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);
v4.6
 
   1/*
   2 * fs/dax.c - Direct Access filesystem code
   3 * Copyright (c) 2013-2014 Intel Corporation
   4 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
   5 * Author: Ross Zwisler <ross.zwisler@linux.intel.com>
   6 *
   7 * This program is free software; you can redistribute it and/or modify it
   8 * under the terms and conditions of the GNU General Public License,
   9 * version 2, as published by the Free Software Foundation.
  10 *
  11 * This program is distributed in the hope it will be useful, but WITHOUT
  12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14 * more details.
  15 */
  16
  17#include <linux/atomic.h>
  18#include <linux/blkdev.h>
  19#include <linux/buffer_head.h>
  20#include <linux/dax.h>
  21#include <linux/fs.h>
  22#include <linux/genhd.h>
  23#include <linux/highmem.h>
  24#include <linux/memcontrol.h>
  25#include <linux/mm.h>
  26#include <linux/mutex.h>
  27#include <linux/pagevec.h>
  28#include <linux/pmem.h>
  29#include <linux/sched.h>
 
  30#include <linux/uio.h>
  31#include <linux/vmstat.h>
  32#include <linux/pfn_t.h>
  33#include <linux/sizes.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  34
  35static long dax_map_atomic(struct block_device *bdev, struct blk_dax_ctl *dax)
  36{
  37	struct request_queue *q = bdev->bd_queue;
  38	long rc = -EIO;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  39
  40	dax->addr = (void __pmem *) ERR_PTR(-EIO);
  41	if (blk_queue_enter(q, true) != 0)
  42		return rc;
 
  43
  44	rc = bdev_direct_access(bdev, dax);
  45	if (rc < 0) {
  46		dax->addr = (void __pmem *) ERR_PTR(rc);
  47		blk_queue_exit(q);
  48		return rc;
  49	}
  50	return rc;
  51}
  52
  53static void dax_unmap_atomic(struct block_device *bdev,
  54		const struct blk_dax_ctl *dax)
 
 
 
  55{
  56	if (IS_ERR(dax->addr))
  57		return;
  58	blk_queue_exit(bdev->bd_queue);
  59}
  60
  61struct page *read_dax_sector(struct block_device *bdev, sector_t n)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  62{
  63	struct page *page = alloc_pages(GFP_KERNEL, 0);
  64	struct blk_dax_ctl dax = {
  65		.size = PAGE_SIZE,
  66		.sector = n & ~((((int) PAGE_SIZE) / 512) - 1),
  67	};
  68	long rc;
 
 
 
 
 
 
 
 
 
 
  69
  70	if (!page)
  71		return ERR_PTR(-ENOMEM);
 
 
 
 
  72
  73	rc = dax_map_atomic(bdev, &dax);
  74	if (rc < 0)
  75		return ERR_PTR(rc);
  76	memcpy_from_pmem(page_address(page), dax.addr, PAGE_SIZE);
  77	dax_unmap_atomic(bdev, &dax);
  78	return page;
  79}
  80
  81/*
  82 * dax_clear_sectors() is called from within transaction context from XFS,
  83 * and hence this means the stack from this point must follow GFP_NOFS
  84 * semantics for all operations.
  85 */
  86int dax_clear_sectors(struct block_device *bdev, sector_t _sector, long _size)
  87{
  88	struct blk_dax_ctl dax = {
  89		.sector = _sector,
  90		.size = _size,
  91	};
 
 
 
 
 
 
 
 
 
 
 
  92
  93	might_sleep();
  94	do {
  95		long count, sz;
 
 
 
 
 
 
 
 
 
 
 
 
  96
  97		count = dax_map_atomic(bdev, &dax);
  98		if (count < 0)
  99			return count;
 100		sz = min_t(long, count, SZ_128K);
 101		clear_pmem(dax.addr, sz);
 102		dax.size -= sz;
 103		dax.sector += sz / 512;
 104		dax_unmap_atomic(bdev, &dax);
 105		cond_resched();
 106	} while (dax.size);
 107
 108	wmb_pmem();
 109	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 110}
 111EXPORT_SYMBOL_GPL(dax_clear_sectors);
 112
 113/* the clear_pmem() calls are ordered by a wmb_pmem() in the caller */
 114static void dax_new_buf(void __pmem *addr, unsigned size, unsigned first,
 115		loff_t pos, loff_t end)
 
 
 
 116{
 117	loff_t final = end - pos + first; /* The final byte of the buffer */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 118
 119	if (first > 0)
 120		clear_pmem(addr, first);
 121	if (final < size)
 122		clear_pmem(addr + final, size - final);
 
 123}
 124
 125static bool buffer_written(struct buffer_head *bh)
 
 
 
 
 
 126{
 127	return buffer_mapped(bh) && !buffer_unwritten(bh);
 
 
 
 
 
 
 
 
 128}
 129
 130/*
 131 * When ext4 encounters a hole, it returns without modifying the buffer_head
 132 * which means that we can't trust b_size.  To cope with this, we set b_state
 133 * to 0 before calling get_block and, if any bit is set, we know we can trust
 134 * b_size.  Unfortunate, really, since ext4 knows precisely how long a hole is
 135 * and would save us time calling get_block repeatedly.
 136 */
 137static bool buffer_size_valid(struct buffer_head *bh)
 
 
 
 
 
 
 138{
 139	return bh->b_state != 0;
 
 
 
 
 
 
 
 
 
 
 
 
 140}
 141
 
 
 
 
 
 
 
 
 
 
 
 
 142
 143static sector_t to_sector(const struct buffer_head *bh,
 144		const struct inode *inode)
 
 
 
 145{
 146	sector_t sector = bh->b_blocknr << (inode->i_blkbits - 9);
 
 
 
 
 
 
 
 
 
 
 147
 148	return sector;
 
 
 149}
 150
 151static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 152		      loff_t start, loff_t end, get_block_t get_block,
 153		      struct buffer_head *bh)
 
 
 
 
 154{
 155	loff_t pos = start, max = start, bh_max = start;
 156	bool hole = false, need_wmb = false;
 157	struct block_device *bdev = NULL;
 158	int rw = iov_iter_rw(iter), rc;
 159	long map_len = 0;
 160	struct blk_dax_ctl dax = {
 161		.addr = (void __pmem *) ERR_PTR(-EIO),
 162	};
 163
 164	if (rw == READ)
 165		end = min(end, i_size_read(inode));
 166
 167	while (pos < end) {
 168		size_t len;
 169		if (pos == max) {
 170			unsigned blkbits = inode->i_blkbits;
 171			long page = pos >> PAGE_SHIFT;
 172			sector_t block = page << (PAGE_SHIFT - blkbits);
 173			unsigned first = pos - (block << blkbits);
 174			long size;
 175
 176			if (pos == bh_max) {
 177				bh->b_size = PAGE_ALIGN(end - pos);
 178				bh->b_state = 0;
 179				rc = get_block(inode, block, bh, rw == WRITE);
 180				if (rc)
 181					break;
 182				if (!buffer_size_valid(bh))
 183					bh->b_size = 1 << blkbits;
 184				bh_max = pos - first + bh->b_size;
 185				bdev = bh->b_bdev;
 186			} else {
 187				unsigned done = bh->b_size -
 188						(bh_max - (pos - first));
 189				bh->b_blocknr += done >> blkbits;
 190				bh->b_size -= done;
 191			}
 192
 193			hole = rw == READ && !buffer_written(bh);
 194			if (hole) {
 195				size = bh->b_size - first;
 196			} else {
 197				dax_unmap_atomic(bdev, &dax);
 198				dax.sector = to_sector(bh, inode);
 199				dax.size = bh->b_size;
 200				map_len = dax_map_atomic(bdev, &dax);
 201				if (map_len < 0) {
 202					rc = map_len;
 203					break;
 204				}
 205				if (buffer_unwritten(bh) || buffer_new(bh)) {
 206					dax_new_buf(dax.addr, map_len, first,
 207							pos, end);
 208					need_wmb = true;
 209				}
 210				dax.addr += first;
 211				size = map_len - first;
 212			}
 213			max = min(pos + size, end);
 214		}
 
 
 215
 216		if (iov_iter_rw(iter) == WRITE) {
 217			len = copy_from_iter_pmem(dax.addr, max - pos, iter);
 218			need_wmb = true;
 219		} else if (!hole)
 220			len = copy_to_iter((void __force *) dax.addr, max - pos,
 221					iter);
 222		else
 223			len = iov_iter_zero(max - pos, iter);
 224
 225		if (!len) {
 226			rc = -EFAULT;
 227			break;
 228		}
 229
 230		pos += len;
 231		if (!IS_ERR(dax.addr))
 232			dax.addr += len;
 
 
 
 
 
 
 233	}
 
 
 
 
 
 234
 235	if (need_wmb)
 236		wmb_pmem();
 237	dax_unmap_atomic(bdev, &dax);
 238
 239	return (pos == start) ? rc : pos - start;
 
 
 
 240}
 241
 242/**
 243 * dax_do_io - Perform I/O to a DAX file
 244 * @iocb: The control block for this I/O
 245 * @inode: The file which the I/O is directed at
 246 * @iter: The addresses to do I/O from or to
 247 * @pos: The file offset where the I/O starts
 248 * @get_block: The filesystem method used to translate file offsets to blocks
 249 * @end_io: A filesystem callback for I/O completion
 250 * @flags: See below
 251 *
 252 * This function uses the same locking scheme as do_blockdev_direct_IO:
 253 * If @flags has DIO_LOCKING set, we assume that the i_mutex is held by the
 254 * caller for writes.  For reads, we take and release the i_mutex ourselves.
 255 * If DIO_LOCKING is not set, the filesystem takes care of its own locking.
 256 * As with do_blockdev_direct_IO(), we increment i_dio_count while the I/O
 257 * is in progress.
 258 */
 259ssize_t dax_do_io(struct kiocb *iocb, struct inode *inode,
 260		  struct iov_iter *iter, loff_t pos, get_block_t get_block,
 261		  dio_iodone_t end_io, int flags)
 262{
 263	struct buffer_head bh;
 264	ssize_t retval = -EINVAL;
 265	loff_t end = pos + iov_iter_count(iter);
 266
 267	memset(&bh, 0, sizeof(bh));
 268	bh.b_bdev = inode->i_sb->s_bdev;
 269
 270	if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ) {
 271		struct address_space *mapping = inode->i_mapping;
 272		inode_lock(inode);
 273		retval = filemap_write_and_wait_range(mapping, pos, end - 1);
 274		if (retval) {
 275			inode_unlock(inode);
 276			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 277		}
 
 
 
 278	}
 
 
 
 279
 280	/* Protects against truncate */
 281	if (!(flags & DIO_SKIP_DIO_COUNT))
 282		inode_dio_begin(inode);
 
 283
 284	retval = dax_io(inode, iter, pos, end, get_block, &bh);
 
 285
 286	if ((flags & DIO_LOCKING) && iov_iter_rw(iter) == READ)
 287		inode_unlock(inode);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 288
 289	if (end_io) {
 290		int err;
 
 
 
 291
 292		err = end_io(iocb, pos, retval, bh.b_private);
 293		if (err)
 294			retval = err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 295	}
 
 
 
 
 
 
 
 
 
 
 
 296
 297	if (!(flags & DIO_SKIP_DIO_COUNT))
 298		inode_dio_end(inode);
 299 out:
 300	return retval;
 301}
 302EXPORT_SYMBOL_GPL(dax_do_io);
 303
 304/*
 305 * The user has performed a load from a hole in the file.  Allocating
 306 * a new page in the file would cause excessive storage usage for
 307 * workloads with sparse files.  We allocate a page cache page instead.
 308 * We'll kick it out of the page cache if it's ever written to,
 309 * otherwise it will simply fall out of the page cache under memory
 310 * pressure without ever having been dirtied.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 311 */
 312static int dax_load_hole(struct address_space *mapping, struct page *page,
 313							struct vm_fault *vmf)
 314{
 315	unsigned long size;
 316	struct inode *inode = mapping->host;
 317	if (!page)
 318		page = find_or_create_page(mapping, vmf->pgoff,
 319						GFP_KERNEL | __GFP_ZERO);
 320	if (!page)
 321		return VM_FAULT_OOM;
 322	/* Recheck i_size under page lock to avoid truncate race */
 323	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 324	if (vmf->pgoff >= size) {
 325		unlock_page(page);
 326		put_page(page);
 327		return VM_FAULT_SIGBUS;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 328	}
 329
 330	vmf->page = page;
 331	return VM_FAULT_LOCKED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 332}
 333
 334static int copy_user_bh(struct page *to, struct inode *inode,
 335		struct buffer_head *bh, unsigned long vaddr)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 336{
 337	struct blk_dax_ctl dax = {
 338		.sector = to_sector(bh, inode),
 339		.size = bh->b_size,
 340	};
 341	struct block_device *bdev = bh->b_bdev;
 342	void *vto;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 343
 344	if (dax_map_atomic(bdev, &dax) < 0)
 345		return PTR_ERR(dax.addr);
 346	vto = kmap_atomic(to);
 347	copy_user_page(vto, (void __force *)dax.addr, vaddr, to);
 348	kunmap_atomic(vto);
 349	dax_unmap_atomic(bdev, &dax);
 350	return 0;
 351}
 
 352
 353#define NO_SECTOR -1
 354#define DAX_PMD_INDEX(page_index) (page_index & (PMD_MASK >> PAGE_SHIFT))
 
 
 
 355
 356static int dax_radix_entry(struct address_space *mapping, pgoff_t index,
 357		sector_t sector, bool pmd_entry, bool dirty)
 358{
 359	struct radix_tree_root *page_tree = &mapping->page_tree;
 360	pgoff_t pmd_index = DAX_PMD_INDEX(index);
 361	int type, error = 0;
 362	void *entry;
 363
 364	WARN_ON_ONCE(pmd_entry && !dirty);
 365	if (dirty)
 366		__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 367
 368	spin_lock_irq(&mapping->tree_lock);
 
 
 
 
 
 369
 370	entry = radix_tree_lookup(page_tree, pmd_index);
 371	if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD) {
 372		index = pmd_index;
 373		goto dirty;
 
 
 374	}
 
 
 
 
 
 
 375
 376	entry = radix_tree_lookup(page_tree, index);
 377	if (entry) {
 378		type = RADIX_DAX_TYPE(entry);
 379		if (WARN_ON_ONCE(type != RADIX_DAX_PTE &&
 380					type != RADIX_DAX_PMD)) {
 381			error = -EIO;
 382			goto unlock;
 383		}
 
 
 384
 385		if (!pmd_entry || type == RADIX_DAX_PMD)
 386			goto dirty;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 387
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 388		/*
 389		 * We only insert dirty PMD entries into the radix tree.  This
 390		 * means we don't need to worry about removing a dirty PTE
 391		 * entry and inserting a clean PMD entry, thus reducing the
 392		 * range we would flush with a follow-up fsync/msync call.
 
 
 393		 */
 394		radix_tree_delete(&mapping->page_tree, index);
 395		mapping->nrexceptional--;
 
 
 
 
 396	}
 397
 398	if (sector == NO_SECTOR) {
 399		/*
 400		 * This can happen during correct operation if our pfn_mkwrite
 401		 * fault raced against a hole punch operation.  If this
 402		 * happens the pte that was hole punched will have been
 403		 * unmapped and the radix tree entry will have been removed by
 404		 * the time we are called, but the call will still happen.  We
 405		 * will return all the way up to wp_pfn_shared(), where the
 406		 * pte_same() check will fail, eventually causing page fault
 407		 * to be retried by the CPU.
 408		 */
 409		goto unlock;
 410	}
 411
 412	error = radix_tree_insert(page_tree, index,
 413			RADIX_DAX_ENTRY(sector, pmd_entry));
 414	if (error)
 415		goto unlock;
 416
 417	mapping->nrexceptional++;
 418 dirty:
 419	if (dirty)
 420		radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
 421 unlock:
 422	spin_unlock_irq(&mapping->tree_lock);
 423	return error;
 424}
 425
 426static int dax_writeback_one(struct block_device *bdev,
 427		struct address_space *mapping, pgoff_t index, void *entry)
 428{
 429	struct radix_tree_root *page_tree = &mapping->page_tree;
 430	int type = RADIX_DAX_TYPE(entry);
 431	struct radix_tree_node *node;
 432	struct blk_dax_ctl dax;
 433	void **slot;
 434	int ret = 0;
 435
 436	spin_lock_irq(&mapping->tree_lock);
 437	/*
 438	 * Regular page slots are stabilized by the page lock even
 439	 * without the tree itself locked.  These unlocked entries
 440	 * need verification under the tree lock.
 441	 */
 442	if (!__radix_tree_lookup(page_tree, index, &node, &slot))
 443		goto unlock;
 444	if (*slot != entry)
 445		goto unlock;
 446
 447	/* another fsync thread may have already written back this entry */
 448	if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
 449		goto unlock;
 450
 451	if (WARN_ON_ONCE(type != RADIX_DAX_PTE && type != RADIX_DAX_PMD)) {
 452		ret = -EIO;
 453		goto unlock;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 454	}
 455
 456	dax.sector = RADIX_DAX_SECTOR(entry);
 457	dax.size = (type == RADIX_DAX_PMD ? PMD_SIZE : PAGE_SIZE);
 458	spin_unlock_irq(&mapping->tree_lock);
 
 
 
 
 
 
 
 
 
 459
 460	/*
 461	 * We cannot hold tree_lock while calling dax_map_atomic() because it
 462	 * eventually calls cond_resched().
 
 
 
 463	 */
 464	ret = dax_map_atomic(bdev, &dax);
 465	if (ret < 0)
 466		return ret;
 
 467
 468	if (WARN_ON_ONCE(ret < dax.size)) {
 469		ret = -EIO;
 470		goto unmap;
 
 
 471	}
 
 472
 473	wb_cache_pmem(dax.addr, dax.size);
 
 
 
 
 
 
 
 
 
 
 
 474
 475	spin_lock_irq(&mapping->tree_lock);
 476	radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
 477	spin_unlock_irq(&mapping->tree_lock);
 478 unmap:
 479	dax_unmap_atomic(bdev, &dax);
 480	return ret;
 481
 482 unlock:
 483	spin_unlock_irq(&mapping->tree_lock);
 484	return ret;
 485}
 486
 487/*
 488 * Flush the mapping to the persistent domain within the byte range of [start,
 489 * end]. This is required by data integrity operations to ensure file data is
 490 * on persistent storage prior to completion of the operation.
 491 */
 492int dax_writeback_mapping_range(struct address_space *mapping,
 493		struct block_device *bdev, struct writeback_control *wbc)
 494{
 
 495	struct inode *inode = mapping->host;
 496	pgoff_t start_index, end_index, pmd_index;
 497	pgoff_t indices[PAGEVEC_SIZE];
 498	struct pagevec pvec;
 499	bool done = false;
 500	int i, ret = 0;
 501	void *entry;
 
 
 502
 503	if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
 504		return -EIO;
 505
 506	if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
 507		return 0;
 508
 509	start_index = wbc->range_start >> PAGE_SHIFT;
 510	end_index = wbc->range_end >> PAGE_SHIFT;
 511	pmd_index = DAX_PMD_INDEX(start_index);
 512
 513	rcu_read_lock();
 514	entry = radix_tree_lookup(&mapping->page_tree, pmd_index);
 515	rcu_read_unlock();
 516
 517	/* see if the start of our range is covered by a PMD entry */
 518	if (entry && RADIX_DAX_TYPE(entry) == RADIX_DAX_PMD)
 519		start_index = pmd_index;
 520
 521	tag_pages_for_writeback(mapping, start_index, end_index);
 522
 523	pagevec_init(&pvec, 0);
 524	while (!done) {
 525		pvec.nr = find_get_entries_tag(mapping, start_index,
 526				PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
 527				pvec.pages, indices);
 528
 529		if (pvec.nr == 0)
 530			break;
 
 
 
 531
 532		for (i = 0; i < pvec.nr; i++) {
 533			if (indices[i] > end_index) {
 534				done = true;
 535				break;
 536			}
 537
 538			ret = dax_writeback_one(bdev, mapping, indices[i],
 539					pvec.pages[i]);
 540			if (ret < 0)
 541				return ret;
 542		}
 543	}
 544	wmb_pmem();
 545	return 0;
 
 546}
 547EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
 548
 549static int dax_insert_mapping(struct inode *inode, struct buffer_head *bh,
 550			struct vm_area_struct *vma, struct vm_fault *vmf)
 551{
 552	unsigned long vaddr = (unsigned long)vmf->virtual_address;
 553	struct address_space *mapping = inode->i_mapping;
 554	struct block_device *bdev = bh->b_bdev;
 555	struct blk_dax_ctl dax = {
 556		.sector = to_sector(bh, inode),
 557		.size = bh->b_size,
 558	};
 559	pgoff_t size;
 560	int error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 561
 562	i_mmap_lock_read(mapping);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 563
 564	/*
 565	 * Check truncate didn't happen while we were allocating a block.
 566	 * If it did, this block may or may not be still allocated to the
 567	 * file.  We can't tell the filesystem to free it because we can't
 568	 * take i_mutex here.  In the worst case, the file still has blocks
 569	 * allocated past the end of the file.
 570	 */
 571	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 572	if (unlikely(vmf->pgoff >= size)) {
 573		error = -EIO;
 574		goto out;
 575	}
 576
 577	if (dax_map_atomic(bdev, &dax) < 0) {
 578		error = PTR_ERR(dax.addr);
 
 
 
 579		goto out;
 580	}
 581
 582	if (buffer_unwritten(bh) || buffer_new(bh)) {
 583		clear_pmem(dax.addr, PAGE_SIZE);
 584		wmb_pmem();
 
 
 
 
 
 
 585	}
 586	dax_unmap_atomic(bdev, &dax);
 587
 588	error = dax_radix_entry(mapping, vmf->pgoff, dax.sector, false,
 589			vmf->flags & FAULT_FLAG_WRITE);
 590	if (error)
 591		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 592
 593	error = vm_insert_mixed(vma, vaddr, dax.pfn);
 
 
 
 
 
 
 
 
 
 
 
 
 
 594
 595 out:
 596	i_mmap_unlock_read(mapping);
 597
 598	return error;
 
 
 599}
 600
 601/**
 602 * __dax_fault - handle a page fault on a DAX file
 603 * @vma: The virtual memory area where the fault occurred
 604 * @vmf: The description of the fault
 605 * @get_block: The filesystem method used to translate file offsets to blocks
 606 * @complete_unwritten: The filesystem method used to convert unwritten blocks
 607 *	to written so the data written to them is exposed. This is required for
 608 *	required by write faults for filesystems that will return unwritten
 609 *	extent mappings from @get_block, but it is optional for reads as
 610 *	dax_insert_mapping() will always zero unwritten blocks. If the fs does
 611 *	not support unwritten extents, the it should pass NULL.
 612 *
 613 * When a page fault occurs, filesystems may call this helper in their
 614 * fault handler for DAX files. __dax_fault() assumes the caller has done all
 615 * the necessary locking for the page fault to proceed successfully.
 616 */
 617int __dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 618			get_block_t get_block, dax_iodone_t complete_unwritten)
 619{
 620	struct file *file = vma->vm_file;
 621	struct address_space *mapping = file->f_mapping;
 622	struct inode *inode = mapping->host;
 623	struct page *page;
 624	struct buffer_head bh;
 625	unsigned long vaddr = (unsigned long)vmf->virtual_address;
 626	unsigned blkbits = inode->i_blkbits;
 627	sector_t block;
 628	pgoff_t size;
 629	int error;
 630	int major = 0;
 
 
 631
 632	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 633	if (vmf->pgoff >= size)
 634		return VM_FAULT_SIGBUS;
 
 
 
 
 
 
 635
 636	memset(&bh, 0, sizeof(bh));
 637	block = (sector_t)vmf->pgoff << (PAGE_SHIFT - blkbits);
 638	bh.b_bdev = inode->i_sb->s_bdev;
 639	bh.b_size = PAGE_SIZE;
 640
 641 repeat:
 642	page = find_get_page(mapping, vmf->pgoff);
 643	if (page) {
 644		if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
 645			put_page(page);
 646			return VM_FAULT_RETRY;
 647		}
 648		if (unlikely(page->mapping != mapping)) {
 649			unlock_page(page);
 650			put_page(page);
 651			goto repeat;
 652		}
 653		size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 654		if (unlikely(vmf->pgoff >= size)) {
 655			/*
 656			 * We have a struct page covering a hole in the file
 657			 * from a read fault and we've raced with a truncate
 658			 */
 659			error = -EIO;
 660			goto unlock_page;
 661		}
 662	}
 663
 664	error = get_block(inode, block, &bh, 0);
 665	if (!error && (bh.b_size < PAGE_SIZE))
 666		error = -EIO;		/* fs corruption? */
 667	if (error)
 668		goto unlock_page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 669
 670	if (!buffer_mapped(&bh) && !buffer_unwritten(&bh) && !vmf->cow_page) {
 671		if (vmf->flags & FAULT_FLAG_WRITE) {
 672			error = get_block(inode, block, &bh, 1);
 673			count_vm_event(PGMAJFAULT);
 674			mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 675			major = VM_FAULT_MAJOR;
 676			if (!error && (bh.b_size < PAGE_SIZE))
 677				error = -EIO;
 678			if (error)
 679				goto unlock_page;
 680		} else {
 681			return dax_load_hole(mapping, page, vmf);
 682		}
 683	}
 684
 685	if (vmf->cow_page) {
 686		struct page *new_page = vmf->cow_page;
 687		if (buffer_written(&bh))
 688			error = copy_user_bh(new_page, inode, &bh, vaddr);
 
 
 
 
 
 
 689		else
 690			clear_user_highpage(new_page, vaddr);
 691		if (error)
 692			goto unlock_page;
 693		vmf->page = page;
 694		if (!page) {
 695			i_mmap_lock_read(mapping);
 696			/* Check we didn't race with truncate */
 697			size = (i_size_read(inode) + PAGE_SIZE - 1) >>
 698								PAGE_SHIFT;
 699			if (vmf->pgoff >= size) {
 700				i_mmap_unlock_read(mapping);
 701				error = -EIO;
 702				goto out;
 703			}
 704		}
 705		return VM_FAULT_LOCKED;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 706	}
 707
 708	/* Check we didn't race with a read fault installing a new page */
 709	if (!page && major)
 710		page = find_lock_page(mapping, vmf->pgoff);
 711
 712	if (page) {
 713		unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
 714							PAGE_SIZE, 0);
 715		delete_from_page_cache(page);
 716		unlock_page(page);
 717		put_page(page);
 718		page = NULL;
 719	}
 720
 721	/*
 722	 * If we successfully insert the new mapping over an unwritten extent,
 723	 * we need to ensure we convert the unwritten extent. If there is an
 724	 * error inserting the mapping, the filesystem needs to leave it as
 725	 * unwritten to prevent exposure of the stale underlying data to
 726	 * userspace, but we still need to call the completion function so
 727	 * the private resources on the mapping buffer can be released. We
 728	 * indicate what the callback should do via the uptodate variable, same
 729	 * as for normal BH based IO completions.
 730	 */
 731	error = dax_insert_mapping(inode, &bh, vma, vmf);
 732	if (buffer_unwritten(&bh)) {
 733		if (complete_unwritten)
 734			complete_unwritten(&bh, !error);
 735		else
 736			WARN_ON_ONCE(!(vmf->flags & FAULT_FLAG_WRITE));
 737	}
 738
 739 out:
 740	if (error == -ENOMEM)
 741		return VM_FAULT_OOM | major;
 742	/* -EBUSY is fine, somebody else faulted on the same PTE */
 743	if ((error < 0) && (error != -EBUSY))
 744		return VM_FAULT_SIGBUS | major;
 745	return VM_FAULT_NOPAGE | major;
 
 746
 747 unlock_page:
 748	if (page) {
 749		unlock_page(page);
 750		put_page(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 751	}
 752	goto out;
 
 
 753}
 754EXPORT_SYMBOL(__dax_fault);
 755
 756/**
 757 * dax_fault - handle a page fault on a DAX file
 758 * @vma: The virtual memory area where the fault occurred
 759 * @vmf: The description of the fault
 760 * @get_block: The filesystem method used to translate file offsets to blocks
 761 *
 762 * When a page fault occurs, filesystems may call this helper in their
 763 * fault handler for DAX files.
 
 764 */
 765int dax_fault(struct vm_area_struct *vma, struct vm_fault *vmf,
 766	      get_block_t get_block, dax_iodone_t complete_unwritten)
 767{
 768	int result;
 769	struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 
 
 
 
 
 
 
 
 
 
 770
 771	if (vmf->flags & FAULT_FLAG_WRITE) {
 772		sb_start_pagefault(sb);
 773		file_update_time(vma->vm_file);
 
 
 774	}
 775	result = __dax_fault(vma, vmf, get_block, complete_unwritten);
 776	if (vmf->flags & FAULT_FLAG_WRITE)
 777		sb_end_pagefault(sb);
 778
 779	return result;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 780}
 781EXPORT_SYMBOL_GPL(dax_fault);
 782
 783#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 784/*
 785 * The 'colour' (ie low bits) within a PMD of a page offset.  This comes up
 786 * more often than one might expect in the below function.
 
 
 787 */
 788#define PG_PMD_COLOUR	((PMD_SIZE >> PAGE_SHIFT) - 1)
 
 
 
 
 
 
 789
 790static void __dax_dbg(struct buffer_head *bh, unsigned long address,
 791		const char *reason, const char *fn)
 792{
 793	if (bh) {
 794		char bname[BDEVNAME_SIZE];
 795		bdevname(bh->b_bdev, bname);
 796		pr_debug("%s: %s addr: %lx dev %s state %lx start %lld "
 797			"length %zd fallback: %s\n", fn, current->comm,
 798			address, bname, bh->b_state, (u64)bh->b_blocknr,
 799			bh->b_size, reason);
 800	} else {
 801		pr_debug("%s: %s addr: %lx fallback: %s\n", fn,
 802			current->comm, address, reason);
 
 
 
 
 
 803	}
 804}
 805
 806#define dax_pmd_dbg(bh, address, reason)	__dax_dbg(bh, address, reason, "dax_pmd")
 
 807
 808int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
 809		pmd_t *pmd, unsigned int flags, get_block_t get_block,
 810		dax_iodone_t complete_unwritten)
 811{
 812	struct file *file = vma->vm_file;
 813	struct address_space *mapping = file->f_mapping;
 814	struct inode *inode = mapping->host;
 815	struct buffer_head bh;
 816	unsigned blkbits = inode->i_blkbits;
 817	unsigned long pmd_addr = address & PMD_MASK;
 818	bool write = flags & FAULT_FLAG_WRITE;
 819	struct block_device *bdev;
 820	pgoff_t size, pgoff;
 821	sector_t block;
 822	int error, result = 0;
 823	bool alloc = false;
 824
 825	/* dax pmd mappings require pfn_t_devmap() */
 826	if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
 827		return VM_FAULT_FALLBACK;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 828
 829	/* Fall back to PTEs if we're going to COW */
 830	if (write && !(vma->vm_flags & VM_SHARED)) {
 831		split_huge_pmd(vma, pmd, address);
 832		dax_pmd_dbg(NULL, address, "cow write");
 833		return VM_FAULT_FALLBACK;
 834	}
 835	/* If the PMD would extend outside the VMA */
 836	if (pmd_addr < vma->vm_start) {
 837		dax_pmd_dbg(NULL, address, "vma start unaligned");
 838		return VM_FAULT_FALLBACK;
 839	}
 840	if ((pmd_addr + PMD_SIZE) > vma->vm_end) {
 841		dax_pmd_dbg(NULL, address, "vma end unaligned");
 842		return VM_FAULT_FALLBACK;
 843	}
 844
 845	pgoff = linear_page_index(vma, pmd_addr);
 846	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 847	if (pgoff >= size)
 848		return VM_FAULT_SIGBUS;
 849	/* If the PMD would cover blocks out of the file */
 850	if ((pgoff | PG_PMD_COLOUR) >= size) {
 851		dax_pmd_dbg(NULL, address,
 852				"offset + huge page size > file size");
 853		return VM_FAULT_FALLBACK;
 854	}
 855
 856	memset(&bh, 0, sizeof(bh));
 857	bh.b_bdev = inode->i_sb->s_bdev;
 858	block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
 859
 860	bh.b_size = PMD_SIZE;
 
 
 861
 862	if (get_block(inode, block, &bh, 0) != 0)
 863		return VM_FAULT_SIGBUS;
 
 
 
 864
 865	if (!buffer_mapped(&bh) && write) {
 866		if (get_block(inode, block, &bh, 1) != 0)
 867			return VM_FAULT_SIGBUS;
 868		alloc = true;
 869	}
 870
 871	bdev = bh.b_bdev;
 
 
 
 
 
 
 
 872
 
 873	/*
 874	 * If the filesystem isn't willing to tell us the length of a hole,
 875	 * just fall back to PTEs.  Calling get_block 512 times in a loop
 876	 * would be silly.
 877	 */
 878	if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
 879		dax_pmd_dbg(&bh, address, "allocated block too small");
 880		return VM_FAULT_FALLBACK;
 881	}
 882
 883	/*
 884	 * If we allocated new storage, make sure no process has any
 885	 * zero pages covering this hole
 886	 */
 887	if (alloc) {
 888		loff_t lstart = pgoff << PAGE_SHIFT;
 889		loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
 890
 891		truncate_pagecache_range(inode, lstart, lend);
 
 
 
 892	}
 893
 894	i_mmap_lock_read(mapping);
 895
 896	/*
 897	 * If a truncate happened while we were allocating blocks, we may
 898	 * leave blocks allocated to the file that are beyond EOF.  We can't
 899	 * take i_mutex here, so just leave them hanging; they'll be freed
 900	 * when the file is deleted.
 901	 */
 902	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
 903	if (pgoff >= size) {
 904		result = VM_FAULT_SIGBUS;
 905		goto out;
 906	}
 907	if ((pgoff | PG_PMD_COLOUR) >= size) {
 908		dax_pmd_dbg(&bh, address,
 909				"offset + huge page size > file size");
 910		goto fallback;
 
 
 
 
 
 
 
 
 
 
 
 
 
 911	}
 912
 913	if (!write && !buffer_mapped(&bh) && buffer_uptodate(&bh)) {
 914		spinlock_t *ptl;
 915		pmd_t entry;
 916		struct page *zero_page = get_huge_zero_page();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 917
 918		if (unlikely(!zero_page)) {
 919			dax_pmd_dbg(&bh, address, "no zero page");
 920			goto fallback;
 921		}
 
 
 
 
 
 922
 923		ptl = pmd_lock(vma->vm_mm, pmd);
 924		if (!pmd_none(*pmd)) {
 925			spin_unlock(ptl);
 926			dax_pmd_dbg(&bh, address, "pmd already present");
 927			goto fallback;
 928		}
 929
 930		dev_dbg(part_to_dev(bdev->bd_part),
 931				"%s: %s addr: %lx pfn: <zero> sect: %llx\n",
 932				__func__, current->comm, address,
 933				(unsigned long long) to_sector(&bh, inode));
 934
 935		entry = mk_pmd(zero_page, vma->vm_page_prot);
 936		entry = pmd_mkhuge(entry);
 937		set_pmd_at(vma->vm_mm, pmd_addr, pmd, entry);
 938		result = VM_FAULT_NOPAGE;
 939		spin_unlock(ptl);
 940	} else {
 941		struct blk_dax_ctl dax = {
 942			.sector = to_sector(&bh, inode),
 943			.size = PMD_SIZE,
 944		};
 945		long length = dax_map_atomic(bdev, &dax);
 946
 947		if (length < 0) {
 948			result = VM_FAULT_SIGBUS;
 949			goto out;
 950		}
 951		if (length < PMD_SIZE) {
 952			dax_pmd_dbg(&bh, address, "dax-length too small");
 953			dax_unmap_atomic(bdev, &dax);
 954			goto fallback;
 955		}
 956		if (pfn_t_to_pfn(dax.pfn) & PG_PMD_COLOUR) {
 957			dax_pmd_dbg(&bh, address, "pfn unaligned");
 958			dax_unmap_atomic(bdev, &dax);
 959			goto fallback;
 960		}
 961
 962		if (!pfn_t_devmap(dax.pfn)) {
 963			dax_unmap_atomic(bdev, &dax);
 964			dax_pmd_dbg(&bh, address, "pfn not in memmap");
 965			goto fallback;
 966		}
 967
 968		if (buffer_unwritten(&bh) || buffer_new(&bh)) {
 969			clear_pmem(dax.addr, PMD_SIZE);
 970			wmb_pmem();
 971			count_vm_event(PGMAJFAULT);
 972			mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
 973			result |= VM_FAULT_MAJOR;
 974		}
 975		dax_unmap_atomic(bdev, &dax);
 976
 977		/*
 978		 * For PTE faults we insert a radix tree entry for reads, and
 979		 * leave it clean.  Then on the first write we dirty the radix
 980		 * tree entry via the dax_pfn_mkwrite() path.  This sequence
 981		 * allows the dax_pfn_mkwrite() call to be simpler and avoid a
 982		 * call into get_block() to translate the pgoff to a sector in
 983		 * order to be able to create a new radix tree entry.
 984		 *
 985		 * The PMD path doesn't have an equivalent to
 986		 * dax_pfn_mkwrite(), though, so for a read followed by a
 987		 * write we traverse all the way through __dax_pmd_fault()
 988		 * twice.  This means we can just skip inserting a radix tree
 989		 * entry completely on the initial read and just wait until
 990		 * the write to insert a dirty entry.
 991		 */
 992		if (write) {
 993			error = dax_radix_entry(mapping, pgoff, dax.sector,
 994					true, true);
 995			if (error) {
 996				dax_pmd_dbg(&bh, address,
 997						"PMD radix insertion failed");
 998				goto fallback;
 999			}
1000		}
1001
1002		dev_dbg(part_to_dev(bdev->bd_part),
1003				"%s: %s addr: %lx pfn: %lx sect: %llx\n",
1004				__func__, current->comm, address,
1005				pfn_t_to_pfn(dax.pfn),
1006				(unsigned long long) dax.sector);
1007		result |= vmf_insert_pfn_pmd(vma, address, pmd,
1008				dax.pfn, write);
1009	}
1010
1011 out:
1012	i_mmap_unlock_read(mapping);
1013
1014	if (buffer_unwritten(&bh))
1015		complete_unwritten(&bh, !(result & VM_FAULT_ERROR));
 
 
 
 
 
 
 
 
 
1016
1017	return result;
1018
1019 fallback:
1020	count_vm_event(THP_FAULT_FALLBACK);
1021	result = VM_FAULT_FALLBACK;
1022	goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1023}
1024EXPORT_SYMBOL_GPL(__dax_pmd_fault);
1025
1026/**
1027 * dax_pmd_fault - handle a PMD fault on a DAX file
1028 * @vma: The virtual memory area where the fault occurred
1029 * @vmf: The description of the fault
1030 * @get_block: The filesystem method used to translate file offsets to blocks
 
 
 
1031 *
1032 * When a page fault occurs, filesystems may call this helper in their
1033 * pmd_fault handler for DAX files.
 
 
1034 */
1035int dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
1036			pmd_t *pmd, unsigned int flags, get_block_t get_block,
1037			dax_iodone_t complete_unwritten)
1038{
1039	int result;
1040	struct super_block *sb = file_inode(vma->vm_file)->i_sb;
 
 
 
 
 
 
 
 
1041
1042	if (flags & FAULT_FLAG_WRITE) {
1043		sb_start_pagefault(sb);
1044		file_update_time(vma->vm_file);
1045	}
1046	result = __dax_pmd_fault(vma, address, pmd, flags, get_block,
1047				complete_unwritten);
1048	if (flags & FAULT_FLAG_WRITE)
1049		sb_end_pagefault(sb);
 
 
 
 
 
 
 
 
1050
1051	return result;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1052}
1053EXPORT_SYMBOL_GPL(dax_pmd_fault);
1054#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1055
1056/**
1057 * dax_pfn_mkwrite - handle first write to DAX page
1058 * @vma: The virtual memory area where the fault occurred
1059 * @vmf: The description of the fault
 
 
 
 
 
 
1060 */
1061int dax_pfn_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
 
1062{
1063	struct file *file = vma->vm_file;
1064	int error;
 
 
1065
1066	/*
1067	 * We pass NO_SECTOR to dax_radix_entry() because we expect that a
1068	 * RADIX_DAX_PTE entry already exists in the radix tree from a
1069	 * previous call to __dax_fault().  We just want to look up that PTE
1070	 * entry using vmf->pgoff and make sure the dirty tag is set.  This
1071	 * saves us from having to make a call to get_block() here to look
1072	 * up the sector.
1073	 */
1074	error = dax_radix_entry(file->f_mapping, vmf->pgoff, NO_SECTOR, false,
1075			true);
1076
1077	if (error == -ENOMEM)
1078		return VM_FAULT_OOM;
1079	if (error)
1080		return VM_FAULT_SIGBUS;
1081	return VM_FAULT_NOPAGE;
1082}
1083EXPORT_SYMBOL_GPL(dax_pfn_mkwrite);
 
 
 
 
 
 
 
 
 
1084
1085/**
1086 * dax_zero_page_range - zero a range within a page of a DAX file
1087 * @inode: The file being truncated
1088 * @from: The file offset that is being truncated to
1089 * @length: The number of bytes to zero
1090 * @get_block: The filesystem method used to translate file offsets to blocks
1091 *
1092 * This function can be called by a filesystem when it is zeroing part of a
1093 * page in a DAX file.  This is intended for hole-punch operations.  If
1094 * you are truncating a file, the helper function dax_truncate_page() may be
1095 * more convenient.
1096 *
1097 * We work in terms of PAGE_SIZE here for commonality with
1098 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1099 * took care of disposing of the unnecessary blocks.  Even if the filesystem
1100 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1101 * since the file might be mmapped.
1102 */
1103int dax_zero_page_range(struct inode *inode, loff_t from, unsigned length,
1104							get_block_t get_block)
1105{
1106	struct buffer_head bh;
1107	pgoff_t index = from >> PAGE_SHIFT;
1108	unsigned offset = from & (PAGE_SIZE-1);
1109	int err;
1110
1111	/* Block boundary? Nothing to do */
1112	if (!length)
1113		return 0;
1114	BUG_ON((offset + length) > PAGE_SIZE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1115
1116	memset(&bh, 0, sizeof(bh));
1117	bh.b_bdev = inode->i_sb->s_bdev;
1118	bh.b_size = PAGE_SIZE;
1119	err = get_block(inode, index, &bh, 0);
1120	if (err < 0)
1121		return err;
1122	if (buffer_written(&bh)) {
1123		struct block_device *bdev = bh.b_bdev;
1124		struct blk_dax_ctl dax = {
1125			.sector = to_sector(&bh, inode),
1126			.size = PAGE_SIZE,
1127		};
1128
1129		if (dax_map_atomic(bdev, &dax) < 0)
1130			return PTR_ERR(dax.addr);
1131		clear_pmem(dax.addr + offset, length);
1132		wmb_pmem();
1133		dax_unmap_atomic(bdev, &dax);
1134	}
1135
1136	return 0;
1137}
1138EXPORT_SYMBOL_GPL(dax_zero_page_range);
1139
1140/**
1141 * dax_truncate_page - handle a partial page being truncated in a DAX file
1142 * @inode: The file being truncated
1143 * @from: The file offset that is being truncated to
1144 * @get_block: The filesystem method used to translate file offsets to blocks
1145 *
1146 * Similar to block_truncate_page(), this function can be called by a
1147 * filesystem when it is truncating a DAX file to handle the partial page.
1148 *
1149 * We work in terms of PAGE_SIZE here for commonality with
1150 * block_truncate_page(), but we could go down to PAGE_SIZE if the filesystem
1151 * took care of disposing of the unnecessary blocks.  Even if the filesystem
1152 * block size is smaller than PAGE_SIZE, we have to zero the rest of the page
1153 * since the file might be mmapped.
1154 */
1155int dax_truncate_page(struct inode *inode, loff_t from, get_block_t get_block)
1156{
1157	unsigned length = PAGE_ALIGN(from) - from;
1158	return dax_zero_page_range(inode, from, length, get_block);
1159}
1160EXPORT_SYMBOL_GPL(dax_truncate_page);