ret = -EIO;
1278
1279out_unlock:
1280	dax_read_unlock(id);
1281	return ret;
1282}
1283
1284int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1285		const struct iomap_ops *ops)
1286{
1287	struct iomap_iter iter = {
1288		.inode		= inode,
1289		.pos		= pos,
1290		.len		= len,
1291		.flags		= IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1292	};
1293	int ret;
1294
1295	while ((ret = iomap_iter(&iter, ops)) > 0)
1296		iter.processed = dax_unshare_iter(&iter);
1297	return ret;
1298}
1299EXPORT_SYMBOL_GPL(dax_file_unshare);
1300
1301static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1302{
1303	const struct iomap *iomap = &iter->iomap;
1304	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1305	unsigned offset = offset_in_page(pos);
1306	pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1307	void *kaddr;
1308	long ret;
1309
1310	ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1311				NULL);
1312	if (ret < 0)
1313		return ret;
 
1314	memset(kaddr + offset, 0, size);
1315	if (iomap->flags & IOMAP_F_SHARED)
1316		ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1317					    kaddr);
1318	else
1319		dax_flush(iomap->dax_dev, kaddr + offset, size);
1320	return ret;
1321}
1322
1323static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1324{
1325	const struct iomap *iomap = &iter->iomap;
1326	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1327	loff_t pos = iter->pos;
1328	u64 length = iomap_length(iter);
1329	s64 written = 0;
1330
1331	/* already zeroed?  we're done. */
1332	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1333		return length;
1334
1335	/*
1336	 * invalidate the pages whose sharing state is to be changed
1337	 * because of CoW.
1338	 */
1339	if (iomap->flags & IOMAP_F_SHARED)
1340		invalidate_inode_pages2_range(iter->inode->i_mapping,
1341					      pos >> PAGE_SHIFT,
1342					      (pos + length - 1) >> PAGE_SHIFT);
1343
1344	do {
1345		unsigned offset = offset_in_page(pos);
1346		unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1347		pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1348		long rc;
1349		int id;
1350
1351		id = dax_read_lock();
1352		if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1353			rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1354		else
1355			rc = dax_memzero(iter, pos, size);
1356		dax_read_unlock(id);
1357
1358		if (rc < 0)
1359			return rc;
1360		pos += size;
1361		length -= size;
1362		written += size;
1363	} while (length > 0);
1364
1365	if (did_zero)
1366		*did_zero = true;
1367	return written;
1368}
1369
1370int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1371		const struct iomap_ops *ops)
1372{
1373	struct iomap_iter iter = {
1374		.inode		= inode,
1375		.pos		= pos,
1376		.len		= len,
1377		.flags		= IOMAP_DAX | IOMAP_ZERO,
1378	};
1379	int ret;
1380
1381	while ((ret = iomap_iter(&iter, ops)) > 0)
1382		iter.processed = dax_zero_iter(&iter, did_zero);
1383	return ret;
1384}
1385EXPORT_SYMBOL_GPL(dax_zero_range);
1386
1387int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1388		const struct iomap_ops *ops)
1389{
1390	unsigned int blocksize = i_blocksize(inode);
1391	unsigned int off = pos & (blocksize - 1);
1392
1393	/* Block boundary? Nothing to do */
1394	if (!off)
1395		return 0;
1396	return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1397}
1398EXPORT_SYMBOL_GPL(dax_truncate_page);
1399
1400static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1401		struct iov_iter *iter)
1402{
1403	const struct iomap *iomap = &iomi->iomap;
1404	const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1405	loff_t length = iomap_length(iomi);
1406	loff_t pos = iomi->pos;
1407	struct dax_device *dax_dev = iomap->dax_dev;
1408	loff_t end = pos + length, done = 0;
1409	bool write = iov_iter_rw(iter) == WRITE;
1410	bool cow = write && iomap->flags & IOMAP_F_SHARED;
1411	ssize_t ret = 0;
1412	size_t xfer;
1413	int id;
1414
1415	if (!write) {
1416		end = min(end, i_size_read(iomi->inode));
1417		if (pos >= end)
1418			return 0;
1419
1420		if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1421			return iov_iter_zero(min(length, end - pos), iter);
1422	}
1423
1424	/*
1425	 * In DAX mode, enforce either pure overwrites of written extents, or
1426	 * writes to unwritten extents as part of a copy-on-write operation.
1427	 */
1428	if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1429			!(iomap->flags & IOMAP_F_SHARED)))
1430		return -EIO;
1431
1432	/*
1433	 * Write can allocate block for an area which has a hole page mapped
1434	 * into page tables. We have to tear down these mappings so that data
1435	 * written by write(2) is visible in mmap.
1436	 */
1437	if (iomap->flags & IOMAP_F_NEW || cow) {
 
 
 
 
 
 
 
 
 
 
1438		invalidate_inode_pages2_range(iomi->inode->i_mapping,
1439					      pos >> PAGE_SHIFT,
1440					      (end - 1) >> PAGE_SHIFT);
1441	}
1442
1443	id = dax_read_lock();
1444	while (pos < end) {
1445		unsigned offset = pos & (PAGE_SIZE - 1);
1446		const size_t size = ALIGN(length + offset, PAGE_SIZE);
1447		pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1448		ssize_t map_len;
1449		bool recovery = false;
1450		void *kaddr;
1451
1452		if (fatal_signal_pending(current)) {
1453			ret = -EINTR;
1454			break;
1455		}
1456
1457		map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1458				DAX_ACCESS, &kaddr, NULL);
1459		if (map_len == -EIO && iov_iter_rw(iter) == WRITE) {
1460			map_len = dax_direct_access(dax_dev, pgoff,
1461					PHYS_PFN(size), DAX_RECOVERY_WRITE,
1462					&kaddr, NULL);
1463			if (map_len > 0)
1464				recovery = true;
1465		}
1466		if (map_len < 0) {
1467			ret = map_len;
1468			break;
1469		}
1470
1471		if (cow) {
1472			ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1473						    srcmap, kaddr);
1474			if (ret)
1475				break;
1476		}
1477
1478		map_len = PFN_PHYS(map_len);
1479		kaddr += offset;
1480		map_len -= offset;
1481		if (map_len > end - pos)
1482			map_len = end - pos;
1483
1484		if (recovery)
1485			xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1486					map_len, iter);
1487		else if (write)
1488			xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1489					map_len, iter);
1490		else
1491			xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1492					map_len, iter);
1493
1494		pos += xfer;
1495		length -= xfer;
1496		done += xfer;
1497
1498		if (xfer == 0)
1499			ret = -EFAULT;
1500		if (xfer < map_len)
1501			break;
1502	}
1503	dax_read_unlock(id);
1504
1505	return done ? done : ret;
1506}
1507
1508/**
1509 * dax_iomap_rw - Perform I/O to a DAX file
1510 * @iocb:	The control block for this I/O
1511 * @iter:	The addresses to do I/O from or to
1512 * @ops:	iomap ops passed from the file system
1513 *
1514 * This function performs read and write operations to directly mapped
1515 * persistent memory.  The callers needs to take care of read/write exclusion
1516 * and evicting any page cache pages in the region under I/O.
1517 */
1518ssize_t
1519dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1520		const struct iomap_ops *ops)
1521{
1522	struct iomap_iter iomi = {
1523		.inode		= iocb->ki_filp->f_mapping->host,
1524		.pos		= iocb->ki_pos,
1525		.len		= iov_iter_count(iter),
1526		.flags		= IOMAP_DAX,
1527	};
1528	loff_t done = 0;
1529	int ret;
1530
1531	if (!iomi.len)
1532		return 0;
1533
1534	if (iov_iter_rw(iter) == WRITE) {
1535		lockdep_assert_held_write(&iomi.inode->i_rwsem);
1536		iomi.flags |= IOMAP_WRITE;
1537	} else {
1538		lockdep_assert_held(&iomi.inode->i_rwsem);
1539	}
1540
1541	if (iocb->ki_flags & IOCB_NOWAIT)
1542		iomi.flags |= IOMAP_NOWAIT;
1543
1544	while ((ret = iomap_iter(&iomi, ops)) > 0)
1545		iomi.processed = dax_iomap_iter(&iomi, iter);
1546
1547	done = iomi.pos - iocb->ki_pos;
1548	iocb->ki_pos = iomi.pos;
1549	return done ? done : ret;
1550}
1551EXPORT_SYMBOL_GPL(dax_iomap_rw);
1552
1553static vm_fault_t dax_fault_return(int error)
1554{
1555	if (error == 0)
1556		return VM_FAULT_NOPAGE;
1557	return vmf_error(error);
1558}
1559
1560/*
1561 * When handling a synchronous page fault and the inode need a fsync, we can
1562 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1563 * insertion for now and return the pfn so that caller can insert it after the
1564 * fsync is done.
1565 */
1566static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1567{
1568	if (WARN_ON_ONCE(!pfnp))
1569		return VM_FAULT_SIGBUS;
1570	*pfnp = pfn;
1571	return VM_FAULT_NEEDDSYNC;
1572}
1573
1574static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1575		const struct iomap_iter *iter)
1576{
1577	vm_fault_t ret;
1578	int error = 0;
1579
1580	switch (iter->iomap.type) {
1581	case IOMAP_HOLE:
1582	case IOMAP_UNWRITTEN:
1583		clear_user_highpage(vmf->cow_page, vmf->address);
1584		break;
1585	case IOMAP_MAPPED:
1586		error = copy_cow_page_dax(vmf, iter);
1587		break;
1588	default:
1589		WARN_ON_ONCE(1);
1590		error = -EIO;
1591		break;
1592	}
1593
1594	if (error)
1595		return dax_fault_return(error);
1596
1597	__SetPageUptodate(vmf->cow_page);
1598	ret = finish_fault(vmf);
1599	if (!ret)
1600		return VM_FAULT_DONE_COW;
1601	return ret;
1602}
1603
1604/**
1605 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1606 * @vmf:	vm fault instance
1607 * @iter:	iomap iter
1608 * @pfnp:	pfn to be returned
1609 * @xas:	the dax mapping tree of a file
1610 * @entry:	an unlocked dax entry to be inserted
1611 * @pmd:	distinguish whether it is a pmd fault
1612 */
1613static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1614		const struct iomap_iter *iter, pfn_t *pfnp,
1615		struct xa_state *xas, void **entry, bool pmd)
1616{
1617	const struct iomap *iomap = &iter->iomap;
1618	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1619	size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1620	loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1621	bool write = iter->flags & IOMAP_WRITE;
1622	unsigned long entry_flags = pmd ? DAX_PMD : 0;
1623	int err = 0;
1624	pfn_t pfn;
1625	void *kaddr;
1626
1627	if (!pmd && vmf->cow_page)
1628		return dax_fault_cow_page(vmf, iter);
1629
1630	/* if we are reading UNWRITTEN and HOLE, return a hole. */
1631	if (!write &&
1632	    (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1633		if (!pmd)
1634			return dax_load_hole(xas, vmf, iter, entry);
1635		return dax_pmd_load_hole(xas, vmf, iter, entry);
1636	}
1637
1638	if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1639		WARN_ON_ONCE(1);
1640		return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1641	}
1642
1643	err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1644	if (err)
1645		return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1646
1647	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1648
1649	if (write && iomap->flags & IOMAP_F_SHARED) {
1650		err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1651		if (err)
1652			return dax_fault_return(err);
1653	}
1654
1655	if (dax_fault_is_synchronous(iter, vmf->vma))
1656		return dax_fault_synchronous_pfnp(pfnp, pfn);
1657
1658	/* insert PMD pfn */
1659	if (pmd)
1660		return vmf_insert_pfn_pmd(vmf, pfn, write);
1661
1662	/* insert PTE pfn */
1663	if (write)
1664		return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1665	return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1666}
1667
1668static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1669			       int *iomap_errp, const struct iomap_ops *ops)
1670{
1671	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1672	XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1673	struct iomap_iter iter = {
1674		.inode		= mapping->host,
1675		.pos		= (loff_t)vmf->pgoff << PAGE_SHIFT,
1676		.len		= PAGE_SIZE,
1677		.flags		= IOMAP_DAX | IOMAP_FAULT,
1678	};
1679	vm_fault_t ret = 0;
1680	void *entry;
1681	int error;
1682
1683	trace_dax_pte_fault(iter.inode, vmf, ret);
1684	/*
1685	 * Check whether offset isn't beyond end of file now. Caller is supposed
1686	 * to hold locks serializing us with truncate / punch hole so this is
1687	 * a reliable test.
1688	 */
1689	if (iter.pos >= i_size_read(iter.inode)) {
1690		ret = VM_FAULT_SIGBUS;
1691		goto out;
1692	}
1693
1694	if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1695		iter.flags |= IOMAP_WRITE;
1696
1697	entry = grab_mapping_entry(&xas, mapping, 0);
1698	if (xa_is_internal(entry)) {
1699		ret = xa_to_internal(entry);
1700		goto out;
1701	}
1702
1703	/*
1704	 * It is possible, particularly with mixed reads & writes to private
1705	 * mappings, that we have raced with a PMD fault that overlaps with
1706	 * the PTE we need to set up.  If so just return and the fault will be
1707	 * retried.
1708	 */
1709	if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1710		ret = VM_FAULT_NOPAGE;
1711		goto unlock_entry;
1712	}
1713
1714	while ((error = iomap_iter(&iter, ops)) > 0) {
1715		if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1716			iter.processed = -EIO;	/* fs corruption? */
1717			continue;
1718		}
1719
1720		ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1721		if (ret != VM_FAULT_SIGBUS &&
1722		    (iter.iomap.flags & IOMAP_F_NEW)) {
1723			count_vm_event(PGMAJFAULT);
1724			count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1725			ret |= VM_FAULT_MAJOR;
1726		}
1727
1728		if (!(ret & VM_FAULT_ERROR))
1729			iter.processed = PAGE_SIZE;
1730	}
1731
1732	if (iomap_errp)
1733		*iomap_errp = error;
1734	if (!ret && error)
1735		ret = dax_fault_return(error);
1736
1737unlock_entry:
1738	dax_unlock_entry(&xas, entry);
1739out:
1740	trace_dax_pte_fault_done(iter.inode, vmf, ret);
1741	return ret;
1742}
1743
1744#ifdef CONFIG_FS_DAX_PMD
1745static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1746		pgoff_t max_pgoff)
1747{
1748	unsigned long pmd_addr = vmf->address & PMD_MASK;
1749	bool write = vmf->flags & FAULT_FLAG_WRITE;
1750
1751	/*
1752	 * Make sure that the faulting address's PMD offset (color) matches
1753	 * the PMD offset from the start of the file.  This is necessary so
1754	 * that a PMD range in the page table overlaps exactly with a PMD
1755	 * range in the page cache.
1756	 */
1757	if ((vmf->pgoff & PG_PMD_COLOUR) !=
1758	    ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1759		return true;
1760
1761	/* Fall back to PTEs if we're going to COW */
1762	if (write && !(vmf->vma->vm_flags & VM_SHARED))
1763		return true;
1764
1765	/* If the PMD would extend outside the VMA */
1766	if (pmd_addr < vmf->vma->vm_start)
1767		return true;
1768	if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1769		return true;
1770
1771	/* If the PMD would extend beyond the file size */
1772	if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1773		return true;
1774
1775	return false;
1776}
1777
1778static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1779			       const struct iomap_ops *ops)
1780{
1781	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1782	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1783	struct iomap_iter iter = {
1784		.inode		= mapping->host,
1785		.len		= PMD_SIZE,
1786		.flags		= IOMAP_DAX | IOMAP_FAULT,
1787	};
1788	vm_fault_t ret = VM_FAULT_FALLBACK;
1789	pgoff_t max_pgoff;
1790	void *entry;
1791	int error;
1792
1793	if (vmf->flags & FAULT_FLAG_WRITE)
1794		iter.flags |= IOMAP_WRITE;
1795
1796	/*
1797	 * Check whether offset isn't beyond end of file now. Caller is
1798	 * supposed to hold locks serializing us with truncate / punch hole so
1799	 * this is a reliable test.
1800	 */
1801	max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1802
1803	trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1804
1805	if (xas.xa_index >= max_pgoff) {
1806		ret = VM_FAULT_SIGBUS;
1807		goto out;
1808	}
1809
1810	if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1811		goto fallback;
1812
1813	/*
1814	 * grab_mapping_entry() will make sure we get an empty PMD entry,
1815	 * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1816	 * entry is already in the array, for instance), it will return
1817	 * VM_FAULT_FALLBACK.
1818	 */
1819	entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1820	if (xa_is_internal(entry)) {
1821		ret = xa_to_internal(entry);
1822		goto fallback;
1823	}
1824
1825	/*
1826	 * It is possible, particularly with mixed reads & writes to private
1827	 * mappings, that we have raced with a PTE fault that overlaps with
1828	 * the PMD we need to set up.  If so just return and the fault will be
1829	 * retried.
1830	 */
1831	if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1832			!pmd_devmap(*vmf->pmd)) {
1833		ret = 0;
1834		goto unlock_entry;
1835	}
1836
1837	iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1838	while ((error = iomap_iter(&iter, ops)) > 0) {
1839		if (iomap_length(&iter) < PMD_SIZE)
1840			continue; /* actually breaks out of the loop */
1841
1842		ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1843		if (ret != VM_FAULT_FALLBACK)
1844			iter.processed = PMD_SIZE;
1845	}
1846
1847unlock_entry:
1848	dax_unlock_entry(&xas, entry);
1849fallback:
1850	if (ret == VM_FAULT_FALLBACK) {
1851		split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1852		count_vm_event(THP_FAULT_FALLBACK);
1853	}
1854out:
1855	trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1856	return ret;
1857}
1858#else
1859static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1860			       const struct iomap_ops *ops)
1861{
1862	return VM_FAULT_FALLBACK;
1863}
1864#endif /* CONFIG_FS_DAX_PMD */
1865
1866/**
1867 * dax_iomap_fault - handle a page fault on a DAX file
1868 * @vmf: The description of the fault
1869 * @pe_size: Size of the page to fault in
1870 * @pfnp: PFN to insert for synchronous faults if fsync is required
1871 * @iomap_errp: Storage for detailed error code in case of error
1872 * @ops: Iomap ops passed from the file system
1873 *
1874 * When a page fault occurs, filesystems may call this helper in
1875 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1876 * has done all the necessary locking for page fault to proceed
1877 * successfully.
1878 */
1879vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
1880		    pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1881{
1882	switch (pe_size) {
1883	case PE_SIZE_PTE:
1884		return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1885	case PE_SIZE_PMD:
1886		return dax_iomap_pmd_fault(vmf, pfnp, ops);
1887	default:
1888		return VM_FAULT_FALLBACK;
1889	}
1890}
1891EXPORT_SYMBOL_GPL(dax_iomap_fault);
1892
1893/*
1894 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1895 * @vmf: The description of the fault
1896 * @pfn: PFN to insert
1897 * @order: Order of entry to insert.
1898 *
1899 * This function inserts a writeable PTE or PMD entry into the page tables
1900 * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1901 */
1902static vm_fault_t
1903dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1904{
1905	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1906	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1907	void *entry;
1908	vm_fault_t ret;
1909
1910	xas_lock_irq(&xas);
1911	entry = get_unlocked_entry(&xas, order);
1912	/* Did we race with someone splitting entry or so? */
1913	if (!entry || dax_is_conflict(entry) ||
1914	    (order == 0 && !dax_is_pte_entry(entry))) {
1915		put_unlocked_entry(&xas, entry, WAKE_NEXT);
1916		xas_unlock_irq(&xas);
1917		trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1918						      VM_FAULT_NOPAGE);
1919		return VM_FAULT_NOPAGE;
1920	}
1921	xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1922	dax_lock_entry(&xas, entry);
1923	xas_unlock_irq(&xas);
1924	if (order == 0)
1925		ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1926#ifdef CONFIG_FS_DAX_PMD
1927	else if (order == PMD_ORDER)
1928		ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1929#endif
1930	else
1931		ret = VM_FAULT_FALLBACK;
1932	dax_unlock_entry(&xas, entry);
1933	trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1934	return ret;
1935}
1936
1937/**
1938 * dax_finish_sync_fault - finish synchronous page fault
1939 * @vmf: The description of the fault
1940 * @pe_size: Size of entry to be inserted
1941 * @pfn: PFN to insert
1942 *
1943 * This function ensures that the file range touched by the page fault is
1944 * stored persistently on the media and handles inserting of appropriate page
1945 * table entry.
1946 */
1947vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf,
1948		enum page_entry_size pe_size, pfn_t pfn)
1949{
1950	int err;
1951	loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
1952	unsigned int order = pe_order(pe_size);
1953	size_t len = PAGE_SIZE << order;
1954
1955	err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1956	if (err)
1957		return VM_FAULT_SIGBUS;
1958	return dax_insert_pfn_mkwrite(vmf, pfn, order);
1959}
1960EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1961
1962static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
1963		struct iomap_iter *it_dest, u64 len, bool *same)
1964{
1965	const struct iomap *smap = &it_src->iomap;
1966	const struct iomap *dmap = &it_dest->iomap;
1967	loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
1968	void *saddr, *daddr;
1969	int id, ret;
1970
1971	len = min(len, min(smap->length, dmap->length));
1972
1973	if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
1974		*same = true;
1975		return len;
1976	}
1977
1978	if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
1979		*same = false;
1980		return 0;
1981	}
1982
1983	id = dax_read_lock();
1984	ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
1985				      &saddr, NULL);
1986	if (ret < 0)
1987		goto out_unlock;
1988
1989	ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
1990				      &daddr, NULL);
1991	if (ret < 0)
1992		goto out_unlock;
1993
1994	*same = !memcmp(saddr, daddr, len);
1995	if (!*same)
1996		len = 0;
1997	dax_read_unlock(id);
1998	return len;
1999
2000out_unlock:
2001	dax_read_unlock(id);
2002	return -EIO;
2003}
2004
2005int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2006		struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2007		const struct iomap_ops *ops)
2008{
2009	struct iomap_iter src_iter = {
2010		.inode		= src,
2011		.pos		= srcoff,
2012		.len		= len,
2013		.flags		= IOMAP_DAX,
2014	};
2015	struct iomap_iter dst_iter = {
2016		.inode		= dst,
2017		.pos		= dstoff,
2018		.len		= len,
2019		.flags		= IOMAP_DAX,
2020	};
2021	int ret, compared = 0;
2022
2023	while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2024	       (ret = iomap_iter(&dst_iter, ops)) > 0) {
2025		compared = dax_range_compare_iter(&src_iter, &dst_iter, len,
2026						  same);
2027		if (compared < 0)
2028			return ret;
2029		src_iter.processed = dst_iter.processed = compared;
2030	}
2031	return ret;
2032}
2033
2034int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2035			      struct file *file_out, loff_t pos_out,
2036			      loff_t *len, unsigned int remap_flags,
2037			      const struct iomap_ops *ops)
2038{
2039	return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2040					       pos_out, len, remap_flags, ops);
2041}
2042EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * fs/dax.c - Direct Access filesystem code
   4 * Copyright (c) 2013-2014 Intel Corporation
   5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   7 */
   8
   9#include <linux/atomic.h>
  10#include <linux/blkdev.h>
  11#include <linux/buffer_head.h>
  12#include <linux/dax.h>
  13#include <linux/fs.h>
  14#include <linux/highmem.h>
  15#include <linux/memcontrol.h>
  16#include <linux/mm.h>
  17#include <linux/mutex.h>
  18#include <linux/pagevec.h>
  19#include <linux/sched.h>
  20#include <linux/sched/signal.h>
  21#include <linux/uio.h>
  22#include <linux/vmstat.h>
  23#include <linux/pfn_t.h>
  24#include <linux/sizes.h>
  25#include <linux/mmu_notifier.h>
  26#include <linux/iomap.h>
  27#include <linux/rmap.h>
  28#include <asm/pgalloc.h>
  29
  30#define CREATE_TRACE_POINTS
  31#include <trace/events/fs_dax.h>
  32
 
 
 
 
 
 
 
 
 
 
 
  33/* We choose 4096 entries - same as per-zone page wait tables */
  34#define DAX_WAIT_TABLE_BITS 12
  35#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
  36
  37/* The 'colour' (ie low bits) within a PMD of a page offset.  */
  38#define PG_PMD_COLOUR	((PMD_SIZE >> PAGE_SHIFT) - 1)
  39#define PG_PMD_NR	(PMD_SIZE >> PAGE_SHIFT)
  40
 
 
 
  41static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
  42
  43static int __init init_dax_wait_table(void)
  44{
  45	int i;
  46
  47	for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
  48		init_waitqueue_head(wait_table + i);
  49	return 0;
  50}
  51fs_initcall(init_dax_wait_table);
  52
  53/*
  54 * DAX pagecache entries use XArray value entries so they can't be mistaken
  55 * for pages.  We use one bit for locking, one bit for the entry size (PMD)
  56 * and two more to tell us if the entry is a zero page or an empty entry that
  57 * is just used for locking.  In total four special bits.
  58 *
  59 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
  60 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
  61 * block allocation.
  62 */
  63#define DAX_SHIFT	(4)
  64#define DAX_LOCKED	(1UL << 0)
  65#define DAX_PMD		(1UL << 1)
  66#define DAX_ZERO_PAGE	(1UL << 2)
  67#define DAX_EMPTY	(1UL << 3)
  68
  69static unsigned long dax_to_pfn(void *entry)
  70{
  71	return xa_to_value(entry) >> DAX_SHIFT;
  72}
  73
  74static void *dax_make_entry(pfn_t pfn, unsigned long flags)
  75{
  76	return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT));
  77}
  78
  79static bool dax_is_locked(void *entry)
  80{
  81	return xa_to_value(entry) & DAX_LOCKED;
  82}
  83
  84static unsigned int dax_entry_order(void *entry)
  85{
  86	if (xa_to_value(entry) & DAX_PMD)
  87		return PMD_ORDER;
  88	return 0;
  89}
  90
  91static unsigned long dax_is_pmd_entry(void *entry)
  92{
  93	return xa_to_value(entry) & DAX_PMD;
  94}
  95
  96static bool dax_is_pte_entry(void *entry)
  97{
  98	return !(xa_to_value(entry) & DAX_PMD);
  99}
 100
 101static int dax_is_zero_entry(void *entry)
 102{
 103	return xa_to_value(entry) & DAX_ZERO_PAGE;
 104}
 105
 106static int dax_is_empty_entry(void *entry)
 107{
 108	return xa_to_value(entry) & DAX_EMPTY;
 109}
 110
 111/*
 112 * true if the entry that was found is of a smaller order than the entry
 113 * we were looking for
 114 */
 115static bool dax_is_conflict(void *entry)
 116{
 117	return entry == XA_RETRY_ENTRY;
 118}
 119
 120/*
 121 * DAX page cache entry locking
 122 */
 123struct exceptional_entry_key {
 124	struct xarray *xa;
 125	pgoff_t entry_start;
 126};
 127
 128struct wait_exceptional_entry_queue {
 129	wait_queue_entry_t wait;
 130	struct exceptional_entry_key key;
 131};
 132
 133/**
 134 * enum dax_wake_mode: waitqueue wakeup behaviour
 135 * @WAKE_ALL: wake all waiters in the waitqueue
 136 * @WAKE_NEXT: wake only the first waiter in the waitqueue
 137 */
 138enum dax_wake_mode {
 139	WAKE_ALL,
 140	WAKE_NEXT,
 141};
 142
 143static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas,
 144		void *entry, struct exceptional_entry_key *key)
 145{
 146	unsigned long hash;
 147	unsigned long index = xas->xa_index;
 148
 149	/*
 150	 * If 'entry' is a PMD, align the 'index' that we use for the wait
 151	 * queue to the start of that PMD.  This ensures that all offsets in
 152	 * the range covered by the PMD map to the same bit lock.
 153	 */
 154	if (dax_is_pmd_entry(entry))
 155		index &= ~PG_PMD_COLOUR;
 156	key->xa = xas->xa;
 157	key->entry_start = index;
 158
 159	hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS);
 160	return wait_table + hash;
 161}
 162
 163static int wake_exceptional_entry_func(wait_queue_entry_t *wait,
 164		unsigned int mode, int sync, void *keyp)
 165{
 166	struct exceptional_entry_key *key = keyp;
 167	struct wait_exceptional_entry_queue *ewait =
 168		container_of(wait, struct wait_exceptional_entry_queue, wait);
 169
 170	if (key->xa != ewait->key.xa ||
 171	    key->entry_start != ewait->key.entry_start)
 172		return 0;
 173	return autoremove_wake_function(wait, mode, sync, NULL);
 174}
 175
 176/*
 177 * @entry may no longer be the entry at the index in the mapping.
 178 * The important information it's conveying is whether the entry at
 179 * this index used to be a PMD entry.
 180 */
 181static void dax_wake_entry(struct xa_state *xas, void *entry,
 182			   enum dax_wake_mode mode)
 183{
 184	struct exceptional_entry_key key;
 185	wait_queue_head_t *wq;
 186
 187	wq = dax_entry_waitqueue(xas, entry, &key);
 188
 189	/*
 190	 * Checking for locked entry and prepare_to_wait_exclusive() happens
 191	 * under the i_pages lock, ditto for entry handling in our callers.
 192	 * So at this point all tasks that could have seen our entry locked
 193	 * must be in the waitqueue and the following check will see them.
 194	 */
 195	if (waitqueue_active(wq))
 196		__wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key);
 197}
 198
 199/*
 200 * Look up entry in page cache, wait for it to become unlocked if it
 201 * is a DAX entry and return it.  The caller must subsequently call
 202 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry()
 203 * if it did.  The entry returned may have a larger order than @order.
 204 * If @order is larger than the order of the entry found in i_pages, this
 205 * function returns a dax_is_conflict entry.
 206 *
 207 * Must be called with the i_pages lock held.
 208 */
 209static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
 210{
 211	void *entry;
 212	struct wait_exceptional_entry_queue ewait;
 213	wait_queue_head_t *wq;
 214
 215	init_wait(&ewait.wait);
 216	ewait.wait.func = wake_exceptional_entry_func;
 217
 218	for (;;) {
 219		entry = xas_find_conflict(xas);
 220		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 221			return entry;
 222		if (dax_entry_order(entry) < order)
 223			return XA_RETRY_ENTRY;
 224		if (!dax_is_locked(entry))
 225			return entry;
 226
 227		wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 228		prepare_to_wait_exclusive(wq, &ewait.wait,
 229					  TASK_UNINTERRUPTIBLE);
 230		xas_unlock_irq(xas);
 231		xas_reset(xas);
 232		schedule();
 233		finish_wait(wq, &ewait.wait);
 234		xas_lock_irq(xas);
 235	}
 236}
 237
 238/*
 239 * The only thing keeping the address space around is the i_pages lock
 240 * (it's cycled in clear_inode() after removing the entries from i_pages)
 241 * After we call xas_unlock_irq(), we cannot touch xas->xa.
 242 */
 243static void wait_entry_unlocked(struct xa_state *xas, void *entry)
 244{
 245	struct wait_exceptional_entry_queue ewait;
 246	wait_queue_head_t *wq;
 247
 248	init_wait(&ewait.wait);
 249	ewait.wait.func = wake_exceptional_entry_func;
 250
 251	wq = dax_entry_waitqueue(xas, entry, &ewait.key);
 252	/*
 253	 * Unlike get_unlocked_entry() there is no guarantee that this
 254	 * path ever successfully retrieves an unlocked entry before an
 255	 * inode dies. Perform a non-exclusive wait in case this path
 256	 * never successfully performs its own wake up.
 257	 */
 258	prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE);
 259	xas_unlock_irq(xas);
 260	schedule();
 261	finish_wait(wq, &ewait.wait);
 262}
 263
 264static void put_unlocked_entry(struct xa_state *xas, void *entry,
 265			       enum dax_wake_mode mode)
 266{
 267	if (entry && !dax_is_conflict(entry))
 268		dax_wake_entry(xas, entry, mode);
 269}
 270
 271/*
 272 * We used the xa_state to get the entry, but then we locked the entry and
 273 * dropped the xa_lock, so we know the xa_state is stale and must be reset
 274 * before use.
 275 */
 276static void dax_unlock_entry(struct xa_state *xas, void *entry)
 277{
 278	void *old;
 279
 280	BUG_ON(dax_is_locked(entry));
 281	xas_reset(xas);
 282	xas_lock_irq(xas);
 283	old = xas_store(xas, entry);
 284	xas_unlock_irq(xas);
 285	BUG_ON(!dax_is_locked(old));
 286	dax_wake_entry(xas, entry, WAKE_NEXT);
 287}
 288
 289/*
 290 * Return: The entry stored at this location before it was locked.
 291 */
 292static void *dax_lock_entry(struct xa_state *xas, void *entry)
 293{
 294	unsigned long v = xa_to_value(entry);
 295	return xas_store(xas, xa_mk_value(v | DAX_LOCKED));
 296}
 297
 298static unsigned long dax_entry_size(void *entry)
 299{
 300	if (dax_is_zero_entry(entry))
 301		return 0;
 302	else if (dax_is_empty_entry(entry))
 303		return 0;
 304	else if (dax_is_pmd_entry(entry))
 305		return PMD_SIZE;
 306	else
 307		return PAGE_SIZE;
 308}
 309
 310static unsigned long dax_end_pfn(void *entry)
 311{
 312	return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE;
 313}
 314
 315/*
 316 * Iterate through all mapped pfns represented by an entry, i.e. skip
 317 * 'empty' and 'zero' entries.
 318 */
 319#define for_each_mapped_pfn(entry, pfn) \
 320	for (pfn = dax_to_pfn(entry); \
 321			pfn < dax_end_pfn(entry); pfn++)
 322
 323static inline bool dax_page_is_shared(struct page *page)
 324{
 325	return page->mapping == PAGE_MAPPING_DAX_SHARED;
 326}
 327
 328/*
 329 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the
 330 * refcount.
 331 */
 332static inline void dax_page_share_get(struct page *page)
 333{
 334	if (page->mapping != PAGE_MAPPING_DAX_SHARED) {
 335		/*
 336		 * Reset the index if the page was already mapped
 337		 * regularly before.
 338		 */
 339		if (page->mapping)
 340			page->share = 1;
 341		page->mapping = PAGE_MAPPING_DAX_SHARED;
 342	}
 343	page->share++;
 344}
 345
 346static inline unsigned long dax_page_share_put(struct page *page)
 347{
 348	return --page->share;
 349}
 350
 351/*
 352 * When it is called in dax_insert_entry(), the shared flag will indicate that
 353 * whether this entry is shared by multiple files.  If so, set the page->mapping
 354 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount.
 355 */
 356static void dax_associate_entry(void *entry, struct address_space *mapping,
 357		struct vm_area_struct *vma, unsigned long address, bool shared)
 358{
 359	unsigned long size = dax_entry_size(entry), pfn, index;
 360	int i = 0;
 361
 362	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 363		return;
 364
 365	index = linear_page_index(vma, address & ~(size - 1));
 366	for_each_mapped_pfn(entry, pfn) {
 367		struct page *page = pfn_to_page(pfn);
 368
 369		if (shared) {
 370			dax_page_share_get(page);
 371		} else {
 372			WARN_ON_ONCE(page->mapping);
 373			page->mapping = mapping;
 374			page->index = index + i++;
 375		}
 376	}
 377}
 378
 379static void dax_disassociate_entry(void *entry, struct address_space *mapping,
 380		bool trunc)
 381{
 382	unsigned long pfn;
 383
 384	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 385		return;
 386
 387	for_each_mapped_pfn(entry, pfn) {
 388		struct page *page = pfn_to_page(pfn);
 389
 390		WARN_ON_ONCE(trunc && page_ref_count(page) > 1);
 391		if (dax_page_is_shared(page)) {
 392			/* keep the shared flag if this page is still shared */
 393			if (dax_page_share_put(page) > 0)
 394				continue;
 395		} else
 396			WARN_ON_ONCE(page->mapping && page->mapping != mapping);
 397		page->mapping = NULL;
 398		page->index = 0;
 399	}
 400}
 401
 402static struct page *dax_busy_page(void *entry)
 403{
 404	unsigned long pfn;
 405
 406	for_each_mapped_pfn(entry, pfn) {
 407		struct page *page = pfn_to_page(pfn);
 408
 409		if (page_ref_count(page) > 1)
 410			return page;
 411	}
 412	return NULL;
 413}
 414
 415/**
 416 * dax_lock_folio - Lock the DAX entry corresponding to a folio
 417 * @folio: The folio whose entry we want to lock
 418 *
 419 * Context: Process context.
 420 * Return: A cookie to pass to dax_unlock_folio() or 0 if the entry could
 421 * not be locked.
 422 */
 423dax_entry_t dax_lock_folio(struct folio *folio)
 424{
 425	XA_STATE(xas, NULL, 0);
 426	void *entry;
 427
 428	/* Ensure folio->mapping isn't freed while we look at it */
 429	rcu_read_lock();
 430	for (;;) {
 431		struct address_space *mapping = READ_ONCE(folio->mapping);
 432
 433		entry = NULL;
 434		if (!mapping || !dax_mapping(mapping))
 435			break;
 436
 437		/*
 438		 * In the device-dax case there's no need to lock, a
 439		 * struct dev_pagemap pin is sufficient to keep the
 440		 * inode alive, and we assume we have dev_pagemap pin
 441		 * otherwise we would not have a valid pfn_to_page()
 442		 * translation.
 443		 */
 444		entry = (void *)~0UL;
 445		if (S_ISCHR(mapping->host->i_mode))
 446			break;
 447
 448		xas.xa = &mapping->i_pages;
 449		xas_lock_irq(&xas);
 450		if (mapping != folio->mapping) {
 451			xas_unlock_irq(&xas);
 452			continue;
 453		}
 454		xas_set(&xas, folio->index);
 455		entry = xas_load(&xas);
 456		if (dax_is_locked(entry)) {
 457			rcu_read_unlock();
 458			wait_entry_unlocked(&xas, entry);
 459			rcu_read_lock();
 460			continue;
 461		}
 462		dax_lock_entry(&xas, entry);
 463		xas_unlock_irq(&xas);
 464		break;
 465	}
 466	rcu_read_unlock();
 467	return (dax_entry_t)entry;
 468}
 469
 470void dax_unlock_folio(struct folio *folio, dax_entry_t cookie)
 471{
 472	struct address_space *mapping = folio->mapping;
 473	XA_STATE(xas, &mapping->i_pages, folio->index);
 474
 475	if (S_ISCHR(mapping->host->i_mode))
 476		return;
 477
 478	dax_unlock_entry(&xas, (void *)cookie);
 479}
 480
 481/*
 482 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping
 483 * @mapping: the file's mapping whose entry we want to lock
 484 * @index: the offset within this file
 485 * @page: output the dax page corresponding to this dax entry
 486 *
 487 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry
 488 * could not be locked.
 489 */
 490dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index,
 491		struct page **page)
 492{
 493	XA_STATE(xas, NULL, 0);
 494	void *entry;
 495
 496	rcu_read_lock();
 497	for (;;) {
 498		entry = NULL;
 499		if (!dax_mapping(mapping))
 500			break;
 501
 502		xas.xa = &mapping->i_pages;
 503		xas_lock_irq(&xas);
 504		xas_set(&xas, index);
 505		entry = xas_load(&xas);
 506		if (dax_is_locked(entry)) {
 507			rcu_read_unlock();
 508			wait_entry_unlocked(&xas, entry);
 509			rcu_read_lock();
 510			continue;
 511		}
 512		if (!entry ||
 513		    dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 514			/*
 515			 * Because we are looking for entry from file's mapping
 516			 * and index, so the entry may not be inserted for now,
 517			 * or even a zero/empty entry.  We don't think this is
 518			 * an error case.  So, return a special value and do
 519			 * not output @page.
 520			 */
 521			entry = (void *)~0UL;
 522		} else {
 523			*page = pfn_to_page(dax_to_pfn(entry));
 524			dax_lock_entry(&xas, entry);
 525		}
 526		xas_unlock_irq(&xas);
 527		break;
 528	}
 529	rcu_read_unlock();
 530	return (dax_entry_t)entry;
 531}
 532
 533void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index,
 534		dax_entry_t cookie)
 535{
 536	XA_STATE(xas, &mapping->i_pages, index);
 537
 538	if (cookie == ~0UL)
 539		return;
 540
 541	dax_unlock_entry(&xas, (void *)cookie);
 542}
 543
 544/*
 545 * Find page cache entry at given index. If it is a DAX entry, return it
 546 * with the entry locked. If the page cache doesn't contain an entry at
 547 * that index, add a locked empty entry.
 548 *
 549 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will
 550 * either return that locked entry or will return VM_FAULT_FALLBACK.
 551 * This will happen if there are any PTE entries within the PMD range
 552 * that we are requesting.
 553 *
 554 * We always favor PTE entries over PMD entries. There isn't a flow where we
 555 * evict PTE entries in order to 'upgrade' them to a PMD entry.  A PMD
 556 * insertion will fail if it finds any PTE entries already in the tree, and a
 557 * PTE insertion will cause an existing PMD entry to be unmapped and
 558 * downgraded to PTE entries.  This happens for both PMD zero pages as
 559 * well as PMD empty entries.
 560 *
 561 * The exception to this downgrade path is for PMD entries that have
 562 * real storage backing them.  We will leave these real PMD entries in
 563 * the tree, and PTE writes will simply dirty the entire PMD entry.
 564 *
 565 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
 566 * persistent memory the benefit is doubtful. We can add that later if we can
 567 * show it helps.
 568 *
 569 * On error, this function does not return an ERR_PTR.  Instead it returns
 570 * a VM_FAULT code, encoded as an xarray internal entry.  The ERR_PTR values
 571 * overlap with xarray value entries.
 572 */
 573static void *grab_mapping_entry(struct xa_state *xas,
 574		struct address_space *mapping, unsigned int order)
 575{
 576	unsigned long index = xas->xa_index;
 577	bool pmd_downgrade;	/* splitting PMD entry into PTE entries? */
 578	void *entry;
 579
 580retry:
 581	pmd_downgrade = false;
 582	xas_lock_irq(xas);
 583	entry = get_unlocked_entry(xas, order);
 584
 585	if (entry) {
 586		if (dax_is_conflict(entry))
 587			goto fallback;
 588		if (!xa_is_value(entry)) {
 589			xas_set_err(xas, -EIO);
 590			goto out_unlock;
 591		}
 592
 593		if (order == 0) {
 594			if (dax_is_pmd_entry(entry) &&
 595			    (dax_is_zero_entry(entry) ||
 596			     dax_is_empty_entry(entry))) {
 597				pmd_downgrade = true;
 598			}
 599		}
 600	}
 601
 602	if (pmd_downgrade) {
 603		/*
 604		 * Make sure 'entry' remains valid while we drop
 605		 * the i_pages lock.
 606		 */
 607		dax_lock_entry(xas, entry);
 608
 609		/*
 610		 * Besides huge zero pages the only other thing that gets
 611		 * downgraded are empty entries which don't need to be
 612		 * unmapped.
 613		 */
 614		if (dax_is_zero_entry(entry)) {
 615			xas_unlock_irq(xas);
 616			unmap_mapping_pages(mapping,
 617					xas->xa_index & ~PG_PMD_COLOUR,
 618					PG_PMD_NR, false);
 619			xas_reset(xas);
 620			xas_lock_irq(xas);
 621		}
 622
 623		dax_disassociate_entry(entry, mapping, false);
 624		xas_store(xas, NULL);	/* undo the PMD join */
 625		dax_wake_entry(xas, entry, WAKE_ALL);
 626		mapping->nrpages -= PG_PMD_NR;
 627		entry = NULL;
 628		xas_set(xas, index);
 629	}
 630
 631	if (entry) {
 632		dax_lock_entry(xas, entry);
 633	} else {
 634		unsigned long flags = DAX_EMPTY;
 635
 636		if (order > 0)
 637			flags |= DAX_PMD;
 638		entry = dax_make_entry(pfn_to_pfn_t(0), flags);
 639		dax_lock_entry(xas, entry);
 640		if (xas_error(xas))
 641			goto out_unlock;
 642		mapping->nrpages += 1UL << order;
 643	}
 644
 645out_unlock:
 646	xas_unlock_irq(xas);
 647	if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM))
 648		goto retry;
 649	if (xas->xa_node == XA_ERROR(-ENOMEM))
 650		return xa_mk_internal(VM_FAULT_OOM);
 651	if (xas_error(xas))
 652		return xa_mk_internal(VM_FAULT_SIGBUS);
 653	return entry;
 654fallback:
 655	xas_unlock_irq(xas);
 656	return xa_mk_internal(VM_FAULT_FALLBACK);
 657}
 658
 659/**
 660 * dax_layout_busy_page_range - find first pinned page in @mapping
 661 * @mapping: address space to scan for a page with ref count > 1
 662 * @start: Starting offset. Page containing 'start' is included.
 663 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX,
 664 *       pages from 'start' till the end of file are included.
 665 *
 666 * DAX requires ZONE_DEVICE mapped pages. These pages are never
 667 * 'onlined' to the page allocator so they are considered idle when
 668 * page->count == 1. A filesystem uses this interface to determine if
 669 * any page in the mapping is busy, i.e. for DMA, or other
 670 * get_user_pages() usages.
 671 *
 672 * It is expected that the filesystem is holding locks to block the
 673 * establishment of new mappings in this address_space. I.e. it expects
 674 * to be able to run unmap_mapping_range() and subsequently not race
 675 * mapping_mapped() becoming true.
 676 */
 677struct page *dax_layout_busy_page_range(struct address_space *mapping,
 678					loff_t start, loff_t end)
 679{
 680	void *entry;
 681	unsigned int scanned = 0;
 682	struct page *page = NULL;
 683	pgoff_t start_idx = start >> PAGE_SHIFT;
 684	pgoff_t end_idx;
 685	XA_STATE(xas, &mapping->i_pages, start_idx);
 686
 687	/*
 688	 * In the 'limited' case get_user_pages() for dax is disabled.
 689	 */
 690	if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
 691		return NULL;
 692
 693	if (!dax_mapping(mapping) || !mapping_mapped(mapping))
 694		return NULL;
 695
 696	/* If end == LLONG_MAX, all pages from start to till end of file */
 697	if (end == LLONG_MAX)
 698		end_idx = ULONG_MAX;
 699	else
 700		end_idx = end >> PAGE_SHIFT;
 701	/*
 702	 * If we race get_user_pages_fast() here either we'll see the
 703	 * elevated page count in the iteration and wait, or
 704	 * get_user_pages_fast() will see that the page it took a reference
 705	 * against is no longer mapped in the page tables and bail to the
 706	 * get_user_pages() slow path.  The slow path is protected by
 707	 * pte_lock() and pmd_lock(). New references are not taken without
 708	 * holding those locks, and unmap_mapping_pages() will not zero the
 709	 * pte or pmd without holding the respective lock, so we are
 710	 * guaranteed to either see new references or prevent new
 711	 * references from being established.
 712	 */
 713	unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0);
 714
 715	xas_lock_irq(&xas);
 716	xas_for_each(&xas, entry, end_idx) {
 717		if (WARN_ON_ONCE(!xa_is_value(entry)))
 718			continue;
 719		if (unlikely(dax_is_locked(entry)))
 720			entry = get_unlocked_entry(&xas, 0);
 721		if (entry)
 722			page = dax_busy_page(entry);
 723		put_unlocked_entry(&xas, entry, WAKE_NEXT);
 724		if (page)
 725			break;
 726		if (++scanned % XA_CHECK_SCHED)
 727			continue;
 728
 729		xas_pause(&xas);
 730		xas_unlock_irq(&xas);
 731		cond_resched();
 732		xas_lock_irq(&xas);
 733	}
 734	xas_unlock_irq(&xas);
 735	return page;
 736}
 737EXPORT_SYMBOL_GPL(dax_layout_busy_page_range);
 738
 739struct page *dax_layout_busy_page(struct address_space *mapping)
 740{
 741	return dax_layout_busy_page_range(mapping, 0, LLONG_MAX);
 742}
 743EXPORT_SYMBOL_GPL(dax_layout_busy_page);
 744
 745static int __dax_invalidate_entry(struct address_space *mapping,
 746					  pgoff_t index, bool trunc)
 747{
 748	XA_STATE(xas, &mapping->i_pages, index);
 749	int ret = 0;
 750	void *entry;
 751
 752	xas_lock_irq(&xas);
 753	entry = get_unlocked_entry(&xas, 0);
 754	if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 755		goto out;
 756	if (!trunc &&
 757	    (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) ||
 758	     xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE)))
 759		goto out;
 760	dax_disassociate_entry(entry, mapping, trunc);
 761	xas_store(&xas, NULL);
 762	mapping->nrpages -= 1UL << dax_entry_order(entry);
 763	ret = 1;
 764out:
 765	put_unlocked_entry(&xas, entry, WAKE_ALL);
 766	xas_unlock_irq(&xas);
 767	return ret;
 768}
 769
 770static int __dax_clear_dirty_range(struct address_space *mapping,
 771		pgoff_t start, pgoff_t end)
 772{
 773	XA_STATE(xas, &mapping->i_pages, start);
 774	unsigned int scanned = 0;
 775	void *entry;
 776
 777	xas_lock_irq(&xas);
 778	xas_for_each(&xas, entry, end) {
 779		entry = get_unlocked_entry(&xas, 0);
 780		xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
 781		xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
 782		put_unlocked_entry(&xas, entry, WAKE_NEXT);
 783
 784		if (++scanned % XA_CHECK_SCHED)
 785			continue;
 786
 787		xas_pause(&xas);
 788		xas_unlock_irq(&xas);
 789		cond_resched();
 790		xas_lock_irq(&xas);
 791	}
 792	xas_unlock_irq(&xas);
 793
 794	return 0;
 795}
 796
 797/*
 798 * Delete DAX entry at @index from @mapping.  Wait for it
 799 * to be unlocked before deleting it.
 800 */
 801int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
 802{
 803	int ret = __dax_invalidate_entry(mapping, index, true);
 804
 805	/*
 806	 * This gets called from truncate / punch_hole path. As such, the caller
 807	 * must hold locks protecting against concurrent modifications of the
 808	 * page cache (usually fs-private i_mmap_sem for writing). Since the
 809	 * caller has seen a DAX entry for this index, we better find it
 810	 * at that index as well...
 811	 */
 812	WARN_ON_ONCE(!ret);
 813	return ret;
 814}
 815
 816/*
 817 * Invalidate DAX entry if it is clean.
 818 */
 819int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
 820				      pgoff_t index)
 821{
 822	return __dax_invalidate_entry(mapping, index, false);
 823}
 824
 825static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos)
 826{
 827	return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset);
 828}
 829
 830static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter)
 831{
 832	pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos);
 833	void *vto, *kaddr;
 834	long rc;
 835	int id;
 836
 837	id = dax_read_lock();
 838	rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS,
 839				&kaddr, NULL);
 840	if (rc < 0) {
 841		dax_read_unlock(id);
 842		return rc;
 843	}
 844	vto = kmap_atomic(vmf->cow_page);
 845	copy_user_page(vto, kaddr, vmf->address, vmf->cow_page);
 846	kunmap_atomic(vto);
 847	dax_read_unlock(id);
 848	return 0;
 849}
 850
 851/*
 852 * MAP_SYNC on a dax mapping guarantees dirty metadata is
 853 * flushed on write-faults (non-cow), but not read-faults.
 854 */
 855static bool dax_fault_is_synchronous(const struct iomap_iter *iter,
 856		struct vm_area_struct *vma)
 857{
 858	return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) &&
 859		(iter->iomap.flags & IOMAP_F_DIRTY);
 860}
 861
 862/*
 863 * By this point grab_mapping_entry() has ensured that we have a locked entry
 864 * of the appropriate size so we don't have to worry about downgrading PMDs to
 865 * PTEs.  If we happen to be trying to insert a PTE and there is a PMD
 866 * already in the tree, we will skip the insertion and just dirty the PMD as
 867 * appropriate.
 868 */
 869static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf,
 870		const struct iomap_iter *iter, void *entry, pfn_t pfn,
 871		unsigned long flags)
 872{
 873	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
 874	void *new_entry = dax_make_entry(pfn, flags);
 875	bool write = iter->flags & IOMAP_WRITE;
 876	bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma);
 877	bool shared = iter->iomap.flags & IOMAP_F_SHARED;
 878
 879	if (dirty)
 880		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 881
 882	if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) {
 883		unsigned long index = xas->xa_index;
 884		/* we are replacing a zero page with block mapping */
 885		if (dax_is_pmd_entry(entry))
 886			unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR,
 887					PG_PMD_NR, false);
 888		else /* pte entry */
 889			unmap_mapping_pages(mapping, index, 1, false);
 890	}
 891
 892	xas_reset(xas);
 893	xas_lock_irq(xas);
 894	if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
 895		void *old;
 896
 897		dax_disassociate_entry(entry, mapping, false);
 898		dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address,
 899				shared);
 900		/*
 901		 * Only swap our new entry into the page cache if the current
 902		 * entry is a zero page or an empty entry.  If a normal PTE or
 903		 * PMD entry is already in the cache, we leave it alone.  This
 904		 * means that if we are trying to insert a PTE and the
 905		 * existing entry is a PMD, we will just leave the PMD in the
 906		 * tree and dirty it if necessary.
 907		 */
 908		old = dax_lock_entry(xas, new_entry);
 909		WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) |
 910					DAX_LOCKED));
 911		entry = new_entry;
 912	} else {
 913		xas_load(xas);	/* Walk the xa_state */
 914	}
 915
 916	if (dirty)
 917		xas_set_mark(xas, PAGECACHE_TAG_DIRTY);
 918
 919	if (write && shared)
 920		xas_set_mark(xas, PAGECACHE_TAG_TOWRITE);
 921
 922	xas_unlock_irq(xas);
 923	return entry;
 924}
 925
 926static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev,
 927		struct address_space *mapping, void *entry)
 928{
 929	unsigned long pfn, index, count, end;
 930	long ret = 0;
 931	struct vm_area_struct *vma;
 932
 933	/*
 934	 * A page got tagged dirty in DAX mapping? Something is seriously
 935	 * wrong.
 936	 */
 937	if (WARN_ON(!xa_is_value(entry)))
 938		return -EIO;
 939
 940	if (unlikely(dax_is_locked(entry))) {
 941		void *old_entry = entry;
 942
 943		entry = get_unlocked_entry(xas, 0);
 944
 945		/* Entry got punched out / reallocated? */
 946		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
 947			goto put_unlocked;
 948		/*
 949		 * Entry got reallocated elsewhere? No need to writeback.
 950		 * We have to compare pfns as we must not bail out due to
 951		 * difference in lockbit or entry type.
 952		 */
 953		if (dax_to_pfn(old_entry) != dax_to_pfn(entry))
 954			goto put_unlocked;
 955		if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
 956					dax_is_zero_entry(entry))) {
 957			ret = -EIO;
 958			goto put_unlocked;
 959		}
 960
 961		/* Another fsync thread may have already done this entry */
 962		if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE))
 963			goto put_unlocked;
 964	}
 965
 966	/* Lock the entry to serialize with page faults */
 967	dax_lock_entry(xas, entry);
 968
 969	/*
 970	 * We can clear the tag now but we have to be careful so that concurrent
 971	 * dax_writeback_one() calls for the same index cannot finish before we
 972	 * actually flush the caches. This is achieved as the calls will look
 973	 * at the entry only under the i_pages lock and once they do that
 974	 * they will see the entry locked and wait for it to unlock.
 975	 */
 976	xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE);
 977	xas_unlock_irq(xas);
 978
 979	/*
 980	 * If dax_writeback_mapping_range() was given a wbc->range_start
 981	 * in the middle of a PMD, the 'index' we use needs to be
 982	 * aligned to the start of the PMD.
 983	 * This allows us to flush for PMD_SIZE and not have to worry about
 984	 * partial PMD writebacks.
 985	 */
 986	pfn = dax_to_pfn(entry);
 987	count = 1UL << dax_entry_order(entry);
 988	index = xas->xa_index & ~(count - 1);
 989	end = index + count - 1;
 990
 991	/* Walk all mappings of a given index of a file and writeprotect them */
 992	i_mmap_lock_read(mapping);
 993	vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) {
 994		pfn_mkclean_range(pfn, count, index, vma);
 995		cond_resched();
 996	}
 997	i_mmap_unlock_read(mapping);
 998
 999	dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE);
1000	/*
1001	 * After we have flushed the cache, we can clear the dirty tag. There
1002	 * cannot be new dirty data in the pfn after the flush has completed as
1003	 * the pfn mappings are writeprotected and fault waits for mapping
1004	 * entry lock.
1005	 */
1006	xas_reset(xas);
1007	xas_lock_irq(xas);
1008	xas_store(xas, entry);
1009	xas_clear_mark(xas, PAGECACHE_TAG_DIRTY);
1010	dax_wake_entry(xas, entry, WAKE_NEXT);
1011
1012	trace_dax_writeback_one(mapping->host, index, count);
1013	return ret;
1014
1015 put_unlocked:
1016	put_unlocked_entry(xas, entry, WAKE_NEXT);
1017	return ret;
1018}
1019
1020/*
1021 * Flush the mapping to the persistent domain within the byte range of [start,
1022 * end]. This is required by data integrity operations to ensure file data is
1023 * on persistent storage prior to completion of the operation.
1024 */
1025int dax_writeback_mapping_range(struct address_space *mapping,
1026		struct dax_device *dax_dev, struct writeback_control *wbc)
1027{
1028	XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT);
1029	struct inode *inode = mapping->host;
1030	pgoff_t end_index = wbc->range_end >> PAGE_SHIFT;
1031	void *entry;
1032	int ret = 0;
1033	unsigned int scanned = 0;
1034
1035	if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
1036		return -EIO;
1037
1038	if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL)
1039		return 0;
1040
1041	trace_dax_writeback_range(inode, xas.xa_index, end_index);
1042
1043	tag_pages_for_writeback(mapping, xas.xa_index, end_index);
1044
1045	xas_lock_irq(&xas);
1046	xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) {
1047		ret = dax_writeback_one(&xas, dax_dev, mapping, entry);
1048		if (ret < 0) {
1049			mapping_set_error(mapping, ret);
1050			break;
1051		}
1052		if (++scanned % XA_CHECK_SCHED)
1053			continue;
1054
1055		xas_pause(&xas);
1056		xas_unlock_irq(&xas);
1057		cond_resched();
1058		xas_lock_irq(&xas);
1059	}
1060	xas_unlock_irq(&xas);
1061	trace_dax_writeback_range_done(inode, xas.xa_index, end_index);
1062	return ret;
1063}
1064EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
1065
1066static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos,
1067		size_t size, void **kaddr, pfn_t *pfnp)
1068{
1069	pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1070	int id, rc = 0;
1071	long length;
1072
1073	id = dax_read_lock();
1074	length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size),
1075				   DAX_ACCESS, kaddr, pfnp);
1076	if (length < 0) {
1077		rc = length;
1078		goto out;
1079	}
1080	if (!pfnp)
1081		goto out_check_addr;
1082	rc = -EINVAL;
1083	if (PFN_PHYS(length) < size)
1084		goto out;
1085	if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1))
1086		goto out;
1087	/* For larger pages we need devmap */
1088	if (length > 1 && !pfn_t_devmap(*pfnp))
1089		goto out;
1090	rc = 0;
1091
1092out_check_addr:
1093	if (!kaddr)
1094		goto out;
1095	if (!*kaddr)
1096		rc = -EFAULT;
1097out:
1098	dax_read_unlock(id);
1099	return rc;
1100}
1101
1102/**
1103 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page
1104 * by copying the data before and after the range to be written.
1105 * @pos:	address to do copy from.
1106 * @length:	size of copy operation.
1107 * @align_size:	aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE)
1108 * @srcmap:	iomap srcmap
1109 * @daddr:	destination address to copy to.
1110 *
1111 * This can be called from two places. Either during DAX write fault (page
1112 * aligned), to copy the length size data to daddr. Or, while doing normal DAX
1113 * write operation, dax_iomap_iter() might call this to do the copy of either
1114 * start or end unaligned address. In the latter case the rest of the copy of
1115 * aligned ranges is taken care by dax_iomap_iter() itself.
1116 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the
1117 * area to make sure no old data remains.
1118 */
1119static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size,
1120		const struct iomap *srcmap, void *daddr)
1121{
1122	loff_t head_off = pos & (align_size - 1);
1123	size_t size = ALIGN(head_off + length, align_size);
1124	loff_t end = pos + length;
1125	loff_t pg_end = round_up(end, align_size);
1126	/* copy_all is usually in page fault case */
1127	bool copy_all = head_off == 0 && end == pg_end;
1128	/* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */
1129	bool zero_edge = srcmap->flags & IOMAP_F_SHARED ||
1130			 srcmap->type == IOMAP_UNWRITTEN;
1131	void *saddr = NULL;
1132	int ret = 0;
1133
1134	if (!zero_edge) {
1135		ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL);
1136		if (ret)
1137			return dax_mem2blk_err(ret);
1138	}
1139
1140	if (copy_all) {
1141		if (zero_edge)
1142			memset(daddr, 0, size);
1143		else
1144			ret = copy_mc_to_kernel(daddr, saddr, length);
1145		goto out;
1146	}
1147
1148	/* Copy the head part of the range */
1149	if (head_off) {
1150		if (zero_edge)
1151			memset(daddr, 0, head_off);
1152		else {
1153			ret = copy_mc_to_kernel(daddr, saddr, head_off);
1154			if (ret)
1155				return -EIO;
1156		}
1157	}
1158
1159	/* Copy the tail part of the range */
1160	if (end < pg_end) {
1161		loff_t tail_off = head_off + length;
1162		loff_t tail_len = pg_end - end;
1163
1164		if (zero_edge)
1165			memset(daddr + tail_off, 0, tail_len);
1166		else {
1167			ret = copy_mc_to_kernel(daddr + tail_off,
1168						saddr + tail_off, tail_len);
1169			if (ret)
1170				return -EIO;
1171		}
1172	}
1173out:
1174	if (zero_edge)
1175		dax_flush(srcmap->dax_dev, daddr, size);
1176	return ret ? -EIO : 0;
1177}
1178
1179/*
1180 * The user has performed a load from a hole in the file.  Allocating a new
1181 * page in the file would cause excessive storage usage for workloads with
1182 * sparse files.  Instead we insert a read-only mapping of the 4k zero page.
1183 * If this page is ever written to we will re-fault and change the mapping to
1184 * point to real DAX storage instead.
1185 */
1186static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1187		const struct iomap_iter *iter, void **entry)
1188{
1189	struct inode *inode = iter->inode;
1190	unsigned long vaddr = vmf->address;
1191	pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr));
1192	vm_fault_t ret;
1193
1194	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE);
1195
1196	ret = vmf_insert_mixed(vmf->vma, vaddr, pfn);
1197	trace_dax_load_hole(inode, vmf, ret);
1198	return ret;
1199}
1200
1201#ifdef CONFIG_FS_DAX_PMD
1202static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1203		const struct iomap_iter *iter, void **entry)
1204{
1205	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1206	unsigned long pmd_addr = vmf->address & PMD_MASK;
1207	struct vm_area_struct *vma = vmf->vma;
1208	struct inode *inode = mapping->host;
1209	pgtable_t pgtable = NULL;
1210	struct page *zero_page;
1211	spinlock_t *ptl;
1212	pmd_t pmd_entry;
1213	pfn_t pfn;
1214
1215	zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
1216
1217	if (unlikely(!zero_page))
1218		goto fallback;
1219
1220	pfn = page_to_pfn_t(zero_page);
1221	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn,
1222				  DAX_PMD | DAX_ZERO_PAGE);
1223
1224	if (arch_needs_pgtable_deposit()) {
1225		pgtable = pte_alloc_one(vma->vm_mm);
1226		if (!pgtable)
1227			return VM_FAULT_OOM;
1228	}
1229
1230	ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1231	if (!pmd_none(*(vmf->pmd))) {
1232		spin_unlock(ptl);
1233		goto fallback;
1234	}
1235
1236	if (pgtable) {
1237		pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
1238		mm_inc_nr_ptes(vma->vm_mm);
1239	}
1240	pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
1241	pmd_entry = pmd_mkhuge(pmd_entry);
1242	set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
1243	spin_unlock(ptl);
1244	trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry);
1245	return VM_FAULT_NOPAGE;
1246
1247fallback:
1248	if (pgtable)
1249		pte_free(vma->vm_mm, pgtable);
1250	trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry);
1251	return VM_FAULT_FALLBACK;
1252}
1253#else
1254static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf,
1255		const struct iomap_iter *iter, void **entry)
1256{
1257	return VM_FAULT_FALLBACK;
1258}
1259#endif /* CONFIG_FS_DAX_PMD */
1260
1261static s64 dax_unshare_iter(struct iomap_iter *iter)
1262{
1263	struct iomap *iomap = &iter->iomap;
1264	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1265	loff_t pos = iter->pos;
1266	loff_t length = iomap_length(iter);
1267	int id = 0;
1268	s64 ret = 0;
1269	void *daddr = NULL, *saddr = NULL;
1270
1271	/* don't bother with blocks that are not shared to start with */
1272	if (!(iomap->flags & IOMAP_F_SHARED))
1273		return length;
 
 
 
1274
1275	id = dax_read_lock();
1276	ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL);
1277	if (ret < 0)
1278		goto out_unlock;
1279
1280	/* zero the distance if srcmap is HOLE or UNWRITTEN */
1281	if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) {
1282		memset(daddr, 0, length);
1283		dax_flush(iomap->dax_dev, daddr, length);
1284		ret = length;
1285		goto out_unlock;
1286	}
1287
1288	ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL);
1289	if (ret < 0)
1290		goto out_unlock;
1291
1292	if (copy_mc_to_kernel(daddr, saddr, length) == 0)
1293		ret = length;
1294	else
1295		ret = -EIO;
1296
1297out_unlock:
1298	dax_read_unlock(id);
1299	return dax_mem2blk_err(ret);
1300}
1301
1302int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len,
1303		const struct iomap_ops *ops)
1304{
1305	struct iomap_iter iter = {
1306		.inode		= inode,
1307		.pos		= pos,
1308		.len		= len,
1309		.flags		= IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX,
1310	};
1311	int ret;
1312
1313	while ((ret = iomap_iter(&iter, ops)) > 0)
1314		iter.processed = dax_unshare_iter(&iter);
1315	return ret;
1316}
1317EXPORT_SYMBOL_GPL(dax_file_unshare);
1318
1319static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size)
1320{
1321	const struct iomap *iomap = &iter->iomap;
1322	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1323	unsigned offset = offset_in_page(pos);
1324	pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1325	void *kaddr;
1326	long ret;
1327
1328	ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr,
1329				NULL);
1330	if (ret < 0)
1331		return dax_mem2blk_err(ret);
1332
1333	memset(kaddr + offset, 0, size);
1334	if (iomap->flags & IOMAP_F_SHARED)
1335		ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap,
1336					    kaddr);
1337	else
1338		dax_flush(iomap->dax_dev, kaddr + offset, size);
1339	return ret;
1340}
1341
1342static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero)
1343{
1344	const struct iomap *iomap = &iter->iomap;
1345	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1346	loff_t pos = iter->pos;
1347	u64 length = iomap_length(iter);
1348	s64 written = 0;
1349
1350	/* already zeroed?  we're done. */
1351	if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN)
1352		return length;
1353
1354	/*
1355	 * invalidate the pages whose sharing state is to be changed
1356	 * because of CoW.
1357	 */
1358	if (iomap->flags & IOMAP_F_SHARED)
1359		invalidate_inode_pages2_range(iter->inode->i_mapping,
1360					      pos >> PAGE_SHIFT,
1361					      (pos + length - 1) >> PAGE_SHIFT);
1362
1363	do {
1364		unsigned offset = offset_in_page(pos);
1365		unsigned size = min_t(u64, PAGE_SIZE - offset, length);
1366		pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1367		long rc;
1368		int id;
1369
1370		id = dax_read_lock();
1371		if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE)
1372			rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1);
1373		else
1374			rc = dax_memzero(iter, pos, size);
1375		dax_read_unlock(id);
1376
1377		if (rc < 0)
1378			return rc;
1379		pos += size;
1380		length -= size;
1381		written += size;
1382	} while (length > 0);
1383
1384	if (did_zero)
1385		*did_zero = true;
1386	return written;
1387}
1388
1389int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero,
1390		const struct iomap_ops *ops)
1391{
1392	struct iomap_iter iter = {
1393		.inode		= inode,
1394		.pos		= pos,
1395		.len		= len,
1396		.flags		= IOMAP_DAX | IOMAP_ZERO,
1397	};
1398	int ret;
1399
1400	while ((ret = iomap_iter(&iter, ops)) > 0)
1401		iter.processed = dax_zero_iter(&iter, did_zero);
1402	return ret;
1403}
1404EXPORT_SYMBOL_GPL(dax_zero_range);
1405
1406int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero,
1407		const struct iomap_ops *ops)
1408{
1409	unsigned int blocksize = i_blocksize(inode);
1410	unsigned int off = pos & (blocksize - 1);
1411
1412	/* Block boundary? Nothing to do */
1413	if (!off)
1414		return 0;
1415	return dax_zero_range(inode, pos, blocksize - off, did_zero, ops);
1416}
1417EXPORT_SYMBOL_GPL(dax_truncate_page);
1418
1419static loff_t dax_iomap_iter(const struct iomap_iter *iomi,
1420		struct iov_iter *iter)
1421{
1422	const struct iomap *iomap = &iomi->iomap;
1423	const struct iomap *srcmap = iomap_iter_srcmap(iomi);
1424	loff_t length = iomap_length(iomi);
1425	loff_t pos = iomi->pos;
1426	struct dax_device *dax_dev = iomap->dax_dev;
1427	loff_t end = pos + length, done = 0;
1428	bool write = iov_iter_rw(iter) == WRITE;
1429	bool cow = write && iomap->flags & IOMAP_F_SHARED;
1430	ssize_t ret = 0;
1431	size_t xfer;
1432	int id;
1433
1434	if (!write) {
1435		end = min(end, i_size_read(iomi->inode));
1436		if (pos >= end)
1437			return 0;
1438
1439		if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
1440			return iov_iter_zero(min(length, end - pos), iter);
1441	}
1442
1443	/*
1444	 * In DAX mode, enforce either pure overwrites of written extents, or
1445	 * writes to unwritten extents as part of a copy-on-write operation.
1446	 */
1447	if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED &&
1448			!(iomap->flags & IOMAP_F_SHARED)))
1449		return -EIO;
1450
1451	/*
1452	 * Write can allocate block for an area which has a hole page mapped
1453	 * into page tables. We have to tear down these mappings so that data
1454	 * written by write(2) is visible in mmap.
1455	 */
1456	if (iomap->flags & IOMAP_F_NEW || cow) {
1457		/*
1458		 * Filesystem allows CoW on non-shared extents. The src extents
1459		 * may have been mmapped with dirty mark before. To be able to
1460		 * invalidate its dax entries, we need to clear the dirty mark
1461		 * in advance.
1462		 */
1463		if (cow)
1464			__dax_clear_dirty_range(iomi->inode->i_mapping,
1465						pos >> PAGE_SHIFT,
1466						(end - 1) >> PAGE_SHIFT);
1467		invalidate_inode_pages2_range(iomi->inode->i_mapping,
1468					      pos >> PAGE_SHIFT,
1469					      (end - 1) >> PAGE_SHIFT);
1470	}
1471
1472	id = dax_read_lock();
1473	while (pos < end) {
1474		unsigned offset = pos & (PAGE_SIZE - 1);
1475		const size_t size = ALIGN(length + offset, PAGE_SIZE);
1476		pgoff_t pgoff = dax_iomap_pgoff(iomap, pos);
1477		ssize_t map_len;
1478		bool recovery = false;
1479		void *kaddr;
1480
1481		if (fatal_signal_pending(current)) {
1482			ret = -EINTR;
1483			break;
1484		}
1485
1486		map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
1487				DAX_ACCESS, &kaddr, NULL);
1488		if (map_len == -EHWPOISON && iov_iter_rw(iter) == WRITE) {
1489			map_len = dax_direct_access(dax_dev, pgoff,
1490					PHYS_PFN(size), DAX_RECOVERY_WRITE,
1491					&kaddr, NULL);
1492			if (map_len > 0)
1493				recovery = true;
1494		}
1495		if (map_len < 0) {
1496			ret = dax_mem2blk_err(map_len);
1497			break;
1498		}
1499
1500		if (cow) {
1501			ret = dax_iomap_copy_around(pos, length, PAGE_SIZE,
1502						    srcmap, kaddr);
1503			if (ret)
1504				break;
1505		}
1506
1507		map_len = PFN_PHYS(map_len);
1508		kaddr += offset;
1509		map_len -= offset;
1510		if (map_len > end - pos)
1511			map_len = end - pos;
1512
1513		if (recovery)
1514			xfer = dax_recovery_write(dax_dev, pgoff, kaddr,
1515					map_len, iter);
1516		else if (write)
1517			xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr,
1518					map_len, iter);
1519		else
1520			xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr,
1521					map_len, iter);
1522
1523		pos += xfer;
1524		length -= xfer;
1525		done += xfer;
1526
1527		if (xfer == 0)
1528			ret = -EFAULT;
1529		if (xfer < map_len)
1530			break;
1531	}
1532	dax_read_unlock(id);
1533
1534	return done ? done : ret;
1535}
1536
1537/**
1538 * dax_iomap_rw - Perform I/O to a DAX file
1539 * @iocb:	The control block for this I/O
1540 * @iter:	The addresses to do I/O from or to
1541 * @ops:	iomap ops passed from the file system
1542 *
1543 * This function performs read and write operations to directly mapped
1544 * persistent memory.  The callers needs to take care of read/write exclusion
1545 * and evicting any page cache pages in the region under I/O.
1546 */
1547ssize_t
1548dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
1549		const struct iomap_ops *ops)
1550{
1551	struct iomap_iter iomi = {
1552		.inode		= iocb->ki_filp->f_mapping->host,
1553		.pos		= iocb->ki_pos,
1554		.len		= iov_iter_count(iter),
1555		.flags		= IOMAP_DAX,
1556	};
1557	loff_t done = 0;
1558	int ret;
1559
1560	if (!iomi.len)
1561		return 0;
1562
1563	if (iov_iter_rw(iter) == WRITE) {
1564		lockdep_assert_held_write(&iomi.inode->i_rwsem);
1565		iomi.flags |= IOMAP_WRITE;
1566	} else {
1567		lockdep_assert_held(&iomi.inode->i_rwsem);
1568	}
1569
1570	if (iocb->ki_flags & IOCB_NOWAIT)
1571		iomi.flags |= IOMAP_NOWAIT;
1572
1573	while ((ret = iomap_iter(&iomi, ops)) > 0)
1574		iomi.processed = dax_iomap_iter(&iomi, iter);
1575
1576	done = iomi.pos - iocb->ki_pos;
1577	iocb->ki_pos = iomi.pos;
1578	return done ? done : ret;
1579}
1580EXPORT_SYMBOL_GPL(dax_iomap_rw);
1581
1582static vm_fault_t dax_fault_return(int error)
1583{
1584	if (error == 0)
1585		return VM_FAULT_NOPAGE;
1586	return vmf_error(error);
1587}
1588
1589/*
1590 * When handling a synchronous page fault and the inode need a fsync, we can
1591 * insert the PTE/PMD into page tables only after that fsync happened. Skip
1592 * insertion for now and return the pfn so that caller can insert it after the
1593 * fsync is done.
1594 */
1595static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn)
1596{
1597	if (WARN_ON_ONCE(!pfnp))
1598		return VM_FAULT_SIGBUS;
1599	*pfnp = pfn;
1600	return VM_FAULT_NEEDDSYNC;
1601}
1602
1603static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf,
1604		const struct iomap_iter *iter)
1605{
1606	vm_fault_t ret;
1607	int error = 0;
1608
1609	switch (iter->iomap.type) {
1610	case IOMAP_HOLE:
1611	case IOMAP_UNWRITTEN:
1612		clear_user_highpage(vmf->cow_page, vmf->address);
1613		break;
1614	case IOMAP_MAPPED:
1615		error = copy_cow_page_dax(vmf, iter);
1616		break;
1617	default:
1618		WARN_ON_ONCE(1);
1619		error = -EIO;
1620		break;
1621	}
1622
1623	if (error)
1624		return dax_fault_return(error);
1625
1626	__SetPageUptodate(vmf->cow_page);
1627	ret = finish_fault(vmf);
1628	if (!ret)
1629		return VM_FAULT_DONE_COW;
1630	return ret;
1631}
1632
1633/**
1634 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault.
1635 * @vmf:	vm fault instance
1636 * @iter:	iomap iter
1637 * @pfnp:	pfn to be returned
1638 * @xas:	the dax mapping tree of a file
1639 * @entry:	an unlocked dax entry to be inserted
1640 * @pmd:	distinguish whether it is a pmd fault
1641 */
1642static vm_fault_t dax_fault_iter(struct vm_fault *vmf,
1643		const struct iomap_iter *iter, pfn_t *pfnp,
1644		struct xa_state *xas, void **entry, bool pmd)
1645{
1646	const struct iomap *iomap = &iter->iomap;
1647	const struct iomap *srcmap = iomap_iter_srcmap(iter);
1648	size_t size = pmd ? PMD_SIZE : PAGE_SIZE;
1649	loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT;
1650	bool write = iter->flags & IOMAP_WRITE;
1651	unsigned long entry_flags = pmd ? DAX_PMD : 0;
1652	int err = 0;
1653	pfn_t pfn;
1654	void *kaddr;
1655
1656	if (!pmd && vmf->cow_page)
1657		return dax_fault_cow_page(vmf, iter);
1658
1659	/* if we are reading UNWRITTEN and HOLE, return a hole. */
1660	if (!write &&
1661	    (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) {
1662		if (!pmd)
1663			return dax_load_hole(xas, vmf, iter, entry);
1664		return dax_pmd_load_hole(xas, vmf, iter, entry);
1665	}
1666
1667	if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) {
1668		WARN_ON_ONCE(1);
1669		return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS;
1670	}
1671
1672	err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn);
1673	if (err)
1674		return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err);
1675
1676	*entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags);
1677
1678	if (write && iomap->flags & IOMAP_F_SHARED) {
1679		err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr);
1680		if (err)
1681			return dax_fault_return(err);
1682	}
1683
1684	if (dax_fault_is_synchronous(iter, vmf->vma))
1685		return dax_fault_synchronous_pfnp(pfnp, pfn);
1686
1687	/* insert PMD pfn */
1688	if (pmd)
1689		return vmf_insert_pfn_pmd(vmf, pfn, write);
1690
1691	/* insert PTE pfn */
1692	if (write)
1693		return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1694	return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
1695}
1696
1697static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp,
1698			       int *iomap_errp, const struct iomap_ops *ops)
1699{
1700	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1701	XA_STATE(xas, &mapping->i_pages, vmf->pgoff);
1702	struct iomap_iter iter = {
1703		.inode		= mapping->host,
1704		.pos		= (loff_t)vmf->pgoff << PAGE_SHIFT,
1705		.len		= PAGE_SIZE,
1706		.flags		= IOMAP_DAX | IOMAP_FAULT,
1707	};
1708	vm_fault_t ret = 0;
1709	void *entry;
1710	int error;
1711
1712	trace_dax_pte_fault(iter.inode, vmf, ret);
1713	/*
1714	 * Check whether offset isn't beyond end of file now. Caller is supposed
1715	 * to hold locks serializing us with truncate / punch hole so this is
1716	 * a reliable test.
1717	 */
1718	if (iter.pos >= i_size_read(iter.inode)) {
1719		ret = VM_FAULT_SIGBUS;
1720		goto out;
1721	}
1722
1723	if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
1724		iter.flags |= IOMAP_WRITE;
1725
1726	entry = grab_mapping_entry(&xas, mapping, 0);
1727	if (xa_is_internal(entry)) {
1728		ret = xa_to_internal(entry);
1729		goto out;
1730	}
1731
1732	/*
1733	 * It is possible, particularly with mixed reads & writes to private
1734	 * mappings, that we have raced with a PMD fault that overlaps with
1735	 * the PTE we need to set up.  If so just return and the fault will be
1736	 * retried.
1737	 */
1738	if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
1739		ret = VM_FAULT_NOPAGE;
1740		goto unlock_entry;
1741	}
1742
1743	while ((error = iomap_iter(&iter, ops)) > 0) {
1744		if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) {
1745			iter.processed = -EIO;	/* fs corruption? */
1746			continue;
1747		}
1748
1749		ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false);
1750		if (ret != VM_FAULT_SIGBUS &&
1751		    (iter.iomap.flags & IOMAP_F_NEW)) {
1752			count_vm_event(PGMAJFAULT);
1753			count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
1754			ret |= VM_FAULT_MAJOR;
1755		}
1756
1757		if (!(ret & VM_FAULT_ERROR))
1758			iter.processed = PAGE_SIZE;
1759	}
1760
1761	if (iomap_errp)
1762		*iomap_errp = error;
1763	if (!ret && error)
1764		ret = dax_fault_return(error);
1765
1766unlock_entry:
1767	dax_unlock_entry(&xas, entry);
1768out:
1769	trace_dax_pte_fault_done(iter.inode, vmf, ret);
1770	return ret;
1771}
1772
1773#ifdef CONFIG_FS_DAX_PMD
1774static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas,
1775		pgoff_t max_pgoff)
1776{
1777	unsigned long pmd_addr = vmf->address & PMD_MASK;
1778	bool write = vmf->flags & FAULT_FLAG_WRITE;
1779
1780	/*
1781	 * Make sure that the faulting address's PMD offset (color) matches
1782	 * the PMD offset from the start of the file.  This is necessary so
1783	 * that a PMD range in the page table overlaps exactly with a PMD
1784	 * range in the page cache.
1785	 */
1786	if ((vmf->pgoff & PG_PMD_COLOUR) !=
1787	    ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
1788		return true;
1789
1790	/* Fall back to PTEs if we're going to COW */
1791	if (write && !(vmf->vma->vm_flags & VM_SHARED))
1792		return true;
1793
1794	/* If the PMD would extend outside the VMA */
1795	if (pmd_addr < vmf->vma->vm_start)
1796		return true;
1797	if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
1798		return true;
1799
1800	/* If the PMD would extend beyond the file size */
1801	if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff)
1802		return true;
1803
1804	return false;
1805}
1806
1807static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1808			       const struct iomap_ops *ops)
1809{
1810	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1811	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER);
1812	struct iomap_iter iter = {
1813		.inode		= mapping->host,
1814		.len		= PMD_SIZE,
1815		.flags		= IOMAP_DAX | IOMAP_FAULT,
1816	};
1817	vm_fault_t ret = VM_FAULT_FALLBACK;
1818	pgoff_t max_pgoff;
1819	void *entry;
 
1820
1821	if (vmf->flags & FAULT_FLAG_WRITE)
1822		iter.flags |= IOMAP_WRITE;
1823
1824	/*
1825	 * Check whether offset isn't beyond end of file now. Caller is
1826	 * supposed to hold locks serializing us with truncate / punch hole so
1827	 * this is a reliable test.
1828	 */
1829	max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE);
1830
1831	trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0);
1832
1833	if (xas.xa_index >= max_pgoff) {
1834		ret = VM_FAULT_SIGBUS;
1835		goto out;
1836	}
1837
1838	if (dax_fault_check_fallback(vmf, &xas, max_pgoff))
1839		goto fallback;
1840
1841	/*
1842	 * grab_mapping_entry() will make sure we get an empty PMD entry,
1843	 * a zero PMD entry or a DAX PMD.  If it can't (because a PTE
1844	 * entry is already in the array, for instance), it will return
1845	 * VM_FAULT_FALLBACK.
1846	 */
1847	entry = grab_mapping_entry(&xas, mapping, PMD_ORDER);
1848	if (xa_is_internal(entry)) {
1849		ret = xa_to_internal(entry);
1850		goto fallback;
1851	}
1852
1853	/*
1854	 * It is possible, particularly with mixed reads & writes to private
1855	 * mappings, that we have raced with a PTE fault that overlaps with
1856	 * the PMD we need to set up.  If so just return and the fault will be
1857	 * retried.
1858	 */
1859	if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
1860			!pmd_devmap(*vmf->pmd)) {
1861		ret = 0;
1862		goto unlock_entry;
1863	}
1864
1865	iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT;
1866	while (iomap_iter(&iter, ops) > 0) {
1867		if (iomap_length(&iter) < PMD_SIZE)
1868			continue; /* actually breaks out of the loop */
1869
1870		ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true);
1871		if (ret != VM_FAULT_FALLBACK)
1872			iter.processed = PMD_SIZE;
1873	}
1874
1875unlock_entry:
1876	dax_unlock_entry(&xas, entry);
1877fallback:
1878	if (ret == VM_FAULT_FALLBACK) {
1879		split_huge_pmd(vmf->vma, vmf->pmd, vmf->address);
1880		count_vm_event(THP_FAULT_FALLBACK);
1881	}
1882out:
1883	trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret);
1884	return ret;
1885}
1886#else
1887static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp,
1888			       const struct iomap_ops *ops)
1889{
1890	return VM_FAULT_FALLBACK;
1891}
1892#endif /* CONFIG_FS_DAX_PMD */
1893
1894/**
1895 * dax_iomap_fault - handle a page fault on a DAX file
1896 * @vmf: The description of the fault
1897 * @order: Order of the page to fault in
1898 * @pfnp: PFN to insert for synchronous faults if fsync is required
1899 * @iomap_errp: Storage for detailed error code in case of error
1900 * @ops: Iomap ops passed from the file system
1901 *
1902 * When a page fault occurs, filesystems may call this helper in
1903 * their fault handler for DAX files. dax_iomap_fault() assumes the caller
1904 * has done all the necessary locking for page fault to proceed
1905 * successfully.
1906 */
1907vm_fault_t dax_iomap_fault(struct vm_fault *vmf, unsigned int order,
1908		    pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops)
1909{
1910	if (order == 0)
 
1911		return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops);
1912	else if (order == PMD_ORDER)
1913		return dax_iomap_pmd_fault(vmf, pfnp, ops);
1914	else
1915		return VM_FAULT_FALLBACK;
 
1916}
1917EXPORT_SYMBOL_GPL(dax_iomap_fault);
1918
1919/*
1920 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables
1921 * @vmf: The description of the fault
1922 * @pfn: PFN to insert
1923 * @order: Order of entry to insert.
1924 *
1925 * This function inserts a writeable PTE or PMD entry into the page tables
1926 * for an mmaped DAX file.  It also marks the page cache entry as dirty.
1927 */
1928static vm_fault_t
1929dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order)
1930{
1931	struct address_space *mapping = vmf->vma->vm_file->f_mapping;
1932	XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order);
1933	void *entry;
1934	vm_fault_t ret;
1935
1936	xas_lock_irq(&xas);
1937	entry = get_unlocked_entry(&xas, order);
1938	/* Did we race with someone splitting entry or so? */
1939	if (!entry || dax_is_conflict(entry) ||
1940	    (order == 0 && !dax_is_pte_entry(entry))) {
1941		put_unlocked_entry(&xas, entry, WAKE_NEXT);
1942		xas_unlock_irq(&xas);
1943		trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf,
1944						      VM_FAULT_NOPAGE);
1945		return VM_FAULT_NOPAGE;
1946	}
1947	xas_set_mark(&xas, PAGECACHE_TAG_DIRTY);
1948	dax_lock_entry(&xas, entry);
1949	xas_unlock_irq(&xas);
1950	if (order == 0)
1951		ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn);
1952#ifdef CONFIG_FS_DAX_PMD
1953	else if (order == PMD_ORDER)
1954		ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE);
1955#endif
1956	else
1957		ret = VM_FAULT_FALLBACK;
1958	dax_unlock_entry(&xas, entry);
1959	trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret);
1960	return ret;
1961}
1962
1963/**
1964 * dax_finish_sync_fault - finish synchronous page fault
1965 * @vmf: The description of the fault
1966 * @order: Order of entry to be inserted
1967 * @pfn: PFN to insert
1968 *
1969 * This function ensures that the file range touched by the page fault is
1970 * stored persistently on the media and handles inserting of appropriate page
1971 * table entry.
1972 */
1973vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, unsigned int order,
1974		pfn_t pfn)
1975{
1976	int err;
1977	loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT;
 
1978	size_t len = PAGE_SIZE << order;
1979
1980	err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1);
1981	if (err)
1982		return VM_FAULT_SIGBUS;
1983	return dax_insert_pfn_mkwrite(vmf, pfn, order);
1984}
1985EXPORT_SYMBOL_GPL(dax_finish_sync_fault);
1986
1987static loff_t dax_range_compare_iter(struct iomap_iter *it_src,
1988		struct iomap_iter *it_dest, u64 len, bool *same)
1989{
1990	const struct iomap *smap = &it_src->iomap;
1991	const struct iomap *dmap = &it_dest->iomap;
1992	loff_t pos1 = it_src->pos, pos2 = it_dest->pos;
1993	void *saddr, *daddr;
1994	int id, ret;
1995
1996	len = min(len, min(smap->length, dmap->length));
1997
1998	if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) {
1999		*same = true;
2000		return len;
2001	}
2002
2003	if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) {
2004		*same = false;
2005		return 0;
2006	}
2007
2008	id = dax_read_lock();
2009	ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE),
2010				      &saddr, NULL);
2011	if (ret < 0)
2012		goto out_unlock;
2013
2014	ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE),
2015				      &daddr, NULL);
2016	if (ret < 0)
2017		goto out_unlock;
2018
2019	*same = !memcmp(saddr, daddr, len);
2020	if (!*same)
2021		len = 0;
2022	dax_read_unlock(id);
2023	return len;
2024
2025out_unlock:
2026	dax_read_unlock(id);
2027	return -EIO;
2028}
2029
2030int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
2031		struct inode *dst, loff_t dstoff, loff_t len, bool *same,
2032		const struct iomap_ops *ops)
2033{
2034	struct iomap_iter src_iter = {
2035		.inode		= src,
2036		.pos		= srcoff,
2037		.len		= len,
2038		.flags		= IOMAP_DAX,
2039	};
2040	struct iomap_iter dst_iter = {
2041		.inode		= dst,
2042		.pos		= dstoff,
2043		.len		= len,
2044		.flags		= IOMAP_DAX,
2045	};
2046	int ret, compared = 0;
2047
2048	while ((ret = iomap_iter(&src_iter, ops)) > 0 &&
2049	       (ret = iomap_iter(&dst_iter, ops)) > 0) {
2050		compared = dax_range_compare_iter(&src_iter, &dst_iter,
2051				min(src_iter.len, dst_iter.len), same);
2052		if (compared < 0)
2053			return ret;
2054		src_iter.processed = dst_iter.processed = compared;
2055	}
2056	return ret;
2057}
2058
2059int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in,
2060			      struct file *file_out, loff_t pos_out,
2061			      loff_t *len, unsigned int remap_flags,
2062			      const struct iomap_ops *ops)
2063{
2064	return __generic_remap_file_range_prep(file_in, pos_in, file_out,
2065					       pos_out, len, remap_flags, ops);
2066}
2067EXPORT_SYMBOL_GPL(dax_remap_file_range_prep);