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 int 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        int ret;
1043
1044        /*
1045         * Don't copy ptes where a page fault will fill them correctly.
1046         * Fork becomes much lighter when there are big shared or private
1047         * readonly mappings. The tradeoff is that copy_page_range is more
1048         * efficient than faulting.
1049         */
1050        if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1051                if (!vma->anon_vma)
1052                        return 0;
1053        }
1054
1055        if (is_vm_hugetlb_page(vma))
1056                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1057
1058        if (unlikely(is_pfn_mapping(vma))) {
1059                /*
1060                 * We do not free on error cases below as remove_vma
1061                 * gets called on error from higher level routine
1062                 */
1063                ret = track_pfn_vma_copy(vma);
1064                if (ret)
1065                        return ret;
1066        }
1067
1068        /*
1069         * We need to invalidate the secondary MMU mappings only when
1070         * there could be a permission downgrade on the ptes of the
1071         * parent mm. And a permission downgrade will only happen if
1072         * is_cow_mapping() returns true.
1073         */
1074        if (is_cow_mapping(vma->vm_flags))
1075                mmu_notifier_invalidate_range_start(src_mm, addr, end);
1076
1077        ret = 0;
1078        dst_pgd = pgd_offset(dst_mm, addr);
1079        src_pgd = pgd_offset(src_mm, addr);
1080        do {
1081                next = pgd_addr_end(addr, end);
1082                if (pgd_none_or_clear_bad(src_pgd))
1083                        continue;
1084                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1085                                            vma, addr, next))) {
1086                        ret = -ENOMEM;
1087                        break;
1088                }
1089        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1090
1091        if (is_cow_mapping(vma->vm_flags))
1092                mmu_notifier_invalidate_range_end(src_mm,
1093                                                  vma->vm_start, end);
1094        return ret;
1095}
1096
1097static unsigned long zap_pte_range(struct mmu_gather *tlb,
1098                                struct vm_area_struct *vma, pmd_t *pmd,
1099                                unsigned long addr, unsigned long end,
1100                                struct zap_details *details)
1101{
1102        struct mm_struct *mm = tlb->mm;
1103        int force_flush = 0;
1104        int rss[NR_MM_COUNTERS];
1105        spinlock_t *ptl;
1106        pte_t *start_pte;
1107        pte_t *pte;
1108
1109again:
1110        init_rss_vec(rss);
1111        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1112        pte = start_pte;
1113        arch_enter_lazy_mmu_mode();
1114        do {
1115                pte_t ptent = *pte;
1116                if (pte_none(ptent)) {
1117                        continue;
1118                }
1119
1120                if (pte_present(ptent)) {
1121                        struct page *page;
1122
1123                        page = vm_normal_page(vma, addr, ptent);
1124                        if (unlikely(details) && page) {
1125                                /*
1126                                 * unmap_shared_mapping_pages() wants to
1127                                 * invalidate cache without truncating:
1128                                 * unmap shared but keep private pages.
1129                                 */
1130                                if (details->check_mapping &&
1131                                    details->check_mapping != page->mapping)
1132                                        continue;
1133                                /*
1134                                 * Each page->index must be checked when
1135                                 * invalidating or truncating nonlinear.
1136                                 */
1137                                if (details->nonlinear_vma &&
1138                                    (page->index < details->first_index ||
1139                                     page->index > details->last_index))
1140                                        continue;
1141                        }
1142                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1143                                                        tlb->fullmm);
1144                        tlb_remove_tlb_entry(tlb, pte, addr);
1145                        if (unlikely(!page))
1146                                continue;
1147                        if (unlikely(details) && details->nonlinear_vma
1148                            && linear_page_index(details->nonlinear_vma,
1149                                                addr) != page->index)
1150                                set_pte_at(mm, addr, pte,
1151                                           pgoff_to_pte(page->index));
1152                        if (PageAnon(page))
1153                                rss[MM_ANONPAGES]--;
1154                        else {
1155                                if (pte_dirty(ptent))
1156                                        set_page_dirty(page);
1157                                if (pte_young(ptent) &&
1158                                    likely(!VM_SequentialReadHint(vma)))
1159                                        mark_page_accessed(page);
1160                                rss[MM_FILEPAGES]--;
1161                        }
1162                        page_remove_rmap(page);
1163                        if (unlikely(page_mapcount(page) < 0))
1164                                print_bad_pte(vma, addr, ptent, page);
1165                        force_flush = !__tlb_remove_page(tlb, page);
1166                        if (force_flush)
1167                                break;
1168                        continue;
1169                }
1170                /*
1171                 * If details->check_mapping, we leave swap entries;
1172                 * if details->nonlinear_vma, we leave file entries.
1173                 */
1174                if (unlikely(details))
1175                        continue;
1176                if (pte_file(ptent)) {
1177                        if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1178                                print_bad_pte(vma, addr, ptent, NULL);
1179                } else {
1180                        swp_entry_t entry = pte_to_swp_entry(ptent);
1181
1182                        if (!non_swap_entry(entry))
1183                                rss[MM_SWAPENTS]--;
1184                        else if (is_migration_entry(entry)) {
1185                                struct page *page;
1186
1187                                page = migration_entry_to_page(entry);
1188
1189                                if (PageAnon(page))
1190                                        rss[MM_ANONPAGES]--;
1191                                else
1192                                        rss[MM_FILEPAGES]--;
1193                        }
1194                        if (unlikely(!free_swap_and_cache(entry)))
1195                                print_bad_pte(vma, addr, ptent, NULL);
1196                }
1197                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198        } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200        add_mm_rss_vec(mm, rss);
1201        arch_leave_lazy_mmu_mode();
1202        pte_unmap_unlock(start_pte, ptl);
1203
1204        /*
1205         * mmu_gather ran out of room to batch pages, we break out of
1206         * the PTE lock to avoid doing the potential expensive TLB invalidate
1207         * and page-free while holding it.
1208         */
1209        if (force_flush) {
1210                force_flush = 0;
1211
1212#ifdef HAVE_GENERIC_MMU_GATHER
1213                tlb->start = addr;
1214                tlb->end = end;
1215#endif
1216                tlb_flush_mmu(tlb);
1217                if (addr != end)
1218                        goto again;
1219        }
1220
1221        return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225                                struct vm_area_struct *vma, pud_t *pud,
1226                                unsigned long addr, unsigned long end,
1227                                struct zap_details *details)
1228{
1229        pmd_t *pmd;
1230        unsigned long next;
1231
1232        pmd = pmd_offset(pud, addr);
1233        do {
1234                next = pmd_addr_end(addr, end);
1235                if (pmd_trans_huge(*pmd)) {
1236                        if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238                                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239                                        pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240                                                __func__, addr, end,
1241                                                vma->vm_start,
1242                                                vma->vm_end);
1243                                        BUG();
1244                                }
1245#endif
1246                                split_huge_page_pmd(vma->vm_mm, pmd);
1247                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248                                goto next;
1249                        /* fall through */
1250                }
1251                /*
1252                 * Here there can be other concurrent MADV_DONTNEED or
1253                 * trans huge page faults running, and if the pmd is
1254                 * none or trans huge it can change under us. This is
1255                 * because MADV_DONTNEED holds the mmap_sem in read
1256                 * mode.
1257                 */
1258                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259                        goto next;
1260                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262                cond_resched();
1263        } while (pmd++, addr = next, addr != end);
1264
1265        return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269                                struct vm_area_struct *vma, pgd_t *pgd,
1270                                unsigned long addr, unsigned long end,
1271                                struct zap_details *details)
1272{
1273        pud_t *pud;
1274        unsigned long next;
1275
1276        pud = pud_offset(pgd, addr);
1277        do {
1278                next = pud_addr_end(addr, end);
1279                if (pud_none_or_clear_bad(pud))
1280                        continue;
1281                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282        } while (pud++, addr = next, addr != end);
1283
1284        return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288                             struct vm_area_struct *vma,
1289                             unsigned long addr, unsigned long end,
1290                             struct zap_details *details)
1291{
1292        pgd_t *pgd;
1293        unsigned long next;
1294
1295        if (details && !details->check_mapping && !details->nonlinear_vma)
1296                details = NULL;
1297
1298        BUG_ON(addr >= end);
1299        mem_cgroup_uncharge_start();
1300        tlb_start_vma(tlb, vma);
1301        pgd = pgd_offset(vma->vm_mm, addr);
1302        do {
1303                next = pgd_addr_end(addr, end);
1304                if (pgd_none_or_clear_bad(pgd))
1305                        continue;
1306                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307        } while (pgd++, addr = next, addr != end);
1308        tlb_end_vma(tlb, vma);
1309        mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314                struct vm_area_struct *vma, unsigned long start_addr,
1315                unsigned long end_addr,
1316                struct zap_details *details)
1317{
1318        unsigned long start = max(vma->vm_start, start_addr);
1319        unsigned long end;
1320
1321        if (start >= vma->vm_end)
1322                return;
1323        end = min(vma->vm_end, end_addr);
1324        if (end <= vma->vm_start)
1325                return;
1326
1327        if (vma->vm_file)
1328                uprobe_munmap(vma, start, end);
1329
1330        if (unlikely(is_pfn_mapping(vma)))
1331                untrack_pfn_vma(vma, 0, 0);
1332
1333        if (start != end) {
1334                if (unlikely(is_vm_hugetlb_page(vma))) {
1335                        /*
1336                         * It is undesirable to test vma->vm_file as it
1337                         * should be non-null for valid hugetlb area.
1338                         * However, vm_file will be NULL in the error
1339                         * cleanup path of do_mmap_pgoff. When
1340                         * hugetlbfs ->mmap method fails,
1341                         * do_mmap_pgoff() nullifies vma->vm_file
1342                         * before calling this function to clean up.
1343                         * Since no pte has actually been setup, it is
1344                         * safe to do nothing in this case.
1345                         */
1346                        if (vma->vm_file) {
1347                                mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348                                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349                                mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350                        }
1351                } else
1352                        unmap_page_range(tlb, vma, start, end, details);
1353        }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns.  So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375                struct vm_area_struct *vma, unsigned long start_addr,
1376                unsigned long end_addr)
1377{
1378        struct mm_struct *mm = vma->vm_mm;
1379
1380        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396                unsigned long size, struct zap_details *details)
1397{
1398        struct mm_struct *mm = vma->vm_mm;
1399        struct mmu_gather tlb;
1400        unsigned long end = start + size;
1401
1402        lru_add_drain();
1403        tlb_gather_mmu(&tlb, mm, 0);
1404        update_hiwater_rss(mm);
1405        mmu_notifier_invalidate_range_start(mm, start, end);
1406        for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407                unmap_single_vma(&tlb, vma, start, end, details);
1408        mmu_notifier_invalidate_range_end(mm, start, end);
1409        tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422                unsigned long size, struct zap_details *details)
1423{
1424        struct mm_struct *mm = vma->vm_mm;
1425        struct mmu_gather tlb;
1426        unsigned long end = address + size;
1427
1428        lru_add_drain();
1429        tlb_gather_mmu(&tlb, mm, 0);
1430        update_hiwater_rss(mm);
1431        mmu_notifier_invalidate_range_start(mm, address, end);
1432        unmap_single_vma(&tlb, vma, address, end, details);
1433        mmu_notifier_invalidate_range_end(mm, address, end);
1434        tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450                unsigned long size)
1451{
1452        if (address < vma->vm_start || address + size > vma->vm_end ||
1453                        !(vma->vm_flags & VM_PFNMAP))
1454                return -1;
1455        zap_page_range_single(vma, address, size, NULL);
1456        return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 *
1466 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1467 *
1468 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1469 * an error pointer if there is a mapping to something not represented
1470 * by a page descriptor (see also vm_normal_page()).
1471 */
1472struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1473                        unsigned int flags)
1474{
1475        pgd_t *pgd;
1476        pud_t *pud;
1477        pmd_t *pmd;
1478        pte_t *ptep, pte;
1479        spinlock_t *ptl;
1480        struct page *page;
1481        struct mm_struct *mm = vma->vm_mm;
1482
1483        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1484        if (!IS_ERR(page)) {
1485                BUG_ON(flags & FOLL_GET);
1486                goto out;
1487        }
1488
1489        page = NULL;
1490        pgd = pgd_offset(mm, address);
1491        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1492                goto no_page_table;
1493
1494        pud = pud_offset(pgd, address);
1495        if (pud_none(*pud))
1496                goto no_page_table;
1497        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1498                BUG_ON(flags & FOLL_GET);
1499                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1500                goto out;
1501        }
1502        if (unlikely(pud_bad(*pud)))
1503                goto no_page_table;
1504
1505        pmd = pmd_offset(pud, address);
1506        if (pmd_none(*pmd))
1507                goto no_page_table;
1508        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1509                BUG_ON(flags & FOLL_GET);
1510                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1511                goto out;
1512        }
1513        if (pmd_trans_huge(*pmd)) {
1514                if (flags & FOLL_SPLIT) {
1515                        split_huge_page_pmd(mm, pmd);
1516                        goto split_fallthrough;
1517                }
1518                spin_lock(&mm->page_table_lock);
1519                if (likely(pmd_trans_huge(*pmd))) {
1520                        if (unlikely(pmd_trans_splitting(*pmd))) {
1521                                spin_unlock(&mm->page_table_lock);
1522                                wait_split_huge_page(vma->anon_vma, pmd);
1523                        } else {
1524                                page = follow_trans_huge_pmd(mm, address,
1525                                                             pmd, flags);
1526                                spin_unlock(&mm->page_table_lock);
1527                                goto out;
1528                        }
1529                } else
1530                        spin_unlock(&mm->page_table_lock);
1531                /* fall through */
1532        }
1533split_fallthrough:
1534        if (unlikely(pmd_bad(*pmd)))
1535                goto no_page_table;
1536
1537        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1538
1539        pte = *ptep;
1540        if (!pte_present(pte))
1541                goto no_page;
1542        if ((flags & FOLL_WRITE) && !pte_write(pte))
1543                goto unlock;
1544
1545        page = vm_normal_page(vma, address, pte);
1546        if (unlikely(!page)) {
1547                if ((flags & FOLL_DUMP) ||
1548                    !is_zero_pfn(pte_pfn(pte)))
1549                        goto bad_page;
1550                page = pte_page(pte);
1551        }
1552
1553        if (flags & FOLL_GET)
1554                get_page_foll(page);
1555        if (flags & FOLL_TOUCH) {
1556                if ((flags & FOLL_WRITE) &&
1557                    !pte_dirty(pte) && !PageDirty(page))
1558                        set_page_dirty(page);
1559                /*
1560                 * pte_mkyoung() would be more correct here, but atomic care
1561                 * is needed to avoid losing the dirty bit: it is easier to use
1562                 * mark_page_accessed().
1563                 */
1564                mark_page_accessed(page);
1565        }
1566        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1567                /*
1568                 * The preliminary mapping check is mainly to avoid the
1569                 * pointless overhead of lock_page on the ZERO_PAGE
1570                 * which might bounce very badly if there is contention.
1571                 *
1572                 * If the page is already locked, we don't need to
1573                 * handle it now - vmscan will handle it later if and
1574                 * when it attempts to reclaim the page.
1575                 */
1576                if (page->mapping && trylock_page(page)) {
1577                        lru_add_drain();  /* push cached pages to LRU */
1578                        /*
1579                         * Because we lock page here and migration is
1580                         * blocked by the pte's page reference, we need
1581                         * only check for file-cache page truncation.
1582                         */
1583                        if (page->mapping)
1584                                mlock_vma_page(page);
1585                        unlock_page(page);
1586                }
1587        }
1588unlock:
1589        pte_unmap_unlock(ptep, ptl);
1590out:
1591        return page;
1592
1593bad_page:
1594        pte_unmap_unlock(ptep, ptl);
1595        return ERR_PTR(-EFAULT);
1596
1597no_page:
1598        pte_unmap_unlock(ptep, ptl);
1599        if (!pte_none(pte))
1600                return page;
1601
1602no_page_table:
1603        /*
1604         * When core dumping an enormous anonymous area that nobody
1605         * has touched so far, we don't want to allocate unnecessary pages or
1606         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1607         * then get_dump_page() will return NULL to leave a hole in the dump.
1608         * But we can only make this optimization where a hole would surely
1609         * be zero-filled if handle_mm_fault() actually did handle it.
1610         */
1611        if ((flags & FOLL_DUMP) &&
1612            (!vma->vm_ops || !vma->vm_ops->fault))
1613                return ERR_PTR(-EFAULT);
1614        return page;
1615}
1616
1617static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1618{
1619        return stack_guard_page_start(vma, addr) ||
1620               stack_guard_page_end(vma, addr+PAGE_SIZE);
1621}
1622
1623/**
1624 * __get_user_pages() - pin user pages in memory
1625 * @tsk:        task_struct of target task
1626 * @mm:         mm_struct of target mm
1627 * @start:      starting user address
1628 * @nr_pages:   number of pages from start to pin
1629 * @gup_flags:  flags modifying pin behaviour
1630 * @pages:      array that receives pointers to the pages pinned.
1631 *              Should be at least nr_pages long. Or NULL, if caller
1632 *              only intends to ensure the pages are faulted in.
1633 * @vmas:       array of pointers to vmas corresponding to each page.
1634 *              Or NULL if the caller does not require them.
1635 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1636 *
1637 * Returns number of pages pinned. This may be fewer than the number
1638 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1639 * were pinned, returns -errno. Each page returned must be released
1640 * with a put_page() call when it is finished with. vmas will only
1641 * remain valid while mmap_sem is held.
1642 *
1643 * Must be called with mmap_sem held for read or write.
1644 *
1645 * __get_user_pages walks a process's page tables and takes a reference to
1646 * each struct page that each user address corresponds to at a given
1647 * instant. That is, it takes the page that would be accessed if a user
1648 * thread accesses the given user virtual address at that instant.
1649 *
1650 * This does not guarantee that the page exists in the user mappings when
1651 * __get_user_pages returns, and there may even be a completely different
1652 * page there in some cases (eg. if mmapped pagecache has been invalidated
1653 * and subsequently re faulted). However it does guarantee that the page
1654 * won't be freed completely. And mostly callers simply care that the page
1655 * contains data that was valid *at some point in time*. Typically, an IO
1656 * or similar operation cannot guarantee anything stronger anyway because
1657 * locks can't be held over the syscall boundary.
1658 *
1659 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1660 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1661 * appropriate) must be called after the page is finished with, and
1662 * before put_page is called.
1663 *
1664 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1665 * or mmap_sem contention, and if waiting is needed to pin all pages,
1666 * *@nonblocking will be set to 0.
1667 *
1668 * In most cases, get_user_pages or get_user_pages_fast should be used
1669 * instead of __get_user_pages. __get_user_pages should be used only if
1670 * you need some special @gup_flags.
1671 */
1672int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1673                     unsigned long start, int nr_pages, unsigned int gup_flags,
1674                     struct page **pages, struct vm_area_struct **vmas,
1675                     int *nonblocking)
1676{
1677        int i;
1678        unsigned long vm_flags;
1679
1680        if (nr_pages <= 0)
1681                return 0;
1682
1683        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1684
1685        /* 
1686         * Require read or write permissions.
1687         * If FOLL_FORCE is set, we only require the "MAY" flags.
1688         */
1689        vm_flags  = (gup_flags & FOLL_WRITE) ?
1690                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1691        vm_flags &= (gup_flags & FOLL_FORCE) ?
1692                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1693        i = 0;
1694
1695        do {
1696                struct vm_area_struct *vma;
1697
1698                vma = find_extend_vma(mm, start);
1699                if (!vma && in_gate_area(mm, start)) {
1700                        unsigned long pg = start & PAGE_MASK;
1701                        pgd_t *pgd;
1702                        pud_t *pud;
1703                        pmd_t *pmd;
1704                        pte_t *pte;
1705
1706                        /* user gate pages are read-only */
1707                        if (gup_flags & FOLL_WRITE)
1708                                return i ? : -EFAULT;
1709                        if (pg > TASK_SIZE)
1710                                pgd = pgd_offset_k(pg);
1711                        else
1712                                pgd = pgd_offset_gate(mm, pg);
1713                        BUG_ON(pgd_none(*pgd));
1714                        pud = pud_offset(pgd, pg);
1715                        BUG_ON(pud_none(*pud));
1716                        pmd = pmd_offset(pud, pg);
1717                        if (pmd_none(*pmd))
1718                                return i ? : -EFAULT;
1719                        VM_BUG_ON(pmd_trans_huge(*pmd));
1720                        pte = pte_offset_map(pmd, pg);
1721                        if (pte_none(*pte)) {
1722                                pte_unmap(pte);
1723                                return i ? : -EFAULT;
1724                        }
1725                        vma = get_gate_vma(mm);
1726                        if (pages) {
1727                                struct page *page;
1728
1729                                page = vm_normal_page(vma, start, *pte);
1730                                if (!page) {
1731                                        if (!(gup_flags & FOLL_DUMP) &&
1732                                             is_zero_pfn(pte_pfn(*pte)))
1733                                                page = pte_page(*pte);
1734                                        else {
1735                                                pte_unmap(pte);
1736                                                return i ? : -EFAULT;
1737                                        }
1738                                }
1739                                pages[i] = page;
1740                                get_page(page);
1741                        }
1742                        pte_unmap(pte);
1743                        goto next_page;
1744                }
1745
1746                if (!vma ||
1747                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1748                    !(vm_flags & vma->vm_flags))
1749                        return i ? : -EFAULT;
1750
1751                if (is_vm_hugetlb_page(vma)) {
1752                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1753                                        &start, &nr_pages, i, gup_flags);
1754                        continue;
1755                }
1756
1757                do {
1758                        struct page *page;
1759                        unsigned int foll_flags = gup_flags;
1760
1761                        /*
1762                         * If we have a pending SIGKILL, don't keep faulting
1763                         * pages and potentially allocating memory.
1764                         */
1765                        if (unlikely(fatal_signal_pending(current)))
1766                                return i ? i : -ERESTARTSYS;
1767
1768                        cond_resched();
1769                        while (!(page = follow_page(vma, start, foll_flags))) {
1770                                int ret;
1771                                unsigned int fault_flags = 0;
1772
1773                                /* For mlock, just skip the stack guard page. */
1774                                if (foll_flags & FOLL_MLOCK) {
1775                                        if (stack_guard_page(vma, start))
1776                                                goto next_page;
1777                                }
1778                                if (foll_flags & FOLL_WRITE)
1779                                        fault_flags |= FAULT_FLAG_WRITE;
1780                                if (nonblocking)
1781                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1782                                if (foll_flags & FOLL_NOWAIT)
1783                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1784
1785                                ret = handle_mm_fault(mm, vma, start,
1786                                                        fault_flags);
1787
1788                                if (ret & VM_FAULT_ERROR) {
1789                                        if (ret & VM_FAULT_OOM)
1790                                                return i ? i : -ENOMEM;
1791                                        if (ret & (VM_FAULT_HWPOISON |
1792                                                   VM_FAULT_HWPOISON_LARGE)) {
1793                                                if (i)
1794                                                        return i;
1795                                                else if (gup_flags & FOLL_HWPOISON)
1796                                                        return -EHWPOISON;
1797                                                else
1798                                                        return -EFAULT;
1799                                        }
1800                                        if (ret & VM_FAULT_SIGBUS)
1801                                                return i ? i : -EFAULT;
1802                                        BUG();
1803                                }
1804
1805                                if (tsk) {
1806                                        if (ret & VM_FAULT_MAJOR)
1807                                                tsk->maj_flt++;
1808                                        else
1809                                                tsk->min_flt++;
1810                                }
1811
1812                                if (ret & VM_FAULT_RETRY) {
1813                                        if (nonblocking)
1814                                                *nonblocking = 0;
1815                                        return i;
1816                                }
1817
1818                                /*
1819                                 * The VM_FAULT_WRITE bit tells us that
1820                                 * do_wp_page has broken COW when necessary,
1821                                 * even if maybe_mkwrite decided not to set
1822                                 * pte_write. We can thus safely do subsequent
1823                                 * page lookups as if they were reads. But only
1824                                 * do so when looping for pte_write is futile:
1825                                 * in some cases userspace may also be wanting
1826                                 * to write to the gotten user page, which a
1827                                 * read fault here might prevent (a readonly
1828                                 * page might get reCOWed by userspace write).
1829                                 */
1830                                if ((ret & VM_FAULT_WRITE) &&
1831                                    !(vma->vm_flags & VM_WRITE))
1832                                        foll_flags &= ~FOLL_WRITE;
1833
1834                                cond_resched();
1835                        }
1836                        if (IS_ERR(page))
1837                                return i ? i : PTR_ERR(page);
1838                        if (pages) {
1839                                pages[i] = page;
1840
1841                                flush_anon_page(vma, page, start);
1842                                flush_dcache_page(page);
1843                        }
1844next_page:
1845                        if (vmas)
1846                                vmas[i] = vma;
1847                        i++;
1848                        start += PAGE_SIZE;
1849                        nr_pages--;
1850                } while (nr_pages && start < vma->vm_end);
1851        } while (nr_pages);
1852        return i;
1853}
1854EXPORT_SYMBOL(__get_user_pages);
1855
1856/*
1857 * fixup_user_fault() - manually resolve a user page fault
1858 * @tsk:        the task_struct to use for page fault accounting, or
1859 *              NULL if faults are not to be recorded.
1860 * @mm:         mm_struct of target mm
1861 * @address:    user address
1862 * @fault_flags:flags to pass down to handle_mm_fault()
1863 *
1864 * This is meant to be called in the specific scenario where for locking reasons
1865 * we try to access user memory in atomic context (within a pagefault_disable()
1866 * section), this returns -EFAULT, and we want to resolve the user fault before
1867 * trying again.
1868 *
1869 * Typically this is meant to be used by the futex code.
1870 *
1871 * The main difference with get_user_pages() is that this function will
1872 * unconditionally call handle_mm_fault() which will in turn perform all the
1873 * necessary SW fixup of the dirty and young bits in the PTE, while
1874 * handle_mm_fault() only guarantees to update these in the struct page.
1875 *
1876 * This is important for some architectures where those bits also gate the
1877 * access permission to the page because they are maintained in software.  On
1878 * such architectures, gup() will not be enough to make a subsequent access
1879 * succeed.
1880 *
1881 * This should be called with the mm_sem held for read.
1882 */
1883int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1884                     unsigned long address, unsigned int fault_flags)
1885{
1886        struct vm_area_struct *vma;
1887        int ret;
1888
1889        vma = find_extend_vma(mm, address);
1890        if (!vma || address < vma->vm_start)
1891                return -EFAULT;
1892
1893        ret = handle_mm_fault(mm, vma, address, fault_flags);
1894        if (ret & VM_FAULT_ERROR) {
1895                if (ret & VM_FAULT_OOM)
1896                        return -ENOMEM;
1897                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1898                        return -EHWPOISON;
1899                if (ret & VM_FAULT_SIGBUS)
1900                        return -EFAULT;
1901                BUG();
1902        }
1903        if (tsk) {
1904                if (ret & VM_FAULT_MAJOR)
1905                        tsk->maj_flt++;
1906                else
1907                        tsk->min_flt++;
1908        }
1909        return 0;
1910}
1911
1912/*
1913 * get_user_pages() - pin user pages in memory
1914 * @tsk:        the task_struct to use for page fault accounting, or
1915 *              NULL if faults are not to be recorded.
1916 * @mm:         mm_struct of target mm
1917 * @start:      starting user address
1918 * @nr_pages:   number of pages from start to pin
1919 * @write:      whether pages will be written to by the caller
1920 * @force:      whether to force write access even if user mapping is
1921 *              readonly. This will result in the page being COWed even
1922 *              in MAP_SHARED mappings. You do not want this.
1923 * @pages:      array that receives pointers to the pages pinned.
1924 *              Should be at least nr_pages long. Or NULL, if caller
1925 *              only intends to ensure the pages are faulted in.
1926 * @vmas:       array of pointers to vmas corresponding to each page.
1927 *              Or NULL if the caller does not require them.
1928 *
1929 * Returns number of pages pinned. This may be fewer than the number
1930 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1931 * were pinned, returns -errno. Each page returned must be released
1932 * with a put_page() call when it is finished with. vmas will only
1933 * remain valid while mmap_sem is held.
1934 *
1935 * Must be called with mmap_sem held for read or write.
1936 *
1937 * get_user_pages walks a process's page tables and takes a reference to
1938 * each struct page that each user address corresponds to at a given
1939 * instant. That is, it takes the page that would be accessed if a user
1940 * thread accesses the given user virtual address at that instant.
1941 *
1942 * This does not guarantee that the page exists in the user mappings when
1943 * get_user_pages returns, and there may even be a completely different
1944 * page there in some cases (eg. if mmapped pagecache has been invalidated
1945 * and subsequently re faulted). However it does guarantee that the page
1946 * won't be freed completely. And mostly callers simply care that the page
1947 * contains data that was valid *at some point in time*. Typically, an IO
1948 * or similar operation cannot guarantee anything stronger anyway because
1949 * locks can't be held over the syscall boundary.
1950 *
1951 * If write=0, the page must not be written to. If the page is written to,
1952 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1953 * after the page is finished with, and before put_page is called.
1954 *
1955 * get_user_pages is typically used for fewer-copy IO operations, to get a
1956 * handle on the memory by some means other than accesses via the user virtual
1957 * addresses. The pages may be submitted for DMA to devices or accessed via
1958 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1959 * use the correct cache flushing APIs.
1960 *
1961 * See also get_user_pages_fast, for performance critical applications.
1962 */
1963int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1964                unsigned long start, int nr_pages, int write, int force,
1965                struct page **pages, struct vm_area_struct **vmas)
1966{
1967        int flags = FOLL_TOUCH;
1968
1969        if (pages)
1970                flags |= FOLL_GET;
1971        if (write)
1972                flags |= FOLL_WRITE;
1973        if (force)
1974                flags |= FOLL_FORCE;
1975
1976        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1977                                NULL);
1978}
1979EXPORT_SYMBOL(get_user_pages);
1980
1981/**
1982 * get_dump_page() - pin user page in memory while writing it to core dump
1983 * @addr: user address
1984 *
1985 * Returns struct page pointer of user page pinned for dump,
1986 * to be freed afterwards by page_cache_release() or put_page().
1987 *
1988 * Returns NULL on any kind of failure - a hole must then be inserted into
1989 * the corefile, to preserve alignment with its headers; and also returns
1990 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1991 * allowing a hole to be left in the corefile to save diskspace.
1992 *
1993 * Called without mmap_sem, but after all other threads have been killed.
1994 */
1995#ifdef CONFIG_ELF_CORE
1996struct page *get_dump_page(unsigned long addr)
1997{
1998        struct vm_area_struct *vma;
1999        struct page *page;
2000
2001        if (__get_user_pages(current, current->mm, addr, 1,
2002                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2003                             NULL) < 1)
2004                return NULL;
2005        flush_cache_page(vma, addr, page_to_pfn(page));
2006        return page;
2007}
2008#endif /* CONFIG_ELF_CORE */
2009
2010pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2011                        spinlock_t **ptl)
2012{
2013        pgd_t * pgd = pgd_offset(mm, addr);
2014        pud_t * pud = pud_alloc(mm, pgd, addr);
2015        if (pud) {
2016                pmd_t * pmd = pmd_alloc(mm, pud, addr);
2017                if (pmd) {
2018                        VM_BUG_ON(pmd_trans_huge(*pmd));
2019                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
2020                }
2021        }
2022        return NULL;
2023}
2024
2025/*
2026 * This is the old fallback for page remapping.
2027 *
2028 * For historical reasons, it only allows reserved pages. Only
2029 * old drivers should use this, and they needed to mark their
2030 * pages reserved for the old functions anyway.
2031 */
2032static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2033                        struct page *page, pgprot_t prot)
2034{
2035        struct mm_struct *mm = vma->vm_mm;
2036        int retval;
2037        pte_t *pte;
2038        spinlock_t *ptl;
2039
2040        retval = -EINVAL;
2041        if (PageAnon(page))
2042                goto out;
2043        retval = -ENOMEM;
2044        flush_dcache_page(page);
2045        pte = get_locked_pte(mm, addr, &ptl);
2046        if (!pte)
2047                goto out;
2048        retval = -EBUSY;
2049        if (!pte_none(*pte))
2050                goto out_unlock;
2051
2052        /* Ok, finally just insert the thing.. */
2053        get_page(page);
2054        inc_mm_counter_fast(mm, MM_FILEPAGES);
2055        page_add_file_rmap(page);
2056        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2057
2058        retval = 0;
2059        pte_unmap_unlock(pte, ptl);
2060        return retval;
2061out_unlock:
2062        pte_unmap_unlock(pte, ptl);
2063out:
2064        return retval;
2065}
2066
2067/**
2068 * vm_insert_page - insert single page into user vma
2069 * @vma: user vma to map to
2070 * @addr: target user address of this page
2071 * @page: source kernel page
2072 *
2073 * This allows drivers to insert individual pages they've allocated
2074 * into a user vma.
2075 *
2076 * The page has to be a nice clean _individual_ kernel allocation.
2077 * If you allocate a compound page, you need to have marked it as
2078 * such (__GFP_COMP), or manually just split the page up yourself
2079 * (see split_page()).
2080 *
2081 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2082 * took an arbitrary page protection parameter. This doesn't allow
2083 * that. Your vma protection will have to be set up correctly, which
2084 * means that if you want a shared writable mapping, you'd better
2085 * ask for a shared writable mapping!
2086 *
2087 * The page does not need to be reserved.
2088 */
2089int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2090                        struct page *page)
2091{
2092        if (addr < vma->vm_start || addr >= vma->vm_end)
2093                return -EFAULT;
2094        if (!page_count(page))
2095                return -EINVAL;
2096        vma->vm_flags |= VM_INSERTPAGE;
2097        return insert_page(vma, addr, page, vma->vm_page_prot);
2098}
2099EXPORT_SYMBOL(vm_insert_page);
2100
2101static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2102                        unsigned long pfn, pgprot_t prot)
2103{
2104        struct mm_struct *mm = vma->vm_mm;
2105        int retval;
2106        pte_t *pte, entry;
2107        spinlock_t *ptl;
2108
2109        retval = -ENOMEM;
2110        pte = get_locked_pte(mm, addr, &ptl);
2111        if (!pte)
2112                goto out;
2113        retval = -EBUSY;
2114        if (!pte_none(*pte))
2115                goto out_unlock;
2116
2117        /* Ok, finally just insert the thing.. */
2118        entry = pte_mkspecial(pfn_pte(pfn, prot));
2119        set_pte_at(mm, addr, pte, entry);
2120        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2121
2122        retval = 0;
2123out_unlock:
2124        pte_unmap_unlock(pte, ptl);
2125out:
2126        return retval;
2127}
2128
2129/**
2130 * vm_insert_pfn - insert single pfn into user vma
2131 * @vma: user vma to map to
2132 * @addr: target user address of this page
2133 * @pfn: source kernel pfn
2134 *
2135 * Similar to vm_inert_page, this allows drivers to insert individual pages
2136 * they've allocated into a user vma. Same comments apply.
2137 *
2138 * This function should only be called from a vm_ops->fault handler, and
2139 * in that case the handler should return NULL.
2140 *
2141 * vma cannot be a COW mapping.
2142 *
2143 * As this is called only for pages that do not currently exist, we
2144 * do not need to flush old virtual caches or the TLB.
2145 */
2146int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2147                        unsigned long pfn)
2148{
2149        int ret;
2150        pgprot_t pgprot = vma->vm_page_prot;
2151        /*
2152         * Technically, architectures with pte_special can avoid all these
2153         * restrictions (same for remap_pfn_range).  However we would like
2154         * consistency in testing and feature parity among all, so we should
2155         * try to keep these invariants in place for everybody.
2156         */
2157        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2158        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2159                                                (VM_PFNMAP|VM_MIXEDMAP));
2160        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2161        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2162
2163        if (addr < vma->vm_start || addr >= vma->vm_end)
2164                return -EFAULT;
2165        if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2166                return -EINVAL;
2167
2168        ret = insert_pfn(vma, addr, pfn, pgprot);
2169
2170        if (ret)
2171                untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2172
2173        return ret;
2174}
2175EXPORT_SYMBOL(vm_insert_pfn);
2176
2177int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2178                        unsigned long pfn)
2179{
2180        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2181
2182        if (addr < vma->vm_start || addr >= vma->vm_end)
2183                return -EFAULT;
2184
2185        /*
2186         * If we don't have pte special, then we have to use the pfn_valid()
2187         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2188         * refcount the page if pfn_valid is true (hence insert_page rather
2189         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2190         * without pte special, it would there be refcounted as a normal page.
2191         */
2192        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2193                struct page *page;
2194
2195                page = pfn_to_page(pfn);
2196                return insert_page(vma, addr, page, vma->vm_page_prot);
2197        }
2198        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2199}
2200EXPORT_SYMBOL(vm_insert_mixed);
2201
2202/*
2203 * maps a range of physical memory into the requested pages. the old
2204 * mappings are removed. any references to nonexistent pages results
2205 * in null mappings (currently treated as "copy-on-access")
2206 */
2207static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2208                        unsigned long addr, unsigned long end,
2209                        unsigned long pfn, pgprot_t prot)
2210{
2211        pte_t *pte;
2212        spinlock_t *ptl;
2213
2214        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2215        if (!pte)
2216                return -ENOMEM;
2217        arch_enter_lazy_mmu_mode();
2218        do {
2219                BUG_ON(!pte_none(*pte));
2220                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2221                pfn++;
2222        } while (pte++, addr += PAGE_SIZE, addr != end);
2223        arch_leave_lazy_mmu_mode();
2224        pte_unmap_unlock(pte - 1, ptl);
2225        return 0;
2226}
2227
2228static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2229                        unsigned long addr, unsigned long end,
2230                        unsigned long pfn, pgprot_t prot)
2231{
2232        pmd_t *pmd;
2233        unsigned long next;
2234
2235        pfn -= addr >> PAGE_SHIFT;
2236        pmd = pmd_alloc(mm, pud, addr);
2237        if (!pmd)
2238                return -ENOMEM;
2239        VM_BUG_ON(pmd_trans_huge(*pmd));
2240        do {
2241                next = pmd_addr_end(addr, end);
2242                if (remap_pte_range(mm, pmd, addr, next,
2243                                pfn + (addr >> PAGE_SHIFT), prot))
2244                        return -ENOMEM;
2245        } while (pmd++, addr = next, addr != end);
2246        return 0;
2247}
2248
2249static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2250                        unsigned long addr, unsigned long end,
2251                        unsigned long pfn, pgprot_t prot)
2252{
2253        pud_t *pud;
2254        unsigned long next;
2255
2256        pfn -= addr >> PAGE_SHIFT;
2257        pud = pud_alloc(mm, pgd, addr);
2258        if (!pud)
2259                return -ENOMEM;
2260        do {
2261                next = pud_addr_end(addr, end);
2262                if (remap_pmd_range(mm, pud, addr, next,
2263                                pfn + (addr >> PAGE_SHIFT), prot))
2264                        return -ENOMEM;
2265        } while (pud++, addr = next, addr != end);
2266        return 0;
2267}
2268
2269/**
2270 * remap_pfn_range - remap kernel memory to userspace
2271 * @vma: user vma to map to
2272 * @addr: target user address to start at
2273 * @pfn: physical address of kernel memory
2274 * @size: size of map area
2275 * @prot: page protection flags for this mapping
2276 *
2277 *  Note: this is only safe if the mm semaphore is held when called.
2278 */
2279int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2280                    unsigned long pfn, unsigned long size, pgprot_t prot)
2281{
2282        pgd_t *pgd;
2283        unsigned long next;
2284        unsigned long end = addr + PAGE_ALIGN(size);
2285        struct mm_struct *mm = vma->vm_mm;
2286        int err;
2287
2288        /*
2289         * Physically remapped pages are special. Tell the
2290         * rest of the world about it:
2291         *   VM_IO tells people not to look at these pages
2292         *      (accesses can have side effects).
2293         *   VM_RESERVED is specified all over the place, because
2294         *      in 2.4 it kept swapout's vma scan off this vma; but
2295         *      in 2.6 the LRU scan won't even find its pages, so this
2296         *      flag means no more than count its pages in reserved_vm,
2297         *      and omit it from core dump, even when VM_IO turned off.
2298         *   VM_PFNMAP tells the core MM that the base pages are just
2299         *      raw PFN mappings, and do not have a "struct page" associated
2300         *      with them.
2301         *
2302         * There's a horrible special case to handle copy-on-write
2303         * behaviour that some programs depend on. We mark the "original"
2304         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2305         */
2306        if (addr == vma->vm_start && end == vma->vm_end) {
2307                vma->vm_pgoff = pfn;
2308                vma->vm_flags |= VM_PFN_AT_MMAP;
2309        } else if (is_cow_mapping(vma->vm_flags))
2310                return -EINVAL;
2311
2312        vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2313
2314        err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2315        if (err) {
2316                /*
2317                 * To indicate that track_pfn related cleanup is not
2318                 * needed from higher level routine calling unmap_vmas
2319                 */
2320                vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2321                vma->vm_flags &= ~VM_PFN_AT_MMAP;
2322                return -EINVAL;
2323        }
2324
2325        BUG_ON(addr >= end);
2326        pfn -= addr >> PAGE_SHIFT;
2327        pgd = pgd_offset(mm, addr);
2328        flush_cache_range(vma, addr, end);
2329        do {
2330                next = pgd_addr_end(addr, end);
2331                err = remap_pud_range(mm, pgd, addr, next,
2332                                pfn + (addr >> PAGE_SHIFT), prot);
2333                if (err)
2334                        break;
2335        } while (pgd++, addr = next, addr != end);
2336
2337        if (err)
2338                untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2339
2340        return err;
2341}
2342EXPORT_SYMBOL(remap_pfn_range);
2343
2344static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2345                                     unsigned long addr, unsigned long end,
2346                                     pte_fn_t fn, void *data)
2347{
2348        pte_t *pte;
2349        int err;
2350        pgtable_t token;
2351        spinlock_t *uninitialized_var(ptl);
2352
2353        pte = (mm == &init_mm) ?
2354                pte_alloc_kernel(pmd, addr) :
2355                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2356        if (!pte)
2357                return -ENOMEM;
2358
2359        BUG_ON(pmd_huge(*pmd));
2360
2361        arch_enter_lazy_mmu_mode();
2362
2363        token = pmd_pgtable(*pmd);
2364
2365        do {
2366                err = fn(pte++, token, addr, data);
2367                if (err)
2368                        break;
2369        } while (addr += PAGE_SIZE, addr != end);
2370
2371        arch_leave_lazy_mmu_mode();
2372
2373        if (mm != &init_mm)
2374                pte_unmap_unlock(pte-1, ptl);
2375        return err;
2376}
2377
2378static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2379                                     unsigned long addr, unsigned long end,
2380                                     pte_fn_t fn, void *data)
2381{
2382        pmd_t *pmd;
2383        unsigned long next;
2384        int err;
2385
2386        BUG_ON(pud_huge(*pud));
2387
2388        pmd = pmd_alloc(mm, pud, addr);
2389        if (!pmd)
2390                return -ENOMEM;
2391        do {
2392                next = pmd_addr_end(addr, end);
2393                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2394                if (err)
2395                        break;
2396        } while (pmd++, addr = next, addr != end);
2397        return err;
2398}
2399
2400static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2401                                     unsigned long addr, unsigned long end,
2402                                     pte_fn_t fn, void *data)
2403{
2404        pud_t *pud;
2405        unsigned long next;
2406        int err;
2407
2408        pud = pud_alloc(mm, pgd, addr);
2409        if (!pud)
2410                return -ENOMEM;
2411        do {
2412                next = pud_addr_end(addr, end);
2413                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2414                if (err)
2415                        break;
2416        } while (pud++, addr = next, addr != end);
2417        return err;
2418}
2419
2420/*
2421 * Scan a region of virtual memory, filling in page tables as necessary
2422 * and calling a provided function on each leaf page table.
2423 */
2424int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2425                        unsigned long size, pte_fn_t fn, void *data)
2426{
2427        pgd_t *pgd;
2428        unsigned long next;
2429        unsigned long end = addr + size;
2430        int err;
2431
2432        BUG_ON(addr >= end);
2433        pgd = pgd_offset(mm, addr);
2434        do {
2435                next = pgd_addr_end(addr, end);
2436                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2437                if (err)
2438                        break;
2439        } while (pgd++, addr = next, addr != end);
2440
2441        return err;
2442}
2443EXPORT_SYMBOL_GPL(apply_to_page_range);
2444
2445/*
2446 * handle_pte_fault chooses page fault handler according to an entry
2447 * which was read non-atomically.  Before making any commitment, on
2448 * those architectures or configurations (e.g. i386 with PAE) which
2449 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2450 * must check under lock before unmapping the pte and proceeding
2451 * (but do_wp_page is only called after already making such a check;
2452 * and do_anonymous_page can safely check later on).
2453 */
2454static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2455                                pte_t *page_table, pte_t orig_pte)
2456{
2457        int same = 1;
2458#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2459        if (sizeof(pte_t) > sizeof(unsigned long)) {
2460                spinlock_t *ptl = pte_lockptr(mm, pmd);
2461                spin_lock(ptl);
2462                same = pte_same(*page_table, orig_pte);
2463                spin_unlock(ptl);
2464        }
2465#endif
2466        pte_unmap(page_table);
2467        return same;
2468}
2469
2470static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2471{
2472        /*
2473         * If the source page was a PFN mapping, we don't have
2474         * a "struct page" for it. We do a best-effort copy by
2475         * just copying from the original user address. If that
2476         * fails, we just zero-fill it. Live with it.
2477         */
2478        if (unlikely(!src)) {
2479                void *kaddr = kmap_atomic(dst);
2480                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2481
2482                /*
2483                 * This really shouldn't fail, because the page is there
2484                 * in the page tables. But it might just be unreadable,
2485                 * in which case we just give up and fill the result with
2486                 * zeroes.
2487                 */
2488                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2489                        clear_page(kaddr);
2490                kunmap_atomic(kaddr);
2491                flush_dcache_page(dst);
2492        } else
2493                copy_user_highpage(dst, src, va, vma);
2494}
2495
2496/*
2497 * This routine handles present pages, when users try to write
2498 * to a shared page. It is done by copying the page to a new address
2499 * and decrementing the shared-page counter for the old page.
2500 *
2501 * Note that this routine assumes that the protection checks have been
2502 * done by the caller (the low-level page fault routine in most cases).
2503 * Thus we can safely just mark it writable once we've done any necessary
2504 * COW.
2505 *
2506 * We also mark the page dirty at this point even though the page will
2507 * change only once the write actually happens. This avoids a few races,
2508 * and potentially makes it more efficient.
2509 *
2510 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2511 * but allow concurrent faults), with pte both mapped and locked.
2512 * We return with mmap_sem still held, but pte unmapped and unlocked.
2513 */
2514static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2515                unsigned long address, pte_t *page_table, pmd_t *pmd,
2516                spinlock_t *ptl, pte_t orig_pte)
2517        __releases(ptl)
2518{
2519        struct page *old_page, *new_page;
2520        pte_t entry;
2521        int ret = 0;
2522        int page_mkwrite = 0;
2523        struct page *dirty_page = NULL;
2524
2525        old_page = vm_normal_page(vma, address, orig_pte);
2526        if (!old_page) {
2527                /*
2528                 * VM_MIXEDMAP !pfn_valid() case
2529                 *
2530                 * We should not cow pages in a shared writeable mapping.
2531                 * Just mark the pages writable as we can't do any dirty
2532                 * accounting on raw pfn maps.
2533                 */
2534                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2535                                     (VM_WRITE|VM_SHARED))
2536                        goto reuse;
2537                goto gotten;
2538        }
2539
2540        /*
2541         * Take out anonymous pages first, anonymous shared vmas are
2542         * not dirty accountable.
2543         */
2544        if (PageAnon(old_page) && !PageKsm(old_page)) {
2545                if (!trylock_page(old_page)) {
2546                        page_cache_get(old_page);
2547                        pte_unmap_unlock(page_table, ptl);
2548                        lock_page(old_page);
2549                        page_table = pte_offset_map_lock(mm, pmd, address,
2550                                                         &ptl);
2551                        if (!pte_same(*page_table, orig_pte)) {
2552                                unlock_page(old_page);
2553                                goto unlock;
2554                        }
2555                        page_cache_release(old_page);
2556                }
2557                if (reuse_swap_page(old_page)) {
2558                        /*
2559                         * The page is all ours.  Move it to our anon_vma so
2560                         * the rmap code will not search our parent or siblings.
2561                         * Protected against the rmap code by the page lock.
2562                         */
2563                        page_move_anon_rmap(old_page, vma, address);
2564                        unlock_page(old_page);
2565                        goto reuse;
2566                }
2567                unlock_page(old_page);
2568        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2569                                        (VM_WRITE|VM_SHARED))) {
2570                /*
2571                 * Only catch write-faults on shared writable pages,
2572                 * read-only shared pages can get COWed by
2573                 * get_user_pages(.write=1, .force=1).
2574                 */
2575                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2576                        struct vm_fault vmf;
2577                        int tmp;
2578
2579                        vmf.virtual_address = (void __user *)(address &
2580                                                                PAGE_MASK);
2581                        vmf.pgoff = old_page->index;
2582                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2583                        vmf.page = old_page;
2584
2585                        /*
2586                         * Notify the address space that the page is about to
2587                         * become writable so that it can prohibit this or wait
2588                         * for the page to get into an appropriate state.
2589                         *
2590                         * We do this without the lock held, so that it can
2591                         * sleep if it needs to.
2592                         */
2593                        page_cache_get(old_page);
2594                        pte_unmap_unlock(page_table, ptl);
2595
2596                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2597                        if (unlikely(tmp &
2598                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2599                                ret = tmp;
2600                                goto unwritable_page;
2601                        }
2602                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2603                                lock_page(old_page);
2604                                if (!old_page->mapping) {
2605                                        ret = 0; /* retry the fault */
2606                                        unlock_page(old_page);
2607                                        goto unwritable_page;
2608                                }
2609                        } else
2610                                VM_BUG_ON(!PageLocked(old_page));
2611
2612                        /*
2613                         * Since we dropped the lock we need to revalidate
2614                         * the PTE as someone else may have changed it.  If
2615                         * they did, we just return, as we can count on the
2616                         * MMU to tell us if they didn't also make it writable.
2617                         */
2618                        page_table = pte_offset_map_lock(mm, pmd, address,
2619                                                         &ptl);
2620                        if (!pte_same(*page_table, orig_pte)) {
2621                                unlock_page(old_page);
2622                                goto unlock;
2623                        }
2624
2625                        page_mkwrite = 1;
2626                }
2627                dirty_page = old_page;
2628                get_page(dirty_page);
2629
2630reuse:
2631                flush_cache_page(vma, address, pte_pfn(orig_pte));
2632                entry = pte_mkyoung(orig_pte);
2633                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2634                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2635                        update_mmu_cache(vma, address, page_table);
2636                pte_unmap_unlock(page_table, ptl);
2637                ret |= VM_FAULT_WRITE;
2638
2639                if (!dirty_page)
2640                        return ret;
2641
2642                /*
2643                 * Yes, Virginia, this is actually required to prevent a race
2644                 * with clear_page_dirty_for_io() from clearing the page dirty
2645                 * bit after it clear all dirty ptes, but before a racing
2646                 * do_wp_page installs a dirty pte.
2647                 *
2648                 * __do_fault is protected similarly.
2649                 */
2650                if (!page_mkwrite) {
2651                        wait_on_page_locked(dirty_page);
2652                        set_page_dirty_balance(dirty_page, page_mkwrite);
2653                        /* file_update_time outside page_lock */
2654                        if (vma->vm_file)
2655                                file_update_time(vma->vm_file);
2656                }
2657                put_page(dirty_page);
2658                if (page_mkwrite) {
2659                        struct address_space *mapping = dirty_page->mapping;
2660
2661                        set_page_dirty(dirty_page);
2662                        unlock_page(dirty_page);
2663                        page_cache_release(dirty_page);
2664                        if (mapping)    {
2665                                /*
2666                                 * Some device drivers do not set page.mapping
2667                                 * but still dirty their pages
2668                                 */
2669                                balance_dirty_pages_ratelimited(mapping);
2670                        }
2671                }
2672
2673                return ret;
2674        }
2675
2676        /*
2677         * Ok, we need to copy. Oh, well..
2678         */
2679        page_cache_get(old_page);
2680gotten:
2681        pte_unmap_unlock(page_table, ptl);
2682
2683        if (unlikely(anon_vma_prepare(vma)))
2684                goto oom;
2685
2686        if (is_zero_pfn(pte_pfn(orig_pte))) {
2687                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2688                if (!new_page)
2689                        goto oom;
2690        } else {
2691                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2692                if (!new_page)
2693                        goto oom;
2694                cow_user_page(new_page, old_page, address, vma);
2695        }
2696        __SetPageUptodate(new_page);
2697
2698        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2699                goto oom_free_new;
2700
2701        /*
2702         * Re-check the pte - we dropped the lock
2703         */
2704        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2705        if (likely(pte_same(*page_table, orig_pte))) {
2706                if (old_page) {
2707                        if (!PageAnon(old_page)) {
2708                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2709                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2710                        }
2711                } else
2712                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2713                flush_cache_page(vma, address, pte_pfn(orig_pte));
2714                entry = mk_pte(new_page, vma->vm_page_prot);
2715                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2716                /*
2717                 * Clear the pte entry and flush it first, before updating the
2718                 * pte with the new entry. This will avoid a race condition
2719                 * seen in the presence of one thread doing SMC and another
2720                 * thread doing COW.
2721                 */
2722                ptep_clear_flush(vma, address, page_table);
2723                page_add_new_anon_rmap(new_page, vma, address);
2724                /*
2725                 * We call the notify macro here because, when using secondary
2726                 * mmu page tables (such as kvm shadow page tables), we want the
2727                 * new page to be mapped directly into the secondary page table.
2728                 */
2729                set_pte_at_notify(mm, address, page_table, entry);
2730                update_mmu_cache(vma, address, page_table);
2731                if (old_page) {
2732                        /*
2733                         * Only after switching the pte to the new page may
2734                         * we remove the mapcount here. Otherwise another
2735                         * process may come and find the rmap count decremented
2736                         * before the pte is switched to the new page, and
2737                         * "reuse" the old page writing into it while our pte
2738                         * here still points into it and can be read by other
2739                         * threads.
2740                         *
2741                         * The critical issue is to order this
2742                         * page_remove_rmap with the ptp_clear_flush above.
2743                         * Those stores are ordered by (if nothing else,)
2744                         * the barrier present in the atomic_add_negative
2745                         * in page_remove_rmap.
2746                         *
2747                         * Then the TLB flush in ptep_clear_flush ensures that
2748                         * no process can access the old page before the
2749                         * decremented mapcount is visible. And the old page
2750                         * cannot be reused until after the decremented
2751                         * mapcount is visible. So transitively, TLBs to
2752                         * old page will be flushed before it can be reused.
2753                         */
2754                        page_remove_rmap(old_page);
2755                }
2756
2757                /* Free the old page.. */
2758                new_page = old_page;
2759                ret |= VM_FAULT_WRITE;
2760        } else
2761                mem_cgroup_uncharge_page(new_page);
2762
2763        if (new_page)
2764                page_cache_release(new_page);
2765unlock:
2766        pte_unmap_unlock(page_table, ptl);
2767        if (old_page) {
2768                /*
2769                 * Don't let another task, with possibly unlocked vma,
2770                 * keep the mlocked page.
2771                 */
2772                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2773                        lock_page(old_page);    /* LRU manipulation */
2774                        munlock_vma_page(old_page);
2775                        unlock_page(old_page);
2776                }
2777                page_cache_release(old_page);
2778        }
2779        return ret;
2780oom_free_new:
2781        page_cache_release(new_page);
2782oom:
2783        if (old_page) {
2784                if (page_mkwrite) {
2785                        unlock_page(old_page);
2786                        page_cache_release(old_page);
2787                }
2788                page_cache_release(old_page);
2789        }
2790        return VM_FAULT_OOM;
2791
2792unwritable_page:
2793        page_cache_release(old_page);
2794        return ret;
2795}
2796
2797static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2798                unsigned long start_addr, unsigned long end_addr,
2799                struct zap_details *details)
2800{
2801        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2802}
2803
2804static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2805                                            struct zap_details *details)
2806{
2807        struct vm_area_struct *vma;
2808        struct prio_tree_iter iter;
2809        pgoff_t vba, vea, zba, zea;
2810
2811        vma_prio_tree_foreach(vma, &iter, root,
2812                        details->first_index, details->last_index) {
2813
2814                vba = vma->vm_pgoff;
2815                vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2816                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2817                zba = details->first_index;
2818                if (zba < vba)
2819                        zba = vba;
2820                zea = details->last_index;
2821                if (zea > vea)
2822                        zea = vea;
2823
2824                unmap_mapping_range_vma(vma,
2825                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2826                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2827                                details);
2828        }
2829}
2830
2831static inline void unmap_mapping_range_list(struct list_head *head,
2832                                            struct zap_details *details)
2833{
2834        struct vm_area_struct *vma;
2835
2836        /*
2837         * In nonlinear VMAs there is no correspondence between virtual address
2838         * offset and file offset.  So we must perform an exhaustive search
2839         * across *all* the pages in each nonlinear VMA, not just the pages
2840         * whose virtual address lies outside the file truncation point.
2841         */
2842        list_for_each_entry(vma, head, shared.vm_set.list) {
2843                details->nonlinear_vma = vma;
2844                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2845        }
2846}
2847
2848/**
2849 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2850 * @mapping: the address space containing mmaps to be unmapped.
2851 * @holebegin: byte in first page to unmap, relative to the start of
2852 * the underlying file.  This will be rounded down to a PAGE_SIZE
2853 * boundary.  Note that this is different from truncate_pagecache(), which
2854 * must keep the partial page.  In contrast, we must get rid of
2855 * partial pages.
2856 * @holelen: size of prospective hole in bytes.  This will be rounded
2857 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2858 * end of the file.
2859 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2860 * but 0 when invalidating pagecache, don't throw away private data.
2861 */
2862void unmap_mapping_range(struct address_space *mapping,
2863                loff_t const holebegin, loff_t const holelen, int even_cows)
2864{
2865        struct zap_details details;
2866        pgoff_t hba = holebegin >> PAGE_SHIFT;
2867        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2868
2869        /* Check for overflow. */
2870        if (sizeof(holelen) > sizeof(hlen)) {
2871                long long holeend =
2872                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2873                if (holeend & ~(long long)ULONG_MAX)
2874                        hlen = ULONG_MAX - hba + 1;
2875        }
2876
2877        details.check_mapping = even_cows? NULL: mapping;
2878        details.nonlinear_vma = NULL;
2879        details.first_index = hba;
2880        details.last_index = hba + hlen - 1;
2881        if (details.last_index < details.first_index)
2882                details.last_index = ULONG_MAX;
2883
2884
2885        mutex_lock(&mapping->i_mmap_mutex);
2886        if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2887                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2888        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2889                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2890        mutex_unlock(&mapping->i_mmap_mutex);
2891}
2892EXPORT_SYMBOL(unmap_mapping_range);
2893
2894/*
2895 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2896 * but allow concurrent faults), and pte mapped but not yet locked.
2897 * We return with mmap_sem still held, but pte unmapped and unlocked.
2898 */
2899static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2900                unsigned long address, pte_t *page_table, pmd_t *pmd,
2901                unsigned int flags, pte_t orig_pte)
2902{
2903        spinlock_t *ptl;
2904        struct page *page, *swapcache = NULL;
2905        swp_entry_t entry;
2906        pte_t pte;
2907        int locked;
2908        struct mem_cgroup *ptr;
2909        int exclusive = 0;
2910        int ret = 0;
2911
2912        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2913                goto out;
2914
2915        entry = pte_to_swp_entry(orig_pte);
2916        if (unlikely(non_swap_entry(entry))) {
2917                if (is_migration_entry(entry)) {
2918                        migration_entry_wait(mm, pmd, address);
2919                } else if (is_hwpoison_entry(entry)) {
2920                        ret = VM_FAULT_HWPOISON;
2921                } else {
2922                        print_bad_pte(vma, address, orig_pte, NULL);
2923                        ret = VM_FAULT_SIGBUS;
2924                }
2925                goto out;
2926        }
2927        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2928        page = lookup_swap_cache(entry);
2929        if (!page) {
2930                page = swapin_readahead(entry,
2931                                        GFP_HIGHUSER_MOVABLE, vma, address);
2932                if (!page) {
2933                        /*
2934                         * Back out if somebody else faulted in this pte
2935                         * while we released the pte lock.
2936                         */
2937                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2938                        if (likely(pte_same(*page_table, orig_pte)))
2939                                ret = VM_FAULT_OOM;
2940                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2941                        goto unlock;
2942                }
2943
2944                /* Had to read the page from swap area: Major fault */
2945                ret = VM_FAULT_MAJOR;
2946                count_vm_event(PGMAJFAULT);
2947                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2948        } else if (PageHWPoison(page)) {
2949                /*
2950                 * hwpoisoned dirty swapcache pages are kept for killing
2951                 * owner processes (which may be unknown at hwpoison time)
2952                 */
2953                ret = VM_FAULT_HWPOISON;
2954                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2955                goto out_release;
2956        }
2957
2958        locked = lock_page_or_retry(page, mm, flags);
2959
2960        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2961        if (!locked) {
2962                ret |= VM_FAULT_RETRY;
2963                goto out_release;
2964        }
2965
2966        /*
2967         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2968         * release the swapcache from under us.  The page pin, and pte_same
2969         * test below, are not enough to exclude that.  Even if it is still
2970         * swapcache, we need to check that the page's swap has not changed.
2971         */
2972        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2973                goto out_page;
2974
2975        if (ksm_might_need_to_copy(page, vma, address)) {
2976                swapcache = page;
2977                page = ksm_does_need_to_copy(page, vma, address);
2978
2979                if (unlikely(!page)) {
2980                        ret = VM_FAULT_OOM;
2981                        page = swapcache;
2982                        swapcache = NULL;
2983                        goto out_page;
2984                }
2985        }
2986
2987        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2988                ret = VM_FAULT_OOM;
2989                goto out_page;
2990        }
2991
2992        /*
2993         * Back out if somebody else already faulted in this pte.
2994         */
2995        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2996        if (unlikely(!pte_same(*page_table, orig_pte)))
2997                goto out_nomap;
2998
2999        if (unlikely(!PageUptodate(page))) {
3000                ret = VM_FAULT_SIGBUS;
3001                goto out_nomap;
3002        }
3003
3004        /*
3005         * The page isn't present yet, go ahead with the fault.
3006         *
3007         * Be careful about the sequence of operations here.
3008         * To get its accounting right, reuse_swap_page() must be called
3009         * while the page is counted on swap but not yet in mapcount i.e.
3010         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3011         * must be called after the swap_free(), or it will never succeed.
3012         * Because delete_from_swap_page() may be called by reuse_swap_page(),
3013         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3014         * in page->private. In this case, a record in swap_cgroup  is silently
3015         * discarded at swap_free().
3016         */
3017
3018        inc_mm_counter_fast(mm, MM_ANONPAGES);
3019        dec_mm_counter_fast(mm, MM_SWAPENTS);
3020        pte = mk_pte(page, vma->vm_page_prot);
3021        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3022                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3023                flags &= ~FAULT_FLAG_WRITE;
3024                ret |= VM_FAULT_WRITE;
3025                exclusive = 1;
3026        }
3027        flush_icache_page(vma, page);
3028        set_pte_at(mm, address, page_table, pte);
3029        do_page_add_anon_rmap(page, vma, address, exclusive);
3030        /* It's better to call commit-charge after rmap is established */
3031        mem_cgroup_commit_charge_swapin(page, ptr);
3032
3033        swap_free(entry);
3034        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3035                try_to_free_swap(page);
3036        unlock_page(page);
3037        if (swapcache) {
3038                /*
3039                 * Hold the lock to avoid the swap entry to be reused
3040                 * until we take the PT lock for the pte_same() check
3041                 * (to avoid false positives from pte_same). For
3042                 * further safety release the lock after the swap_free
3043                 * so that the swap count won't change under a
3044                 * parallel locked swapcache.
3045                 */
3046                unlock_page(swapcache);
3047                page_cache_release(swapcache);
3048        }
3049
3050        if (flags & FAULT_FLAG_WRITE) {
3051                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3052                if (ret & VM_FAULT_ERROR)
3053                        ret &= VM_FAULT_ERROR;
3054                goto out;
3055        }
3056
3057        /* No need to invalidate - it was non-present before */
3058        update_mmu_cache(vma, address, page_table);
3059unlock:
3060        pte_unmap_unlock(page_table, ptl);
3061out:
3062        return ret;
3063out_nomap:
3064        mem_cgroup_cancel_charge_swapin(ptr);
3065        pte_unmap_unlock(page_table, ptl);
3066out_page:
3067        unlock_page(page);
3068out_release:
3069        page_cache_release(page);
3070        if (swapcache) {
3071                unlock_page(swapcache);
3072                page_cache_release(swapcache);
3073        }
3074        return ret;
3075}
3076
3077/*
3078 * This is like a special single-page "expand_{down|up}wards()",
3079 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3080 * doesn't hit another vma.
3081 */
3082static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3083{
3084        address &= PAGE_MASK;
3085        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3086                struct vm_area_struct *prev = vma->vm_prev;
3087
3088                /*
3089                 * Is there a mapping abutting this one below?
3090                 *
3091                 * That's only ok if it's the same stack mapping
3092                 * that has gotten split..
3093                 */
3094                if (prev && prev->vm_end == address)
3095                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3096
3097                expand_downwards(vma, address - PAGE_SIZE);
3098        }
3099        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3100                struct vm_area_struct *next = vma->vm_next;
3101
3102                /* As VM_GROWSDOWN but s/below/above/ */
3103                if (next && next->vm_start == address + PAGE_SIZE)
3104                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3105
3106                expand_upwards(vma, address + PAGE_SIZE);
3107        }
3108        return 0;
3109}
3110
3111/*
3112 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3113 * but allow concurrent faults), and pte mapped but not yet locked.
3114 * We return with mmap_sem still held, but pte unmapped and unlocked.
3115 */
3116static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3117                unsigned long address, pte_t *page_table, pmd_t *pmd,
3118                unsigned int flags)
3119{
3120        struct page *page;
3121        spinlock_t *ptl;
3122        pte_t entry;
3123
3124        pte_unmap(page_table);
3125
3126        /* Check if we need to add a guard page to the stack */
3127        if (check_stack_guard_page(vma, address) < 0)
3128                return VM_FAULT_SIGBUS;
3129
3130        /* Use the zero-page for reads */
3131        if (!(flags & FAULT_FLAG_WRITE)) {
3132                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3133                                                vma->vm_page_prot));
3134                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3135                if (!pte_none(*page_table))
3136                        goto unlock;
3137                goto setpte;
3138        }
3139
3140        /* Allocate our own private page. */
3141        if (unlikely(anon_vma_prepare(vma)))
3142                goto oom;
3143        page = alloc_zeroed_user_highpage_movable(vma, address);
3144        if (!page)
3145                goto oom;
3146        __SetPageUptodate(page);
3147
3148        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3149                goto oom_free_page;
3150
3151        entry = mk_pte(page, vma->vm_page_prot);
3152        if (vma->vm_flags & VM_WRITE)
3153                entry = pte_mkwrite(pte_mkdirty(entry));
3154
3155        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3156        if (!pte_none(*page_table))
3157                goto release;
3158
3159        inc_mm_counter_fast(mm, MM_ANONPAGES);
3160        page_add_new_anon_rmap(page, vma, address);
3161setpte:
3162        set_pte_at(mm, address, page_table, entry);
3163
3164        /* No need to invalidate - it was non-present before */
3165        update_mmu_cache(vma, address, page_table);
3166unlock:
3167        pte_unmap_unlock(page_table, ptl);
3168        return 0;
3169release:
3170        mem_cgroup_uncharge_page(page);
3171        page_cache_release(page);
3172        goto unlock;
3173oom_free_page:
3174        page_cache_release(page);
3175oom:
3176        return VM_FAULT_OOM;
3177}
3178
3179/*
3180 * __do_fault() tries to create a new page mapping. It aggressively
3181 * tries to share with existing pages, but makes a separate copy if
3182 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3183 * the next page fault.
3184 *
3185 * As this is called only for pages that do not currently exist, we
3186 * do not need to flush old virtual caches or the TLB.
3187 *
3188 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3189 * but allow concurrent faults), and pte neither mapped nor locked.
3190 * We return with mmap_sem still held, but pte unmapped and unlocked.
3191 */
3192static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3193                unsigned long address, pmd_t *pmd,
3194                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3195{
3196        pte_t *page_table;
3197        spinlock_t *ptl;
3198        struct page *page;
3199        struct page *cow_page;
3200        pte_t entry;
3201        int anon = 0;
3202        struct page *dirty_page = NULL;
3203        struct vm_fault vmf;
3204        int ret;
3205        int page_mkwrite = 0;
3206
3207        /*
3208         * If we do COW later, allocate page befor taking lock_page()
3209         * on the file cache page. This will reduce lock holding time.
3210         */
3211        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3212
3213                if (unlikely(anon_vma_prepare(vma)))
3214                        return VM_FAULT_OOM;
3215
3216                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3217                if (!cow_page)
3218                        return VM_FAULT_OOM;
3219
3220                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3221                        page_cache_release(cow_page);
3222                        return VM_FAULT_OOM;
3223                }
3224        } else
3225                cow_page = NULL;
3226
3227        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3228        vmf.pgoff = pgoff;
3229        vmf.flags = flags;
3230        vmf.page = NULL;
3231
3232        ret = vma->vm_ops->fault(vma, &vmf);
3233        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3234                            VM_FAULT_RETRY)))
3235                goto uncharge_out;
3236
3237        if (unlikely(PageHWPoison(vmf.page))) {
3238                if (ret & VM_FAULT_LOCKED)
3239                        unlock_page(vmf.page);
3240                ret = VM_FAULT_HWPOISON;
3241                goto uncharge_out;
3242        }
3243
3244        /*
3245         * For consistency in subsequent calls, make the faulted page always
3246         * locked.
3247         */
3248        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3249                lock_page(vmf.page);
3250        else
3251                VM_BUG_ON(!PageLocked(vmf.page));
3252
3253        /*
3254         * Should we do an early C-O-W break?
3255         */
3256        page = vmf.page;
3257        if (flags & FAULT_FLAG_WRITE) {
3258                if (!(vma->vm_flags & VM_SHARED)) {
3259                        page = cow_page;
3260                        anon = 1;
3261                        copy_user_highpage(page, vmf.page, address, vma);
3262                        __SetPageUptodate(page);
3263                } else {
3264                        /*
3265                         * If the page will be shareable, see if the backing
3266                         * address space wants to know that the page is about
3267                         * to become writable
3268                         */
3269                        if (vma->vm_ops->page_mkwrite) {
3270                                int tmp;
3271
3272                                unlock_page(page);
3273                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3274                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3275                                if (unlikely(tmp &
3276                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3277                                        ret = tmp;
3278                                        goto unwritable_page;
3279                                }
3280                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3281                                        lock_page(page);
3282                                        if (!page->mapping) {
3283                                                ret = 0; /* retry the fault */
3284                                                unlock_page(page);
3285                                                goto unwritable_page;
3286                                        }
3287                                } else
3288                                        VM_BUG_ON(!PageLocked(page));
3289                                page_mkwrite = 1;
3290                        }
3291                }
3292
3293        }
3294
3295        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3296
3297        /*
3298         * This silly early PAGE_DIRTY setting removes a race
3299         * due to the bad i386 page protection. But it's valid
3300         * for other architectures too.
3301         *
3302         * Note that if FAULT_FLAG_WRITE is set, we either now have
3303         * an exclusive copy of the page, or this is a shared mapping,
3304         * so we can make it writable and dirty to avoid having to
3305         * handle that later.
3306         */
3307        /* Only go through if we didn't race with anybody else... */
3308        if (likely(pte_same(*page_table, orig_pte))) {
3309                flush_icache_page(vma, page);
3310                entry = mk_pte(page, vma->vm_page_prot);
3311                if (flags & FAULT_FLAG_WRITE)
3312                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3313                if (anon) {
3314                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3315                        page_add_new_anon_rmap(page, vma, address);
3316                } else {
3317                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3318                        page_add_file_rmap(page);
3319                        if (flags & FAULT_FLAG_WRITE) {
3320                                dirty_page = page;
3321                                get_page(dirty_page);
3322                        }
3323                }
3324                set_pte_at(mm, address, page_table, entry);
3325
3326                /* no need to invalidate: a not-present page won't be cached */
3327                update_mmu_cache(vma, address, page_table);
3328        } else {
3329                if (cow_page)
3330                        mem_cgroup_uncharge_page(cow_page);
3331                if (anon)
3332                        page_cache_release(page);
3333                else
3334                        anon = 1; /* no anon but release faulted_page */
3335        }
3336
3337        pte_unmap_unlock(page_table, ptl);
3338
3339        if (dirty_page) {
3340                struct address_space *mapping = page->mapping;
3341                int dirtied = 0;
3342
3343                if (set_page_dirty(dirty_page))
3344                        dirtied = 1;
3345                unlock_page(dirty_page);
3346                put_page(dirty_page);
3347                if ((dirtied || page_mkwrite) && mapping) {
3348                        /*
3349                         * Some device drivers do not set page.mapping but still
3350                         * dirty their pages
3351                         */
3352                        balance_dirty_pages_ratelimited(mapping);
3353                }
3354
3355                /* file_update_time outside page_lock */
3356                if (vma->vm_file && !page_mkwrite)
3357                        file_update_time(vma->vm_file);
3358        } else {
3359                unlock_page(vmf.page);
3360                if (anon)
3361                        page_cache_release(vmf.page);
3362        }
3363
3364        return ret;
3365
3366unwritable_page:
3367        page_cache_release(page);
3368        return ret;
3369uncharge_out:
3370        /* fs's fault handler get error */
3371        if (cow_page) {
3372                mem_cgroup_uncharge_page(cow_page);
3373                page_cache_release(cow_page);
3374        }
3375        return ret;
3376}
3377
3378static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3379                unsigned long address, pte_t *page_table, pmd_t *pmd,
3380                unsigned int flags, pte_t orig_pte)
3381{
3382        pgoff_t pgoff = (((address & PAGE_MASK)
3383                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3384
3385        pte_unmap(page_table);
3386        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3387}
3388
3389/*
3390 * Fault of a previously existing named mapping. Repopulate the pte
3391 * from the encoded file_pte if possible. This enables swappable
3392 * nonlinear vmas.
3393 *
3394 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3395 * but allow concurrent faults), and pte mapped but not yet locked.
3396 * We return with mmap_sem still held, but pte unmapped and unlocked.
3397 */
3398static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3399                unsigned long address, pte_t *page_table, pmd_t *pmd,
3400                unsigned int flags, pte_t orig_pte)
3401{
3402        pgoff_t pgoff;
3403
3404        flags |= FAULT_FLAG_NONLINEAR;
3405
3406        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3407                return 0;
3408
3409        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3410                /*
3411                 * Page table corrupted: show pte and kill process.
3412                 */
3413                print_bad_pte(vma, address, orig_pte, NULL);
3414                return VM_FAULT_SIGBUS;
3415        }
3416
3417        pgoff = pte_to_pgoff(orig_pte);
3418        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3419}
3420
3421/*
3422 * These routines also need to handle stuff like marking pages dirty
3423 * and/or accessed for architectures that don't do it in hardware (most
3424 * RISC architectures).  The early dirtying is also good on the i386.
3425 *
3426 * There is also a hook called "update_mmu_cache()" that architectures
3427 * with external mmu caches can use to update those (ie the Sparc or
3428 * PowerPC hashed page tables that act as extended TLBs).
3429 *
3430 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3431 * but allow concurrent faults), and pte mapped but not yet locked.
3432 * We return with mmap_sem still held, but pte unmapped and unlocked.
3433 */
3434int handle_pte_fault(struct mm_struct *mm,
3435                     struct vm_area_struct *vma, unsigned long address,
3436                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3437{
3438        pte_t entry;
3439        spinlock_t *ptl;
3440
3441        entry = *pte;
3442        if (!pte_present(entry)) {
3443                if (pte_none(entry)) {
3444                        if (vma->vm_ops) {
3445                                if (likely(vma->vm_ops->fault))
3446                                        return do_linear_fault(mm, vma, address,
3447                                                pte, pmd, flags, entry);
3448                        }
3449                        return do_anonymous_page(mm, vma, address,
3450                                                 pte, pmd, flags);
3451                }
3452                if (pte_file(entry))
3453                        return do_nonlinear_fault(mm, vma, address,
3454                                        pte, pmd, flags, entry);
3455                return do_swap_page(mm, vma, address,
3456                                        pte, pmd, flags, entry);
3457        }
3458
3459        ptl = pte_lockptr(mm, pmd);
3460        spin_lock(ptl);
3461        if (unlikely(!pte_same(*pte, entry)))
3462                goto unlock;
3463        if (flags & FAULT_FLAG_WRITE) {
3464                if (!pte_write(entry))
3465                        return do_wp_page(mm, vma, address,
3466                                        pte, pmd, ptl, entry);
3467                entry = pte_mkdirty(entry);
3468        }
3469        entry = pte_mkyoung(entry);
3470        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3471                update_mmu_cache(vma, address, pte);
3472        } else {
3473                /*
3474                 * This is needed only for protection faults but the arch code
3475                 * is not yet telling us if this is a protection fault or not.
3476                 * This still avoids useless tlb flushes for .text page faults
3477                 * with threads.
3478                 */
3479                if (flags & FAULT_FLAG_WRITE)
3480                        flush_tlb_fix_spurious_fault(vma, address);
3481        }
3482unlock:
3483        pte_unmap_unlock(pte, ptl);
3484        return 0;
3485}
3486
3487/*
3488 * By the time we get here, we already hold the mm semaphore
3489 */
3490int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3491                unsigned long address, unsigned int flags)
3492{
3493        pgd_t *pgd;
3494        pud_t *pud;
3495        pmd_t *pmd;
3496        pte_t *pte;
3497
3498        __set_current_state(TASK_RUNNING);
3499
3500        count_vm_event(PGFAULT);
3501        mem_cgroup_count_vm_event(mm, PGFAULT);
3502
3503        /* do counter updates before entering really critical section. */
3504        check_sync_rss_stat(current);
3505
3506        if (unlikely(is_vm_hugetlb_page(vma)))
3507                return hugetlb_fault(mm, vma, address, flags);
3508
3509retry:
3510        pgd = pgd_offset(mm, address);
3511        pud = pud_alloc(mm, pgd, address);
3512        if (!pud)
3513                return VM_FAULT_OOM;
3514        pmd = pmd_alloc(mm, pud, address);
3515        if (!pmd)
3516                return VM_FAULT_OOM;
3517        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3518                if (!vma->vm_ops)
3519                        return do_huge_pmd_anonymous_page(mm, vma, address,
3520                                                          pmd, flags);
3521        } else {
3522                pmd_t orig_pmd = *pmd;
3523                int ret;
3524
3525                barrier();
3526                if (pmd_trans_huge(orig_pmd)) {
3527                        if (flags & FAULT_FLAG_WRITE &&
3528                            !pmd_write(orig_pmd) &&
3529                            !pmd_trans_splitting(orig_pmd)) {
3530                                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3531                                                          orig_pmd);
3532                                /*
3533                                 * If COW results in an oom, the huge pmd will
3534                                 * have been split, so retry the fault on the
3535                                 * pte for a smaller charge.
3536                                 */
3537                                if (unlikely(ret & VM_FAULT_OOM))
3538                                        goto retry;
3539                                return ret;
3540                        }
3541                        return 0;
3542                }
3543        }
3544
3545        /*
3546         * Use __pte_alloc instead of pte_alloc_map, because we can't
3547         * run pte_offset_map on the pmd, if an huge pmd could
3548         * materialize from under us from a different thread.
3549         */
3550        if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3551                return VM_FAULT_OOM;
3552        /* if an huge pmd materialized from under us just retry later */
3553        if (unlikely(pmd_trans_huge(*pmd)))
3554                return 0;
3555        /*
3556         * A regular pmd is established and it can't morph into a huge pmd
3557         * from under us anymore at this point because we hold the mmap_sem
3558         * read mode and khugepaged takes it in write mode. So now it's
3559         * safe to run pte_offset_map().
3560         */
3561        pte = pte_offset_map(pmd, address);
3562
3563        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3564}
3565
3566#ifndef __PAGETABLE_PUD_FOLDED
3567/*
3568 * Allocate page upper directory.
3569 * We've already handled the fast-path in-line.
3570 */
3571int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3572{
3573        pud_t *new = pud_alloc_one(mm, address);
3574        if (!new)
3575                return -ENOMEM;
3576
3577        smp_wmb(); /* See comment in __pte_alloc */
3578
3579        spin_lock(&mm->page_table_lock);
3580        if (pgd_present(*pgd))          /* Another has populated it */
3581                pud_free(mm, new);
3582        else
3583                pgd_populate(mm, pgd, new);
3584        spin_unlock(&mm->page_table_lock);
3585        return 0;
3586}
3587#endif /* __PAGETABLE_PUD_FOLDED */
3588
3589#ifndef __PAGETABLE_PMD_FOLDED
3590/*
3591 * Allocate page middle directory.
3592 * We've already handled the fast-path in-line.
3593 */
3594int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3595{
3596        pmd_t *new = pmd_alloc_one(mm, address);
3597        if (!new)
3598                return -ENOMEM;
3599
3600        smp_wmb(); /* See comment in __pte_alloc */
3601
3602        spin_lock(&mm->page_table_lock);
3603#ifndef __ARCH_HAS_4LEVEL_HACK
3604        if (pud_present(*pud))          /* Another has populated it */
3605                pmd_free(mm, new);
3606        else
3607                pud_populate(mm, pud, new);
3608#else
3609        if (pgd_present(*pud))          /* Another has populated it */
3610                pmd_free(mm, new);
3611        else
3612                pgd_populate(mm, pud, new);
3613#endif /* __ARCH_HAS_4LEVEL_HACK */
3614        spin_unlock(&mm->page_table_lock);
3615        return 0;
3616}
3617#endif /* __PAGETABLE_PMD_FOLDED */
3618
3619int make_pages_present(unsigned long addr, unsigned long end)
3620{
3621        int ret, len, write;
3622        struct vm_area_struct * vma;
3623
3624        vma = find_vma(current->mm, addr);
3625        if (!vma)
3626                return -ENOMEM;
3627        /*
3628         * We want to touch writable mappings with a write fault in order
3629         * to break COW, except for shared mappings because these don't COW
3630         * and we would not want to dirty them for nothing.
3631         */
3632        write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3633        BUG_ON(addr >= end);
3634        BUG_ON(end > vma->vm_end);
3635        len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3636        ret = get_user_pages(current, current->mm, addr,
3637                        len, write, 0, NULL, NULL);
3638        if (ret < 0)
3639                return ret;
3640        return ret == len ? 0 : -EFAULT;
3641}
3642
3643#if !defined(__HAVE_ARCH_GATE_AREA)
3644
3645#if defined(AT_SYSINFO_EHDR)
3646static struct vm_area_struct gate_vma;
3647
3648static int __init gate_vma_init(void)
3649{
3650        gate_vma.vm_mm = NULL;
3651        gate_vma.vm_start = FIXADDR_USER_START;
3652        gate_vma.vm_end = FIXADDR_USER_END;
3653        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3654        gate_vma.vm_page_prot = __P101;
3655
3656        return 0;
3657}
3658__initcall(gate_vma_init);
3659#endif
3660
3661struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3662{
3663#ifdef AT_SYSINFO_EHDR
3664        return &gate_vma;
3665#else
3666        return NULL;
3667#endif
3668}
3669
3670int in_gate_area_no_mm(unsigned long addr)
3671{
3672#ifdef AT_SYSINFO_EHDR
3673        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3674                return 1;
3675#endif
3676        return 0;
3677}
3678
3679#endif  /* __HAVE_ARCH_GATE_AREA */
3680
3681static int __follow_pte(struct mm_struct *mm, unsigned long address,
3682                pte_t **ptepp, spinlock_t **ptlp)
3683{
3684        pgd_t *pgd;
3685        pud_t *pud;
3686        pmd_t *pmd;
3687        pte_t *ptep;
3688
3689        pgd = pgd_offset(mm, address);
3690        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3691                goto out;
3692
3693        pud = pud_offset(pgd, address);
3694        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3695                goto out;
3696
3697        pmd = pmd_offset(pud, address);
3698        VM_BUG_ON(pmd_trans_huge(*pmd));
3699        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3700                goto out;
3701
3702        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3703        if (pmd_huge(*pmd))
3704                goto out;
3705
3706        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3707        if (!ptep)
3708                goto out;
3709        if (!pte_present(*ptep))
3710                goto unlock;
3711        *ptepp = ptep;
3712        return 0;
3713unlock:
3714        pte_unmap_unlock(ptep, *ptlp);
3715out:
3716        return -EINVAL;
3717}
3718
3719static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3720                             pte_t **ptepp, spinlock_t **ptlp)
3721{
3722        int res;
3723
3724        /* (void) is needed to make gcc happy */
3725        (void) __cond_lock(*ptlp,
3726                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
3727        return res;
3728}
3729
3730/**
3731 * follow_pfn - look up PFN at a user virtual address
3732 * @vma: memory mapping
3733 * @address: user virtual address
3734 * @pfn: location to store found PFN
3735 *
3736 * Only IO mappings and raw PFN mappings are allowed.
3737 *
3738 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3739 */
3740int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3741        unsigned long *pfn)
3742{
3743        int ret = -EINVAL;
3744        spinlock_t *ptl;
3745        pte_t *ptep;
3746
3747        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3748                return ret;
3749
3750        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3751        if (ret)
3752                return ret;
3753        *pfn = pte_pfn(*ptep);
3754        pte_unmap_unlock(ptep, ptl);
3755        return 0;
3756}
3757EXPORT_SYMBOL(follow_pfn);
3758
3759#ifdef CONFIG_HAVE_IOREMAP_PROT
3760int follow_phys(struct vm_area_struct *vma,
3761                unsigned long address, unsigned int flags,
3762                unsigned long *prot, resource_size_t *phys)
3763{
3764        int ret = -EINVAL;
3765        pte_t *ptep, pte;
3766        spinlock_t *ptl;
3767
3768        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3769                goto out;
3770
3771        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3772                goto out;
3773        pte = *ptep;
3774
3775        if ((flags & FOLL_WRITE) && !pte_write(pte))
3776                goto unlock;
3777
3778        *prot = pgprot_val(pte_pgprot(pte));
3779        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3780
3781        ret = 0;
3782unlock:
3783        pte_unmap_unlock(ptep, ptl);
3784out:
3785        return ret;
3786}
3787
3788int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3789                        void *buf, int len, int write)
3790{
3791        resource_size_t phys_addr;
3792        unsigned long prot = 0;
3793        void __iomem *maddr;
3794        int offset = addr & (PAGE_SIZE-1);
3795
3796        if (follow_phys(vma, addr, write, &prot, &phys_addr))
3797                return -EINVAL;
3798
3799        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3800        if (write)
3801                memcpy_toio(maddr + offset, buf, len);
3802        else
3803                memcpy_fromio(buf, maddr + offset, len);
3804        iounmap(maddr);
3805
3806        return len;
3807}
3808#endif
3809
3810/*
3811 * Access another process' address space as given in mm.  If non-NULL, use the
3812 * given task for page fault accounting.
3813 */
3814static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3815                unsigned long addr, void *buf, int len, int write)
3816{
3817        struct vm_area_struct *vma;
3818        void *old_buf = buf;
3819
3820        down_read(&mm->mmap_sem);
3821        /* ignore errors, just check how much was successfully transferred */
3822        while (len) {
3823                int bytes, ret, offset;
3824                void *maddr;
3825                struct page *page = NULL;
3826
3827                ret = get_user_pages(tsk, mm, addr, 1,
3828                                write, 1, &page, &vma);
3829                if (ret <= 0) {
3830                        /*
3831                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
3832                         * we can access using slightly different code.
3833                         */
3834#ifdef CONFIG_HAVE_IOREMAP_PROT
3835                        vma = find_vma(mm, addr);
3836                        if (!vma || vma->vm_start > addr)
3837                                break;
3838                        if (vma->vm_ops && vma->vm_ops->access)
3839                                ret = vma->vm_ops->access(vma, addr, buf,
3840                                                          len, write);
3841                        if (ret <= 0)
3842#endif
3843                                break;
3844                        bytes = ret;
3845                } else {
3846                        bytes = len;
3847                        offset = addr & (PAGE_SIZE-1);
3848                        if (bytes > PAGE_SIZE-offset)
3849                                bytes = PAGE_SIZE-offset;
3850
3851                        maddr = kmap(page);
3852                        if (write) {
3853                                copy_to_user_page(vma, page, addr,
3854                                                  maddr + offset, buf, bytes);
3855                                set_page_dirty_lock(page);
3856                        } else {
3857                                copy_from_user_page(vma, page, addr,
3858                                                    buf, maddr + offset, bytes);
3859                        }
3860                        kunmap(page);
3861                        page_cache_release(page);
3862                }
3863                len -= bytes;
3864                buf += bytes;
3865                addr += bytes;
3866        }
3867        up_read(&mm->mmap_sem);
3868
3869        return buf - old_buf;
3870}
3871
3872/**
3873 * access_remote_vm - access another process' address space
3874 * @mm:         the mm_struct of the target address space
3875 * @addr:       start address to access
3876 * @buf:        source or destination buffer
3877 * @len:        number of bytes to transfer
3878 * @write:      whether the access is a write
3879 *
3880 * The caller must hold a reference on @mm.
3881 */
3882int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3883                void *buf, int len, int write)
3884{
3885        return __access_remote_vm(NULL, mm, addr, buf, len, write);
3886}
3887
3888/*
3889 * Access another process' address space.
3890 * Source/target buffer must be kernel space,
3891 * Do not walk the page table directly, use get_user_pages
3892 */
3893int access_process_vm(struct task_struct *tsk, unsigned long addr,
3894                void *buf, int len, int write)
3895{
3896        struct mm_struct *mm;
3897        int ret;
3898
3899        mm = get_task_mm(tsk);
3900        if (!mm)
3901                return 0;
3902
3903        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3904        mmput(mm);
3905
3906        return ret;
3907}
3908
3909/*
3910 * Print the name of a VMA.
3911 */
3912void print_vma_addr(char *prefix, unsigned long ip)
3913{
3914        struct mm_struct *mm = current->mm;
3915        struct vm_area_struct *vma;
3916
3917        /*
3918         * Do not print if we are in atomic
3919         * contexts (in exception stacks, etc.):
3920         */
3921        if (preempt_count())
3922                return;
3923
3924        down_read(&mm->mmap_sem);
3925        vma = find_vma(mm, ip);
3926        if (vma && vma->vm_file) {
3927                struct file *f = vma->vm_file;
3928                char *buf = (char *)__get_free_page(GFP_KERNEL);
3929                if (buf) {
3930                        char *p, *s;
3931
3932                        p = d_path(&f->f_path, buf, PAGE_SIZE);
3933                        if (IS_ERR(p))
3934                                p = "?";
3935                        s = strrchr(p, '/');
3936                        if (s)
3937                                p = s+1;
3938                        printk("%s%s[%lx+%lx]", prefix, p,
3939                                        vma->vm_start,
3940                                        vma->vm_end - vma->vm_start);
3941                        free_page((unsigned long)buf);
3942                }
3943        }
3944        up_read(&mm->mmap_sem);
3945}
3946
3947#ifdef CONFIG_PROVE_LOCKING
3948void might_fault(void)
3949{
3950        /*
3951         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3952         * holding the mmap_sem, this is safe because kernel memory doesn't
3953         * get paged out, therefore we'll never actually fault, and the
3954         * below annotations will generate false positives.
3955         */
3956        if (segment_eq(get_fs(), KERNEL_DS))
3957                return;
3958
3959        might_sleep();
3960        /*
3961         * it would be nicer only to annotate paths which are not under
3962         * pagefault_disable, however that requires a larger audit and
3963         * providing helpers like get_user_atomic.
3964         */
3965        if (!in_atomic() && current->mm)
3966                might_lock_read(&current->mm->mmap_sem);
3967}
3968EXPORT_SYMBOL(might_fault);
3969#endif
3970
3971#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3972static void clear_gigantic_page(struct page *page,
3973                                unsigned long addr,
3974                                unsigned int pages_per_huge_page)
3975{
3976        int i;
3977        struct page *p = page;
3978
3979        might_sleep();
3980        for (i = 0; i < pages_per_huge_page;
3981             i++, p = mem_map_next(p, page, i)) {
3982                cond_resched();
3983                clear_user_highpage(p, addr + i * PAGE_SIZE);
3984        }
3985}
3986void clear_huge_page(struct page *page,
3987                     unsigned long addr, unsigned int pages_per_huge_page)
3988{
3989        int i;
3990
3991        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3992                clear_gigantic_page(page, addr, pages_per_huge_page);
3993                return;
3994        }
3995
3996        might_sleep();
3997        for (i = 0; i < pages_per_huge_page; i++) {
3998                cond_resched();
3999                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4000        }
4001}
4002
4003static void copy_user_gigantic_page(struct page *dst, struct page *src,
4004                                    unsigned long addr,
4005                                    struct vm_area_struct *vma,
4006                                    unsigned int pages_per_huge_page)
4007{
4008        int i;
4009        struct page *dst_base = dst;
4010        struct page *src_base = src;
4011
4012        for (i = 0; i < pages_per_huge_page; ) {
4013                cond_resched();
4014                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4015
4016                i++;
4017                dst = mem_map_next(dst, dst_base, i);
4018                src = mem_map_next(src, src_base, i);
4019        }
4020}
4021
4022void copy_user_huge_page(struct page *dst, struct page *src,
4023                         unsigned long addr, struct vm_area_struct *vma,
4024                         unsigned int pages_per_huge_page)
4025{
4026        int i;
4027
4028        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4029                copy_user_gigantic_page(dst, src, addr, vma,
4030                                        pages_per_huge_page);
4031                return;
4032        }
4033
4034        might_sleep();
4035        for (i = 0; i < pages_per_huge_page; i++) {
4036                cond_resched();
4037                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4038        }
4039}
4040#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4041
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.