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
 128static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
 129{
 130        int i;
 131
 132        for (i = 0; i < NR_MM_COUNTERS; i++) {
 133                if (task->rss_stat.count[i]) {
 134                        add_mm_counter(mm, i, task->rss_stat.count[i]);
 135                        task->rss_stat.count[i] = 0;
 136                }
 137        }
 138        task->rss_stat.events = 0;
 139}
 140
 141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 142{
 143        struct task_struct *task = current;
 144
 145        if (likely(task->mm == mm))
 146                task->rss_stat.count[member] += val;
 147        else
 148                add_mm_counter(mm, member, val);
 149}
 150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 152
 153/* sync counter once per 64 page faults */
 154#define TASK_RSS_EVENTS_THRESH  (64)
 155static void check_sync_rss_stat(struct task_struct *task)
 156{
 157        if (unlikely(task != current))
 158                return;
 159        if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 160                __sync_task_rss_stat(task, task->mm);
 161}
 162
 163unsigned long get_mm_counter(struct mm_struct *mm, int member)
 164{
 165        long val = 0;
 166
 167        /*
 168         * Don't use task->mm here...for avoiding to use task_get_mm()..
 169         * The caller must guarantee task->mm is not invalid.
 170         */
 171        val = atomic_long_read(&mm->rss_stat.count[member]);
 172        /*
 173         * counter is updated in asynchronous manner and may go to minus.
 174         * But it's never be expected number for users.
 175         */
 176        if (val < 0)
 177                return 0;
 178        return (unsigned long)val;
 179}
 180
 181void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
 182{
 183        __sync_task_rss_stat(task, mm);
 184}
 185#else /* SPLIT_RSS_COUNTING */
 186
 187#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 188#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 189
 190static void check_sync_rss_stat(struct task_struct *task)
 191{
 192}
 193
 194#endif /* SPLIT_RSS_COUNTING */
 195
 196#ifdef HAVE_GENERIC_MMU_GATHER
 197
 198static int tlb_next_batch(struct mmu_gather *tlb)
 199{
 200        struct mmu_gather_batch *batch;
 201
 202        batch = tlb->active;
 203        if (batch->next) {
 204                tlb->active = batch->next;
 205                return 1;
 206        }
 207
 208        batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 209        if (!batch)
 210                return 0;
 211
 212        batch->next = NULL;
 213        batch->nr   = 0;
 214        batch->max  = MAX_GATHER_BATCH;
 215
 216        tlb->active->next = batch;
 217        tlb->active = batch;
 218
 219        return 1;
 220}
 221
 222/* tlb_gather_mmu
 223 *      Called to initialize an (on-stack) mmu_gather structure for page-table
 224 *      tear-down from @mm. The @fullmm argument is used when @mm is without
 225 *      users and we're going to destroy the full address space (exit/execve).
 226 */
 227void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 228{
 229        tlb->mm = mm;
 230
 231        tlb->fullmm     = fullmm;
 232        tlb->need_flush = 0;
 233        tlb->fast_mode  = (num_possible_cpus() == 1);
 234        tlb->local.next = NULL;
 235        tlb->local.nr   = 0;
 236        tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 237        tlb->active     = &tlb->local;
 238
 239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 240        tlb->batch = NULL;
 241#endif
 242}
 243
 244void tlb_flush_mmu(struct mmu_gather *tlb)
 245{
 246        struct mmu_gather_batch *batch;
 247
 248        if (!tlb->need_flush)
 249                return;
 250        tlb->need_flush = 0;
 251        tlb_flush(tlb);
 252#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 253        tlb_table_flush(tlb);
 254#endif
 255
 256        if (tlb_fast_mode(tlb))
 257                return;
 258
 259        for (batch = &tlb->local; batch; batch = batch->next) {
 260                free_pages_and_swap_cache(batch->pages, batch->nr);
 261                batch->nr = 0;
 262        }
 263        tlb->active = &tlb->local;
 264}
 265
 266/* tlb_finish_mmu
 267 *      Called at the end of the shootdown operation to free up any resources
 268 *      that were required.
 269 */
 270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 271{
 272        struct mmu_gather_batch *batch, *next;
 273
 274        tlb_flush_mmu(tlb);
 275
 276        /* keep the page table cache within bounds */
 277        check_pgt_cache();
 278
 279        for (batch = tlb->local.next; batch; batch = next) {
 280                next = batch->next;
 281                free_pages((unsigned long)batch, 0);
 282        }
 283        tlb->local.next = NULL;
 284}
 285
 286/* __tlb_remove_page
 287 *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 288 *      handling the additional races in SMP caused by other CPUs caching valid
 289 *      mappings in their TLBs. Returns the number of free page slots left.
 290 *      When out of page slots we must call tlb_flush_mmu().
 291 */
 292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 293{
 294        struct mmu_gather_batch *batch;
 295
 296        VM_BUG_ON(!tlb->need_flush);
 297
 298        if (tlb_fast_mode(tlb)) {
 299                free_page_and_swap_cache(page);
 300                return 1; /* avoid calling tlb_flush_mmu() */
 301        }
 302
 303        batch = tlb->active;
 304        batch->pages[batch->nr++] = page;
 305        if (batch->nr == batch->max) {
 306                if (!tlb_next_batch(tlb))
 307                        return 0;
 308                batch = tlb->active;
 309        }
 310        VM_BUG_ON(batch->nr > batch->max);
 311
 312        return batch->max - batch->nr;
 313}
 314
 315#endif /* HAVE_GENERIC_MMU_GATHER */
 316
 317#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 318
 319/*
 320 * See the comment near struct mmu_table_batch.
 321 */
 322
 323static void tlb_remove_table_smp_sync(void *arg)
 324{
 325        /* Simply deliver the interrupt */
 326}
 327
 328static void tlb_remove_table_one(void *table)
 329{
 330        /*
 331         * This isn't an RCU grace period and hence the page-tables cannot be
 332         * assumed to be actually RCU-freed.
 333         *
 334         * It is however sufficient for software page-table walkers that rely on
 335         * IRQ disabling. See the comment near struct mmu_table_batch.
 336         */
 337        smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 338        __tlb_remove_table(table);
 339}
 340
 341static void tlb_remove_table_rcu(struct rcu_head *head)
 342{
 343        struct mmu_table_batch *batch;
 344        int i;
 345
 346        batch = container_of(head, struct mmu_table_batch, rcu);
 347
 348        for (i = 0; i < batch->nr; i++)
 349                __tlb_remove_table(batch->tables[i]);
 350
 351        free_page((unsigned long)batch);
 352}
 353
 354void tlb_table_flush(struct mmu_gather *tlb)
 355{
 356        struct mmu_table_batch **batch = &tlb->batch;
 357
 358        if (*batch) {
 359                call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 360                *batch = NULL;
 361        }
 362}
 363
 364void tlb_remove_table(struct mmu_gather *tlb, void *table)
 365{
 366        struct mmu_table_batch **batch = &tlb->batch;
 367
 368        tlb->need_flush = 1;
 369
 370        /*
 371         * When there's less then two users of this mm there cannot be a
 372         * concurrent page-table walk.
 373         */
 374        if (atomic_read(&tlb->mm->mm_users) < 2) {
 375                __tlb_remove_table(table);
 376                return;
 377        }
 378
 379        if (*batch == NULL) {
 380                *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 381                if (*batch == NULL) {
 382                        tlb_remove_table_one(table);
 383                        return;
 384                }
 385                (*batch)->nr = 0;
 386        }
 387        (*batch)->tables[(*batch)->nr++] = table;
 388        if ((*batch)->nr == MAX_TABLE_BATCH)
 389                tlb_table_flush(tlb);
 390}
 391
 392#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 393
 394/*
 395 * If a p?d_bad entry is found while walking page tables, report
 396 * the error, before resetting entry to p?d_none.  Usually (but
 397 * very seldom) called out from the p?d_none_or_clear_bad macros.
 398 */
 399
 400void pgd_clear_bad(pgd_t *pgd)
 401{
 402        pgd_ERROR(*pgd);
 403        pgd_clear(pgd);
 404}
 405
 406void pud_clear_bad(pud_t *pud)
 407{
 408        pud_ERROR(*pud);
 409        pud_clear(pud);
 410}
 411
 412void pmd_clear_bad(pmd_t *pmd)
 413{
 414        pmd_ERROR(*pmd);
 415        pmd_clear(pmd);
 416}
 417
 418/*
 419 * Note: this doesn't free the actual pages themselves. That
 420 * has been handled earlier when unmapping all the memory regions.
 421 */
 422static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 423                           unsigned long addr)
 424{
 425        pgtable_t token = pmd_pgtable(*pmd);
 426        pmd_clear(pmd);
 427        pte_free_tlb(tlb, token, addr);
 428        tlb->mm->nr_ptes--;
 429}
 430
 431static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 432                                unsigned long addr, unsigned long end,
 433                                unsigned long floor, unsigned long ceiling)
 434{
 435        pmd_t *pmd;
 436        unsigned long next;
 437        unsigned long start;
 438
 439        start = addr;
 440        pmd = pmd_offset(pud, addr);
 441        do {
 442                next = pmd_addr_end(addr, end);
 443                if (pmd_none_or_clear_bad(pmd))
 444                        continue;
 445                free_pte_range(tlb, pmd, addr);
 446        } while (pmd++, addr = next, addr != end);
 447
 448        start &= PUD_MASK;
 449        if (start < floor)
 450                return;
 451        if (ceiling) {
 452                ceiling &= PUD_MASK;
 453                if (!ceiling)
 454                        return;
 455        }
 456        if (end - 1 > ceiling - 1)
 457                return;
 458
 459        pmd = pmd_offset(pud, start);
 460        pud_clear(pud);
 461        pmd_free_tlb(tlb, pmd, start);
 462}
 463
 464static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 465                                unsigned long addr, unsigned long end,
 466                                unsigned long floor, unsigned long ceiling)
 467{
 468        pud_t *pud;
 469        unsigned long next;
 470        unsigned long start;
 471
 472        start = addr;
 473        pud = pud_offset(pgd, addr);
 474        do {
 475                next = pud_addr_end(addr, end);
 476                if (pud_none_or_clear_bad(pud))
 477                        continue;
 478                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 479        } while (pud++, addr = next, addr != end);
 480
 481        start &= PGDIR_MASK;
 482        if (start < floor)
 483                return;
 484        if (ceiling) {
 485                ceiling &= PGDIR_MASK;
 486                if (!ceiling)
 487                        return;
 488        }
 489        if (end - 1 > ceiling - 1)
 490                return;
 491
 492        pud = pud_offset(pgd, start);
 493        pgd_clear(pgd);
 494        pud_free_tlb(tlb, pud, start);
 495}
 496
 497/*
 498 * This function frees user-level page tables of a process.
 499 *
 500 * Must be called with pagetable lock held.
 501 */
 502void free_pgd_range(struct mmu_gather *tlb,
 503                        unsigned long addr, unsigned long end,
 504                        unsigned long floor, unsigned long ceiling)
 505{
 506        pgd_t *pgd;
 507        unsigned long next;
 508
 509        /*
 510         * The next few lines have given us lots of grief...
 511         *
 512         * Why are we testing PMD* at this top level?  Because often
 513         * there will be no work to do at all, and we'd prefer not to
 514         * go all the way down to the bottom just to discover that.
 515         *
 516         * Why all these "- 1"s?  Because 0 represents both the bottom
 517         * of the address space and the top of it (using -1 for the
 518         * top wouldn't help much: the masks would do the wrong thing).
 519         * The rule is that addr 0 and floor 0 refer to the bottom of
 520         * the address space, but end 0 and ceiling 0 refer to the top
 521         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 522         * that end 0 case should be mythical).
 523         *
 524         * Wherever addr is brought up or ceiling brought down, we must
 525         * be careful to reject "the opposite 0" before it confuses the
 526         * subsequent tests.  But what about where end is brought down
 527         * by PMD_SIZE below? no, end can't go down to 0 there.
 528         *
 529         * Whereas we round start (addr) and ceiling down, by different
 530         * masks at different levels, in order to test whether a table
 531         * now has no other vmas using it, so can be freed, we don't
 532         * bother to round floor or end up - the tests don't need that.
 533         */
 534
 535        addr &= PMD_MASK;
 536        if (addr < floor) {
 537                addr += PMD_SIZE;
 538                if (!addr)
 539                        return;
 540        }
 541        if (ceiling) {
 542                ceiling &= PMD_MASK;
 543                if (!ceiling)
 544                        return;
 545        }
 546        if (end - 1 > ceiling - 1)
 547                end -= PMD_SIZE;
 548        if (addr > end - 1)
 549                return;
 550
 551        pgd = pgd_offset(tlb->mm, addr);
 552        do {
 553                next = pgd_addr_end(addr, end);
 554                if (pgd_none_or_clear_bad(pgd))
 555                        continue;
 556                free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 557        } while (pgd++, addr = next, addr != end);
 558}
 559
 560void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 561                unsigned long floor, unsigned long ceiling)
 562{
 563        while (vma) {
 564                struct vm_area_struct *next = vma->vm_next;
 565                unsigned long addr = vma->vm_start;
 566
 567                /*
 568                 * Hide vma from rmap and truncate_pagecache before freeing
 569                 * pgtables
 570                 */
 571                unlink_anon_vmas(vma);
 572                unlink_file_vma(vma);
 573
 574                if (is_vm_hugetlb_page(vma)) {
 575                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 576                                floor, next? next->vm_start: ceiling);
 577                } else {
 578                        /*
 579                         * Optimization: gather nearby vmas into one call down
 580                         */
 581                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 582                               && !is_vm_hugetlb_page(next)) {
 583                                vma = next;
 584                                next = vma->vm_next;
 585                                unlink_anon_vmas(vma);
 586                                unlink_file_vma(vma);
 587                        }
 588                        free_pgd_range(tlb, addr, vma->vm_end,
 589                                floor, next? next->vm_start: ceiling);
 590                }
 591                vma = next;
 592        }
 593}
 594
 595int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 596                pmd_t *pmd, unsigned long address)
 597{
 598        pgtable_t new = pte_alloc_one(mm, address);
 599        int wait_split_huge_page;
 600        if (!new)
 601                return -ENOMEM;
 602
 603        /*
 604         * Ensure all pte setup (eg. pte page lock and page clearing) are
 605         * visible before the pte is made visible to other CPUs by being
 606         * put into page tables.
 607         *
 608         * The other side of the story is the pointer chasing in the page
 609         * table walking code (when walking the page table without locking;
 610         * ie. most of the time). Fortunately, these data accesses consist
 611         * of a chain of data-dependent loads, meaning most CPUs (alpha
 612         * being the notable exception) will already guarantee loads are
 613         * seen in-order. See the alpha page table accessors for the
 614         * smp_read_barrier_depends() barriers in page table walking code.
 615         */
 616        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 617
 618        spin_lock(&mm->page_table_lock);
 619        wait_split_huge_page = 0;
 620        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 621                mm->nr_ptes++;
 622                pmd_populate(mm, pmd, new);
 623                new = NULL;
 624        } else if (unlikely(pmd_trans_splitting(*pmd)))
 625                wait_split_huge_page = 1;
 626        spin_unlock(&mm->page_table_lock);
 627        if (new)
 628                pte_free(mm, new);
 629        if (wait_split_huge_page)
 630                wait_split_huge_page(vma->anon_vma, pmd);
 631        return 0;
 632}
 633
 634int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 635{
 636        pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 637        if (!new)
 638                return -ENOMEM;
 639
 640        smp_wmb(); /* See comment in __pte_alloc */
 641
 642        spin_lock(&init_mm.page_table_lock);
 643        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 644                pmd_populate_kernel(&init_mm, pmd, new);
 645                new = NULL;
 646        } else
 647                VM_BUG_ON(pmd_trans_splitting(*pmd));
 648        spin_unlock(&init_mm.page_table_lock);
 649        if (new)
 650                pte_free_kernel(&init_mm, new);
 651        return 0;
 652}
 653
 654static inline void init_rss_vec(int *rss)
 655{
 656        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 657}
 658
 659static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 660{
 661        int i;
 662
 663        if (current->mm == mm)
 664                sync_mm_rss(current, mm);
 665        for (i = 0; i < NR_MM_COUNTERS; i++)
 666                if (rss[i])
 667                        add_mm_counter(mm, i, rss[i]);
 668}
 669
 670/*
 671 * This function is called to print an error when a bad pte
 672 * is found. For example, we might have a PFN-mapped pte in
 673 * a region that doesn't allow it.
 674 *
 675 * The calling function must still handle the error.
 676 */
 677static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 678                          pte_t pte, struct page *page)
 679{
 680        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 681        pud_t *pud = pud_offset(pgd, addr);
 682        pmd_t *pmd = pmd_offset(pud, addr);
 683        struct address_space *mapping;
 684        pgoff_t index;
 685        static unsigned long resume;
 686        static unsigned long nr_shown;
 687        static unsigned long nr_unshown;
 688
 689        /*
 690         * Allow a burst of 60 reports, then keep quiet for that minute;
 691         * or allow a steady drip of one report per second.
 692         */
 693        if (nr_shown == 60) {
 694                if (time_before(jiffies, resume)) {
 695                        nr_unshown++;
 696                        return;
 697                }
 698                if (nr_unshown) {
 699                        printk(KERN_ALERT
 700                                "BUG: Bad page map: %lu messages suppressed\n",
 701                                nr_unshown);
 702                        nr_unshown = 0;
 703                }
 704                nr_shown = 0;
 705        }
 706        if (nr_shown++ == 0)
 707                resume = jiffies + 60 * HZ;
 708
 709        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 710        index = linear_page_index(vma, addr);
 711
 712        printk(KERN_ALERT
 713                "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 714                current->comm,
 715                (long long)pte_val(pte), (long long)pmd_val(*pmd));
 716        if (page)
 717                dump_page(page);
 718        printk(KERN_ALERT
 719                "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 720                (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 721        /*
 722         * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 723         */
 724        if (vma->vm_ops)
 725                print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
 726                                (unsigned long)vma->vm_ops->fault);
 727        if (vma->vm_file && vma->vm_file->f_op)
 728                print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
 729                                (unsigned long)vma->vm_file->f_op->mmap);
 730        dump_stack();
 731        add_taint(TAINT_BAD_PAGE);
 732}
 733
 734static inline int is_cow_mapping(vm_flags_t flags)
 735{
 736        return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 737}
 738
 739#ifndef is_zero_pfn
 740static inline int is_zero_pfn(unsigned long pfn)
 741{
 742        return pfn == zero_pfn;
 743}
 744#endif
 745
 746#ifndef my_zero_pfn
 747static inline unsigned long my_zero_pfn(unsigned long addr)
 748{
 749        return zero_pfn;
 750}
 751#endif
 752
 753/*
 754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 755 *
 756 * "Special" mappings do not wish to be associated with a "struct page" (either
 757 * it doesn't exist, or it exists but they don't want to touch it). In this
 758 * case, NULL is returned here. "Normal" mappings do have a struct page.
 759 *
 760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 761 * pte bit, in which case this function is trivial. Secondly, an architecture
 762 * may not have a spare pte bit, which requires a more complicated scheme,
 763 * described below.
 764 *
 765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 766 * special mapping (even if there are underlying and valid "struct pages").
 767 * COWed pages of a VM_PFNMAP are always normal.
 768 *
 769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 772 * mapping will always honor the rule
 773 *
 774 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 775 *
 776 * And for normal mappings this is false.
 777 *
 778 * This restricts such mappings to be a linear translation from virtual address
 779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 780 * as the vma is not a COW mapping; in that case, we know that all ptes are
 781 * special (because none can have been COWed).
 782 *
 783 *
 784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 785 *
 786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 787 * page" backing, however the difference is that _all_ pages with a struct
 788 * page (that is, those where pfn_valid is true) are refcounted and considered
 789 * normal pages by the VM. The disadvantage is that pages are refcounted
 790 * (which can be slower and simply not an option for some PFNMAP users). The
 791 * advantage is that we don't have to follow the strict linearity rule of
 792 * PFNMAP mappings in order to support COWable mappings.
 793 *
 794 */
 795#ifdef __HAVE_ARCH_PTE_SPECIAL
 796# define HAVE_PTE_SPECIAL 1
 797#else
 798# define HAVE_PTE_SPECIAL 0
 799#endif
 800struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 801                                pte_t pte)
 802{
 803        unsigned long pfn = pte_pfn(pte);
 804
 805        if (HAVE_PTE_SPECIAL) {
 806                if (likely(!pte_special(pte)))
 807                        goto check_pfn;
 808                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 809                        return NULL;
 810                if (!is_zero_pfn(pfn))
 811                        print_bad_pte(vma, addr, pte, NULL);
 812                return NULL;
 813        }
 814
 815        /* !HAVE_PTE_SPECIAL case follows: */
 816
 817        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 818                if (vma->vm_flags & VM_MIXEDMAP) {
 819                        if (!pfn_valid(pfn))
 820                                return NULL;
 821                        goto out;
 822                } else {
 823                        unsigned long off;
 824                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 825                        if (pfn == vma->vm_pgoff + off)
 826                                return NULL;
 827                        if (!is_cow_mapping(vma->vm_flags))
 828                                return NULL;
 829                }
 830        }
 831
 832        if (is_zero_pfn(pfn))
 833                return NULL;
 834check_pfn:
 835        if (unlikely(pfn > highest_memmap_pfn)) {
 836                print_bad_pte(vma, addr, pte, NULL);
 837                return NULL;
 838        }
 839
 840        /*
 841         * NOTE! We still have PageReserved() pages in the page tables.
 842         * eg. VDSO mappings can cause them to exist.
 843         */
 844out:
 845        return pfn_to_page(pfn);
 846}
 847
 848/*
 849 * copy one vm_area from one task to the other. Assumes the page tables
 850 * already present in the new task to be cleared in the whole range
 851 * covered by this vma.
 852 */
 853
 854static inline unsigned long
 855copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 856                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 857                unsigned long addr, int *rss)
 858{
 859        unsigned long vm_flags = vma->vm_flags;
 860        pte_t pte = *src_pte;
 861        struct page *page;
 862
 863        /* pte contains position in swap or file, so copy. */
 864        if (unlikely(!pte_present(pte))) {
 865                if (!pte_file(pte)) {
 866                        swp_entry_t entry = pte_to_swp_entry(pte);
 867
 868                        if (swap_duplicate(entry) < 0)
 869                                return entry.val;
 870
 871                        /* make sure dst_mm is on swapoff's mmlist. */
 872                        if (unlikely(list_empty(&dst_mm->mmlist))) {
 873                                spin_lock(&mmlist_lock);
 874                                if (list_empty(&dst_mm->mmlist))
 875                                        list_add(&dst_mm->mmlist,
 876                                                 &src_mm->mmlist);
 877                                spin_unlock(&mmlist_lock);
 878                        }
 879                        if (likely(!non_swap_entry(entry)))
 880                                rss[MM_SWAPENTS]++;
 881                        else if (is_migration_entry(entry)) {
 882                                page = migration_entry_to_page(entry);
 883
 884                                if (PageAnon(page))
 885                                        rss[MM_ANONPAGES]++;
 886                                else
 887                                        rss[MM_FILEPAGES]++;
 888
 889                                if (is_write_migration_entry(entry) &&
 890                                    is_cow_mapping(vm_flags)) {
 891                                        /*
 892                                         * COW mappings require pages in both
 893                                         * parent and child to be set to read.
 894                                         */
 895                                        make_migration_entry_read(&entry);
 896                                        pte = swp_entry_to_pte(entry);
 897                                        set_pte_at(src_mm, addr, src_pte, pte);
 898                                }
 899                        }
 900                }
 901                goto out_set_pte;
 902        }
 903
 904        /*
 905         * If it's a COW mapping, write protect it both
 906         * in the parent and the child
 907         */
 908        if (is_cow_mapping(vm_flags)) {
 909                ptep_set_wrprotect(src_mm, addr, src_pte);
 910                pte = pte_wrprotect(pte);
 911        }
 912
 913        /*
 914         * If it's a shared mapping, mark it clean in
 915         * the child
 916         */
 917        if (vm_flags & VM_SHARED)
 918                pte = pte_mkclean(pte);
 919        pte = pte_mkold(pte);
 920
 921        page = vm_normal_page(vma, addr, pte);
 922        if (page) {
 923                get_page(page);
 924                page_dup_rmap(page);
 925                if (PageAnon(page))
 926                        rss[MM_ANONPAGES]++;
 927                else
 928                        rss[MM_FILEPAGES]++;
 929        }
 930
 931out_set_pte:
 932        set_pte_at(dst_mm, addr, dst_pte, pte);
 933        return 0;
 934}
 935
 936int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 937                   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 938                   unsigned long addr, unsigned long end)
 939{
 940        pte_t *orig_src_pte, *orig_dst_pte;
 941        pte_t *src_pte, *dst_pte;
 942        spinlock_t *src_ptl, *dst_ptl;
 943        int progress = 0;
 944        int rss[NR_MM_COUNTERS];
 945        swp_entry_t entry = (swp_entry_t){0};
 946
 947again:
 948        init_rss_vec(rss);
 949
 950        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 951        if (!dst_pte)
 952                return -ENOMEM;
 953        src_pte = pte_offset_map(src_pmd, addr);
 954        src_ptl = pte_lockptr(src_mm, src_pmd);
 955        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 956        orig_src_pte = src_pte;
 957        orig_dst_pte = dst_pte;
 958        arch_enter_lazy_mmu_mode();
 959
 960        do {
 961                /*
 962                 * We are holding two locks at this point - either of them
 963                 * could generate latencies in another task on another CPU.
 964                 */
 965                if (progress >= 32) {
 966                        progress = 0;
 967                        if (need_resched() ||
 968                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 969                                break;
 970                }
 971                if (pte_none(*src_pte)) {
 972                        progress++;
 973                        continue;
 974                }
 975                entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 976                                                        vma, addr, rss);
 977                if (entry.val)
 978                        break;
 979                progress += 8;
 980        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 981
 982        arch_leave_lazy_mmu_mode();
 983        spin_unlock(src_ptl);
 984        pte_unmap(orig_src_pte);
 985        add_mm_rss_vec(dst_mm, rss);
 986        pte_unmap_unlock(orig_dst_pte, dst_ptl);
 987        cond_resched();
 988
 989        if (entry.val) {
 990                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 991                        return -ENOMEM;
 992                progress = 0;
 993        }
 994        if (addr != end)
 995                goto again;
 996        return 0;
 997}
 998
 999static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1000                pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
1001                unsigned long addr, unsigned long end)
1002{
1003        pmd_t *src_pmd, *dst_pmd;
1004        unsigned long next;
1005
1006        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1007        if (!dst_pmd)
1008                return -ENOMEM;
1009        src_pmd = pmd_offset(src_pud, addr);
1010        do {
1011                next = pmd_addr_end(addr, end);
1012                if (pmd_trans_huge(*src_pmd)) {
1013                        int err;
1014                        VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1015                        err = copy_huge_pmd(dst_mm, src_mm,
1016                                            dst_pmd, src_pmd, addr, vma);
1017                        if (err == -ENOMEM)
1018                                return -ENOMEM;
1019                        if (!err)
1020                                continue;
1021                        /* fall through */
1022                }
1023                if (pmd_none_or_clear_bad(src_pmd))
1024                        continue;
1025                if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1026                                                vma, addr, next))
1027                        return -ENOMEM;
1028        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1029        return 0;
1030}
1031
1032static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033                pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1034                unsigned long addr, unsigned long end)
1035{
1036        pud_t *src_pud, *dst_pud;
1037        unsigned long next;
1038
1039        dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1040        if (!dst_pud)
1041                return -ENOMEM;
1042        src_pud = pud_offset(src_pgd, addr);
1043        do {
1044                next = pud_addr_end(addr, end);
1045                if (pud_none_or_clear_bad(src_pud))
1046                        continue;
1047                if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1048                                                vma, addr, next))
1049                        return -ENOMEM;
1050        } while (dst_pud++, src_pud++, addr = next, addr != end);
1051        return 0;
1052}
1053
1054int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1055                struct vm_area_struct *vma)
1056{
1057        pgd_t *src_pgd, *dst_pgd;
1058        unsigned long next;
1059        unsigned long addr = vma->vm_start;
1060        unsigned long end = vma->vm_end;
1061        int ret;
1062
1063        /*
1064         * Don't copy ptes where a page fault will fill them correctly.
1065         * Fork becomes much lighter when there are big shared or private
1066         * readonly mappings. The tradeoff is that copy_page_range is more
1067         * efficient than faulting.
1068         */
1069        if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1070                if (!vma->anon_vma)
1071                        return 0;
1072        }
1073
1074        if (is_vm_hugetlb_page(vma))
1075                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1076
1077        if (unlikely(is_pfn_mapping(vma))) {
1078                /*
1079                 * We do not free on error cases below as remove_vma
1080                 * gets called on error from higher level routine
1081                 */
1082                ret = track_pfn_vma_copy(vma);
1083                if (ret)
1084                        return ret;
1085        }
1086
1087        /*
1088         * We need to invalidate the secondary MMU mappings only when
1089         * there could be a permission downgrade on the ptes of the
1090         * parent mm. And a permission downgrade will only happen if
1091         * is_cow_mapping() returns true.
1092         */
1093        if (is_cow_mapping(vma->vm_flags))
1094                mmu_notifier_invalidate_range_start(src_mm, addr, end);
1095
1096        ret = 0;
1097        dst_pgd = pgd_offset(dst_mm, addr);
1098        src_pgd = pgd_offset(src_mm, addr);
1099        do {
1100                next = pgd_addr_end(addr, end);
1101                if (pgd_none_or_clear_bad(src_pgd))
1102                        continue;
1103                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1104                                            vma, addr, next))) {
1105                        ret = -ENOMEM;
1106                        break;
1107                }
1108        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1109
1110        if (is_cow_mapping(vma->vm_flags))
1111                mmu_notifier_invalidate_range_end(src_mm,
1112                                                  vma->vm_start, end);
1113        return ret;
1114}
1115
1116static unsigned long zap_pte_range(struct mmu_gather *tlb,
1117                                struct vm_area_struct *vma, pmd_t *pmd,
1118                                unsigned long addr, unsigned long end,
1119                                struct zap_details *details)
1120{
1121        struct mm_struct *mm = tlb->mm;
1122        int force_flush = 0;
1123        int rss[NR_MM_COUNTERS];
1124        spinlock_t *ptl;
1125        pte_t *start_pte;
1126        pte_t *pte;
1127
1128again:
1129        init_rss_vec(rss);
1130        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1131        pte = start_pte;
1132        arch_enter_lazy_mmu_mode();
1133        do {
1134                pte_t ptent = *pte;
1135                if (pte_none(ptent)) {
1136                        continue;
1137                }
1138
1139                if (pte_present(ptent)) {
1140                        struct page *page;
1141
1142                        page = vm_normal_page(vma, addr, ptent);
1143                        if (unlikely(details) && page) {
1144                                /*
1145                                 * unmap_shared_mapping_pages() wants to
1146                                 * invalidate cache without truncating:
1147                                 * unmap shared but keep private pages.
1148                                 */
1149                                if (details->check_mapping &&
1150                                    details->check_mapping != page->mapping)
1151                                        continue;
1152                                /*
1153                                 * Each page->index must be checked when
1154                                 * invalidating or truncating nonlinear.
1155                                 */
1156                                if (details->nonlinear_vma &&
1157                                    (page->index < details->first_index ||
1158                                     page->index > details->last_index))
1159                                        continue;
1160                        }
1161                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1162                                                        tlb->fullmm);
1163                        tlb_remove_tlb_entry(tlb, pte, addr);
1164                        if (unlikely(!page))
1165                                continue;
1166                        if (unlikely(details) && details->nonlinear_vma
1167                            && linear_page_index(details->nonlinear_vma,
1168                                                addr) != page->index)
1169                                set_pte_at(mm, addr, pte,
1170                                           pgoff_to_pte(page->index));
1171                        if (PageAnon(page))
1172                                rss[MM_ANONPAGES]--;
1173                        else {
1174                                if (pte_dirty(ptent))
1175                                        set_page_dirty(page);
1176                                if (pte_young(ptent) &&
1177                                    likely(!VM_SequentialReadHint(vma)))
1178                                        mark_page_accessed(page);
1179                                rss[MM_FILEPAGES]--;
1180                        }
1181                        page_remove_rmap(page);
1182                        if (unlikely(page_mapcount(page) < 0))
1183                                print_bad_pte(vma, addr, ptent, page);
1184                        force_flush = !__tlb_remove_page(tlb, page);
1185                        if (force_flush)
1186                                break;
1187                        continue;
1188                }
1189                /*
1190                 * If details->check_mapping, we leave swap entries;
1191                 * if details->nonlinear_vma, we leave file entries.
1192                 */
1193                if (unlikely(details))
1194                        continue;
1195                if (pte_file(ptent)) {
1196                        if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1197                                print_bad_pte(vma, addr, ptent, NULL);
1198                } else {
1199                        swp_entry_t entry = pte_to_swp_entry(ptent);
1200
1201                        if (!non_swap_entry(entry))
1202                                rss[MM_SWAPENTS]--;
1203                        else if (is_migration_entry(entry)) {
1204                                struct page *page;
1205
1206                                page = migration_entry_to_page(entry);
1207
1208                                if (PageAnon(page))
1209                                        rss[MM_ANONPAGES]--;
1210                                else
1211                                        rss[MM_FILEPAGES]--;
1212                        }
1213                        if (unlikely(!free_swap_and_cache(entry)))
1214                                print_bad_pte(vma, addr, ptent, NULL);
1215                }
1216                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1217        } while (pte++, addr += PAGE_SIZE, addr != end);
1218
1219        add_mm_rss_vec(mm, rss);
1220        arch_leave_lazy_mmu_mode();
1221        pte_unmap_unlock(start_pte, ptl);
1222
1223        /*
1224         * mmu_gather ran out of room to batch pages, we break out of
1225         * the PTE lock to avoid doing the potential expensive TLB invalidate
1226         * and page-free while holding it.
1227         */
1228        if (force_flush) {
1229                force_flush = 0;
1230                tlb_flush_mmu(tlb);
1231                if (addr != end)
1232                        goto again;
1233        }
1234
1235        return addr;
1236}
1237
1238static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1239                                struct vm_area_struct *vma, pud_t *pud,
1240                                unsigned long addr, unsigned long end,
1241                                struct zap_details *details)
1242{
1243        pmd_t *pmd;
1244        unsigned long next;
1245
1246        pmd = pmd_offset(pud, addr);
1247        do {
1248                next = pmd_addr_end(addr, end);
1249                if (pmd_trans_huge(*pmd)) {
1250                        if (next - addr != HPAGE_PMD_SIZE) {
1251                                VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1252                                split_huge_page_pmd(vma->vm_mm, pmd);
1253                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1254                                goto next;
1255                        /* fall through */
1256                }
1257                /*
1258                 * Here there can be other concurrent MADV_DONTNEED or
1259                 * trans huge page faults running, and if the pmd is
1260                 * none or trans huge it can change under us. This is
1261                 * because MADV_DONTNEED holds the mmap_sem in read
1262                 * mode.
1263                 */
1264                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1265                        goto next;
1266                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1267next:
1268                cond_resched();
1269        } while (pmd++, addr = next, addr != end);
1270
1271        return addr;
1272}
1273
1274static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1275                                struct vm_area_struct *vma, pgd_t *pgd,
1276                                unsigned long addr, unsigned long end,
1277                                struct zap_details *details)
1278{
1279        pud_t *pud;
1280        unsigned long next;
1281
1282        pud = pud_offset(pgd, addr);
1283        do {
1284                next = pud_addr_end(addr, end);
1285                if (pud_none_or_clear_bad(pud))
1286                        continue;
1287                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1288        } while (pud++, addr = next, addr != end);
1289
1290        return addr;
1291}
1292
1293static unsigned long unmap_page_range(struct mmu_gather *tlb,
1294                                struct vm_area_struct *vma,
1295                                unsigned long addr, unsigned long end,
1296                                struct zap_details *details)
1297{
1298        pgd_t *pgd;
1299        unsigned long next;
1300
1301        if (details && !details->check_mapping && !details->nonlinear_vma)
1302                details = NULL;
1303
1304        BUG_ON(addr >= end);
1305        mem_cgroup_uncharge_start();
1306        tlb_start_vma(tlb, vma);
1307        pgd = pgd_offset(vma->vm_mm, addr);
1308        do {
1309                next = pgd_addr_end(addr, end);
1310                if (pgd_none_or_clear_bad(pgd))
1311                        continue;
1312                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1313        } while (pgd++, addr = next, addr != end);
1314        tlb_end_vma(tlb, vma);
1315        mem_cgroup_uncharge_end();
1316
1317        return addr;
1318}
1319
1320/**
1321 * unmap_vmas - unmap a range of memory covered by a list of vma's
1322 * @tlb: address of the caller's struct mmu_gather
1323 * @vma: the starting vma
1324 * @start_addr: virtual address at which to start unmapping
1325 * @end_addr: virtual address at which to end unmapping
1326 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1327 * @details: details of nonlinear truncation or shared cache invalidation
1328 *
1329 * Returns the end address of the unmapping (restart addr if interrupted).
1330 *
1331 * Unmap all pages in the vma list.
1332 *
1333 * Only addresses between `start' and `end' will be unmapped.
1334 *
1335 * The VMA list must be sorted in ascending virtual address order.
1336 *
1337 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1338 * range after unmap_vmas() returns.  So the only responsibility here is to
1339 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1340 * drops the lock and schedules.
1341 */
1342unsigned long unmap_vmas(struct mmu_gather *tlb,
1343                struct vm_area_struct *vma, unsigned long start_addr,
1344                unsigned long end_addr, unsigned long *nr_accounted,
1345                struct zap_details *details)
1346{
1347        unsigned long start = start_addr;
1348        struct mm_struct *mm = vma->vm_mm;
1349
1350        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1351        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1352                unsigned long end;
1353
1354                start = max(vma->vm_start, start_addr);
1355                if (start >= vma->vm_end)
1356                        continue;
1357                end = min(vma->vm_end, end_addr);
1358                if (end <= vma->vm_start)
1359                        continue;
1360
1361                if (vma->vm_flags & VM_ACCOUNT)
1362                        *nr_accounted += (end - start) >> PAGE_SHIFT;
1363
1364                if (unlikely(is_pfn_mapping(vma)))
1365                        untrack_pfn_vma(vma, 0, 0);
1366
1367                while (start != end) {
1368                        if (unlikely(is_vm_hugetlb_page(vma))) {
1369                                /*
1370                                 * It is undesirable to test vma->vm_file as it
1371                                 * should be non-null for valid hugetlb area.
1372                                 * However, vm_file will be NULL in the error
1373                                 * cleanup path of do_mmap_pgoff. When
1374                                 * hugetlbfs ->mmap method fails,
1375                                 * do_mmap_pgoff() nullifies vma->vm_file
1376                                 * before calling this function to clean up.
1377                                 * Since no pte has actually been setup, it is
1378                                 * safe to do nothing in this case.
1379                                 */
1380                                if (vma->vm_file)
1381                                        unmap_hugepage_range(vma, start, end, NULL);
1382
1383                                start = end;
1384                        } else
1385                                start = unmap_page_range(tlb, vma, start, end, details);
1386                }
1387        }
1388
1389        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1390        return start;   /* which is now the end (or restart) address */
1391}
1392
1393/**
1394 * zap_page_range - remove user pages in a given range
1395 * @vma: vm_area_struct holding the applicable pages
1396 * @address: starting address of pages to zap
1397 * @size: number of bytes to zap
1398 * @details: details of nonlinear truncation or shared cache invalidation
1399 */
1400unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1401                unsigned long size, struct zap_details *details)
1402{
1403        struct mm_struct *mm = vma->vm_mm;
1404        struct mmu_gather tlb;
1405        unsigned long end = address + size;
1406        unsigned long nr_accounted = 0;
1407
1408        lru_add_drain();
1409        tlb_gather_mmu(&tlb, mm, 0);
1410        update_hiwater_rss(mm);
1411        end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1412        tlb_finish_mmu(&tlb, address, end);
1413        return end;
1414}
1415
1416/**
1417 * zap_vma_ptes - remove ptes mapping the vma
1418 * @vma: vm_area_struct holding ptes to be zapped
1419 * @address: starting address of pages to zap
1420 * @size: number of bytes to zap
1421 *
1422 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1423 *
1424 * The entire address range must be fully contained within the vma.
1425 *
1426 * Returns 0 if successful.
1427 */
1428int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1429                unsigned long size)
1430{
1431        if (address < vma->vm_start || address + size > vma->vm_end ||
1432                        !(vma->vm_flags & VM_PFNMAP))
1433                return -1;
1434        zap_page_range(vma, address, size, NULL);
1435        return 0;
1436}
1437EXPORT_SYMBOL_GPL(zap_vma_ptes);
1438
1439/**
1440 * follow_page - look up a page descriptor from a user-virtual address
1441 * @vma: vm_area_struct mapping @address
1442 * @address: virtual address to look up
1443 * @flags: flags modifying lookup behaviour
1444 *
1445 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1446 *
1447 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1448 * an error pointer if there is a mapping to something not represented
1449 * by a page descriptor (see also vm_normal_page()).
1450 */
1451struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1452                        unsigned int flags)
1453{
1454        pgd_t *pgd;
1455        pud_t *pud;
1456        pmd_t *pmd;
1457        pte_t *ptep, pte;
1458        spinlock_t *ptl;
1459        struct page *page;
1460        struct mm_struct *mm = vma->vm_mm;
1461
1462        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1463        if (!IS_ERR(page)) {
1464                BUG_ON(flags & FOLL_GET);
1465                goto out;
1466        }
1467
1468        page = NULL;
1469        pgd = pgd_offset(mm, address);
1470        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1471                goto no_page_table;
1472
1473        pud = pud_offset(pgd, address);
1474        if (pud_none(*pud))
1475                goto no_page_table;
1476        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1477                BUG_ON(flags & FOLL_GET);
1478                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1479                goto out;
1480        }
1481        if (unlikely(pud_bad(*pud)))
1482                goto no_page_table;
1483
1484        pmd = pmd_offset(pud, address);
1485        if (pmd_none(*pmd))
1486                goto no_page_table;
1487        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1488                BUG_ON(flags & FOLL_GET);
1489                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1490                goto out;
1491        }
1492        if (pmd_trans_huge(*pmd)) {
1493                if (flags & FOLL_SPLIT) {
1494                        split_huge_page_pmd(mm, pmd);
1495                        goto split_fallthrough;
1496                }
1497                spin_lock(&mm->page_table_lock);
1498                if (likely(pmd_trans_huge(*pmd))) {
1499                        if (unlikely(pmd_trans_splitting(*pmd))) {
1500                                spin_unlock(&mm->page_table_lock);
1501                                wait_split_huge_page(vma->anon_vma, pmd);
1502                        } else {
1503                                page = follow_trans_huge_pmd(mm, address,
1504                                                             pmd, flags);
1505                                spin_unlock(&mm->page_table_lock);
1506                                goto out;
1507                        }
1508                } else
1509                        spin_unlock(&mm->page_table_lock);
1510                /* fall through */
1511        }
1512split_fallthrough:
1513        if (unlikely(pmd_bad(*pmd)))
1514                goto no_page_table;
1515
1516        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1517
1518        pte = *ptep;
1519        if (!pte_present(pte))
1520                goto no_page;
1521        if ((flags & FOLL_WRITE) && !pte_write(pte))
1522                goto unlock;
1523
1524        page = vm_normal_page(vma, address, pte);
1525        if (unlikely(!page)) {
1526                if ((flags & FOLL_DUMP) ||
1527                    !is_zero_pfn(pte_pfn(pte)))
1528                        goto bad_page;
1529                page = pte_page(pte);
1530        }
1531
1532        if (flags & FOLL_GET)
1533                get_page_foll(page);
1534        if (flags & FOLL_TOUCH) {
1535                if ((flags & FOLL_WRITE) &&
1536                    !pte_dirty(pte) && !PageDirty(page))
1537                        set_page_dirty(page);
1538                /*
1539                 * pte_mkyoung() would be more correct here, but atomic care
1540                 * is needed to avoid losing the dirty bit: it is easier to use
1541                 * mark_page_accessed().
1542                 */
1543                mark_page_accessed(page);
1544        }
1545        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1546                /*
1547                 * The preliminary mapping check is mainly to avoid the
1548                 * pointless overhead of lock_page on the ZERO_PAGE
1549                 * which might bounce very badly if there is contention.
1550                 *
1551                 * If the page is already locked, we don't need to
1552                 * handle it now - vmscan will handle it later if and
1553                 * when it attempts to reclaim the page.
1554                 */
1555                if (page->mapping && trylock_page(page)) {
1556                        lru_add_drain();  /* push cached pages to LRU */
1557                        /*
1558                         * Because we lock page here and migration is
1559                         * blocked by the pte's page reference, we need
1560                         * only check for file-cache page truncation.
1561                         */
1562                        if (page->mapping)
1563                                mlock_vma_page(page);
1564                        unlock_page(page);
1565                }
1566        }
1567unlock:
1568        pte_unmap_unlock(ptep, ptl);
1569out:
1570        return page;
1571
1572bad_page:
1573        pte_unmap_unlock(ptep, ptl);
1574        return ERR_PTR(-EFAULT);
1575
1576no_page:
1577        pte_unmap_unlock(ptep, ptl);
1578        if (!pte_none(pte))
1579                return page;
1580
1581no_page_table:
1582        /*
1583         * When core dumping an enormous anonymous area that nobody
1584         * has touched so far, we don't want to allocate unnecessary pages or
1585         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1586         * then get_dump_page() will return NULL to leave a hole in the dump.
1587         * But we can only make this optimization where a hole would surely
1588         * be zero-filled if handle_mm_fault() actually did handle it.
1589         */
1590        if ((flags & FOLL_DUMP) &&
1591            (!vma->vm_ops || !vma->vm_ops->fault))
1592                return ERR_PTR(-EFAULT);
1593        return page;
1594}
1595
1596static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1597{
1598        return stack_guard_page_start(vma, addr) ||
1599               stack_guard_page_end(vma, addr+PAGE_SIZE);
1600}
1601
1602/**
1603 * __get_user_pages() - pin user pages in memory
1604 * @tsk:        task_struct of target task
1605 * @mm:         mm_struct of target mm
1606 * @start:      starting user address
1607 * @nr_pages:   number of pages from start to pin
1608 * @gup_flags:  flags modifying pin behaviour
1609 * @pages:      array that receives pointers to the pages pinned.
1610 *              Should be at least nr_pages long. Or NULL, if caller
1611 *              only intends to ensure the pages are faulted in.
1612 * @vmas:       array of pointers to vmas corresponding to each page.
1613 *              Or NULL if the caller does not require them.
1614 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1615 *
1616 * Returns number of pages pinned. This may be fewer than the number
1617 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1618 * were pinned, returns -errno. Each page returned must be released
1619 * with a put_page() call when it is finished with. vmas will only
1620 * remain valid while mmap_sem is held.
1621 *
1622 * Must be called with mmap_sem held for read or write.
1623 *
1624 * __get_user_pages walks a process's page tables and takes a reference to
1625 * each struct page that each user address corresponds to at a given
1626 * instant. That is, it takes the page that would be accessed if a user
1627 * thread accesses the given user virtual address at that instant.
1628 *
1629 * This does not guarantee that the page exists in the user mappings when
1630 * __get_user_pages returns, and there may even be a completely different
1631 * page there in some cases (eg. if mmapped pagecache has been invalidated
1632 * and subsequently re faulted). However it does guarantee that the page
1633 * won't be freed completely. And mostly callers simply care that the page
1634 * contains data that was valid *at some point in time*. Typically, an IO
1635 * or similar operation cannot guarantee anything stronger anyway because
1636 * locks can't be held over the syscall boundary.
1637 *
1638 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1639 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1640 * appropriate) must be called after the page is finished with, and
1641 * before put_page is called.
1642 *
1643 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1644 * or mmap_sem contention, and if waiting is needed to pin all pages,
1645 * *@nonblocking will be set to 0.
1646 *
1647 * In most cases, get_user_pages or get_user_pages_fast should be used
1648 * instead of __get_user_pages. __get_user_pages should be used only if
1649 * you need some special @gup_flags.
1650 */
1651int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1652                     unsigned long start, int nr_pages, unsigned int gup_flags,
1653                     struct page **pages, struct vm_area_struct **vmas,
1654                     int *nonblocking)
1655{
1656        int i;
1657        unsigned long vm_flags;
1658
1659        if (nr_pages <= 0)
1660                return 0;
1661
1662        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1663
1664        /* 
1665         * Require read or write permissions.
1666         * If FOLL_FORCE is set, we only require the "MAY" flags.
1667         */
1668        vm_flags  = (gup_flags & FOLL_WRITE) ?
1669                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1670        vm_flags &= (gup_flags & FOLL_FORCE) ?
1671                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1672        i = 0;
1673
1674        do {
1675                struct vm_area_struct *vma;
1676
1677                vma = find_extend_vma(mm, start);
1678                if (!vma && in_gate_area(mm, start)) {
1679                        unsigned long pg = start & PAGE_MASK;
1680                        pgd_t *pgd;
1681                        pud_t *pud;
1682                        pmd_t *pmd;
1683                        pte_t *pte;
1684
1685                        /* user gate pages are read-only */
1686                        if (gup_flags & FOLL_WRITE)
1687                                return i ? : -EFAULT;
1688                        if (pg > TASK_SIZE)
1689                                pgd = pgd_offset_k(pg);
1690                        else
1691                                pgd = pgd_offset_gate(mm, pg);
1692                        BUG_ON(pgd_none(*pgd));
1693                        pud = pud_offset(pgd, pg);
1694                        BUG_ON(pud_none(*pud));
1695                        pmd = pmd_offset(pud, pg);
1696                        if (pmd_none(*pmd))
1697                                return i ? : -EFAULT;
1698                        VM_BUG_ON(pmd_trans_huge(*pmd));
1699                        pte = pte_offset_map(pmd, pg);
1700                        if (pte_none(*pte)) {
1701                                pte_unmap(pte);
1702                                return i ? : -EFAULT;
1703                        }
1704                        vma = get_gate_vma(mm);
1705                        if (pages) {
1706                                struct page *page;
1707
1708                                page = vm_normal_page(vma, start, *pte);
1709                                if (!page) {
1710                                        if (!(gup_flags & FOLL_DUMP) &&
1711                                             is_zero_pfn(pte_pfn(*pte)))
1712                                                page = pte_page(*pte);
1713                                        else {
1714                                                pte_unmap(pte);
1715                                                return i ? : -EFAULT;
1716                                        }
1717                                }
1718                                pages[i] = page;
1719                                get_page(page);
1720                        }
1721                        pte_unmap(pte);
1722                        goto next_page;
1723                }
1724
1725                if (!vma ||
1726                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1727                    !(vm_flags & vma->vm_flags))
1728                        return i ? : -EFAULT;
1729
1730                if (is_vm_hugetlb_page(vma)) {
1731                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1732                                        &start, &nr_pages, i, gup_flags);
1733                        continue;
1734                }
1735
1736                do {
1737                        struct page *page;
1738                        unsigned int foll_flags = gup_flags;
1739
1740                        /*
1741                         * If we have a pending SIGKILL, don't keep faulting
1742                         * pages and potentially allocating memory.
1743                         */
1744                        if (unlikely(fatal_signal_pending(current)))
1745                                return i ? i : -ERESTARTSYS;
1746
1747                        cond_resched();
1748                        while (!(page = follow_page(vma, start, foll_flags))) {
1749                                int ret;
1750                                unsigned int fault_flags = 0;
1751
1752                                /* For mlock, just skip the stack guard page. */
1753                                if (foll_flags & FOLL_MLOCK) {
1754                                        if (stack_guard_page(vma, start))
1755                                                goto next_page;
1756                                }
1757                                if (foll_flags & FOLL_WRITE)
1758                                        fault_flags |= FAULT_FLAG_WRITE;
1759                                if (nonblocking)
1760                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1761                                if (foll_flags & FOLL_NOWAIT)
1762                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1763
1764                                ret = handle_mm_fault(mm, vma, start,
1765                                                        fault_flags);
1766
1767                                if (ret & VM_FAULT_ERROR) {
1768                                        if (ret & VM_FAULT_OOM)
1769                                                return i ? i : -ENOMEM;
1770                                        if (ret & (VM_FAULT_HWPOISON |
1771                                                   VM_FAULT_HWPOISON_LARGE)) {
1772                                                if (i)
1773                                                        return i;
1774                                                else if (gup_flags & FOLL_HWPOISON)
1775                                                        return -EHWPOISON;
1776                                                else
1777                                                        return -EFAULT;
1778                                        }
1779                                        if (ret & VM_FAULT_SIGBUS)
1780                                                return i ? i : -EFAULT;
1781                                        BUG();
1782                                }
1783
1784                                if (tsk) {
1785                                        if (ret & VM_FAULT_MAJOR)
1786                                                tsk->maj_flt++;
1787                                        else
1788                                                tsk->min_flt++;
1789                                }
1790
1791                                if (ret & VM_FAULT_RETRY) {
1792                                        if (nonblocking)
1793                                                *nonblocking = 0;
1794                                        return i;
1795                                }
1796
1797                                /*
1798                                 * The VM_FAULT_WRITE bit tells us that
1799                                 * do_wp_page has broken COW when necessary,
1800                                 * even if maybe_mkwrite decided not to set
1801                                 * pte_write. We can thus safely do subsequent
1802                                 * page lookups as if they were reads. But only
1803                                 * do so when looping for pte_write is futile:
1804                                 * in some cases userspace may also be wanting
1805                                 * to write to the gotten user page, which a
1806                                 * read fault here might prevent (a readonly
1807                                 * page might get reCOWed by userspace write).
1808                                 */
1809                                if ((ret & VM_FAULT_WRITE) &&
1810                                    !(vma->vm_flags & VM_WRITE))
1811                                        foll_flags &= ~FOLL_WRITE;
1812
1813                                cond_resched();
1814                        }
1815                        if (IS_ERR(page))
1816                                return i ? i : PTR_ERR(page);
1817                        if (pages) {
1818                                pages[i] = page;
1819
1820                                flush_anon_page(vma, page, start);
1821                                flush_dcache_page(page);
1822                        }
1823next_page:
1824                        if (vmas)
1825                                vmas[i] = vma;
1826                        i++;
1827                        start += PAGE_SIZE;
1828                        nr_pages--;
1829                } while (nr_pages && start < vma->vm_end);
1830        } while (nr_pages);
1831        return i;
1832}
1833EXPORT_SYMBOL(__get_user_pages);
1834
1835/*
1836 * fixup_user_fault() - manually resolve a user page fault
1837 * @tsk:        the task_struct to use for page fault accounting, or
1838 *              NULL if faults are not to be recorded.
1839 * @mm:         mm_struct of target mm
1840 * @address:    user address
1841 * @fault_flags:flags to pass down to handle_mm_fault()
1842 *
1843 * This is meant to be called in the specific scenario where for locking reasons
1844 * we try to access user memory in atomic context (within a pagefault_disable()
1845 * section), this returns -EFAULT, and we want to resolve the user fault before
1846 * trying again.
1847 *
1848 * Typically this is meant to be used by the futex code.
1849 *
1850 * The main difference with get_user_pages() is that this function will
1851 * unconditionally call handle_mm_fault() which will in turn perform all the
1852 * necessary SW fixup of the dirty and young bits in the PTE, while
1853 * handle_mm_fault() only guarantees to update these in the struct page.
1854 *
1855 * This is important for some architectures where those bits also gate the
1856 * access permission to the page because they are maintained in software.  On
1857 * such architectures, gup() will not be enough to make a subsequent access
1858 * succeed.
1859 *
1860 * This should be called with the mm_sem held for read.
1861 */
1862int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1863                     unsigned long address, unsigned int fault_flags)
1864{
1865        struct vm_area_struct *vma;
1866        int ret;
1867
1868        vma = find_extend_vma(mm, address);
1869        if (!vma || address < vma->vm_start)
1870                return -EFAULT;
1871
1872        ret = handle_mm_fault(mm, vma, address, fault_flags);
1873        if (ret & VM_FAULT_ERROR) {
1874                if (ret & VM_FAULT_OOM)
1875                        return -ENOMEM;
1876                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1877                        return -EHWPOISON;
1878                if (ret & VM_FAULT_SIGBUS)
1879                        return -EFAULT;
1880                BUG();
1881        }
1882        if (tsk) {
1883                if (ret & VM_FAULT_MAJOR)
1884                        tsk->maj_flt++;
1885                else
1886                        tsk->min_flt++;
1887        }
1888        return 0;
1889}
1890
1891/*
1892 * get_user_pages() - pin user pages in memory
1893 * @tsk:        the task_struct to use for page fault accounting, or
1894 *              NULL if faults are not to be recorded.
1895 * @mm:         mm_struct of target mm
1896 * @start:      starting user address
1897 * @nr_pages:   number of pages from start to pin
1898 * @write:      whether pages will be written to by the caller
1899 * @force:      whether to force write access even if user mapping is
1900 *              readonly. This will result in the page being COWed even
1901 *              in MAP_SHARED mappings. You do not want this.
1902 * @pages:      array that receives pointers to the pages pinned.
1903 *              Should be at least nr_pages long. Or NULL, if caller
1904 *              only intends to ensure the pages are faulted in.
1905 * @vmas:       array of pointers to vmas corresponding to each page.
1906 *              Or NULL if the caller does not require them.
1907 *
1908 * Returns number of pages pinned. This may be fewer than the number
1909 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1910 * were pinned, returns -errno. Each page returned must be released
1911 * with a put_page() call when it is finished with. vmas will only
1912 * remain valid while mmap_sem is held.
1913 *
1914 * Must be called with mmap_sem held for read or write.
1915 *
1916 * get_user_pages walks a process's page tables and takes a reference to
1917 * each struct page that each user address corresponds to at a given
1918 * instant. That is, it takes the page that would be accessed if a user
1919 * thread accesses the given user virtual address at that instant.
1920 *
1921 * This does not guarantee that the page exists in the user mappings when
1922 * get_user_pages returns, and there may even be a completely different
1923 * page there in some cases (eg. if mmapped pagecache has been invalidated
1924 * and subsequently re faulted). However it does guarantee that the page
1925 * won't be freed completely. And mostly callers simply care that the page
1926 * contains data that was valid *at some point in time*. Typically, an IO
1927 * or similar operation cannot guarantee anything stronger anyway because
1928 * locks can't be held over the syscall boundary.
1929 *
1930 * If write=0, the page must not be written to. If the page is written to,
1931 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1932 * after the page is finished with, and before put_page is called.
1933 *
1934 * get_user_pages is typically used for fewer-copy IO operations, to get a
1935 * handle on the memory by some means other than accesses via the user virtual
1936 * addresses. The pages may be submitted for DMA to devices or accessed via
1937 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1938 * use the correct cache flushing APIs.
1939 *
1940 * See also get_user_pages_fast, for performance critical applications.
1941 */
1942int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1943                unsigned long start, int nr_pages, int write, int force,
1944                struct page **pages, struct vm_area_struct **vmas)
1945{
1946        int flags = FOLL_TOUCH;
1947
1948        if (pages)
1949                flags |= FOLL_GET;
1950        if (write)
1951                flags |= FOLL_WRITE;
1952        if (force)
1953                flags |= FOLL_FORCE;
1954
1955        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1956                                NULL);
1957}
1958EXPORT_SYMBOL(get_user_pages);
1959
1960/**
1961 * get_dump_page() - pin user page in memory while writing it to core dump
1962 * @addr: user address
1963 *
1964 * Returns struct page pointer of user page pinned for dump,
1965 * to be freed afterwards by page_cache_release() or put_page().
1966 *
1967 * Returns NULL on any kind of failure - a hole must then be inserted into
1968 * the corefile, to preserve alignment with its headers; and also returns
1969 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1970 * allowing a hole to be left in the corefile to save diskspace.
1971 *
1972 * Called without mmap_sem, but after all other threads have been killed.
1973 */
1974#ifdef CONFIG_ELF_CORE
1975struct page *get_dump_page(unsigned long addr)
1976{
1977        struct vm_area_struct *vma;
1978        struct page *page;
1979
1980        if (__get_user_pages(current, current->mm, addr, 1,
1981                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1982                             NULL) < 1)
1983                return NULL;
1984        flush_cache_page(vma, addr, page_to_pfn(page));
1985        return page;
1986}
1987#endif /* CONFIG_ELF_CORE */
1988
1989pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1990                        spinlock_t **ptl)
1991{
1992        pgd_t * pgd = pgd_offset(mm, addr);
1993        pud_t * pud = pud_alloc(mm, pgd, addr);
1994        if (pud) {
1995                pmd_t * pmd = pmd_alloc(mm, pud, addr);
1996                if (pmd) {
1997                        VM_BUG_ON(pmd_trans_huge(*pmd));
1998                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
1999                }
2000        }
2001        return NULL;
2002}
2003
2004/*
2005 * This is the old fallback for page remapping.
2006 *
2007 * For historical reasons, it only allows reserved pages. Only
2008 * old drivers should use this, and they needed to mark their
2009 * pages reserved for the old functions anyway.
2010 */
2011static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2012                        struct page *page, pgprot_t prot)
2013{
2014        struct mm_struct *mm = vma->vm_mm;
2015        int retval;
2016        pte_t *pte;
2017        spinlock_t *ptl;
2018
2019        retval = -EINVAL;
2020        if (PageAnon(page))
2021                goto out;
2022        retval = -ENOMEM;
2023        flush_dcache_page(page);
2024        pte = get_locked_pte(mm, addr, &ptl);
2025        if (!pte)
2026                goto out;
2027        retval = -EBUSY;
2028        if (!pte_none(*pte))
2029                goto out_unlock;
2030
2031        /* Ok, finally just insert the thing.. */
2032        get_page(page);
2033        inc_mm_counter_fast(mm, MM_FILEPAGES);
2034        page_add_file_rmap(page);
2035        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2036
2037        retval = 0;
2038        pte_unmap_unlock(pte, ptl);
2039        return retval;
2040out_unlock:
2041        pte_unmap_unlock(pte, ptl);
2042out:
2043        return retval;
2044}
2045
2046/**
2047 * vm_insert_page - insert single page into user vma
2048 * @vma: user vma to map to
2049 * @addr: target user address of this page
2050 * @page: source kernel page
2051 *
2052 * This allows drivers to insert individual pages they've allocated
2053 * into a user vma.
2054 *
2055 * The page has to be a nice clean _individual_ kernel allocation.
2056 * If you allocate a compound page, you need to have marked it as
2057 * such (__GFP_COMP), or manually just split the page up yourself
2058 * (see split_page()).
2059 *
2060 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2061 * took an arbitrary page protection parameter. This doesn't allow
2062 * that. Your vma protection will have to be set up correctly, which
2063 * means that if you want a shared writable mapping, you'd better
2064 * ask for a shared writable mapping!
2065 *
2066 * The page does not need to be reserved.
2067 */
2068int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2069                        struct page *page)
2070{
2071        if (addr < vma->vm_start || addr >= vma->vm_end)
2072                return -EFAULT;
2073        if (!page_count(page))
2074                return -EINVAL;
2075        vma->vm_flags |= VM_INSERTPAGE;
2076        return insert_page(vma, addr, page, vma->vm_page_prot);
2077}
2078EXPORT_SYMBOL(vm_insert_page);
2079
2080static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2081                        unsigned long pfn, pgprot_t prot)
2082{
2083        struct mm_struct *mm = vma->vm_mm;
2084        int retval;
2085        pte_t *pte, entry;
2086        spinlock_t *ptl;
2087
2088        retval = -ENOMEM;
2089        pte = get_locked_pte(mm, addr, &ptl);
2090        if (!pte)
2091                goto out;
2092        retval = -EBUSY;
2093        if (!pte_none(*pte))
2094                goto out_unlock;
2095
2096        /* Ok, finally just insert the thing.. */
2097        entry = pte_mkspecial(pfn_pte(pfn, prot));
2098        set_pte_at(mm, addr, pte, entry);
2099        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2100
2101        retval = 0;
2102out_unlock:
2103        pte_unmap_unlock(pte, ptl);
2104out:
2105        return retval;
2106}
2107
2108/**
2109 * vm_insert_pfn - insert single pfn into user vma
2110 * @vma: user vma to map to
2111 * @addr: target user address of this page
2112 * @pfn: source kernel pfn
2113 *
2114 * Similar to vm_inert_page, this allows drivers to insert individual pages
2115 * they've allocated into a user vma. Same comments apply.
2116 *
2117 * This function should only be called from a vm_ops->fault handler, and
2118 * in that case the handler should return NULL.
2119 *
2120 * vma cannot be a COW mapping.
2121 *
2122 * As this is called only for pages that do not currently exist, we
2123 * do not need to flush old virtual caches or the TLB.
2124 */
2125int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2126                        unsigned long pfn)
2127{
2128        int ret;
2129        pgprot_t pgprot = vma->vm_page_prot;
2130        /*
2131         * Technically, architectures with pte_special can avoid all these
2132         * restrictions (same for remap_pfn_range).  However we would like
2133         * consistency in testing and feature parity among all, so we should
2134         * try to keep these invariants in place for everybody.
2135         */
2136        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2137        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2138                                                (VM_PFNMAP|VM_MIXEDMAP));
2139        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2140        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2141
2142        if (addr < vma->vm_start || addr >= vma->vm_end)
2143                return -EFAULT;
2144        if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2145                return -EINVAL;
2146
2147        ret = insert_pfn(vma, addr, pfn, pgprot);
2148
2149        if (ret)
2150                untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2151
2152        return ret;
2153}
2154EXPORT_SYMBOL(vm_insert_pfn);
2155
2156int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2157                        unsigned long pfn)
2158{
2159        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2160
2161        if (addr < vma->vm_start || addr >= vma->vm_end)
2162                return -EFAULT;
2163
2164        /*
2165         * If we don't have pte special, then we have to use the pfn_valid()
2166         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2167         * refcount the page if pfn_valid is true (hence insert_page rather
2168         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2169         * without pte special, it would there be refcounted as a normal page.
2170         */
2171        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2172                struct page *page;
2173
2174                page = pfn_to_page(pfn);
2175                return insert_page(vma, addr, page, vma->vm_page_prot);
2176        }
2177        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2178}
2179EXPORT_SYMBOL(vm_insert_mixed);
2180
2181/*
2182 * maps a range of physical memory into the requested pages. the old
2183 * mappings are removed. any references to nonexistent pages results
2184 * in null mappings (currently treated as "copy-on-access")
2185 */
2186static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2187                        unsigned long addr, unsigned long end,
2188                        unsigned long pfn, pgprot_t prot)
2189{
2190        pte_t *pte;
2191        spinlock_t *ptl;
2192
2193        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2194        if (!pte)
2195                return -ENOMEM;
2196        arch_enter_lazy_mmu_mode();
2197        do {
2198                BUG_ON(!pte_none(*pte));
2199                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2200                pfn++;
2201        } while (pte++, addr += PAGE_SIZE, addr != end);
2202        arch_leave_lazy_mmu_mode();
2203        pte_unmap_unlock(pte - 1, ptl);
2204        return 0;
2205}
2206
2207static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2208                        unsigned long addr, unsigned long end,
2209                        unsigned long pfn, pgprot_t prot)
2210{
2211        pmd_t *pmd;
2212        unsigned long next;
2213
2214        pfn -= addr >> PAGE_SHIFT;
2215        pmd = pmd_alloc(mm, pud, addr);
2216        if (!pmd)
2217                return -ENOMEM;
2218        VM_BUG_ON(pmd_trans_huge(*pmd));
2219        do {
2220                next = pmd_addr_end(addr, end);
2221                if (remap_pte_range(mm, pmd, addr, next,
2222                                pfn + (addr >> PAGE_SHIFT), prot))
2223                        return -ENOMEM;
2224        } while (pmd++, addr = next, addr != end);
2225        return 0;
2226}
2227
2228static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2229                        unsigned long addr, unsigned long end,
2230                        unsigned long pfn, pgprot_t prot)
2231{
2232        pud_t *pud;
2233        unsigned long next;
2234
2235        pfn -= addr >> PAGE_SHIFT;
2236        pud = pud_alloc(mm, pgd, addr);
2237        if (!pud)
2238                return -ENOMEM;
2239        do {
2240                next = pud_addr_end(addr, end);
2241                if (remap_pmd_range(mm, pud, addr, next,
2242                                pfn + (addr >> PAGE_SHIFT), prot))
2243                        return -ENOMEM;
2244        } while (pud++, addr = next, addr != end);
2245        return 0;
2246}
2247
2248/**
2249 * remap_pfn_range - remap kernel memory to userspace
2250 * @vma: user vma to map to
2251 * @addr: target user address to start at
2252 * @pfn: physical address of kernel memory
2253 * @size: size of map area
2254 * @prot: page protection flags for this mapping
2255 *
2256 *  Note: this is only safe if the mm semaphore is held when called.
2257 */
2258int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2259                    unsigned long pfn, unsigned long size, pgprot_t prot)
2260{
2261        pgd_t *pgd;
2262        unsigned long next;
2263        unsigned long end = addr + PAGE_ALIGN(size);
2264        struct mm_struct *mm = vma->vm_mm;
2265        int err;
2266
2267        /*
2268         * Physically remapped pages are special. Tell the
2269         * rest of the world about it:
2270         *   VM_IO tells people not to look at these pages
2271         *      (accesses can have side effects).
2272         *   VM_RESERVED is specified all over the place, because
2273         *      in 2.4 it kept swapout's vma scan off this vma; but
2274         *      in 2.6 the LRU scan won't even find its pages, so this
2275         *      flag means no more than count its pages in reserved_vm,
2276         *      and omit it from core dump, even when VM_IO turned off.
2277         *   VM_PFNMAP tells the core MM that the base pages are just
2278         *      raw PFN mappings, and do not have a "struct page" associated
2279         *      with them.
2280         *
2281         * There's a horrible special case to handle copy-on-write
2282         * behaviour that some programs depend on. We mark the "original"
2283         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2284         */
2285        if (addr == vma->vm_start && end == vma->vm_end) {
2286                vma->vm_pgoff = pfn;
2287                vma->vm_flags |= VM_PFN_AT_MMAP;
2288        } else if (is_cow_mapping(vma->vm_flags))
2289                return -EINVAL;
2290
2291        vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2292
2293        err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2294        if (err) {
2295                /*
2296                 * To indicate that track_pfn related cleanup is not
2297                 * needed from higher level routine calling unmap_vmas
2298                 */
2299                vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2300                vma->vm_flags &= ~VM_PFN_AT_MMAP;
2301                return -EINVAL;
2302        }
2303
2304        BUG_ON(addr >= end);
2305        pfn -= addr >> PAGE_SHIFT;
2306        pgd = pgd_offset(mm, addr);
2307        flush_cache_range(vma, addr, end);
2308        do {
2309                next = pgd_addr_end(addr, end);
2310                err = remap_pud_range(mm, pgd, addr, next,
2311                                pfn + (addr >> PAGE_SHIFT), prot);
2312                if (err)
2313                        break;
2314        } while (pgd++, addr = next, addr != end);
2315
2316        if (err)
2317                untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2318
2319        return err;
2320}
2321EXPORT_SYMBOL(remap_pfn_range);
2322
2323static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2324                                     unsigned long addr, unsigned long end,
2325                                     pte_fn_t fn, void *data)
2326{
2327        pte_t *pte;
2328        int err;
2329        pgtable_t token;
2330        spinlock_t *uninitialized_var(ptl);
2331
2332        pte = (mm == &init_mm) ?
2333                pte_alloc_kernel(pmd, addr) :
2334                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2335        if (!pte)
2336                return -ENOMEM;
2337
2338        BUG_ON(pmd_huge(*pmd));
2339
2340        arch_enter_lazy_mmu_mode();
2341
2342        token = pmd_pgtable(*pmd);
2343
2344        do {
2345                err = fn(pte++, token, addr, data);
2346                if (err)
2347                        break;
2348        } while (addr += PAGE_SIZE, addr != end);
2349
2350        arch_leave_lazy_mmu_mode();
2351
2352        if (mm != &init_mm)
2353                pte_unmap_unlock(pte-1, ptl);
2354        return err;
2355}
2356
2357static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2358                                     unsigned long addr, unsigned long end,
2359                                     pte_fn_t fn, void *data)
2360{
2361        pmd_t *pmd;
2362        unsigned long next;
2363        int err;
2364
2365        BUG_ON(pud_huge(*pud));
2366
2367        pmd = pmd_alloc(mm, pud, addr);
2368        if (!pmd)
2369                return -ENOMEM;
2370        do {
2371                next = pmd_addr_end(addr, end);
2372                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2373                if (err)
2374                        break;
2375        } while (pmd++, addr = next, addr != end);
2376        return err;
2377}
2378
2379static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2380                                     unsigned long addr, unsigned long end,
2381                                     pte_fn_t fn, void *data)
2382{
2383        pud_t *pud;
2384        unsigned long next;
2385        int err;
2386
2387        pud = pud_alloc(mm, pgd, addr);
2388        if (!pud)
2389                return -ENOMEM;
2390        do {
2391                next = pud_addr_end(addr, end);
2392                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2393                if (err)
2394                        break;
2395        } while (pud++, addr = next, addr != end);
2396        return err;
2397}
2398
2399/*
2400 * Scan a region of virtual memory, filling in page tables as necessary
2401 * and calling a provided function on each leaf page table.
2402 */
2403int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2404                        unsigned long size, pte_fn_t fn, void *data)
2405{
2406        pgd_t *pgd;
2407        unsigned long next;
2408        unsigned long end = addr + size;
2409        int err;
2410
2411        BUG_ON(addr >= end);
2412        pgd = pgd_offset(mm, addr);
2413        do {
2414                next = pgd_addr_end(addr, end);
2415                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2416                if (err)
2417                        break;
2418        } while (pgd++, addr = next, addr != end);
2419
2420        return err;
2421}
2422EXPORT_SYMBOL_GPL(apply_to_page_range);
2423
2424/*
2425 * handle_pte_fault chooses page fault handler according to an entry
2426 * which was read non-atomically.  Before making any commitment, on
2427 * those architectures or configurations (e.g. i386 with PAE) which
2428 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2429 * must check under lock before unmapping the pte and proceeding
2430 * (but do_wp_page is only called after already making such a check;
2431 * and do_anonymous_page can safely check later on).
2432 */
2433static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2434                                pte_t *page_table, pte_t orig_pte)
2435{
2436        int same = 1;
2437#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2438        if (sizeof(pte_t) > sizeof(unsigned long)) {
2439                spinlock_t *ptl = pte_lockptr(mm, pmd);
2440                spin_lock(ptl);
2441                same = pte_same(*page_table, orig_pte);
2442                spin_unlock(ptl);
2443        }
2444#endif
2445        pte_unmap(page_table);
2446        return same;
2447}
2448
2449static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2450{
2451        /*
2452         * If the source page was a PFN mapping, we don't have
2453         * a "struct page" for it. We do a best-effort copy by
2454         * just copying from the original user address. If that
2455         * fails, we just zero-fill it. Live with it.
2456         */
2457        if (unlikely(!src)) {
2458                void *kaddr = kmap_atomic(dst, KM_USER0);
2459                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2460
2461                /*
2462                 * This really shouldn't fail, because the page is there
2463                 * in the page tables. But it might just be unreadable,
2464                 * in which case we just give up and fill the result with
2465                 * zeroes.
2466                 */
2467                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2468                        clear_page(kaddr);
2469                kunmap_atomic(kaddr, KM_USER0);
2470                flush_dcache_page(dst);
2471        } else
2472                copy_user_highpage(dst, src, va, vma);
2473}
2474
2475/*
2476 * This routine handles present pages, when users try to write
2477 * to a shared page. It is done by copying the page to a new address
2478 * and decrementing the shared-page counter for the old page.
2479 *
2480 * Note that this routine assumes that the protection checks have been
2481 * done by the caller (the low-level page fault routine in most cases).
2482 * Thus we can safely just mark it writable once we've done any necessary
2483 * COW.
2484 *
2485 * We also mark the page dirty at this point even though the page will
2486 * change only once the write actually happens. This avoids a few races,
2487 * and potentially makes it more efficient.
2488 *
2489 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2490 * but allow concurrent faults), with pte both mapped and locked.
2491 * We return with mmap_sem still held, but pte unmapped and unlocked.
2492 */
2493static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2494                unsigned long address, pte_t *page_table, pmd_t *pmd,
2495                spinlock_t *ptl, pte_t orig_pte)
2496        __releases(ptl)
2497{
2498        struct page *old_page, *new_page;
2499        pte_t entry;
2500        int ret = 0;
2501        int page_mkwrite = 0;
2502        struct page *dirty_page = NULL;
2503
2504        old_page = vm_normal_page(vma, address, orig_pte);
2505        if (!old_page) {
2506                /*
2507                 * VM_MIXEDMAP !pfn_valid() case
2508                 *
2509                 * We should not cow pages in a shared writeable mapping.
2510                 * Just mark the pages writable as we can't do any dirty
2511                 * accounting on raw pfn maps.
2512                 */
2513                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2514                                     (VM_WRITE|VM_SHARED))
2515                        goto reuse;
2516                goto gotten;
2517        }
2518
2519        /*
2520         * Take out anonymous pages first, anonymous shared vmas are
2521         * not dirty accountable.
2522         */
2523        if (PageAnon(old_page) && !PageKsm(old_page)) {
2524                if (!trylock_page(old_page)) {
2525                        page_cache_get(old_page);
2526                        pte_unmap_unlock(page_table, ptl);
2527                        lock_page(old_page);
2528                        page_table = pte_offset_map_lock(mm, pmd, address,
2529                                                         &ptl);
2530                        if (!pte_same(*page_table, orig_pte)) {
2531                                unlock_page(old_page);
2532                                goto unlock;
2533                        }
2534                        page_cache_release(old_page);
2535                }
2536                if (reuse_swap_page(old_page)) {
2537                        /*
2538                         * The page is all ours.  Move it to our anon_vma so
2539                         * the rmap code will not search our parent or siblings.
2540                         * Protected against the rmap code by the page lock.
2541                         */
2542                        page_move_anon_rmap(old_page, vma, address);
2543                        unlock_page(old_page);
2544                        goto reuse;
2545                }
2546                unlock_page(old_page);
2547        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2548                                        (VM_WRITE|VM_SHARED))) {
2549                /*
2550                 * Only catch write-faults on shared writable pages,
2551                 * read-only shared pages can get COWed by
2552                 * get_user_pages(.write=1, .force=1).
2553                 */
2554                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2555                        struct vm_fault vmf;
2556                        int tmp;
2557
2558                        vmf.virtual_address = (void __user *)(address &
2559                                                                PAGE_MASK);
2560                        vmf.pgoff = old_page->index;
2561                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2562                        vmf.page = old_page;
2563
2564                        /*
2565                         * Notify the address space that the page is about to
2566                         * become writable so that it can prohibit this or wait
2567                         * for the page to get into an appropriate state.
2568                         *
2569                         * We do this without the lock held, so that it can
2570                         * sleep if it needs to.
2571                         */
2572                        page_cache_get(old_page);
2573                        pte_unmap_unlock(page_table, ptl);
2574
2575                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2576                        if (unlikely(tmp &
2577                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2578                                ret = tmp;
2579                                goto unwritable_page;
2580                        }
2581                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2582                                lock_page(old_page);
2583                                if (!old_page->mapping) {
2584                                        ret = 0; /* retry the fault */
2585                                        unlock_page(old_page);
2586                                        goto unwritable_page;
2587                                }
2588                        } else
2589                                VM_BUG_ON(!PageLocked(old_page));
2590
2591                        /*
2592                         * Since we dropped the lock we need to revalidate
2593                         * the PTE as someone else may have changed it.  If
2594                         * they did, we just return, as we can count on the
2595                         * MMU to tell us if they didn't also make it writable.
2596                         */
2597                        page_table = pte_offset_map_lock(mm, pmd, address,
2598                                                         &ptl);
2599                        if (!pte_same(*page_table, orig_pte)) {
2600                                unlock_page(old_page);
2601                                goto unlock;
2602                        }
2603
2604                        page_mkwrite = 1;
2605                }
2606                dirty_page = old_page;
2607                get_page(dirty_page);
2608
2609reuse:
2610                flush_cache_page(vma, address, pte_pfn(orig_pte));
2611                entry = pte_mkyoung(orig_pte);
2612                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2613                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2614                        update_mmu_cache(vma, address, page_table);
2615                pte_unmap_unlock(page_table, ptl);
2616                ret |= VM_FAULT_WRITE;
2617
2618                if (!dirty_page)
2619                        return ret;
2620
2621                /*
2622                 * Yes, Virginia, this is actually required to prevent a race
2623                 * with clear_page_dirty_for_io() from clearing the page dirty
2624                 * bit after it clear all dirty ptes, but before a racing
2625                 * do_wp_page installs a dirty pte.
2626                 *
2627                 * __do_fault is protected similarly.
2628                 */
2629                if (!page_mkwrite) {
2630                        wait_on_page_locked(dirty_page);
2631                        set_page_dirty_balance(dirty_page, page_mkwrite);
2632                }
2633                put_page(dirty_page);
2634                if (page_mkwrite) {
2635                        struct address_space *mapping = dirty_page->mapping;
2636
2637                        set_page_dirty(dirty_page);
2638                        unlock_page(dirty_page);
2639                        page_cache_release(dirty_page);
2640                        if (mapping)    {
2641                                /*
2642                                 * Some device drivers do not set page.mapping
2643                                 * but still dirty their pages
2644                                 */
2645                                balance_dirty_pages_ratelimited(mapping);
2646                        }
2647                }
2648
2649                /* file_update_time outside page_lock */
2650                if (vma->vm_file)
2651                        file_update_time(vma->vm_file);
2652
2653                return ret;
2654        }
2655
2656        /*
2657         * Ok, we need to copy. Oh, well..
2658         */
2659        page_cache_get(old_page);
2660gotten:
2661        pte_unmap_unlock(page_table, ptl);
2662
2663        if (unlikely(anon_vma_prepare(vma)))
2664                goto oom;
2665
2666        if (is_zero_pfn(pte_pfn(orig_pte))) {
2667                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2668                if (!new_page)
2669                        goto oom;
2670        } else {
2671                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2672                if (!new_page)
2673                        goto oom;
2674                cow_user_page(new_page, old_page, address, vma);
2675        }
2676        __SetPageUptodate(new_page);
2677
2678        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2679                goto oom_free_new;
2680
2681        /*
2682         * Re-check the pte - we dropped the lock
2683         */
2684        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2685        if (likely(pte_same(*page_table, orig_pte))) {
2686                if (old_page) {
2687                        if (!PageAnon(old_page)) {
2688                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2689                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2690                        }
2691                } else
2692                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2693                flush_cache_page(vma, address, pte_pfn(orig_pte));
2694                entry = mk_pte(new_page, vma->vm_page_prot);
2695                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2696                /*
2697                 * Clear the pte entry and flush it first, before updating the
2698                 * pte with the new entry. This will avoid a race condition
2699                 * seen in the presence of one thread doing SMC and another
2700                 * thread doing COW.
2701                 */
2702                ptep_clear_flush(vma, address, page_table);
2703                page_add_new_anon_rmap(new_page, vma, address);
2704                /*
2705                 * We call the notify macro here because, when using secondary
2706                 * mmu page tables (such as kvm shadow page tables), we want the
2707                 * new page to be mapped directly into the secondary page table.
2708                 */
2709                set_pte_at_notify(mm, address, page_table, entry);
2710                update_mmu_cache(vma, address, page_table);
2711                if (old_page) {
2712                        /*
2713                         * Only after switching the pte to the new page may
2714                         * we remove the mapcount here. Otherwise another
2715                         * process may come and find the rmap count decremented
2716                         * before the pte is switched to the new page, and
2717                         * "reuse" the old page writing into it while our pte
2718                         * here still points into it and can be read by other
2719                         * threads.
2720                         *
2721                         * The critical issue is to order this
2722                         * page_remove_rmap with the ptp_clear_flush above.
2723                         * Those stores are ordered by (if nothing else,)
2724                         * the barrier present in the atomic_add_negative
2725                         * in page_remove_rmap.
2726                         *
2727                         * Then the TLB flush in ptep_clear_flush ensures that
2728                         * no process can access the old page before the
2729                         * decremented mapcount is visible. And the old page
2730                         * cannot be reused until after the decremented
2731                         * mapcount is visible. So transitively, TLBs to
2732                         * old page will be flushed before it can be reused.
2733                         */
2734                        page_remove_rmap(old_page);
2735                }
2736
2737                /* Free the old page.. */
2738                new_page = old_page;
2739                ret |= VM_FAULT_WRITE;
2740        } else
2741                mem_cgroup_uncharge_page(new_page);
2742
2743        if (new_page)
2744                page_cache_release(new_page);
2745unlock:
2746        pte_unmap_unlock(page_table, ptl);
2747        if (old_page) {
2748                /*
2749                 * Don't let another task, with possibly unlocked vma,
2750                 * keep the mlocked page.
2751                 */
2752                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2753                        lock_page(old_page);    /* LRU manipulation */
2754                        munlock_vma_page(old_page);
2755                        unlock_page(old_page);
2756                }
2757                page_cache_release(old_page);
2758        }
2759        return ret;
2760oom_free_new:
2761        page_cache_release(new_page);
2762oom:
2763        if (old_page) {
2764                if (page_mkwrite) {
2765                        unlock_page(old_page);
2766                        page_cache_release(old_page);
2767                }
2768                page_cache_release(old_page);
2769        }
2770        return VM_FAULT_OOM;
2771
2772unwritable_page:
2773        page_cache_release(old_page);
2774        return ret;
2775}
2776
2777static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2778                unsigned long start_addr, unsigned long end_addr,
2779                struct zap_details *details)
2780{
2781        zap_page_range(vma, start_addr, end_addr - start_addr, details);
2782}
2783
2784static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2785                                            struct zap_details *details)
2786{
2787        struct vm_area_struct *vma;
2788        struct prio_tree_iter iter;
2789        pgoff_t vba, vea, zba, zea;
2790
2791        vma_prio_tree_foreach(vma, &iter, root,
2792                        details->first_index, details->last_index) {
2793
2794                vba = vma->vm_pgoff;
2795                vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2796                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2797                zba = details->first_index;
2798                if (zba < vba)
2799                        zba = vba;
2800                zea = details->last_index;
2801                if (zea > vea)
2802                        zea = vea;
2803
2804                unmap_mapping_range_vma(vma,
2805                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2806                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2807                                details);
2808        }
2809}
2810
2811static inline void unmap_mapping_range_list(struct list_head *head,
2812                                            struct zap_details *details)
2813{
2814        struct vm_area_struct *vma;
2815
2816        /*
2817         * In nonlinear VMAs there is no correspondence between virtual address
2818         * offset and file offset.  So we must perform an exhaustive search
2819         * across *all* the pages in each nonlinear VMA, not just the pages
2820         * whose virtual address lies outside the file truncation point.
2821         */
2822        list_for_each_entry(vma, head, shared.vm_set.list) {
2823                details->nonlinear_vma = vma;
2824                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2825        }
2826}
2827
2828/**
2829 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2830 * @mapping: the address space containing mmaps to be unmapped.
2831 * @holebegin: byte in first page to unmap, relative to the start of
2832 * the underlying file.  This will be rounded down to a PAGE_SIZE
2833 * boundary.  Note that this is different from truncate_pagecache(), which
2834 * must keep the partial page.  In contrast, we must get rid of
2835 * partial pages.
2836 * @holelen: size of prospective hole in bytes.  This will be rounded
2837 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2838 * end of the file.
2839 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2840 * but 0 when invalidating pagecache, don't throw away private data.
2841 */
2842void unmap_mapping_range(struct address_space *mapping,
2843                loff_t const holebegin, loff_t const holelen, int even_cows)
2844{
2845        struct zap_details details;
2846        pgoff_t hba = holebegin >> PAGE_SHIFT;
2847        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2848
2849        /* Check for overflow. */
2850        if (sizeof(holelen) > sizeof(hlen)) {
2851                long long holeend =
2852                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2853                if (holeend & ~(long long)ULONG_MAX)
2854                        hlen = ULONG_MAX - hba + 1;
2855        }
2856
2857        details.check_mapping = even_cows? NULL: mapping;
2858        details.nonlinear_vma = NULL;
2859        details.first_index = hba;
2860        details.last_index = hba + hlen - 1;
2861        if (details.last_index < details.first_index)
2862                details.last_index = ULONG_MAX;
2863
2864
2865        mutex_lock(&mapping->i_mmap_mutex);
2866        if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2867                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2868        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2869                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2870        mutex_unlock(&mapping->i_mmap_mutex);
2871}
2872EXPORT_SYMBOL(unmap_mapping_range);
2873
2874/*
2875 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2876 * but allow concurrent faults), and pte mapped but not yet locked.
2877 * We return with mmap_sem still held, but pte unmapped and unlocked.
2878 */
2879static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2880                unsigned long address, pte_t *page_table, pmd_t *pmd,
2881                unsigned int flags, pte_t orig_pte)
2882{
2883        spinlock_t *ptl;
2884        struct page *page, *swapcache = NULL;
2885        swp_entry_t entry;
2886        pte_t pte;
2887        int locked;
2888        struct mem_cgroup *ptr;
2889        int exclusive = 0;
2890        int ret = 0;
2891
2892        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2893                goto out;
2894
2895        entry = pte_to_swp_entry(orig_pte);
2896        if (unlikely(non_swap_entry(entry))) {
2897                if (is_migration_entry(entry)) {
2898                        migration_entry_wait(mm, pmd, address);
2899                } else if (is_hwpoison_entry(entry)) {
2900                        ret = VM_FAULT_HWPOISON;
2901                } else {
2902                        print_bad_pte(vma, address, orig_pte, NULL);
2903                        ret = VM_FAULT_SIGBUS;
2904                }
2905                goto out;
2906        }
2907        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2908        page = lookup_swap_cache(entry);
2909        if (!page) {
2910                grab_swap_token(mm); /* Contend for token _before_ read-in */
2911                page = swapin_readahead(entry,
2912                                        GFP_HIGHUSER_MOVABLE, vma, address);
2913                if (!page) {
2914                        /*
2915                         * Back out if somebody else faulted in this pte
2916                         * while we released the pte lock.
2917                         */
2918                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2919                        if (likely(pte_same(*page_table, orig_pte)))
2920                                ret = VM_FAULT_OOM;
2921                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2922                        goto unlock;
2923                }
2924
2925                /* Had to read the page from swap area: Major fault */
2926                ret = VM_FAULT_MAJOR;
2927                count_vm_event(PGMAJFAULT);
2928                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2929        } else if (PageHWPoison(page)) {
2930                /*
2931                 * hwpoisoned dirty swapcache pages are kept for killing
2932                 * owner processes (which may be unknown at hwpoison time)
2933                 */
2934                ret = VM_FAULT_HWPOISON;
2935                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2936                goto out_release;
2937        }
2938
2939        locked = lock_page_or_retry(page, mm, flags);
2940        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2941        if (!locked) {
2942                ret |= VM_FAULT_RETRY;
2943                goto out_release;
2944        }
2945
2946        /*
2947         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2948         * release the swapcache from under us.  The page pin, and pte_same
2949         * test below, are not enough to exclude that.  Even if it is still
2950         * swapcache, we need to check that the page's swap has not changed.
2951         */
2952        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2953                goto out_page;
2954
2955        if (ksm_might_need_to_copy(page, vma, address)) {
2956                swapcache = page;
2957                page = ksm_does_need_to_copy(page, vma, address);
2958
2959                if (unlikely(!page)) {
2960                        ret = VM_FAULT_OOM;
2961                        page = swapcache;
2962                        swapcache = NULL;
2963                        goto out_page;
2964                }
2965        }
2966
2967        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2968                ret = VM_FAULT_OOM;
2969                goto out_page;
2970        }
2971
2972        /*
2973         * Back out if somebody else already faulted in this pte.
2974         */
2975        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2976        if (unlikely(!pte_same(*page_table, orig_pte)))
2977                goto out_nomap;
2978
2979        if (unlikely(!PageUptodate(page))) {
2980                ret = VM_FAULT_SIGBUS;
2981                goto out_nomap;
2982        }
2983
2984        /*
2985         * The page isn't present yet, go ahead with the fault.
2986         *
2987         * Be careful about the sequence of operations here.
2988         * To get its accounting right, reuse_swap_page() must be called
2989         * while the page is counted on swap but not yet in mapcount i.e.
2990         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2991         * must be called after the swap_free(), or it will never succeed.
2992         * Because delete_from_swap_page() may be called by reuse_swap_page(),
2993         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2994         * in page->private. In this case, a record in swap_cgroup  is silently
2995         * discarded at swap_free().
2996         */
2997
2998        inc_mm_counter_fast(mm, MM_ANONPAGES);
2999        dec_mm_counter_fast(mm, MM_SWAPENTS);
3000        pte = mk_pte(page, vma->vm_page_prot);
3001        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3002                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3003                flags &= ~FAULT_FLAG_WRITE;
3004                ret |= VM_FAULT_WRITE;
3005                exclusive = 1;
3006        }
3007        flush_icache_page(vma, page);
3008        set_pte_at(mm, address, page_table, pte);
3009        do_page_add_anon_rmap(page, vma, address, exclusive);
3010        /* It's better to call commit-charge after rmap is established */
3011        mem_cgroup_commit_charge_swapin(page, ptr);
3012
3013        swap_free(entry);
3014        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3015                try_to_free_swap(page);
3016        unlock_page(page);
3017        if (swapcache) {
3018                /*
3019                 * Hold the lock to avoid the swap entry to be reused
3020                 * until we take the PT lock for the pte_same() check
3021                 * (to avoid false positives from pte_same). For
3022                 * further safety release the lock after the swap_free
3023                 * so that the swap count won't change under a
3024                 * parallel locked swapcache.
3025                 */
3026                unlock_page(swapcache);
3027                page_cache_release(swapcache);
3028        }
3029
3030        if (flags & FAULT_FLAG_WRITE) {
3031                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3032                if (ret & VM_FAULT_ERROR)
3033                        ret &= VM_FAULT_ERROR;
3034                goto out;
3035        }
3036
3037        /* No need to invalidate - it was non-present before */
3038        update_mmu_cache(vma, address, page_table);
3039unlock:
3040        pte_unmap_unlock(page_table, ptl);
3041out:
3042        return ret;
3043out_nomap:
3044        mem_cgroup_cancel_charge_swapin(ptr);
3045        pte_unmap_unlock(page_table, ptl);
3046out_page:
3047        unlock_page(page);
3048out_release:
3049        page_cache_release(page);
3050        if (swapcache) {
3051                unlock_page(swapcache);
3052                page_cache_release(swapcache);
3053        }
3054        return ret;
3055}
3056
3057/*
3058 * This is like a special single-page "expand_{down|up}wards()",
3059 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3060 * doesn't hit another vma.
3061 */
3062static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3063{
3064        address &= PAGE_MASK;
3065        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3066                struct vm_area_struct *prev = vma->vm_prev;
3067
3068                /*
3069                 * Is there a mapping abutting this one below?
3070                 *
3071                 * That's only ok if it's the same stack mapping
3072                 * that has gotten split..
3073                 */
3074                if (prev && prev->vm_end == address)
3075                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3076
3077                expand_downwards(vma, address - PAGE_SIZE);
3078        }
3079        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3080                struct vm_area_struct *next = vma->vm_next;
3081
3082                /* As VM_GROWSDOWN but s/below/above/ */
3083                if (next && next->vm_start == address + PAGE_SIZE)
3084                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3085
3086                expand_upwards(vma, address + PAGE_SIZE);
3087        }
3088        return 0;
3089}
3090
3091/*
3092 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3093 * but allow concurrent faults), and pte mapped but not yet locked.
3094 * We return with mmap_sem still held, but pte unmapped and unlocked.
3095 */
3096static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3097                unsigned long address, pte_t *page_table, pmd_t *pmd,
3098                unsigned int flags)
3099{
3100        struct page *page;
3101        spinlock_t *ptl;
3102        pte_t entry;
3103
3104        pte_unmap(page_table);
3105
3106        /* Check if we need to add a guard page to the stack */
3107        if (check_stack_guard_page(vma, address) < 0)
3108                return VM_FAULT_SIGBUS;
3109
3110        /* Use the zero-page for reads */
3111        if (!(flags & FAULT_FLAG_WRITE)) {
3112                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3113                                                vma->vm_page_prot));
3114                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3115                if (!pte_none(*page_table))
3116                        goto unlock;
3117                goto setpte;
3118        }
3119
3120        /* Allocate our own private page. */
3121        if (unlikely(anon_vma_prepare(vma)))
3122                goto oom;
3123        page = alloc_zeroed_user_highpage_movable(vma, address);
3124        if (!page)
3125                goto oom;
3126        __SetPageUptodate(page);
3127
3128        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3129                goto oom_free_page;
3130
3131        entry = mk_pte(page, vma->vm_page_prot);
3132        if (vma->vm_flags & VM_WRITE)
3133                entry = pte_mkwrite(pte_mkdirty(entry));
3134
3135        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3136        if (!pte_none(*page_table))
3137                goto release;
3138
3139        inc_mm_counter_fast(mm, MM_ANONPAGES);
3140        page_add_new_anon_rmap(page, vma, address);
3141setpte:
3142        set_pte_at(mm, address, page_table, entry);
3143
3144        /* No need to invalidate - it was non-present before */
3145        update_mmu_cache(vma, address, page_table);
3146unlock:
3147        pte_unmap_unlock(page_table, ptl);
3148        return 0;
3149release:
3150        mem_cgroup_uncharge_page(page);
3151        page_cache_release(page);
3152        goto unlock;
3153oom_free_page:
3154        page_cache_release(page);
3155oom:
3156        return VM_FAULT_OOM;
3157}
3158
3159/*
3160 * __do_fault() tries to create a new page mapping. It aggressively
3161 * tries to share with existing pages, but makes a separate copy if
3162 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3163 * the next page fault.
3164 *
3165 * As this is called only for pages that do not currently exist, we
3166 * do not need to flush old virtual caches or the TLB.
3167 *
3168 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3169 * but allow concurrent faults), and pte neither mapped nor locked.
3170 * We return with mmap_sem still held, but pte unmapped and unlocked.
3171 */
3172static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3173                unsigned long address, pmd_t *pmd,
3174                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3175{
3176        pte_t *page_table;
3177        spinlock_t *ptl;
3178        struct page *page;
3179        struct page *cow_page;
3180        pte_t entry;
3181        int anon = 0;
3182        struct page *dirty_page = NULL;
3183        struct vm_fault vmf;
3184        int ret;
3185        int page_mkwrite = 0;
3186
3187        /*
3188         * If we do COW later, allocate page befor taking lock_page()
3189         * on the file cache page. This will reduce lock holding time.
3190         */
3191        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3192
3193                if (unlikely(anon_vma_prepare(vma)))
3194                        return VM_FAULT_OOM;
3195
3196                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3197                if (!cow_page)
3198                        return VM_FAULT_OOM;
3199
3200                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3201                        page_cache_release(cow_page);
3202                        return VM_FAULT_OOM;
3203                }
3204        } else
3205                cow_page = NULL;
3206
3207        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3208        vmf.pgoff = pgoff;
3209        vmf.flags = flags;
3210        vmf.page = NULL;
3211
3212        ret = vma->vm_ops->fault(vma, &vmf);
3213        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3214                            VM_FAULT_RETRY)))
3215                goto uncharge_out;
3216
3217        if (unlikely(PageHWPoison(vmf.page))) {
3218                if (ret & VM_FAULT_LOCKED)
3219                        unlock_page(vmf.page);
3220                ret = VM_FAULT_HWPOISON;
3221                goto uncharge_out;
3222        }
3223
3224        /*
3225         * For consistency in subsequent calls, make the faulted page always
3226         * locked.
3227         */
3228        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3229                lock_page(vmf.page);
3230        else
3231                VM_BUG_ON(!PageLocked(vmf.page));
3232
3233        /*
3234         * Should we do an early C-O-W break?
3235         */
3236        page = vmf.page;
3237        if (flags & FAULT_FLAG_WRITE) {
3238                if (!(vma->vm_flags & VM_SHARED)) {
3239                        page = cow_page;
3240                        anon = 1;
3241                        copy_user_highpage(page, vmf.page, address, vma);
3242                        __SetPageUptodate(page);
3243                } else {
3244                        /*
3245                         * If the page will be shareable, see if the backing
3246                         * address space wants to know that the page is about
3247                         * to become writable
3248                         */
3249                        if (vma->vm_ops->page_mkwrite) {
3250                                int tmp;
3251
3252                                unlock_page(page);
3253                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3254                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3255                                if (unlikely(tmp &
3256                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3257                                        ret = tmp;
3258                                        goto unwritable_page;
3259                                }
3260                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3261                                        lock_page(page);
3262                                        if (!page->mapping) {
3263                                                ret = 0; /* retry the fault */
3264                                                unlock_page(page);
3265                                                goto unwritable_page;
3266                                        }
3267                                } else
3268                                        VM_BUG_ON(!PageLocked(page));
3269                                page_mkwrite = 1;
3270                        }
3271                }
3272
3273        }
3274
3275        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3276
3277        /*
3278         * This silly early PAGE_DIRTY setting removes a race
3279         * due to the bad i386 page protection. But it's valid
3280         * for other architectures too.
3281         *
3282         * Note that if FAULT_FLAG_WRITE is set, we either now have
3283         * an exclusive copy of the page, or this is a shared mapping,
3284         * so we can make it writable and dirty to avoid having to
3285         * handle that later.
3286         */
3287        /* Only go through if we didn't race with anybody else... */
3288        if (likely(pte_same(*page_table, orig_pte))) {
3289                flush_icache_page(vma, page);
3290                entry = mk_pte(page, vma->vm_page_prot);
3291                if (flags & FAULT_FLAG_WRITE)
3292                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3293                if (anon) {
3294                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3295                        page_add_new_anon_rmap(page, vma, address);
3296                } else {
3297                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3298                        page_add_file_rmap(page);
3299                        if (flags & FAULT_FLAG_WRITE) {
3300                                dirty_page = page;
3301                                get_page(dirty_page);
3302                        }
3303                }
3304                set_pte_at(mm, address, page_table, entry);
3305
3306                /* no need to invalidate: a not-present page won't be cached */
3307                update_mmu_cache(vma, address, page_table);
3308        } else {
3309                if (cow_page)
3310                        mem_cgroup_uncharge_page(cow_page);
3311                if (anon)
3312                        page_cache_release(page);
3313                else
3314                        anon = 1; /* no anon but release faulted_page */
3315        }
3316
3317        pte_unmap_unlock(page_table, ptl);
3318
3319        if (dirty_page) {
3320                struct address_space *mapping = page->mapping;
3321
3322                if (set_page_dirty(dirty_page))
3323                        page_mkwrite = 1;
3324                unlock_page(dirty_page);
3325                put_page(dirty_page);
3326                if (page_mkwrite && mapping) {
3327                        /*
3328                         * Some device drivers do not set page.mapping but still
3329                         * dirty their pages
3330                         */
3331                        balance_dirty_pages_ratelimited(mapping);
3332                }
3333
3334                /* file_update_time outside page_lock */
3335                if (vma->vm_file)
3336                        file_update_time(vma->vm_file);
3337        } else {
3338                unlock_page(vmf.page);
3339                if (anon)
3340                        page_cache_release(vmf.page);
3341        }
3342
3343        return ret;
3344
3345unwritable_page:
3346        page_cache_release(page);
3347        return ret;
3348uncharge_out:
3349        /* fs's fault handler get error */
3350        if (cow_page) {
3351                mem_cgroup_uncharge_page(cow_page);
3352                page_cache_release(cow_page);
3353        }
3354        return ret;
3355}
3356
3357static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3358                unsigned long address, pte_t *page_table, pmd_t *pmd,
3359                unsigned int flags, pte_t orig_pte)
3360{
3361        pgoff_t pgoff = (((address & PAGE_MASK)
3362                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3363
3364        pte_unmap(page_table);
3365        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3366}
3367
3368/*
3369 * Fault of a previously existing named mapping. Repopulate the pte
3370 * from the encoded file_pte if possible. This enables swappable
3371 * nonlinear vmas.
3372 *
3373 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3374 * but allow concurrent faults), and pte mapped but not yet locked.
3375 * We return with mmap_sem still held, but pte unmapped and unlocked.
3376 */
3377static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3378                unsigned long address, pte_t *page_table, pmd_t *pmd,
3379                unsigned int flags, pte_t orig_pte)
3380{
3381        pgoff_t pgoff;
3382
3383        flags |= FAULT_FLAG_NONLINEAR;
3384
3385        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3386                return 0;
3387
3388        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3389                /*
3390                 * Page table corrupted: show pte and kill process.
3391                 */
3392                print_bad_pte(vma, address, orig_pte, NULL);
3393                return VM_FAULT_SIGBUS;
3394        }
3395
3396        pgoff = pte_to_pgoff(orig_pte);
3397        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3398}
3399
3400/*
3401 * These routines also need to handle stuff like marking pages dirty
3402 * and/or accessed for architectures that don't do it in hardware (most
3403 * RISC architectures).  The early dirtying is also good on the i386.
3404 *
3405 * There is also a hook called "update_mmu_cache()" that architectures
3406 * with external mmu caches can use to update those (ie the Sparc or
3407 * PowerPC hashed page tables that act as extended TLBs).
3408 *
3409 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3410 * but allow concurrent faults), and pte mapped but not yet locked.
3411 * We return with mmap_sem still held, but pte unmapped and unlocked.
3412 */
3413int handle_pte_fault(struct mm_struct *mm,
3414                     struct vm_area_struct *vma, unsigned long address,
3415                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3416{
3417        pte_t entry;
3418        spinlock_t *ptl;
3419
3420        entry = *pte;
3421        if (!pte_present(entry)) {
3422                if (pte_none(entry)) {
3423                        if (vma->vm_ops) {
3424                                if (likely(vma->vm_ops->fault))
3425                                        return do_linear_fault(mm, vma, address,
3426                                                pte, pmd, flags, entry);
3427                        }
3428                        return do_anonymous_page(mm, vma, address,
3429                                                 pte, pmd, flags);
3430                }
3431                if (pte_file(entry))
3432                        return do_nonlinear_fault(mm, vma, address,
3433                                        pte, pmd, flags, entry);
3434                return do_swap_page(mm, vma, address,
3435                                        pte, pmd, flags, entry);
3436        }
3437
3438        ptl = pte_lockptr(mm, pmd);
3439        spin_lock(ptl);
3440        if (unlikely(!pte_same(*pte, entry)))
3441                goto unlock;
3442        if (flags & FAULT_FLAG_WRITE) {
3443                if (!pte_write(entry))
3444                        return do_wp_page(mm, vma, address,
3445                                        pte, pmd, ptl, entry);
3446                entry = pte_mkdirty(entry);
3447        }
3448        entry = pte_mkyoung(entry);
3449        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3450                update_mmu_cache(vma, address, pte);
3451        } else {
3452                /*
3453                 * This is needed only for protection faults but the arch code
3454                 * is not yet telling us if this is a protection fault or not.
3455                 * This still avoids useless tlb flushes for .text page faults
3456                 * with threads.
3457                 */
3458                if (flags & FAULT_FLAG_WRITE)
3459                        flush_tlb_fix_spurious_fault(vma, address);
3460        }
3461unlock:
3462        pte_unmap_unlock(pte, ptl);
3463        return 0;
3464}
3465
3466/*
3467 * By the time we get here, we already hold the mm semaphore
3468 */
3469int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3470                unsigned long address, unsigned int flags)
3471{
3472        pgd_t *pgd;
3473        pud_t *pud;
3474        pmd_t *pmd;
3475        pte_t *pte;
3476
3477        __set_current_state(TASK_RUNNING);
3478
3479        count_vm_event(PGFAULT);
3480        mem_cgroup_count_vm_event(mm, PGFAULT);
3481
3482        /* do counter updates before entering really critical section. */
3483        check_sync_rss_stat(current);
3484
3485        if (unlikely(is_vm_hugetlb_page(vma)))
3486                return hugetlb_fault(mm, vma, address, flags);
3487
3488        pgd = pgd_offset(mm, address);
3489        pud = pud_alloc(mm, pgd, address);
3490        if (!pud)
3491                return VM_FAULT_OOM;
3492        pmd = pmd_alloc(mm, pud, address);
3493        if (!pmd)
3494                return VM_FAULT_OOM;
3495        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3496                if (!vma->vm_ops)
3497                        return do_huge_pmd_anonymous_page(mm, vma, address,
3498                                                          pmd, flags);
3499        } else {
3500                pmd_t orig_pmd = *pmd;
3501                barrier();
3502                if (pmd_trans_huge(orig_pmd)) {
3503                        if (flags & FAULT_FLAG_WRITE &&
3504                            !pmd_write(orig_pmd) &&
3505                            !pmd_trans_splitting(orig_pmd))
3506                                return do_huge_pmd_wp_page(mm, vma, address,
3507                                                           pmd, orig_pmd);
3508                        return 0;
3509                }
3510        }
3511
3512        /*
3513         * Use __pte_alloc instead of pte_alloc_map, because we can't
3514         * run pte_offset_map on the pmd, if an huge pmd could
3515         * materialize from under us from a different thread.
3516         */
3517        if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3518                return VM_FAULT_OOM;
3519        /* if an huge pmd materialized from under us just retry later */
3520        if (unlikely(pmd_trans_huge(*pmd)))
3521                return 0;
3522        /*
3523         * A regular pmd is established and it can't morph into a huge pmd
3524         * from under us anymore at this point because we hold the mmap_sem
3525         * read mode and khugepaged takes it in write mode. So now it's
3526         * safe to run pte_offset_map().
3527         */
3528        pte = pte_offset_map(pmd, address);
3529
3530        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3531}
3532
3533#ifndef __PAGETABLE_PUD_FOLDED
3534/*
3535 * Allocate page upper directory.
3536 * We've already handled the fast-path in-line.
3537 */
3538int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3539{
3540        pud_t *new = pud_alloc_one(mm, address);
3541        if (!new)
3542                return -ENOMEM;
3543
3544        smp_wmb(); /* See comment in __pte_alloc */
3545
3546        spin_lock(&mm->page_table_lock);
3547        if (pgd_present(*pgd))          /* Another has populated it */
3548                pud_free(mm, new);
3549        else
3550                pgd_populate(mm, pgd, new);
3551        spin_unlock(&mm->page_table_lock);
3552        return 0;
3553}
3554#endif /* __PAGETABLE_PUD_FOLDED */
3555
3556#ifndef __PAGETABLE_PMD_FOLDED
3557/*
3558 * Allocate page middle directory.
3559 * We've already handled the fast-path in-line.
3560 */
3561int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3562{
3563        pmd_t *new = pmd_alloc_one(mm, address);
3564        if (!new)
3565                return -ENOMEM;
3566
3567        smp_wmb(); /* See comment in __pte_alloc */
3568
3569        spin_lock(&mm->page_table_lock);
3570#ifndef __ARCH_HAS_4LEVEL_HACK
3571        if (pud_present(*pud))          /* Another has populated it */
3572                pmd_free(mm, new);
3573        else
3574                pud_populate(mm, pud, new);
3575#else
3576        if (pgd_present(*pud))          /* Another has populated it */
3577                pmd_free(mm, new);
3578        else
3579                pgd_populate(mm, pud, new);
3580#endif /* __ARCH_HAS_4LEVEL_HACK */
3581        spin_unlock(&mm->page_table_lock);
3582        return 0;
3583}
3584#endif /* __PAGETABLE_PMD_FOLDED */
3585
3586int make_pages_present(unsigned long addr, unsigned long end)
3587{
3588        int ret, len, write;
3589        struct vm_area_struct * vma;
3590
3591        vma = find_vma(current->mm, addr);
3592        if (!vma)
3593                return -ENOMEM;
3594        /*
3595         * We want to touch writable mappings with a write fault in order
3596         * to break COW, except for shared mappings because these don't COW
3597         * and we would not want to dirty them for nothing.
3598         */
3599        write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3600        BUG_ON(addr >= end);
3601        BUG_ON(end > vma->vm_end);
3602        len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3603        ret = get_user_pages(current, current->mm, addr,
3604                        len, write, 0, NULL, NULL);
3605        if (ret < 0)
3606                return ret;
3607        return ret == len ? 0 : -EFAULT;
3608}
3609
3610#if !defined(__HAVE_ARCH_GATE_AREA)
3611
3612#if defined(AT_SYSINFO_EHDR)
3613static struct vm_area_struct gate_vma;
3614
3615static int __init gate_vma_init(void)
3616{
3617        gate_vma.vm_mm = NULL;
3618        gate_vma.vm_start = FIXADDR_USER_START;
3619        gate_vma.vm_end = FIXADDR_USER_END;
3620        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3621        gate_vma.vm_page_prot = __P101;
3622        /*
3623         * Make sure the vDSO gets into every core dump.
3624         * Dumping its contents makes post-mortem fully interpretable later
3625         * without matching up the same kernel and hardware config to see
3626         * what PC values meant.
3627         */
3628        gate_vma.vm_flags |= VM_ALWAYSDUMP;
3629        return 0;
3630}
3631__initcall(gate_vma_init);
3632#endif
3633
3634struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3635{
3636#ifdef AT_SYSINFO_EHDR
3637        return &gate_vma;
3638#else
3639        return NULL;
3640#endif
3641}
3642
3643int in_gate_area_no_mm(unsigned long addr)
3644{
3645#ifdef AT_SYSINFO_EHDR
3646        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3647                return 1;
3648#endif
3649        return 0;
3650}
3651
3652#endif  /* __HAVE_ARCH_GATE_AREA */
3653
3654static int __follow_pte(struct mm_struct *mm, unsigned long address,
3655                pte_t **ptepp, spinlock_t **ptlp)
3656{
3657        pgd_t *pgd;
3658        pud_t *pud;
3659        pmd_t *pmd;
3660        pte_t *ptep;
3661
3662        pgd = pgd_offset(mm, address);
3663        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3664                goto out;
3665
3666        pud = pud_offset(pgd, address);
3667        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3668                goto out;
3669
3670        pmd = pmd_offset(pud, address);
3671        VM_BUG_ON(pmd_trans_huge(*pmd));
3672        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3673                goto out;
3674
3675        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3676        if (pmd_huge(*pmd))
3677                goto out;
3678
3679        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3680        if (!ptep)
3681                goto out;
3682        if (!pte_present(*ptep))
3683                goto unlock;
3684        *ptepp = ptep;
3685        return 0;
3686unlock:
3687        pte_unmap_unlock(ptep, *ptlp);
3688out:
3689        return -EINVAL;
3690}
3691
3692static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3693                             pte_t **ptepp, spinlock_t **ptlp)
3694{
3695        int res;
3696
3697        /* (void) is needed to make gcc happy */
3698        (void) __cond_lock(*ptlp,
3699                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
3700        return res;
3701}
3702
3703/**
3704 * follow_pfn - look up PFN at a user virtual address
3705 * @vma: memory mapping
3706 * @address: user virtual address
3707 * @pfn: location to store found PFN
3708 *
3709 * Only IO mappings and raw PFN mappings are allowed.
3710 *
3711 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3712 */
3713int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3714        unsigned long *pfn)
3715{
3716        int ret = -EINVAL;
3717        spinlock_t *ptl;
3718        pte_t *ptep;
3719
3720        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3721                return ret;
3722
3723        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3724        if (ret)
3725                return ret;
3726        *pfn = pte_pfn(*ptep);
3727        pte_unmap_unlock(ptep, ptl);
3728        return 0;
3729}
3730EXPORT_SYMBOL(follow_pfn);
3731
3732#ifdef CONFIG_HAVE_IOREMAP_PROT
3733int follow_phys(struct vm_area_struct *vma,
3734                unsigned long address, unsigned int flags,
3735                unsigned long *prot, resource_size_t *phys)
3736{
3737        int ret = -EINVAL;
3738        pte_t *ptep, pte;
3739        spinlock_t *ptl;
3740
3741        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3742                goto out;
3743
3744        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3745                goto out;
3746        pte = *ptep;
3747
3748        if ((flags & FOLL_WRITE) && !pte_write(pte))
3749                goto unlock;
3750
3751        *prot = pgprot_val(pte_pgprot(pte));
3752        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3753
3754        ret = 0;
3755unlock:
3756        pte_unmap_unlock(ptep, ptl);
3757out:
3758        return ret;
3759}
3760
3761int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3762                        void *buf, int len, int write)
3763{
3764        resource_size_t phys_addr;
3765        unsigned long prot = 0;
3766        void __iomem *maddr;
3767        int offset = addr & (PAGE_SIZE-1);
3768
3769        if (follow_phys(vma, addr, write, &prot, &phys_addr))
3770                return -EINVAL;
3771
3772        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3773        if (write)
3774                memcpy_toio(maddr + offset, buf, len);
3775        else
3776                memcpy_fromio(buf, maddr + offset, len);
3777        iounmap(maddr);
3778
3779        return len;
3780}
3781#endif
3782
3783/*
3784 * Access another process' address space as given in mm.  If non-NULL, use the
3785 * given task for page fault accounting.
3786 */
3787static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3788                unsigned long addr, void *buf, int len, int write)
3789{
3790        struct vm_area_struct *vma;
3791        void *old_buf = buf;
3792
3793        down_read(&mm->mmap_sem);
3794        /* ignore errors, just check how much was successfully transferred */
3795        while (len) {
3796                int bytes, ret, offset;
3797                void *maddr;
3798                struct page *page = NULL;
3799
3800                ret = get_user_pages(tsk, mm, addr, 1,
3801                                write, 1, &page, &vma);
3802                if (ret <= 0) {
3803                        /*
3804                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
3805                         * we can access using slightly different code.
3806                         */
3807#ifdef CONFIG_HAVE_IOREMAP_PROT
3808                        vma = find_vma(mm, addr);
3809                        if (!vma || vma->vm_start > addr)
3810                                break;
3811                        if (vma->vm_ops && vma->vm_ops->access)
3812                                ret = vma->vm_ops->access(vma, addr, buf,
3813                                                          len, write);
3814                        if (ret <= 0)
3815#endif
3816                                break;
3817                        bytes = ret;
3818                } else {
3819                        bytes = len;
3820                        offset = addr & (PAGE_SIZE-1);
3821                        if (bytes > PAGE_SIZE-offset)
3822                                bytes = PAGE_SIZE-offset;
3823
3824                        maddr = kmap(page);
3825                        if (write) {
3826                                copy_to_user_page(vma, page, addr,
3827                                                  maddr + offset, buf, bytes);
3828                                set_page_dirty_lock(page);
3829                        } else {
3830                                copy_from_user_page(vma, page, addr,
3831                                                    buf, maddr + offset, bytes);
3832                        }
3833                        kunmap(page);
3834                        page_cache_release(page);
3835                }
3836                len -= bytes;
3837                buf += bytes;
3838                addr += bytes;
3839        }
3840        up_read(&mm->mmap_sem);
3841
3842        return buf - old_buf;
3843}
3844
3845/**
3846 * access_remote_vm - access another process' address space
3847 * @mm:         the mm_struct of the target address space
3848 * @addr:       start address to access
3849 * @buf:        source or destination buffer
3850 * @len:        number of bytes to transfer
3851 * @write:      whether the access is a write
3852 *
3853 * The caller must hold a reference on @mm.
3854 */
3855int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3856                void *buf, int len, int write)
3857{
3858        return __access_remote_vm(NULL, mm, addr, buf, len, write);
3859}
3860
3861/*
3862 * Access another process' address space.
3863 * Source/target buffer must be kernel space,
3864 * Do not walk the page table directly, use get_user_pages
3865 */
3866int access_process_vm(struct task_struct *tsk, unsigned long addr,
3867                void *buf, int len, int write)
3868{
3869        struct mm_struct *mm;
3870        int ret;
3871
3872        mm = get_task_mm(tsk);
3873        if (!mm)
3874                return 0;
3875
3876        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3877        mmput(mm);
3878
3879        return ret;
3880}
3881
3882/*
3883 * Print the name of a VMA.
3884 */
3885void print_vma_addr(char *prefix, unsigned long ip)
3886{
3887        struct mm_struct *mm = current->mm;
3888        struct vm_area_struct *vma;
3889
3890        /*
3891         * Do not print if we are in atomic
3892         * contexts (in exception stacks, etc.):
3893         */
3894        if (preempt_count())
3895                return;
3896
3897        down_read(&mm->mmap_sem);
3898        vma = find_vma(mm, ip);
3899        if (vma && vma->vm_file) {
3900                struct file *f = vma->vm_file;
3901                char *buf = (char *)__get_free_page(GFP_KERNEL);
3902                if (buf) {
3903                        char *p, *s;
3904
3905                        p = d_path(&f->f_path, buf, PAGE_SIZE);
3906                        if (IS_ERR(p))
3907                                p = "?";
3908                        s = strrchr(p, '/');
3909                        if (s)
3910                                p = s+1;
3911                        printk("%s%s[%lx+%lx]", prefix, p,
3912                                        vma->vm_start,
3913                                        vma->vm_end - vma->vm_start);
3914                        free_page((unsigned long)buf);
3915                }
3916        }
3917        up_read(&current->mm->mmap_sem);
3918}
3919
3920#ifdef CONFIG_PROVE_LOCKING
3921void might_fault(void)
3922{
3923        /*
3924         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3925         * holding the mmap_sem, this is safe because kernel memory doesn't
3926         * get paged out, therefore we'll never actually fault, and the
3927         * below annotations will generate false positives.
3928         */
3929        if (segment_eq(get_fs(), KERNEL_DS))
3930                return;
3931
3932        might_sleep();
3933        /*
3934         * it would be nicer only to annotate paths which are not under
3935         * pagefault_disable, however that requires a larger audit and
3936         * providing helpers like get_user_atomic.
3937         */
3938        if (!in_atomic() && current->mm)
3939                might_lock_read(&current->mm->mmap_sem);
3940}
3941EXPORT_SYMBOL(might_fault);
3942#endif
3943
3944#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3945static void clear_gigantic_page(struct page *page,
3946                                unsigned long addr,
3947                                unsigned int pages_per_huge_page)
3948{
3949        int i;
3950        struct page *p = page;
3951
3952        might_sleep();
3953        for (i = 0; i < pages_per_huge_page;
3954             i++, p = mem_map_next(p, page, i)) {
3955                cond_resched();
3956                clear_user_highpage(p, addr + i * PAGE_SIZE);
3957        }
3958}
3959void clear_huge_page(struct page *page,
3960                     unsigned long addr, unsigned int pages_per_huge_page)
3961{
3962        int i;
3963
3964        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3965                clear_gigantic_page(page, addr, pages_per_huge_page);
3966                return;
3967        }
3968
3969        might_sleep();
3970        for (i = 0; i < pages_per_huge_page; i++) {
3971                cond_resched();
3972                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3973        }
3974}
3975
3976static void copy_user_gigantic_page(struct page *dst, struct page *src,
3977                                    unsigned long addr,
3978                                    struct vm_area_struct *vma,
3979                                    unsigned int pages_per_huge_page)
3980{
3981        int i;
3982        struct page *dst_base = dst;
3983        struct page *src_base = src;
3984
3985        for (i = 0; i < pages_per_huge_page; ) {
3986                cond_resched();
3987                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3988
3989                i++;
3990                dst = mem_map_next(dst, dst_base, i);
3991                src = mem_map_next(src, src_base, i);
3992        }
3993}
3994
3995void copy_user_huge_page(struct page *dst, struct page *src,
3996                         unsigned long addr, struct vm_area_struct *vma,
3997                         unsigned int pages_per_huge_page)
3998{
3999        int i;
4000
4001        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4002                copy_user_gigantic_page(dst, src, addr, vma,
4003                                        pages_per_huge_page);
4004                return;
4005        }
4006
4007        might_sleep();
4008        for (i = 0; i < pages_per_huge_page; i++) {
4009                cond_resched();
4010                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4011        }
4012}
4013#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4014
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.