linux/mm/memory.c
<<
>>
Prefs
   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/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
 128void sync_mm_rss(struct mm_struct *mm)
 129{
 130        int i;
 131
 132        for (i = 0; i < NR_MM_COUNTERS; i++) {
 133                if (current->rss_stat.count[i]) {
 134                        add_mm_counter(mm, i, current->rss_stat.count[i]);
 135                        current->rss_stat.count[i] = 0;
 136                }
 137        }
 138        current->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_mm_rss(task->mm);
 161}
 162#else /* SPLIT_RSS_COUNTING */
 163
 164#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 165#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 166
 167static void check_sync_rss_stat(struct task_struct *task)
 168{
 169}
 170
 171#endif /* SPLIT_RSS_COUNTING */
 172
 173#ifdef HAVE_GENERIC_MMU_GATHER
 174
 175static int tlb_next_batch(struct mmu_gather *tlb)
 176{
 177        struct mmu_gather_batch *batch;
 178
 179        batch = tlb->active;
 180        if (batch->next) {
 181                tlb->active = batch->next;
 182                return 1;
 183        }
 184
 185        batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 186        if (!batch)
 187                return 0;
 188
 189        batch->next = NULL;
 190        batch->nr   = 0;
 191        batch->max  = MAX_GATHER_BATCH;
 192
 193        tlb->active->next = batch;
 194        tlb->active = batch;
 195
 196        return 1;
 197}
 198
 199/* tlb_gather_mmu
 200 *      Called to initialize an (on-stack) mmu_gather structure for page-table
 201 *      tear-down from @mm. The @fullmm argument is used when @mm is without
 202 *      users and we're going to destroy the full address space (exit/execve).
 203 */
 204void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 205{
 206        tlb->mm = mm;
 207
 208        tlb->fullmm     = fullmm;
 209        tlb->start      = -1UL;
 210        tlb->end        = 0;
 211        tlb->need_flush = 0;
 212        tlb->fast_mode  = (num_possible_cpus() == 1);
 213        tlb->local.next = NULL;
 214        tlb->local.nr   = 0;
 215        tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 216        tlb->active     = &tlb->local;
 217
 218#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 219        tlb->batch = NULL;
 220#endif
 221}
 222
 223void tlb_flush_mmu(struct mmu_gather *tlb)
 224{
 225        struct mmu_gather_batch *batch;
 226
 227        if (!tlb->need_flush)
 228                return;
 229        tlb->need_flush = 0;
 230        tlb_flush(tlb);
 231#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 232        tlb_table_flush(tlb);
 233#endif
 234
 235        if (tlb_fast_mode(tlb))
 236                return;
 237
 238        for (batch = &tlb->local; batch; batch = batch->next) {
 239                free_pages_and_swap_cache(batch->pages, batch->nr);
 240                batch->nr = 0;
 241        }
 242        tlb->active = &tlb->local;
 243}
 244
 245/* tlb_finish_mmu
 246 *      Called at the end of the shootdown operation to free up any resources
 247 *      that were required.
 248 */
 249void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 250{
 251        struct mmu_gather_batch *batch, *next;
 252
 253        tlb->start = start;
 254        tlb->end   = end;
 255        tlb_flush_mmu(tlb);
 256
 257        /* keep the page table cache within bounds */
 258        check_pgt_cache();
 259
 260        for (batch = tlb->local.next; batch; batch = next) {
 261                next = batch->next;
 262                free_pages((unsigned long)batch, 0);
 263        }
 264        tlb->local.next = NULL;
 265}
 266
 267/* __tlb_remove_page
 268 *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 269 *      handling the additional races in SMP caused by other CPUs caching valid
 270 *      mappings in their TLBs. Returns the number of free page slots left.
 271 *      When out of page slots we must call tlb_flush_mmu().
 272 */
 273int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 274{
 275        struct mmu_gather_batch *batch;
 276
 277        VM_BUG_ON(!tlb->need_flush);
 278
 279        if (tlb_fast_mode(tlb)) {
 280                free_page_and_swap_cache(page);
 281                return 1; /* avoid calling tlb_flush_mmu() */
 282        }
 283
 284        batch = tlb->active;
 285        batch->pages[batch->nr++] = page;
 286        if (batch->nr == batch->max) {
 287                if (!tlb_next_batch(tlb))
 288                        return 0;
 289                batch = tlb->active;
 290        }
 291        VM_BUG_ON(batch->nr > batch->max);
 292
 293        return batch->max - batch->nr;
 294}
 295
 296#endif /* HAVE_GENERIC_MMU_GATHER */
 297
 298#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 299
 300/*
 301 * See the comment near struct mmu_table_batch.
 302 */
 303
 304static void tlb_remove_table_smp_sync(void *arg)
 305{
 306        /* Simply deliver the interrupt */
 307}
 308
 309static void tlb_remove_table_one(void *table)
 310{
 311        /*
 312         * This isn't an RCU grace period and hence the page-tables cannot be
 313         * assumed to be actually RCU-freed.
 314         *
 315         * It is however sufficient for software page-table walkers that rely on
 316         * IRQ disabling. See the comment near struct mmu_table_batch.
 317         */
 318        smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 319        __tlb_remove_table(table);
 320}
 321
 322static void tlb_remove_table_rcu(struct rcu_head *head)
 323{
 324        struct mmu_table_batch *batch;
 325        int i;
 326
 327        batch = container_of(head, struct mmu_table_batch, rcu);
 328
 329        for (i = 0; i < batch->nr; i++)
 330                __tlb_remove_table(batch->tables[i]);
 331
 332        free_page((unsigned long)batch);
 333}
 334
 335void tlb_table_flush(struct mmu_gather *tlb)
 336{
 337        struct mmu_table_batch **batch = &tlb->batch;
 338
 339        if (*batch) {
 340                call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 341                *batch = NULL;
 342        }
 343}
 344
 345void tlb_remove_table(struct mmu_gather *tlb, void *table)
 346{
 347        struct mmu_table_batch **batch = &tlb->batch;
 348
 349        tlb->need_flush = 1;
 350
 351        /*
 352         * When there's less then two users of this mm there cannot be a
 353         * concurrent page-table walk.
 354         */
 355        if (atomic_read(&tlb->mm->mm_users) < 2) {
 356                __tlb_remove_table(table);
 357                return;
 358        }
 359
 360        if (*batch == NULL) {
 361                *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 362                if (*batch == NULL) {
 363                        tlb_remove_table_one(table);
 364                        return;
 365                }
 366                (*batch)->nr = 0;
 367        }
 368        (*batch)->tables[(*batch)->nr++] = table;
 369        if ((*batch)->nr == MAX_TABLE_BATCH)
 370                tlb_table_flush(tlb);
 371}
 372
 373#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 374
 375/*
 376 * If a p?d_bad entry is found while walking page tables, report
 377 * the error, before resetting entry to p?d_none.  Usually (but
 378 * very seldom) called out from the p?d_none_or_clear_bad macros.
 379 */
 380
 381void pgd_clear_bad(pgd_t *pgd)
 382{
 383        pgd_ERROR(*pgd);
 384        pgd_clear(pgd);
 385}
 386
 387void pud_clear_bad(pud_t *pud)
 388{
 389        pud_ERROR(*pud);
 390        pud_clear(pud);
 391}
 392
 393void pmd_clear_bad(pmd_t *pmd)
 394{
 395        pmd_ERROR(*pmd);
 396        pmd_clear(pmd);
 397}
 398
 399/*
 400 * Note: this doesn't free the actual pages themselves. That
 401 * has been handled earlier when unmapping all the memory regions.
 402 */
 403static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 404                           unsigned long addr)
 405{
 406        pgtable_t token = pmd_pgtable(*pmd);
 407        pmd_clear(pmd);
 408        pte_free_tlb(tlb, token, addr);
 409        tlb->mm->nr_ptes--;
 410}
 411
 412static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 413                                unsigned long addr, unsigned long end,
 414                                unsigned long floor, unsigned long ceiling)
 415{
 416        pmd_t *pmd;
 417        unsigned long next;
 418        unsigned long start;
 419
 420        start = addr;
 421        pmd = pmd_offset(pud, addr);
 422        do {
 423                next = pmd_addr_end(addr, end);
 424                if (pmd_none_or_clear_bad(pmd))
 425                        continue;
 426                free_pte_range(tlb, pmd, addr);
 427        } while (pmd++, addr = next, addr != end);
 428
 429        start &= PUD_MASK;
 430        if (start < floor)
 431                return;
 432        if (ceiling) {
 433                ceiling &= PUD_MASK;
 434                if (!ceiling)
 435                        return;
 436        }
 437        if (end - 1 > ceiling - 1)
 438                return;
 439
 440        pmd = pmd_offset(pud, start);
 441        pud_clear(pud);
 442        pmd_free_tlb(tlb, pmd, start);
 443}
 444
 445static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 446                                unsigned long addr, unsigned long end,
 447                                unsigned long floor, unsigned long ceiling)
 448{
 449        pud_t *pud;
 450        unsigned long next;
 451        unsigned long start;
 452
 453        start = addr;
 454        pud = pud_offset(pgd, addr);
 455        do {
 456                next = pud_addr_end(addr, end);
 457                if (pud_none_or_clear_bad(pud))
 458                        continue;
 459                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 460        } while (pud++, addr = next, addr != end);
 461
 462        start &= PGDIR_MASK;
 463        if (start < floor)
 464                return;
 465        if (ceiling) {
 466                ceiling &= PGDIR_MASK;
 467                if (!ceiling)
 468                        return;
 469        }
 470        if (end - 1 > ceiling - 1)
 471                return;
 472
 473        pud = pud_offset(pgd, start);
 474        pgd_clear(pgd);
 475        pud_free_tlb(tlb, pud, start);
 476}
 477
 478/*
 479 * This function frees user-level page tables of a process.
 480 *
 481 * Must be called with pagetable lock held.
 482 */
 483void free_pgd_range(struct mmu_gather *tlb,
 484                        unsigned long addr, unsigned long end,
 485                        unsigned long floor, unsigned long ceiling)
 486{
 487        pgd_t *pgd;
 488        unsigned long next;
 489
 490        /*
 491         * The next few lines have given us lots of grief...
 492         *
 493         * Why are we testing PMD* at this top level?  Because often
 494         * there will be no work to do at all, and we'd prefer not to
 495         * go all the way down to the bottom just to discover that.
 496         *
 497         * Why all these "- 1"s?  Because 0 represents both the bottom
 498         * of the address space and the top of it (using -1 for the
 499         * top wouldn't help much: the masks would do the wrong thing).
 500         * The rule is that addr 0 and floor 0 refer to the bottom of
 501         * the address space, but end 0 and ceiling 0 refer to the top
 502         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 503         * that end 0 case should be mythical).
 504         *
 505         * Wherever addr is brought up or ceiling brought down, we must
 506         * be careful to reject "the opposite 0" before it confuses the
 507         * subsequent tests.  But what about where end is brought down
 508         * by PMD_SIZE below? no, end can't go down to 0 there.
 509         *
 510         * Whereas we round start (addr) and ceiling down, by different
 511         * masks at different levels, in order to test whether a table
 512         * now has no other vmas using it, so can be freed, we don't
 513         * bother to round floor or end up - the tests don't need that.
 514         */
 515
 516        addr &= PMD_MASK;
 517        if (addr < floor) {
 518                addr += PMD_SIZE;
 519                if (!addr)
 520                        return;
 521        }
 522        if (ceiling) {
 523                ceiling &= PMD_MASK;
 524                if (!ceiling)
 525                        return;
 526        }
 527        if (end - 1 > ceiling - 1)
 528                end -= PMD_SIZE;
 529        if (addr > end - 1)
 530                return;
 531
 532        pgd = pgd_offset(tlb->mm, addr);
 533        do {
 534                next = pgd_addr_end(addr, end);
 535                if (pgd_none_or_clear_bad(pgd))
 536                        continue;
 537                free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 538        } while (pgd++, addr = next, addr != end);
 539}
 540
 541void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 542                unsigned long floor, unsigned long ceiling)
 543{
 544        while (vma) {
 545                struct vm_area_struct *next = vma->vm_next;
 546                unsigned long addr = vma->vm_start;
 547
 548                /*
 549                 * Hide vma from rmap and truncate_pagecache before freeing
 550                 * pgtables
 551                 */
 552                unlink_anon_vmas(vma);
 553                unlink_file_vma(vma);
 554
 555                if (is_vm_hugetlb_page(vma)) {
 556                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 557                                floor, next? next->vm_start: ceiling);
 558                } else {
 559                        /*
 560                         * Optimization: gather nearby vmas into one call down
 561                         */
 562                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 563                               && !is_vm_hugetlb_page(next)) {
 564                                vma = next;
 565                                next = vma->vm_next;
 566                                unlink_anon_vmas(vma);
 567                                unlink_file_vma(vma);
 568                        }
 569                        free_pgd_range(tlb, addr, vma->vm_end,
 570                                floor, next? next->vm_start: ceiling);
 571                }
 572                vma = next;
 573        }
 574}
 575
 576int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 577                pmd_t *pmd, unsigned long address)
 578{
 579        pgtable_t new = pte_alloc_one(mm, address);
 580        int wait_split_huge_page;
 581        if (!new)
 582                return -ENOMEM;
 583
 584        /*
 585         * Ensure all pte setup (eg. pte page lock and page clearing) are
 586         * visible before the pte is made visible to other CPUs by being
 587         * put into page tables.
 588         *
 589         * The other side of the story is the pointer chasing in the page
 590         * table walking code (when walking the page table without locking;
 591         * ie. most of the time). Fortunately, these data accesses consist
 592         * of a chain of data-dependent loads, meaning most CPUs (alpha
 593         * being the notable exception) will already guarantee loads are
 594         * seen in-order. See the alpha page table accessors for the
 595         * smp_read_barrier_depends() barriers in page table walking code.
 596         */
 597        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 598
 599        spin_lock(&mm->page_table_lock);
 600        wait_split_huge_page = 0;
 601        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 602                mm->nr_ptes++;
 603                pmd_populate(mm, pmd, new);
 604                new = NULL;
 605        } else if (unlikely(pmd_trans_splitting(*pmd)))
 606                wait_split_huge_page = 1;
 607        spin_unlock(&mm->page_table_lock);
 608        if (new)
 609                pte_free(mm, new);
 610        if (wait_split_huge_page)
 611                wait_split_huge_page(vma->anon_vma, pmd);
 612        return 0;
 613}
 614
 615int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 616{
 617        pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 618        if (!new)
 619                return -ENOMEM;
 620
 621        smp_wmb(); /* See comment in __pte_alloc */
 622
 623        spin_lock(&init_mm.page_table_lock);
 624        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 625                pmd_populate_kernel(&init_mm, pmd, new);
 626                new = NULL;
 627        } else
 628                VM_BUG_ON(pmd_trans_splitting(*pmd));
 629        spin_unlock(&init_mm.page_table_lock);
 630        if (new)
 631                pte_free_kernel(&init_mm, new);
 632        return 0;
 633}
 634
 635static inline void init_rss_vec(int *rss)
 636{
 637        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 638}
 639
 640static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 641{
 642        int i;
 643
 644        if (current->mm == mm)
 645                sync_mm_rss(mm);
 646        for (i = 0; i < NR_MM_COUNTERS; i++)
 647                if (rss[i])
 648                        add_mm_counter(mm, i, rss[i]);
 649}
 650
 651/*
 652 * This function is called to print an error when a bad pte
 653 * is found. For example, we might have a PFN-mapped pte in
 654 * a region that doesn't allow it.
 655 *
 656 * The calling function must still handle the error.
 657 */
 658static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 659                          pte_t pte, struct page *page)
 660{
 661        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 662        pud_t *pud = pud_offset(pgd, addr);
 663        pmd_t *pmd = pmd_offset(pud, addr);
 664        struct address_space *mapping;
 665        pgoff_t index;
 666        static unsigned long resume;
 667        static unsigned long nr_shown;
 668        static unsigned long nr_unshown;
 669
 670        /*
 671         * Allow a burst of 60 reports, then keep quiet for that minute;
 672         * or allow a steady drip of one report per second.
 673         */
 674        if (nr_shown == 60) {
 675                if (time_before(jiffies, resume)) {
 676                        nr_unshown++;
 677                        return;
 678                }
 679                if (nr_unshown) {
 680                        printk(KERN_ALERT
 681                                "BUG: Bad page map: %lu messages suppressed\n",
 682                                nr_unshown);
 683                        nr_unshown = 0;
 684                }
 685                nr_shown = 0;
 686        }
 687        if (nr_shown++ == 0)
 688                resume = jiffies + 60 * HZ;
 689
 690        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 691        index = linear_page_index(vma, addr);
 692
 693        printk(KERN_ALERT
 694                "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 695                current->comm,
 696                (long long)pte_val(pte), (long long)pmd_val(*pmd));
 697        if (page)
 698                dump_page(page);
 699        printk(KERN_ALERT
 700                "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 701                (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 702        /*
 703         * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 704         */
 705        if (vma->vm_ops)
 706                print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
 707                                (unsigned long)vma->vm_ops->fault);
 708        if (vma->vm_file && vma->vm_file->f_op)
 709                print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
 710                                (unsigned long)vma->vm_file->f_op->mmap);
 711        dump_stack();
 712        add_taint(TAINT_BAD_PAGE);
 713}
 714
 715static inline bool is_cow_mapping(vm_flags_t flags)
 716{
 717        return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 718}
 719
 720#ifndef is_zero_pfn
 721static inline int is_zero_pfn(unsigned long pfn)
 722{
 723        return pfn == zero_pfn;
 724}
 725#endif
 726
 727#ifndef my_zero_pfn
 728static inline unsigned long my_zero_pfn(unsigned long addr)
 729{
 730        return zero_pfn;
 731}
 732#endif
 733
 734/*
 735 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 736 *
 737 * "Special" mappings do not wish to be associated with a "struct page" (either
 738 * it doesn't exist, or it exists but they don't want to touch it). In this
 739 * case, NULL is returned here. "Normal" mappings do have a struct page.
 740 *
 741 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 742 * pte bit, in which case this function is trivial. Secondly, an architecture
 743 * may not have a spare pte bit, which requires a more complicated scheme,
 744 * described below.
 745 *
 746 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 747 * special mapping (even if there are underlying and valid "struct pages").
 748 * COWed pages of a VM_PFNMAP are always normal.
 749 *
 750 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 751 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 752 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 753 * mapping will always honor the rule
 754 *
 755 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 756 *
 757 * And for normal mappings this is false.
 758 *
 759 * This restricts such mappings to be a linear translation from virtual address
 760 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 761 * as the vma is not a COW mapping; in that case, we know that all ptes are
 762 * special (because none can have been COWed).
 763 *
 764 *
 765 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 766 *
 767 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 768 * page" backing, however the difference is that _all_ pages with a struct
 769 * page (that is, those where pfn_valid is true) are refcounted and considered
 770 * normal pages by the VM. The disadvantage is that pages are refcounted
 771 * (which can be slower and simply not an option for some PFNMAP users). The
 772 * advantage is that we don't have to follow the strict linearity rule of
 773 * PFNMAP mappings in order to support COWable mappings.
 774 *
 775 */
 776#ifdef __HAVE_ARCH_PTE_SPECIAL
 777# define HAVE_PTE_SPECIAL 1
 778#else
 779# define HAVE_PTE_SPECIAL 0
 780#endif
 781struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 782                                pte_t pte)
 783{
 784        unsigned long pfn = pte_pfn(pte);
 785
 786        if (HAVE_PTE_SPECIAL) {
 787                if (likely(!pte_special(pte)))
 788                        goto check_pfn;
 789                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 790                        return NULL;
 791                if (!is_zero_pfn(pfn))
 792                        print_bad_pte(vma, addr, pte, NULL);
 793                return NULL;
 794        }
 795
 796        /* !HAVE_PTE_SPECIAL case follows: */
 797
 798        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 799                if (vma->vm_flags & VM_MIXEDMAP) {
 800                        if (!pfn_valid(pfn))
 801                                return NULL;
 802                        goto out;
 803                } else {
 804                        unsigned long off;
 805                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 806                        if (pfn == vma->vm_pgoff + off)
 807                                return NULL;
 808                        if (!is_cow_mapping(vma->vm_flags))
 809                                return NULL;
 810                }
 811        }
 812
 813        if (is_zero_pfn(pfn))
 814                return NULL;
 815check_pfn:
 816        if (unlikely(pfn > highest_memmap_pfn)) {
 817                print_bad_pte(vma, addr, pte, NULL);
 818                return NULL;
 819        }
 820
 821        /*
 822         * NOTE! We still have PageReserved() pages in the page tables.
 823         * eg. VDSO mappings can cause them to exist.
 824         */
 825out:
 826        return pfn_to_page(pfn);
 827}
 828
 829/*
 830 * copy one vm_area from one task to the other. Assumes the page tables
 831 * already present in the new task to be cleared in the whole range
 832 * covered by this vma.
 833 */
 834
 835static inline unsigned long
 836copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 837                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 838                unsigned long addr, int *rss)
 839{
 840        unsigned long vm_flags = vma->vm_flags;
 841        pte_t pte = *src_pte;
 842        struct page *page;
 843
 844        /* pte contains position in swap or file, so copy. */
 845        if (unlikely(!pte_present(pte))) {
 846                if (!pte_file(pte)) {
 847                        swp_entry_t entry = pte_to_swp_entry(pte);
 848
 849                        if (swap_duplicate(entry) < 0)
 850                                return entry.val;
 851
 852                        /* make sure dst_mm is on swapoff's mmlist. */
 853                        if (unlikely(list_empty(&dst_mm->mmlist))) {
 854                                spin_lock(&mmlist_lock);
 855                                if (list_empty(&dst_mm->mmlist))
 856                                        list_add(&dst_mm->mmlist,
 857                                                 &src_mm->mmlist);
 858                                spin_unlock(&mmlist_lock);
 859                        }
 860                        if (likely(!non_swap_entry(entry)))
 861                                rss[MM_SWAPENTS]++;
 862                        else if (is_migration_entry(entry)) {
 863                                page = migration_entry_to_page(entry);
 864
 865                                if (PageAnon(page))
 866                                        rss[MM_ANONPAGES]++;
 867                                else
 868                                        rss[MM_FILEPAGES]++;
 869
 870                                if (is_write_migration_entry(entry) &&
 871                                    is_cow_mapping(vm_flags)) {
 872                                        /*
 873                                         * COW mappings require pages in both
 874                                         * parent and child to be set to read.
 875                                         */
 876                                        make_migration_entry_read(&entry);
 877                                        pte = swp_entry_to_pte(entry);
 878                                        set_pte_at(src_mm, addr, src_pte, pte);
 879                                }
 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);
 906                if (PageAnon(page))
 907                        rss[MM_ANONPAGES]++;
 908                else
 909                        rss[MM_FILEPAGES]++;
 910        }
 911
 912out_set_pte:
 913        set_pte_at(dst_mm, addr, dst_pte, pte);
 914        return 0;
 915}
 916
 917int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 918                   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 919                   unsigned long addr, unsigned long end)
 920{
 921        pte_t *orig_src_pte, *orig_dst_pte;
 922        pte_t *src_pte, *dst_pte;
 923        spinlock_t *src_ptl, *dst_ptl;
 924        int progress = 0;
 925        int rss[NR_MM_COUNTERS];
 926        swp_entry_t entry = (swp_entry_t){0};
 927
 928again:
 929        init_rss_vec(rss);
 930
 931        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 932        if (!dst_pte)
 933                return -ENOMEM;
 934        src_pte = pte_offset_map(src_pmd, addr);
 935        src_ptl = pte_lockptr(src_mm, src_pmd);
 936        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 937        orig_src_pte = src_pte;
 938        orig_dst_pte = dst_pte;
 939        arch_enter_lazy_mmu_mode();
 940
 941        do {
 942                /*
 943                 * We are holding two locks at this point - either of them
 944                 * could generate latencies in another task on another CPU.
 945                 */
 946                if (progress >= 32) {
 947                        progress = 0;
 948                        if (need_resched() ||
 949                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 950                                break;
 951                }
 952                if (pte_none(*src_pte)) {
 953                        progress++;
 954                        continue;
 955                }
 956                entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 957                                                        vma, addr, rss);
 958                if (entry.val)
 959                        break;
 960                progress += 8;
 961        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 962
 963        arch_leave_lazy_mmu_mode();
 964        spin_unlock(src_ptl);
 965        pte_unmap(orig_src_pte);
 966        add_mm_rss_vec(dst_mm, rss);
 967        pte_unmap_unlock(orig_dst_pte, dst_ptl);
 968        cond_resched();
 969
 970        if (entry.val) {
 971                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 972                        return -ENOMEM;
 973                progress = 0;
 974        }
 975        if (addr != end)
 976                goto again;
 977        return 0;
 978}
 979
 980static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 981                pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 982                unsigned long addr, unsigned long end)
 983{
 984        pmd_t *src_pmd, *dst_pmd;
 985        unsigned long next;
 986
 987        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 988        if (!dst_pmd)
 989                return -ENOMEM;
 990        src_pmd = pmd_offset(src_pud, addr);
 991        do {
 992                next = pmd_addr_end(addr, end);
 993                if (pmd_trans_huge(*src_pmd)) {
 994                        int err;
 995                        VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
 996                        err = copy_huge_pmd(dst_mm, src_mm,
 997                                            dst_pmd, src_pmd, addr, vma);
 998                        if (err == -ENOMEM)
 999                                return -ENOMEM;
1000                        if (!err)
1001                                continue;
1002                        /* fall through */
1003                }
1004                if (pmd_none_or_clear_bad(src_pmd))
1005                        continue;
1006                if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1007                                                vma, addr, next))
1008                        return -ENOMEM;
1009        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1010        return 0;
1011}
1012
1013static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1014                pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1015                unsigned long addr, unsigned long end)
1016{
1017        pud_t *src_pud, *dst_pud;
1018        unsigned long next;
1019
1020        dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1021        if (!dst_pud)
1022                return -ENOMEM;
1023        src_pud = pud_offset(src_pgd, addr);
1024        do {
1025                next = pud_addr_end(addr, end);
1026                if (pud_none_or_clear_bad(src_pud))
1027                        continue;
1028                if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1029                                                vma, addr, next))
1030                        return -ENOMEM;
1031        } while (dst_pud++, src_pud++, addr = next, addr != end);
1032        return 0;
1033}
1034
1035int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1036                struct vm_area_struct *vma)
1037{
1038        pgd_t *src_pgd, *dst_pgd;
1039        unsigned long next;
1040        unsigned long addr = vma->vm_start;
1041        unsigned long end = vma->vm_end;
1042        unsigned long mmun_start;       /* For mmu_notifiers */
1043        unsigned long mmun_end;         /* For mmu_notifiers */
1044        bool is_cow;
1045        int ret;
1046
1047        /*
1048         * Don't copy ptes where a page fault will fill them correctly.
1049         * Fork becomes much lighter when there are big shared or private
1050         * readonly mappings. The tradeoff is that copy_page_range is more
1051         * efficient than faulting.
1052         */
1053        if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1054                               VM_PFNMAP | VM_MIXEDMAP))) {
1055                if (!vma->anon_vma)
1056                        return 0;
1057        }
1058
1059        if (is_vm_hugetlb_page(vma))
1060                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1061
1062        if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1063                /*
1064                 * We do not free on error cases below as remove_vma
1065                 * gets called on error from higher level routine
1066                 */
1067                ret = track_pfn_copy(vma);
1068                if (ret)
1069                        return ret;
1070        }
1071
1072        /*
1073         * We need to invalidate the secondary MMU mappings only when
1074         * there could be a permission downgrade on the ptes of the
1075         * parent mm. And a permission downgrade will only happen if
1076         * is_cow_mapping() returns true.
1077         */
1078        is_cow = is_cow_mapping(vma->vm_flags);
1079        mmun_start = addr;
1080        mmun_end   = end;
1081        if (is_cow)
1082                mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1083                                                    mmun_end);
1084
1085        ret = 0;
1086        dst_pgd = pgd_offset(dst_mm, addr);
1087        src_pgd = pgd_offset(src_mm, addr);
1088        do {
1089                next = pgd_addr_end(addr, end);
1090                if (pgd_none_or_clear_bad(src_pgd))
1091                        continue;
1092                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1093                                            vma, addr, next))) {
1094                        ret = -ENOMEM;
1095                        break;
1096                }
1097        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1098
1099        if (is_cow)
1100                mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1101        return ret;
1102}
1103
1104static unsigned long zap_pte_range(struct mmu_gather *tlb,
1105                                struct vm_area_struct *vma, pmd_t *pmd,
1106                                unsigned long addr, unsigned long end,
1107                                struct zap_details *details)
1108{
1109        struct mm_struct *mm = tlb->mm;
1110        int force_flush = 0;
1111        int rss[NR_MM_COUNTERS];
1112        spinlock_t *ptl;
1113        pte_t *start_pte;
1114        pte_t *pte;
1115
1116again:
1117        init_rss_vec(rss);
1118        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1119        pte = start_pte;
1120        arch_enter_lazy_mmu_mode();
1121        do {
1122                pte_t ptent = *pte;
1123                if (pte_none(ptent)) {
1124                        continue;
1125                }
1126
1127                if (pte_present(ptent)) {
1128                        struct page *page;
1129
1130                        page = vm_normal_page(vma, addr, ptent);
1131                        if (unlikely(details) && page) {
1132                                /*
1133                                 * unmap_shared_mapping_pages() wants to
1134                                 * invalidate cache without truncating:
1135                                 * unmap shared but keep private pages.
1136                                 */
1137                                if (details->check_mapping &&
1138                                    details->check_mapping != page->mapping)
1139                                        continue;
1140                                /*
1141                                 * Each page->index must be checked when
1142                                 * invalidating or truncating nonlinear.
1143                                 */
1144                                if (details->nonlinear_vma &&
1145                                    (page->index < details->first_index ||
1146                                     page->index > details->last_index))
1147                                        continue;
1148                        }
1149                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1150                                                        tlb->fullmm);
1151                        tlb_remove_tlb_entry(tlb, pte, addr);
1152                        if (unlikely(!page))
1153                                continue;
1154                        if (unlikely(details) && details->nonlinear_vma
1155                            && linear_page_index(details->nonlinear_vma,
1156                                                addr) != page->index)
1157                                set_pte_at(mm, addr, pte,
1158                                           pgoff_to_pte(page->index));
1159                        if (PageAnon(page))
1160                                rss[MM_ANONPAGES]--;
1161                        else {
1162                                if (pte_dirty(ptent))
1163                                        set_page_dirty(page);
1164                                if (pte_young(ptent) &&
1165                                    likely(!VM_SequentialReadHint(vma)))
1166                                        mark_page_accessed(page);
1167                                rss[MM_FILEPAGES]--;
1168                        }
1169                        page_remove_rmap(page);
1170                        if (unlikely(page_mapcount(page) < 0))
1171                                print_bad_pte(vma, addr, ptent, page);
1172                        force_flush = !__tlb_remove_page(tlb, page);
1173                        if (force_flush)
1174                                break;
1175                        continue;
1176                }
1177                /*
1178                 * If details->check_mapping, we leave swap entries;
1179                 * if details->nonlinear_vma, we leave file entries.
1180                 */
1181                if (unlikely(details))
1182                        continue;
1183                if (pte_file(ptent)) {
1184                        if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1185                                print_bad_pte(vma, addr, ptent, NULL);
1186                } else {
1187                        swp_entry_t entry = pte_to_swp_entry(ptent);
1188
1189                        if (!non_swap_entry(entry))
1190                                rss[MM_SWAPENTS]--;
1191                        else if (is_migration_entry(entry)) {
1192                                struct page *page;
1193
1194                                page = migration_entry_to_page(entry);
1195
1196                                if (PageAnon(page))
1197                                        rss[MM_ANONPAGES]--;
1198                                else
1199                                        rss[MM_FILEPAGES]--;
1200                        }
1201                        if (unlikely(!free_swap_and_cache(entry)))
1202                                print_bad_pte(vma, addr, ptent, NULL);
1203                }
1204                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1205        } while (pte++, addr += PAGE_SIZE, addr != end);
1206
1207        add_mm_rss_vec(mm, rss);
1208        arch_leave_lazy_mmu_mode();
1209        pte_unmap_unlock(start_pte, ptl);
1210
1211        /*
1212         * mmu_gather ran out of room to batch pages, we break out of
1213         * the PTE lock to avoid doing the potential expensive TLB invalidate
1214         * and page-free while holding it.
1215         */
1216        if (force_flush) {
1217                force_flush = 0;
1218
1219#ifdef HAVE_GENERIC_MMU_GATHER
1220                tlb->start = addr;
1221                tlb->end = end;
1222#endif
1223                tlb_flush_mmu(tlb);
1224                if (addr != end)
1225                        goto again;
1226        }
1227
1228        return addr;
1229}
1230
1231static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1232                                struct vm_area_struct *vma, pud_t *pud,
1233                                unsigned long addr, unsigned long end,
1234                                struct zap_details *details)
1235{
1236        pmd_t *pmd;
1237        unsigned long next;
1238
1239        pmd = pmd_offset(pud, addr);
1240        do {
1241                next = pmd_addr_end(addr, end);
1242                if (pmd_trans_huge(*pmd)) {
1243                        if (next - addr != HPAGE_PMD_SIZE) {
1244#ifdef CONFIG_DEBUG_VM
1245                                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1246                                        pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1247                                                __func__, addr, end,
1248                                                vma->vm_start,
1249                                                vma->vm_end);
1250                                        BUG();
1251                                }
1252#endif
1253                                split_huge_page_pmd(vma->vm_mm, pmd);
1254                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1255                                goto next;
1256                        /* fall through */
1257                }
1258                /*
1259                 * Here there can be other concurrent MADV_DONTNEED or
1260                 * trans huge page faults running, and if the pmd is
1261                 * none or trans huge it can change under us. This is
1262                 * because MADV_DONTNEED holds the mmap_sem in read
1263                 * mode.
1264                 */
1265                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1266                        goto next;
1267                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1268next:
1269                cond_resched();
1270        } while (pmd++, addr = next, addr != end);
1271
1272        return addr;
1273}
1274
1275static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1276                                struct vm_area_struct *vma, pgd_t *pgd,
1277                                unsigned long addr, unsigned long end,
1278                                struct zap_details *details)
1279{
1280        pud_t *pud;
1281        unsigned long next;
1282
1283        pud = pud_offset(pgd, addr);
1284        do {
1285                next = pud_addr_end(addr, end);
1286                if (pud_none_or_clear_bad(pud))
1287                        continue;
1288                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1289        } while (pud++, addr = next, addr != end);
1290
1291        return addr;
1292}
1293
1294static void unmap_page_range(struct mmu_gather *tlb,
1295                             struct vm_area_struct *vma,
1296                             unsigned long addr, unsigned long end,
1297                             struct zap_details *details)
1298{
1299        pgd_t *pgd;
1300        unsigned long next;
1301
1302        if (details && !details->check_mapping && !details->nonlinear_vma)
1303                details = NULL;
1304
1305        BUG_ON(addr >= end);
1306        mem_cgroup_uncharge_start();
1307        tlb_start_vma(tlb, vma);
1308        pgd = pgd_offset(vma->vm_mm, addr);
1309        do {
1310                next = pgd_addr_end(addr, end);
1311                if (pgd_none_or_clear_bad(pgd))
1312                        continue;
1313                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1314        } while (pgd++, addr = next, addr != end);
1315        tlb_end_vma(tlb, vma);
1316        mem_cgroup_uncharge_end();
1317}
1318
1319
1320static void unmap_single_vma(struct mmu_gather *tlb,
1321                struct vm_area_struct *vma, unsigned long start_addr,
1322                unsigned long end_addr,
1323                struct zap_details *details)
1324{
1325        unsigned long start = max(vma->vm_start, start_addr);
1326        unsigned long end;
1327
1328        if (start >= vma->vm_end)
1329                return;
1330        end = min(vma->vm_end, end_addr);
1331        if (end <= vma->vm_start)
1332                return;
1333
1334        if (vma->vm_file)
1335                uprobe_munmap(vma, start, end);
1336
1337        if (unlikely(vma->vm_flags & VM_PFNMAP))
1338                untrack_pfn(vma, 0, 0);
1339
1340        if (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                                mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1355                                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1356                                mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1357                        }
1358                } else
1359                        unmap_page_range(tlb, vma, start, end, details);
1360        }
1361}
1362
1363/**
1364 * unmap_vmas - unmap a range of memory covered by a list of vma's
1365 * @tlb: address of the caller's struct mmu_gather
1366 * @vma: the starting vma
1367 * @start_addr: virtual address at which to start unmapping
1368 * @end_addr: virtual address at which to end unmapping
1369 *
1370 * Unmap all pages in the vma list.
1371 *
1372 * Only addresses between `start' and `end' will be unmapped.
1373 *
1374 * The VMA list must be sorted in ascending virtual address order.
1375 *
1376 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1377 * range after unmap_vmas() returns.  So the only responsibility here is to
1378 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1379 * drops the lock and schedules.
1380 */
1381void unmap_vmas(struct mmu_gather *tlb,
1382                struct vm_area_struct *vma, unsigned long start_addr,
1383                unsigned long end_addr)
1384{
1385        struct mm_struct *mm = vma->vm_mm;
1386
1387        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1388        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1389                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1390        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1391}
1392
1393/**
1394 * zap_page_range - remove user pages in a given range
1395 * @vma: vm_area_struct holding the applicable pages
1396 * @start: starting address of pages to zap
1397 * @size: number of bytes to zap
1398 * @details: details of nonlinear truncation or shared cache invalidation
1399 *
1400 * Caller must protect the VMA list
1401 */
1402void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1403                unsigned long size, struct zap_details *details)
1404{
1405        struct mm_struct *mm = vma->vm_mm;
1406        struct mmu_gather tlb;
1407        unsigned long end = start + size;
1408
1409        lru_add_drain();
1410        tlb_gather_mmu(&tlb, mm, 0);
1411        update_hiwater_rss(mm);
1412        mmu_notifier_invalidate_range_start(mm, start, end);
1413        for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1414                unmap_single_vma(&tlb, vma, start, end, details);
1415        mmu_notifier_invalidate_range_end(mm, start, end);
1416        tlb_finish_mmu(&tlb, start, end);
1417}
1418
1419/**
1420 * zap_page_range_single - remove user pages in a given range
1421 * @vma: vm_area_struct holding the applicable pages
1422 * @address: starting address of pages to zap
1423 * @size: number of bytes to zap
1424 * @details: details of nonlinear truncation or shared cache invalidation
1425 *
1426 * The range must fit into one VMA.
1427 */
1428static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1429                unsigned long size, struct zap_details *details)
1430{
1431        struct mm_struct *mm = vma->vm_mm;
1432        struct mmu_gather tlb;
1433        unsigned long end = address + size;
1434
1435        lru_add_drain();
1436        tlb_gather_mmu(&tlb, mm, 0);
1437        update_hiwater_rss(mm);
1438        mmu_notifier_invalidate_range_start(mm, address, end);
1439        unmap_single_vma(&tlb, vma, address, end, details);
1440        mmu_notifier_invalidate_range_end(mm, address, end);
1441        tlb_finish_mmu(&tlb, address, end);
1442}
1443
1444/**
1445 * zap_vma_ptes - remove ptes mapping the vma
1446 * @vma: vm_area_struct holding ptes to be zapped
1447 * @address: starting address of pages to zap
1448 * @size: number of bytes to zap
1449 *
1450 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1451 *
1452 * The entire address range must be fully contained within the vma.
1453 *
1454 * Returns 0 if successful.
1455 */
1456int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1457                unsigned long size)
1458{
1459        if (address < vma->vm_start || address + size > vma->vm_end ||
1460                        !(vma->vm_flags & VM_PFNMAP))
1461                return -1;
1462        zap_page_range_single(vma, address, size, NULL);
1463        return 0;
1464}
1465EXPORT_SYMBOL_GPL(zap_vma_ptes);
1466
1467/**
1468 * follow_page - look up a page descriptor from a user-virtual address
1469 * @vma: vm_area_struct mapping @address
1470 * @address: virtual address to look up
1471 * @flags: flags modifying lookup behaviour
1472 *
1473 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1474 *
1475 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1476 * an error pointer if there is a mapping to something not represented
1477 * by a page descriptor (see also vm_normal_page()).
1478 */
1479struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1480                        unsigned int flags)
1481{
1482        pgd_t *pgd;
1483        pud_t *pud;
1484        pmd_t *pmd;
1485        pte_t *ptep, pte;
1486        spinlock_t *ptl;
1487        struct page *page;
1488        struct mm_struct *mm = vma->vm_mm;
1489
1490        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1491        if (!IS_ERR(page)) {
1492                BUG_ON(flags & FOLL_GET);
1493                goto out;
1494        }
1495
1496        page = NULL;
1497        pgd = pgd_offset(mm, address);
1498        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1499                goto no_page_table;
1500
1501        pud = pud_offset(pgd, address);
1502        if (pud_none(*pud))
1503                goto no_page_table;
1504        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1505                BUG_ON(flags & FOLL_GET);
1506                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1507                goto out;
1508        }
1509        if (unlikely(pud_bad(*pud)))
1510                goto no_page_table;
1511
1512        pmd = pmd_offset(pud, address);
1513        if (pmd_none(*pmd))
1514                goto no_page_table;
1515        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1516                BUG_ON(flags & FOLL_GET);
1517                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1518                goto out;
1519        }
1520        if (pmd_trans_huge(*pmd)) {
1521                if (flags & FOLL_SPLIT) {
1522                        split_huge_page_pmd(mm, pmd);
1523                        goto split_fallthrough;
1524                }
1525                spin_lock(&mm->page_table_lock);
1526                if (likely(pmd_trans_huge(*pmd))) {
1527                        if (unlikely(pmd_trans_splitting(*pmd))) {
1528                                spin_unlock(&mm->page_table_lock);
1529                                wait_split_huge_page(vma->anon_vma, pmd);
1530                        } else {
1531                                page = follow_trans_huge_pmd(vma, address,
1532                                                             pmd, flags);
1533                                spin_unlock(&mm->page_table_lock);
1534                                goto out;
1535                        }
1536                } else
1537                        spin_unlock(&mm->page_table_lock);
1538                /* fall through */
1539        }
1540split_fallthrough:
1541        if (unlikely(pmd_bad(*pmd)))
1542                goto no_page_table;
1543
1544        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1545
1546        pte = *ptep;
1547        if (!pte_present(pte))
1548                goto no_page;
1549        if ((flags & FOLL_WRITE) && !pte_write(pte))
1550                goto unlock;
1551
1552        page = vm_normal_page(vma, address, pte);
1553        if (unlikely(!page)) {
1554                if ((flags & FOLL_DUMP) ||
1555                    !is_zero_pfn(pte_pfn(pte)))
1556                        goto bad_page;
1557                page = pte_page(pte);
1558        }
1559
1560        if (flags & FOLL_GET)
1561                get_page_foll(page);
1562        if (flags & FOLL_TOUCH) {
1563                if ((flags & FOLL_WRITE) &&
1564                    !pte_dirty(pte) && !PageDirty(page))
1565                        set_page_dirty(page);
1566                /*
1567                 * pte_mkyoung() would be more correct here, but atomic care
1568                 * is needed to avoid losing the dirty bit: it is easier to use
1569                 * mark_page_accessed().
1570                 */
1571                mark_page_accessed(page);
1572        }
1573        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1574                /*
1575                 * The preliminary mapping check is mainly to avoid the
1576                 * pointless overhead of lock_page on the ZERO_PAGE
1577                 * which might bounce very badly if there is contention.
1578                 *
1579                 * If the page is already locked, we don't need to
1580                 * handle it now - vmscan will handle it later if and
1581                 * when it attempts to reclaim the page.
1582                 */
1583                if (page->mapping && trylock_page(page)) {
1584                        lru_add_drain();  /* push cached pages to LRU */
1585                        /*
1586                         * Because we lock page here, and migration is
1587                         * blocked by the pte's page reference, and we
1588                         * know the page is still mapped, we don't even
1589                         * need to check for file-cache page truncation.
1590                         */
1591                        mlock_vma_page(page);
1592                        unlock_page(page);
1593                }
1594        }
1595unlock:
1596        pte_unmap_unlock(ptep, ptl);
1597out:
1598        return page;
1599
1600bad_page:
1601        pte_unmap_unlock(ptep, ptl);
1602        return ERR_PTR(-EFAULT);
1603
1604no_page:
1605        pte_unmap_unlock(ptep, ptl);
1606        if (!pte_none(pte))
1607                return page;
1608
1609no_page_table:
1610        /*
1611         * When core dumping an enormous anonymous area that nobody
1612         * has touched so far, we don't want to allocate unnecessary pages or
1613         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1614         * then get_dump_page() will return NULL to leave a hole in the dump.
1615         * But we can only make this optimization where a hole would surely
1616         * be zero-filled if handle_mm_fault() actually did handle it.
1617         */
1618        if ((flags & FOLL_DUMP) &&
1619            (!vma->vm_ops || !vma->vm_ops->fault))
1620                return ERR_PTR(-EFAULT);
1621        return page;
1622}
1623
1624static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1625{
1626        return stack_guard_page_start(vma, addr) ||
1627               stack_guard_page_end(vma, addr+PAGE_SIZE);
1628}
1629
1630/**
1631 * __get_user_pages() - pin user pages in memory
1632 * @tsk:        task_struct of target task
1633 * @mm:         mm_struct of target mm
1634 * @start:      starting user address
1635 * @nr_pages:   number of pages from start to pin
1636 * @gup_flags:  flags modifying pin behaviour
1637 * @pages:      array that receives pointers to the pages pinned.
1638 *              Should be at least nr_pages long. Or NULL, if caller
1639 *              only intends to ensure the pages are faulted in.
1640 * @vmas:       array of pointers to vmas corresponding to each page.
1641 *              Or NULL if the caller does not require them.
1642 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1643 *
1644 * Returns number of pages pinned. This may be fewer than the number
1645 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1646 * were pinned, returns -errno. Each page returned must be released
1647 * with a put_page() call when it is finished with. vmas will only
1648 * remain valid while mmap_sem is held.
1649 *
1650 * Must be called with mmap_sem held for read or write.
1651 *
1652 * __get_user_pages walks a process's page tables and takes a reference to
1653 * each struct page that each user address corresponds to at a given
1654 * instant. That is, it takes the page that would be accessed if a user
1655 * thread accesses the given user virtual address at that instant.
1656 *
1657 * This does not guarantee that the page exists in the user mappings when
1658 * __get_user_pages returns, and there may even be a completely different
1659 * page there in some cases (eg. if mmapped pagecache has been invalidated
1660 * and subsequently re faulted). However it does guarantee that the page
1661 * won't be freed completely. And mostly callers simply care that the page
1662 * contains data that was valid *at some point in time*. Typically, an IO
1663 * or similar operation cannot guarantee anything stronger anyway because
1664 * locks can't be held over the syscall boundary.
1665 *
1666 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1667 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1668 * appropriate) must be called after the page is finished with, and
1669 * before put_page is called.
1670 *
1671 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1672 * or mmap_sem contention, and if waiting is needed to pin all pages,
1673 * *@nonblocking will be set to 0.
1674 *
1675 * In most cases, get_user_pages or get_user_pages_fast should be used
1676 * instead of __get_user_pages. __get_user_pages should be used only if
1677 * you need some special @gup_flags.
1678 */
1679int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1680                     unsigned long start, int nr_pages, unsigned int gup_flags,
1681                     struct page **pages, struct vm_area_struct **vmas,
1682                     int *nonblocking)
1683{
1684        int i;
1685        unsigned long vm_flags;
1686
1687        if (nr_pages <= 0)
1688                return 0;
1689
1690        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1691
1692        /* 
1693         * Require read or write permissions.
1694         * If FOLL_FORCE is set, we only require the "MAY" flags.
1695         */
1696        vm_flags  = (gup_flags & FOLL_WRITE) ?
1697                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1698        vm_flags &= (gup_flags & FOLL_FORCE) ?
1699                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1700        i = 0;
1701
1702        do {
1703                struct vm_area_struct *vma;
1704
1705                vma = find_extend_vma(mm, start);
1706                if (!vma && in_gate_area(mm, start)) {
1707                        unsigned long pg = start & PAGE_MASK;
1708                        pgd_t *pgd;
1709                        pud_t *pud;
1710                        pmd_t *pmd;
1711                        pte_t *pte;
1712
1713                        /* user gate pages are read-only */
1714                        if (gup_flags & FOLL_WRITE)
1715                                return i ? : -EFAULT;
1716                        if (pg > TASK_SIZE)
1717                                pgd = pgd_offset_k(pg);
1718                        else
1719                                pgd = pgd_offset_gate(mm, pg);
1720                        BUG_ON(pgd_none(*pgd));
1721                        pud = pud_offset(pgd, pg);
1722                        BUG_ON(pud_none(*pud));
1723                        pmd = pmd_offset(pud, pg);
1724                        if (pmd_none(*pmd))
1725                                return i ? : -EFAULT;
1726                        VM_BUG_ON(pmd_trans_huge(*pmd));
1727                        pte = pte_offset_map(pmd, pg);
1728                        if (pte_none(*pte)) {
1729                                pte_unmap(pte);
1730                                return i ? : -EFAULT;
1731                        }
1732                        vma = get_gate_vma(mm);
1733                        if (pages) {
1734                                struct page *page;
1735
1736                                page = vm_normal_page(vma, start, *pte);
1737                                if (!page) {
1738                                        if (!(gup_flags & FOLL_DUMP) &&
1739                                             is_zero_pfn(pte_pfn(*pte)))
1740                                                page = pte_page(*pte);
1741                                        else {
1742                                                pte_unmap(pte);
1743                                                return i ? : -EFAULT;
1744                                        }
1745                                }
1746                                pages[i] = page;
1747                                get_page(page);
1748                        }
1749                        pte_unmap(pte);
1750                        goto next_page;
1751                }
1752
1753                if (!vma ||
1754                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1755                    !(vm_flags & vma->vm_flags))
1756                        return i ? : -EFAULT;
1757
1758                if (is_vm_hugetlb_page(vma)) {
1759                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1760                                        &start, &nr_pages, i, gup_flags);
1761                        continue;
1762                }
1763
1764                do {
1765                        struct page *page;
1766                        unsigned int foll_flags = gup_flags;
1767
1768                        /*
1769                         * If we have a pending SIGKILL, don't keep faulting
1770                         * pages and potentially allocating memory.
1771                         */
1772                        if (unlikely(fatal_signal_pending(current)))
1773                                return i ? i : -ERESTARTSYS;
1774
1775                        cond_resched();
1776                        while (!(page = follow_page(vma, start, foll_flags))) {
1777                                int ret;
1778                                unsigned int fault_flags = 0;
1779
1780                                /* For mlock, just skip the stack guard page. */
1781                                if (foll_flags & FOLL_MLOCK) {
1782                                        if (stack_guard_page(vma, start))
1783                                                goto next_page;
1784                                }
1785                                if (foll_flags & FOLL_WRITE)
1786                                        fault_flags |= FAULT_FLAG_WRITE;
1787                                if (nonblocking)
1788                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1789                                if (foll_flags & FOLL_NOWAIT)
1790                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1791
1792                                ret = handle_mm_fault(mm, vma, start,
1793                                                        fault_flags);
1794
1795                                if (ret & VM_FAULT_ERROR) {
1796                                        if (ret & VM_FAULT_OOM)
1797                                                return i ? i : -ENOMEM;
1798                                        if (ret & (VM_FAULT_HWPOISON |
1799                                                   VM_FAULT_HWPOISON_LARGE)) {
1800                                                if (i)
1801                                                        return i;
1802                                                else if (gup_flags & FOLL_HWPOISON)
1803                                                        return -EHWPOISON;
1804                                                else
1805                                                        return -EFAULT;
1806                                        }
1807                                        if (ret & VM_FAULT_SIGBUS)
1808                                                return i ? i : -EFAULT;
1809                                        BUG();
1810                                }
1811
1812                                if (tsk) {
1813                                        if (ret & VM_FAULT_MAJOR)
1814                                                tsk->maj_flt++;
1815                                        else
1816                                                tsk->min_flt++;
1817                                }
1818
1819                                if (ret & VM_FAULT_RETRY) {
1820                                        if (nonblocking)
1821                                                *nonblocking = 0;
1822                                        return i;
1823                                }
1824
1825                                /*
1826                                 * The VM_FAULT_WRITE bit tells us that
1827                                 * do_wp_page has broken COW when necessary,
1828                                 * even if maybe_mkwrite decided not to set
1829                                 * pte_write. We can thus safely do subsequent
1830                                 * page lookups as if they were reads. But only
1831                                 * do so when looping for pte_write is futile:
1832                                 * in some cases userspace may also be wanting
1833                                 * to write to the gotten user page, which a
1834                                 * read fault here might prevent (a readonly
1835                                 * page might get reCOWed by userspace write).
1836                                 */
1837                                if ((ret & VM_FAULT_WRITE) &&
1838                                    !(vma->vm_flags & VM_WRITE))
1839                                        foll_flags &= ~FOLL_WRITE;
1840
1841                                cond_resched();
1842                        }
1843                        if (IS_ERR(page))
1844                                return i ? i : PTR_ERR(page);
1845                        if (pages) {
1846                                pages[i] = page;
1847
1848                                flush_anon_page(vma, page, start);
1849                                flush_dcache_page(page);
1850                        }
1851next_page:
1852                        if (vmas)
1853                                vmas[i] = vma;
1854                        i++;
1855                        start += PAGE_SIZE;
1856                        nr_pages--;
1857                } while (nr_pages && start < vma->vm_end);
1858        } while (nr_pages);
1859        return i;
1860}
1861EXPORT_SYMBOL(__get_user_pages);
1862
1863/*
1864 * fixup_user_fault() - manually resolve a user page fault
1865 * @tsk:        the task_struct to use for page fault accounting, or
1866 *              NULL if faults are not to be recorded.
1867 * @mm:         mm_struct of target mm
1868 * @address:    user address
1869 * @fault_flags:flags to pass down to handle_mm_fault()
1870 *
1871 * This is meant to be called in the specific scenario where for locking reasons
1872 * we try to access user memory in atomic context (within a pagefault_disable()
1873 * section), this returns -EFAULT, and we want to resolve the user fault before
1874 * trying again.
1875 *
1876 * Typically this is meant to be used by the futex code.
1877 *
1878 * The main difference with get_user_pages() is that this function will
1879 * unconditionally call handle_mm_fault() which will in turn perform all the
1880 * necessary SW fixup of the dirty and young bits in the PTE, while
1881 * handle_mm_fault() only guarantees to update these in the struct page.
1882 *
1883 * This is important for some architectures where those bits also gate the
1884 * access permission to the page because they are maintained in software.  On
1885 * such architectures, gup() will not be enough to make a subsequent access
1886 * succeed.
1887 *
1888 * This should be called with the mm_sem held for read.
1889 */
1890int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1891                     unsigned long address, unsigned int fault_flags)
1892{
1893        struct vm_area_struct *vma;
1894        int ret;
1895
1896        vma = find_extend_vma(mm, address);
1897        if (!vma || address < vma->vm_start)
1898                return -EFAULT;
1899
1900        ret = handle_mm_fault(mm, vma, address, fault_flags);
1901        if (ret & VM_FAULT_ERROR) {
1902                if (ret & VM_FAULT_OOM)
1903                        return -ENOMEM;
1904                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1905                        return -EHWPOISON;
1906                if (ret & VM_FAULT_SIGBUS)
1907                        return -EFAULT;
1908                BUG();
1909        }
1910        if (tsk) {
1911                if (ret & VM_FAULT_MAJOR)
1912                        tsk->maj_flt++;
1913                else
1914                        tsk->min_flt++;
1915        }
1916        return 0;
1917}
1918
1919/*
1920 * get_user_pages() - pin user pages in memory
1921 * @tsk:        the task_struct to use for page fault accounting, or
1922 *              NULL if faults are not to be recorded.
1923 * @mm:         mm_struct of target mm
1924 * @start:      starting user address
1925 * @nr_pages:   number of pages from start to pin
1926 * @write:      whether pages will be written to by the caller
1927 * @force:      whether to force write access even if user mapping is
1928 *              readonly. This will result in the page being COWed even
1929 *              in MAP_SHARED mappings. You do not want this.
1930 * @pages:      array that receives pointers to the pages pinned.
1931 *              Should be at least nr_pages long. Or NULL, if caller
1932 *              only intends to ensure the pages are faulted in.
1933 * @vmas:       array of pointers to vmas corresponding to each page.
1934 *              Or NULL if the caller does not require them.
1935 *
1936 * Returns number of pages pinned. This may be fewer than the number
1937 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1938 * were pinned, returns -errno. Each page returned must be released
1939 * with a put_page() call when it is finished with. vmas will only
1940 * remain valid while mmap_sem is held.
1941 *
1942 * Must be called with mmap_sem held for read or write.
1943 *
1944 * get_user_pages walks a process's page tables and takes a reference to
1945 * each struct page that each user address corresponds to at a given
1946 * instant. That is, it takes the page that would be accessed if a user
1947 * thread accesses the given user virtual address at that instant.
1948 *
1949 * This does not guarantee that the page exists in the user mappings when
1950 * get_user_pages returns, and there may even be a completely different
1951 * page there in some cases (eg. if mmapped pagecache has been invalidated
1952 * and subsequently re faulted). However it does guarantee that the page
1953 * won't be freed completely. And mostly callers simply care that the page
1954 * contains data that was valid *at some point in time*. Typically, an IO
1955 * or similar operation cannot guarantee anything stronger anyway because
1956 * locks can't be held over the syscall boundary.
1957 *
1958 * If write=0, the page must not be written to. If the page is written to,
1959 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1960 * after the page is finished with, and before put_page is called.
1961 *
1962 * get_user_pages is typically used for fewer-copy IO operations, to get a
1963 * handle on the memory by some means other than accesses via the user virtual
1964 * addresses. The pages may be submitted for DMA to devices or accessed via
1965 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1966 * use the correct cache flushing APIs.
1967 *
1968 * See also get_user_pages_fast, for performance critical applications.
1969 */
1970int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1971                unsigned long start, int nr_pages, int write, int force,
1972                struct page **pages, struct vm_area_struct **vmas)
1973{
1974        int flags = FOLL_TOUCH;
1975
1976        if (pages)
1977                flags |= FOLL_GET;
1978        if (write)
1979                flags |= FOLL_WRITE;
1980        if (force)
1981                flags |= FOLL_FORCE;
1982
1983        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1984                                NULL);
1985}
1986EXPORT_SYMBOL(get_user_pages);
1987
1988/**
1989 * get_dump_page() - pin user page in memory while writing it to core dump
1990 * @addr: user address
1991 *
1992 * Returns struct page pointer of user page pinned for dump,
1993 * to be freed afterwards by page_cache_release() or put_page().
1994 *
1995 * Returns NULL on any kind of failure - a hole must then be inserted into
1996 * the corefile, to preserve alignment with its headers; and also returns
1997 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1998 * allowing a hole to be left in the corefile to save diskspace.
1999 *
2000 * Called without mmap_sem, but after all other threads have been killed.
2001 */
2002#ifdef CONFIG_ELF_CORE
2003struct page *get_dump_page(unsigned long addr)
2004{
2005        struct vm_area_struct *vma;
2006        struct page *page;
2007
2008        if (__get_user_pages(current, current->mm, addr, 1,
2009                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2010                             NULL) < 1)
2011                return NULL;
2012        flush_cache_page(vma, addr, page_to_pfn(page));
2013        return page;
2014}
2015#endif /* CONFIG_ELF_CORE */
2016
2017pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2018                        spinlock_t **ptl)
2019{
2020        pgd_t * pgd = pgd_offset(mm, addr);
2021        pud_t * pud = pud_alloc(mm, pgd, addr);
2022        if (pud) {
2023                pmd_t * pmd = pmd_alloc(mm, pud, addr);
2024                if (pmd) {
2025                        VM_BUG_ON(pmd_trans_huge(*pmd));
2026                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
2027                }
2028        }
2029        return NULL;
2030}
2031
2032/*
2033 * This is the old fallback for page remapping.
2034 *
2035 * For historical reasons, it only allows reserved pages. Only
2036 * old drivers should use this, and they needed to mark their
2037 * pages reserved for the old functions anyway.
2038 */
2039static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2040                        struct page *page, pgprot_t prot)
2041{
2042        struct mm_struct *mm = vma->vm_mm;
2043        int retval;
2044        pte_t *pte;
2045        spinlock_t *ptl;
2046
2047        retval = -EINVAL;
2048        if (PageAnon(page))
2049                goto out;
2050        retval = -ENOMEM;
2051        flush_dcache_page(page);
2052        pte = get_locked_pte(mm, addr, &ptl);
2053        if (!pte)
2054                goto out;
2055        retval = -EBUSY;
2056        if (!pte_none(*pte))
2057                goto out_unlock;
2058
2059        /* Ok, finally just insert the thing.. */
2060        get_page(page);
2061        inc_mm_counter_fast(mm, MM_FILEPAGES);
2062        page_add_file_rmap(page);
2063        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2064
2065        retval = 0;
2066        pte_unmap_unlock(pte, ptl);
2067        return retval;
2068out_unlock:
2069        pte_unmap_unlock(pte, ptl);
2070out:
2071        return retval;
2072}
2073
2074/**
2075 * vm_insert_page - insert single page into user vma
2076 * @vma: user vma to map to
2077 * @addr: target user address of this page
2078 * @page: source kernel page
2079 *
2080 * This allows drivers to insert individual pages they've allocated
2081 * into a user vma.
2082 *
2083 * The page has to be a nice clean _individual_ kernel allocation.
2084 * If you allocate a compound page, you need to have marked it as
2085 * such (__GFP_COMP), or manually just split the page up yourself
2086 * (see split_page()).
2087 *
2088 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2089 * took an arbitrary page protection parameter. This doesn't allow
2090 * that. Your vma protection will have to be set up correctly, which
2091 * means that if you want a shared writable mapping, you'd better
2092 * ask for a shared writable mapping!
2093 *
2094 * The page does not need to be reserved.
2095 *
2096 * Usually this function is called from f_op->mmap() handler
2097 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2098 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2099 * function from other places, for example from page-fault handler.
2100 */
2101int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2102                        struct page *page)
2103{
2104        if (addr < vma->vm_start || addr >= vma->vm_end)
2105                return -EFAULT;
2106        if (!page_count(page))
2107                return -EINVAL;
2108        if (!(vma->vm_flags & VM_MIXEDMAP)) {
2109                BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2110                BUG_ON(vma->vm_flags & VM_PFNMAP);
2111                vma->vm_flags |= VM_MIXEDMAP;
2112        }
2113        return insert_page(vma, addr, page, vma->vm_page_prot);
2114}
2115EXPORT_SYMBOL(vm_insert_page);
2116
2117static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2118                        unsigned long pfn, pgprot_t prot)
2119{
2120        struct mm_struct *mm = vma->vm_mm;
2121        int retval;
2122        pte_t *pte, entry;
2123        spinlock_t *ptl;
2124
2125        retval = -ENOMEM;
2126        pte = get_locked_pte(mm, addr, &ptl);
2127        if (!pte)
2128                goto out;
2129        retval = -EBUSY;
2130        if (!pte_none(*pte))
2131                goto out_unlock;
2132
2133        /* Ok, finally just insert the thing.. */
2134        entry = pte_mkspecial(pfn_pte(pfn, prot));
2135        set_pte_at(mm, addr, pte, entry);
2136        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2137
2138        retval = 0;
2139out_unlock:
2140        pte_unmap_unlock(pte, ptl);
2141out:
2142        return retval;
2143}
2144
2145/**
2146 * vm_insert_pfn - insert single pfn into user vma
2147 * @vma: user vma to map to
2148 * @addr: target user address of this page
2149 * @pfn: source kernel pfn
2150 *
2151 * Similar to vm_insert_page, this allows drivers to insert individual pages
2152 * they've allocated into a user vma. Same comments apply.
2153 *
2154 * This function should only be called from a vm_ops->fault handler, and
2155 * in that case the handler should return NULL.
2156 *
2157 * vma cannot be a COW mapping.
2158 *
2159 * As this is called only for pages that do not currently exist, we
2160 * do not need to flush old virtual caches or the TLB.
2161 */
2162int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2163                        unsigned long pfn)
2164{
2165        int ret;
2166        pgprot_t pgprot = vma->vm_page_prot;
2167        /*
2168         * Technically, architectures with pte_special can avoid all these
2169         * restrictions (same for remap_pfn_range).  However we would like
2170         * consistency in testing and feature parity among all, so we should
2171         * try to keep these invariants in place for everybody.
2172         */
2173        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2174        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2175                                                (VM_PFNMAP|VM_MIXEDMAP));
2176        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2177        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2178
2179        if (addr < vma->vm_start || addr >= vma->vm_end)
2180                return -EFAULT;
2181        if (track_pfn_insert(vma, &pgprot, pfn))
2182                return -EINVAL;
2183
2184        ret = insert_pfn(vma, addr, pfn, pgprot);
2185
2186        return ret;
2187}
2188EXPORT_SYMBOL(vm_insert_pfn);
2189
2190int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2191                        unsigned long pfn)
2192{
2193        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2194
2195        if (addr < vma->vm_start || addr >= vma->vm_end)
2196                return -EFAULT;
2197
2198        /*
2199         * If we don't have pte special, then we have to use the pfn_valid()
2200         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2201         * refcount the page if pfn_valid is true (hence insert_page rather
2202         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2203         * without pte special, it would there be refcounted as a normal page.
2204         */
2205        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2206                struct page *page;
2207
2208                page = pfn_to_page(pfn);
2209                return insert_page(vma, addr, page, vma->vm_page_prot);
2210        }
2211        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2212}
2213EXPORT_SYMBOL(vm_insert_mixed);
2214
2215/*
2216 * maps a range of physical memory into the requested pages. the old
2217 * mappings are removed. any references to nonexistent pages results
2218 * in null mappings (currently treated as "copy-on-access")
2219 */
2220static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2221                        unsigned long addr, unsigned long end,
2222                        unsigned long pfn, pgprot_t prot)
2223{
2224        pte_t *pte;
2225        spinlock_t *ptl;
2226
2227        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2228        if (!pte)
2229                return -ENOMEM;
2230        arch_enter_lazy_mmu_mode();
2231        do {
2232                BUG_ON(!pte_none(*pte));
2233                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2234                pfn++;
2235        } while (pte++, addr += PAGE_SIZE, addr != end);
2236        arch_leave_lazy_mmu_mode();
2237        pte_unmap_unlock(pte - 1, ptl);
2238        return 0;
2239}
2240
2241static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2242                        unsigned long addr, unsigned long end,
2243                        unsigned long pfn, pgprot_t prot)
2244{
2245        pmd_t *pmd;
2246        unsigned long next;
2247
2248        pfn -= addr >> PAGE_SHIFT;
2249        pmd = pmd_alloc(mm, pud, addr);
2250        if (!pmd)
2251                return -ENOMEM;
2252        VM_BUG_ON(pmd_trans_huge(*pmd));
2253        do {
2254                next = pmd_addr_end(addr, end);
2255                if (remap_pte_range(mm, pmd, addr, next,
2256                                pfn + (addr >> PAGE_SHIFT), prot))
2257                        return -ENOMEM;
2258        } while (pmd++, addr = next, addr != end);
2259        return 0;
2260}
2261
2262static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2263                        unsigned long addr, unsigned long end,
2264                        unsigned long pfn, pgprot_t prot)
2265{
2266        pud_t *pud;
2267        unsigned long next;
2268
2269        pfn -= addr >> PAGE_SHIFT;
2270        pud = pud_alloc(mm, pgd, addr);
2271        if (!pud)
2272                return -ENOMEM;
2273        do {
2274                next = pud_addr_end(addr, end);
2275                if (remap_pmd_range(mm, pud, addr, next,
2276                                pfn + (addr >> PAGE_SHIFT), prot))
2277                        return -ENOMEM;
2278        } while (pud++, addr = next, addr != end);
2279        return 0;
2280}
2281
2282/**
2283 * remap_pfn_range - remap kernel memory to userspace
2284 * @vma: user vma to map to
2285 * @addr: target user address to start at
2286 * @pfn: physical address of kernel memory
2287 * @size: size of map area
2288 * @prot: page protection flags for this mapping
2289 *
2290 *  Note: this is only safe if the mm semaphore is held when called.
2291 */
2292int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2293                    unsigned long pfn, unsigned long size, pgprot_t prot)
2294{
2295        pgd_t *pgd;
2296        unsigned long next;
2297        unsigned long end = addr + PAGE_ALIGN(size);
2298        struct mm_struct *mm = vma->vm_mm;
2299        int err;
2300
2301        /*
2302         * Physically remapped pages are special. Tell the
2303         * rest of the world about it:
2304         *   VM_IO tells people not to look at these pages
2305         *      (accesses can have side effects).
2306         *   VM_PFNMAP tells the core MM that the base pages are just
2307         *      raw PFN mappings, and do not have a "struct page" associated
2308         *      with them.
2309         *   VM_DONTEXPAND
2310         *      Disable vma merging and expanding with mremap().
2311         *   VM_DONTDUMP
2312         *      Omit vma from core dump, even when VM_IO turned off.
2313         *
2314         * There's a horrible special case to handle copy-on-write
2315         * behaviour that some programs depend on. We mark the "original"
2316         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2317         * See vm_normal_page() for details.
2318         */
2319        if (is_cow_mapping(vma->vm_flags)) {
2320                if (addr != vma->vm_start || end != vma->vm_end)
2321                        return -EINVAL;
2322                vma->vm_pgoff = pfn;
2323        }
2324
2325        err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2326        if (err)
2327                return -EINVAL;
2328
2329        vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2330
2331        BUG_ON(addr >= end);
2332        pfn -= addr >> PAGE_SHIFT;
2333        pgd = pgd_offset(mm, addr);
2334        flush_cache_range(vma, addr, end);
2335        do {
2336                next = pgd_addr_end(addr, end);
2337                err = remap_pud_range(mm, pgd, addr, next,
2338                                pfn + (addr >> PAGE_SHIFT), prot);
2339                if (err)
2340                        break;
2341        } while (pgd++, addr = next, addr != end);
2342
2343        if (err)
2344                untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2345
2346        return err;
2347}
2348EXPORT_SYMBOL(remap_pfn_range);
2349
2350static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2351                                     unsigned long addr, unsigned long end,
2352                                     pte_fn_t fn, void *data)
2353{
2354        pte_t *pte;
2355        int err;
2356        pgtable_t token;
2357        spinlock_t *uninitialized_var(ptl);
2358
2359        pte = (mm == &init_mm) ?
2360                pte_alloc_kernel(pmd, addr) :
2361                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2362        if (!pte)
2363                return -ENOMEM;
2364
2365        BUG_ON(pmd_huge(*pmd));
2366
2367        arch_enter_lazy_mmu_mode();
2368
2369        token = pmd_pgtable(*pmd);
2370
2371        do {
2372                err = fn(pte++, token, addr, data);
2373                if (err)
2374                        break;
2375        } while (addr += PAGE_SIZE, addr != end);
2376
2377        arch_leave_lazy_mmu_mode();
2378
2379        if (mm != &init_mm)
2380                pte_unmap_unlock(pte-1, ptl);
2381        return err;
2382}
2383
2384static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2385                                     unsigned long addr, unsigned long end,
2386                                     pte_fn_t fn, void *data)
2387{
2388        pmd_t *pmd;
2389        unsigned long next;
2390        int err;
2391
2392        BUG_ON(pud_huge(*pud));
2393
2394        pmd = pmd_alloc(mm, pud, addr);
2395        if (!pmd)
2396                return -ENOMEM;
2397        do {
2398                next = pmd_addr_end(addr, end);
2399                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2400                if (err)
2401                        break;
2402        } while (pmd++, addr = next, addr != end);
2403        return err;
2404}
2405
2406static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2407                                     unsigned long addr, unsigned long end,
2408                                     pte_fn_t fn, void *data)
2409{
2410        pud_t *pud;
2411        unsigned long next;
2412        int err;
2413
2414        pud = pud_alloc(mm, pgd, addr);
2415        if (!pud)
2416                return -ENOMEM;
2417        do {
2418                next = pud_addr_end(addr, end);
2419                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2420                if (err)
2421                        break;
2422        } while (pud++, addr = next, addr != end);
2423        return err;
2424}
2425
2426/*
2427 * Scan a region of virtual memory, filling in page tables as necessary
2428 * and calling a provided function on each leaf page table.
2429 */
2430int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2431                        unsigned long size, pte_fn_t fn, void *data)
2432{
2433        pgd_t *pgd;
2434        unsigned long next;
2435        unsigned long end = addr + size;
2436        int err;
2437
2438        BUG_ON(addr >= end);
2439        pgd = pgd_offset(mm, addr);
2440        do {
2441                next = pgd_addr_end(addr, end);
2442                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2443                if (err)
2444                        break;
2445        } while (pgd++, addr = next, addr != end);
2446
2447        return err;
2448}
2449EXPORT_SYMBOL_GPL(apply_to_page_range);
2450
2451/*
2452 * handle_pte_fault chooses page fault handler according to an entry
2453 * which was read non-atomically.  Before making any commitment, on
2454 * those architectures or configurations (e.g. i386 with PAE) which
2455 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2456 * must check under lock before unmapping the pte and proceeding
2457 * (but do_wp_page is only called after already making such a check;
2458 * and do_anonymous_page can safely check later on).
2459 */
2460static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2461                                pte_t *page_table, pte_t orig_pte)
2462{
2463        int same = 1;
2464#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2465        if (sizeof(pte_t) > sizeof(unsigned long)) {
2466                spinlock_t *ptl = pte_lockptr(mm, pmd);
2467                spin_lock(ptl);
2468                same = pte_same(*page_table, orig_pte);
2469                spin_unlock(ptl);
2470        }
2471#endif
2472        pte_unmap(page_table);
2473        return same;
2474}
2475
2476static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2477{
2478        /*
2479         * If the source page was a PFN mapping, we don't have
2480         * a "struct page" for it. We do a best-effort copy by
2481         * just copying from the original user address. If that
2482         * fails, we just zero-fill it. Live with it.
2483         */
2484        if (unlikely(!src)) {
2485                void *kaddr = kmap_atomic(dst);
2486                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2487
2488                /*
2489                 * This really shouldn't fail, because the page is there
2490                 * in the page tables. But it might just be unreadable,
2491                 * in which case we just give up and fill the result with
2492                 * zeroes.
2493                 */
2494                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2495                        clear_page(kaddr);
2496                kunmap_atomic(kaddr);
2497                flush_dcache_page(dst);
2498        } else
2499                copy_user_highpage(dst, src, va, vma);
2500}
2501
2502/*
2503 * This routine handles present pages, when users try to write
2504 * to a shared page. It is done by copying the page to a new address
2505 * and decrementing the shared-page counter for the old page.
2506 *
2507 * Note that this routine assumes that the protection checks have been
2508 * done by the caller (the low-level page fault routine in most cases).
2509 * Thus we can safely just mark it writable once we've done any necessary
2510 * COW.
2511 *
2512 * We also mark the page dirty at this point even though the page will
2513 * change only once the write actually happens. This avoids a few races,
2514 * and potentially makes it more efficient.
2515 *
2516 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2517 * but allow concurrent faults), with pte both mapped and locked.
2518 * We return with mmap_sem still held, but pte unmapped and unlocked.
2519 */
2520static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2521                unsigned long address, pte_t *page_table, pmd_t *pmd,
2522                spinlock_t *ptl, pte_t orig_pte)
2523        __releases(ptl)
2524{
2525        struct page *old_page, *new_page = NULL;
2526        pte_t entry;
2527        int ret = 0;
2528        int page_mkwrite = 0;
2529        struct page *dirty_page = NULL;
2530        unsigned long mmun_start = 0;   /* For mmu_notifiers */
2531        unsigned long mmun_end = 0;     /* For mmu_notifiers */
2532
2533        old_page = vm_normal_page(vma, address, orig_pte);
2534        if (!old_page) {
2535                /*
2536                 * VM_MIXEDMAP !pfn_valid() case
2537                 *
2538                 * We should not cow pages in a shared writeable mapping.
2539                 * Just mark the pages writable as we can't do any dirty
2540                 * accounting on raw pfn maps.
2541                 */
2542                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2543                                     (VM_WRITE|VM_SHARED))
2544                        goto reuse;
2545                goto gotten;
2546        }
2547
2548        /*
2549         * Take out anonymous pages first, anonymous shared vmas are
2550         * not dirty accountable.
2551         */
2552        if (PageAnon(old_page) && !PageKsm(old_page)) {
2553                if (!trylock_page(old_page)) {
2554                        page_cache_get(old_page);
2555                        pte_unmap_unlock(page_table, ptl);
2556                        lock_page(old_page);
2557                        page_table = pte_offset_map_lock(mm, pmd, address,
2558                                                         &ptl);
2559                        if (!pte_same(*page_table, orig_pte)) {
2560                                unlock_page(old_page);
2561                                goto unlock;
2562                        }
2563                        page_cache_release(old_page);
2564                }
2565                if (reuse_swap_page(old_page)) {
2566                        /*
2567                         * The page is all ours.  Move it to our anon_vma so
2568                         * the rmap code will not search our parent or siblings.
2569                         * Protected against the rmap code by the page lock.
2570                         */
2571                        page_move_anon_rmap(old_page, vma, address);
2572                        unlock_page(old_page);
2573                        goto reuse;
2574                }
2575                unlock_page(old_page);
2576        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2577                                        (VM_WRITE|VM_SHARED))) {
2578                /*
2579                 * Only catch write-faults on shared writable pages,
2580                 * read-only shared pages can get COWed by
2581                 * get_user_pages(.write=1, .force=1).
2582                 */
2583                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2584                        struct vm_fault vmf;
2585                        int tmp;
2586
2587                        vmf.virtual_address = (void __user *)(address &
2588                                                                PAGE_MASK);
2589                        vmf.pgoff = old_page->index;
2590                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2591                        vmf.page = old_page;
2592
2593                        /*
2594                         * Notify the address space that the page is about to
2595                         * become writable so that it can prohibit this or wait
2596                         * for the page to get into an appropriate state.
2597                         *
2598                         * We do this without the lock held, so that it can
2599                         * sleep if it needs to.
2600                         */
2601                        page_cache_get(old_page);
2602                        pte_unmap_unlock(page_table, ptl);
2603
2604                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2605                        if (unlikely(tmp &
2606                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2607                                ret = tmp;
2608                                goto unwritable_page;
2609                        }
2610                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2611                                lock_page(old_page);
2612                                if (!old_page->mapping) {
2613                                        ret = 0; /* retry the fault */
2614                                        unlock_page(old_page);
2615                                        goto unwritable_page;
2616                                }
2617                        } else
2618                                VM_BUG_ON(!PageLocked(old_page));
2619
2620                        /*
2621                         * Since we dropped the lock we need to revalidate
2622                         * the PTE as someone else may have changed it.  If
2623                         * they did, we just return, as we can count on the
2624                         * MMU to tell us if they didn't also make it writable.
2625                         */
2626                        page_table = pte_offset_map_lock(mm, pmd, address,
2627                                                         &ptl);
2628                        if (!pte_same(*page_table, orig_pte)) {
2629                                unlock_page(old_page);
2630                                goto unlock;
2631                        }
2632
2633                        page_mkwrite = 1;
2634                }
2635                dirty_page = old_page;
2636                get_page(dirty_page);
2637
2638reuse:
2639                flush_cache_page(vma, address, pte_pfn(orig_pte));
2640                entry = pte_mkyoung(orig_pte);
2641                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2642                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2643                        update_mmu_cache(vma, address, page_table);
2644                pte_unmap_unlock(page_table, ptl);
2645                ret |= VM_FAULT_WRITE;
2646
2647                if (!dirty_page)
2648                        return ret;
2649
2650                /*
2651                 * Yes, Virginia, this is actually required to prevent a race
2652                 * with clear_page_dirty_for_io() from clearing the page dirty
2653                 * bit after it clear all dirty ptes, but before a racing
2654                 * do_wp_page installs a dirty pte.
2655                 *
2656                 * __do_fault is protected similarly.
2657                 */
2658                if (!page_mkwrite) {
2659                        wait_on_page_locked(dirty_page);
2660                        set_page_dirty_balance(dirty_page, page_mkwrite);
2661                        /* file_update_time outside page_lock */
2662                        if (vma->vm_file)
2663                                file_update_time(vma->vm_file);
2664                }
2665                put_page(dirty_page);
2666                if (page_mkwrite) {
2667                        struct address_space *mapping = dirty_page->mapping;
2668
2669                        set_page_dirty(dirty_page);
2670                        unlock_page(dirty_page);
2671                        page_cache_release(dirty_page);
2672                        if (mapping)    {
2673                                /*
2674                                 * Some device drivers do not set page.mapping
2675                                 * but still dirty their pages
2676                                 */
2677                                balance_dirty_pages_ratelimited(mapping);
2678                        }
2679                }
2680
2681                return ret;
2682        }
2683
2684        /*
2685         * Ok, we need to copy. Oh, well..
2686         */
2687        page_cache_get(old_page);
2688gotten:
2689        pte_unmap_unlock(page_table, ptl);
2690
2691        if (unlikely(anon_vma_prepare(vma)))
2692                goto oom;
2693
2694        if (is_zero_pfn(pte_pfn(orig_pte))) {
2695                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2696                if (!new_page)
2697                        goto oom;
2698        } else {
2699                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2700                if (!new_page)
2701                        goto oom;
2702                cow_user_page(new_page, old_page, address, vma);
2703        }
2704        __SetPageUptodate(new_page);
2705
2706        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2707                goto oom_free_new;
2708
2709        mmun_start  = address & PAGE_MASK;
2710        mmun_end    = mmun_start + PAGE_SIZE;
2711        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2712
2713        /*
2714         * Re-check the pte - we dropped the lock
2715         */
2716        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2717        if (likely(pte_same(*page_table, orig_pte))) {
2718                if (old_page) {
2719                        if (!PageAnon(old_page)) {
2720                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2721                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2722                        }
2723                } else
2724                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2725                flush_cache_page(vma, address, pte_pfn(orig_pte));
2726                entry = mk_pte(new_page, vma->vm_page_prot);
2727                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2728                /*
2729                 * Clear the pte entry and flush it first, before updating the
2730                 * pte with the new entry. This will avoid a race condition
2731                 * seen in the presence of one thread doing SMC and another
2732                 * thread doing COW.
2733                 */
2734                ptep_clear_flush(vma, address, page_table);
2735                page_add_new_anon_rmap(new_page, vma, address);
2736                /*
2737                 * We call the notify macro here because, when using secondary
2738                 * mmu page tables (such as kvm shadow page tables), we want the
2739                 * new page to be mapped directly into the secondary page table.
2740                 */
2741                set_pte_at_notify(mm, address, page_table, entry);
2742                update_mmu_cache(vma, address, page_table);
2743                if (old_page) {
2744                        /*
2745                         * Only after switching the pte to the new page may
2746                         * we remove the mapcount here. Otherwise another
2747                         * process may come and find the rmap count decremented
2748                         * before the pte is switched to the new page, and
2749                         * "reuse" the old page writing into it while our pte
2750                         * here still points into it and can be read by other
2751                         * threads.
2752                         *
2753                         * The critical issue is to order this
2754                         * page_remove_rmap with the ptp_clear_flush above.
2755                         * Those stores are ordered by (if nothing else,)
2756                         * the barrier present in the atomic_add_negative
2757                         * in page_remove_rmap.
2758                         *
2759                         * Then the TLB flush in ptep_clear_flush ensures that
2760                         * no process can access the old page before the
2761                         * decremented mapcount is visible. And the old page
2762                         * cannot be reused until after the decremented
2763                         * mapcount is visible. So transitively, TLBs to
2764                         * old page will be flushed before it can be reused.
2765                         */
2766                        page_remove_rmap(old_page);
2767                }
2768
2769                /* Free the old page.. */
2770                new_page = old_page;
2771                ret |= VM_FAULT_WRITE;
2772        } else
2773                mem_cgroup_uncharge_page(new_page);
2774
2775        if (new_page)
2776                page_cache_release(new_page);
2777unlock:
2778        pte_unmap_unlock(page_table, ptl);
2779        if (mmun_end > mmun_start)
2780                mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2781        if (old_page) {
2782                /*
2783                 * Don't let another task, with possibly unlocked vma,
2784                 * keep the mlocked page.
2785                 */
2786                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2787                        lock_page(old_page);    /* LRU manipulation */
2788                        munlock_vma_page(old_page);
2789                        unlock_page(old_page);
2790                }
2791                page_cache_release(old_page);
2792        }
2793        return ret;
2794oom_free_new:
2795        page_cache_release(new_page);
2796oom:
2797        if (old_page) {
2798                if (page_mkwrite) {
2799                        unlock_page(old_page);
2800                        page_cache_release(old_page);
2801                }
2802                page_cache_release(old_page);
2803        }
2804        return VM_FAULT_OOM;
2805
2806unwritable_page:
2807        page_cache_release(old_page);
2808        return ret;
2809}
2810
2811static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2812                unsigned long start_addr, unsigned long end_addr,
2813                struct zap_details *details)
2814{
2815        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2816}
2817
2818static inline void unmap_mapping_range_tree(struct rb_root *root,
2819                                            struct zap_details *details)
2820{
2821        struct vm_area_struct *vma;
2822        pgoff_t vba, vea, zba, zea;
2823
2824        vma_interval_tree_foreach(vma, root,
2825                        details->first_index, details->last_index) {
2826
2827                vba = vma->vm_pgoff;
2828                vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2829                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2830                zba = details->first_index;
2831                if (zba < vba)
2832                        zba = vba;
2833                zea = details->last_index;
2834                if (zea > vea)
2835                        zea = vea;
2836
2837                unmap_mapping_range_vma(vma,
2838                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2839                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2840                                details);
2841        }
2842}
2843
2844static inline void unmap_mapping_range_list(struct list_head *head,
2845                                            struct zap_details *details)
2846{
2847        struct vm_area_struct *vma;
2848
2849        /*
2850         * In nonlinear VMAs there is no correspondence between virtual address
2851         * offset and file offset.  So we must perform an exhaustive search
2852         * across *all* the pages in each nonlinear VMA, not just the pages
2853         * whose virtual address lies outside the file truncation point.
2854         */
2855        list_for_each_entry(vma, head, shared.nonlinear) {
2856                details->nonlinear_vma = vma;
2857                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2858        }
2859}
2860
2861/**
2862 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2863 * @mapping: the address space containing mmaps to be unmapped.
2864 * @holebegin: byte in first page to unmap, relative to the start of
2865 * the underlying file.  This will be rounded down to a PAGE_SIZE
2866 * boundary.  Note that this is different from truncate_pagecache(), which
2867 * must keep the partial page.  In contrast, we must get rid of
2868 * partial pages.
2869 * @holelen: size of prospective hole in bytes.  This will be rounded
2870 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2871 * end of the file.
2872 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2873 * but 0 when invalidating pagecache, don't throw away private data.
2874 */
2875void unmap_mapping_range(struct address_space *mapping,
2876                loff_t const holebegin, loff_t const holelen, int even_cows)
2877{
2878        struct zap_details details;
2879        pgoff_t hba = holebegin >> PAGE_SHIFT;
2880        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2881
2882        /* Check for overflow. */
2883        if (sizeof(holelen) > sizeof(hlen)) {
2884                long long holeend =
2885                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2886                if (holeend & ~(long long)ULONG_MAX)
2887                        hlen = ULONG_MAX - hba + 1;
2888        }
2889
2890        details.check_mapping = even_cows? NULL: mapping;
2891        details.nonlinear_vma = NULL;
2892        details.first_index = hba;
2893        details.last_index = hba + hlen - 1;
2894        if (details.last_index < details.first_index)
2895                details.last_index = ULONG_MAX;
2896
2897
2898        mutex_lock(&mapping->i_mmap_mutex);
2899        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2900                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2901        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2902                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2903        mutex_unlock(&mapping->i_mmap_mutex);
2904}
2905EXPORT_SYMBOL(unmap_mapping_range);
2906
2907/*
2908 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2909 * but allow concurrent faults), and pte mapped but not yet locked.
2910 * We return with mmap_sem still held, but pte unmapped and unlocked.
2911 */
2912static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2913                unsigned long address, pte_t *page_table, pmd_t *pmd,
2914                unsigned int flags, pte_t orig_pte)
2915{
2916        spinlock_t *ptl;
2917        struct page *page, *swapcache = NULL;
2918        swp_entry_t entry;
2919        pte_t pte;
2920        int locked;
2921        struct mem_cgroup *ptr;
2922        int exclusive = 0;
2923        int ret = 0;
2924
2925        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2926                goto out;
2927
2928        entry = pte_to_swp_entry(orig_pte);
2929        if (unlikely(non_swap_entry(entry))) {
2930                if (is_migration_entry(entry)) {
2931                        migration_entry_wait(mm, pmd, address);
2932                } else if (is_hwpoison_entry(entry)) {
2933                        ret = VM_FAULT_HWPOISON;
2934                } else {
2935                        print_bad_pte(vma, address, orig_pte, NULL);
2936                        ret = VM_FAULT_SIGBUS;
2937                }
2938                goto out;
2939        }
2940        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2941        page = lookup_swap_cache(entry);
2942        if (!page) {
2943                page = swapin_readahead(entry,
2944                                        GFP_HIGHUSER_MOVABLE, vma, address);
2945                if (!page) {
2946                        /*
2947                         * Back out if somebody else faulted in this pte
2948                         * while we released the pte lock.
2949                         */
2950                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2951                        if (likely(pte_same(*page_table, orig_pte)))
2952                                ret = VM_FAULT_OOM;
2953                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2954                        goto unlock;
2955                }
2956
2957                /* Had to read the page from swap area: Major fault */
2958                ret = VM_FAULT_MAJOR;
2959                count_vm_event(PGMAJFAULT);
2960                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2961        } else if (PageHWPoison(page)) {
2962                /*
2963                 * hwpoisoned dirty swapcache pages are kept for killing
2964                 * owner processes (which may be unknown at hwpoison time)
2965                 */
2966                ret = VM_FAULT_HWPOISON;
2967                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2968                goto out_release;
2969        }
2970
2971        locked = lock_page_or_retry(page, mm, flags);
2972
2973        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2974        if (!locked) {
2975                ret |= VM_FAULT_RETRY;
2976                goto out_release;
2977        }
2978
2979        /*
2980         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2981         * release the swapcache from under us.  The page pin, and pte_same
2982         * test below, are not enough to exclude that.  Even if it is still
2983         * swapcache, we need to check that the page's swap has not changed.
2984         */
2985        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2986                goto out_page;
2987
2988        if (ksm_might_need_to_copy(page, vma, address)) {
2989                swapcache = page;
2990                page = ksm_does_need_to_copy(page, vma, address);
2991
2992                if (unlikely(!page)) {
2993                        ret = VM_FAULT_OOM;
2994                        page = swapcache;
2995                        swapcache = NULL;
2996                        goto out_page;
2997                }
2998        }
2999
3000        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3001                ret = VM_FAULT_OOM;
3002                goto out_page;
3003        }
3004
3005        /*
3006         * Back out if somebody else already faulted in this pte.
3007         */
3008        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3009        if (unlikely(!pte_same(*page_table, orig_pte)))
3010                goto out_nomap;
3011
3012        if (unlikely(!PageUptodate(page))) {
3013                ret = VM_FAULT_SIGBUS;
3014                goto out_nomap;
3015        }
3016
3017        /*
3018         * The page isn't present yet, go ahead with the fault.
3019         *
3020         * Be careful about the sequence of operations here.
3021         * To get its accounting right, reuse_swap_page() must be called
3022         * while the page is counted on swap but not yet in mapcount i.e.
3023         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3024         * must be called after the swap_free(), or it will never succeed.
3025         * Because delete_from_swap_page() may be called by reuse_swap_page(),
3026         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3027         * in page->private. In this case, a record in swap_cgroup  is silently
3028         * discarded at swap_free().
3029         */
3030
3031        inc_mm_counter_fast(mm, MM_ANONPAGES);
3032        dec_mm_counter_fast(mm, MM_SWAPENTS);
3033        pte = mk_pte(page, vma->vm_page_prot);
3034        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3035                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3036                flags &= ~FAULT_FLAG_WRITE;
3037                ret |= VM_FAULT_WRITE;
3038                exclusive = 1;
3039        }
3040        flush_icache_page(vma, page);
3041        set_pte_at(mm, address, page_table, pte);
3042        do_page_add_anon_rmap(page, vma, address, exclusive);
3043        /* It's better to call commit-charge after rmap is established */
3044        mem_cgroup_commit_charge_swapin(page, ptr);
3045
3046        swap_free(entry);
3047        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3048                try_to_free_swap(page);
3049        unlock_page(page);
3050        if (swapcache) {
3051                /*
3052                 * Hold the lock to avoid the swap entry to be reused
3053                 * until we take the PT lock for the pte_same() check
3054                 * (to avoid false positives from pte_same). For
3055                 * further safety release the lock after the swap_free
3056                 * so that the swap count won't change under a
3057                 * parallel locked swapcache.
3058                 */
3059                unlock_page(swapcache);
3060                page_cache_release(swapcache);
3061        }
3062
3063        if (flags & FAULT_FLAG_WRITE) {
3064                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3065                if (ret & VM_FAULT_ERROR)
3066                        ret &= VM_FAULT_ERROR;
3067                goto out;
3068        }
3069
3070        /* No need to invalidate - it was non-present before */
3071        update_mmu_cache(vma, address, page_table);
3072unlock:
3073        pte_unmap_unlock(page_table, ptl);
3074out:
3075        return ret;
3076out_nomap:
3077        mem_cgroup_cancel_charge_swapin(ptr);
3078        pte_unmap_unlock(page_table, ptl);
3079out_page:
3080        unlock_page(page);
3081out_release:
3082        page_cache_release(page);
3083        if (swapcache) {
3084                unlock_page(swapcache);
3085                page_cache_release(swapcache);
3086        }
3087        return ret;
3088}
3089
3090/*
3091 * This is like a special single-page "expand_{down|up}wards()",
3092 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3093 * doesn't hit another vma.
3094 */
3095static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3096{
3097        address &= PAGE_MASK;
3098        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3099                struct vm_area_struct *prev = vma->vm_prev;
3100
3101                /*
3102                 * Is there a mapping abutting this one below?
3103                 *
3104                 * That's only ok if it's the same stack mapping
3105                 * that has gotten split..
3106                 */
3107                if (prev && prev->vm_end == address)
3108                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3109
3110                expand_downwards(vma, address - PAGE_SIZE);
3111        }
3112        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3113                struct vm_area_struct *next = vma->vm_next;
3114
3115                /* As VM_GROWSDOWN but s/below/above/ */
3116                if (next && next->vm_start == address + PAGE_SIZE)
3117                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3118
3119                expand_upwards(vma, address + PAGE_SIZE);
3120        }
3121        return 0;
3122}
3123
3124/*
3125 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3126 * but allow concurrent faults), and pte mapped but not yet locked.
3127 * We return with mmap_sem still held, but pte unmapped and unlocked.
3128 */
3129static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3130                unsigned long address, pte_t *page_table, pmd_t *pmd,
3131                unsigned int flags)
3132{
3133        struct page *page;
3134        spinlock_t *ptl;
3135        pte_t entry;
3136
3137        pte_unmap(page_table);
3138
3139        /* Check if we need to add a guard page to the stack */
3140        if (check_stack_guard_page(vma, address) < 0)
3141                return VM_FAULT_SIGBUS;
3142
3143        /* Use the zero-page for reads */
3144        if (!(flags & FAULT_FLAG_WRITE)) {
3145                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3146                                                vma->vm_page_prot));
3147                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3148                if (!pte_none(*page_table))
3149                        goto unlock;
3150                goto setpte;
3151        }
3152
3153        /* Allocate our own private page. */
3154        if (unlikely(anon_vma_prepare(vma)))
3155                goto oom;
3156        page = alloc_zeroed_user_highpage_movable(vma, address);
3157        if (!page)
3158                goto oom;
3159        __SetPageUptodate(page);
3160
3161        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3162                goto oom_free_page;
3163
3164        entry = mk_pte(page, vma->vm_page_prot);
3165        if (vma->vm_flags & VM_WRITE)
3166                entry = pte_mkwrite(pte_mkdirty(entry));
3167
3168        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3169        if (!pte_none(*page_table))
3170                goto release;
3171
3172        inc_mm_counter_fast(mm, MM_ANONPAGES);
3173        page_add_new_anon_rmap(page, vma, address);
3174setpte:
3175        set_pte_at(mm, address, page_table, entry);
3176
3177        /* No need to invalidate - it was non-present before */
3178        update_mmu_cache(vma, address, page_table);
3179unlock:
3180        pte_unmap_unlock(page_table, ptl);
3181        return 0;
3182release:
3183        mem_cgroup_uncharge_page(page);
3184        page_cache_release(page);
3185        goto unlock;
3186oom_free_page:
3187        page_cache_release(page);
3188oom:
3189        return VM_FAULT_OOM;
3190}
3191
3192/*
3193 * __do_fault() tries to create a new page mapping. It aggressively
3194 * tries to share with existing pages, but makes a separate copy if
3195 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3196 * the next page fault.
3197 *
3198 * As this is called only for pages that do not currently exist, we
3199 * do not need to flush old virtual caches or the TLB.
3200 *
3201 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3202 * but allow concurrent faults), and pte neither mapped nor locked.
3203 * We return with mmap_sem still held, but pte unmapped and unlocked.
3204 */
3205static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3206                unsigned long address, pmd_t *pmd,
3207                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3208{
3209        pte_t *page_table;
3210        spinlock_t *ptl;
3211        struct page *page;
3212        struct page *cow_page;
3213        pte_t entry;
3214        int anon = 0;
3215        struct page *dirty_page = NULL;
3216        struct vm_fault vmf;
3217        int ret;
3218        int page_mkwrite = 0;
3219
3220        /*
3221         * If we do COW later, allocate page befor taking lock_page()
3222         * on the file cache page. This will reduce lock holding time.
3223         */
3224        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3225
3226                if (unlikely(anon_vma_prepare(vma)))
3227                        return VM_FAULT_OOM;
3228
3229                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3230                if (!cow_page)
3231                        return VM_FAULT_OOM;
3232
3233                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3234                        page_cache_release(cow_page);
3235                        return VM_FAULT_OOM;
3236                }
3237        } else
3238                cow_page = NULL;
3239
3240        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3241        vmf.pgoff = pgoff;
3242        vmf.flags = flags;
3243        vmf.page = NULL;
3244
3245        ret = vma->vm_ops->fault(vma, &vmf);
3246        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3247                            VM_FAULT_RETRY)))
3248                goto uncharge_out;
3249
3250        if (unlikely(PageHWPoison(vmf.page))) {
3251                if (ret & VM_FAULT_LOCKED)
3252                        unlock_page(vmf.page);
3253                ret = VM_FAULT_HWPOISON;
3254                goto uncharge_out;
3255        }
3256
3257        /*
3258         * For consistency in subsequent calls, make the faulted page always
3259         * locked.
3260         */
3261        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3262                lock_page(vmf.page);
3263        else
3264                VM_BUG_ON(!PageLocked(vmf.page));
3265
3266        /*
3267         * Should we do an early C-O-W break?
3268         */
3269        page = vmf.page;
3270        if (flags & FAULT_FLAG_WRITE) {
3271                if (!(vma->vm_flags & VM_SHARED)) {
3272                        page = cow_page;
3273                        anon = 1;
3274                        copy_user_highpage(page, vmf.page, address, vma);
3275                        __SetPageUptodate(page);
3276                } else {
3277                        /*
3278                         * If the page will be shareable, see if the backing
3279                         * address space wants to know that the page is about
3280                         * to become writable
3281                         */
3282                        if (vma->vm_ops->page_mkwrite) {
3283                                int tmp;
3284
3285                                unlock_page(page);
3286                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3287                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3288                                if (unlikely(tmp &
3289                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3290                                        ret = tmp;
3291                                        goto unwritable_page;
3292                                }
3293                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3294                                        lock_page(page);
3295                                        if (!page->mapping) {
3296                                                ret = 0; /* retry the fault */
3297                                                unlock_page(page);
3298                                                goto unwritable_page;
3299                                        }
3300                                } else
3301                                        VM_BUG_ON(!PageLocked(page));
3302                                page_mkwrite = 1;
3303                        }
3304                }
3305
3306        }
3307
3308        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3309
3310        /*
3311         * This silly early PAGE_DIRTY setting removes a race
3312         * due to the bad i386 page protection. But it's valid
3313         * for other architectures too.
3314         *
3315         * Note that if FAULT_FLAG_WRITE is set, we either now have
3316         * an exclusive copy of the page, or this is a shared mapping,
3317         * so we can make it writable and dirty to avoid having to
3318         * handle that later.
3319         */
3320        /* Only go through if we didn't race with anybody else... */
3321        if (likely(pte_same(*page_table, orig_pte))) {
3322                flush_icache_page(vma, page);
3323                entry = mk_pte(page, vma->vm_page_prot);
3324                if (flags & FAULT_FLAG_WRITE)
3325                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3326                if (anon) {
3327                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3328                        page_add_new_anon_rmap(page, vma, address);
3329                } else {
3330                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3331                        page_add_file_rmap(page);
3332                        if (flags & FAULT_FLAG_WRITE) {
3333                                dirty_page = page;
3334                                get_page(dirty_page);
3335                        }
3336                }
3337                set_pte_at(mm, address, page_table, entry);
3338
3339                /* no need to invalidate: a not-present page won't be cached */
3340                update_mmu_cache(vma, address, page_table);
3341        } else {
3342                if (cow_page)
3343                        mem_cgroup_uncharge_page(cow_page);
3344                if (anon)
3345                        page_cache_release(page);
3346                else
3347                        anon = 1; /* no anon but release faulted_page */
3348        }
3349
3350        pte_unmap_unlock(page_table, ptl);
3351
3352        if (dirty_page) {
3353                struct address_space *mapping = page->mapping;
3354                int dirtied = 0;
3355
3356                if (set_page_dirty(dirty_page))
3357                        dirtied = 1;
3358                unlock_page(dirty_page);
3359                put_page(dirty_page);
3360                if ((dirtied || page_mkwrite) && mapping) {
3361                        /*
3362                         * Some device drivers do not set page.mapping but still
3363                         * dirty their pages
3364                         */
3365                        balance_dirty_pages_ratelimited(mapping);
3366                }
3367
3368                /* file_update_time outside page_lock */
3369                if (vma->vm_file && !page_mkwrite)
3370                        file_update_time(vma->vm_file);
3371        } else {
3372                unlock_page(vmf.page);
3373                if (anon)
3374                        page_cache_release(vmf.page);
3375        }
3376
3377        return ret;
3378
3379unwritable_page:
3380        page_cache_release(page);
3381        return ret;
3382uncharge_out:
3383        /* fs's fault handler get error */
3384        if (cow_page) {
3385                mem_cgroup_uncharge_page(cow_page);
3386                page_cache_release(cow_page);
3387        }
3388        return ret;
3389}
3390
3391static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3392                unsigned long address, pte_t *page_table, pmd_t *pmd,
3393                unsigned int flags, pte_t orig_pte)
3394{
3395        pgoff_t pgoff = (((address & PAGE_MASK)
3396                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3397
3398        pte_unmap(page_table);
3399        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3400}
3401
3402/*
3403 * Fault of a previously existing named mapping. Repopulate the pte
3404 * from the encoded file_pte if possible. This enables swappable
3405 * nonlinear vmas.
3406 *
3407 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3408 * but allow concurrent faults), and pte mapped but not yet locked.
3409 * We return with mmap_sem still held, but pte unmapped and unlocked.
3410 */
3411static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3412                unsigned long address, pte_t *page_table, pmd_t *pmd,
3413                unsigned int flags, pte_t orig_pte)
3414{
3415        pgoff_t pgoff;
3416
3417        flags |= FAULT_FLAG_NONLINEAR;
3418
3419        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3420                return 0;
3421
3422        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3423                /*
3424                 * Page table corrupted: show pte and kill process.
3425                 */
3426                print_bad_pte(vma, address, orig_pte, NULL);
3427                return VM_FAULT_SIGBUS;
3428        }
3429
3430        pgoff = pte_to_pgoff(orig_pte);
3431        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3432}
3433
3434/*
3435 * These routines also need to handle stuff like marking pages dirty
3436 * and/or accessed for architectures that don't do it in hardware (most
3437 * RISC architectures).  The early dirtying is also good on the i386.
3438 *
3439 * There is also a hook called "update_mmu_cache()" that architectures
3440 * with external mmu caches can use to update those (ie the Sparc or
3441 * PowerPC hashed page tables that act as extended TLBs).
3442 *
3443 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3444 * but allow concurrent faults), and pte mapped but not yet locked.
3445 * We return with mmap_sem still held, but pte unmapped and unlocked.
3446 */
3447int handle_pte_fault(struct mm_struct *mm,
3448                     struct vm_area_struct *vma, unsigned long address,
3449                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3450{
3451        pte_t entry;
3452        spinlock_t *ptl;
3453
3454        entry = *pte;
3455        if (!pte_present(entry)) {
3456                if (pte_none(entry)) {
3457                        if (vma->vm_ops) {
3458                                if (likely(vma->vm_ops->fault))
3459                                        return do_linear_fault(mm, vma, address,
3460                                                pte, pmd, flags, entry);
3461                        }
3462                        return do_anonymous_page(mm, vma, address,
3463                                                 pte, pmd, flags);
3464                }
3465                if (pte_file(entry))
3466                        return do_nonlinear_fault(mm, vma, address,
3467                                        pte, pmd, flags, entry);
3468                return do_swap_page(mm, vma, address,
3469                                        pte, pmd, flags, entry);
3470        }
3471
3472        ptl = pte_lockptr(mm, pmd);
3473        spin_lock(ptl);
3474        if (unlikely(!pte_same(*pte, entry)))
3475                goto unlock;
3476        if (flags & FAULT_FLAG_WRITE) {
3477                if (!pte_write(entry))
3478                        return do_wp_page(mm, vma, address,
3479                                        pte, pmd, ptl, entry);
3480                entry = pte_mkdirty(entry);
3481        }
3482        entry = pte_mkyoung(entry);
3483        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3484                update_mmu_cache(vma, address, pte);
3485        } else {
3486                /*
3487                 * This is needed only for protection faults but the arch code
3488                 * is not yet telling us if this is a protection fault or not.
3489                 * This still avoids useless tlb flushes for .text page faults
3490                 * with threads.
3491                 */
3492                if (flags & FAULT_FLAG_WRITE)
3493                        flush_tlb_fix_spurious_fault(vma, address);
3494        }
3495unlock:
3496        pte_unmap_unlock(pte, ptl);
3497        return 0;
3498}
3499
3500/*
3501 * By the time we get here, we already hold the mm semaphore
3502 */
3503int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3504                unsigned long address, unsigned int flags)
3505{
3506        pgd_t *pgd;
3507        pud_t *pud;
3508        pmd_t *pmd;
3509        pte_t *pte;
3510
3511        __set_current_state(TASK_RUNNING);
3512
3513        count_vm_event(PGFAULT);
3514        mem_cgroup_count_vm_event(mm, PGFAULT);
3515
3516        /* do counter updates before entering really critical section. */
3517        check_sync_rss_stat(current);
3518
3519        if (unlikely(is_vm_hugetlb_page(vma)))
3520                return hugetlb_fault(mm, vma, address, flags);
3521
3522retry:
3523        pgd = pgd_offset(mm, address);
3524        pud = pud_alloc(mm, pgd, address);
3525        if (!pud)
3526                return VM_FAULT_OOM;
3527        pmd = pmd_alloc(mm, pud, address);
3528        if (!pmd)
3529                return VM_FAULT_OOM;
3530        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3531                if (!vma->vm_ops)
3532                        return do_huge_pmd_anonymous_page(mm, vma, address,
3533                                                          pmd, flags);
3534        } else {
3535                pmd_t orig_pmd = *pmd;
3536                int ret;
3537
3538                barrier();
3539                if (pmd_trans_huge(orig_pmd)) {
3540                        if (flags & FAULT_FLAG_WRITE &&
3541                            !pmd_write(orig_pmd) &&
3542                            !pmd_trans_splitting(orig_pmd)) {
3543                                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3544                                                          orig_pmd);
3545                                /*
3546                                 * If COW results in an oom, the huge pmd will
3547                                 * have been split, so retry the fault on the
3548                                 * pte for a smaller charge.
3549                                 */
3550                                if (unlikely(ret & VM_FAULT_OOM))
3551                                        goto retry;
3552                                return ret;
3553                        }
3554                        return 0;
3555                }
3556        }
3557
3558        /*
3559         * Use __pte_alloc instead of pte_alloc_map, because we can't
3560         * run pte_offset_map on the pmd, if an huge pmd could
3561         * materialize from under us from a different thread.
3562         */
3563        if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3564                return VM_FAULT_OOM;
3565        /* if an huge pmd materialized from under us just retry later */
3566        if (unlikely(pmd_trans_huge(*pmd)))
3567                return 0;
3568        /*
3569         * A regular pmd is established and it can't morph into a huge pmd
3570         * from under us anymore at this point because we hold the mmap_sem
3571         * read mode and khugepaged takes it in write mode. So now it's
3572         * safe to run pte_offset_map().
3573         */
3574        pte = pte_offset_map(pmd, address);
3575
3576        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3577}
3578
3579#ifndef __PAGETABLE_PUD_FOLDED
3580/*
3581 * Allocate page upper directory.
3582 * We've already handled the fast-path in-line.
3583 */
3584int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3585{
3586        pud_t *new = pud_alloc_one(mm, address);
3587        if (!new)
3588                return -ENOMEM;
3589
3590        smp_wmb(); /* See comment in __pte_alloc */
3591
3592        spin_lock(&mm->page_table_lock);
3593        if (pgd_present(*pgd))          /* Another has populated it */
3594                pud_free(mm, new);
3595        else
3596                pgd_populate(mm, pgd, new);
3597        spin_unlock(&mm->page_table_lock);
3598        return 0;
3599}
3600#endif /* __PAGETABLE_PUD_FOLDED */
3601
3602#ifndef __PAGETABLE_PMD_FOLDED
3603/*
3604 * Allocate page middle directory.
3605 * We've already handled the fast-path in-line.
3606 */
3607int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3608{
3609        pmd_t *new = pmd_alloc_one(mm, address);
3610        if (!new)
3611                return -ENOMEM;
3612
3613        smp_wmb(); /* See comment in __pte_alloc */
3614
3615        spin_lock(&mm->page_table_lock);
3616#ifndef __ARCH_HAS_4LEVEL_HACK
3617        if (pud_present(*pud))          /* Another has populated it */
3618                pmd_free(mm, new);
3619        else
3620                pud_populate(mm, pud, new);
3621#else
3622        if (pgd_present(*pud))          /* Another has populated it */
3623                pmd_free(mm, new);
3624        else
3625                pgd_populate(mm, pud, new);
3626#endif /* __ARCH_HAS_4LEVEL_HACK */
3627        spin_unlock(&mm->page_table_lock);
3628        return 0;
3629}
3630#endif /* __PAGETABLE_PMD_FOLDED */
3631
3632int make_pages_present(unsigned long addr, unsigned long end)
3633{
3634        int ret, len, write;
3635        struct vm_area_struct * vma;
3636
3637        vma = find_vma(current->mm, addr);
3638        if (!vma)
3639                return -ENOMEM;
3640        /*
3641         * We want to touch writable mappings with a write fault in order
3642         * to break COW, except for shared mappings because these don't COW
3643         * and we would not want to dirty them for nothing.
3644         */
3645        write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3646        BUG_ON(addr >= end);
3647        BUG_ON(end > vma->vm_end);
3648        len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3649        ret = get_user_pages(current, current->mm, addr,
3650                        len, write, 0, NULL, NULL);
3651        if (ret < 0)
3652                return ret;
3653        return ret == len ? 0 : -EFAULT;
3654}
3655
3656#if !defined(__HAVE_ARCH_GATE_AREA)
3657
3658#if defined(AT_SYSINFO_EHDR)
3659static struct vm_area_struct gate_vma;
3660
3661static int __init gate_vma_init(void)
3662{
3663        gate_vma.vm_mm = NULL;
3664        gate_vma.vm_start = FIXADDR_USER_START;
3665        gate_vma.vm_end = FIXADDR_USER_END;
3666        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3667        gate_vma.vm_page_prot = __P101;
3668
3669        return 0;
3670}
3671__initcall(gate_vma_init);
3672#endif
3673
3674struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3675{
3676#ifdef AT_SYSINFO_EHDR
3677        return &gate_vma;
3678#else
3679        return NULL;
3680#endif
3681}
3682
3683int in_gate_area_no_mm(unsigned long addr)
3684{
3685#ifdef AT_SYSINFO_EHDR
3686        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3687                return 1;
3688#endif
3689        return 0;
3690}
3691
3692#endif  /* __HAVE_ARCH_GATE_AREA */
3693
3694static int __follow_pte(struct mm_struct *mm, unsigned long address,
3695                pte_t **ptepp, spinlock_t **ptlp)
3696{
3697        pgd_t *pgd;
3698        pud_t *pud;
3699        pmd_t *pmd;
3700        pte_t *ptep;
3701
3702        pgd = pgd_offset(mm, address);
3703        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3704                goto out;
3705
3706        pud = pud_offset(pgd, address);
3707        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3708                goto out;
3709
3710        pmd = pmd_offset(pud, address);
3711        VM_BUG_ON(pmd_trans_huge(*pmd));
3712        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3713                goto out;
3714
3715        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3716        if (pmd_huge(*pmd))
3717                goto out;
3718
3719        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3720        if (!ptep)
3721                goto out;
3722        if (!pte_present(*ptep))
3723                goto unlock;
3724        *ptepp = ptep;
3725        return 0;
3726unlock:
3727        pte_unmap_unlock(ptep, *ptlp);
3728out:
3729        return -EINVAL;
3730}
3731
3732static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3733                             pte_t **ptepp, spinlock_t **ptlp)
3734{
3735        int res;
3736
3737        /* (void) is needed to make gcc happy */
3738        (void) __cond_lock(*ptlp,
3739                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
3740        return res;
3741}
3742
3743/**
3744 * follow_pfn - look up PFN at a user virtual address
3745 * @vma: memory mapping
3746 * @address: user virtual address
3747 * @pfn: location to store found PFN
3748 *
3749 * Only IO mappings and raw PFN mappings are allowed.
3750 *
3751 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3752 */
3753int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3754        unsigned long *pfn)
3755{
3756        int ret = -EINVAL;
3757        spinlock_t *ptl;
3758        pte_t *ptep;
3759
3760        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3761                return ret;
3762
3763        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3764        if (ret)
3765                return ret;
3766        *pfn = pte_pfn(*ptep);
3767        pte_unmap_unlock(ptep, ptl);
3768        return 0;
3769}
3770EXPORT_SYMBOL(follow_pfn);
3771
3772#ifdef CONFIG_HAVE_IOREMAP_PROT
3773int follow_phys(struct vm_area_struct *vma,
3774                unsigned long address, unsigned int flags,
3775                unsigned long *prot, resource_size_t *phys)
3776{
3777        int ret = -EINVAL;
3778        pte_t *ptep, pte;
3779        spinlock_t *ptl;
3780
3781        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3782                goto out;
3783
3784        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3785                goto out;
3786        pte = *ptep;
3787
3788        if ((flags & FOLL_WRITE) && !pte_write(pte))
3789                goto unlock;
3790
3791        *prot = pgprot_val(pte_pgprot(pte));
3792        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3793
3794        ret = 0;
3795unlock:
3796        pte_unmap_unlock(ptep, ptl);
3797out:
3798        return ret;
3799}
3800
3801int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3802                        void *buf, int len, int write)
3803{
3804        resource_size_t phys_addr;
3805        unsigned long prot = 0;
3806        void __iomem *maddr;
3807        int offset = addr & (PAGE_SIZE-1);
3808
3809        if (follow_phys(vma, addr, write, &prot, &phys_addr))
3810                return -EINVAL;
3811
3812        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3813        if (write)
3814                memcpy_toio(maddr + offset, buf, len);
3815        else
3816                memcpy_fromio(buf, maddr + offset, len);
3817        iounmap(maddr);
3818
3819        return len;
3820}
3821#endif
3822
3823/*
3824 * Access another process' address space as given in mm.  If non-NULL, use the
3825 * given task for page fault accounting.
3826 */
3827static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3828                unsigned long addr, void *buf, int len, int write)
3829{
3830        struct vm_area_struct *vma;
3831        void *old_buf = buf;
3832
3833        down_read(&mm->mmap_sem);
3834        /* ignore errors, just check how much was successfully transferred */
3835        while (len) {
3836                int bytes, ret, offset;
3837                void *maddr;
3838                struct page *page = NULL;
3839
3840                ret = get_user_pages(tsk, mm, addr, 1,
3841                                write, 1, &page, &vma);
3842                if (ret <= 0) {
3843                        /*
3844                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
3845                         * we can access using slightly different code.
3846                         */
3847#ifdef CONFIG_HAVE_IOREMAP_PROT
3848                        vma = find_vma(mm, addr);
3849                        if (!vma || vma->vm_start > addr)
3850                                break;
3851                        if (vma->vm_ops && vma->vm_ops->access)
3852                                ret = vma->vm_ops->access(vma, addr, buf,
3853                                                          len, write);
3854                        if (ret <= 0)
3855#endif
3856                                break;
3857                        bytes = ret;
3858                } else {
3859                        bytes = len;
3860                        offset = addr & (PAGE_SIZE-1);
3861                        if (bytes > PAGE_SIZE-offset)
3862                                bytes = PAGE_SIZE-offset;
3863
3864                        maddr = kmap(page);
3865                        if (write) {
3866                                copy_to_user_page(vma, page, addr,
3867                                                  maddr + offset, buf, bytes);
3868                                set_page_dirty_lock(page);
3869                        } else {
3870                                copy_from_user_page(vma, page, addr,
3871                                                    buf, maddr + offset, bytes);
3872                        }
3873                        kunmap(page);
3874                        page_cache_release(page);
3875                }
3876                len -= bytes;
3877                buf += bytes;
3878                addr += bytes;
3879        }
3880        up_read(&mm->mmap_sem);
3881
3882        return buf - old_buf;
3883}
3884
3885/**
3886 * access_remote_vm - access another process' address space
3887 * @mm:         the mm_struct of the target address space
3888 * @addr:       start address to access
3889 * @buf:        source or destination buffer
3890 * @len:        number of bytes to transfer
3891 * @write:      whether the access is a write
3892 *
3893 * The caller must hold a reference on @mm.
3894 */
3895int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3896                void *buf, int len, int write)
3897{
3898        return __access_remote_vm(NULL, mm, addr, buf, len, write);
3899}
3900
3901/*
3902 * Access another process' address space.
3903 * Source/target buffer must be kernel space,
3904 * Do not walk the page table directly, use get_user_pages
3905 */
3906int access_process_vm(struct task_struct *tsk, unsigned long addr,
3907                void *buf, int len, int write)
3908{
3909        struct mm_struct *mm;
3910        int ret;
3911
3912        mm = get_task_mm(tsk);
3913        if (!mm)
3914                return 0;
3915
3916        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3917        mmput(mm);
3918
3919        return ret;
3920}
3921
3922/*
3923 * Print the name of a VMA.
3924 */
3925void print_vma_addr(char *prefix, unsigned long ip)
3926{
3927        struct mm_struct *mm = current->mm;
3928        struct vm_area_struct *vma;
3929
3930        /*
3931         * Do not print if we are in atomic
3932         * contexts (in exception stacks, etc.):
3933         */
3934        if (preempt_count())
3935                return;
3936
3937        down_read(&mm->mmap_sem);
3938        vma = find_vma(mm, ip);
3939        if (vma && vma->vm_file) {
3940                struct file *f = vma->vm_file;
3941                char *buf = (char *)__get_free_page(GFP_KERNEL);
3942                if (buf) {
3943                        char *p, *s;
3944
3945                        p = d_path(&f->f_path, buf, PAGE_SIZE);
3946                        if (IS_ERR(p))
3947                                p = "?";
3948                        s = strrchr(p, '/');
3949                        if (s)
3950                                p = s+1;
3951                        printk("%s%s[%lx+%lx]", prefix, p,
3952                                        vma->vm_start,
3953                                        vma->vm_end - vma->vm_start);
3954                        free_page((unsigned long)buf);
3955                }
3956        }
3957        up_read(&mm->mmap_sem);
3958}
3959
3960#ifdef CONFIG_PROVE_LOCKING
3961void might_fault(void)
3962{
3963        /*
3964         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3965         * holding the mmap_sem, this is safe because kernel memory doesn't
3966         * get paged out, therefore we'll never actually fault, and the
3967         * below annotations will generate false positives.
3968         */
3969        if (segment_eq(get_fs(), KERNEL_DS))
3970                return;
3971
3972        might_sleep();
3973        /*
3974         * it would be nicer only to annotate paths which are not under
3975         * pagefault_disable, however that requires a larger audit and
3976         * providing helpers like get_user_atomic.
3977         */
3978        if (!in_atomic() && current->mm)
3979                might_lock_read(&current->mm->mmap_sem);
3980}
3981EXPORT_SYMBOL(might_fault);
3982#endif
3983
3984#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3985static void clear_gigantic_page(struct page *page,
3986                                unsigned long addr,
3987                                unsigned int pages_per_huge_page)
3988{
3989        int i;
3990        struct page *p = page;
3991
3992        might_sleep();
3993        for (i = 0; i < pages_per_huge_page;
3994             i++, p = mem_map_next(p, page, i)) {
3995                cond_resched();
3996                clear_user_highpage(p, addr + i * PAGE_SIZE);
3997        }
3998}
3999void clear_huge_page(struct page *page,
4000                     unsigned long addr, unsigned int pages_per_huge_page)
4001{
4002        int i;
4003
4004        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4005                clear_gigantic_page(page, addr, pages_per_huge_page);
4006                return;
4007        }
4008
4009        might_sleep();
4010        for (i = 0; i < pages_per_huge_page; i++) {
4011                cond_resched();
4012                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4013        }
4014}
4015
4016static void copy_user_gigantic_page(struct page *dst, struct page *src,
4017                                    unsigned long addr,
4018                                    struct vm_area_struct *vma,
4019                                    unsigned int pages_per_huge_page)
4020{
4021        int i;
4022        struct page *dst_base = dst;
4023        struct page *src_base = src;
4024
4025        for (i = 0; i < pages_per_huge_page; ) {
4026                cond_resched();
4027                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4028
4029                i++;
4030                dst = mem_map_next(dst, dst_base, i);
4031                src = mem_map_next(src, src_base, i);
4032        }
4033}
4034
4035void copy_user_huge_page(struct page *dst, struct page *src,
4036                         unsigned long addr, struct vm_area_struct *vma,
4037                         unsigned int pages_per_huge_page)
4038{
4039        int i;
4040
4041        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4042                copy_user_gigantic_page(dst, src, addr, vma,
4043                                        pages_per_huge_page);
4044                return;
4045        }
4046
4047        might_sleep();
4048        for (i = 0; i < pages_per_huge_page; i++) {
4049                cond_resched();
4050                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4051        }
4052}
4053#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4054
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.