Linux Audio

Check our new training course

Loading...
v4.6
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *		Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 * 		Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *		(Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/hugetlb.h>
  44#include <linux/mman.h>
  45#include <linux/swap.h>
  46#include <linux/highmem.h>
  47#include <linux/pagemap.h>
  48#include <linux/ksm.h>
  49#include <linux/rmap.h>
  50#include <linux/export.h>
  51#include <linux/delayacct.h>
  52#include <linux/init.h>
  53#include <linux/pfn_t.h>
  54#include <linux/writeback.h>
  55#include <linux/memcontrol.h>
  56#include <linux/mmu_notifier.h>
  57#include <linux/kallsyms.h>
  58#include <linux/swapops.h>
  59#include <linux/elf.h>
  60#include <linux/gfp.h>
  61#include <linux/migrate.h>
  62#include <linux/string.h>
  63#include <linux/dma-debug.h>
  64#include <linux/debugfs.h>
  65#include <linux/userfaultfd_k.h>
  66
  67#include <asm/io.h>
  68#include <asm/mmu_context.h>
  69#include <asm/pgalloc.h>
  70#include <asm/uaccess.h>
  71#include <asm/tlb.h>
  72#include <asm/tlbflush.h>
  73#include <asm/pgtable.h>
  74
  75#include "internal.h"
  76
  77#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
  78#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  79#endif
  80
  81#ifndef CONFIG_NEED_MULTIPLE_NODES
  82/* use the per-pgdat data instead for discontigmem - mbligh */
  83unsigned long max_mapnr;
  84struct page *mem_map;
  85
  86EXPORT_SYMBOL(max_mapnr);
  87EXPORT_SYMBOL(mem_map);
  88#endif
  89
 
  90/*
  91 * A number of key systems in x86 including ioremap() rely on the assumption
  92 * that high_memory defines the upper bound on direct map memory, then end
  93 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  95 * and ZONE_HIGHMEM.
  96 */
  97void * high_memory;
  98
 
  99EXPORT_SYMBOL(high_memory);
 100
 101/*
 102 * Randomize the address space (stacks, mmaps, brk, etc.).
 103 *
 104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 105 *   as ancient (libc5 based) binaries can segfault. )
 106 */
 107int randomize_va_space __read_mostly =
 108#ifdef CONFIG_COMPAT_BRK
 109					1;
 110#else
 111					2;
 112#endif
 113
 114static int __init disable_randmaps(char *s)
 115{
 116	randomize_va_space = 0;
 117	return 1;
 118}
 119__setup("norandmaps", disable_randmaps);
 120
 121unsigned long zero_pfn __read_mostly;
 122unsigned long highest_memmap_pfn __read_mostly;
 123
 124EXPORT_SYMBOL(zero_pfn);
 125
 126/*
 127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 128 */
 129static int __init init_zero_pfn(void)
 130{
 131	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 132	return 0;
 133}
 134core_initcall(init_zero_pfn);
 135
 136
 137#if defined(SPLIT_RSS_COUNTING)
 138
 139void sync_mm_rss(struct mm_struct *mm)
 140{
 141	int i;
 142
 143	for (i = 0; i < NR_MM_COUNTERS; i++) {
 144		if (current->rss_stat.count[i]) {
 145			add_mm_counter(mm, i, current->rss_stat.count[i]);
 146			current->rss_stat.count[i] = 0;
 147		}
 148	}
 149	current->rss_stat.events = 0;
 150}
 151
 152static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 153{
 154	struct task_struct *task = current;
 155
 156	if (likely(task->mm == mm))
 157		task->rss_stat.count[member] += val;
 158	else
 159		add_mm_counter(mm, member, val);
 160}
 161#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 162#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 163
 164/* sync counter once per 64 page faults */
 165#define TASK_RSS_EVENTS_THRESH	(64)
 166static void check_sync_rss_stat(struct task_struct *task)
 167{
 168	if (unlikely(task != current))
 169		return;
 170	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 171		sync_mm_rss(task->mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 172}
 173#else /* SPLIT_RSS_COUNTING */
 174
 175#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 176#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 177
 178static void check_sync_rss_stat(struct task_struct *task)
 179{
 180}
 181
 182#endif /* SPLIT_RSS_COUNTING */
 183
 184#ifdef HAVE_GENERIC_MMU_GATHER
 185
 186static bool tlb_next_batch(struct mmu_gather *tlb)
 187{
 188	struct mmu_gather_batch *batch;
 189
 190	batch = tlb->active;
 191	if (batch->next) {
 192		tlb->active = batch->next;
 193		return true;
 194	}
 195
 196	if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
 197		return false;
 198
 199	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 200	if (!batch)
 201		return false;
 202
 203	tlb->batch_count++;
 204	batch->next = NULL;
 205	batch->nr   = 0;
 206	batch->max  = MAX_GATHER_BATCH;
 207
 208	tlb->active->next = batch;
 209	tlb->active = batch;
 210
 211	return true;
 212}
 213
 214/* tlb_gather_mmu
 215 *	Called to initialize an (on-stack) mmu_gather structure for page-table
 216 *	tear-down from @mm. The @fullmm argument is used when @mm is without
 217 *	users and we're going to destroy the full address space (exit/execve).
 218 */
 219void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
 220{
 221	tlb->mm = mm;
 222
 223	/* Is it from 0 to ~0? */
 224	tlb->fullmm     = !(start | (end+1));
 225	tlb->need_flush_all = 0;
 226	tlb->local.next = NULL;
 227	tlb->local.nr   = 0;
 228	tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 229	tlb->active     = &tlb->local;
 230	tlb->batch_count = 0;
 231
 232#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 233	tlb->batch = NULL;
 234#endif
 235
 236	__tlb_reset_range(tlb);
 237}
 238
 239static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
 240{
 241	if (!tlb->end)
 242		return;
 243
 
 
 
 244	tlb_flush(tlb);
 245	mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
 246#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 247	tlb_table_flush(tlb);
 248#endif
 249	__tlb_reset_range(tlb);
 250}
 251
 252static void tlb_flush_mmu_free(struct mmu_gather *tlb)
 253{
 254	struct mmu_gather_batch *batch;
 255
 256	for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
 257		free_pages_and_swap_cache(batch->pages, batch->nr);
 258		batch->nr = 0;
 259	}
 260	tlb->active = &tlb->local;
 261}
 262
 263void tlb_flush_mmu(struct mmu_gather *tlb)
 264{
 265	tlb_flush_mmu_tlbonly(tlb);
 266	tlb_flush_mmu_free(tlb);
 267}
 268
 269/* tlb_finish_mmu
 270 *	Called at the end of the shootdown operation to free up any resources
 271 *	that were required.
 272 */
 273void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 274{
 275	struct mmu_gather_batch *batch, *next;
 276
 277	tlb_flush_mmu(tlb);
 278
 279	/* keep the page table cache within bounds */
 280	check_pgt_cache();
 281
 282	for (batch = tlb->local.next; batch; batch = next) {
 283		next = batch->next;
 284		free_pages((unsigned long)batch, 0);
 285	}
 286	tlb->local.next = NULL;
 287}
 288
 289/* __tlb_remove_page
 290 *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 291 *	handling the additional races in SMP caused by other CPUs caching valid
 292 *	mappings in their TLBs. Returns the number of free page slots left.
 293 *	When out of page slots we must call tlb_flush_mmu().
 294 */
 295int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 296{
 297	struct mmu_gather_batch *batch;
 298
 299	VM_BUG_ON(!tlb->end);
 
 
 
 
 
 300
 301	batch = tlb->active;
 302	batch->pages[batch->nr++] = page;
 303	if (batch->nr == batch->max) {
 304		if (!tlb_next_batch(tlb))
 305			return 0;
 306		batch = tlb->active;
 307	}
 308	VM_BUG_ON_PAGE(batch->nr > batch->max, page);
 309
 310	return batch->max - batch->nr;
 311}
 312
 313#endif /* HAVE_GENERIC_MMU_GATHER */
 314
 315#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 316
 317/*
 318 * See the comment near struct mmu_table_batch.
 319 */
 320
 321static void tlb_remove_table_smp_sync(void *arg)
 322{
 323	/* Simply deliver the interrupt */
 324}
 325
 326static void tlb_remove_table_one(void *table)
 327{
 328	/*
 329	 * This isn't an RCU grace period and hence the page-tables cannot be
 330	 * assumed to be actually RCU-freed.
 331	 *
 332	 * It is however sufficient for software page-table walkers that rely on
 333	 * IRQ disabling. See the comment near struct mmu_table_batch.
 334	 */
 335	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 336	__tlb_remove_table(table);
 337}
 338
 339static void tlb_remove_table_rcu(struct rcu_head *head)
 340{
 341	struct mmu_table_batch *batch;
 342	int i;
 343
 344	batch = container_of(head, struct mmu_table_batch, rcu);
 345
 346	for (i = 0; i < batch->nr; i++)
 347		__tlb_remove_table(batch->tables[i]);
 348
 349	free_page((unsigned long)batch);
 350}
 351
 352void tlb_table_flush(struct mmu_gather *tlb)
 353{
 354	struct mmu_table_batch **batch = &tlb->batch;
 355
 356	if (*batch) {
 357		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 358		*batch = NULL;
 359	}
 360}
 361
 362void tlb_remove_table(struct mmu_gather *tlb, void *table)
 363{
 364	struct mmu_table_batch **batch = &tlb->batch;
 365
 
 
 366	/*
 367	 * When there's less then two users of this mm there cannot be a
 368	 * concurrent page-table walk.
 369	 */
 370	if (atomic_read(&tlb->mm->mm_users) < 2) {
 371		__tlb_remove_table(table);
 372		return;
 373	}
 374
 375	if (*batch == NULL) {
 376		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 377		if (*batch == NULL) {
 378			tlb_remove_table_one(table);
 379			return;
 380		}
 381		(*batch)->nr = 0;
 382	}
 383	(*batch)->tables[(*batch)->nr++] = table;
 384	if ((*batch)->nr == MAX_TABLE_BATCH)
 385		tlb_table_flush(tlb);
 386}
 387
 388#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 389
 390/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 391 * Note: this doesn't free the actual pages themselves. That
 392 * has been handled earlier when unmapping all the memory regions.
 393 */
 394static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 395			   unsigned long addr)
 396{
 397	pgtable_t token = pmd_pgtable(*pmd);
 398	pmd_clear(pmd);
 399	pte_free_tlb(tlb, token, addr);
 400	atomic_long_dec(&tlb->mm->nr_ptes);
 401}
 402
 403static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 404				unsigned long addr, unsigned long end,
 405				unsigned long floor, unsigned long ceiling)
 406{
 407	pmd_t *pmd;
 408	unsigned long next;
 409	unsigned long start;
 410
 411	start = addr;
 412	pmd = pmd_offset(pud, addr);
 413	do {
 414		next = pmd_addr_end(addr, end);
 415		if (pmd_none_or_clear_bad(pmd))
 416			continue;
 417		free_pte_range(tlb, pmd, addr);
 418	} while (pmd++, addr = next, addr != end);
 419
 420	start &= PUD_MASK;
 421	if (start < floor)
 422		return;
 423	if (ceiling) {
 424		ceiling &= PUD_MASK;
 425		if (!ceiling)
 426			return;
 427	}
 428	if (end - 1 > ceiling - 1)
 429		return;
 430
 431	pmd = pmd_offset(pud, start);
 432	pud_clear(pud);
 433	pmd_free_tlb(tlb, pmd, start);
 434	mm_dec_nr_pmds(tlb->mm);
 435}
 436
 437static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 438				unsigned long addr, unsigned long end,
 439				unsigned long floor, unsigned long ceiling)
 440{
 441	pud_t *pud;
 442	unsigned long next;
 443	unsigned long start;
 444
 445	start = addr;
 446	pud = pud_offset(pgd, addr);
 447	do {
 448		next = pud_addr_end(addr, end);
 449		if (pud_none_or_clear_bad(pud))
 450			continue;
 451		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 452	} while (pud++, addr = next, addr != end);
 453
 454	start &= PGDIR_MASK;
 455	if (start < floor)
 456		return;
 457	if (ceiling) {
 458		ceiling &= PGDIR_MASK;
 459		if (!ceiling)
 460			return;
 461	}
 462	if (end - 1 > ceiling - 1)
 463		return;
 464
 465	pud = pud_offset(pgd, start);
 466	pgd_clear(pgd);
 467	pud_free_tlb(tlb, pud, start);
 468}
 469
 470/*
 471 * This function frees user-level page tables of a process.
 
 
 472 */
 473void free_pgd_range(struct mmu_gather *tlb,
 474			unsigned long addr, unsigned long end,
 475			unsigned long floor, unsigned long ceiling)
 476{
 477	pgd_t *pgd;
 478	unsigned long next;
 479
 480	/*
 481	 * The next few lines have given us lots of grief...
 482	 *
 483	 * Why are we testing PMD* at this top level?  Because often
 484	 * there will be no work to do at all, and we'd prefer not to
 485	 * go all the way down to the bottom just to discover that.
 486	 *
 487	 * Why all these "- 1"s?  Because 0 represents both the bottom
 488	 * of the address space and the top of it (using -1 for the
 489	 * top wouldn't help much: the masks would do the wrong thing).
 490	 * The rule is that addr 0 and floor 0 refer to the bottom of
 491	 * the address space, but end 0 and ceiling 0 refer to the top
 492	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 493	 * that end 0 case should be mythical).
 494	 *
 495	 * Wherever addr is brought up or ceiling brought down, we must
 496	 * be careful to reject "the opposite 0" before it confuses the
 497	 * subsequent tests.  But what about where end is brought down
 498	 * by PMD_SIZE below? no, end can't go down to 0 there.
 499	 *
 500	 * Whereas we round start (addr) and ceiling down, by different
 501	 * masks at different levels, in order to test whether a table
 502	 * now has no other vmas using it, so can be freed, we don't
 503	 * bother to round floor or end up - the tests don't need that.
 504	 */
 505
 506	addr &= PMD_MASK;
 507	if (addr < floor) {
 508		addr += PMD_SIZE;
 509		if (!addr)
 510			return;
 511	}
 512	if (ceiling) {
 513		ceiling &= PMD_MASK;
 514		if (!ceiling)
 515			return;
 516	}
 517	if (end - 1 > ceiling - 1)
 518		end -= PMD_SIZE;
 519	if (addr > end - 1)
 520		return;
 521
 522	pgd = pgd_offset(tlb->mm, addr);
 523	do {
 524		next = pgd_addr_end(addr, end);
 525		if (pgd_none_or_clear_bad(pgd))
 526			continue;
 527		free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 528	} while (pgd++, addr = next, addr != end);
 529}
 530
 531void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 532		unsigned long floor, unsigned long ceiling)
 533{
 534	while (vma) {
 535		struct vm_area_struct *next = vma->vm_next;
 536		unsigned long addr = vma->vm_start;
 537
 538		/*
 539		 * Hide vma from rmap and truncate_pagecache before freeing
 540		 * pgtables
 541		 */
 542		unlink_anon_vmas(vma);
 543		unlink_file_vma(vma);
 544
 545		if (is_vm_hugetlb_page(vma)) {
 546			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 547				floor, next? next->vm_start: ceiling);
 548		} else {
 549			/*
 550			 * Optimization: gather nearby vmas into one call down
 551			 */
 552			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 553			       && !is_vm_hugetlb_page(next)) {
 554				vma = next;
 555				next = vma->vm_next;
 556				unlink_anon_vmas(vma);
 557				unlink_file_vma(vma);
 558			}
 559			free_pgd_range(tlb, addr, vma->vm_end,
 560				floor, next? next->vm_start: ceiling);
 561		}
 562		vma = next;
 563	}
 564}
 565
 566int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 
 567{
 568	spinlock_t *ptl;
 569	pgtable_t new = pte_alloc_one(mm, address);
 
 570	if (!new)
 571		return -ENOMEM;
 572
 573	/*
 574	 * Ensure all pte setup (eg. pte page lock and page clearing) are
 575	 * visible before the pte is made visible to other CPUs by being
 576	 * put into page tables.
 577	 *
 578	 * The other side of the story is the pointer chasing in the page
 579	 * table walking code (when walking the page table without locking;
 580	 * ie. most of the time). Fortunately, these data accesses consist
 581	 * of a chain of data-dependent loads, meaning most CPUs (alpha
 582	 * being the notable exception) will already guarantee loads are
 583	 * seen in-order. See the alpha page table accessors for the
 584	 * smp_read_barrier_depends() barriers in page table walking code.
 585	 */
 586	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 587
 588	ptl = pmd_lock(mm, pmd);
 
 589	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 590		atomic_long_inc(&mm->nr_ptes);
 591		pmd_populate(mm, pmd, new);
 592		new = NULL;
 593	}
 594	spin_unlock(ptl);
 
 595	if (new)
 596		pte_free(mm, new);
 
 
 597	return 0;
 598}
 599
 600int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 601{
 602	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 603	if (!new)
 604		return -ENOMEM;
 605
 606	smp_wmb(); /* See comment in __pte_alloc */
 607
 608	spin_lock(&init_mm.page_table_lock);
 609	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 610		pmd_populate_kernel(&init_mm, pmd, new);
 611		new = NULL;
 612	}
 
 613	spin_unlock(&init_mm.page_table_lock);
 614	if (new)
 615		pte_free_kernel(&init_mm, new);
 616	return 0;
 617}
 618
 619static inline void init_rss_vec(int *rss)
 620{
 621	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 622}
 623
 624static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 625{
 626	int i;
 627
 628	if (current->mm == mm)
 629		sync_mm_rss(mm);
 630	for (i = 0; i < NR_MM_COUNTERS; i++)
 631		if (rss[i])
 632			add_mm_counter(mm, i, rss[i]);
 633}
 634
 635/*
 636 * This function is called to print an error when a bad pte
 637 * is found. For example, we might have a PFN-mapped pte in
 638 * a region that doesn't allow it.
 639 *
 640 * The calling function must still handle the error.
 641 */
 642static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 643			  pte_t pte, struct page *page)
 644{
 645	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 646	pud_t *pud = pud_offset(pgd, addr);
 647	pmd_t *pmd = pmd_offset(pud, addr);
 648	struct address_space *mapping;
 649	pgoff_t index;
 650	static unsigned long resume;
 651	static unsigned long nr_shown;
 652	static unsigned long nr_unshown;
 653
 654	/*
 655	 * Allow a burst of 60 reports, then keep quiet for that minute;
 656	 * or allow a steady drip of one report per second.
 657	 */
 658	if (nr_shown == 60) {
 659		if (time_before(jiffies, resume)) {
 660			nr_unshown++;
 661			return;
 662		}
 663		if (nr_unshown) {
 664			pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 665				 nr_unshown);
 
 666			nr_unshown = 0;
 667		}
 668		nr_shown = 0;
 669	}
 670	if (nr_shown++ == 0)
 671		resume = jiffies + 60 * HZ;
 672
 673	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 674	index = linear_page_index(vma, addr);
 675
 676	pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 677		 current->comm,
 678		 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 
 679	if (page)
 680		dump_page(page, "bad pte");
 681	pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 682		 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 
 683	/*
 684	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 685	 */
 686	pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
 687		 vma->vm_file,
 688		 vma->vm_ops ? vma->vm_ops->fault : NULL,
 689		 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 690		 mapping ? mapping->a_ops->readpage : NULL);
 
 691	dump_stack();
 692	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 693}
 
 694
 695/*
 696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 697 *
 698 * "Special" mappings do not wish to be associated with a "struct page" (either
 699 * it doesn't exist, or it exists but they don't want to touch it). In this
 700 * case, NULL is returned here. "Normal" mappings do have a struct page.
 701 *
 702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 703 * pte bit, in which case this function is trivial. Secondly, an architecture
 704 * may not have a spare pte bit, which requires a more complicated scheme,
 705 * described below.
 706 *
 707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 708 * special mapping (even if there are underlying and valid "struct pages").
 709 * COWed pages of a VM_PFNMAP are always normal.
 710 *
 711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 714 * mapping will always honor the rule
 715 *
 716 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 717 *
 718 * And for normal mappings this is false.
 719 *
 720 * This restricts such mappings to be a linear translation from virtual address
 721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 722 * as the vma is not a COW mapping; in that case, we know that all ptes are
 723 * special (because none can have been COWed).
 724 *
 725 *
 726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 727 *
 728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 729 * page" backing, however the difference is that _all_ pages with a struct
 730 * page (that is, those where pfn_valid is true) are refcounted and considered
 731 * normal pages by the VM. The disadvantage is that pages are refcounted
 732 * (which can be slower and simply not an option for some PFNMAP users). The
 733 * advantage is that we don't have to follow the strict linearity rule of
 734 * PFNMAP mappings in order to support COWable mappings.
 735 *
 736 */
 737#ifdef __HAVE_ARCH_PTE_SPECIAL
 738# define HAVE_PTE_SPECIAL 1
 739#else
 740# define HAVE_PTE_SPECIAL 0
 741#endif
 742struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 743				pte_t pte)
 744{
 745	unsigned long pfn = pte_pfn(pte);
 746
 747	if (HAVE_PTE_SPECIAL) {
 748		if (likely(!pte_special(pte)))
 749			goto check_pfn;
 750		if (vma->vm_ops && vma->vm_ops->find_special_page)
 751			return vma->vm_ops->find_special_page(vma, addr);
 752		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 753			return NULL;
 754		if (!is_zero_pfn(pfn))
 755			print_bad_pte(vma, addr, pte, NULL);
 756		return NULL;
 757	}
 758
 759	/* !HAVE_PTE_SPECIAL case follows: */
 760
 761	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 762		if (vma->vm_flags & VM_MIXEDMAP) {
 763			if (!pfn_valid(pfn))
 764				return NULL;
 765			goto out;
 766		} else {
 767			unsigned long off;
 768			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 769			if (pfn == vma->vm_pgoff + off)
 770				return NULL;
 771			if (!is_cow_mapping(vma->vm_flags))
 772				return NULL;
 773		}
 774	}
 775
 776	if (is_zero_pfn(pfn))
 777		return NULL;
 778check_pfn:
 779	if (unlikely(pfn > highest_memmap_pfn)) {
 780		print_bad_pte(vma, addr, pte, NULL);
 781		return NULL;
 782	}
 783
 784	/*
 785	 * NOTE! We still have PageReserved() pages in the page tables.
 786	 * eg. VDSO mappings can cause them to exist.
 787	 */
 788out:
 789	return pfn_to_page(pfn);
 790}
 791
 792#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 793struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 794				pmd_t pmd)
 795{
 796	unsigned long pfn = pmd_pfn(pmd);
 797
 798	/*
 799	 * There is no pmd_special() but there may be special pmds, e.g.
 800	 * in a direct-access (dax) mapping, so let's just replicate the
 801	 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
 802	 */
 803	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 804		if (vma->vm_flags & VM_MIXEDMAP) {
 805			if (!pfn_valid(pfn))
 806				return NULL;
 807			goto out;
 808		} else {
 809			unsigned long off;
 810			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 811			if (pfn == vma->vm_pgoff + off)
 812				return NULL;
 813			if (!is_cow_mapping(vma->vm_flags))
 814				return NULL;
 815		}
 816	}
 817
 818	if (is_zero_pfn(pfn))
 819		return NULL;
 820	if (unlikely(pfn > highest_memmap_pfn))
 821		return NULL;
 822
 823	/*
 824	 * NOTE! We still have PageReserved() pages in the page tables.
 825	 * eg. VDSO mappings can cause them to exist.
 826	 */
 827out:
 828	return pfn_to_page(pfn);
 829}
 830#endif
 831
 832/*
 833 * copy one vm_area from one task to the other. Assumes the page tables
 834 * already present in the new task to be cleared in the whole range
 835 * covered by this vma.
 836 */
 837
 838static inline unsigned long
 839copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 840		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 841		unsigned long addr, int *rss)
 842{
 843	unsigned long vm_flags = vma->vm_flags;
 844	pte_t pte = *src_pte;
 845	struct page *page;
 846
 847	/* pte contains position in swap or file, so copy. */
 848	if (unlikely(!pte_present(pte))) {
 849		swp_entry_t entry = pte_to_swp_entry(pte);
 
 850
 851		if (likely(!non_swap_entry(entry))) {
 852			if (swap_duplicate(entry) < 0)
 853				return entry.val;
 854
 855			/* make sure dst_mm is on swapoff's mmlist. */
 856			if (unlikely(list_empty(&dst_mm->mmlist))) {
 857				spin_lock(&mmlist_lock);
 858				if (list_empty(&dst_mm->mmlist))
 859					list_add(&dst_mm->mmlist,
 860							&src_mm->mmlist);
 861				spin_unlock(&mmlist_lock);
 862			}
 863			rss[MM_SWAPENTS]++;
 864		} else if (is_migration_entry(entry)) {
 865			page = migration_entry_to_page(entry);
 866
 867			rss[mm_counter(page)]++;
 868
 869			if (is_write_migration_entry(entry) &&
 870					is_cow_mapping(vm_flags)) {
 871				/*
 872				 * COW mappings require pages in both
 873				 * parent and child to be set to read.
 874				 */
 875				make_migration_entry_read(&entry);
 876				pte = swp_entry_to_pte(entry);
 877				if (pte_swp_soft_dirty(*src_pte))
 878					pte = pte_swp_mksoft_dirty(pte);
 879				set_pte_at(src_mm, addr, src_pte, pte);
 880			}
 881		}
 882		goto out_set_pte;
 883	}
 884
 885	/*
 886	 * If it's a COW mapping, write protect it both
 887	 * in the parent and the child
 888	 */
 889	if (is_cow_mapping(vm_flags)) {
 890		ptep_set_wrprotect(src_mm, addr, src_pte);
 891		pte = pte_wrprotect(pte);
 892	}
 893
 894	/*
 895	 * If it's a shared mapping, mark it clean in
 896	 * the child
 897	 */
 898	if (vm_flags & VM_SHARED)
 899		pte = pte_mkclean(pte);
 900	pte = pte_mkold(pte);
 901
 902	page = vm_normal_page(vma, addr, pte);
 903	if (page) {
 904		get_page(page);
 905		page_dup_rmap(page, false);
 906		rss[mm_counter(page)]++;
 
 
 
 907	}
 908
 909out_set_pte:
 910	set_pte_at(dst_mm, addr, dst_pte, pte);
 911	return 0;
 912}
 913
 914static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 915		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 916		   unsigned long addr, unsigned long end)
 917{
 918	pte_t *orig_src_pte, *orig_dst_pte;
 919	pte_t *src_pte, *dst_pte;
 920	spinlock_t *src_ptl, *dst_ptl;
 921	int progress = 0;
 922	int rss[NR_MM_COUNTERS];
 923	swp_entry_t entry = (swp_entry_t){0};
 924
 925again:
 926	init_rss_vec(rss);
 927
 928	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 929	if (!dst_pte)
 930		return -ENOMEM;
 931	src_pte = pte_offset_map(src_pmd, addr);
 932	src_ptl = pte_lockptr(src_mm, src_pmd);
 933	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 934	orig_src_pte = src_pte;
 935	orig_dst_pte = dst_pte;
 936	arch_enter_lazy_mmu_mode();
 937
 938	do {
 939		/*
 940		 * We are holding two locks at this point - either of them
 941		 * could generate latencies in another task on another CPU.
 942		 */
 943		if (progress >= 32) {
 944			progress = 0;
 945			if (need_resched() ||
 946			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 947				break;
 948		}
 949		if (pte_none(*src_pte)) {
 950			progress++;
 951			continue;
 952		}
 953		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 954							vma, addr, rss);
 955		if (entry.val)
 956			break;
 957		progress += 8;
 958	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 959
 960	arch_leave_lazy_mmu_mode();
 961	spin_unlock(src_ptl);
 962	pte_unmap(orig_src_pte);
 963	add_mm_rss_vec(dst_mm, rss);
 964	pte_unmap_unlock(orig_dst_pte, dst_ptl);
 965	cond_resched();
 966
 967	if (entry.val) {
 968		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 969			return -ENOMEM;
 970		progress = 0;
 971	}
 972	if (addr != end)
 973		goto again;
 974	return 0;
 975}
 976
 977static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 978		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 979		unsigned long addr, unsigned long end)
 980{
 981	pmd_t *src_pmd, *dst_pmd;
 982	unsigned long next;
 983
 984	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 985	if (!dst_pmd)
 986		return -ENOMEM;
 987	src_pmd = pmd_offset(src_pud, addr);
 988	do {
 989		next = pmd_addr_end(addr, end);
 990		if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
 991			int err;
 992			VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
 993			err = copy_huge_pmd(dst_mm, src_mm,
 994					    dst_pmd, src_pmd, addr, vma);
 995			if (err == -ENOMEM)
 996				return -ENOMEM;
 997			if (!err)
 998				continue;
 999			/* fall through */
1000		}
1001		if (pmd_none_or_clear_bad(src_pmd))
1002			continue;
1003		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1004						vma, addr, next))
1005			return -ENOMEM;
1006	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1007	return 0;
1008}
1009
1010static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1011		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1012		unsigned long addr, unsigned long end)
1013{
1014	pud_t *src_pud, *dst_pud;
1015	unsigned long next;
1016
1017	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1018	if (!dst_pud)
1019		return -ENOMEM;
1020	src_pud = pud_offset(src_pgd, addr);
1021	do {
1022		next = pud_addr_end(addr, end);
1023		if (pud_none_or_clear_bad(src_pud))
1024			continue;
1025		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1026						vma, addr, next))
1027			return -ENOMEM;
1028	} while (dst_pud++, src_pud++, addr = next, addr != end);
1029	return 0;
1030}
1031
1032int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033		struct vm_area_struct *vma)
1034{
1035	pgd_t *src_pgd, *dst_pgd;
1036	unsigned long next;
1037	unsigned long addr = vma->vm_start;
1038	unsigned long end = vma->vm_end;
1039	unsigned long mmun_start;	/* For mmu_notifiers */
1040	unsigned long mmun_end;		/* For mmu_notifiers */
1041	bool is_cow;
1042	int ret;
1043
1044	/*
1045	 * Don't copy ptes where a page fault will fill them correctly.
1046	 * Fork becomes much lighter when there are big shared or private
1047	 * readonly mappings. The tradeoff is that copy_page_range is more
1048	 * efficient than faulting.
1049	 */
1050	if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1051			!vma->anon_vma)
1052		return 0;
 
1053
1054	if (is_vm_hugetlb_page(vma))
1055		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1056
1057	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1058		/*
1059		 * We do not free on error cases below as remove_vma
1060		 * gets called on error from higher level routine
1061		 */
1062		ret = track_pfn_copy(vma);
1063		if (ret)
1064			return ret;
1065	}
1066
1067	/*
1068	 * We need to invalidate the secondary MMU mappings only when
1069	 * there could be a permission downgrade on the ptes of the
1070	 * parent mm. And a permission downgrade will only happen if
1071	 * is_cow_mapping() returns true.
1072	 */
1073	is_cow = is_cow_mapping(vma->vm_flags);
1074	mmun_start = addr;
1075	mmun_end   = end;
1076	if (is_cow)
1077		mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1078						    mmun_end);
1079
1080	ret = 0;
1081	dst_pgd = pgd_offset(dst_mm, addr);
1082	src_pgd = pgd_offset(src_mm, addr);
1083	do {
1084		next = pgd_addr_end(addr, end);
1085		if (pgd_none_or_clear_bad(src_pgd))
1086			continue;
1087		if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1088					    vma, addr, next))) {
1089			ret = -ENOMEM;
1090			break;
1091		}
1092	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1093
1094	if (is_cow)
1095		mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
 
1096	return ret;
1097}
1098
1099static unsigned long zap_pte_range(struct mmu_gather *tlb,
1100				struct vm_area_struct *vma, pmd_t *pmd,
1101				unsigned long addr, unsigned long end,
1102				struct zap_details *details)
1103{
1104	struct mm_struct *mm = tlb->mm;
1105	int force_flush = 0;
1106	int rss[NR_MM_COUNTERS];
1107	spinlock_t *ptl;
1108	pte_t *start_pte;
1109	pte_t *pte;
1110	swp_entry_t entry;
1111
1112again:
1113	init_rss_vec(rss);
1114	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1115	pte = start_pte;
1116	arch_enter_lazy_mmu_mode();
1117	do {
1118		pte_t ptent = *pte;
1119		if (pte_none(ptent)) {
1120			continue;
1121		}
1122
1123		if (pte_present(ptent)) {
1124			struct page *page;
1125
1126			page = vm_normal_page(vma, addr, ptent);
1127			if (unlikely(details) && page) {
1128				/*
1129				 * unmap_shared_mapping_pages() wants to
1130				 * invalidate cache without truncating:
1131				 * unmap shared but keep private pages.
1132				 */
1133				if (details->check_mapping &&
1134				    details->check_mapping != page->mapping)
1135					continue;
 
 
 
 
 
 
 
 
1136			}
1137			ptent = ptep_get_and_clear_full(mm, addr, pte,
1138							tlb->fullmm);
1139			tlb_remove_tlb_entry(tlb, pte, addr);
1140			if (unlikely(!page))
1141				continue;
1142
1143			if (!PageAnon(page)) {
1144				if (pte_dirty(ptent)) {
1145					/*
1146					 * oom_reaper cannot tear down dirty
1147					 * pages
1148					 */
1149					if (unlikely(details && details->ignore_dirty))
1150						continue;
1151					force_flush = 1;
1152					set_page_dirty(page);
1153				}
1154				if (pte_young(ptent) &&
1155				    likely(!(vma->vm_flags & VM_SEQ_READ)))
1156					mark_page_accessed(page);
 
1157			}
1158			rss[mm_counter(page)]--;
1159			page_remove_rmap(page, false);
1160			if (unlikely(page_mapcount(page) < 0))
1161				print_bad_pte(vma, addr, ptent, page);
1162			if (unlikely(!__tlb_remove_page(tlb, page))) {
1163				force_flush = 1;
1164				addr += PAGE_SIZE;
1165				break;
1166			}
1167			continue;
1168		}
1169		/* only check swap_entries if explicitly asked for in details */
1170		if (unlikely(details && !details->check_swap_entries))
 
 
 
1171			continue;
 
 
 
 
 
1172
1173		entry = pte_to_swp_entry(ptent);
1174		if (!non_swap_entry(entry))
1175			rss[MM_SWAPENTS]--;
1176		else if (is_migration_entry(entry)) {
1177			struct page *page;
1178
1179			page = migration_entry_to_page(entry);
1180			rss[mm_counter(page)]--;
1181		}
1182		if (unlikely(!free_swap_and_cache(entry)))
1183			print_bad_pte(vma, addr, ptent, NULL);
1184		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185	} while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187	add_mm_rss_vec(mm, rss);
1188	arch_leave_lazy_mmu_mode();
1189
1190	/* Do the actual TLB flush before dropping ptl */
1191	if (force_flush)
1192		tlb_flush_mmu_tlbonly(tlb);
1193	pte_unmap_unlock(start_pte, ptl);
1194
1195	/*
1196	 * If we forced a TLB flush (either due to running out of
1197	 * batch buffers or because we needed to flush dirty TLB
1198	 * entries before releasing the ptl), free the batched
1199	 * memory too. Restart if we didn't do everything.
1200	 */
1201	if (force_flush) {
1202		force_flush = 0;
1203		tlb_flush_mmu_free(tlb);
1204
1205		if (addr != end)
1206			goto again;
1207	}
1208
1209	return addr;
1210}
1211
1212static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1213				struct vm_area_struct *vma, pud_t *pud,
1214				unsigned long addr, unsigned long end,
1215				struct zap_details *details)
1216{
1217	pmd_t *pmd;
1218	unsigned long next;
1219
1220	pmd = pmd_offset(pud, addr);
1221	do {
1222		next = pmd_addr_end(addr, end);
1223		if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1224			if (next - addr != HPAGE_PMD_SIZE) {
1225				VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1226				    !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1227				split_huge_pmd(vma, pmd, addr);
1228			} else if (zap_huge_pmd(tlb, vma, pmd, addr))
1229				goto next;
1230			/* fall through */
1231		}
1232		/*
1233		 * Here there can be other concurrent MADV_DONTNEED or
1234		 * trans huge page faults running, and if the pmd is
1235		 * none or trans huge it can change under us. This is
1236		 * because MADV_DONTNEED holds the mmap_sem in read
1237		 * mode.
1238		 */
1239		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1240			goto next;
1241		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1242next:
1243		cond_resched();
1244	} while (pmd++, addr = next, addr != end);
1245
1246	return addr;
1247}
1248
1249static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1250				struct vm_area_struct *vma, pgd_t *pgd,
1251				unsigned long addr, unsigned long end,
1252				struct zap_details *details)
1253{
1254	pud_t *pud;
1255	unsigned long next;
1256
1257	pud = pud_offset(pgd, addr);
1258	do {
1259		next = pud_addr_end(addr, end);
1260		if (pud_none_or_clear_bad(pud))
1261			continue;
1262		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1263	} while (pud++, addr = next, addr != end);
1264
1265	return addr;
1266}
1267
1268void unmap_page_range(struct mmu_gather *tlb,
1269			     struct vm_area_struct *vma,
1270			     unsigned long addr, unsigned long end,
1271			     struct zap_details *details)
1272{
1273	pgd_t *pgd;
1274	unsigned long next;
1275
 
 
 
1276	BUG_ON(addr >= end);
 
1277	tlb_start_vma(tlb, vma);
1278	pgd = pgd_offset(vma->vm_mm, addr);
1279	do {
1280		next = pgd_addr_end(addr, end);
1281		if (pgd_none_or_clear_bad(pgd))
1282			continue;
1283		next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1284	} while (pgd++, addr = next, addr != end);
1285	tlb_end_vma(tlb, vma);
1286}
1287
1288
1289static void unmap_single_vma(struct mmu_gather *tlb,
1290		struct vm_area_struct *vma, unsigned long start_addr,
1291		unsigned long end_addr,
1292		struct zap_details *details)
1293{
1294	unsigned long start = max(vma->vm_start, start_addr);
1295	unsigned long end;
1296
1297	if (start >= vma->vm_end)
1298		return;
1299	end = min(vma->vm_end, end_addr);
1300	if (end <= vma->vm_start)
1301		return;
1302
1303	if (vma->vm_file)
1304		uprobe_munmap(vma, start, end);
1305
1306	if (unlikely(vma->vm_flags & VM_PFNMAP))
1307		untrack_pfn(vma, 0, 0);
1308
1309	if (start != end) {
1310		if (unlikely(is_vm_hugetlb_page(vma))) {
1311			/*
1312			 * It is undesirable to test vma->vm_file as it
1313			 * should be non-null for valid hugetlb area.
1314			 * However, vm_file will be NULL in the error
1315			 * cleanup path of mmap_region. When
1316			 * hugetlbfs ->mmap method fails,
1317			 * mmap_region() nullifies vma->vm_file
1318			 * before calling this function to clean up.
1319			 * Since no pte has actually been setup, it is
1320			 * safe to do nothing in this case.
1321			 */
1322			if (vma->vm_file) {
1323				i_mmap_lock_write(vma->vm_file->f_mapping);
1324				__unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1325				i_mmap_unlock_write(vma->vm_file->f_mapping);
1326			}
1327		} else
1328			unmap_page_range(tlb, vma, start, end, details);
1329	}
1330}
1331
1332/**
1333 * unmap_vmas - unmap a range of memory covered by a list of vma's
1334 * @tlb: address of the caller's struct mmu_gather
1335 * @vma: the starting vma
1336 * @start_addr: virtual address at which to start unmapping
1337 * @end_addr: virtual address at which to end unmapping
 
 
 
 
1338 *
1339 * Unmap all pages in the vma list.
1340 *
1341 * Only addresses between `start' and `end' will be unmapped.
1342 *
1343 * The VMA list must be sorted in ascending virtual address order.
1344 *
1345 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1346 * range after unmap_vmas() returns.  So the only responsibility here is to
1347 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1348 * drops the lock and schedules.
1349 */
1350void unmap_vmas(struct mmu_gather *tlb,
1351		struct vm_area_struct *vma, unsigned long start_addr,
1352		unsigned long end_addr)
 
1353{
 
1354	struct mm_struct *mm = vma->vm_mm;
1355
1356	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1357	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1358		unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1359	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1360}
1361
1362/**
1363 * zap_page_range - remove user pages in a given range
1364 * @vma: vm_area_struct holding the applicable pages
1365 * @start: starting address of pages to zap
1366 * @size: number of bytes to zap
1367 * @details: details of shared cache invalidation
1368 *
1369 * Caller must protect the VMA list
1370 */
1371void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1372		unsigned long size, struct zap_details *details)
1373{
1374	struct mm_struct *mm = vma->vm_mm;
1375	struct mmu_gather tlb;
1376	unsigned long end = start + size;
1377
1378	lru_add_drain();
1379	tlb_gather_mmu(&tlb, mm, start, end);
1380	update_hiwater_rss(mm);
1381	mmu_notifier_invalidate_range_start(mm, start, end);
1382	for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1383		unmap_single_vma(&tlb, vma, start, end, details);
1384	mmu_notifier_invalidate_range_end(mm, start, end);
1385	tlb_finish_mmu(&tlb, start, end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1386}
1387
1388/**
1389 * zap_page_range_single - remove user pages in a given range
1390 * @vma: vm_area_struct holding the applicable pages
1391 * @address: starting address of pages to zap
1392 * @size: number of bytes to zap
1393 * @details: details of shared cache invalidation
1394 *
1395 * The range must fit into one VMA.
1396 */
1397static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1398		unsigned long size, struct zap_details *details)
1399{
1400	struct mm_struct *mm = vma->vm_mm;
1401	struct mmu_gather tlb;
1402	unsigned long end = address + size;
 
1403
1404	lru_add_drain();
1405	tlb_gather_mmu(&tlb, mm, address, end);
1406	update_hiwater_rss(mm);
1407	mmu_notifier_invalidate_range_start(mm, address, end);
1408	unmap_single_vma(&tlb, vma, address, end, details);
1409	mmu_notifier_invalidate_range_end(mm, address, end);
1410	tlb_finish_mmu(&tlb, address, end);
 
1411}
1412
1413/**
1414 * zap_vma_ptes - remove ptes mapping the vma
1415 * @vma: vm_area_struct holding ptes to be zapped
1416 * @address: starting address of pages to zap
1417 * @size: number of bytes to zap
1418 *
1419 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1420 *
1421 * The entire address range must be fully contained within the vma.
1422 *
1423 * Returns 0 if successful.
1424 */
1425int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1426		unsigned long size)
1427{
1428	if (address < vma->vm_start || address + size > vma->vm_end ||
1429	    		!(vma->vm_flags & VM_PFNMAP))
1430		return -1;
1431	zap_page_range_single(vma, address, size, NULL);
1432	return 0;
1433}
1434EXPORT_SYMBOL_GPL(zap_vma_ptes);
1435
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1436pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1437			spinlock_t **ptl)
1438{
1439	pgd_t * pgd = pgd_offset(mm, addr);
1440	pud_t * pud = pud_alloc(mm, pgd, addr);
1441	if (pud) {
1442		pmd_t * pmd = pmd_alloc(mm, pud, addr);
1443		if (pmd) {
1444			VM_BUG_ON(pmd_trans_huge(*pmd));
1445			return pte_alloc_map_lock(mm, pmd, addr, ptl);
1446		}
1447	}
1448	return NULL;
1449}
1450
1451/*
1452 * This is the old fallback for page remapping.
1453 *
1454 * For historical reasons, it only allows reserved pages. Only
1455 * old drivers should use this, and they needed to mark their
1456 * pages reserved for the old functions anyway.
1457 */
1458static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1459			struct page *page, pgprot_t prot)
1460{
1461	struct mm_struct *mm = vma->vm_mm;
1462	int retval;
1463	pte_t *pte;
1464	spinlock_t *ptl;
1465
1466	retval = -EINVAL;
1467	if (PageAnon(page))
1468		goto out;
1469	retval = -ENOMEM;
1470	flush_dcache_page(page);
1471	pte = get_locked_pte(mm, addr, &ptl);
1472	if (!pte)
1473		goto out;
1474	retval = -EBUSY;
1475	if (!pte_none(*pte))
1476		goto out_unlock;
1477
1478	/* Ok, finally just insert the thing.. */
1479	get_page(page);
1480	inc_mm_counter_fast(mm, mm_counter_file(page));
1481	page_add_file_rmap(page);
1482	set_pte_at(mm, addr, pte, mk_pte(page, prot));
1483
1484	retval = 0;
1485	pte_unmap_unlock(pte, ptl);
1486	return retval;
1487out_unlock:
1488	pte_unmap_unlock(pte, ptl);
1489out:
1490	return retval;
1491}
1492
1493/**
1494 * vm_insert_page - insert single page into user vma
1495 * @vma: user vma to map to
1496 * @addr: target user address of this page
1497 * @page: source kernel page
1498 *
1499 * This allows drivers to insert individual pages they've allocated
1500 * into a user vma.
1501 *
1502 * The page has to be a nice clean _individual_ kernel allocation.
1503 * If you allocate a compound page, you need to have marked it as
1504 * such (__GFP_COMP), or manually just split the page up yourself
1505 * (see split_page()).
1506 *
1507 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1508 * took an arbitrary page protection parameter. This doesn't allow
1509 * that. Your vma protection will have to be set up correctly, which
1510 * means that if you want a shared writable mapping, you'd better
1511 * ask for a shared writable mapping!
1512 *
1513 * The page does not need to be reserved.
1514 *
1515 * Usually this function is called from f_op->mmap() handler
1516 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1517 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1518 * function from other places, for example from page-fault handler.
1519 */
1520int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1521			struct page *page)
1522{
1523	if (addr < vma->vm_start || addr >= vma->vm_end)
1524		return -EFAULT;
1525	if (!page_count(page))
1526		return -EINVAL;
1527	if (!(vma->vm_flags & VM_MIXEDMAP)) {
1528		BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1529		BUG_ON(vma->vm_flags & VM_PFNMAP);
1530		vma->vm_flags |= VM_MIXEDMAP;
1531	}
1532	return insert_page(vma, addr, page, vma->vm_page_prot);
1533}
1534EXPORT_SYMBOL(vm_insert_page);
1535
1536static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1537			pfn_t pfn, pgprot_t prot)
1538{
1539	struct mm_struct *mm = vma->vm_mm;
1540	int retval;
1541	pte_t *pte, entry;
1542	spinlock_t *ptl;
1543
1544	retval = -ENOMEM;
1545	pte = get_locked_pte(mm, addr, &ptl);
1546	if (!pte)
1547		goto out;
1548	retval = -EBUSY;
1549	if (!pte_none(*pte))
1550		goto out_unlock;
1551
1552	/* Ok, finally just insert the thing.. */
1553	if (pfn_t_devmap(pfn))
1554		entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1555	else
1556		entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1557	set_pte_at(mm, addr, pte, entry);
1558	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1559
1560	retval = 0;
1561out_unlock:
1562	pte_unmap_unlock(pte, ptl);
1563out:
1564	return retval;
1565}
1566
1567/**
1568 * vm_insert_pfn - insert single pfn into user vma
1569 * @vma: user vma to map to
1570 * @addr: target user address of this page
1571 * @pfn: source kernel pfn
1572 *
1573 * Similar to vm_insert_page, this allows drivers to insert individual pages
1574 * they've allocated into a user vma. Same comments apply.
1575 *
1576 * This function should only be called from a vm_ops->fault handler, and
1577 * in that case the handler should return NULL.
1578 *
1579 * vma cannot be a COW mapping.
1580 *
1581 * As this is called only for pages that do not currently exist, we
1582 * do not need to flush old virtual caches or the TLB.
1583 */
1584int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1585			unsigned long pfn)
1586{
1587	return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1588}
1589EXPORT_SYMBOL(vm_insert_pfn);
1590
1591/**
1592 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1593 * @vma: user vma to map to
1594 * @addr: target user address of this page
1595 * @pfn: source kernel pfn
1596 * @pgprot: pgprot flags for the inserted page
1597 *
1598 * This is exactly like vm_insert_pfn, except that it allows drivers to
1599 * to override pgprot on a per-page basis.
1600 *
1601 * This only makes sense for IO mappings, and it makes no sense for
1602 * cow mappings.  In general, using multiple vmas is preferable;
1603 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1604 * impractical.
1605 */
1606int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1607			unsigned long pfn, pgprot_t pgprot)
1608{
1609	int ret;
 
1610	/*
1611	 * Technically, architectures with pte_special can avoid all these
1612	 * restrictions (same for remap_pfn_range).  However we would like
1613	 * consistency in testing and feature parity among all, so we should
1614	 * try to keep these invariants in place for everybody.
1615	 */
1616	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1617	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1618						(VM_PFNMAP|VM_MIXEDMAP));
1619	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1620	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1621
1622	if (addr < vma->vm_start || addr >= vma->vm_end)
1623		return -EFAULT;
1624	if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
1625		return -EINVAL;
1626
1627	ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
 
 
 
1628
1629	return ret;
1630}
1631EXPORT_SYMBOL(vm_insert_pfn_prot);
1632
1633int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1634			pfn_t pfn)
1635{
1636	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1637
1638	if (addr < vma->vm_start || addr >= vma->vm_end)
1639		return -EFAULT;
1640
1641	/*
1642	 * If we don't have pte special, then we have to use the pfn_valid()
1643	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1644	 * refcount the page if pfn_valid is true (hence insert_page rather
1645	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1646	 * without pte special, it would there be refcounted as a normal page.
1647	 */
1648	if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1649		struct page *page;
1650
1651		/*
1652		 * At this point we are committed to insert_page()
1653		 * regardless of whether the caller specified flags that
1654		 * result in pfn_t_has_page() == false.
1655		 */
1656		page = pfn_to_page(pfn_t_to_pfn(pfn));
1657		return insert_page(vma, addr, page, vma->vm_page_prot);
1658	}
1659	return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1660}
1661EXPORT_SYMBOL(vm_insert_mixed);
1662
1663/*
1664 * maps a range of physical memory into the requested pages. the old
1665 * mappings are removed. any references to nonexistent pages results
1666 * in null mappings (currently treated as "copy-on-access")
1667 */
1668static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1669			unsigned long addr, unsigned long end,
1670			unsigned long pfn, pgprot_t prot)
1671{
1672	pte_t *pte;
1673	spinlock_t *ptl;
1674
1675	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1676	if (!pte)
1677		return -ENOMEM;
1678	arch_enter_lazy_mmu_mode();
1679	do {
1680		BUG_ON(!pte_none(*pte));
1681		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1682		pfn++;
1683	} while (pte++, addr += PAGE_SIZE, addr != end);
1684	arch_leave_lazy_mmu_mode();
1685	pte_unmap_unlock(pte - 1, ptl);
1686	return 0;
1687}
1688
1689static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1690			unsigned long addr, unsigned long end,
1691			unsigned long pfn, pgprot_t prot)
1692{
1693	pmd_t *pmd;
1694	unsigned long next;
1695
1696	pfn -= addr >> PAGE_SHIFT;
1697	pmd = pmd_alloc(mm, pud, addr);
1698	if (!pmd)
1699		return -ENOMEM;
1700	VM_BUG_ON(pmd_trans_huge(*pmd));
1701	do {
1702		next = pmd_addr_end(addr, end);
1703		if (remap_pte_range(mm, pmd, addr, next,
1704				pfn + (addr >> PAGE_SHIFT), prot))
1705			return -ENOMEM;
1706	} while (pmd++, addr = next, addr != end);
1707	return 0;
1708}
1709
1710static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1711			unsigned long addr, unsigned long end,
1712			unsigned long pfn, pgprot_t prot)
1713{
1714	pud_t *pud;
1715	unsigned long next;
1716
1717	pfn -= addr >> PAGE_SHIFT;
1718	pud = pud_alloc(mm, pgd, addr);
1719	if (!pud)
1720		return -ENOMEM;
1721	do {
1722		next = pud_addr_end(addr, end);
1723		if (remap_pmd_range(mm, pud, addr, next,
1724				pfn + (addr >> PAGE_SHIFT), prot))
1725			return -ENOMEM;
1726	} while (pud++, addr = next, addr != end);
1727	return 0;
1728}
1729
1730/**
1731 * remap_pfn_range - remap kernel memory to userspace
1732 * @vma: user vma to map to
1733 * @addr: target user address to start at
1734 * @pfn: physical address of kernel memory
1735 * @size: size of map area
1736 * @prot: page protection flags for this mapping
1737 *
1738 *  Note: this is only safe if the mm semaphore is held when called.
1739 */
1740int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1741		    unsigned long pfn, unsigned long size, pgprot_t prot)
1742{
1743	pgd_t *pgd;
1744	unsigned long next;
1745	unsigned long end = addr + PAGE_ALIGN(size);
1746	struct mm_struct *mm = vma->vm_mm;
1747	int err;
1748
1749	/*
1750	 * Physically remapped pages are special. Tell the
1751	 * rest of the world about it:
1752	 *   VM_IO tells people not to look at these pages
1753	 *	(accesses can have side effects).
 
 
 
 
 
1754	 *   VM_PFNMAP tells the core MM that the base pages are just
1755	 *	raw PFN mappings, and do not have a "struct page" associated
1756	 *	with them.
1757	 *   VM_DONTEXPAND
1758	 *      Disable vma merging and expanding with mremap().
1759	 *   VM_DONTDUMP
1760	 *      Omit vma from core dump, even when VM_IO turned off.
1761	 *
1762	 * There's a horrible special case to handle copy-on-write
1763	 * behaviour that some programs depend on. We mark the "original"
1764	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1765	 * See vm_normal_page() for details.
1766	 */
1767	if (is_cow_mapping(vma->vm_flags)) {
1768		if (addr != vma->vm_start || end != vma->vm_end)
1769			return -EINVAL;
1770		vma->vm_pgoff = pfn;
1771	}
1772
1773	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1774	if (err)
1775		return -EINVAL;
1776
1777	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
 
 
 
 
 
 
 
 
 
 
 
1778
1779	BUG_ON(addr >= end);
1780	pfn -= addr >> PAGE_SHIFT;
1781	pgd = pgd_offset(mm, addr);
1782	flush_cache_range(vma, addr, end);
1783	do {
1784		next = pgd_addr_end(addr, end);
1785		err = remap_pud_range(mm, pgd, addr, next,
1786				pfn + (addr >> PAGE_SHIFT), prot);
1787		if (err)
1788			break;
1789	} while (pgd++, addr = next, addr != end);
1790
1791	if (err)
1792		untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1793
1794	return err;
1795}
1796EXPORT_SYMBOL(remap_pfn_range);
1797
1798/**
1799 * vm_iomap_memory - remap memory to userspace
1800 * @vma: user vma to map to
1801 * @start: start of area
1802 * @len: size of area
1803 *
1804 * This is a simplified io_remap_pfn_range() for common driver use. The
1805 * driver just needs to give us the physical memory range to be mapped,
1806 * we'll figure out the rest from the vma information.
1807 *
1808 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1809 * whatever write-combining details or similar.
1810 */
1811int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1812{
1813	unsigned long vm_len, pfn, pages;
1814
1815	/* Check that the physical memory area passed in looks valid */
1816	if (start + len < start)
1817		return -EINVAL;
1818	/*
1819	 * You *really* shouldn't map things that aren't page-aligned,
1820	 * but we've historically allowed it because IO memory might
1821	 * just have smaller alignment.
1822	 */
1823	len += start & ~PAGE_MASK;
1824	pfn = start >> PAGE_SHIFT;
1825	pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1826	if (pfn + pages < pfn)
1827		return -EINVAL;
1828
1829	/* We start the mapping 'vm_pgoff' pages into the area */
1830	if (vma->vm_pgoff > pages)
1831		return -EINVAL;
1832	pfn += vma->vm_pgoff;
1833	pages -= vma->vm_pgoff;
1834
1835	/* Can we fit all of the mapping? */
1836	vm_len = vma->vm_end - vma->vm_start;
1837	if (vm_len >> PAGE_SHIFT > pages)
1838		return -EINVAL;
1839
1840	/* Ok, let it rip */
1841	return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1842}
1843EXPORT_SYMBOL(vm_iomap_memory);
1844
1845static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1846				     unsigned long addr, unsigned long end,
1847				     pte_fn_t fn, void *data)
1848{
1849	pte_t *pte;
1850	int err;
1851	pgtable_t token;
1852	spinlock_t *uninitialized_var(ptl);
1853
1854	pte = (mm == &init_mm) ?
1855		pte_alloc_kernel(pmd, addr) :
1856		pte_alloc_map_lock(mm, pmd, addr, &ptl);
1857	if (!pte)
1858		return -ENOMEM;
1859
1860	BUG_ON(pmd_huge(*pmd));
1861
1862	arch_enter_lazy_mmu_mode();
1863
1864	token = pmd_pgtable(*pmd);
1865
1866	do {
1867		err = fn(pte++, token, addr, data);
1868		if (err)
1869			break;
1870	} while (addr += PAGE_SIZE, addr != end);
1871
1872	arch_leave_lazy_mmu_mode();
1873
1874	if (mm != &init_mm)
1875		pte_unmap_unlock(pte-1, ptl);
1876	return err;
1877}
1878
1879static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1880				     unsigned long addr, unsigned long end,
1881				     pte_fn_t fn, void *data)
1882{
1883	pmd_t *pmd;
1884	unsigned long next;
1885	int err;
1886
1887	BUG_ON(pud_huge(*pud));
1888
1889	pmd = pmd_alloc(mm, pud, addr);
1890	if (!pmd)
1891		return -ENOMEM;
1892	do {
1893		next = pmd_addr_end(addr, end);
1894		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1895		if (err)
1896			break;
1897	} while (pmd++, addr = next, addr != end);
1898	return err;
1899}
1900
1901static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1902				     unsigned long addr, unsigned long end,
1903				     pte_fn_t fn, void *data)
1904{
1905	pud_t *pud;
1906	unsigned long next;
1907	int err;
1908
1909	pud = pud_alloc(mm, pgd, addr);
1910	if (!pud)
1911		return -ENOMEM;
1912	do {
1913		next = pud_addr_end(addr, end);
1914		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1915		if (err)
1916			break;
1917	} while (pud++, addr = next, addr != end);
1918	return err;
1919}
1920
1921/*
1922 * Scan a region of virtual memory, filling in page tables as necessary
1923 * and calling a provided function on each leaf page table.
1924 */
1925int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1926			unsigned long size, pte_fn_t fn, void *data)
1927{
1928	pgd_t *pgd;
1929	unsigned long next;
1930	unsigned long end = addr + size;
1931	int err;
1932
1933	if (WARN_ON(addr >= end))
1934		return -EINVAL;
1935
1936	pgd = pgd_offset(mm, addr);
1937	do {
1938		next = pgd_addr_end(addr, end);
1939		err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1940		if (err)
1941			break;
1942	} while (pgd++, addr = next, addr != end);
1943
1944	return err;
1945}
1946EXPORT_SYMBOL_GPL(apply_to_page_range);
1947
1948/*
1949 * handle_pte_fault chooses page fault handler according to an entry which was
1950 * read non-atomically.  Before making any commitment, on those architectures
1951 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1952 * parts, do_swap_page must check under lock before unmapping the pte and
1953 * proceeding (but do_wp_page is only called after already making such a check;
 
1954 * and do_anonymous_page can safely check later on).
1955 */
1956static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1957				pte_t *page_table, pte_t orig_pte)
1958{
1959	int same = 1;
1960#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1961	if (sizeof(pte_t) > sizeof(unsigned long)) {
1962		spinlock_t *ptl = pte_lockptr(mm, pmd);
1963		spin_lock(ptl);
1964		same = pte_same(*page_table, orig_pte);
1965		spin_unlock(ptl);
1966	}
1967#endif
1968	pte_unmap(page_table);
1969	return same;
1970}
1971
1972static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1973{
1974	debug_dma_assert_idle(src);
1975
1976	/*
1977	 * If the source page was a PFN mapping, we don't have
1978	 * a "struct page" for it. We do a best-effort copy by
1979	 * just copying from the original user address. If that
1980	 * fails, we just zero-fill it. Live with it.
1981	 */
1982	if (unlikely(!src)) {
1983		void *kaddr = kmap_atomic(dst);
1984		void __user *uaddr = (void __user *)(va & PAGE_MASK);
1985
1986		/*
1987		 * This really shouldn't fail, because the page is there
1988		 * in the page tables. But it might just be unreadable,
1989		 * in which case we just give up and fill the result with
1990		 * zeroes.
1991		 */
1992		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1993			clear_page(kaddr);
1994		kunmap_atomic(kaddr);
1995		flush_dcache_page(dst);
1996	} else
1997		copy_user_highpage(dst, src, va, vma);
1998}
1999
2000static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2001{
2002	struct file *vm_file = vma->vm_file;
2003
2004	if (vm_file)
2005		return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2006
2007	/*
2008	 * Special mappings (e.g. VDSO) do not have any file so fake
2009	 * a default GFP_KERNEL for them.
2010	 */
2011	return GFP_KERNEL;
2012}
2013
2014/*
2015 * Notify the address space that the page is about to become writable so that
2016 * it can prohibit this or wait for the page to get into an appropriate state.
 
2017 *
2018 * We do this without the lock held, so that it can sleep if it needs to.
2019 */
2020static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2021	       unsigned long address)
2022{
2023	struct vm_fault vmf;
2024	int ret;
2025
2026	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2027	vmf.pgoff = page->index;
2028	vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2029	vmf.gfp_mask = __get_fault_gfp_mask(vma);
2030	vmf.page = page;
2031	vmf.cow_page = NULL;
2032
2033	ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2034	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2035		return ret;
2036	if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2037		lock_page(page);
2038		if (!page->mapping) {
2039			unlock_page(page);
2040			return 0; /* retry */
2041		}
2042		ret |= VM_FAULT_LOCKED;
2043	} else
2044		VM_BUG_ON_PAGE(!PageLocked(page), page);
2045	return ret;
2046}
2047
2048/*
2049 * Handle write page faults for pages that can be reused in the current vma
2050 *
2051 * This can happen either due to the mapping being with the VM_SHARED flag,
2052 * or due to us being the last reference standing to the page. In either
2053 * case, all we need to do here is to mark the page as writable and update
2054 * any related book-keeping.
2055 */
2056static inline int wp_page_reuse(struct mm_struct *mm,
2057			struct vm_area_struct *vma, unsigned long address,
2058			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2059			struct page *page, int page_mkwrite,
2060			int dirty_shared)
2061	__releases(ptl)
2062{
 
2063	pte_t entry;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2064	/*
2065	 * Clear the pages cpupid information as the existing
2066	 * information potentially belongs to a now completely
2067	 * unrelated process.
2068	 */
2069	if (page)
2070		page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2071
2072	flush_cache_page(vma, address, pte_pfn(orig_pte));
2073	entry = pte_mkyoung(orig_pte);
2074	entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2075	if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2076		update_mmu_cache(vma, address, page_table);
2077	pte_unmap_unlock(page_table, ptl);
 
 
 
 
2078
2079	if (dirty_shared) {
2080		struct address_space *mapping;
2081		int dirtied;
2082
2083		if (!page_mkwrite)
2084			lock_page(page);
2085
2086		dirtied = set_page_dirty(page);
2087		VM_BUG_ON_PAGE(PageAnon(page), page);
2088		mapping = page->mapping;
2089		unlock_page(page);
2090		put_page(page);
 
 
 
2091
2092		if ((dirtied || page_mkwrite) && mapping) {
2093			/*
2094			 * Some device drivers do not set page.mapping
2095			 * but still dirty their pages
 
 
2096			 */
2097			balance_dirty_pages_ratelimited(mapping);
 
 
 
 
 
 
 
2098		}
 
 
2099
2100		if (!page_mkwrite)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2101			file_update_time(vma->vm_file);
2102	}
2103
2104	return VM_FAULT_WRITE;
2105}
2106
2107/*
2108 * Handle the case of a page which we actually need to copy to a new page.
2109 *
2110 * Called with mmap_sem locked and the old page referenced, but
2111 * without the ptl held.
2112 *
2113 * High level logic flow:
2114 *
2115 * - Allocate a page, copy the content of the old page to the new one.
2116 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2117 * - Take the PTL. If the pte changed, bail out and release the allocated page
2118 * - If the pte is still the way we remember it, update the page table and all
2119 *   relevant references. This includes dropping the reference the page-table
2120 *   held to the old page, as well as updating the rmap.
2121 * - In any case, unlock the PTL and drop the reference we took to the old page.
2122 */
2123static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2124			unsigned long address, pte_t *page_table, pmd_t *pmd,
2125			pte_t orig_pte, struct page *old_page)
2126{
2127	struct page *new_page = NULL;
2128	spinlock_t *ptl = NULL;
2129	pte_t entry;
2130	int page_copied = 0;
2131	const unsigned long mmun_start = address & PAGE_MASK;	/* For mmu_notifiers */
2132	const unsigned long mmun_end = mmun_start + PAGE_SIZE;	/* For mmu_notifiers */
2133	struct mem_cgroup *memcg;
2134
2135	if (unlikely(anon_vma_prepare(vma)))
2136		goto oom;
2137
2138	if (is_zero_pfn(pte_pfn(orig_pte))) {
2139		new_page = alloc_zeroed_user_highpage_movable(vma, address);
2140		if (!new_page)
2141			goto oom;
2142	} else {
2143		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2144		if (!new_page)
2145			goto oom;
2146		cow_user_page(new_page, old_page, address, vma);
2147	}
2148
2149	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2150		goto oom_free_new;
2151
2152	__SetPageUptodate(new_page);
2153
2154	mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 
2155
2156	/*
2157	 * Re-check the pte - we dropped the lock
2158	 */
2159	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2160	if (likely(pte_same(*page_table, orig_pte))) {
2161		if (old_page) {
2162			if (!PageAnon(old_page)) {
2163				dec_mm_counter_fast(mm,
2164						mm_counter_file(old_page));
2165				inc_mm_counter_fast(mm, MM_ANONPAGES);
2166			}
2167		} else {
2168			inc_mm_counter_fast(mm, MM_ANONPAGES);
2169		}
2170		flush_cache_page(vma, address, pte_pfn(orig_pte));
2171		entry = mk_pte(new_page, vma->vm_page_prot);
2172		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2173		/*
2174		 * Clear the pte entry and flush it first, before updating the
2175		 * pte with the new entry. This will avoid a race condition
2176		 * seen in the presence of one thread doing SMC and another
2177		 * thread doing COW.
2178		 */
2179		ptep_clear_flush_notify(vma, address, page_table);
2180		page_add_new_anon_rmap(new_page, vma, address, false);
2181		mem_cgroup_commit_charge(new_page, memcg, false, false);
2182		lru_cache_add_active_or_unevictable(new_page, vma);
2183		/*
2184		 * We call the notify macro here because, when using secondary
2185		 * mmu page tables (such as kvm shadow page tables), we want the
2186		 * new page to be mapped directly into the secondary page table.
2187		 */
2188		set_pte_at_notify(mm, address, page_table, entry);
2189		update_mmu_cache(vma, address, page_table);
2190		if (old_page) {
2191			/*
2192			 * Only after switching the pte to the new page may
2193			 * we remove the mapcount here. Otherwise another
2194			 * process may come and find the rmap count decremented
2195			 * before the pte is switched to the new page, and
2196			 * "reuse" the old page writing into it while our pte
2197			 * here still points into it and can be read by other
2198			 * threads.
2199			 *
2200			 * The critical issue is to order this
2201			 * page_remove_rmap with the ptp_clear_flush above.
2202			 * Those stores are ordered by (if nothing else,)
2203			 * the barrier present in the atomic_add_negative
2204			 * in page_remove_rmap.
2205			 *
2206			 * Then the TLB flush in ptep_clear_flush ensures that
2207			 * no process can access the old page before the
2208			 * decremented mapcount is visible. And the old page
2209			 * cannot be reused until after the decremented
2210			 * mapcount is visible. So transitively, TLBs to
2211			 * old page will be flushed before it can be reused.
2212			 */
2213			page_remove_rmap(old_page, false);
2214		}
2215
2216		/* Free the old page.. */
2217		new_page = old_page;
2218		page_copied = 1;
2219	} else {
2220		mem_cgroup_cancel_charge(new_page, memcg, false);
2221	}
2222
2223	if (new_page)
2224		put_page(new_page);
2225
2226	pte_unmap_unlock(page_table, ptl);
2227	mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2228	if (old_page) {
2229		/*
2230		 * Don't let another task, with possibly unlocked vma,
2231		 * keep the mlocked page.
2232		 */
2233		if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2234			lock_page(old_page);	/* LRU manipulation */
2235			if (PageMlocked(old_page))
2236				munlock_vma_page(old_page);
2237			unlock_page(old_page);
2238		}
2239		put_page(old_page);
2240	}
2241	return page_copied ? VM_FAULT_WRITE : 0;
2242oom_free_new:
2243	put_page(new_page);
2244oom:
2245	if (old_page)
2246		put_page(old_page);
2247	return VM_FAULT_OOM;
2248}
2249
2250/*
2251 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2252 * mapping
2253 */
2254static int wp_pfn_shared(struct mm_struct *mm,
2255			struct vm_area_struct *vma, unsigned long address,
2256			pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2257			pmd_t *pmd)
2258{
2259	if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2260		struct vm_fault vmf = {
2261			.page = NULL,
2262			.pgoff = linear_page_index(vma, address),
2263			.virtual_address = (void __user *)(address & PAGE_MASK),
2264			.flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2265		};
2266		int ret;
2267
2268		pte_unmap_unlock(page_table, ptl);
2269		ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2270		if (ret & VM_FAULT_ERROR)
2271			return ret;
2272		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2273		/*
2274		 * We might have raced with another page fault while we
2275		 * released the pte_offset_map_lock.
2276		 */
2277		if (!pte_same(*page_table, orig_pte)) {
2278			pte_unmap_unlock(page_table, ptl);
2279			return 0;
2280		}
2281	}
2282	return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2283			     NULL, 0, 0);
2284}
2285
2286static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2287			  unsigned long address, pte_t *page_table,
2288			  pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2289			  struct page *old_page)
2290	__releases(ptl)
2291{
2292	int page_mkwrite = 0;
2293
2294	get_page(old_page);
2295
2296	if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2297		int tmp;
2298
2299		pte_unmap_unlock(page_table, ptl);
2300		tmp = do_page_mkwrite(vma, old_page, address);
2301		if (unlikely(!tmp || (tmp &
2302				      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2303			put_page(old_page);
2304			return tmp;
2305		}
2306		/*
2307		 * Since we dropped the lock we need to revalidate
2308		 * the PTE as someone else may have changed it.  If
2309		 * they did, we just return, as we can count on the
2310		 * MMU to tell us if they didn't also make it writable.
2311		 */
2312		page_table = pte_offset_map_lock(mm, pmd, address,
2313						 &ptl);
2314		if (!pte_same(*page_table, orig_pte)) {
2315			unlock_page(old_page);
2316			pte_unmap_unlock(page_table, ptl);
2317			put_page(old_page);
2318			return 0;
2319		}
2320		page_mkwrite = 1;
2321	}
2322
2323	return wp_page_reuse(mm, vma, address, page_table, ptl,
2324			     orig_pte, old_page, page_mkwrite, 1);
2325}
2326
2327/*
2328 * This routine handles present pages, when users try to write
2329 * to a shared page. It is done by copying the page to a new address
2330 * and decrementing the shared-page counter for the old page.
2331 *
2332 * Note that this routine assumes that the protection checks have been
2333 * done by the caller (the low-level page fault routine in most cases).
2334 * Thus we can safely just mark it writable once we've done any necessary
2335 * COW.
2336 *
2337 * We also mark the page dirty at this point even though the page will
2338 * change only once the write actually happens. This avoids a few races,
2339 * and potentially makes it more efficient.
2340 *
2341 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2342 * but allow concurrent faults), with pte both mapped and locked.
2343 * We return with mmap_sem still held, but pte unmapped and unlocked.
2344 */
2345static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2346		unsigned long address, pte_t *page_table, pmd_t *pmd,
2347		spinlock_t *ptl, pte_t orig_pte)
2348	__releases(ptl)
2349{
2350	struct page *old_page;
2351
2352	old_page = vm_normal_page(vma, address, orig_pte);
2353	if (!old_page) {
2354		/*
2355		 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2356		 * VM_PFNMAP VMA.
2357		 *
2358		 * We should not cow pages in a shared writeable mapping.
2359		 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2360		 */
2361		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2362				     (VM_WRITE|VM_SHARED))
2363			return wp_pfn_shared(mm, vma, address, page_table, ptl,
2364					     orig_pte, pmd);
2365
2366		pte_unmap_unlock(page_table, ptl);
2367		return wp_page_copy(mm, vma, address, page_table, pmd,
2368				    orig_pte, old_page);
2369	}
2370
2371	/*
2372	 * Take out anonymous pages first, anonymous shared vmas are
2373	 * not dirty accountable.
2374	 */
2375	if (PageAnon(old_page) && !PageKsm(old_page)) {
2376		int total_mapcount;
2377		if (!trylock_page(old_page)) {
2378			get_page(old_page);
2379			pte_unmap_unlock(page_table, ptl);
2380			lock_page(old_page);
2381			page_table = pte_offset_map_lock(mm, pmd, address,
2382							 &ptl);
2383			if (!pte_same(*page_table, orig_pte)) {
2384				unlock_page(old_page);
2385				pte_unmap_unlock(page_table, ptl);
2386				put_page(old_page);
2387				return 0;
2388			}
2389			put_page(old_page);
2390		}
2391		if (reuse_swap_page(old_page, &total_mapcount)) {
2392			if (total_mapcount == 1) {
2393				/*
2394				 * The page is all ours. Move it to
2395				 * our anon_vma so the rmap code will
2396				 * not search our parent or siblings.
2397				 * Protected against the rmap code by
2398				 * the page lock.
2399				 */
2400				page_move_anon_rmap(compound_head(old_page),
2401						    vma, address);
2402			}
2403			unlock_page(old_page);
2404			return wp_page_reuse(mm, vma, address, page_table, ptl,
2405					     orig_pte, old_page, 0, 0);
2406		}
2407		unlock_page(old_page);
2408	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2409					(VM_WRITE|VM_SHARED))) {
2410		return wp_page_shared(mm, vma, address, page_table, pmd,
2411				      ptl, orig_pte, old_page);
2412	}
 
2413
2414	/*
2415	 * Ok, we need to copy. Oh, well..
2416	 */
2417	get_page(old_page);
2418
2419	pte_unmap_unlock(page_table, ptl);
2420	return wp_page_copy(mm, vma, address, page_table, pmd,
2421			    orig_pte, old_page);
2422}
2423
2424static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2425		unsigned long start_addr, unsigned long end_addr,
2426		struct zap_details *details)
2427{
2428	zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2429}
2430
2431static inline void unmap_mapping_range_tree(struct rb_root *root,
2432					    struct zap_details *details)
2433{
2434	struct vm_area_struct *vma;
 
2435	pgoff_t vba, vea, zba, zea;
2436
2437	vma_interval_tree_foreach(vma, root,
2438			details->first_index, details->last_index) {
2439
2440		vba = vma->vm_pgoff;
2441		vea = vba + vma_pages(vma) - 1;
 
2442		zba = details->first_index;
2443		if (zba < vba)
2444			zba = vba;
2445		zea = details->last_index;
2446		if (zea > vea)
2447			zea = vea;
2448
2449		unmap_mapping_range_vma(vma,
2450			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2451			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2452				details);
2453	}
2454}
2455
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2456/**
2457 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2458 * address_space corresponding to the specified page range in the underlying
2459 * file.
2460 *
2461 * @mapping: the address space containing mmaps to be unmapped.
2462 * @holebegin: byte in first page to unmap, relative to the start of
2463 * the underlying file.  This will be rounded down to a PAGE_SIZE
2464 * boundary.  Note that this is different from truncate_pagecache(), which
2465 * must keep the partial page.  In contrast, we must get rid of
2466 * partial pages.
2467 * @holelen: size of prospective hole in bytes.  This will be rounded
2468 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2469 * end of the file.
2470 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2471 * but 0 when invalidating pagecache, don't throw away private data.
2472 */
2473void unmap_mapping_range(struct address_space *mapping,
2474		loff_t const holebegin, loff_t const holelen, int even_cows)
2475{
2476	struct zap_details details = { };
2477	pgoff_t hba = holebegin >> PAGE_SHIFT;
2478	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2479
2480	/* Check for overflow. */
2481	if (sizeof(holelen) > sizeof(hlen)) {
2482		long long holeend =
2483			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2484		if (holeend & ~(long long)ULONG_MAX)
2485			hlen = ULONG_MAX - hba + 1;
2486	}
2487
2488	details.check_mapping = even_cows? NULL: mapping;
 
2489	details.first_index = hba;
2490	details.last_index = hba + hlen - 1;
2491	if (details.last_index < details.first_index)
2492		details.last_index = ULONG_MAX;
2493
2494
2495	/* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2496	i_mmap_lock_write(mapping);
2497	if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2498		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2499	i_mmap_unlock_write(mapping);
 
 
2500}
2501EXPORT_SYMBOL(unmap_mapping_range);
2502
2503/*
2504 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2505 * but allow concurrent faults), and pte mapped but not yet locked.
2506 * We return with pte unmapped and unlocked.
2507 *
2508 * We return with the mmap_sem locked or unlocked in the same cases
2509 * as does filemap_fault().
2510 */
2511static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2512		unsigned long address, pte_t *page_table, pmd_t *pmd,
2513		unsigned int flags, pte_t orig_pte)
2514{
2515	spinlock_t *ptl;
2516	struct page *page, *swapcache;
2517	struct mem_cgroup *memcg;
2518	swp_entry_t entry;
2519	pte_t pte;
2520	int locked;
 
2521	int exclusive = 0;
2522	int ret = 0;
2523
2524	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2525		goto out;
2526
2527	entry = pte_to_swp_entry(orig_pte);
2528	if (unlikely(non_swap_entry(entry))) {
2529		if (is_migration_entry(entry)) {
2530			migration_entry_wait(mm, pmd, address);
2531		} else if (is_hwpoison_entry(entry)) {
2532			ret = VM_FAULT_HWPOISON;
2533		} else {
2534			print_bad_pte(vma, address, orig_pte, NULL);
2535			ret = VM_FAULT_SIGBUS;
2536		}
2537		goto out;
2538	}
2539	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2540	page = lookup_swap_cache(entry);
2541	if (!page) {
 
2542		page = swapin_readahead(entry,
2543					GFP_HIGHUSER_MOVABLE, vma, address);
2544		if (!page) {
2545			/*
2546			 * Back out if somebody else faulted in this pte
2547			 * while we released the pte lock.
2548			 */
2549			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2550			if (likely(pte_same(*page_table, orig_pte)))
2551				ret = VM_FAULT_OOM;
2552			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2553			goto unlock;
2554		}
2555
2556		/* Had to read the page from swap area: Major fault */
2557		ret = VM_FAULT_MAJOR;
2558		count_vm_event(PGMAJFAULT);
2559		mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2560	} else if (PageHWPoison(page)) {
2561		/*
2562		 * hwpoisoned dirty swapcache pages are kept for killing
2563		 * owner processes (which may be unknown at hwpoison time)
2564		 */
2565		ret = VM_FAULT_HWPOISON;
2566		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2567		swapcache = page;
2568		goto out_release;
2569	}
2570
2571	swapcache = page;
2572	locked = lock_page_or_retry(page, mm, flags);
2573
2574	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2575	if (!locked) {
2576		ret |= VM_FAULT_RETRY;
2577		goto out_release;
2578	}
2579
2580	/*
2581	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2582	 * release the swapcache from under us.  The page pin, and pte_same
2583	 * test below, are not enough to exclude that.  Even if it is still
2584	 * swapcache, we need to check that the page's swap has not changed.
2585	 */
2586	if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2587		goto out_page;
2588
2589	page = ksm_might_need_to_copy(page, vma, address);
2590	if (unlikely(!page)) {
2591		ret = VM_FAULT_OOM;
2592		page = swapcache;
2593		goto out_page;
 
 
 
 
 
2594	}
2595
2596	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2597		ret = VM_FAULT_OOM;
2598		goto out_page;
2599	}
2600
2601	/*
2602	 * Back out if somebody else already faulted in this pte.
2603	 */
2604	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2605	if (unlikely(!pte_same(*page_table, orig_pte)))
2606		goto out_nomap;
2607
2608	if (unlikely(!PageUptodate(page))) {
2609		ret = VM_FAULT_SIGBUS;
2610		goto out_nomap;
2611	}
2612
2613	/*
2614	 * The page isn't present yet, go ahead with the fault.
2615	 *
2616	 * Be careful about the sequence of operations here.
2617	 * To get its accounting right, reuse_swap_page() must be called
2618	 * while the page is counted on swap but not yet in mapcount i.e.
2619	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2620	 * must be called after the swap_free(), or it will never succeed.
 
 
 
 
2621	 */
2622
2623	inc_mm_counter_fast(mm, MM_ANONPAGES);
2624	dec_mm_counter_fast(mm, MM_SWAPENTS);
2625	pte = mk_pte(page, vma->vm_page_prot);
2626	if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2627		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2628		flags &= ~FAULT_FLAG_WRITE;
2629		ret |= VM_FAULT_WRITE;
2630		exclusive = RMAP_EXCLUSIVE;
2631	}
2632	flush_icache_page(vma, page);
2633	if (pte_swp_soft_dirty(orig_pte))
2634		pte = pte_mksoft_dirty(pte);
2635	set_pte_at(mm, address, page_table, pte);
2636	if (page == swapcache) {
2637		do_page_add_anon_rmap(page, vma, address, exclusive);
2638		mem_cgroup_commit_charge(page, memcg, true, false);
2639	} else { /* ksm created a completely new copy */
2640		page_add_new_anon_rmap(page, vma, address, false);
2641		mem_cgroup_commit_charge(page, memcg, false, false);
2642		lru_cache_add_active_or_unevictable(page, vma);
2643	}
2644
2645	swap_free(entry);
2646	if (mem_cgroup_swap_full(page) ||
2647	    (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2648		try_to_free_swap(page);
2649	unlock_page(page);
2650	if (page != swapcache) {
2651		/*
2652		 * Hold the lock to avoid the swap entry to be reused
2653		 * until we take the PT lock for the pte_same() check
2654		 * (to avoid false positives from pte_same). For
2655		 * further safety release the lock after the swap_free
2656		 * so that the swap count won't change under a
2657		 * parallel locked swapcache.
2658		 */
2659		unlock_page(swapcache);
2660		put_page(swapcache);
2661	}
2662
2663	if (flags & FAULT_FLAG_WRITE) {
2664		ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2665		if (ret & VM_FAULT_ERROR)
2666			ret &= VM_FAULT_ERROR;
2667		goto out;
2668	}
2669
2670	/* No need to invalidate - it was non-present before */
2671	update_mmu_cache(vma, address, page_table);
2672unlock:
2673	pte_unmap_unlock(page_table, ptl);
2674out:
2675	return ret;
2676out_nomap:
2677	mem_cgroup_cancel_charge(page, memcg, false);
2678	pte_unmap_unlock(page_table, ptl);
2679out_page:
2680	unlock_page(page);
2681out_release:
2682	put_page(page);
2683	if (page != swapcache) {
2684		unlock_page(swapcache);
2685		put_page(swapcache);
2686	}
2687	return ret;
2688}
2689
2690/*
2691 * This is like a special single-page "expand_{down|up}wards()",
2692 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2693 * doesn't hit another vma.
2694 */
2695static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2696{
2697	address &= PAGE_MASK;
2698	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2699		struct vm_area_struct *prev = vma->vm_prev;
2700
2701		/*
2702		 * Is there a mapping abutting this one below?
2703		 *
2704		 * That's only ok if it's the same stack mapping
2705		 * that has gotten split..
2706		 */
2707		if (prev && prev->vm_end == address)
2708			return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2709
2710		return expand_downwards(vma, address - PAGE_SIZE);
2711	}
2712	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2713		struct vm_area_struct *next = vma->vm_next;
2714
2715		/* As VM_GROWSDOWN but s/below/above/ */
2716		if (next && next->vm_start == address + PAGE_SIZE)
2717			return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2718
2719		return expand_upwards(vma, address + PAGE_SIZE);
2720	}
2721	return 0;
2722}
2723
2724/*
2725 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2726 * but allow concurrent faults), and pte mapped but not yet locked.
2727 * We return with mmap_sem still held, but pte unmapped and unlocked.
2728 */
2729static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2730		unsigned long address, pte_t *page_table, pmd_t *pmd,
2731		unsigned int flags)
2732{
2733	struct mem_cgroup *memcg;
2734	struct page *page;
2735	spinlock_t *ptl;
2736	pte_t entry;
2737
2738	pte_unmap(page_table);
2739
2740	/* File mapping without ->vm_ops ? */
2741	if (vma->vm_flags & VM_SHARED)
2742		return VM_FAULT_SIGBUS;
2743
2744	/* Check if we need to add a guard page to the stack */
2745	if (check_stack_guard_page(vma, address) < 0)
2746		return VM_FAULT_SIGSEGV;
2747
2748	/* Use the zero-page for reads */
2749	if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2750		entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2751						vma->vm_page_prot));
2752		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2753		if (!pte_none(*page_table))
2754			goto unlock;
2755		/* Deliver the page fault to userland, check inside PT lock */
2756		if (userfaultfd_missing(vma)) {
2757			pte_unmap_unlock(page_table, ptl);
2758			return handle_userfault(vma, address, flags,
2759						VM_UFFD_MISSING);
2760		}
2761		goto setpte;
2762	}
2763
2764	/* Allocate our own private page. */
2765	if (unlikely(anon_vma_prepare(vma)))
2766		goto oom;
2767	page = alloc_zeroed_user_highpage_movable(vma, address);
2768	if (!page)
2769		goto oom;
 
2770
2771	if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2772		goto oom_free_page;
2773
2774	/*
2775	 * The memory barrier inside __SetPageUptodate makes sure that
2776	 * preceeding stores to the page contents become visible before
2777	 * the set_pte_at() write.
2778	 */
2779	__SetPageUptodate(page);
2780
2781	entry = mk_pte(page, vma->vm_page_prot);
2782	if (vma->vm_flags & VM_WRITE)
2783		entry = pte_mkwrite(pte_mkdirty(entry));
2784
2785	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2786	if (!pte_none(*page_table))
2787		goto release;
2788
2789	/* Deliver the page fault to userland, check inside PT lock */
2790	if (userfaultfd_missing(vma)) {
2791		pte_unmap_unlock(page_table, ptl);
2792		mem_cgroup_cancel_charge(page, memcg, false);
2793		put_page(page);
2794		return handle_userfault(vma, address, flags,
2795					VM_UFFD_MISSING);
2796	}
2797
2798	inc_mm_counter_fast(mm, MM_ANONPAGES);
2799	page_add_new_anon_rmap(page, vma, address, false);
2800	mem_cgroup_commit_charge(page, memcg, false, false);
2801	lru_cache_add_active_or_unevictable(page, vma);
2802setpte:
2803	set_pte_at(mm, address, page_table, entry);
2804
2805	/* No need to invalidate - it was non-present before */
2806	update_mmu_cache(vma, address, page_table);
2807unlock:
2808	pte_unmap_unlock(page_table, ptl);
2809	return 0;
2810release:
2811	mem_cgroup_cancel_charge(page, memcg, false);
2812	put_page(page);
2813	goto unlock;
2814oom_free_page:
2815	put_page(page);
2816oom:
2817	return VM_FAULT_OOM;
2818}
2819
2820/*
2821 * The mmap_sem must have been held on entry, and may have been
2822 * released depending on flags and vma->vm_ops->fault() return value.
2823 * See filemap_fault() and __lock_page_retry().
2824 */
2825static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2826			pgoff_t pgoff, unsigned int flags,
2827			struct page *cow_page, struct page **page)
 
 
 
 
 
 
 
 
2828{
 
 
 
 
 
 
 
2829	struct vm_fault vmf;
2830	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2831
2832	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2833	vmf.pgoff = pgoff;
2834	vmf.flags = flags;
2835	vmf.page = NULL;
2836	vmf.gfp_mask = __get_fault_gfp_mask(vma);
2837	vmf.cow_page = cow_page;
2838
2839	ret = vma->vm_ops->fault(vma, &vmf);
2840	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2841		return ret;
2842	if (!vmf.page)
2843		goto out;
2844
2845	if (unlikely(PageHWPoison(vmf.page))) {
2846		if (ret & VM_FAULT_LOCKED)
2847			unlock_page(vmf.page);
2848		put_page(vmf.page);
2849		return VM_FAULT_HWPOISON;
2850	}
2851
 
 
 
 
2852	if (unlikely(!(ret & VM_FAULT_LOCKED)))
2853		lock_page(vmf.page);
2854	else
2855		VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2856
2857 out:
2858	*page = vmf.page;
2859	return ret;
2860}
2861
2862/**
2863 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2864 *
2865 * @vma: virtual memory area
2866 * @address: user virtual address
2867 * @page: page to map
2868 * @pte: pointer to target page table entry
2869 * @write: true, if new entry is writable
2870 * @anon: true, if it's anonymous page
2871 *
2872 * Caller must hold page table lock relevant for @pte.
2873 *
2874 * Target users are page handler itself and implementations of
2875 * vm_ops->map_pages.
2876 */
2877void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2878		struct page *page, pte_t *pte, bool write, bool anon)
2879{
2880	pte_t entry;
2881
2882	flush_icache_page(vma, page);
2883	entry = mk_pte(page, vma->vm_page_prot);
2884	if (write)
2885		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2886	if (anon) {
2887		inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2888		page_add_new_anon_rmap(page, vma, address, false);
2889	} else {
2890		inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
2891		page_add_file_rmap(page);
2892	}
2893	set_pte_at(vma->vm_mm, address, pte, entry);
2894
2895	/* no need to invalidate: a not-present page won't be cached */
2896	update_mmu_cache(vma, address, pte);
2897}
2898
2899static unsigned long fault_around_bytes __read_mostly =
2900	rounddown_pow_of_two(65536);
2901
2902#ifdef CONFIG_DEBUG_FS
2903static int fault_around_bytes_get(void *data, u64 *val)
2904{
2905	*val = fault_around_bytes;
2906	return 0;
2907}
2908
2909/*
2910 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2911 * rounded down to nearest page order. It's what do_fault_around() expects to
2912 * see.
2913 */
2914static int fault_around_bytes_set(void *data, u64 val)
2915{
2916	if (val / PAGE_SIZE > PTRS_PER_PTE)
2917		return -EINVAL;
2918	if (val > PAGE_SIZE)
2919		fault_around_bytes = rounddown_pow_of_two(val);
2920	else
2921		fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2922	return 0;
2923}
2924DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2925		fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2926
2927static int __init fault_around_debugfs(void)
2928{
2929	void *ret;
2930
2931	ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2932			&fault_around_bytes_fops);
2933	if (!ret)
2934		pr_warn("Failed to create fault_around_bytes in debugfs");
2935	return 0;
2936}
2937late_initcall(fault_around_debugfs);
2938#endif
2939
2940/*
2941 * do_fault_around() tries to map few pages around the fault address. The hope
2942 * is that the pages will be needed soon and this will lower the number of
2943 * faults to handle.
2944 *
2945 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2946 * not ready to be mapped: not up-to-date, locked, etc.
2947 *
2948 * This function is called with the page table lock taken. In the split ptlock
2949 * case the page table lock only protects only those entries which belong to
2950 * the page table corresponding to the fault address.
2951 *
2952 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2953 * only once.
2954 *
2955 * fault_around_pages() defines how many pages we'll try to map.
2956 * do_fault_around() expects it to return a power of two less than or equal to
2957 * PTRS_PER_PTE.
2958 *
2959 * The virtual address of the area that we map is naturally aligned to the
2960 * fault_around_pages() value (and therefore to page order).  This way it's
2961 * easier to guarantee that we don't cross page table boundaries.
2962 */
2963static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2964		pte_t *pte, pgoff_t pgoff, unsigned int flags)
2965{
2966	unsigned long start_addr, nr_pages, mask;
2967	pgoff_t max_pgoff;
2968	struct vm_fault vmf;
2969	int off;
2970
2971	nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2972	mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2973
2974	start_addr = max(address & mask, vma->vm_start);
2975	off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2976	pte -= off;
2977	pgoff -= off;
2978
2979	/*
2980	 *  max_pgoff is either end of page table or end of vma
2981	 *  or fault_around_pages() from pgoff, depending what is nearest.
2982	 */
2983	max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2984		PTRS_PER_PTE - 1;
2985	max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2986			pgoff + nr_pages - 1);
2987
2988	/* Check if it makes any sense to call ->map_pages */
2989	while (!pte_none(*pte)) {
2990		if (++pgoff > max_pgoff)
2991			return;
2992		start_addr += PAGE_SIZE;
2993		if (start_addr >= vma->vm_end)
2994			return;
2995		pte++;
2996	}
2997
2998	vmf.virtual_address = (void __user *) start_addr;
2999	vmf.pte = pte;
3000	vmf.pgoff = pgoff;
3001	vmf.max_pgoff = max_pgoff;
3002	vmf.flags = flags;
3003	vmf.gfp_mask = __get_fault_gfp_mask(vma);
3004	vma->vm_ops->map_pages(vma, &vmf);
3005}
3006
3007static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3008		unsigned long address, pmd_t *pmd,
3009		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3010{
3011	struct page *fault_page;
3012	spinlock_t *ptl;
3013	pte_t *pte;
3014	int ret = 0;
3015
3016	/*
3017	 * Let's call ->map_pages() first and use ->fault() as fallback
3018	 * if page by the offset is not ready to be mapped (cold cache or
3019	 * something).
3020	 */
3021	if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3022		pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3023		do_fault_around(vma, address, pte, pgoff, flags);
3024		if (!pte_same(*pte, orig_pte))
3025			goto unlock_out;
3026		pte_unmap_unlock(pte, ptl);
3027	}
3028
3029	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3030	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3031		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3032
3033	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3034	if (unlikely(!pte_same(*pte, orig_pte))) {
3035		pte_unmap_unlock(pte, ptl);
3036		unlock_page(fault_page);
3037		put_page(fault_page);
3038		return ret;
3039	}
3040	do_set_pte(vma, address, fault_page, pte, false, false);
3041	unlock_page(fault_page);
3042unlock_out:
3043	pte_unmap_unlock(pte, ptl);
3044	return ret;
3045}
3046
3047static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3048		unsigned long address, pmd_t *pmd,
3049		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3050{
3051	struct page *fault_page, *new_page;
3052	struct mem_cgroup *memcg;
3053	spinlock_t *ptl;
3054	pte_t *pte;
3055	int ret;
3056
3057	if (unlikely(anon_vma_prepare(vma)))
3058		return VM_FAULT_OOM;
3059
3060	new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3061	if (!new_page)
3062		return VM_FAULT_OOM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3063
3064	if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
3065		put_page(new_page);
3066		return VM_FAULT_OOM;
 
 
 
 
 
 
3067	}
3068
3069	ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3070	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3071		goto uncharge_out;
3072
3073	if (fault_page)
3074		copy_user_highpage(new_page, fault_page, address, vma);
3075	__SetPageUptodate(new_page);
3076
3077	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3078	if (unlikely(!pte_same(*pte, orig_pte))) {
3079		pte_unmap_unlock(pte, ptl);
3080		if (fault_page) {
3081			unlock_page(fault_page);
3082			put_page(fault_page);
3083		} else {
3084			/*
3085			 * The fault handler has no page to lock, so it holds
3086			 * i_mmap_lock for read to protect against truncate.
3087			 */
3088			i_mmap_unlock_read(vma->vm_file->f_mapping);
3089		}
3090		goto uncharge_out;
3091	}
3092	do_set_pte(vma, address, new_page, pte, true, true);
3093	mem_cgroup_commit_charge(new_page, memcg, false, false);
3094	lru_cache_add_active_or_unevictable(new_page, vma);
3095	pte_unmap_unlock(pte, ptl);
3096	if (fault_page) {
3097		unlock_page(fault_page);
3098		put_page(fault_page);
3099	} else {
3100		/*
3101		 * The fault handler has no page to lock, so it holds
3102		 * i_mmap_lock for read to protect against truncate.
3103		 */
3104		i_mmap_unlock_read(vma->vm_file->f_mapping);
3105	}
 
 
 
 
 
3106	return ret;
3107uncharge_out:
3108	mem_cgroup_cancel_charge(new_page, memcg, false);
3109	put_page(new_page);
 
 
 
3110	return ret;
3111}
3112
3113static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3114		unsigned long address, pmd_t *pmd,
3115		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3116{
3117	struct page *fault_page;
3118	struct address_space *mapping;
3119	spinlock_t *ptl;
3120	pte_t *pte;
3121	int dirtied = 0;
3122	int ret, tmp;
3123
3124	ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3125	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3126		return ret;
3127
3128	/*
3129	 * Check if the backing address space wants to know that the page is
3130	 * about to become writable
3131	 */
3132	if (vma->vm_ops->page_mkwrite) {
3133		unlock_page(fault_page);
3134		tmp = do_page_mkwrite(vma, fault_page, address);
3135		if (unlikely(!tmp ||
3136				(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3137			put_page(fault_page);
3138			return tmp;
3139		}
3140	}
3141
3142	pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3143	if (unlikely(!pte_same(*pte, orig_pte))) {
3144		pte_unmap_unlock(pte, ptl);
3145		unlock_page(fault_page);
3146		put_page(fault_page);
3147		return ret;
3148	}
3149	do_set_pte(vma, address, fault_page, pte, true, false);
3150	pte_unmap_unlock(pte, ptl);
3151
3152	if (set_page_dirty(fault_page))
3153		dirtied = 1;
3154	/*
3155	 * Take a local copy of the address_space - page.mapping may be zeroed
3156	 * by truncate after unlock_page().   The address_space itself remains
3157	 * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
3158	 * release semantics to prevent the compiler from undoing this copying.
3159	 */
3160	mapping = page_rmapping(fault_page);
3161	unlock_page(fault_page);
3162	if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3163		/*
3164		 * Some device drivers do not set page.mapping but still
3165		 * dirty their pages
3166		 */
3167		balance_dirty_pages_ratelimited(mapping);
3168	}
3169
3170	if (!vma->vm_ops->page_mkwrite)
3171		file_update_time(vma->vm_file);
3172
3173	return ret;
3174}
3175
3176/*
 
 
 
 
3177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3178 * but allow concurrent faults).
3179 * The mmap_sem may have been released depending on flags and our
3180 * return value.  See filemap_fault() and __lock_page_or_retry().
3181 */
3182static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3183		unsigned long address, pte_t *page_table, pmd_t *pmd,
3184		unsigned int flags, pte_t orig_pte)
3185{
3186	pgoff_t pgoff = linear_page_index(vma, address);
3187
3188	pte_unmap(page_table);
3189	/* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3190	if (!vma->vm_ops->fault)
3191		return VM_FAULT_SIGBUS;
3192	if (!(flags & FAULT_FLAG_WRITE))
3193		return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3194				orig_pte);
3195	if (!(vma->vm_flags & VM_SHARED))
3196		return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3197				orig_pte);
3198	return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3199}
3200
3201static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3202				unsigned long addr, int page_nid,
3203				int *flags)
3204{
3205	get_page(page);
3206
3207	count_vm_numa_event(NUMA_HINT_FAULTS);
3208	if (page_nid == numa_node_id()) {
3209		count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3210		*flags |= TNF_FAULT_LOCAL;
3211	}
3212
3213	return mpol_misplaced(page, vma, addr);
3214}
3215
3216static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3217		   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3218{
3219	struct page *page = NULL;
3220	spinlock_t *ptl;
3221	int page_nid = -1;
3222	int last_cpupid;
3223	int target_nid;
3224	bool migrated = false;
3225	bool was_writable = pte_write(pte);
3226	int flags = 0;
3227
3228	/* A PROT_NONE fault should not end up here */
3229	BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3230
3231	/*
3232	* The "pte" at this point cannot be used safely without
3233	* validation through pte_unmap_same(). It's of NUMA type but
3234	* the pfn may be screwed if the read is non atomic.
3235	*
3236	* We can safely just do a "set_pte_at()", because the old
3237	* page table entry is not accessible, so there would be no
3238	* concurrent hardware modifications to the PTE.
3239	*/
3240	ptl = pte_lockptr(mm, pmd);
3241	spin_lock(ptl);
3242	if (unlikely(!pte_same(*ptep, pte))) {
3243		pte_unmap_unlock(ptep, ptl);
3244		goto out;
3245	}
3246
3247	/* Make it present again */
3248	pte = pte_modify(pte, vma->vm_page_prot);
3249	pte = pte_mkyoung(pte);
3250	if (was_writable)
3251		pte = pte_mkwrite(pte);
3252	set_pte_at(mm, addr, ptep, pte);
3253	update_mmu_cache(vma, addr, ptep);
3254
3255	page = vm_normal_page(vma, addr, pte);
3256	if (!page) {
3257		pte_unmap_unlock(ptep, ptl);
3258		return 0;
3259	}
3260
3261	/* TODO: handle PTE-mapped THP */
3262	if (PageCompound(page)) {
3263		pte_unmap_unlock(ptep, ptl);
3264		return 0;
3265	}
3266
3267	/*
3268	 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3269	 * much anyway since they can be in shared cache state. This misses
3270	 * the case where a mapping is writable but the process never writes
3271	 * to it but pte_write gets cleared during protection updates and
3272	 * pte_dirty has unpredictable behaviour between PTE scan updates,
3273	 * background writeback, dirty balancing and application behaviour.
3274	 */
3275	if (!(vma->vm_flags & VM_WRITE))
3276		flags |= TNF_NO_GROUP;
3277
3278	/*
3279	 * Flag if the page is shared between multiple address spaces. This
3280	 * is later used when determining whether to group tasks together
3281	 */
3282	if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3283		flags |= TNF_SHARED;
3284
3285	last_cpupid = page_cpupid_last(page);
3286	page_nid = page_to_nid(page);
3287	target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3288	pte_unmap_unlock(ptep, ptl);
3289	if (target_nid == -1) {
3290		put_page(page);
3291		goto out;
3292	}
3293
3294	/* Migrate to the requested node */
3295	migrated = migrate_misplaced_page(page, vma, target_nid);
3296	if (migrated) {
3297		page_nid = target_nid;
3298		flags |= TNF_MIGRATED;
3299	} else
3300		flags |= TNF_MIGRATE_FAIL;
3301
3302out:
3303	if (page_nid != -1)
3304		task_numa_fault(last_cpupid, page_nid, 1, flags);
3305	return 0;
3306}
3307
3308static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3309			unsigned long address, pmd_t *pmd, unsigned int flags)
3310{
3311	if (vma_is_anonymous(vma))
3312		return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3313	if (vma->vm_ops->pmd_fault)
3314		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3315	return VM_FAULT_FALLBACK;
3316}
3317
3318static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3319			unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3320			unsigned int flags)
3321{
3322	if (vma_is_anonymous(vma))
3323		return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3324	if (vma->vm_ops->pmd_fault)
3325		return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3326	return VM_FAULT_FALLBACK;
3327}
3328
3329/*
3330 * These routines also need to handle stuff like marking pages dirty
3331 * and/or accessed for architectures that don't do it in hardware (most
3332 * RISC architectures).  The early dirtying is also good on the i386.
3333 *
3334 * There is also a hook called "update_mmu_cache()" that architectures
3335 * with external mmu caches can use to update those (ie the Sparc or
3336 * PowerPC hashed page tables that act as extended TLBs).
3337 *
3338 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3339 * but allow concurrent faults), and pte mapped but not yet locked.
3340 * We return with pte unmapped and unlocked.
3341 *
3342 * The mmap_sem may have been released depending on flags and our
3343 * return value.  See filemap_fault() and __lock_page_or_retry().
3344 */
3345static int handle_pte_fault(struct mm_struct *mm,
3346		     struct vm_area_struct *vma, unsigned long address,
3347		     pte_t *pte, pmd_t *pmd, unsigned int flags)
3348{
3349	pte_t entry;
3350	spinlock_t *ptl;
3351
3352	/*
3353	 * some architectures can have larger ptes than wordsize,
3354	 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3355	 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3356	 * The code below just needs a consistent view for the ifs and
3357	 * we later double check anyway with the ptl lock held. So here
3358	 * a barrier will do.
3359	 */
3360	entry = *pte;
3361	barrier();
3362	if (!pte_present(entry)) {
3363		if (pte_none(entry)) {
3364			if (vma_is_anonymous(vma))
3365				return do_anonymous_page(mm, vma, address,
3366							 pte, pmd, flags);
3367			else
3368				return do_fault(mm, vma, address, pte, pmd,
3369						flags, entry);
 
3370		}
 
 
 
3371		return do_swap_page(mm, vma, address,
3372					pte, pmd, flags, entry);
3373	}
3374
3375	if (pte_protnone(entry))
3376		return do_numa_page(mm, vma, address, entry, pte, pmd);
3377
3378	ptl = pte_lockptr(mm, pmd);
3379	spin_lock(ptl);
3380	if (unlikely(!pte_same(*pte, entry)))
3381		goto unlock;
3382	if (flags & FAULT_FLAG_WRITE) {
3383		if (!pte_write(entry))
3384			return do_wp_page(mm, vma, address,
3385					pte, pmd, ptl, entry);
3386		entry = pte_mkdirty(entry);
3387	}
3388	entry = pte_mkyoung(entry);
3389	if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3390		update_mmu_cache(vma, address, pte);
3391	} else {
3392		/*
3393		 * This is needed only for protection faults but the arch code
3394		 * is not yet telling us if this is a protection fault or not.
3395		 * This still avoids useless tlb flushes for .text page faults
3396		 * with threads.
3397		 */
3398		if (flags & FAULT_FLAG_WRITE)
3399			flush_tlb_fix_spurious_fault(vma, address);
3400	}
3401unlock:
3402	pte_unmap_unlock(pte, ptl);
3403	return 0;
3404}
3405
3406/*
3407 * By the time we get here, we already hold the mm semaphore
3408 *
3409 * The mmap_sem may have been released depending on flags and our
3410 * return value.  See filemap_fault() and __lock_page_or_retry().
3411 */
3412static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3413			     unsigned long address, unsigned int flags)
3414{
3415	pgd_t *pgd;
3416	pud_t *pud;
3417	pmd_t *pmd;
3418	pte_t *pte;
3419
3420	if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3421					    flags & FAULT_FLAG_INSTRUCTION,
3422					    flags & FAULT_FLAG_REMOTE))
3423		return VM_FAULT_SIGSEGV;
 
 
 
3424
3425	if (unlikely(is_vm_hugetlb_page(vma)))
3426		return hugetlb_fault(mm, vma, address, flags);
3427
3428	pgd = pgd_offset(mm, address);
3429	pud = pud_alloc(mm, pgd, address);
3430	if (!pud)
3431		return VM_FAULT_OOM;
3432	pmd = pmd_alloc(mm, pud, address);
3433	if (!pmd)
3434		return VM_FAULT_OOM;
3435	if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3436		int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3437		if (!(ret & VM_FAULT_FALLBACK))
3438			return ret;
3439	} else {
3440		pmd_t orig_pmd = *pmd;
3441		int ret;
3442
3443		barrier();
3444		if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3445			unsigned int dirty = flags & FAULT_FLAG_WRITE;
3446
3447			if (pmd_protnone(orig_pmd))
3448				return do_huge_pmd_numa_page(mm, vma, address,
3449							     orig_pmd, pmd);
3450
3451			if (dirty && !pmd_write(orig_pmd)) {
3452				ret = wp_huge_pmd(mm, vma, address, pmd,
3453							orig_pmd, flags);
3454				if (!(ret & VM_FAULT_FALLBACK))
3455					return ret;
3456			} else {
3457				huge_pmd_set_accessed(mm, vma, address, pmd,
3458						      orig_pmd, dirty);
3459				return 0;
3460			}
3461		}
3462	}
3463
3464	/*
3465	 * Use pte_alloc() instead of pte_alloc_map, because we can't
3466	 * run pte_offset_map on the pmd, if an huge pmd could
3467	 * materialize from under us from a different thread.
3468	 */
3469	if (unlikely(pte_alloc(mm, pmd, address)))
3470		return VM_FAULT_OOM;
3471	/*
3472	 * If a huge pmd materialized under us just retry later.  Use
3473	 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3474	 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3475	 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3476	 * in a different thread of this mm, in turn leading to a misleading
3477	 * pmd_trans_huge() retval.  All we have to ensure is that it is a
3478	 * regular pmd that we can walk with pte_offset_map() and we can do that
3479	 * through an atomic read in C, which is what pmd_trans_unstable()
3480	 * provides.
3481	 */
3482	if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
3483		return 0;
3484	/*
3485	 * A regular pmd is established and it can't morph into a huge pmd
3486	 * from under us anymore at this point because we hold the mmap_sem
3487	 * read mode and khugepaged takes it in write mode. So now it's
3488	 * safe to run pte_offset_map().
3489	 */
3490	pte = pte_offset_map(pmd, address);
3491
3492	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3493}
3494
3495/*
3496 * By the time we get here, we already hold the mm semaphore
3497 *
3498 * The mmap_sem may have been released depending on flags and our
3499 * return value.  See filemap_fault() and __lock_page_or_retry().
3500 */
3501int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3502		    unsigned long address, unsigned int flags)
3503{
3504	int ret;
3505
3506	__set_current_state(TASK_RUNNING);
3507
3508	count_vm_event(PGFAULT);
3509	mem_cgroup_count_vm_event(mm, PGFAULT);
3510
3511	/* do counter updates before entering really critical section. */
3512	check_sync_rss_stat(current);
3513
3514	/*
3515	 * Enable the memcg OOM handling for faults triggered in user
3516	 * space.  Kernel faults are handled more gracefully.
3517	 */
3518	if (flags & FAULT_FLAG_USER)
3519		mem_cgroup_oom_enable();
3520
3521	ret = __handle_mm_fault(mm, vma, address, flags);
3522
3523	if (flags & FAULT_FLAG_USER) {
3524		mem_cgroup_oom_disable();
3525                /*
3526                 * The task may have entered a memcg OOM situation but
3527                 * if the allocation error was handled gracefully (no
3528                 * VM_FAULT_OOM), there is no need to kill anything.
3529                 * Just clean up the OOM state peacefully.
3530                 */
3531                if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3532                        mem_cgroup_oom_synchronize(false);
3533	}
3534
3535	return ret;
3536}
3537EXPORT_SYMBOL_GPL(handle_mm_fault);
3538
3539#ifndef __PAGETABLE_PUD_FOLDED
3540/*
3541 * Allocate page upper directory.
3542 * We've already handled the fast-path in-line.
3543 */
3544int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3545{
3546	pud_t *new = pud_alloc_one(mm, address);
3547	if (!new)
3548		return -ENOMEM;
3549
3550	smp_wmb(); /* See comment in __pte_alloc */
3551
3552	spin_lock(&mm->page_table_lock);
3553	if (pgd_present(*pgd))		/* Another has populated it */
3554		pud_free(mm, new);
3555	else
3556		pgd_populate(mm, pgd, new);
3557	spin_unlock(&mm->page_table_lock);
3558	return 0;
3559}
3560#endif /* __PAGETABLE_PUD_FOLDED */
3561
3562#ifndef __PAGETABLE_PMD_FOLDED
3563/*
3564 * Allocate page middle directory.
3565 * We've already handled the fast-path in-line.
3566 */
3567int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3568{
3569	pmd_t *new = pmd_alloc_one(mm, address);
3570	if (!new)
3571		return -ENOMEM;
3572
3573	smp_wmb(); /* See comment in __pte_alloc */
3574
3575	spin_lock(&mm->page_table_lock);
3576#ifndef __ARCH_HAS_4LEVEL_HACK
3577	if (!pud_present(*pud)) {
3578		mm_inc_nr_pmds(mm);
3579		pud_populate(mm, pud, new);
3580	} else	/* Another has populated it */
3581		pmd_free(mm, new);
 
 
3582#else
3583	if (!pgd_present(*pud)) {
3584		mm_inc_nr_pmds(mm);
3585		pgd_populate(mm, pud, new);
3586	} else /* Another has populated it */
3587		pmd_free(mm, new);
 
 
3588#endif /* __ARCH_HAS_4LEVEL_HACK */
3589	spin_unlock(&mm->page_table_lock);
3590	return 0;
3591}
3592#endif /* __PAGETABLE_PMD_FOLDED */
3593
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3594static int __follow_pte(struct mm_struct *mm, unsigned long address,
3595		pte_t **ptepp, spinlock_t **ptlp)
3596{
3597	pgd_t *pgd;
3598	pud_t *pud;
3599	pmd_t *pmd;
3600	pte_t *ptep;
3601
3602	pgd = pgd_offset(mm, address);
3603	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3604		goto out;
3605
3606	pud = pud_offset(pgd, address);
3607	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3608		goto out;
3609
3610	pmd = pmd_offset(pud, address);
3611	VM_BUG_ON(pmd_trans_huge(*pmd));
3612	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3613		goto out;
3614
3615	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
3616	if (pmd_huge(*pmd))
3617		goto out;
3618
3619	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3620	if (!ptep)
3621		goto out;
3622	if (!pte_present(*ptep))
3623		goto unlock;
3624	*ptepp = ptep;
3625	return 0;
3626unlock:
3627	pte_unmap_unlock(ptep, *ptlp);
3628out:
3629	return -EINVAL;
3630}
3631
3632static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3633			     pte_t **ptepp, spinlock_t **ptlp)
3634{
3635	int res;
3636
3637	/* (void) is needed to make gcc happy */
3638	(void) __cond_lock(*ptlp,
3639			   !(res = __follow_pte(mm, address, ptepp, ptlp)));
3640	return res;
3641}
3642
3643/**
3644 * follow_pfn - look up PFN at a user virtual address
3645 * @vma: memory mapping
3646 * @address: user virtual address
3647 * @pfn: location to store found PFN
3648 *
3649 * Only IO mappings and raw PFN mappings are allowed.
3650 *
3651 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3652 */
3653int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3654	unsigned long *pfn)
3655{
3656	int ret = -EINVAL;
3657	spinlock_t *ptl;
3658	pte_t *ptep;
3659
3660	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3661		return ret;
3662
3663	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3664	if (ret)
3665		return ret;
3666	*pfn = pte_pfn(*ptep);
3667	pte_unmap_unlock(ptep, ptl);
3668	return 0;
3669}
3670EXPORT_SYMBOL(follow_pfn);
3671
3672#ifdef CONFIG_HAVE_IOREMAP_PROT
3673int follow_phys(struct vm_area_struct *vma,
3674		unsigned long address, unsigned int flags,
3675		unsigned long *prot, resource_size_t *phys)
3676{
3677	int ret = -EINVAL;
3678	pte_t *ptep, pte;
3679	spinlock_t *ptl;
3680
3681	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3682		goto out;
3683
3684	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3685		goto out;
3686	pte = *ptep;
3687
3688	if ((flags & FOLL_WRITE) && !pte_write(pte))
3689		goto unlock;
3690
3691	*prot = pgprot_val(pte_pgprot(pte));
3692	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3693
3694	ret = 0;
3695unlock:
3696	pte_unmap_unlock(ptep, ptl);
3697out:
3698	return ret;
3699}
3700
3701int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3702			void *buf, int len, int write)
3703{
3704	resource_size_t phys_addr;
3705	unsigned long prot = 0;
3706	void __iomem *maddr;
3707	int offset = addr & (PAGE_SIZE-1);
3708
3709	if (follow_phys(vma, addr, write, &prot, &phys_addr))
3710		return -EINVAL;
3711
3712	maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3713	if (write)
3714		memcpy_toio(maddr + offset, buf, len);
3715	else
3716		memcpy_fromio(buf, maddr + offset, len);
3717	iounmap(maddr);
3718
3719	return len;
3720}
3721EXPORT_SYMBOL_GPL(generic_access_phys);
3722#endif
3723
3724/*
3725 * Access another process' address space as given in mm.  If non-NULL, use the
3726 * given task for page fault accounting.
3727 */
3728static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3729		unsigned long addr, void *buf, int len, int write)
3730{
3731	struct vm_area_struct *vma;
3732	void *old_buf = buf;
3733
3734	down_read(&mm->mmap_sem);
3735	/* ignore errors, just check how much was successfully transferred */
3736	while (len) {
3737		int bytes, ret, offset;
3738		void *maddr;
3739		struct page *page = NULL;
3740
3741		ret = get_user_pages_remote(tsk, mm, addr, 1,
3742				write, 1, &page, &vma);
3743		if (ret <= 0) {
3744#ifndef CONFIG_HAVE_IOREMAP_PROT
3745			break;
3746#else
3747			/*
3748			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3749			 * we can access using slightly different code.
3750			 */
 
3751			vma = find_vma(mm, addr);
3752			if (!vma || vma->vm_start > addr)
3753				break;
3754			if (vma->vm_ops && vma->vm_ops->access)
3755				ret = vma->vm_ops->access(vma, addr, buf,
3756							  len, write);
3757			if (ret <= 0)
 
3758				break;
3759			bytes = ret;
3760#endif
3761		} else {
3762			bytes = len;
3763			offset = addr & (PAGE_SIZE-1);
3764			if (bytes > PAGE_SIZE-offset)
3765				bytes = PAGE_SIZE-offset;
3766
3767			maddr = kmap(page);
3768			if (write) {
3769				copy_to_user_page(vma, page, addr,
3770						  maddr + offset, buf, bytes);
3771				set_page_dirty_lock(page);
3772			} else {
3773				copy_from_user_page(vma, page, addr,
3774						    buf, maddr + offset, bytes);
3775			}
3776			kunmap(page);
3777			put_page(page);
3778		}
3779		len -= bytes;
3780		buf += bytes;
3781		addr += bytes;
3782	}
3783	up_read(&mm->mmap_sem);
3784
3785	return buf - old_buf;
3786}
3787
3788/**
3789 * access_remote_vm - access another process' address space
3790 * @mm:		the mm_struct of the target address space
3791 * @addr:	start address to access
3792 * @buf:	source or destination buffer
3793 * @len:	number of bytes to transfer
3794 * @write:	whether the access is a write
3795 *
3796 * The caller must hold a reference on @mm.
3797 */
3798int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3799		void *buf, int len, int write)
3800{
3801	return __access_remote_vm(NULL, mm, addr, buf, len, write);
3802}
3803
3804/*
3805 * Access another process' address space.
3806 * Source/target buffer must be kernel space,
3807 * Do not walk the page table directly, use get_user_pages
3808 */
3809int access_process_vm(struct task_struct *tsk, unsigned long addr,
3810		void *buf, int len, int write)
3811{
3812	struct mm_struct *mm;
3813	int ret;
3814
3815	mm = get_task_mm(tsk);
3816	if (!mm)
3817		return 0;
3818
3819	ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3820	mmput(mm);
3821
3822	return ret;
3823}
3824
3825/*
3826 * Print the name of a VMA.
3827 */
3828void print_vma_addr(char *prefix, unsigned long ip)
3829{
3830	struct mm_struct *mm = current->mm;
3831	struct vm_area_struct *vma;
3832
3833	/*
3834	 * Do not print if we are in atomic
3835	 * contexts (in exception stacks, etc.):
3836	 */
3837	if (preempt_count())
3838		return;
3839
3840	down_read(&mm->mmap_sem);
3841	vma = find_vma(mm, ip);
3842	if (vma && vma->vm_file) {
3843		struct file *f = vma->vm_file;
3844		char *buf = (char *)__get_free_page(GFP_KERNEL);
3845		if (buf) {
3846			char *p;
3847
3848			p = file_path(f, buf, PAGE_SIZE);
3849			if (IS_ERR(p))
3850				p = "?";
3851			printk("%s%s[%lx+%lx]", prefix, kbasename(p),
 
 
 
3852					vma->vm_start,
3853					vma->vm_end - vma->vm_start);
3854			free_page((unsigned long)buf);
3855		}
3856	}
3857	up_read(&mm->mmap_sem);
3858}
3859
3860#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3861void __might_fault(const char *file, int line)
3862{
3863	/*
3864	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3865	 * holding the mmap_sem, this is safe because kernel memory doesn't
3866	 * get paged out, therefore we'll never actually fault, and the
3867	 * below annotations will generate false positives.
3868	 */
3869	if (segment_eq(get_fs(), KERNEL_DS))
3870		return;
3871	if (pagefault_disabled())
3872		return;
3873	__might_sleep(file, line, 0);
3874#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3875	if (current->mm)
 
 
 
3876		might_lock_read(&current->mm->mmap_sem);
3877#endif
3878}
3879EXPORT_SYMBOL(__might_fault);
3880#endif
3881
3882#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3883static void clear_gigantic_page(struct page *page,
3884				unsigned long addr,
3885				unsigned int pages_per_huge_page)
3886{
3887	int i;
3888	struct page *p = page;
3889
3890	might_sleep();
3891	for (i = 0; i < pages_per_huge_page;
3892	     i++, p = mem_map_next(p, page, i)) {
3893		cond_resched();
3894		clear_user_highpage(p, addr + i * PAGE_SIZE);
3895	}
3896}
3897void clear_huge_page(struct page *page,
3898		     unsigned long addr, unsigned int pages_per_huge_page)
3899{
3900	int i;
3901
3902	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3903		clear_gigantic_page(page, addr, pages_per_huge_page);
3904		return;
3905	}
3906
3907	might_sleep();
3908	for (i = 0; i < pages_per_huge_page; i++) {
3909		cond_resched();
3910		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3911	}
3912}
3913
3914static void copy_user_gigantic_page(struct page *dst, struct page *src,
3915				    unsigned long addr,
3916				    struct vm_area_struct *vma,
3917				    unsigned int pages_per_huge_page)
3918{
3919	int i;
3920	struct page *dst_base = dst;
3921	struct page *src_base = src;
3922
3923	for (i = 0; i < pages_per_huge_page; ) {
3924		cond_resched();
3925		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3926
3927		i++;
3928		dst = mem_map_next(dst, dst_base, i);
3929		src = mem_map_next(src, src_base, i);
3930	}
3931}
3932
3933void copy_user_huge_page(struct page *dst, struct page *src,
3934			 unsigned long addr, struct vm_area_struct *vma,
3935			 unsigned int pages_per_huge_page)
3936{
3937	int i;
3938
3939	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3940		copy_user_gigantic_page(dst, src, addr, vma,
3941					pages_per_huge_page);
3942		return;
3943	}
3944
3945	might_sleep();
3946	for (i = 0; i < pages_per_huge_page; i++) {
3947		cond_resched();
3948		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3949	}
3950}
3951#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3952
3953#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3954
3955static struct kmem_cache *page_ptl_cachep;
3956
3957void __init ptlock_cache_init(void)
3958{
3959	page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3960			SLAB_PANIC, NULL);
3961}
3962
3963bool ptlock_alloc(struct page *page)
3964{
3965	spinlock_t *ptl;
3966
3967	ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3968	if (!ptl)
3969		return false;
3970	page->ptl = ptl;
3971	return true;
3972}
3973
3974void ptlock_free(struct page *page)
3975{
3976	kmem_cache_free(page_ptl_cachep, page->ptl);
3977}
3978#endif
v3.1
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *		Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 * 		Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *		(Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/hugetlb.h>
  44#include <linux/mman.h>
  45#include <linux/swap.h>
  46#include <linux/highmem.h>
  47#include <linux/pagemap.h>
  48#include <linux/ksm.h>
  49#include <linux/rmap.h>
  50#include <linux/module.h>
  51#include <linux/delayacct.h>
  52#include <linux/init.h>
 
  53#include <linux/writeback.h>
  54#include <linux/memcontrol.h>
  55#include <linux/mmu_notifier.h>
  56#include <linux/kallsyms.h>
  57#include <linux/swapops.h>
  58#include <linux/elf.h>
  59#include <linux/gfp.h>
 
 
 
 
 
  60
  61#include <asm/io.h>
 
  62#include <asm/pgalloc.h>
  63#include <asm/uaccess.h>
  64#include <asm/tlb.h>
  65#include <asm/tlbflush.h>
  66#include <asm/pgtable.h>
  67
  68#include "internal.h"
  69
 
 
 
 
  70#ifndef CONFIG_NEED_MULTIPLE_NODES
  71/* use the per-pgdat data instead for discontigmem - mbligh */
  72unsigned long max_mapnr;
  73struct page *mem_map;
  74
  75EXPORT_SYMBOL(max_mapnr);
  76EXPORT_SYMBOL(mem_map);
  77#endif
  78
  79unsigned long num_physpages;
  80/*
  81 * A number of key systems in x86 including ioremap() rely on the assumption
  82 * that high_memory defines the upper bound on direct map memory, then end
  83 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  85 * and ZONE_HIGHMEM.
  86 */
  87void * high_memory;
  88
  89EXPORT_SYMBOL(num_physpages);
  90EXPORT_SYMBOL(high_memory);
  91
  92/*
  93 * Randomize the address space (stacks, mmaps, brk, etc.).
  94 *
  95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
  96 *   as ancient (libc5 based) binaries can segfault. )
  97 */
  98int randomize_va_space __read_mostly =
  99#ifdef CONFIG_COMPAT_BRK
 100					1;
 101#else
 102					2;
 103#endif
 104
 105static int __init disable_randmaps(char *s)
 106{
 107	randomize_va_space = 0;
 108	return 1;
 109}
 110__setup("norandmaps", disable_randmaps);
 111
 112unsigned long zero_pfn __read_mostly;
 113unsigned long highest_memmap_pfn __read_mostly;
 114
 
 
 115/*
 116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 117 */
 118static int __init init_zero_pfn(void)
 119{
 120	zero_pfn = page_to_pfn(ZERO_PAGE(0));
 121	return 0;
 122}
 123core_initcall(init_zero_pfn);
 124
 125
 126#if defined(SPLIT_RSS_COUNTING)
 127
 128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
 129{
 130	int i;
 131
 132	for (i = 0; i < NR_MM_COUNTERS; i++) {
 133		if (task->rss_stat.count[i]) {
 134			add_mm_counter(mm, i, task->rss_stat.count[i]);
 135			task->rss_stat.count[i] = 0;
 136		}
 137	}
 138	task->rss_stat.events = 0;
 139}
 140
 141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 142{
 143	struct task_struct *task = current;
 144
 145	if (likely(task->mm == mm))
 146		task->rss_stat.count[member] += val;
 147	else
 148		add_mm_counter(mm, member, val);
 149}
 150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 152
 153/* sync counter once per 64 page faults */
 154#define TASK_RSS_EVENTS_THRESH	(64)
 155static void check_sync_rss_stat(struct task_struct *task)
 156{
 157	if (unlikely(task != current))
 158		return;
 159	if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 160		__sync_task_rss_stat(task, task->mm);
 161}
 162
 163unsigned long get_mm_counter(struct mm_struct *mm, int member)
 164{
 165	long val = 0;
 166
 167	/*
 168	 * Don't use task->mm here...for avoiding to use task_get_mm()..
 169	 * The caller must guarantee task->mm is not invalid.
 170	 */
 171	val = atomic_long_read(&mm->rss_stat.count[member]);
 172	/*
 173	 * counter is updated in asynchronous manner and may go to minus.
 174	 * But it's never be expected number for users.
 175	 */
 176	if (val < 0)
 177		return 0;
 178	return (unsigned long)val;
 179}
 180
 181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
 182{
 183	__sync_task_rss_stat(task, mm);
 184}
 185#else /* SPLIT_RSS_COUNTING */
 186
 187#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 188#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 189
 190static void check_sync_rss_stat(struct task_struct *task)
 191{
 192}
 193
 194#endif /* SPLIT_RSS_COUNTING */
 195
 196#ifdef HAVE_GENERIC_MMU_GATHER
 197
 198static int tlb_next_batch(struct mmu_gather *tlb)
 199{
 200	struct mmu_gather_batch *batch;
 201
 202	batch = tlb->active;
 203	if (batch->next) {
 204		tlb->active = batch->next;
 205		return 1;
 206	}
 207
 
 
 
 208	batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 209	if (!batch)
 210		return 0;
 211
 
 212	batch->next = NULL;
 213	batch->nr   = 0;
 214	batch->max  = MAX_GATHER_BATCH;
 215
 216	tlb->active->next = batch;
 217	tlb->active = batch;
 218
 219	return 1;
 220}
 221
 222/* tlb_gather_mmu
 223 *	Called to initialize an (on-stack) mmu_gather structure for page-table
 224 *	tear-down from @mm. The @fullmm argument is used when @mm is without
 225 *	users and we're going to destroy the full address space (exit/execve).
 226 */
 227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 228{
 229	tlb->mm = mm;
 230
 231	tlb->fullmm     = fullmm;
 232	tlb->need_flush = 0;
 233	tlb->fast_mode  = (num_possible_cpus() == 1);
 234	tlb->local.next = NULL;
 235	tlb->local.nr   = 0;
 236	tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 237	tlb->active     = &tlb->local;
 
 238
 239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 240	tlb->batch = NULL;
 241#endif
 
 
 242}
 243
 244void tlb_flush_mmu(struct mmu_gather *tlb)
 245{
 246	struct mmu_gather_batch *batch;
 
 247
 248	if (!tlb->need_flush)
 249		return;
 250	tlb->need_flush = 0;
 251	tlb_flush(tlb);
 
 252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 253	tlb_table_flush(tlb);
 254#endif
 
 
 255
 256	if (tlb_fast_mode(tlb))
 257		return;
 
 258
 259	for (batch = &tlb->local; batch; batch = batch->next) {
 260		free_pages_and_swap_cache(batch->pages, batch->nr);
 261		batch->nr = 0;
 262	}
 263	tlb->active = &tlb->local;
 264}
 265
 
 
 
 
 
 
 266/* tlb_finish_mmu
 267 *	Called at the end of the shootdown operation to free up any resources
 268 *	that were required.
 269 */
 270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 271{
 272	struct mmu_gather_batch *batch, *next;
 273
 274	tlb_flush_mmu(tlb);
 275
 276	/* keep the page table cache within bounds */
 277	check_pgt_cache();
 278
 279	for (batch = tlb->local.next; batch; batch = next) {
 280		next = batch->next;
 281		free_pages((unsigned long)batch, 0);
 282	}
 283	tlb->local.next = NULL;
 284}
 285
 286/* __tlb_remove_page
 287 *	Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 288 *	handling the additional races in SMP caused by other CPUs caching valid
 289 *	mappings in their TLBs. Returns the number of free page slots left.
 290 *	When out of page slots we must call tlb_flush_mmu().
 291 */
 292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 293{
 294	struct mmu_gather_batch *batch;
 295
 296	tlb->need_flush = 1;
 297
 298	if (tlb_fast_mode(tlb)) {
 299		free_page_and_swap_cache(page);
 300		return 1; /* avoid calling tlb_flush_mmu() */
 301	}
 302
 303	batch = tlb->active;
 304	batch->pages[batch->nr++] = page;
 305	if (batch->nr == batch->max) {
 306		if (!tlb_next_batch(tlb))
 307			return 0;
 308		batch = tlb->active;
 309	}
 310	VM_BUG_ON(batch->nr > batch->max);
 311
 312	return batch->max - batch->nr;
 313}
 314
 315#endif /* HAVE_GENERIC_MMU_GATHER */
 316
 317#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 318
 319/*
 320 * See the comment near struct mmu_table_batch.
 321 */
 322
 323static void tlb_remove_table_smp_sync(void *arg)
 324{
 325	/* Simply deliver the interrupt */
 326}
 327
 328static void tlb_remove_table_one(void *table)
 329{
 330	/*
 331	 * This isn't an RCU grace period and hence the page-tables cannot be
 332	 * assumed to be actually RCU-freed.
 333	 *
 334	 * It is however sufficient for software page-table walkers that rely on
 335	 * IRQ disabling. See the comment near struct mmu_table_batch.
 336	 */
 337	smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 338	__tlb_remove_table(table);
 339}
 340
 341static void tlb_remove_table_rcu(struct rcu_head *head)
 342{
 343	struct mmu_table_batch *batch;
 344	int i;
 345
 346	batch = container_of(head, struct mmu_table_batch, rcu);
 347
 348	for (i = 0; i < batch->nr; i++)
 349		__tlb_remove_table(batch->tables[i]);
 350
 351	free_page((unsigned long)batch);
 352}
 353
 354void tlb_table_flush(struct mmu_gather *tlb)
 355{
 356	struct mmu_table_batch **batch = &tlb->batch;
 357
 358	if (*batch) {
 359		call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 360		*batch = NULL;
 361	}
 362}
 363
 364void tlb_remove_table(struct mmu_gather *tlb, void *table)
 365{
 366	struct mmu_table_batch **batch = &tlb->batch;
 367
 368	tlb->need_flush = 1;
 369
 370	/*
 371	 * When there's less then two users of this mm there cannot be a
 372	 * concurrent page-table walk.
 373	 */
 374	if (atomic_read(&tlb->mm->mm_users) < 2) {
 375		__tlb_remove_table(table);
 376		return;
 377	}
 378
 379	if (*batch == NULL) {
 380		*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 381		if (*batch == NULL) {
 382			tlb_remove_table_one(table);
 383			return;
 384		}
 385		(*batch)->nr = 0;
 386	}
 387	(*batch)->tables[(*batch)->nr++] = table;
 388	if ((*batch)->nr == MAX_TABLE_BATCH)
 389		tlb_table_flush(tlb);
 390}
 391
 392#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 393
 394/*
 395 * If a p?d_bad entry is found while walking page tables, report
 396 * the error, before resetting entry to p?d_none.  Usually (but
 397 * very seldom) called out from the p?d_none_or_clear_bad macros.
 398 */
 399
 400void pgd_clear_bad(pgd_t *pgd)
 401{
 402	pgd_ERROR(*pgd);
 403	pgd_clear(pgd);
 404}
 405
 406void pud_clear_bad(pud_t *pud)
 407{
 408	pud_ERROR(*pud);
 409	pud_clear(pud);
 410}
 411
 412void pmd_clear_bad(pmd_t *pmd)
 413{
 414	pmd_ERROR(*pmd);
 415	pmd_clear(pmd);
 416}
 417
 418/*
 419 * Note: this doesn't free the actual pages themselves. That
 420 * has been handled earlier when unmapping all the memory regions.
 421 */
 422static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 423			   unsigned long addr)
 424{
 425	pgtable_t token = pmd_pgtable(*pmd);
 426	pmd_clear(pmd);
 427	pte_free_tlb(tlb, token, addr);
 428	tlb->mm->nr_ptes--;
 429}
 430
 431static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 432				unsigned long addr, unsigned long end,
 433				unsigned long floor, unsigned long ceiling)
 434{
 435	pmd_t *pmd;
 436	unsigned long next;
 437	unsigned long start;
 438
 439	start = addr;
 440	pmd = pmd_offset(pud, addr);
 441	do {
 442		next = pmd_addr_end(addr, end);
 443		if (pmd_none_or_clear_bad(pmd))
 444			continue;
 445		free_pte_range(tlb, pmd, addr);
 446	} while (pmd++, addr = next, addr != end);
 447
 448	start &= PUD_MASK;
 449	if (start < floor)
 450		return;
 451	if (ceiling) {
 452		ceiling &= PUD_MASK;
 453		if (!ceiling)
 454			return;
 455	}
 456	if (end - 1 > ceiling - 1)
 457		return;
 458
 459	pmd = pmd_offset(pud, start);
 460	pud_clear(pud);
 461	pmd_free_tlb(tlb, pmd, start);
 
 462}
 463
 464static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 465				unsigned long addr, unsigned long end,
 466				unsigned long floor, unsigned long ceiling)
 467{
 468	pud_t *pud;
 469	unsigned long next;
 470	unsigned long start;
 471
 472	start = addr;
 473	pud = pud_offset(pgd, addr);
 474	do {
 475		next = pud_addr_end(addr, end);
 476		if (pud_none_or_clear_bad(pud))
 477			continue;
 478		free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 479	} while (pud++, addr = next, addr != end);
 480
 481	start &= PGDIR_MASK;
 482	if (start < floor)
 483		return;
 484	if (ceiling) {
 485		ceiling &= PGDIR_MASK;
 486		if (!ceiling)
 487			return;
 488	}
 489	if (end - 1 > ceiling - 1)
 490		return;
 491
 492	pud = pud_offset(pgd, start);
 493	pgd_clear(pgd);
 494	pud_free_tlb(tlb, pud, start);
 495}
 496
 497/*
 498 * This function frees user-level page tables of a process.
 499 *
 500 * Must be called with pagetable lock held.
 501 */
 502void free_pgd_range(struct mmu_gather *tlb,
 503			unsigned long addr, unsigned long end,
 504			unsigned long floor, unsigned long ceiling)
 505{
 506	pgd_t *pgd;
 507	unsigned long next;
 508
 509	/*
 510	 * The next few lines have given us lots of grief...
 511	 *
 512	 * Why are we testing PMD* at this top level?  Because often
 513	 * there will be no work to do at all, and we'd prefer not to
 514	 * go all the way down to the bottom just to discover that.
 515	 *
 516	 * Why all these "- 1"s?  Because 0 represents both the bottom
 517	 * of the address space and the top of it (using -1 for the
 518	 * top wouldn't help much: the masks would do the wrong thing).
 519	 * The rule is that addr 0 and floor 0 refer to the bottom of
 520	 * the address space, but end 0 and ceiling 0 refer to the top
 521	 * Comparisons need to use "end - 1" and "ceiling - 1" (though
 522	 * that end 0 case should be mythical).
 523	 *
 524	 * Wherever addr is brought up or ceiling brought down, we must
 525	 * be careful to reject "the opposite 0" before it confuses the
 526	 * subsequent tests.  But what about where end is brought down
 527	 * by PMD_SIZE below? no, end can't go down to 0 there.
 528	 *
 529	 * Whereas we round start (addr) and ceiling down, by different
 530	 * masks at different levels, in order to test whether a table
 531	 * now has no other vmas using it, so can be freed, we don't
 532	 * bother to round floor or end up - the tests don't need that.
 533	 */
 534
 535	addr &= PMD_MASK;
 536	if (addr < floor) {
 537		addr += PMD_SIZE;
 538		if (!addr)
 539			return;
 540	}
 541	if (ceiling) {
 542		ceiling &= PMD_MASK;
 543		if (!ceiling)
 544			return;
 545	}
 546	if (end - 1 > ceiling - 1)
 547		end -= PMD_SIZE;
 548	if (addr > end - 1)
 549		return;
 550
 551	pgd = pgd_offset(tlb->mm, addr);
 552	do {
 553		next = pgd_addr_end(addr, end);
 554		if (pgd_none_or_clear_bad(pgd))
 555			continue;
 556		free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 557	} while (pgd++, addr = next, addr != end);
 558}
 559
 560void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 561		unsigned long floor, unsigned long ceiling)
 562{
 563	while (vma) {
 564		struct vm_area_struct *next = vma->vm_next;
 565		unsigned long addr = vma->vm_start;
 566
 567		/*
 568		 * Hide vma from rmap and truncate_pagecache before freeing
 569		 * pgtables
 570		 */
 571		unlink_anon_vmas(vma);
 572		unlink_file_vma(vma);
 573
 574		if (is_vm_hugetlb_page(vma)) {
 575			hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 576				floor, next? next->vm_start: ceiling);
 577		} else {
 578			/*
 579			 * Optimization: gather nearby vmas into one call down
 580			 */
 581			while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 582			       && !is_vm_hugetlb_page(next)) {
 583				vma = next;
 584				next = vma->vm_next;
 585				unlink_anon_vmas(vma);
 586				unlink_file_vma(vma);
 587			}
 588			free_pgd_range(tlb, addr, vma->vm_end,
 589				floor, next? next->vm_start: ceiling);
 590		}
 591		vma = next;
 592	}
 593}
 594
 595int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 596		pmd_t *pmd, unsigned long address)
 597{
 
 598	pgtable_t new = pte_alloc_one(mm, address);
 599	int wait_split_huge_page;
 600	if (!new)
 601		return -ENOMEM;
 602
 603	/*
 604	 * Ensure all pte setup (eg. pte page lock and page clearing) are
 605	 * visible before the pte is made visible to other CPUs by being
 606	 * put into page tables.
 607	 *
 608	 * The other side of the story is the pointer chasing in the page
 609	 * table walking code (when walking the page table without locking;
 610	 * ie. most of the time). Fortunately, these data accesses consist
 611	 * of a chain of data-dependent loads, meaning most CPUs (alpha
 612	 * being the notable exception) will already guarantee loads are
 613	 * seen in-order. See the alpha page table accessors for the
 614	 * smp_read_barrier_depends() barriers in page table walking code.
 615	 */
 616	smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 617
 618	spin_lock(&mm->page_table_lock);
 619	wait_split_huge_page = 0;
 620	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 621		mm->nr_ptes++;
 622		pmd_populate(mm, pmd, new);
 623		new = NULL;
 624	} else if (unlikely(pmd_trans_splitting(*pmd)))
 625		wait_split_huge_page = 1;
 626	spin_unlock(&mm->page_table_lock);
 627	if (new)
 628		pte_free(mm, new);
 629	if (wait_split_huge_page)
 630		wait_split_huge_page(vma->anon_vma, pmd);
 631	return 0;
 632}
 633
 634int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 635{
 636	pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 637	if (!new)
 638		return -ENOMEM;
 639
 640	smp_wmb(); /* See comment in __pte_alloc */
 641
 642	spin_lock(&init_mm.page_table_lock);
 643	if (likely(pmd_none(*pmd))) {	/* Has another populated it ? */
 644		pmd_populate_kernel(&init_mm, pmd, new);
 645		new = NULL;
 646	} else
 647		VM_BUG_ON(pmd_trans_splitting(*pmd));
 648	spin_unlock(&init_mm.page_table_lock);
 649	if (new)
 650		pte_free_kernel(&init_mm, new);
 651	return 0;
 652}
 653
 654static inline void init_rss_vec(int *rss)
 655{
 656	memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 657}
 658
 659static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 660{
 661	int i;
 662
 663	if (current->mm == mm)
 664		sync_mm_rss(current, mm);
 665	for (i = 0; i < NR_MM_COUNTERS; i++)
 666		if (rss[i])
 667			add_mm_counter(mm, i, rss[i]);
 668}
 669
 670/*
 671 * This function is called to print an error when a bad pte
 672 * is found. For example, we might have a PFN-mapped pte in
 673 * a region that doesn't allow it.
 674 *
 675 * The calling function must still handle the error.
 676 */
 677static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 678			  pte_t pte, struct page *page)
 679{
 680	pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 681	pud_t *pud = pud_offset(pgd, addr);
 682	pmd_t *pmd = pmd_offset(pud, addr);
 683	struct address_space *mapping;
 684	pgoff_t index;
 685	static unsigned long resume;
 686	static unsigned long nr_shown;
 687	static unsigned long nr_unshown;
 688
 689	/*
 690	 * Allow a burst of 60 reports, then keep quiet for that minute;
 691	 * or allow a steady drip of one report per second.
 692	 */
 693	if (nr_shown == 60) {
 694		if (time_before(jiffies, resume)) {
 695			nr_unshown++;
 696			return;
 697		}
 698		if (nr_unshown) {
 699			printk(KERN_ALERT
 700				"BUG: Bad page map: %lu messages suppressed\n",
 701				nr_unshown);
 702			nr_unshown = 0;
 703		}
 704		nr_shown = 0;
 705	}
 706	if (nr_shown++ == 0)
 707		resume = jiffies + 60 * HZ;
 708
 709	mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 710	index = linear_page_index(vma, addr);
 711
 712	printk(KERN_ALERT
 713		"BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 714		current->comm,
 715		(long long)pte_val(pte), (long long)pmd_val(*pmd));
 716	if (page)
 717		dump_page(page);
 718	printk(KERN_ALERT
 719		"addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 720		(void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 721	/*
 722	 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 723	 */
 724	if (vma->vm_ops)
 725		print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
 726				(unsigned long)vma->vm_ops->fault);
 727	if (vma->vm_file && vma->vm_file->f_op)
 728		print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
 729				(unsigned long)vma->vm_file->f_op->mmap);
 730	dump_stack();
 731	add_taint(TAINT_BAD_PAGE);
 732}
 733
 734static inline int is_cow_mapping(vm_flags_t flags)
 735{
 736	return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 737}
 738
 739#ifndef is_zero_pfn
 740static inline int is_zero_pfn(unsigned long pfn)
 741{
 742	return pfn == zero_pfn;
 743}
 744#endif
 745
 746#ifndef my_zero_pfn
 747static inline unsigned long my_zero_pfn(unsigned long addr)
 748{
 749	return zero_pfn;
 750}
 751#endif
 752
 753/*
 754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 755 *
 756 * "Special" mappings do not wish to be associated with a "struct page" (either
 757 * it doesn't exist, or it exists but they don't want to touch it). In this
 758 * case, NULL is returned here. "Normal" mappings do have a struct page.
 759 *
 760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 761 * pte bit, in which case this function is trivial. Secondly, an architecture
 762 * may not have a spare pte bit, which requires a more complicated scheme,
 763 * described below.
 764 *
 765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 766 * special mapping (even if there are underlying and valid "struct pages").
 767 * COWed pages of a VM_PFNMAP are always normal.
 768 *
 769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 772 * mapping will always honor the rule
 773 *
 774 *	pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 775 *
 776 * And for normal mappings this is false.
 777 *
 778 * This restricts such mappings to be a linear translation from virtual address
 779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 780 * as the vma is not a COW mapping; in that case, we know that all ptes are
 781 * special (because none can have been COWed).
 782 *
 783 *
 784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 785 *
 786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 787 * page" backing, however the difference is that _all_ pages with a struct
 788 * page (that is, those where pfn_valid is true) are refcounted and considered
 789 * normal pages by the VM. The disadvantage is that pages are refcounted
 790 * (which can be slower and simply not an option for some PFNMAP users). The
 791 * advantage is that we don't have to follow the strict linearity rule of
 792 * PFNMAP mappings in order to support COWable mappings.
 793 *
 794 */
 795#ifdef __HAVE_ARCH_PTE_SPECIAL
 796# define HAVE_PTE_SPECIAL 1
 797#else
 798# define HAVE_PTE_SPECIAL 0
 799#endif
 800struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 801				pte_t pte)
 802{
 803	unsigned long pfn = pte_pfn(pte);
 804
 805	if (HAVE_PTE_SPECIAL) {
 806		if (likely(!pte_special(pte)))
 807			goto check_pfn;
 
 
 808		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 809			return NULL;
 810		if (!is_zero_pfn(pfn))
 811			print_bad_pte(vma, addr, pte, NULL);
 812		return NULL;
 813	}
 814
 815	/* !HAVE_PTE_SPECIAL case follows: */
 816
 817	if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 818		if (vma->vm_flags & VM_MIXEDMAP) {
 819			if (!pfn_valid(pfn))
 820				return NULL;
 821			goto out;
 822		} else {
 823			unsigned long off;
 824			off = (addr - vma->vm_start) >> PAGE_SHIFT;
 825			if (pfn == vma->vm_pgoff + off)
 826				return NULL;
 827			if (!is_cow_mapping(vma->vm_flags))
 828				return NULL;
 829		}
 830	}
 831
 832	if (is_zero_pfn(pfn))
 833		return NULL;
 834check_pfn:
 835	if (unlikely(pfn > highest_memmap_pfn)) {
 836		print_bad_pte(vma, addr, pte, NULL);
 837		return NULL;
 838	}
 839
 840	/*
 841	 * NOTE! We still have PageReserved() pages in the page tables.
 842	 * eg. VDSO mappings can cause them to exist.
 843	 */
 844out:
 845	return pfn_to_page(pfn);
 846}
 847
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 848/*
 849 * copy one vm_area from one task to the other. Assumes the page tables
 850 * already present in the new task to be cleared in the whole range
 851 * covered by this vma.
 852 */
 853
 854static inline unsigned long
 855copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 856		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 857		unsigned long addr, int *rss)
 858{
 859	unsigned long vm_flags = vma->vm_flags;
 860	pte_t pte = *src_pte;
 861	struct page *page;
 862
 863	/* pte contains position in swap or file, so copy. */
 864	if (unlikely(!pte_present(pte))) {
 865		if (!pte_file(pte)) {
 866			swp_entry_t entry = pte_to_swp_entry(pte);
 867
 
 868			if (swap_duplicate(entry) < 0)
 869				return entry.val;
 870
 871			/* make sure dst_mm is on swapoff's mmlist. */
 872			if (unlikely(list_empty(&dst_mm->mmlist))) {
 873				spin_lock(&mmlist_lock);
 874				if (list_empty(&dst_mm->mmlist))
 875					list_add(&dst_mm->mmlist,
 876						 &src_mm->mmlist);
 877				spin_unlock(&mmlist_lock);
 878			}
 879			if (likely(!non_swap_entry(entry)))
 880				rss[MM_SWAPENTS]++;
 881			else if (is_write_migration_entry(entry) &&
 
 
 
 
 882					is_cow_mapping(vm_flags)) {
 883				/*
 884				 * COW mappings require pages in both parent
 885				 * and child to be set to read.
 886				 */
 887				make_migration_entry_read(&entry);
 888				pte = swp_entry_to_pte(entry);
 
 
 889				set_pte_at(src_mm, addr, src_pte, pte);
 890			}
 891		}
 892		goto out_set_pte;
 893	}
 894
 895	/*
 896	 * If it's a COW mapping, write protect it both
 897	 * in the parent and the child
 898	 */
 899	if (is_cow_mapping(vm_flags)) {
 900		ptep_set_wrprotect(src_mm, addr, src_pte);
 901		pte = pte_wrprotect(pte);
 902	}
 903
 904	/*
 905	 * If it's a shared mapping, mark it clean in
 906	 * the child
 907	 */
 908	if (vm_flags & VM_SHARED)
 909		pte = pte_mkclean(pte);
 910	pte = pte_mkold(pte);
 911
 912	page = vm_normal_page(vma, addr, pte);
 913	if (page) {
 914		get_page(page);
 915		page_dup_rmap(page);
 916		if (PageAnon(page))
 917			rss[MM_ANONPAGES]++;
 918		else
 919			rss[MM_FILEPAGES]++;
 920	}
 921
 922out_set_pte:
 923	set_pte_at(dst_mm, addr, dst_pte, pte);
 924	return 0;
 925}
 926
 927int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 928		   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 929		   unsigned long addr, unsigned long end)
 930{
 931	pte_t *orig_src_pte, *orig_dst_pte;
 932	pte_t *src_pte, *dst_pte;
 933	spinlock_t *src_ptl, *dst_ptl;
 934	int progress = 0;
 935	int rss[NR_MM_COUNTERS];
 936	swp_entry_t entry = (swp_entry_t){0};
 937
 938again:
 939	init_rss_vec(rss);
 940
 941	dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 942	if (!dst_pte)
 943		return -ENOMEM;
 944	src_pte = pte_offset_map(src_pmd, addr);
 945	src_ptl = pte_lockptr(src_mm, src_pmd);
 946	spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 947	orig_src_pte = src_pte;
 948	orig_dst_pte = dst_pte;
 949	arch_enter_lazy_mmu_mode();
 950
 951	do {
 952		/*
 953		 * We are holding two locks at this point - either of them
 954		 * could generate latencies in another task on another CPU.
 955		 */
 956		if (progress >= 32) {
 957			progress = 0;
 958			if (need_resched() ||
 959			    spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 960				break;
 961		}
 962		if (pte_none(*src_pte)) {
 963			progress++;
 964			continue;
 965		}
 966		entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 967							vma, addr, rss);
 968		if (entry.val)
 969			break;
 970		progress += 8;
 971	} while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 972
 973	arch_leave_lazy_mmu_mode();
 974	spin_unlock(src_ptl);
 975	pte_unmap(orig_src_pte);
 976	add_mm_rss_vec(dst_mm, rss);
 977	pte_unmap_unlock(orig_dst_pte, dst_ptl);
 978	cond_resched();
 979
 980	if (entry.val) {
 981		if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 982			return -ENOMEM;
 983		progress = 0;
 984	}
 985	if (addr != end)
 986		goto again;
 987	return 0;
 988}
 989
 990static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 991		pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 992		unsigned long addr, unsigned long end)
 993{
 994	pmd_t *src_pmd, *dst_pmd;
 995	unsigned long next;
 996
 997	dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 998	if (!dst_pmd)
 999		return -ENOMEM;
1000	src_pmd = pmd_offset(src_pud, addr);
1001	do {
1002		next = pmd_addr_end(addr, end);
1003		if (pmd_trans_huge(*src_pmd)) {
1004			int err;
1005			VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1006			err = copy_huge_pmd(dst_mm, src_mm,
1007					    dst_pmd, src_pmd, addr, vma);
1008			if (err == -ENOMEM)
1009				return -ENOMEM;
1010			if (!err)
1011				continue;
1012			/* fall through */
1013		}
1014		if (pmd_none_or_clear_bad(src_pmd))
1015			continue;
1016		if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017						vma, addr, next))
1018			return -ENOMEM;
1019	} while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020	return 0;
1021}
1022
1023static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024		pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025		unsigned long addr, unsigned long end)
1026{
1027	pud_t *src_pud, *dst_pud;
1028	unsigned long next;
1029
1030	dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031	if (!dst_pud)
1032		return -ENOMEM;
1033	src_pud = pud_offset(src_pgd, addr);
1034	do {
1035		next = pud_addr_end(addr, end);
1036		if (pud_none_or_clear_bad(src_pud))
1037			continue;
1038		if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039						vma, addr, next))
1040			return -ENOMEM;
1041	} while (dst_pud++, src_pud++, addr = next, addr != end);
1042	return 0;
1043}
1044
1045int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046		struct vm_area_struct *vma)
1047{
1048	pgd_t *src_pgd, *dst_pgd;
1049	unsigned long next;
1050	unsigned long addr = vma->vm_start;
1051	unsigned long end = vma->vm_end;
 
 
 
1052	int ret;
1053
1054	/*
1055	 * Don't copy ptes where a page fault will fill them correctly.
1056	 * Fork becomes much lighter when there are big shared or private
1057	 * readonly mappings. The tradeoff is that copy_page_range is more
1058	 * efficient than faulting.
1059	 */
1060	if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1061		if (!vma->anon_vma)
1062			return 0;
1063	}
1064
1065	if (is_vm_hugetlb_page(vma))
1066		return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067
1068	if (unlikely(is_pfn_mapping(vma))) {
1069		/*
1070		 * We do not free on error cases below as remove_vma
1071		 * gets called on error from higher level routine
1072		 */
1073		ret = track_pfn_vma_copy(vma);
1074		if (ret)
1075			return ret;
1076	}
1077
1078	/*
1079	 * We need to invalidate the secondary MMU mappings only when
1080	 * there could be a permission downgrade on the ptes of the
1081	 * parent mm. And a permission downgrade will only happen if
1082	 * is_cow_mapping() returns true.
1083	 */
1084	if (is_cow_mapping(vma->vm_flags))
1085		mmu_notifier_invalidate_range_start(src_mm, addr, end);
 
 
 
 
1086
1087	ret = 0;
1088	dst_pgd = pgd_offset(dst_mm, addr);
1089	src_pgd = pgd_offset(src_mm, addr);
1090	do {
1091		next = pgd_addr_end(addr, end);
1092		if (pgd_none_or_clear_bad(src_pgd))
1093			continue;
1094		if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095					    vma, addr, next))) {
1096			ret = -ENOMEM;
1097			break;
1098		}
1099	} while (dst_pgd++, src_pgd++, addr = next, addr != end);
1100
1101	if (is_cow_mapping(vma->vm_flags))
1102		mmu_notifier_invalidate_range_end(src_mm,
1103						  vma->vm_start, end);
1104	return ret;
1105}
1106
1107static unsigned long zap_pte_range(struct mmu_gather *tlb,
1108				struct vm_area_struct *vma, pmd_t *pmd,
1109				unsigned long addr, unsigned long end,
1110				struct zap_details *details)
1111{
1112	struct mm_struct *mm = tlb->mm;
1113	int force_flush = 0;
1114	int rss[NR_MM_COUNTERS];
1115	spinlock_t *ptl;
1116	pte_t *start_pte;
1117	pte_t *pte;
 
1118
1119again:
1120	init_rss_vec(rss);
1121	start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1122	pte = start_pte;
1123	arch_enter_lazy_mmu_mode();
1124	do {
1125		pte_t ptent = *pte;
1126		if (pte_none(ptent)) {
1127			continue;
1128		}
1129
1130		if (pte_present(ptent)) {
1131			struct page *page;
1132
1133			page = vm_normal_page(vma, addr, ptent);
1134			if (unlikely(details) && page) {
1135				/*
1136				 * unmap_shared_mapping_pages() wants to
1137				 * invalidate cache without truncating:
1138				 * unmap shared but keep private pages.
1139				 */
1140				if (details->check_mapping &&
1141				    details->check_mapping != page->mapping)
1142					continue;
1143				/*
1144				 * Each page->index must be checked when
1145				 * invalidating or truncating nonlinear.
1146				 */
1147				if (details->nonlinear_vma &&
1148				    (page->index < details->first_index ||
1149				     page->index > details->last_index))
1150					continue;
1151			}
1152			ptent = ptep_get_and_clear_full(mm, addr, pte,
1153							tlb->fullmm);
1154			tlb_remove_tlb_entry(tlb, pte, addr);
1155			if (unlikely(!page))
1156				continue;
1157			if (unlikely(details) && details->nonlinear_vma
1158			    && linear_page_index(details->nonlinear_vma,
1159						addr) != page->index)
1160				set_pte_at(mm, addr, pte,
1161					   pgoff_to_pte(page->index));
1162			if (PageAnon(page))
1163				rss[MM_ANONPAGES]--;
1164			else {
1165				if (pte_dirty(ptent))
 
1166					set_page_dirty(page);
 
1167				if (pte_young(ptent) &&
1168				    likely(!VM_SequentialReadHint(vma)))
1169					mark_page_accessed(page);
1170				rss[MM_FILEPAGES]--;
1171			}
1172			page_remove_rmap(page);
 
1173			if (unlikely(page_mapcount(page) < 0))
1174				print_bad_pte(vma, addr, ptent, page);
1175			force_flush = !__tlb_remove_page(tlb, page);
1176			if (force_flush)
 
1177				break;
 
1178			continue;
1179		}
1180		/*
1181		 * If details->check_mapping, we leave swap entries;
1182		 * if details->nonlinear_vma, we leave file entries.
1183		 */
1184		if (unlikely(details))
1185			continue;
1186		if (pte_file(ptent)) {
1187			if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188				print_bad_pte(vma, addr, ptent, NULL);
1189		} else {
1190			swp_entry_t entry = pte_to_swp_entry(ptent);
1191
1192			if (!non_swap_entry(entry))
1193				rss[MM_SWAPENTS]--;
1194			if (unlikely(!free_swap_and_cache(entry)))
1195				print_bad_pte(vma, addr, ptent, NULL);
 
 
 
 
1196		}
 
 
1197		pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198	} while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200	add_mm_rss_vec(mm, rss);
1201	arch_leave_lazy_mmu_mode();
 
 
 
 
1202	pte_unmap_unlock(start_pte, ptl);
1203
1204	/*
1205	 * mmu_gather ran out of room to batch pages, we break out of
1206	 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207	 * and page-free while holding it.
 
1208	 */
1209	if (force_flush) {
1210		force_flush = 0;
1211		tlb_flush_mmu(tlb);
 
1212		if (addr != end)
1213			goto again;
1214	}
1215
1216	return addr;
1217}
1218
1219static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1220				struct vm_area_struct *vma, pud_t *pud,
1221				unsigned long addr, unsigned long end,
1222				struct zap_details *details)
1223{
1224	pmd_t *pmd;
1225	unsigned long next;
1226
1227	pmd = pmd_offset(pud, addr);
1228	do {
1229		next = pmd_addr_end(addr, end);
1230		if (pmd_trans_huge(*pmd)) {
1231			if (next-addr != HPAGE_PMD_SIZE) {
1232				VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1233				split_huge_page_pmd(vma->vm_mm, pmd);
1234			} else if (zap_huge_pmd(tlb, vma, pmd))
1235				continue;
 
1236			/* fall through */
1237		}
1238		if (pmd_none_or_clear_bad(pmd))
1239			continue;
 
 
 
 
 
 
 
1240		next = zap_pte_range(tlb, vma, pmd, addr, next, details);
 
1241		cond_resched();
1242	} while (pmd++, addr = next, addr != end);
1243
1244	return addr;
1245}
1246
1247static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1248				struct vm_area_struct *vma, pgd_t *pgd,
1249				unsigned long addr, unsigned long end,
1250				struct zap_details *details)
1251{
1252	pud_t *pud;
1253	unsigned long next;
1254
1255	pud = pud_offset(pgd, addr);
1256	do {
1257		next = pud_addr_end(addr, end);
1258		if (pud_none_or_clear_bad(pud))
1259			continue;
1260		next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1261	} while (pud++, addr = next, addr != end);
1262
1263	return addr;
1264}
1265
1266static unsigned long unmap_page_range(struct mmu_gather *tlb,
1267				struct vm_area_struct *vma,
1268				unsigned long addr, unsigned long end,
1269				struct zap_details *details)
1270{
1271	pgd_t *pgd;
1272	unsigned long next;
1273
1274	if (details && !details->check_mapping && !details->nonlinear_vma)
1275		details = NULL;
1276
1277	BUG_ON(addr >= end);
1278	mem_cgroup_uncharge_start();
1279	tlb_start_vma(tlb, vma);
1280	pgd = pgd_offset(vma->vm_mm, addr);
1281	do {
1282		next = pgd_addr_end(addr, end);
1283		if (pgd_none_or_clear_bad(pgd))
1284			continue;
1285		next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1286	} while (pgd++, addr = next, addr != end);
1287	tlb_end_vma(tlb, vma);
1288	mem_cgroup_uncharge_end();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1289
1290	return addr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1291}
1292
1293/**
1294 * unmap_vmas - unmap a range of memory covered by a list of vma's
1295 * @tlb: address of the caller's struct mmu_gather
1296 * @vma: the starting vma
1297 * @start_addr: virtual address at which to start unmapping
1298 * @end_addr: virtual address at which to end unmapping
1299 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1300 * @details: details of nonlinear truncation or shared cache invalidation
1301 *
1302 * Returns the end address of the unmapping (restart addr if interrupted).
1303 *
1304 * Unmap all pages in the vma list.
1305 *
1306 * Only addresses between `start' and `end' will be unmapped.
1307 *
1308 * The VMA list must be sorted in ascending virtual address order.
1309 *
1310 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1311 * range after unmap_vmas() returns.  So the only responsibility here is to
1312 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1313 * drops the lock and schedules.
1314 */
1315unsigned long unmap_vmas(struct mmu_gather *tlb,
1316		struct vm_area_struct *vma, unsigned long start_addr,
1317		unsigned long end_addr, unsigned long *nr_accounted,
1318		struct zap_details *details)
1319{
1320	unsigned long start = start_addr;
1321	struct mm_struct *mm = vma->vm_mm;
1322
1323	mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1324	for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1325		unsigned long end;
 
 
1326
1327		start = max(vma->vm_start, start_addr);
1328		if (start >= vma->vm_end)
1329			continue;
1330		end = min(vma->vm_end, end_addr);
1331		if (end <= vma->vm_start)
1332			continue;
 
 
 
 
 
 
 
 
 
1333
1334		if (vma->vm_flags & VM_ACCOUNT)
1335			*nr_accounted += (end - start) >> PAGE_SHIFT;
1336
1337		if (unlikely(is_pfn_mapping(vma)))
1338			untrack_pfn_vma(vma, 0, 0);
1339
1340		while (start != end) {
1341			if (unlikely(is_vm_hugetlb_page(vma))) {
1342				/*
1343				 * It is undesirable to test vma->vm_file as it
1344				 * should be non-null for valid hugetlb area.
1345				 * However, vm_file will be NULL in the error
1346				 * cleanup path of do_mmap_pgoff. When
1347				 * hugetlbfs ->mmap method fails,
1348				 * do_mmap_pgoff() nullifies vma->vm_file
1349				 * before calling this function to clean up.
1350				 * Since no pte has actually been setup, it is
1351				 * safe to do nothing in this case.
1352				 */
1353				if (vma->vm_file)
1354					unmap_hugepage_range(vma, start, end, NULL);
1355
1356				start = end;
1357			} else
1358				start = unmap_page_range(tlb, vma, start, end, details);
1359		}
1360	}
1361
1362	mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1363	return start;	/* which is now the end (or restart) address */
1364}
1365
1366/**
1367 * zap_page_range - remove user pages in a given range
1368 * @vma: vm_area_struct holding the applicable pages
1369 * @address: starting address of pages to zap
1370 * @size: number of bytes to zap
1371 * @details: details of nonlinear truncation or shared cache invalidation
 
 
1372 */
1373unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1374		unsigned long size, struct zap_details *details)
1375{
1376	struct mm_struct *mm = vma->vm_mm;
1377	struct mmu_gather tlb;
1378	unsigned long end = address + size;
1379	unsigned long nr_accounted = 0;
1380
1381	lru_add_drain();
1382	tlb_gather_mmu(&tlb, mm, 0);
1383	update_hiwater_rss(mm);
1384	end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
 
 
1385	tlb_finish_mmu(&tlb, address, end);
1386	return end;
1387}
1388
1389/**
1390 * zap_vma_ptes - remove ptes mapping the vma
1391 * @vma: vm_area_struct holding ptes to be zapped
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1394 *
1395 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1396 *
1397 * The entire address range must be fully contained within the vma.
1398 *
1399 * Returns 0 if successful.
1400 */
1401int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1402		unsigned long size)
1403{
1404	if (address < vma->vm_start || address + size > vma->vm_end ||
1405	    		!(vma->vm_flags & VM_PFNMAP))
1406		return -1;
1407	zap_page_range(vma, address, size, NULL);
1408	return 0;
1409}
1410EXPORT_SYMBOL_GPL(zap_vma_ptes);
1411
1412/**
1413 * follow_page - look up a page descriptor from a user-virtual address
1414 * @vma: vm_area_struct mapping @address
1415 * @address: virtual address to look up
1416 * @flags: flags modifying lookup behaviour
1417 *
1418 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1419 *
1420 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1421 * an error pointer if there is a mapping to something not represented
1422 * by a page descriptor (see also vm_normal_page()).
1423 */
1424struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1425			unsigned int flags)
1426{
1427	pgd_t *pgd;
1428	pud_t *pud;
1429	pmd_t *pmd;
1430	pte_t *ptep, pte;
1431	spinlock_t *ptl;
1432	struct page *page;
1433	struct mm_struct *mm = vma->vm_mm;
1434
1435	page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1436	if (!IS_ERR(page)) {
1437		BUG_ON(flags & FOLL_GET);
1438		goto out;
1439	}
1440
1441	page = NULL;
1442	pgd = pgd_offset(mm, address);
1443	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1444		goto no_page_table;
1445
1446	pud = pud_offset(pgd, address);
1447	if (pud_none(*pud))
1448		goto no_page_table;
1449	if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1450		BUG_ON(flags & FOLL_GET);
1451		page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1452		goto out;
1453	}
1454	if (unlikely(pud_bad(*pud)))
1455		goto no_page_table;
1456
1457	pmd = pmd_offset(pud, address);
1458	if (pmd_none(*pmd))
1459		goto no_page_table;
1460	if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1461		BUG_ON(flags & FOLL_GET);
1462		page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1463		goto out;
1464	}
1465	if (pmd_trans_huge(*pmd)) {
1466		if (flags & FOLL_SPLIT) {
1467			split_huge_page_pmd(mm, pmd);
1468			goto split_fallthrough;
1469		}
1470		spin_lock(&mm->page_table_lock);
1471		if (likely(pmd_trans_huge(*pmd))) {
1472			if (unlikely(pmd_trans_splitting(*pmd))) {
1473				spin_unlock(&mm->page_table_lock);
1474				wait_split_huge_page(vma->anon_vma, pmd);
1475			} else {
1476				page = follow_trans_huge_pmd(mm, address,
1477							     pmd, flags);
1478				spin_unlock(&mm->page_table_lock);
1479				goto out;
1480			}
1481		} else
1482			spin_unlock(&mm->page_table_lock);
1483		/* fall through */
1484	}
1485split_fallthrough:
1486	if (unlikely(pmd_bad(*pmd)))
1487		goto no_page_table;
1488
1489	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1490
1491	pte = *ptep;
1492	if (!pte_present(pte))
1493		goto no_page;
1494	if ((flags & FOLL_WRITE) && !pte_write(pte))
1495		goto unlock;
1496
1497	page = vm_normal_page(vma, address, pte);
1498	if (unlikely(!page)) {
1499		if ((flags & FOLL_DUMP) ||
1500		    !is_zero_pfn(pte_pfn(pte)))
1501			goto bad_page;
1502		page = pte_page(pte);
1503	}
1504
1505	if (flags & FOLL_GET)
1506		get_page(page);
1507	if (flags & FOLL_TOUCH) {
1508		if ((flags & FOLL_WRITE) &&
1509		    !pte_dirty(pte) && !PageDirty(page))
1510			set_page_dirty(page);
1511		/*
1512		 * pte_mkyoung() would be more correct here, but atomic care
1513		 * is needed to avoid losing the dirty bit: it is easier to use
1514		 * mark_page_accessed().
1515		 */
1516		mark_page_accessed(page);
1517	}
1518	if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1519		/*
1520		 * The preliminary mapping check is mainly to avoid the
1521		 * pointless overhead of lock_page on the ZERO_PAGE
1522		 * which might bounce very badly if there is contention.
1523		 *
1524		 * If the page is already locked, we don't need to
1525		 * handle it now - vmscan will handle it later if and
1526		 * when it attempts to reclaim the page.
1527		 */
1528		if (page->mapping && trylock_page(page)) {
1529			lru_add_drain();  /* push cached pages to LRU */
1530			/*
1531			 * Because we lock page here and migration is
1532			 * blocked by the pte's page reference, we need
1533			 * only check for file-cache page truncation.
1534			 */
1535			if (page->mapping)
1536				mlock_vma_page(page);
1537			unlock_page(page);
1538		}
1539	}
1540unlock:
1541	pte_unmap_unlock(ptep, ptl);
1542out:
1543	return page;
1544
1545bad_page:
1546	pte_unmap_unlock(ptep, ptl);
1547	return ERR_PTR(-EFAULT);
1548
1549no_page:
1550	pte_unmap_unlock(ptep, ptl);
1551	if (!pte_none(pte))
1552		return page;
1553
1554no_page_table:
1555	/*
1556	 * When core dumping an enormous anonymous area that nobody
1557	 * has touched so far, we don't want to allocate unnecessary pages or
1558	 * page tables.  Return error instead of NULL to skip handle_mm_fault,
1559	 * then get_dump_page() will return NULL to leave a hole in the dump.
1560	 * But we can only make this optimization where a hole would surely
1561	 * be zero-filled if handle_mm_fault() actually did handle it.
1562	 */
1563	if ((flags & FOLL_DUMP) &&
1564	    (!vma->vm_ops || !vma->vm_ops->fault))
1565		return ERR_PTR(-EFAULT);
1566	return page;
1567}
1568
1569static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1570{
1571	return stack_guard_page_start(vma, addr) ||
1572	       stack_guard_page_end(vma, addr+PAGE_SIZE);
1573}
1574
1575/**
1576 * __get_user_pages() - pin user pages in memory
1577 * @tsk:	task_struct of target task
1578 * @mm:		mm_struct of target mm
1579 * @start:	starting user address
1580 * @nr_pages:	number of pages from start to pin
1581 * @gup_flags:	flags modifying pin behaviour
1582 * @pages:	array that receives pointers to the pages pinned.
1583 *		Should be at least nr_pages long. Or NULL, if caller
1584 *		only intends to ensure the pages are faulted in.
1585 * @vmas:	array of pointers to vmas corresponding to each page.
1586 *		Or NULL if the caller does not require them.
1587 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1588 *
1589 * Returns number of pages pinned. This may be fewer than the number
1590 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1591 * were pinned, returns -errno. Each page returned must be released
1592 * with a put_page() call when it is finished with. vmas will only
1593 * remain valid while mmap_sem is held.
1594 *
1595 * Must be called with mmap_sem held for read or write.
1596 *
1597 * __get_user_pages walks a process's page tables and takes a reference to
1598 * each struct page that each user address corresponds to at a given
1599 * instant. That is, it takes the page that would be accessed if a user
1600 * thread accesses the given user virtual address at that instant.
1601 *
1602 * This does not guarantee that the page exists in the user mappings when
1603 * __get_user_pages returns, and there may even be a completely different
1604 * page there in some cases (eg. if mmapped pagecache has been invalidated
1605 * and subsequently re faulted). However it does guarantee that the page
1606 * won't be freed completely. And mostly callers simply care that the page
1607 * contains data that was valid *at some point in time*. Typically, an IO
1608 * or similar operation cannot guarantee anything stronger anyway because
1609 * locks can't be held over the syscall boundary.
1610 *
1611 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1612 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1613 * appropriate) must be called after the page is finished with, and
1614 * before put_page is called.
1615 *
1616 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1617 * or mmap_sem contention, and if waiting is needed to pin all pages,
1618 * *@nonblocking will be set to 0.
1619 *
1620 * In most cases, get_user_pages or get_user_pages_fast should be used
1621 * instead of __get_user_pages. __get_user_pages should be used only if
1622 * you need some special @gup_flags.
1623 */
1624int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1625		     unsigned long start, int nr_pages, unsigned int gup_flags,
1626		     struct page **pages, struct vm_area_struct **vmas,
1627		     int *nonblocking)
1628{
1629	int i;
1630	unsigned long vm_flags;
1631
1632	if (nr_pages <= 0)
1633		return 0;
1634
1635	VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1636
1637	/* 
1638	 * Require read or write permissions.
1639	 * If FOLL_FORCE is set, we only require the "MAY" flags.
1640	 */
1641	vm_flags  = (gup_flags & FOLL_WRITE) ?
1642			(VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1643	vm_flags &= (gup_flags & FOLL_FORCE) ?
1644			(VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1645	i = 0;
1646
1647	do {
1648		struct vm_area_struct *vma;
1649
1650		vma = find_extend_vma(mm, start);
1651		if (!vma && in_gate_area(mm, start)) {
1652			unsigned long pg = start & PAGE_MASK;
1653			pgd_t *pgd;
1654			pud_t *pud;
1655			pmd_t *pmd;
1656			pte_t *pte;
1657
1658			/* user gate pages are read-only */
1659			if (gup_flags & FOLL_WRITE)
1660				return i ? : -EFAULT;
1661			if (pg > TASK_SIZE)
1662				pgd = pgd_offset_k(pg);
1663			else
1664				pgd = pgd_offset_gate(mm, pg);
1665			BUG_ON(pgd_none(*pgd));
1666			pud = pud_offset(pgd, pg);
1667			BUG_ON(pud_none(*pud));
1668			pmd = pmd_offset(pud, pg);
1669			if (pmd_none(*pmd))
1670				return i ? : -EFAULT;
1671			VM_BUG_ON(pmd_trans_huge(*pmd));
1672			pte = pte_offset_map(pmd, pg);
1673			if (pte_none(*pte)) {
1674				pte_unmap(pte);
1675				return i ? : -EFAULT;
1676			}
1677			vma = get_gate_vma(mm);
1678			if (pages) {
1679				struct page *page;
1680
1681				page = vm_normal_page(vma, start, *pte);
1682				if (!page) {
1683					if (!(gup_flags & FOLL_DUMP) &&
1684					     is_zero_pfn(pte_pfn(*pte)))
1685						page = pte_page(*pte);
1686					else {
1687						pte_unmap(pte);
1688						return i ? : -EFAULT;
1689					}
1690				}
1691				pages[i] = page;
1692				get_page(page);
1693			}
1694			pte_unmap(pte);
1695			goto next_page;
1696		}
1697
1698		if (!vma ||
1699		    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1700		    !(vm_flags & vma->vm_flags))
1701			return i ? : -EFAULT;
1702
1703		if (is_vm_hugetlb_page(vma)) {
1704			i = follow_hugetlb_page(mm, vma, pages, vmas,
1705					&start, &nr_pages, i, gup_flags);
1706			continue;
1707		}
1708
1709		do {
1710			struct page *page;
1711			unsigned int foll_flags = gup_flags;
1712
1713			/*
1714			 * If we have a pending SIGKILL, don't keep faulting
1715			 * pages and potentially allocating memory.
1716			 */
1717			if (unlikely(fatal_signal_pending(current)))
1718				return i ? i : -ERESTARTSYS;
1719
1720			cond_resched();
1721			while (!(page = follow_page(vma, start, foll_flags))) {
1722				int ret;
1723				unsigned int fault_flags = 0;
1724
1725				/* For mlock, just skip the stack guard page. */
1726				if (foll_flags & FOLL_MLOCK) {
1727					if (stack_guard_page(vma, start))
1728						goto next_page;
1729				}
1730				if (foll_flags & FOLL_WRITE)
1731					fault_flags |= FAULT_FLAG_WRITE;
1732				if (nonblocking)
1733					fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1734				if (foll_flags & FOLL_NOWAIT)
1735					fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1736
1737				ret = handle_mm_fault(mm, vma, start,
1738							fault_flags);
1739
1740				if (ret & VM_FAULT_ERROR) {
1741					if (ret & VM_FAULT_OOM)
1742						return i ? i : -ENOMEM;
1743					if (ret & (VM_FAULT_HWPOISON |
1744						   VM_FAULT_HWPOISON_LARGE)) {
1745						if (i)
1746							return i;
1747						else if (gup_flags & FOLL_HWPOISON)
1748							return -EHWPOISON;
1749						else
1750							return -EFAULT;
1751					}
1752					if (ret & VM_FAULT_SIGBUS)
1753						return i ? i : -EFAULT;
1754					BUG();
1755				}
1756
1757				if (tsk) {
1758					if (ret & VM_FAULT_MAJOR)
1759						tsk->maj_flt++;
1760					else
1761						tsk->min_flt++;
1762				}
1763
1764				if (ret & VM_FAULT_RETRY) {
1765					if (nonblocking)
1766						*nonblocking = 0;
1767					return i;
1768				}
1769
1770				/*
1771				 * The VM_FAULT_WRITE bit tells us that
1772				 * do_wp_page has broken COW when necessary,
1773				 * even if maybe_mkwrite decided not to set
1774				 * pte_write. We can thus safely do subsequent
1775				 * page lookups as if they were reads. But only
1776				 * do so when looping for pte_write is futile:
1777				 * in some cases userspace may also be wanting
1778				 * to write to the gotten user page, which a
1779				 * read fault here might prevent (a readonly
1780				 * page might get reCOWed by userspace write).
1781				 */
1782				if ((ret & VM_FAULT_WRITE) &&
1783				    !(vma->vm_flags & VM_WRITE))
1784					foll_flags &= ~FOLL_WRITE;
1785
1786				cond_resched();
1787			}
1788			if (IS_ERR(page))
1789				return i ? i : PTR_ERR(page);
1790			if (pages) {
1791				pages[i] = page;
1792
1793				flush_anon_page(vma, page, start);
1794				flush_dcache_page(page);
1795			}
1796next_page:
1797			if (vmas)
1798				vmas[i] = vma;
1799			i++;
1800			start += PAGE_SIZE;
1801			nr_pages--;
1802		} while (nr_pages && start < vma->vm_end);
1803	} while (nr_pages);
1804	return i;
1805}
1806EXPORT_SYMBOL(__get_user_pages);
1807
1808/*
1809 * fixup_user_fault() - manually resolve a user page fault
1810 * @tsk:	the task_struct to use for page fault accounting, or
1811 *		NULL if faults are not to be recorded.
1812 * @mm:		mm_struct of target mm
1813 * @address:	user address
1814 * @fault_flags:flags to pass down to handle_mm_fault()
1815 *
1816 * This is meant to be called in the specific scenario where for locking reasons
1817 * we try to access user memory in atomic context (within a pagefault_disable()
1818 * section), this returns -EFAULT, and we want to resolve the user fault before
1819 * trying again.
1820 *
1821 * Typically this is meant to be used by the futex code.
1822 *
1823 * The main difference with get_user_pages() is that this function will
1824 * unconditionally call handle_mm_fault() which will in turn perform all the
1825 * necessary SW fixup of the dirty and young bits in the PTE, while
1826 * handle_mm_fault() only guarantees to update these in the struct page.
1827 *
1828 * This is important for some architectures where those bits also gate the
1829 * access permission to the page because they are maintained in software.  On
1830 * such architectures, gup() will not be enough to make a subsequent access
1831 * succeed.
1832 *
1833 * This should be called with the mm_sem held for read.
1834 */
1835int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1836		     unsigned long address, unsigned int fault_flags)
1837{
1838	struct vm_area_struct *vma;
1839	int ret;
1840
1841	vma = find_extend_vma(mm, address);
1842	if (!vma || address < vma->vm_start)
1843		return -EFAULT;
1844
1845	ret = handle_mm_fault(mm, vma, address, fault_flags);
1846	if (ret & VM_FAULT_ERROR) {
1847		if (ret & VM_FAULT_OOM)
1848			return -ENOMEM;
1849		if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1850			return -EHWPOISON;
1851		if (ret & VM_FAULT_SIGBUS)
1852			return -EFAULT;
1853		BUG();
1854	}
1855	if (tsk) {
1856		if (ret & VM_FAULT_MAJOR)
1857			tsk->maj_flt++;
1858		else
1859			tsk->min_flt++;
1860	}
1861	return 0;
1862}
1863
1864/*
1865 * get_user_pages() - pin user pages in memory
1866 * @tsk:	the task_struct to use for page fault accounting, or
1867 *		NULL if faults are not to be recorded.
1868 * @mm:		mm_struct of target mm
1869 * @start:	starting user address
1870 * @nr_pages:	number of pages from start to pin
1871 * @write:	whether pages will be written to by the caller
1872 * @force:	whether to force write access even if user mapping is
1873 *		readonly. This will result in the page being COWed even
1874 *		in MAP_SHARED mappings. You do not want this.
1875 * @pages:	array that receives pointers to the pages pinned.
1876 *		Should be at least nr_pages long. Or NULL, if caller
1877 *		only intends to ensure the pages are faulted in.
1878 * @vmas:	array of pointers to vmas corresponding to each page.
1879 *		Or NULL if the caller does not require them.
1880 *
1881 * Returns number of pages pinned. This may be fewer than the number
1882 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1883 * were pinned, returns -errno. Each page returned must be released
1884 * with a put_page() call when it is finished with. vmas will only
1885 * remain valid while mmap_sem is held.
1886 *
1887 * Must be called with mmap_sem held for read or write.
1888 *
1889 * get_user_pages walks a process's page tables and takes a reference to
1890 * each struct page that each user address corresponds to at a given
1891 * instant. That is, it takes the page that would be accessed if a user
1892 * thread accesses the given user virtual address at that instant.
1893 *
1894 * This does not guarantee that the page exists in the user mappings when
1895 * get_user_pages returns, and there may even be a completely different
1896 * page there in some cases (eg. if mmapped pagecache has been invalidated
1897 * and subsequently re faulted). However it does guarantee that the page
1898 * won't be freed completely. And mostly callers simply care that the page
1899 * contains data that was valid *at some point in time*. Typically, an IO
1900 * or similar operation cannot guarantee anything stronger anyway because
1901 * locks can't be held over the syscall boundary.
1902 *
1903 * If write=0, the page must not be written to. If the page is written to,
1904 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1905 * after the page is finished with, and before put_page is called.
1906 *
1907 * get_user_pages is typically used for fewer-copy IO operations, to get a
1908 * handle on the memory by some means other than accesses via the user virtual
1909 * addresses. The pages may be submitted for DMA to devices or accessed via
1910 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1911 * use the correct cache flushing APIs.
1912 *
1913 * See also get_user_pages_fast, for performance critical applications.
1914 */
1915int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1916		unsigned long start, int nr_pages, int write, int force,
1917		struct page **pages, struct vm_area_struct **vmas)
1918{
1919	int flags = FOLL_TOUCH;
1920
1921	if (pages)
1922		flags |= FOLL_GET;
1923	if (write)
1924		flags |= FOLL_WRITE;
1925	if (force)
1926		flags |= FOLL_FORCE;
1927
1928	return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1929				NULL);
1930}
1931EXPORT_SYMBOL(get_user_pages);
1932
1933/**
1934 * get_dump_page() - pin user page in memory while writing it to core dump
1935 * @addr: user address
1936 *
1937 * Returns struct page pointer of user page pinned for dump,
1938 * to be freed afterwards by page_cache_release() or put_page().
1939 *
1940 * Returns NULL on any kind of failure - a hole must then be inserted into
1941 * the corefile, to preserve alignment with its headers; and also returns
1942 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1943 * allowing a hole to be left in the corefile to save diskspace.
1944 *
1945 * Called without mmap_sem, but after all other threads have been killed.
1946 */
1947#ifdef CONFIG_ELF_CORE
1948struct page *get_dump_page(unsigned long addr)
1949{
1950	struct vm_area_struct *vma;
1951	struct page *page;
1952
1953	if (__get_user_pages(current, current->mm, addr, 1,
1954			     FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1955			     NULL) < 1)
1956		return NULL;
1957	flush_cache_page(vma, addr, page_to_pfn(page));
1958	return page;
1959}
1960#endif /* CONFIG_ELF_CORE */
1961
1962pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1963			spinlock_t **ptl)
1964{
1965	pgd_t * pgd = pgd_offset(mm, addr);
1966	pud_t * pud = pud_alloc(mm, pgd, addr);
1967	if (pud) {
1968		pmd_t * pmd = pmd_alloc(mm, pud, addr);
1969		if (pmd) {
1970			VM_BUG_ON(pmd_trans_huge(*pmd));
1971			return pte_alloc_map_lock(mm, pmd, addr, ptl);
1972		}
1973	}
1974	return NULL;
1975}
1976
1977/*
1978 * This is the old fallback for page remapping.
1979 *
1980 * For historical reasons, it only allows reserved pages. Only
1981 * old drivers should use this, and they needed to mark their
1982 * pages reserved for the old functions anyway.
1983 */
1984static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1985			struct page *page, pgprot_t prot)
1986{
1987	struct mm_struct *mm = vma->vm_mm;
1988	int retval;
1989	pte_t *pte;
1990	spinlock_t *ptl;
1991
1992	retval = -EINVAL;
1993	if (PageAnon(page))
1994		goto out;
1995	retval = -ENOMEM;
1996	flush_dcache_page(page);
1997	pte = get_locked_pte(mm, addr, &ptl);
1998	if (!pte)
1999		goto out;
2000	retval = -EBUSY;
2001	if (!pte_none(*pte))
2002		goto out_unlock;
2003
2004	/* Ok, finally just insert the thing.. */
2005	get_page(page);
2006	inc_mm_counter_fast(mm, MM_FILEPAGES);
2007	page_add_file_rmap(page);
2008	set_pte_at(mm, addr, pte, mk_pte(page, prot));
2009
2010	retval = 0;
2011	pte_unmap_unlock(pte, ptl);
2012	return retval;
2013out_unlock:
2014	pte_unmap_unlock(pte, ptl);
2015out:
2016	return retval;
2017}
2018
2019/**
2020 * vm_insert_page - insert single page into user vma
2021 * @vma: user vma to map to
2022 * @addr: target user address of this page
2023 * @page: source kernel page
2024 *
2025 * This allows drivers to insert individual pages they've allocated
2026 * into a user vma.
2027 *
2028 * The page has to be a nice clean _individual_ kernel allocation.
2029 * If you allocate a compound page, you need to have marked it as
2030 * such (__GFP_COMP), or manually just split the page up yourself
2031 * (see split_page()).
2032 *
2033 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2034 * took an arbitrary page protection parameter. This doesn't allow
2035 * that. Your vma protection will have to be set up correctly, which
2036 * means that if you want a shared writable mapping, you'd better
2037 * ask for a shared writable mapping!
2038 *
2039 * The page does not need to be reserved.
 
 
 
 
 
2040 */
2041int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2042			struct page *page)
2043{
2044	if (addr < vma->vm_start || addr >= vma->vm_end)
2045		return -EFAULT;
2046	if (!page_count(page))
2047		return -EINVAL;
2048	vma->vm_flags |= VM_INSERTPAGE;
 
 
 
 
2049	return insert_page(vma, addr, page, vma->vm_page_prot);
2050}
2051EXPORT_SYMBOL(vm_insert_page);
2052
2053static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2054			unsigned long pfn, pgprot_t prot)
2055{
2056	struct mm_struct *mm = vma->vm_mm;
2057	int retval;
2058	pte_t *pte, entry;
2059	spinlock_t *ptl;
2060
2061	retval = -ENOMEM;
2062	pte = get_locked_pte(mm, addr, &ptl);
2063	if (!pte)
2064		goto out;
2065	retval = -EBUSY;
2066	if (!pte_none(*pte))
2067		goto out_unlock;
2068
2069	/* Ok, finally just insert the thing.. */
2070	entry = pte_mkspecial(pfn_pte(pfn, prot));
 
 
 
2071	set_pte_at(mm, addr, pte, entry);
2072	update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2073
2074	retval = 0;
2075out_unlock:
2076	pte_unmap_unlock(pte, ptl);
2077out:
2078	return retval;
2079}
2080
2081/**
2082 * vm_insert_pfn - insert single pfn into user vma
2083 * @vma: user vma to map to
2084 * @addr: target user address of this page
2085 * @pfn: source kernel pfn
2086 *
2087 * Similar to vm_inert_page, this allows drivers to insert individual pages
2088 * they've allocated into a user vma. Same comments apply.
2089 *
2090 * This function should only be called from a vm_ops->fault handler, and
2091 * in that case the handler should return NULL.
2092 *
2093 * vma cannot be a COW mapping.
2094 *
2095 * As this is called only for pages that do not currently exist, we
2096 * do not need to flush old virtual caches or the TLB.
2097 */
2098int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2099			unsigned long pfn)
2100{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2101	int ret;
2102	pgprot_t pgprot = vma->vm_page_prot;
2103	/*
2104	 * Technically, architectures with pte_special can avoid all these
2105	 * restrictions (same for remap_pfn_range).  However we would like
2106	 * consistency in testing and feature parity among all, so we should
2107	 * try to keep these invariants in place for everybody.
2108	 */
2109	BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2110	BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2111						(VM_PFNMAP|VM_MIXEDMAP));
2112	BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2113	BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2114
2115	if (addr < vma->vm_start || addr >= vma->vm_end)
2116		return -EFAULT;
2117	if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2118		return -EINVAL;
2119
2120	ret = insert_pfn(vma, addr, pfn, pgprot);
2121
2122	if (ret)
2123		untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2124
2125	return ret;
2126}
2127EXPORT_SYMBOL(vm_insert_pfn);
2128
2129int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2130			unsigned long pfn)
2131{
2132	BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2133
2134	if (addr < vma->vm_start || addr >= vma->vm_end)
2135		return -EFAULT;
2136
2137	/*
2138	 * If we don't have pte special, then we have to use the pfn_valid()
2139	 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2140	 * refcount the page if pfn_valid is true (hence insert_page rather
2141	 * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2142	 * without pte special, it would there be refcounted as a normal page.
2143	 */
2144	if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2145		struct page *page;
2146
2147		page = pfn_to_page(pfn);
 
 
 
 
 
2148		return insert_page(vma, addr, page, vma->vm_page_prot);
2149	}
2150	return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2151}
2152EXPORT_SYMBOL(vm_insert_mixed);
2153
2154/*
2155 * maps a range of physical memory into the requested pages. the old
2156 * mappings are removed. any references to nonexistent pages results
2157 * in null mappings (currently treated as "copy-on-access")
2158 */
2159static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2160			unsigned long addr, unsigned long end,
2161			unsigned long pfn, pgprot_t prot)
2162{
2163	pte_t *pte;
2164	spinlock_t *ptl;
2165
2166	pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2167	if (!pte)
2168		return -ENOMEM;
2169	arch_enter_lazy_mmu_mode();
2170	do {
2171		BUG_ON(!pte_none(*pte));
2172		set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2173		pfn++;
2174	} while (pte++, addr += PAGE_SIZE, addr != end);
2175	arch_leave_lazy_mmu_mode();
2176	pte_unmap_unlock(pte - 1, ptl);
2177	return 0;
2178}
2179
2180static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2181			unsigned long addr, unsigned long end,
2182			unsigned long pfn, pgprot_t prot)
2183{
2184	pmd_t *pmd;
2185	unsigned long next;
2186
2187	pfn -= addr >> PAGE_SHIFT;
2188	pmd = pmd_alloc(mm, pud, addr);
2189	if (!pmd)
2190		return -ENOMEM;
2191	VM_BUG_ON(pmd_trans_huge(*pmd));
2192	do {
2193		next = pmd_addr_end(addr, end);
2194		if (remap_pte_range(mm, pmd, addr, next,
2195				pfn + (addr >> PAGE_SHIFT), prot))
2196			return -ENOMEM;
2197	} while (pmd++, addr = next, addr != end);
2198	return 0;
2199}
2200
2201static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2202			unsigned long addr, unsigned long end,
2203			unsigned long pfn, pgprot_t prot)
2204{
2205	pud_t *pud;
2206	unsigned long next;
2207
2208	pfn -= addr >> PAGE_SHIFT;
2209	pud = pud_alloc(mm, pgd, addr);
2210	if (!pud)
2211		return -ENOMEM;
2212	do {
2213		next = pud_addr_end(addr, end);
2214		if (remap_pmd_range(mm, pud, addr, next,
2215				pfn + (addr >> PAGE_SHIFT), prot))
2216			return -ENOMEM;
2217	} while (pud++, addr = next, addr != end);
2218	return 0;
2219}
2220
2221/**
2222 * remap_pfn_range - remap kernel memory to userspace
2223 * @vma: user vma to map to
2224 * @addr: target user address to start at
2225 * @pfn: physical address of kernel memory
2226 * @size: size of map area
2227 * @prot: page protection flags for this mapping
2228 *
2229 *  Note: this is only safe if the mm semaphore is held when called.
2230 */
2231int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2232		    unsigned long pfn, unsigned long size, pgprot_t prot)
2233{
2234	pgd_t *pgd;
2235	unsigned long next;
2236	unsigned long end = addr + PAGE_ALIGN(size);
2237	struct mm_struct *mm = vma->vm_mm;
2238	int err;
2239
2240	/*
2241	 * Physically remapped pages are special. Tell the
2242	 * rest of the world about it:
2243	 *   VM_IO tells people not to look at these pages
2244	 *	(accesses can have side effects).
2245	 *   VM_RESERVED is specified all over the place, because
2246	 *	in 2.4 it kept swapout's vma scan off this vma; but
2247	 *	in 2.6 the LRU scan won't even find its pages, so this
2248	 *	flag means no more than count its pages in reserved_vm,
2249	 * 	and omit it from core dump, even when VM_IO turned off.
2250	 *   VM_PFNMAP tells the core MM that the base pages are just
2251	 *	raw PFN mappings, and do not have a "struct page" associated
2252	 *	with them.
 
 
 
 
2253	 *
2254	 * There's a horrible special case to handle copy-on-write
2255	 * behaviour that some programs depend on. We mark the "original"
2256	 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
 
2257	 */
2258	if (addr == vma->vm_start && end == vma->vm_end) {
 
 
2259		vma->vm_pgoff = pfn;
2260		vma->vm_flags |= VM_PFN_AT_MMAP;
2261	} else if (is_cow_mapping(vma->vm_flags))
 
 
2262		return -EINVAL;
2263
2264	vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2265
2266	err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2267	if (err) {
2268		/*
2269		 * To indicate that track_pfn related cleanup is not
2270		 * needed from higher level routine calling unmap_vmas
2271		 */
2272		vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2273		vma->vm_flags &= ~VM_PFN_AT_MMAP;
2274		return -EINVAL;
2275	}
2276
2277	BUG_ON(addr >= end);
2278	pfn -= addr >> PAGE_SHIFT;
2279	pgd = pgd_offset(mm, addr);
2280	flush_cache_range(vma, addr, end);
2281	do {
2282		next = pgd_addr_end(addr, end);
2283		err = remap_pud_range(mm, pgd, addr, next,
2284				pfn + (addr >> PAGE_SHIFT), prot);
2285		if (err)
2286			break;
2287	} while (pgd++, addr = next, addr != end);
2288
2289	if (err)
2290		untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2291
2292	return err;
2293}
2294EXPORT_SYMBOL(remap_pfn_range);
2295
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2296static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2297				     unsigned long addr, unsigned long end,
2298				     pte_fn_t fn, void *data)
2299{
2300	pte_t *pte;
2301	int err;
2302	pgtable_t token;
2303	spinlock_t *uninitialized_var(ptl);
2304
2305	pte = (mm == &init_mm) ?
2306		pte_alloc_kernel(pmd, addr) :
2307		pte_alloc_map_lock(mm, pmd, addr, &ptl);
2308	if (!pte)
2309		return -ENOMEM;
2310
2311	BUG_ON(pmd_huge(*pmd));
2312
2313	arch_enter_lazy_mmu_mode();
2314
2315	token = pmd_pgtable(*pmd);
2316
2317	do {
2318		err = fn(pte++, token, addr, data);
2319		if (err)
2320			break;
2321	} while (addr += PAGE_SIZE, addr != end);
2322
2323	arch_leave_lazy_mmu_mode();
2324
2325	if (mm != &init_mm)
2326		pte_unmap_unlock(pte-1, ptl);
2327	return err;
2328}
2329
2330static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2331				     unsigned long addr, unsigned long end,
2332				     pte_fn_t fn, void *data)
2333{
2334	pmd_t *pmd;
2335	unsigned long next;
2336	int err;
2337
2338	BUG_ON(pud_huge(*pud));
2339
2340	pmd = pmd_alloc(mm, pud, addr);
2341	if (!pmd)
2342		return -ENOMEM;
2343	do {
2344		next = pmd_addr_end(addr, end);
2345		err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2346		if (err)
2347			break;
2348	} while (pmd++, addr = next, addr != end);
2349	return err;
2350}
2351
2352static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2353				     unsigned long addr, unsigned long end,
2354				     pte_fn_t fn, void *data)
2355{
2356	pud_t *pud;
2357	unsigned long next;
2358	int err;
2359
2360	pud = pud_alloc(mm, pgd, addr);
2361	if (!pud)
2362		return -ENOMEM;
2363	do {
2364		next = pud_addr_end(addr, end);
2365		err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2366		if (err)
2367			break;
2368	} while (pud++, addr = next, addr != end);
2369	return err;
2370}
2371
2372/*
2373 * Scan a region of virtual memory, filling in page tables as necessary
2374 * and calling a provided function on each leaf page table.
2375 */
2376int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2377			unsigned long size, pte_fn_t fn, void *data)
2378{
2379	pgd_t *pgd;
2380	unsigned long next;
2381	unsigned long end = addr + size;
2382	int err;
2383
2384	BUG_ON(addr >= end);
 
 
2385	pgd = pgd_offset(mm, addr);
2386	do {
2387		next = pgd_addr_end(addr, end);
2388		err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2389		if (err)
2390			break;
2391	} while (pgd++, addr = next, addr != end);
2392
2393	return err;
2394}
2395EXPORT_SYMBOL_GPL(apply_to_page_range);
2396
2397/*
2398 * handle_pte_fault chooses page fault handler according to an entry
2399 * which was read non-atomically.  Before making any commitment, on
2400 * those architectures or configurations (e.g. i386 with PAE) which
2401 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2402 * must check under lock before unmapping the pte and proceeding
2403 * (but do_wp_page is only called after already making such a check;
2404 * and do_anonymous_page can safely check later on).
2405 */
2406static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2407				pte_t *page_table, pte_t orig_pte)
2408{
2409	int same = 1;
2410#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2411	if (sizeof(pte_t) > sizeof(unsigned long)) {
2412		spinlock_t *ptl = pte_lockptr(mm, pmd);
2413		spin_lock(ptl);
2414		same = pte_same(*page_table, orig_pte);
2415		spin_unlock(ptl);
2416	}
2417#endif
2418	pte_unmap(page_table);
2419	return same;
2420}
2421
2422static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2423{
 
 
2424	/*
2425	 * If the source page was a PFN mapping, we don't have
2426	 * a "struct page" for it. We do a best-effort copy by
2427	 * just copying from the original user address. If that
2428	 * fails, we just zero-fill it. Live with it.
2429	 */
2430	if (unlikely(!src)) {
2431		void *kaddr = kmap_atomic(dst, KM_USER0);
2432		void __user *uaddr = (void __user *)(va & PAGE_MASK);
2433
2434		/*
2435		 * This really shouldn't fail, because the page is there
2436		 * in the page tables. But it might just be unreadable,
2437		 * in which case we just give up and fill the result with
2438		 * zeroes.
2439		 */
2440		if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2441			clear_page(kaddr);
2442		kunmap_atomic(kaddr, KM_USER0);
2443		flush_dcache_page(dst);
2444	} else
2445		copy_user_highpage(dst, src, va, vma);
2446}
2447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2448/*
2449 * This routine handles present pages, when users try to write
2450 * to a shared page. It is done by copying the page to a new address
2451 * and decrementing the shared-page counter for the old page.
2452 *
2453 * Note that this routine assumes that the protection checks have been
2454 * done by the caller (the low-level page fault routine in most cases).
2455 * Thus we can safely just mark it writable once we've done any necessary
2456 * COW.
2457 *
2458 * We also mark the page dirty at this point even though the page will
2459 * change only once the write actually happens. This avoids a few races,
2460 * and potentially makes it more efficient.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2461 *
2462 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2463 * but allow concurrent faults), with pte both mapped and locked.
2464 * We return with mmap_sem still held, but pte unmapped and unlocked.
2465 */
2466static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2467		unsigned long address, pte_t *page_table, pmd_t *pmd,
2468		spinlock_t *ptl, pte_t orig_pte)
 
 
 
2469	__releases(ptl)
2470{
2471	struct page *old_page, *new_page;
2472	pte_t entry;
2473	int ret = 0;
2474	int page_mkwrite = 0;
2475	struct page *dirty_page = NULL;
2476
2477	old_page = vm_normal_page(vma, address, orig_pte);
2478	if (!old_page) {
2479		/*
2480		 * VM_MIXEDMAP !pfn_valid() case
2481		 *
2482		 * We should not cow pages in a shared writeable mapping.
2483		 * Just mark the pages writable as we can't do any dirty
2484		 * accounting on raw pfn maps.
2485		 */
2486		if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2487				     (VM_WRITE|VM_SHARED))
2488			goto reuse;
2489		goto gotten;
2490	}
2491
2492	/*
2493	 * Take out anonymous pages first, anonymous shared vmas are
2494	 * not dirty accountable.
 
2495	 */
2496	if (PageAnon(old_page) && !PageKsm(old_page)) {
2497		if (!trylock_page(old_page)) {
2498			page_cache_get(old_page);
2499			pte_unmap_unlock(page_table, ptl);
2500			lock_page(old_page);
2501			page_table = pte_offset_map_lock(mm, pmd, address,
2502							 &ptl);
2503			if (!pte_same(*page_table, orig_pte)) {
2504				unlock_page(old_page);
2505				goto unlock;
2506			}
2507			page_cache_release(old_page);
2508		}
2509		if (reuse_swap_page(old_page)) {
2510			/*
2511			 * The page is all ours.  Move it to our anon_vma so
2512			 * the rmap code will not search our parent or siblings.
2513			 * Protected against the rmap code by the page lock.
2514			 */
2515			page_move_anon_rmap(old_page, vma, address);
2516			unlock_page(old_page);
2517			goto reuse;
2518		}
2519		unlock_page(old_page);
2520	} else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2521					(VM_WRITE|VM_SHARED))) {
2522		/*
2523		 * Only catch write-faults on shared writable pages,
2524		 * read-only shared pages can get COWed by
2525		 * get_user_pages(.write=1, .force=1).
2526		 */
2527		if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2528			struct vm_fault vmf;
2529			int tmp;
2530
2531			vmf.virtual_address = (void __user *)(address &
2532								PAGE_MASK);
2533			vmf.pgoff = old_page->index;
2534			vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2535			vmf.page = old_page;
2536
2537			/*
2538			 * Notify the address space that the page is about to
2539			 * become writable so that it can prohibit this or wait
2540			 * for the page to get into an appropriate state.
2541			 *
2542			 * We do this without the lock held, so that it can
2543			 * sleep if it needs to.
2544			 */
2545			page_cache_get(old_page);
2546			pte_unmap_unlock(page_table, ptl);
2547
2548			tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2549			if (unlikely(tmp &
2550					(VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2551				ret = tmp;
2552				goto unwritable_page;
2553			}
2554			if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2555				lock_page(old_page);
2556				if (!old_page->mapping) {
2557					ret = 0; /* retry the fault */
2558					unlock_page(old_page);
2559					goto unwritable_page;
2560				}
2561			} else
2562				VM_BUG_ON(!PageLocked(old_page));
2563
 
2564			/*
2565			 * Since we dropped the lock we need to revalidate
2566			 * the PTE as someone else may have changed it.  If
2567			 * they did, we just return, as we can count on the
2568			 * MMU to tell us if they didn't also make it writable.
2569			 */
2570			page_table = pte_offset_map_lock(mm, pmd, address,
2571							 &ptl);
2572			if (!pte_same(*page_table, orig_pte)) {
2573				unlock_page(old_page);
2574				goto unlock;
2575			}
2576
2577			page_mkwrite = 1;
2578		}
2579		dirty_page = old_page;
2580		get_page(dirty_page);
2581
2582reuse:
2583		flush_cache_page(vma, address, pte_pfn(orig_pte));
2584		entry = pte_mkyoung(orig_pte);
2585		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2586		if (ptep_set_access_flags(vma, address, page_table, entry,1))
2587			update_mmu_cache(vma, address, page_table);
2588		pte_unmap_unlock(page_table, ptl);
2589		ret |= VM_FAULT_WRITE;
2590
2591		if (!dirty_page)
2592			return ret;
2593
2594		/*
2595		 * Yes, Virginia, this is actually required to prevent a race
2596		 * with clear_page_dirty_for_io() from clearing the page dirty
2597		 * bit after it clear all dirty ptes, but before a racing
2598		 * do_wp_page installs a dirty pte.
2599		 *
2600		 * __do_fault is protected similarly.
2601		 */
2602		if (!page_mkwrite) {
2603			wait_on_page_locked(dirty_page);
2604			set_page_dirty_balance(dirty_page, page_mkwrite);
2605		}
2606		put_page(dirty_page);
2607		if (page_mkwrite) {
2608			struct address_space *mapping = dirty_page->mapping;
2609
2610			set_page_dirty(dirty_page);
2611			unlock_page(dirty_page);
2612			page_cache_release(dirty_page);
2613			if (mapping)	{
2614				/*
2615				 * Some device drivers do not set page.mapping
2616				 * but still dirty their pages
2617				 */
2618				balance_dirty_pages_ratelimited(mapping);
2619			}
2620		}
2621
2622		/* file_update_time outside page_lock */
2623		if (vma->vm_file)
2624			file_update_time(vma->vm_file);
 
2625
2626		return ret;
2627	}
2628
2629	/*
2630	 * Ok, we need to copy. Oh, well..
2631	 */
2632	page_cache_get(old_page);
2633gotten:
2634	pte_unmap_unlock(page_table, ptl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2635
2636	if (unlikely(anon_vma_prepare(vma)))
2637		goto oom;
2638
2639	if (is_zero_pfn(pte_pfn(orig_pte))) {
2640		new_page = alloc_zeroed_user_highpage_movable(vma, address);
2641		if (!new_page)
2642			goto oom;
2643	} else {
2644		new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2645		if (!new_page)
2646			goto oom;
2647		cow_user_page(new_page, old_page, address, vma);
2648	}
 
 
 
 
2649	__SetPageUptodate(new_page);
2650
2651	if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2652		goto oom_free_new;
2653
2654	/*
2655	 * Re-check the pte - we dropped the lock
2656	 */
2657	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2658	if (likely(pte_same(*page_table, orig_pte))) {
2659		if (old_page) {
2660			if (!PageAnon(old_page)) {
2661				dec_mm_counter_fast(mm, MM_FILEPAGES);
 
2662				inc_mm_counter_fast(mm, MM_ANONPAGES);
2663			}
2664		} else
2665			inc_mm_counter_fast(mm, MM_ANONPAGES);
 
2666		flush_cache_page(vma, address, pte_pfn(orig_pte));
2667		entry = mk_pte(new_page, vma->vm_page_prot);
2668		entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2669		/*
2670		 * Clear the pte entry and flush it first, before updating the
2671		 * pte with the new entry. This will avoid a race condition
2672		 * seen in the presence of one thread doing SMC and another
2673		 * thread doing COW.
2674		 */
2675		ptep_clear_flush(vma, address, page_table);
2676		page_add_new_anon_rmap(new_page, vma, address);
 
 
2677		/*
2678		 * We call the notify macro here because, when using secondary
2679		 * mmu page tables (such as kvm shadow page tables), we want the
2680		 * new page to be mapped directly into the secondary page table.
2681		 */
2682		set_pte_at_notify(mm, address, page_table, entry);
2683		update_mmu_cache(vma, address, page_table);
2684		if (old_page) {
2685			/*
2686			 * Only after switching the pte to the new page may
2687			 * we remove the mapcount here. Otherwise another
2688			 * process may come and find the rmap count decremented
2689			 * before the pte is switched to the new page, and
2690			 * "reuse" the old page writing into it while our pte
2691			 * here still points into it and can be read by other
2692			 * threads.
2693			 *
2694			 * The critical issue is to order this
2695			 * page_remove_rmap with the ptp_clear_flush above.
2696			 * Those stores are ordered by (if nothing else,)
2697			 * the barrier present in the atomic_add_negative
2698			 * in page_remove_rmap.
2699			 *
2700			 * Then the TLB flush in ptep_clear_flush ensures that
2701			 * no process can access the old page before the
2702			 * decremented mapcount is visible. And the old page
2703			 * cannot be reused until after the decremented
2704			 * mapcount is visible. So transitively, TLBs to
2705			 * old page will be flushed before it can be reused.
2706			 */
2707			page_remove_rmap(old_page);
2708		}
2709
2710		/* Free the old page.. */
2711		new_page = old_page;
2712		ret |= VM_FAULT_WRITE;
2713	} else
2714		mem_cgroup_uncharge_page(new_page);
 
2715
2716	if (new_page)
2717		page_cache_release(new_page);
2718unlock:
2719	pte_unmap_unlock(page_table, ptl);
 
2720	if (old_page) {
2721		/*
2722		 * Don't let another task, with possibly unlocked vma,
2723		 * keep the mlocked page.
2724		 */
2725		if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2726			lock_page(old_page);	/* LRU manipulation */
2727			munlock_vma_page(old_page);
 
2728			unlock_page(old_page);
2729		}
2730		page_cache_release(old_page);
2731	}
2732	return ret;
2733oom_free_new:
2734	page_cache_release(new_page);
2735oom:
2736	if (old_page) {
2737		if (page_mkwrite) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2738			unlock_page(old_page);
2739			page_cache_release(old_page);
 
2740		}
2741		page_cache_release(old_page);
 
 
 
 
2742	}
2743	return VM_FAULT_OOM;
2744
2745unwritable_page:
2746	page_cache_release(old_page);
2747	return ret;
 
 
 
 
 
2748}
2749
2750static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2751		unsigned long start_addr, unsigned long end_addr,
2752		struct zap_details *details)
2753{
2754	zap_page_range(vma, start_addr, end_addr - start_addr, details);
2755}
2756
2757static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2758					    struct zap_details *details)
2759{
2760	struct vm_area_struct *vma;
2761	struct prio_tree_iter iter;
2762	pgoff_t vba, vea, zba, zea;
2763
2764	vma_prio_tree_foreach(vma, &iter, root,
2765			details->first_index, details->last_index) {
2766
2767		vba = vma->vm_pgoff;
2768		vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2769		/* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2770		zba = details->first_index;
2771		if (zba < vba)
2772			zba = vba;
2773		zea = details->last_index;
2774		if (zea > vea)
2775			zea = vea;
2776
2777		unmap_mapping_range_vma(vma,
2778			((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2779			((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2780				details);
2781	}
2782}
2783
2784static inline void unmap_mapping_range_list(struct list_head *head,
2785					    struct zap_details *details)
2786{
2787	struct vm_area_struct *vma;
2788
2789	/*
2790	 * In nonlinear VMAs there is no correspondence between virtual address
2791	 * offset and file offset.  So we must perform an exhaustive search
2792	 * across *all* the pages in each nonlinear VMA, not just the pages
2793	 * whose virtual address lies outside the file truncation point.
2794	 */
2795	list_for_each_entry(vma, head, shared.vm_set.list) {
2796		details->nonlinear_vma = vma;
2797		unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2798	}
2799}
2800
2801/**
2802 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
 
 
 
2803 * @mapping: the address space containing mmaps to be unmapped.
2804 * @holebegin: byte in first page to unmap, relative to the start of
2805 * the underlying file.  This will be rounded down to a PAGE_SIZE
2806 * boundary.  Note that this is different from truncate_pagecache(), which
2807 * must keep the partial page.  In contrast, we must get rid of
2808 * partial pages.
2809 * @holelen: size of prospective hole in bytes.  This will be rounded
2810 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2811 * end of the file.
2812 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2813 * but 0 when invalidating pagecache, don't throw away private data.
2814 */
2815void unmap_mapping_range(struct address_space *mapping,
2816		loff_t const holebegin, loff_t const holelen, int even_cows)
2817{
2818	struct zap_details details;
2819	pgoff_t hba = holebegin >> PAGE_SHIFT;
2820	pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2821
2822	/* Check for overflow. */
2823	if (sizeof(holelen) > sizeof(hlen)) {
2824		long long holeend =
2825			(holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2826		if (holeend & ~(long long)ULONG_MAX)
2827			hlen = ULONG_MAX - hba + 1;
2828	}
2829
2830	details.check_mapping = even_cows? NULL: mapping;
2831	details.nonlinear_vma = NULL;
2832	details.first_index = hba;
2833	details.last_index = hba + hlen - 1;
2834	if (details.last_index < details.first_index)
2835		details.last_index = ULONG_MAX;
2836
2837
2838	mutex_lock(&mapping->i_mmap_mutex);
2839	if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
 
2840		unmap_mapping_range_tree(&mapping->i_mmap, &details);
2841	if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2842		unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2843	mutex_unlock(&mapping->i_mmap_mutex);
2844}
2845EXPORT_SYMBOL(unmap_mapping_range);
2846
2847/*
2848 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2849 * but allow concurrent faults), and pte mapped but not yet locked.
2850 * We return with mmap_sem still held, but pte unmapped and unlocked.
 
 
 
2851 */
2852static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2853		unsigned long address, pte_t *page_table, pmd_t *pmd,
2854		unsigned int flags, pte_t orig_pte)
2855{
2856	spinlock_t *ptl;
2857	struct page *page, *swapcache = NULL;
 
2858	swp_entry_t entry;
2859	pte_t pte;
2860	int locked;
2861	struct mem_cgroup *ptr;
2862	int exclusive = 0;
2863	int ret = 0;
2864
2865	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2866		goto out;
2867
2868	entry = pte_to_swp_entry(orig_pte);
2869	if (unlikely(non_swap_entry(entry))) {
2870		if (is_migration_entry(entry)) {
2871			migration_entry_wait(mm, pmd, address);
2872		} else if (is_hwpoison_entry(entry)) {
2873			ret = VM_FAULT_HWPOISON;
2874		} else {
2875			print_bad_pte(vma, address, orig_pte, NULL);
2876			ret = VM_FAULT_SIGBUS;
2877		}
2878		goto out;
2879	}
2880	delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2881	page = lookup_swap_cache(entry);
2882	if (!page) {
2883		grab_swap_token(mm); /* Contend for token _before_ read-in */
2884		page = swapin_readahead(entry,
2885					GFP_HIGHUSER_MOVABLE, vma, address);
2886		if (!page) {
2887			/*
2888			 * Back out if somebody else faulted in this pte
2889			 * while we released the pte lock.
2890			 */
2891			page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2892			if (likely(pte_same(*page_table, orig_pte)))
2893				ret = VM_FAULT_OOM;
2894			delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2895			goto unlock;
2896		}
2897
2898		/* Had to read the page from swap area: Major fault */
2899		ret = VM_FAULT_MAJOR;
2900		count_vm_event(PGMAJFAULT);
2901		mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2902	} else if (PageHWPoison(page)) {
2903		/*
2904		 * hwpoisoned dirty swapcache pages are kept for killing
2905		 * owner processes (which may be unknown at hwpoison time)
2906		 */
2907		ret = VM_FAULT_HWPOISON;
2908		delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
 
2909		goto out_release;
2910	}
2911
 
2912	locked = lock_page_or_retry(page, mm, flags);
 
2913	delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2914	if (!locked) {
2915		ret |= VM_FAULT_RETRY;
2916		goto out_release;
2917	}
2918
2919	/*
2920	 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2921	 * release the swapcache from under us.  The page pin, and pte_same
2922	 * test below, are not enough to exclude that.  Even if it is still
2923	 * swapcache, we need to check that the page's swap has not changed.
2924	 */
2925	if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2926		goto out_page;
2927
2928	if (ksm_might_need_to_copy(page, vma, address)) {
2929		swapcache = page;
2930		page = ksm_does_need_to_copy(page, vma, address);
2931
2932		if (unlikely(!page)) {
2933			ret = VM_FAULT_OOM;
2934			page = swapcache;
2935			swapcache = NULL;
2936			goto out_page;
2937		}
2938	}
2939
2940	if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2941		ret = VM_FAULT_OOM;
2942		goto out_page;
2943	}
2944
2945	/*
2946	 * Back out if somebody else already faulted in this pte.
2947	 */
2948	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2949	if (unlikely(!pte_same(*page_table, orig_pte)))
2950		goto out_nomap;
2951
2952	if (unlikely(!PageUptodate(page))) {
2953		ret = VM_FAULT_SIGBUS;
2954		goto out_nomap;
2955	}
2956
2957	/*
2958	 * The page isn't present yet, go ahead with the fault.
2959	 *
2960	 * Be careful about the sequence of operations here.
2961	 * To get its accounting right, reuse_swap_page() must be called
2962	 * while the page is counted on swap but not yet in mapcount i.e.
2963	 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2964	 * must be called after the swap_free(), or it will never succeed.
2965	 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2966	 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2967	 * in page->private. In this case, a record in swap_cgroup  is silently
2968	 * discarded at swap_free().
2969	 */
2970
2971	inc_mm_counter_fast(mm, MM_ANONPAGES);
2972	dec_mm_counter_fast(mm, MM_SWAPENTS);
2973	pte = mk_pte(page, vma->vm_page_prot);
2974	if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2975		pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2976		flags &= ~FAULT_FLAG_WRITE;
2977		ret |= VM_FAULT_WRITE;
2978		exclusive = 1;
2979	}
2980	flush_icache_page(vma, page);
 
 
2981	set_pte_at(mm, address, page_table, pte);
2982	do_page_add_anon_rmap(page, vma, address, exclusive);
2983	/* It's better to call commit-charge after rmap is established */
2984	mem_cgroup_commit_charge_swapin(page, ptr);
 
 
 
 
 
2985
2986	swap_free(entry);
2987	if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
 
2988		try_to_free_swap(page);
2989	unlock_page(page);
2990	if (swapcache) {
2991		/*
2992		 * Hold the lock to avoid the swap entry to be reused
2993		 * until we take the PT lock for the pte_same() check
2994		 * (to avoid false positives from pte_same). For
2995		 * further safety release the lock after the swap_free
2996		 * so that the swap count won't change under a
2997		 * parallel locked swapcache.
2998		 */
2999		unlock_page(swapcache);
3000		page_cache_release(swapcache);
3001	}
3002
3003	if (flags & FAULT_FLAG_WRITE) {
3004		ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3005		if (ret & VM_FAULT_ERROR)
3006			ret &= VM_FAULT_ERROR;
3007		goto out;
3008	}
3009
3010	/* No need to invalidate - it was non-present before */
3011	update_mmu_cache(vma, address, page_table);
3012unlock:
3013	pte_unmap_unlock(page_table, ptl);
3014out:
3015	return ret;
3016out_nomap:
3017	mem_cgroup_cancel_charge_swapin(ptr);
3018	pte_unmap_unlock(page_table, ptl);
3019out_page:
3020	unlock_page(page);
3021out_release:
3022	page_cache_release(page);
3023	if (swapcache) {
3024		unlock_page(swapcache);
3025		page_cache_release(swapcache);
3026	}
3027	return ret;
3028}
3029
3030/*
3031 * This is like a special single-page "expand_{down|up}wards()",
3032 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3033 * doesn't hit another vma.
3034 */
3035static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3036{
3037	address &= PAGE_MASK;
3038	if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3039		struct vm_area_struct *prev = vma->vm_prev;
3040
3041		/*
3042		 * Is there a mapping abutting this one below?
3043		 *
3044		 * That's only ok if it's the same stack mapping
3045		 * that has gotten split..
3046		 */
3047		if (prev && prev->vm_end == address)
3048			return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3049
3050		expand_downwards(vma, address - PAGE_SIZE);
3051	}
3052	if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3053		struct vm_area_struct *next = vma->vm_next;
3054
3055		/* As VM_GROWSDOWN but s/below/above/ */
3056		if (next && next->vm_start == address + PAGE_SIZE)
3057			return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3058
3059		expand_upwards(vma, address + PAGE_SIZE);
3060	}
3061	return 0;
3062}
3063
3064/*
3065 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3066 * but allow concurrent faults), and pte mapped but not yet locked.
3067 * We return with mmap_sem still held, but pte unmapped and unlocked.
3068 */
3069static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3070		unsigned long address, pte_t *page_table, pmd_t *pmd,
3071		unsigned int flags)
3072{
 
3073	struct page *page;
3074	spinlock_t *ptl;
3075	pte_t entry;
3076
3077	pte_unmap(page_table);
3078
 
 
 
 
3079	/* Check if we need to add a guard page to the stack */
3080	if (check_stack_guard_page(vma, address) < 0)
3081		return VM_FAULT_SIGBUS;
3082
3083	/* Use the zero-page for reads */
3084	if (!(flags & FAULT_FLAG_WRITE)) {
3085		entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3086						vma->vm_page_prot));
3087		page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3088		if (!pte_none(*page_table))
3089			goto unlock;
 
 
 
 
 
 
3090		goto setpte;
3091	}
3092
3093	/* Allocate our own private page. */
3094	if (unlikely(anon_vma_prepare(vma)))
3095		goto oom;
3096	page = alloc_zeroed_user_highpage_movable(vma, address);
3097	if (!page)
3098		goto oom;
3099	__SetPageUptodate(page);
3100
3101	if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3102		goto oom_free_page;
3103
 
 
 
 
 
 
 
3104	entry = mk_pte(page, vma->vm_page_prot);
3105	if (vma->vm_flags & VM_WRITE)
3106		entry = pte_mkwrite(pte_mkdirty(entry));
3107
3108	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3109	if (!pte_none(*page_table))
3110		goto release;
3111
 
 
 
 
 
 
 
 
 
3112	inc_mm_counter_fast(mm, MM_ANONPAGES);
3113	page_add_new_anon_rmap(page, vma, address);
 
 
3114setpte:
3115	set_pte_at(mm, address, page_table, entry);
3116
3117	/* No need to invalidate - it was non-present before */
3118	update_mmu_cache(vma, address, page_table);
3119unlock:
3120	pte_unmap_unlock(page_table, ptl);
3121	return 0;
3122release:
3123	mem_cgroup_uncharge_page(page);
3124	page_cache_release(page);
3125	goto unlock;
3126oom_free_page:
3127	page_cache_release(page);
3128oom:
3129	return VM_FAULT_OOM;
3130}
3131
3132/*
3133 * __do_fault() tries to create a new page mapping. It aggressively
3134 * tries to share with existing pages, but makes a separate copy if
3135 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3136 * the next page fault.
3137 *
3138 * As this is called only for pages that do not currently exist, we
3139 * do not need to flush old virtual caches or the TLB.
3140 *
3141 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3142 * but allow concurrent faults), and pte neither mapped nor locked.
3143 * We return with mmap_sem still held, but pte unmapped and unlocked.
3144 */
3145static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3146		unsigned long address, pmd_t *pmd,
3147		pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3148{
3149	pte_t *page_table;
3150	spinlock_t *ptl;
3151	struct page *page;
3152	struct page *cow_page;
3153	pte_t entry;
3154	int anon = 0;
3155	struct page *dirty_page = NULL;
3156	struct vm_fault vmf;
3157	int ret;
3158	int page_mkwrite = 0;
3159
3160	/*
3161	 * If we do COW later, allocate page befor taking lock_page()
3162	 * on the file cache page. This will reduce lock holding time.
3163	 */
3164	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3165
3166		if (unlikely(anon_vma_prepare(vma)))
3167			return VM_FAULT_OOM;
3168
3169		cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3170		if (!cow_page)
3171			return VM_FAULT_OOM;
3172
3173		if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3174			page_cache_release(cow_page);
3175			return VM_FAULT_OOM;
3176		}
3177	} else
3178		cow_page = NULL;
3179
3180	vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3181	vmf.pgoff = pgoff;
3182	vmf.flags = flags;
3183	vmf.page = NULL;
 
 
3184
3185	ret = vma->vm_ops->fault(vma, &vmf);
3186	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3187			    VM_FAULT_RETRY)))
3188		goto uncharge_out;
 
3189
3190	if (unlikely(PageHWPoison(vmf.page))) {
3191		if (ret & VM_FAULT_LOCKED)
3192			unlock_page(vmf.page);
3193		ret = VM_FAULT_HWPOISON;
3194		goto uncharge_out;
3195	}
3196
3197	/*
3198	 * For consistency in subsequent calls, make the faulted page always
3199	 * locked.
3200	 */
3201	if (unlikely(!(ret & VM_FAULT_LOCKED)))
3202		lock_page(vmf.page);
3203	else
3204		VM_BUG_ON(!PageLocked(vmf.page));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3205
3206	/*
3207	 * Should we do an early C-O-W break?
 
3208	 */
3209	page = vmf.page;
3210	if (flags & FAULT_FLAG_WRITE) {
3211		if (!(vma->vm_flags & VM_SHARED)) {
3212			page = cow_page;
3213			anon = 1;
3214			copy_user_highpage(page, vmf.page, address, vma);
3215			__SetPageUptodate(page);
3216		} else {
3217			/*
3218			 * If the page will be shareable, see if the backing
3219			 * address space wants to know that the page is about
3220			 * to become writable
3221			 */
3222			if (vma->vm_ops->page_mkwrite) {
3223				int tmp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3224
3225				unlock_page(page);
3226				vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3227				tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3228				if (unlikely(tmp &
3229					  (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3230					ret = tmp;
3231					goto unwritable_page;
3232				}
3233				if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3234					lock_page(page);
3235					if (!page->mapping) {
3236						ret = 0; /* retry the fault */
3237						unlock_page(page);
3238						goto unwritable_page;
3239					}
3240				} else
3241					VM_BUG_ON(!PageLocked(page));
3242				page_mkwrite = 1;
3243			}
3244		}
3245
 
 
 
 
 
 
3246	}
 
 
 
 
 
 
3247
3248	page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
 
 
 
 
 
 
 
 
 
 
 
3249
3250	/*
3251	 * This silly early PAGE_DIRTY setting removes a race
3252	 * due to the bad i386 page protection. But it's valid
3253	 * for other architectures too.
3254	 *
3255	 * Note that if FAULT_FLAG_WRITE is set, we either now have
3256	 * an exclusive copy of the page, or this is a shared mapping,
3257	 * so we can make it writable and dirty to avoid having to
3258	 * handle that later.
3259	 */
3260	/* Only go through if we didn't race with anybody else... */
3261	if (likely(pte_same(*page_table, orig_pte))) {
3262		flush_icache_page(vma, page);
3263		entry = mk_pte(page, vma->vm_page_prot);
3264		if (flags & FAULT_FLAG_WRITE)
3265			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3266		if (anon) {
3267			inc_mm_counter_fast(mm, MM_ANONPAGES);
3268			page_add_new_anon_rmap(page, vma, address);
3269		} else {
3270			inc_mm_counter_fast(mm, MM_FILEPAGES);
3271			page_add_file_rmap(page);
3272			if (flags & FAULT_FLAG_WRITE) {
3273				dirty_page = page;
3274				get_page(dirty_page);
3275			}
3276		}
3277		set_pte_at(mm, address, page_table, entry);
3278
3279		/* no need to invalidate: a not-present page won't be cached */
3280		update_mmu_cache(vma, address, page_table);
3281	} else {
3282		if (cow_page)
3283			mem_cgroup_uncharge_page(cow_page);
3284		if (anon)
3285			page_cache_release(page);
3286		else
3287			anon = 1; /* no anon but release faulted_page */
3288	}
3289
3290	pte_unmap_unlock(page_table, ptl);
 
 
3291
3292	if (dirty_page) {
3293		struct address_space *mapping = page->mapping;
 
3294
3295		if (set_page_dirty(dirty_page))
3296			page_mkwrite = 1;
3297		unlock_page(dirty_page);
3298		put_page(dirty_page);
3299		if (page_mkwrite && mapping) {
 
 
3300			/*
3301			 * Some device drivers do not set page.mapping but still
3302			 * dirty their pages
3303			 */
3304			balance_dirty_pages_ratelimited(mapping);
3305		}
3306
3307		/* file_update_time outside page_lock */
3308		if (vma->vm_file)
3309			file_update_time(vma->vm_file);
 
 
 
 
 
3310	} else {
3311		unlock_page(vmf.page);
3312		if (anon)
3313			page_cache_release(vmf.page);
 
 
3314	}
3315
3316	return ret;
3317
3318unwritable_page:
3319	page_cache_release(page);
3320	return ret;
3321uncharge_out:
3322	/* fs's fault handler get error */
3323	if (cow_page) {
3324		mem_cgroup_uncharge_page(cow_page);
3325		page_cache_release(cow_page);
3326	}
3327	return ret;
3328}
3329
3330static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3331		unsigned long address, pte_t *page_table, pmd_t *pmd,
3332		unsigned int flags, pte_t orig_pte)
3333{
3334	pgoff_t pgoff = (((address & PAGE_MASK)
3335			- vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3336
3337	pte_unmap(page_table);
3338	return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
 
 
3339}
3340
3341/*
3342 * Fault of a previously existing named mapping. Repopulate the pte
3343 * from the encoded file_pte if possible. This enables swappable
3344 * nonlinear vmas.
3345 *
3346 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3347 * but allow concurrent faults), and pte mapped but not yet locked.
3348 * We return with mmap_sem still held, but pte unmapped and unlocked.
 
3349 */
3350static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3351		unsigned long address, pte_t *page_table, pmd_t *pmd,
3352		unsigned int flags, pte_t orig_pte)
3353{
3354	pgoff_t pgoff;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3355
3356	flags |= FAULT_FLAG_NONLINEAR;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3357
3358	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
 
 
3359		return 0;
 
3360
3361	if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3362		/*
3363		 * Page table corrupted: show pte and kill process.
3364		 */
3365		print_bad_pte(vma, address, orig_pte, NULL);
3366		return VM_FAULT_SIGBUS;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3367	}
3368
3369	pgoff = pte_to_pgoff(orig_pte);
3370	return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3371}
3372
3373/*
3374 * These routines also need to handle stuff like marking pages dirty
3375 * and/or accessed for architectures that don't do it in hardware (most
3376 * RISC architectures).  The early dirtying is also good on the i386.
3377 *
3378 * There is also a hook called "update_mmu_cache()" that architectures
3379 * with external mmu caches can use to update those (ie the Sparc or
3380 * PowerPC hashed page tables that act as extended TLBs).
3381 *
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
 
 
 
3385 */
3386int handle_pte_fault(struct mm_struct *mm,
3387		     struct vm_area_struct *vma, unsigned long address,
3388		     pte_t *pte, pmd_t *pmd, unsigned int flags)
3389{
3390	pte_t entry;
3391	spinlock_t *ptl;
3392
 
 
 
 
 
 
 
 
3393	entry = *pte;
 
3394	if (!pte_present(entry)) {
3395		if (pte_none(entry)) {
3396			if (vma->vm_ops) {
3397				if (likely(vma->vm_ops->fault))
3398					return do_linear_fault(mm, vma, address,
3399						pte, pmd, flags, entry);
3400			}
3401			return do_anonymous_page(mm, vma, address,
3402						 pte, pmd, flags);
3403		}
3404		if (pte_file(entry))
3405			return do_nonlinear_fault(mm, vma, address,
3406					pte, pmd, flags, entry);
3407		return do_swap_page(mm, vma, address,
3408					pte, pmd, flags, entry);
3409	}
3410
 
 
 
3411	ptl = pte_lockptr(mm, pmd);
3412	spin_lock(ptl);
3413	if (unlikely(!pte_same(*pte, entry)))
3414		goto unlock;
3415	if (flags & FAULT_FLAG_WRITE) {
3416		if (!pte_write(entry))
3417			return do_wp_page(mm, vma, address,
3418					pte, pmd, ptl, entry);
3419		entry = pte_mkdirty(entry);
3420	}
3421	entry = pte_mkyoung(entry);
3422	if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3423		update_mmu_cache(vma, address, pte);
3424	} else {
3425		/*
3426		 * This is needed only for protection faults but the arch code
3427		 * is not yet telling us if this is a protection fault or not.
3428		 * This still avoids useless tlb flushes for .text page faults
3429		 * with threads.
3430		 */
3431		if (flags & FAULT_FLAG_WRITE)
3432			flush_tlb_fix_spurious_fault(vma, address);
3433	}
3434unlock:
3435	pte_unmap_unlock(pte, ptl);
3436	return 0;
3437}
3438
3439/*
3440 * By the time we get here, we already hold the mm semaphore
 
 
 
3441 */
3442int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3443		unsigned long address, unsigned int flags)
3444{
3445	pgd_t *pgd;
3446	pud_t *pud;
3447	pmd_t *pmd;
3448	pte_t *pte;
3449
3450	__set_current_state(TASK_RUNNING);
3451
3452	count_vm_event(PGFAULT);
3453	mem_cgroup_count_vm_event(mm, PGFAULT);
3454
3455	/* do counter updates before entering really critical section. */
3456	check_sync_rss_stat(current);
3457
3458	if (unlikely(is_vm_hugetlb_page(vma)))
3459		return hugetlb_fault(mm, vma, address, flags);
3460
3461	pgd = pgd_offset(mm, address);
3462	pud = pud_alloc(mm, pgd, address);
3463	if (!pud)
3464		return VM_FAULT_OOM;
3465	pmd = pmd_alloc(mm, pud, address);
3466	if (!pmd)
3467		return VM_FAULT_OOM;
3468	if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3469		if (!vma->vm_ops)
3470			return do_huge_pmd_anonymous_page(mm, vma, address,
3471							  pmd, flags);
3472	} else {
3473		pmd_t orig_pmd = *pmd;
 
 
3474		barrier();
3475		if (pmd_trans_huge(orig_pmd)) {
3476			if (flags & FAULT_FLAG_WRITE &&
3477			    !pmd_write(orig_pmd) &&
3478			    !pmd_trans_splitting(orig_pmd))
3479				return do_huge_pmd_wp_page(mm, vma, address,
3480							   pmd, orig_pmd);
3481			return 0;
 
 
 
 
 
 
 
 
 
 
3482		}
3483	}
3484
3485	/*
3486	 * Use __pte_alloc instead of pte_alloc_map, because we can't
3487	 * run pte_offset_map on the pmd, if an huge pmd could
3488	 * materialize from under us from a different thread.
3489	 */
3490	if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3491		return VM_FAULT_OOM;
3492	/* if an huge pmd materialized from under us just retry later */
3493	if (unlikely(pmd_trans_huge(*pmd)))
 
 
 
 
 
 
 
 
 
 
3494		return 0;
3495	/*
3496	 * A regular pmd is established and it can't morph into a huge pmd
3497	 * from under us anymore at this point because we hold the mmap_sem
3498	 * read mode and khugepaged takes it in write mode. So now it's
3499	 * safe to run pte_offset_map().
3500	 */
3501	pte = pte_offset_map(pmd, address);
3502
3503	return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3504}
3505
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3506#ifndef __PAGETABLE_PUD_FOLDED
3507/*
3508 * Allocate page upper directory.
3509 * We've already handled the fast-path in-line.
3510 */
3511int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3512{
3513	pud_t *new = pud_alloc_one(mm, address);
3514	if (!new)
3515		return -ENOMEM;
3516
3517	smp_wmb(); /* See comment in __pte_alloc */
3518
3519	spin_lock(&mm->page_table_lock);
3520	if (pgd_present(*pgd))		/* Another has populated it */
3521		pud_free(mm, new);
3522	else
3523		pgd_populate(mm, pgd, new);
3524	spin_unlock(&mm->page_table_lock);
3525	return 0;
3526}
3527#endif /* __PAGETABLE_PUD_FOLDED */
3528
3529#ifndef __PAGETABLE_PMD_FOLDED
3530/*
3531 * Allocate page middle directory.
3532 * We've already handled the fast-path in-line.
3533 */
3534int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3535{
3536	pmd_t *new = pmd_alloc_one(mm, address);
3537	if (!new)
3538		return -ENOMEM;
3539
3540	smp_wmb(); /* See comment in __pte_alloc */
3541
3542	spin_lock(&mm->page_table_lock);
3543#ifndef __ARCH_HAS_4LEVEL_HACK
3544	if (pud_present(*pud))		/* Another has populated it */
 
 
 
3545		pmd_free(mm, new);
3546	else
3547		pud_populate(mm, pud, new);
3548#else
3549	if (pgd_present(*pud))		/* Another has populated it */
 
 
 
3550		pmd_free(mm, new);
3551	else
3552		pgd_populate(mm, pud, new);
3553#endif /* __ARCH_HAS_4LEVEL_HACK */
3554	spin_unlock(&mm->page_table_lock);
3555	return 0;
3556}
3557#endif /* __PAGETABLE_PMD_FOLDED */
3558
3559int make_pages_present(unsigned long addr, unsigned long end)
3560{
3561	int ret, len, write;
3562	struct vm_area_struct * vma;
3563
3564	vma = find_vma(current->mm, addr);
3565	if (!vma)
3566		return -ENOMEM;
3567	/*
3568	 * We want to touch writable mappings with a write fault in order
3569	 * to break COW, except for shared mappings because these don't COW
3570	 * and we would not want to dirty them for nothing.
3571	 */
3572	write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3573	BUG_ON(addr >= end);
3574	BUG_ON(end > vma->vm_end);
3575	len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3576	ret = get_user_pages(current, current->mm, addr,
3577			len, write, 0, NULL, NULL);
3578	if (ret < 0)
3579		return ret;
3580	return ret == len ? 0 : -EFAULT;
3581}
3582
3583#if !defined(__HAVE_ARCH_GATE_AREA)
3584
3585#if defined(AT_SYSINFO_EHDR)
3586static struct vm_area_struct gate_vma;
3587
3588static int __init gate_vma_init(void)
3589{
3590	gate_vma.vm_mm = NULL;
3591	gate_vma.vm_start = FIXADDR_USER_START;
3592	gate_vma.vm_end = FIXADDR_USER_END;
3593	gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3594	gate_vma.vm_page_prot = __P101;
3595	/*
3596	 * Make sure the vDSO gets into every core dump.
3597	 * Dumping its contents makes post-mortem fully interpretable later
3598	 * without matching up the same kernel and hardware config to see
3599	 * what PC values meant.
3600	 */
3601	gate_vma.vm_flags |= VM_ALWAYSDUMP;
3602	return 0;
3603}
3604__initcall(gate_vma_init);
3605#endif
3606
3607struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3608{
3609#ifdef AT_SYSINFO_EHDR
3610	return &gate_vma;
3611#else
3612	return NULL;
3613#endif
3614}
3615
3616int in_gate_area_no_mm(unsigned long addr)
3617{
3618#ifdef AT_SYSINFO_EHDR
3619	if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3620		return 1;
3621#endif
3622	return 0;
3623}
3624
3625#endif	/* __HAVE_ARCH_GATE_AREA */
3626
3627static int __follow_pte(struct mm_struct *mm, unsigned long address,
3628		pte_t **ptepp, spinlock_t **ptlp)
3629{
3630	pgd_t *pgd;
3631	pud_t *pud;
3632	pmd_t *pmd;
3633	pte_t *ptep;
3634
3635	pgd = pgd_offset(mm, address);
3636	if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3637		goto out;
3638
3639	pud = pud_offset(pgd, address);
3640	if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3641		goto out;
3642
3643	pmd = pmd_offset(pud, address);
3644	VM_BUG_ON(pmd_trans_huge(*pmd));
3645	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3646		goto out;
3647
3648	/* We cannot handle huge page PFN maps. Luckily they don't exist. */
3649	if (pmd_huge(*pmd))
3650		goto out;
3651
3652	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3653	if (!ptep)
3654		goto out;
3655	if (!pte_present(*ptep))
3656		goto unlock;
3657	*ptepp = ptep;
3658	return 0;
3659unlock:
3660	pte_unmap_unlock(ptep, *ptlp);
3661out:
3662	return -EINVAL;
3663}
3664
3665static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3666			     pte_t **ptepp, spinlock_t **ptlp)
3667{
3668	int res;
3669
3670	/* (void) is needed to make gcc happy */
3671	(void) __cond_lock(*ptlp,
3672			   !(res = __follow_pte(mm, address, ptepp, ptlp)));
3673	return res;
3674}
3675
3676/**
3677 * follow_pfn - look up PFN at a user virtual address
3678 * @vma: memory mapping
3679 * @address: user virtual address
3680 * @pfn: location to store found PFN
3681 *
3682 * Only IO mappings and raw PFN mappings are allowed.
3683 *
3684 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3685 */
3686int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3687	unsigned long *pfn)
3688{
3689	int ret = -EINVAL;
3690	spinlock_t *ptl;
3691	pte_t *ptep;
3692
3693	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3694		return ret;
3695
3696	ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3697	if (ret)
3698		return ret;
3699	*pfn = pte_pfn(*ptep);
3700	pte_unmap_unlock(ptep, ptl);
3701	return 0;
3702}
3703EXPORT_SYMBOL(follow_pfn);
3704
3705#ifdef CONFIG_HAVE_IOREMAP_PROT
3706int follow_phys(struct vm_area_struct *vma,
3707		unsigned long address, unsigned int flags,
3708		unsigned long *prot, resource_size_t *phys)
3709{
3710	int ret = -EINVAL;
3711	pte_t *ptep, pte;
3712	spinlock_t *ptl;
3713
3714	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3715		goto out;
3716
3717	if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3718		goto out;
3719	pte = *ptep;
3720
3721	if ((flags & FOLL_WRITE) && !pte_write(pte))
3722		goto unlock;
3723
3724	*prot = pgprot_val(pte_pgprot(pte));
3725	*phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3726
3727	ret = 0;
3728unlock:
3729	pte_unmap_unlock(ptep, ptl);
3730out:
3731	return ret;
3732}
3733
3734int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3735			void *buf, int len, int write)
3736{
3737	resource_size_t phys_addr;
3738	unsigned long prot = 0;
3739	void __iomem *maddr;
3740	int offset = addr & (PAGE_SIZE-1);
3741
3742	if (follow_phys(vma, addr, write, &prot, &phys_addr))
3743		return -EINVAL;
3744
3745	maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3746	if (write)
3747		memcpy_toio(maddr + offset, buf, len);
3748	else
3749		memcpy_fromio(buf, maddr + offset, len);
3750	iounmap(maddr);
3751
3752	return len;
3753}
 
3754#endif
3755
3756/*
3757 * Access another process' address space as given in mm.  If non-NULL, use the
3758 * given task for page fault accounting.
3759 */
3760static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3761		unsigned long addr, void *buf, int len, int write)
3762{
3763	struct vm_area_struct *vma;
3764	void *old_buf = buf;
3765
3766	down_read(&mm->mmap_sem);
3767	/* ignore errors, just check how much was successfully transferred */
3768	while (len) {
3769		int bytes, ret, offset;
3770		void *maddr;
3771		struct page *page = NULL;
3772
3773		ret = get_user_pages(tsk, mm, addr, 1,
3774				write, 1, &page, &vma);
3775		if (ret <= 0) {
 
 
 
3776			/*
3777			 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3778			 * we can access using slightly different code.
3779			 */
3780#ifdef CONFIG_HAVE_IOREMAP_PROT
3781			vma = find_vma(mm, addr);
3782			if (!vma || vma->vm_start > addr)
3783				break;
3784			if (vma->vm_ops && vma->vm_ops->access)
3785				ret = vma->vm_ops->access(vma, addr, buf,
3786							  len, write);
3787			if (ret <= 0)
3788#endif
3789				break;
3790			bytes = ret;
 
3791		} else {
3792			bytes = len;
3793			offset = addr & (PAGE_SIZE-1);
3794			if (bytes > PAGE_SIZE-offset)
3795				bytes = PAGE_SIZE-offset;
3796
3797			maddr = kmap(page);
3798			if (write) {
3799				copy_to_user_page(vma, page, addr,
3800						  maddr + offset, buf, bytes);
3801				set_page_dirty_lock(page);
3802			} else {
3803				copy_from_user_page(vma, page, addr,
3804						    buf, maddr + offset, bytes);
3805			}
3806			kunmap(page);
3807			page_cache_release(page);
3808		}
3809		len -= bytes;
3810		buf += bytes;
3811		addr += bytes;
3812	}
3813	up_read(&mm->mmap_sem);
3814
3815	return buf - old_buf;
3816}
3817
3818/**
3819 * access_remote_vm - access another process' address space
3820 * @mm:		the mm_struct of the target address space
3821 * @addr:	start address to access
3822 * @buf:	source or destination buffer
3823 * @len:	number of bytes to transfer
3824 * @write:	whether the access is a write
3825 *
3826 * The caller must hold a reference on @mm.
3827 */
3828int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3829		void *buf, int len, int write)
3830{
3831	return __access_remote_vm(NULL, mm, addr, buf, len, write);
3832}
3833
3834/*
3835 * Access another process' address space.
3836 * Source/target buffer must be kernel space,
3837 * Do not walk the page table directly, use get_user_pages
3838 */
3839int access_process_vm(struct task_struct *tsk, unsigned long addr,
3840		void *buf, int len, int write)
3841{
3842	struct mm_struct *mm;
3843	int ret;
3844
3845	mm = get_task_mm(tsk);
3846	if (!mm)
3847		return 0;
3848
3849	ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3850	mmput(mm);
3851
3852	return ret;
3853}
3854
3855/*
3856 * Print the name of a VMA.
3857 */
3858void print_vma_addr(char *prefix, unsigned long ip)
3859{
3860	struct mm_struct *mm = current->mm;
3861	struct vm_area_struct *vma;
3862
3863	/*
3864	 * Do not print if we are in atomic
3865	 * contexts (in exception stacks, etc.):
3866	 */
3867	if (preempt_count())
3868		return;
3869
3870	down_read(&mm->mmap_sem);
3871	vma = find_vma(mm, ip);
3872	if (vma && vma->vm_file) {
3873		struct file *f = vma->vm_file;
3874		char *buf = (char *)__get_free_page(GFP_KERNEL);
3875		if (buf) {
3876			char *p, *s;
3877
3878			p = d_path(&f->f_path, buf, PAGE_SIZE);
3879			if (IS_ERR(p))
3880				p = "?";
3881			s = strrchr(p, '/');
3882			if (s)
3883				p = s+1;
3884			printk("%s%s[%lx+%lx]", prefix, p,
3885					vma->vm_start,
3886					vma->vm_end - vma->vm_start);
3887			free_page((unsigned long)buf);
3888		}
3889	}
3890	up_read(&current->mm->mmap_sem);
3891}
3892
3893#ifdef CONFIG_PROVE_LOCKING
3894void might_fault(void)
3895{
3896	/*
3897	 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3898	 * holding the mmap_sem, this is safe because kernel memory doesn't
3899	 * get paged out, therefore we'll never actually fault, and the
3900	 * below annotations will generate false positives.
3901	 */
3902	if (segment_eq(get_fs(), KERNEL_DS))
3903		return;
3904
3905	might_sleep();
3906	/*
3907	 * it would be nicer only to annotate paths which are not under
3908	 * pagefault_disable, however that requires a larger audit and
3909	 * providing helpers like get_user_atomic.
3910	 */
3911	if (!in_atomic() && current->mm)
3912		might_lock_read(&current->mm->mmap_sem);
 
3913}
3914EXPORT_SYMBOL(might_fault);
3915#endif
3916
3917#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3918static void clear_gigantic_page(struct page *page,
3919				unsigned long addr,
3920				unsigned int pages_per_huge_page)
3921{
3922	int i;
3923	struct page *p = page;
3924
3925	might_sleep();
3926	for (i = 0; i < pages_per_huge_page;
3927	     i++, p = mem_map_next(p, page, i)) {
3928		cond_resched();
3929		clear_user_highpage(p, addr + i * PAGE_SIZE);
3930	}
3931}
3932void clear_huge_page(struct page *page,
3933		     unsigned long addr, unsigned int pages_per_huge_page)
3934{
3935	int i;
3936
3937	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3938		clear_gigantic_page(page, addr, pages_per_huge_page);
3939		return;
3940	}
3941
3942	might_sleep();
3943	for (i = 0; i < pages_per_huge_page; i++) {
3944		cond_resched();
3945		clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3946	}
3947}
3948
3949static void copy_user_gigantic_page(struct page *dst, struct page *src,
3950				    unsigned long addr,
3951				    struct vm_area_struct *vma,
3952				    unsigned int pages_per_huge_page)
3953{
3954	int i;
3955	struct page *dst_base = dst;
3956	struct page *src_base = src;
3957
3958	for (i = 0; i < pages_per_huge_page; ) {
3959		cond_resched();
3960		copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3961
3962		i++;
3963		dst = mem_map_next(dst, dst_base, i);
3964		src = mem_map_next(src, src_base, i);
3965	}
3966}
3967
3968void copy_user_huge_page(struct page *dst, struct page *src,
3969			 unsigned long addr, struct vm_area_struct *vma,
3970			 unsigned int pages_per_huge_page)
3971{
3972	int i;
3973
3974	if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3975		copy_user_gigantic_page(dst, src, addr, vma,
3976					pages_per_huge_page);
3977		return;
3978	}
3979
3980	might_sleep();
3981	for (i = 0; i < pages_per_huge_page; i++) {
3982		cond_resched();
3983		copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3984	}
3985}
3986#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */