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