linux/mm/memory.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *              Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *              (Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/hugetlb.h>
  44#include <linux/mman.h>
  45#include <linux/swap.h>
  46#include <linux/highmem.h>
  47#include <linux/pagemap.h>
  48#include <linux/ksm.h>
  49#include <linux/rmap.h>
  50#include <linux/export.h>
  51#include <linux/delayacct.h>
  52#include <linux/init.h>
  53#include <linux/writeback.h>
  54#include <linux/memcontrol.h>
  55#include <linux/mmu_notifier.h>
  56#include <linux/kallsyms.h>
  57#include <linux/swapops.h>
  58#include <linux/elf.h>
  59#include <linux/gfp.h>
  60#include <linux/migrate.h>
  61#include <linux/string.h>
  62
  63#include <asm/io.h>
  64#include <asm/pgalloc.h>
  65#include <asm/uaccess.h>
  66#include <asm/tlb.h>
  67#include <asm/tlbflush.h>
  68#include <asm/pgtable.h>
  69
  70#include "internal.h"
  71
  72#ifdef LAST_NID_NOT_IN_PAGE_FLAGS
  73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
  74#endif
  75
  76#ifndef CONFIG_NEED_MULTIPLE_NODES
  77/* use the per-pgdat data instead for discontigmem - mbligh */
  78unsigned long max_mapnr;
  79struct page *mem_map;
  80
  81EXPORT_SYMBOL(max_mapnr);
  82EXPORT_SYMBOL(mem_map);
  83#endif
  84
  85unsigned long num_physpages;
  86/*
  87 * A number of key systems in x86 including ioremap() rely on the assumption
  88 * that high_memory defines the upper bound on direct map memory, then end
  89 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  90 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  91 * and ZONE_HIGHMEM.
  92 */
  93void * high_memory;
  94
  95EXPORT_SYMBOL(num_physpages);
  96EXPORT_SYMBOL(high_memory);
  97
  98/*
  99 * Randomize the address space (stacks, mmaps, brk, etc.).
 100 *
 101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 102 *   as ancient (libc5 based) binaries can segfault. )
 103 */
 104int randomize_va_space __read_mostly =
 105#ifdef CONFIG_COMPAT_BRK
 106                                        1;
 107#else
 108                                        2;
 109#endif
 110
 111static int __init disable_randmaps(char *s)
 112{
 113        randomize_va_space = 0;
 114        return 1;
 115}
 116__setup("norandmaps", disable_randmaps);
 117
 118unsigned long zero_pfn __read_mostly;
 119unsigned long highest_memmap_pfn __read_mostly;
 120
 121/*
 122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 123 */
 124static int __init init_zero_pfn(void)
 125{
 126        zero_pfn = page_to_pfn(ZERO_PAGE(0));
 127        return 0;
 128}
 129core_initcall(init_zero_pfn);
 130
 131
 132#if defined(SPLIT_RSS_COUNTING)
 133
 134void sync_mm_rss(struct mm_struct *mm)
 135{
 136        int i;
 137
 138        for (i = 0; i < NR_MM_COUNTERS; i++) {
 139                if (current->rss_stat.count[i]) {
 140                        add_mm_counter(mm, i, current->rss_stat.count[i]);
 141                        current->rss_stat.count[i] = 0;
 142                }
 143        }
 144        current->rss_stat.events = 0;
 145}
 146
 147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 148{
 149        struct task_struct *task = current;
 150
 151        if (likely(task->mm == mm))
 152                task->rss_stat.count[member] += val;
 153        else
 154                add_mm_counter(mm, member, val);
 155}
 156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 158
 159/* sync counter once per 64 page faults */
 160#define TASK_RSS_EVENTS_THRESH  (64)
 161static void check_sync_rss_stat(struct task_struct *task)
 162{
 163        if (unlikely(task != current))
 164                return;
 165        if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 166                sync_mm_rss(task->mm);
 167}
 168#else /* SPLIT_RSS_COUNTING */
 169
 170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 172
 173static void check_sync_rss_stat(struct task_struct *task)
 174{
 175}
 176
 177#endif /* SPLIT_RSS_COUNTING */
 178
 179#ifdef HAVE_GENERIC_MMU_GATHER
 180
 181static int tlb_next_batch(struct mmu_gather *tlb)
 182{
 183        struct mmu_gather_batch *batch;
 184
 185        batch = tlb->active;
 186        if (batch->next) {
 187                tlb->active = batch->next;
 188                return 1;
 189        }
 190
 191        if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
 192                return 0;
 193
 194        batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 195        if (!batch)
 196                return 0;
 197
 198        tlb->batch_count++;
 199        batch->next = NULL;
 200        batch->nr   = 0;
 201        batch->max  = MAX_GATHER_BATCH;
 202
 203        tlb->active->next = batch;
 204        tlb->active = batch;
 205
 206        return 1;
 207}
 208
 209/* tlb_gather_mmu
 210 *      Called to initialize an (on-stack) mmu_gather structure for page-table
 211 *      tear-down from @mm. The @fullmm argument is used when @mm is without
 212 *      users and we're going to destroy the full address space (exit/execve).
 213 */
 214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 215{
 216        tlb->mm = mm;
 217
 218        tlb->fullmm     = fullmm;
 219        tlb->need_flush_all = 0;
 220        tlb->start      = -1UL;
 221        tlb->end        = 0;
 222        tlb->need_flush = 0;
 223        tlb->fast_mode  = (num_possible_cpus() == 1);
 224        tlb->local.next = NULL;
 225        tlb->local.nr   = 0;
 226        tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 227        tlb->active     = &tlb->local;
 228        tlb->batch_count = 0;
 229
 230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 231        tlb->batch = NULL;
 232#endif
 233}
 234
 235void tlb_flush_mmu(struct mmu_gather *tlb)
 236{
 237        struct mmu_gather_batch *batch;
 238
 239        if (!tlb->need_flush)
 240                return;
 241        tlb->need_flush = 0;
 242        tlb_flush(tlb);
 243#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 244        tlb_table_flush(tlb);
 245#endif
 246
 247        if (tlb_fast_mode(tlb))
 248                return;
 249
 250        for (batch = &tlb->local; batch; batch = batch->next) {
 251                free_pages_and_swap_cache(batch->pages, batch->nr);
 252                batch->nr = 0;
 253        }
 254        tlb->active = &tlb->local;
 255}
 256
 257/* tlb_finish_mmu
 258 *      Called at the end of the shootdown operation to free up any resources
 259 *      that were required.
 260 */
 261void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 262{
 263        struct mmu_gather_batch *batch, *next;
 264
 265        tlb->start = start;
 266        tlb->end   = end;
 267        tlb_flush_mmu(tlb);
 268
 269        /* keep the page table cache within bounds */
 270        check_pgt_cache();
 271
 272        for (batch = tlb->local.next; batch; batch = next) {
 273                next = batch->next;
 274                free_pages((unsigned long)batch, 0);
 275        }
 276        tlb->local.next = NULL;
 277}
 278
 279/* __tlb_remove_page
 280 *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 281 *      handling the additional races in SMP caused by other CPUs caching valid
 282 *      mappings in their TLBs. Returns the number of free page slots left.
 283 *      When out of page slots we must call tlb_flush_mmu().
 284 */
 285int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 286{
 287        struct mmu_gather_batch *batch;
 288
 289        VM_BUG_ON(!tlb->need_flush);
 290
 291        if (tlb_fast_mode(tlb)) {
 292                free_page_and_swap_cache(page);
 293                return 1; /* avoid calling tlb_flush_mmu() */
 294        }
 295
 296        batch = tlb->active;
 297        batch->pages[batch->nr++] = page;
 298        if (batch->nr == batch->max) {
 299                if (!tlb_next_batch(tlb))
 300                        return 0;
 301                batch = tlb->active;
 302        }
 303        VM_BUG_ON(batch->nr > batch->max);
 304
 305        return batch->max - batch->nr;
 306}
 307
 308#endif /* HAVE_GENERIC_MMU_GATHER */
 309
 310#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 311
 312/*
 313 * See the comment near struct mmu_table_batch.
 314 */
 315
 316static void tlb_remove_table_smp_sync(void *arg)
 317{
 318        /* Simply deliver the interrupt */
 319}
 320
 321static void tlb_remove_table_one(void *table)
 322{
 323        /*
 324         * This isn't an RCU grace period and hence the page-tables cannot be
 325         * assumed to be actually RCU-freed.
 326         *
 327         * It is however sufficient for software page-table walkers that rely on
 328         * IRQ disabling. See the comment near struct mmu_table_batch.
 329         */
 330        smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 331        __tlb_remove_table(table);
 332}
 333
 334static void tlb_remove_table_rcu(struct rcu_head *head)
 335{
 336        struct mmu_table_batch *batch;
 337        int i;
 338
 339        batch = container_of(head, struct mmu_table_batch, rcu);
 340
 341        for (i = 0; i < batch->nr; i++)
 342                __tlb_remove_table(batch->tables[i]);
 343
 344        free_page((unsigned long)batch);
 345}
 346
 347void tlb_table_flush(struct mmu_gather *tlb)
 348{
 349        struct mmu_table_batch **batch = &tlb->batch;
 350
 351        if (*batch) {
 352                call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 353                *batch = NULL;
 354        }
 355}
 356
 357void tlb_remove_table(struct mmu_gather *tlb, void *table)
 358{
 359        struct mmu_table_batch **batch = &tlb->batch;
 360
 361        tlb->need_flush = 1;
 362
 363        /*
 364         * When there's less then two users of this mm there cannot be a
 365         * concurrent page-table walk.
 366         */
 367        if (atomic_read(&tlb->mm->mm_users) < 2) {
 368                __tlb_remove_table(table);
 369                return;
 370        }
 371
 372        if (*batch == NULL) {
 373                *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 374                if (*batch == NULL) {
 375                        tlb_remove_table_one(table);
 376                        return;
 377                }
 378                (*batch)->nr = 0;
 379        }
 380        (*batch)->tables[(*batch)->nr++] = table;
 381        if ((*batch)->nr == MAX_TABLE_BATCH)
 382                tlb_table_flush(tlb);
 383}
 384
 385#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 386
 387/*
 388 * If a p?d_bad entry is found while walking page tables, report
 389 * the error, before resetting entry to p?d_none.  Usually (but
 390 * very seldom) called out from the p?d_none_or_clear_bad macros.
 391 */
 392
 393void pgd_clear_bad(pgd_t *pgd)
 394{
 395        pgd_ERROR(*pgd);
 396        pgd_clear(pgd);
 397}
 398
 399void pud_clear_bad(pud_t *pud)
 400{
 401        pud_ERROR(*pud);
 402        pud_clear(pud);
 403}
 404
 405void pmd_clear_bad(pmd_t *pmd)
 406{
 407        pmd_ERROR(*pmd);
 408        pmd_clear(pmd);
 409}
 410
 411/*
 412 * Note: this doesn't free the actual pages themselves. That
 413 * has been handled earlier when unmapping all the memory regions.
 414 */
 415static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 416                           unsigned long addr)
 417{
 418        pgtable_t token = pmd_pgtable(*pmd);
 419        pmd_clear(pmd);
 420        pte_free_tlb(tlb, token, addr);
 421        tlb->mm->nr_ptes--;
 422}
 423
 424static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 425                                unsigned long addr, unsigned long end,
 426                                unsigned long floor, unsigned long ceiling)
 427{
 428        pmd_t *pmd;
 429        unsigned long next;
 430        unsigned long start;
 431
 432        start = addr;
 433        pmd = pmd_offset(pud, addr);
 434        do {
 435                next = pmd_addr_end(addr, end);
 436                if (pmd_none_or_clear_bad(pmd))
 437                        continue;
 438                free_pte_range(tlb, pmd, addr);
 439        } while (pmd++, addr = next, addr != end);
 440
 441        start &= PUD_MASK;
 442        if (start < floor)
 443                return;
 444        if (ceiling) {
 445                ceiling &= PUD_MASK;
 446                if (!ceiling)
 447                        return;
 448        }
 449        if (end - 1 > ceiling - 1)
 450                return;
 451
 452        pmd = pmd_offset(pud, start);
 453        pud_clear(pud);
 454        pmd_free_tlb(tlb, pmd, start);
 455}
 456
 457static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 458                                unsigned long addr, unsigned long end,
 459                                unsigned long floor, unsigned long ceiling)
 460{
 461        pud_t *pud;
 462        unsigned long next;
 463        unsigned long start;
 464
 465        start = addr;
 466        pud = pud_offset(pgd, addr);
 467        do {
 468                next = pud_addr_end(addr, end);
 469                if (pud_none_or_clear_bad(pud))
 470                        continue;
 471                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 472        } while (pud++, addr = next, addr != end);
 473
 474        start &= PGDIR_MASK;
 475        if (start < floor)
 476                return;
 477        if (ceiling) {
 478                ceiling &= PGDIR_MASK;
 479                if (!ceiling)
 480                        return;
 481        }
 482        if (end - 1 > ceiling - 1)
 483                return;
 484
 485        pud = pud_offset(pgd, start);
 486        pgd_clear(pgd);
 487        pud_free_tlb(tlb, pud, start);
 488}
 489
 490/*
 491 * This function frees user-level page tables of a process.
 492 *
 493 * Must be called with pagetable lock held.
 494 */
 495void free_pgd_range(struct mmu_gather *tlb,
 496                        unsigned long addr, unsigned long end,
 497                        unsigned long floor, unsigned long ceiling)
 498{
 499        pgd_t *pgd;
 500        unsigned long next;
 501
 502        /*
 503         * The next few lines have given us lots of grief...
 504         *
 505         * Why are we testing PMD* at this top level?  Because often
 506         * there will be no work to do at all, and we'd prefer not to
 507         * go all the way down to the bottom just to discover that.
 508         *
 509         * Why all these "- 1"s?  Because 0 represents both the bottom
 510         * of the address space and the top of it (using -1 for the
 511         * top wouldn't help much: the masks would do the wrong thing).
 512         * The rule is that addr 0 and floor 0 refer to the bottom of
 513         * the address space, but end 0 and ceiling 0 refer to the top
 514         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 515         * that end 0 case should be mythical).
 516         *
 517         * Wherever addr is brought up or ceiling brought down, we must
 518         * be careful to reject "the opposite 0" before it confuses the
 519         * subsequent tests.  But what about where end is brought down
 520         * by PMD_SIZE below? no, end can't go down to 0 there.
 521         *
 522         * Whereas we round start (addr) and ceiling down, by different
 523         * masks at different levels, in order to test whether a table
 524         * now has no other vmas using it, so can be freed, we don't
 525         * bother to round floor or end up - the tests don't need that.
 526         */
 527
 528        addr &= PMD_MASK;
 529        if (addr < floor) {
 530                addr += PMD_SIZE;
 531                if (!addr)
 532                        return;
 533        }
 534        if (ceiling) {
 535                ceiling &= PMD_MASK;
 536                if (!ceiling)
 537                        return;
 538        }
 539        if (end - 1 > ceiling - 1)
 540                end -= PMD_SIZE;
 541        if (addr > end - 1)
 542                return;
 543
 544        pgd = pgd_offset(tlb->mm, addr);
 545        do {
 546                next = pgd_addr_end(addr, end);
 547                if (pgd_none_or_clear_bad(pgd))
 548                        continue;
 549                free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 550        } while (pgd++, addr = next, addr != end);
 551}
 552
 553void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 554                unsigned long floor, unsigned long ceiling)
 555{
 556        while (vma) {
 557                struct vm_area_struct *next = vma->vm_next;
 558                unsigned long addr = vma->vm_start;
 559
 560                /*
 561                 * Hide vma from rmap and truncate_pagecache before freeing
 562                 * pgtables
 563                 */
 564                unlink_anon_vmas(vma);
 565                unlink_file_vma(vma);
 566
 567                if (is_vm_hugetlb_page(vma)) {
 568                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 569                                floor, next? next->vm_start: ceiling);
 570                } else {
 571                        /*
 572                         * Optimization: gather nearby vmas into one call down
 573                         */
 574                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 575                               && !is_vm_hugetlb_page(next)) {
 576                                vma = next;
 577                                next = vma->vm_next;
 578                                unlink_anon_vmas(vma);
 579                                unlink_file_vma(vma);
 580                        }
 581                        free_pgd_range(tlb, addr, vma->vm_end,
 582                                floor, next? next->vm_start: ceiling);
 583                }
 584                vma = next;
 585        }
 586}
 587
 588int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 589                pmd_t *pmd, unsigned long address)
 590{
 591        pgtable_t new = pte_alloc_one(mm, address);
 592        int wait_split_huge_page;
 593        if (!new)
 594                return -ENOMEM;
 595
 596        /*
 597         * Ensure all pte setup (eg. pte page lock and page clearing) are
 598         * visible before the pte is made visible to other CPUs by being
 599         * put into page tables.
 600         *
 601         * The other side of the story is the pointer chasing in the page
 602         * table walking code (when walking the page table without locking;
 603         * ie. most of the time). Fortunately, these data accesses consist
 604         * of a chain of data-dependent loads, meaning most CPUs (alpha
 605         * being the notable exception) will already guarantee loads are
 606         * seen in-order. See the alpha page table accessors for the
 607         * smp_read_barrier_depends() barriers in page table walking code.
 608         */
 609        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 610
 611        spin_lock(&mm->page_table_lock);
 612        wait_split_huge_page = 0;
 613        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 614                mm->nr_ptes++;
 615                pmd_populate(mm, pmd, new);
 616                new = NULL;
 617        } else if (unlikely(pmd_trans_splitting(*pmd)))
 618                wait_split_huge_page = 1;
 619        spin_unlock(&mm->page_table_lock);
 620        if (new)
 621                pte_free(mm, new);
 622        if (wait_split_huge_page)
 623                wait_split_huge_page(vma->anon_vma, pmd);
 624        return 0;
 625}
 626
 627int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 628{
 629        pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 630        if (!new)
 631                return -ENOMEM;
 632
 633        smp_wmb(); /* See comment in __pte_alloc */
 634
 635        spin_lock(&init_mm.page_table_lock);
 636        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 637                pmd_populate_kernel(&init_mm, pmd, new);
 638                new = NULL;
 639        } else
 640                VM_BUG_ON(pmd_trans_splitting(*pmd));
 641        spin_unlock(&init_mm.page_table_lock);
 642        if (new)
 643                pte_free_kernel(&init_mm, new);
 644        return 0;
 645}
 646
 647static inline void init_rss_vec(int *rss)
 648{
 649        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 650}
 651
 652static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 653{
 654        int i;
 655
 656        if (current->mm == mm)
 657                sync_mm_rss(mm);
 658        for (i = 0; i < NR_MM_COUNTERS; i++)
 659                if (rss[i])
 660                        add_mm_counter(mm, i, rss[i]);
 661}
 662
 663/*
 664 * This function is called to print an error when a bad pte
 665 * is found. For example, we might have a PFN-mapped pte in
 666 * a region that doesn't allow it.
 667 *
 668 * The calling function must still handle the error.
 669 */
 670static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 671                          pte_t pte, struct page *page)
 672{
 673        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 674        pud_t *pud = pud_offset(pgd, addr);
 675        pmd_t *pmd = pmd_offset(pud, addr);
 676        struct address_space *mapping;
 677        pgoff_t index;
 678        static unsigned long resume;
 679        static unsigned long nr_shown;
 680        static unsigned long nr_unshown;
 681
 682        /*
 683         * Allow a burst of 60 reports, then keep quiet for that minute;
 684         * or allow a steady drip of one report per second.
 685         */
 686        if (nr_shown == 60) {
 687                if (time_before(jiffies, resume)) {
 688                        nr_unshown++;
 689                        return;
 690                }
 691                if (nr_unshown) {
 692                        printk(KERN_ALERT
 693                                "BUG: Bad page map: %lu messages suppressed\n",
 694                                nr_unshown);
 695                        nr_unshown = 0;
 696                }
 697                nr_shown = 0;
 698        }
 699        if (nr_shown++ == 0)
 700                resume = jiffies + 60 * HZ;
 701
 702        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 703        index = linear_page_index(vma, addr);
 704
 705        printk(KERN_ALERT
 706                "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 707                current->comm,
 708                (long long)pte_val(pte), (long long)pmd_val(*pmd));
 709        if (page)
 710                dump_page(page);
 711        printk(KERN_ALERT
 712                "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 713                (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 714        /*
 715         * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 716         */
 717        if (vma->vm_ops)
 718                print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
 719                                (unsigned long)vma->vm_ops->fault);
 720        if (vma->vm_file && vma->vm_file->f_op)
 721                print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
 722                                (unsigned long)vma->vm_file->f_op->mmap);
 723        dump_stack();
 724        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 725}
 726
 727static inline bool is_cow_mapping(vm_flags_t flags)
 728{
 729        return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 730}
 731
 732/*
 733 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 734 *
 735 * "Special" mappings do not wish to be associated with a "struct page" (either
 736 * it doesn't exist, or it exists but they don't want to touch it). In this
 737 * case, NULL is returned here. "Normal" mappings do have a struct page.
 738 *
 739 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 740 * pte bit, in which case this function is trivial. Secondly, an architecture
 741 * may not have a spare pte bit, which requires a more complicated scheme,
 742 * described below.
 743 *
 744 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 745 * special mapping (even if there are underlying and valid "struct pages").
 746 * COWed pages of a VM_PFNMAP are always normal.
 747 *
 748 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 749 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 750 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 751 * mapping will always honor the rule
 752 *
 753 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 754 *
 755 * And for normal mappings this is false.
 756 *
 757 * This restricts such mappings to be a linear translation from virtual address
 758 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 759 * as the vma is not a COW mapping; in that case, we know that all ptes are
 760 * special (because none can have been COWed).
 761 *
 762 *
 763 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 764 *
 765 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 766 * page" backing, however the difference is that _all_ pages with a struct
 767 * page (that is, those where pfn_valid is true) are refcounted and considered
 768 * normal pages by the VM. The disadvantage is that pages are refcounted
 769 * (which can be slower and simply not an option for some PFNMAP users). The
 770 * advantage is that we don't have to follow the strict linearity rule of
 771 * PFNMAP mappings in order to support COWable mappings.
 772 *
 773 */
 774#ifdef __HAVE_ARCH_PTE_SPECIAL
 775# define HAVE_PTE_SPECIAL 1
 776#else
 777# define HAVE_PTE_SPECIAL 0
 778#endif
 779struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 780                                pte_t pte)
 781{
 782        unsigned long pfn = pte_pfn(pte);
 783
 784        if (HAVE_PTE_SPECIAL) {
 785                if (likely(!pte_special(pte)))
 786                        goto check_pfn;
 787                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 788                        return NULL;
 789                if (!is_zero_pfn(pfn))
 790                        print_bad_pte(vma, addr, pte, NULL);
 791                return NULL;
 792        }
 793
 794        /* !HAVE_PTE_SPECIAL case follows: */
 795
 796        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 797                if (vma->vm_flags & VM_MIXEDMAP) {
 798                        if (!pfn_valid(pfn))
 799                                return NULL;
 800                        goto out;
 801                } else {
 802                        unsigned long off;
 803                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 804                        if (pfn == vma->vm_pgoff + off)
 805                                return NULL;
 806                        if (!is_cow_mapping(vma->vm_flags))
 807                                return NULL;
 808                }
 809        }
 810
 811        if (is_zero_pfn(pfn))
 812                return NULL;
 813check_pfn:
 814        if (unlikely(pfn > highest_memmap_pfn)) {
 815                print_bad_pte(vma, addr, pte, NULL);
 816                return NULL;
 817        }
 818
 819        /*
 820         * NOTE! We still have PageReserved() pages in the page tables.
 821         * eg. VDSO mappings can cause them to exist.
 822         */
 823out:
 824        return pfn_to_page(pfn);
 825}
 826
 827/*
 828 * copy one vm_area from one task to the other. Assumes the page tables
 829 * already present in the new task to be cleared in the whole range
 830 * covered by this vma.
 831 */
 832
 833static inline unsigned long
 834copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 835                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 836                unsigned long addr, int *rss)
 837{
 838        unsigned long vm_flags = vma->vm_flags;
 839        pte_t pte = *src_pte;
 840        struct page *page;
 841
 842        /* pte contains position in swap or file, so copy. */
 843        if (unlikely(!pte_present(pte))) {
 844                if (!pte_file(pte)) {
 845                        swp_entry_t entry = pte_to_swp_entry(pte);
 846
 847                        if (swap_duplicate(entry) < 0)
 848                                return entry.val;
 849
 850                        /* make sure dst_mm is on swapoff's mmlist. */
 851                        if (unlikely(list_empty(&dst_mm->mmlist))) {
 852                                spin_lock(&mmlist_lock);
 853                                if (list_empty(&dst_mm->mmlist))
 854                                        list_add(&dst_mm->mmlist,
 855                                                 &src_mm->mmlist);
 856                                spin_unlock(&mmlist_lock);
 857                        }
 858                        if (likely(!non_swap_entry(entry)))
 859                                rss[MM_SWAPENTS]++;
 860                        else if (is_migration_entry(entry)) {
 861                                page = migration_entry_to_page(entry);
 862
 863                                if (PageAnon(page))
 864                                        rss[MM_ANONPAGES]++;
 865                                else
 866                                        rss[MM_FILEPAGES]++;
 867
 868                                if (is_write_migration_entry(entry) &&
 869                                    is_cow_mapping(vm_flags)) {
 870                                        /*
 871                                         * COW mappings require pages in both
 872                                         * parent and child to be set to read.
 873                                         */
 874                                        make_migration_entry_read(&entry);
 875                                        pte = swp_entry_to_pte(entry);
 876                                        set_pte_at(src_mm, addr, src_pte, pte);
 877                                }
 878                        }
 879                }
 880                goto out_set_pte;
 881        }
 882
 883        /*
 884         * If it's a COW mapping, write protect it both
 885         * in the parent and the child
 886         */
 887        if (is_cow_mapping(vm_flags)) {
 888                ptep_set_wrprotect(src_mm, addr, src_pte);
 889                pte = pte_wrprotect(pte);
 890        }
 891
 892        /*
 893         * If it's a shared mapping, mark it clean in
 894         * the child
 895         */
 896        if (vm_flags & VM_SHARED)
 897                pte = pte_mkclean(pte);
 898        pte = pte_mkold(pte);
 899
 900        page = vm_normal_page(vma, addr, pte);
 901        if (page) {
 902                get_page(page);
 903                page_dup_rmap(page);
 904                if (PageAnon(page))
 905                        rss[MM_ANONPAGES]++;
 906                else
 907                        rss[MM_FILEPAGES]++;
 908        }
 909
 910out_set_pte:
 911        set_pte_at(dst_mm, addr, dst_pte, pte);
 912        return 0;
 913}
 914
 915int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 916                   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 917                   unsigned long addr, unsigned long end)
 918{
 919        pte_t *orig_src_pte, *orig_dst_pte;
 920        pte_t *src_pte, *dst_pte;
 921        spinlock_t *src_ptl, *dst_ptl;
 922        int progress = 0;
 923        int rss[NR_MM_COUNTERS];
 924        swp_entry_t entry = (swp_entry_t){0};
 925
 926again:
 927        init_rss_vec(rss);
 928
 929        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 930        if (!dst_pte)
 931                return -ENOMEM;
 932        src_pte = pte_offset_map(src_pmd, addr);
 933        src_ptl = pte_lockptr(src_mm, src_pmd);
 934        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 935        orig_src_pte = src_pte;
 936        orig_dst_pte = dst_pte;
 937        arch_enter_lazy_mmu_mode();
 938
 939        do {
 940                /*
 941                 * We are holding two locks at this point - either of them
 942                 * could generate latencies in another task on another CPU.
 943                 */
 944                if (progress >= 32) {
 945                        progress = 0;
 946                        if (need_resched() ||
 947                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 948                                break;
 949                }
 950                if (pte_none(*src_pte)) {
 951                        progress++;
 952                        continue;
 953                }
 954                entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 955                                                        vma, addr, rss);
 956                if (entry.val)
 957                        break;
 958                progress += 8;
 959        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 960
 961        arch_leave_lazy_mmu_mode();
 962        spin_unlock(src_ptl);
 963        pte_unmap(orig_src_pte);
 964        add_mm_rss_vec(dst_mm, rss);
 965        pte_unmap_unlock(orig_dst_pte, dst_ptl);
 966        cond_resched();
 967
 968        if (entry.val) {
 969                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 970                        return -ENOMEM;
 971                progress = 0;
 972        }
 973        if (addr != end)
 974                goto again;
 975        return 0;
 976}
 977
 978static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 979                pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 980                unsigned long addr, unsigned long end)
 981{
 982        pmd_t *src_pmd, *dst_pmd;
 983        unsigned long next;
 984
 985        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 986        if (!dst_pmd)
 987                return -ENOMEM;
 988        src_pmd = pmd_offset(src_pud, addr);
 989        do {
 990                next = pmd_addr_end(addr, end);
 991                if (pmd_trans_huge(*src_pmd)) {
 992                        int err;
 993                        VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
 994                        err = copy_huge_pmd(dst_mm, src_mm,
 995                                            dst_pmd, src_pmd, addr, vma);
 996                        if (err == -ENOMEM)
 997                                return -ENOMEM;
 998                        if (!err)
 999                                continue;
1000                        /* fall through */
1001                }
1002                if (pmd_none_or_clear_bad(src_pmd))
1003                        continue;
1004                if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1005                                                vma, addr, next))
1006                        return -ENOMEM;
1007        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1008        return 0;
1009}
1010
1011static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1012                pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1013                unsigned long addr, unsigned long end)
1014{
1015        pud_t *src_pud, *dst_pud;
1016        unsigned long next;
1017
1018        dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1019        if (!dst_pud)
1020                return -ENOMEM;
1021        src_pud = pud_offset(src_pgd, addr);
1022        do {
1023                next = pud_addr_end(addr, end);
1024                if (pud_none_or_clear_bad(src_pud))
1025                        continue;
1026                if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1027                                                vma, addr, next))
1028                        return -ENOMEM;
1029        } while (dst_pud++, src_pud++, addr = next, addr != end);
1030        return 0;
1031}
1032
1033int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1034                struct vm_area_struct *vma)
1035{
1036        pgd_t *src_pgd, *dst_pgd;
1037        unsigned long next;
1038        unsigned long addr = vma->vm_start;
1039        unsigned long end = vma->vm_end;
1040        unsigned long mmun_start;       /* For mmu_notifiers */
1041        unsigned long mmun_end;         /* For mmu_notifiers */
1042        bool is_cow;
1043        int ret;
1044
1045        /*
1046         * Don't copy ptes where a page fault will fill them correctly.
1047         * Fork becomes much lighter when there are big shared or private
1048         * readonly mappings. The tradeoff is that copy_page_range is more
1049         * efficient than faulting.
1050         */
1051        if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1052                               VM_PFNMAP | VM_MIXEDMAP))) {
1053                if (!vma->anon_vma)
1054                        return 0;
1055        }
1056
1057        if (is_vm_hugetlb_page(vma))
1058                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1059
1060        if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1061                /*
1062                 * We do not free on error cases below as remove_vma
1063                 * gets called on error from higher level routine
1064                 */
1065                ret = track_pfn_copy(vma);
1066                if (ret)
1067                        return ret;
1068        }
1069
1070        /*
1071         * We need to invalidate the secondary MMU mappings only when
1072         * there could be a permission downgrade on the ptes of the
1073         * parent mm. And a permission downgrade will only happen if
1074         * is_cow_mapping() returns true.
1075         */
1076        is_cow = is_cow_mapping(vma->vm_flags);
1077        mmun_start = addr;
1078        mmun_end   = end;
1079        if (is_cow)
1080                mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1081                                                    mmun_end);
1082
1083        ret = 0;
1084        dst_pgd = pgd_offset(dst_mm, addr);
1085        src_pgd = pgd_offset(src_mm, addr);
1086        do {
1087                next = pgd_addr_end(addr, end);
1088                if (pgd_none_or_clear_bad(src_pgd))
1089                        continue;
1090                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1091                                            vma, addr, next))) {
1092                        ret = -ENOMEM;
1093                        break;
1094                }
1095        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1096
1097        if (is_cow)
1098                mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1099        return ret;
1100}
1101
1102static unsigned long zap_pte_range(struct mmu_gather *tlb,
1103                                struct vm_area_struct *vma, pmd_t *pmd,
1104                                unsigned long addr, unsigned long end,
1105                                struct zap_details *details)
1106{
1107        struct mm_struct *mm = tlb->mm;
1108        int force_flush = 0;
1109        int rss[NR_MM_COUNTERS];
1110        spinlock_t *ptl;
1111        pte_t *start_pte;
1112        pte_t *pte;
1113
1114again:
1115        init_rss_vec(rss);
1116        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1117        pte = start_pte;
1118        arch_enter_lazy_mmu_mode();
1119        do {
1120                pte_t ptent = *pte;
1121                if (pte_none(ptent)) {
1122                        continue;
1123                }
1124
1125                if (pte_present(ptent)) {
1126                        struct page *page;
1127
1128                        page = vm_normal_page(vma, addr, ptent);
1129                        if (unlikely(details) && page) {
1130                                /*
1131                                 * unmap_shared_mapping_pages() wants to
1132                                 * invalidate cache without truncating:
1133                                 * unmap shared but keep private pages.
1134                                 */
1135                                if (details->check_mapping &&
1136                                    details->check_mapping != page->mapping)
1137                                        continue;
1138                                /*
1139                                 * Each page->index must be checked when
1140                                 * invalidating or truncating nonlinear.
1141                                 */
1142                                if (details->nonlinear_vma &&
1143                                    (page->index < details->first_index ||
1144                                     page->index > details->last_index))
1145                                        continue;
1146                        }
1147                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1148                                                        tlb->fullmm);
1149                        tlb_remove_tlb_entry(tlb, pte, addr);
1150                        if (unlikely(!page))
1151                                continue;
1152                        if (unlikely(details) && details->nonlinear_vma
1153                            && linear_page_index(details->nonlinear_vma,
1154                                                addr) != page->index)
1155                                set_pte_at(mm, addr, pte,
1156                                           pgoff_to_pte(page->index));
1157                        if (PageAnon(page))
1158                                rss[MM_ANONPAGES]--;
1159                        else {
1160                                if (pte_dirty(ptent))
1161                                        set_page_dirty(page);
1162                                if (pte_young(ptent) &&
1163                                    likely(!VM_SequentialReadHint(vma)))
1164                                        mark_page_accessed(page);
1165                                rss[MM_FILEPAGES]--;
1166                        }
1167                        page_remove_rmap(page);
1168                        if (unlikely(page_mapcount(page) < 0))
1169                                print_bad_pte(vma, addr, ptent, page);
1170                        force_flush = !__tlb_remove_page(tlb, page);
1171                        if (force_flush)
1172                                break;
1173                        continue;
1174                }
1175                /*
1176                 * If details->check_mapping, we leave swap entries;
1177                 * if details->nonlinear_vma, we leave file entries.
1178                 */
1179                if (unlikely(details))
1180                        continue;
1181                if (pte_file(ptent)) {
1182                        if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1183                                print_bad_pte(vma, addr, ptent, NULL);
1184                } else {
1185                        swp_entry_t entry = pte_to_swp_entry(ptent);
1186
1187                        if (!non_swap_entry(entry))
1188                                rss[MM_SWAPENTS]--;
1189                        else if (is_migration_entry(entry)) {
1190                                struct page *page;
1191
1192                                page = migration_entry_to_page(entry);
1193
1194                                if (PageAnon(page))
1195                                        rss[MM_ANONPAGES]--;
1196                                else
1197                                        rss[MM_FILEPAGES]--;
1198                        }
1199                        if (unlikely(!free_swap_and_cache(entry)))
1200                                print_bad_pte(vma, addr, ptent, NULL);
1201                }
1202                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1203        } while (pte++, addr += PAGE_SIZE, addr != end);
1204
1205        add_mm_rss_vec(mm, rss);
1206        arch_leave_lazy_mmu_mode();
1207        pte_unmap_unlock(start_pte, ptl);
1208
1209        /*
1210         * mmu_gather ran out of room to batch pages, we break out of
1211         * the PTE lock to avoid doing the potential expensive TLB invalidate
1212         * and page-free while holding it.
1213         */
1214        if (force_flush) {
1215                force_flush = 0;
1216
1217#ifdef HAVE_GENERIC_MMU_GATHER
1218                tlb->start = addr;
1219                tlb->end = end;
1220#endif
1221                tlb_flush_mmu(tlb);
1222                if (addr != end)
1223                        goto again;
1224        }
1225
1226        return addr;
1227}
1228
1229static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1230                                struct vm_area_struct *vma, pud_t *pud,
1231                                unsigned long addr, unsigned long end,
1232                                struct zap_details *details)
1233{
1234        pmd_t *pmd;
1235        unsigned long next;
1236
1237        pmd = pmd_offset(pud, addr);
1238        do {
1239                next = pmd_addr_end(addr, end);
1240                if (pmd_trans_huge(*pmd)) {
1241                        if (next - addr != HPAGE_PMD_SIZE) {
1242#ifdef CONFIG_DEBUG_VM
1243                                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1244                                        pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1245                                                __func__, addr, end,
1246                                                vma->vm_start,
1247                                                vma->vm_end);
1248                                        BUG();
1249                                }
1250#endif
1251                                split_huge_page_pmd(vma, addr, pmd);
1252                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1253                                goto next;
1254                        /* fall through */
1255                }
1256                /*
1257                 * Here there can be other concurrent MADV_DONTNEED or
1258                 * trans huge page faults running, and if the pmd is
1259                 * none or trans huge it can change under us. This is
1260                 * because MADV_DONTNEED holds the mmap_sem in read
1261                 * mode.
1262                 */
1263                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1264                        goto next;
1265                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1266next:
1267                cond_resched();
1268        } while (pmd++, addr = next, addr != end);
1269
1270        return addr;
1271}
1272
1273static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1274                                struct vm_area_struct *vma, pgd_t *pgd,
1275                                unsigned long addr, unsigned long end,
1276                                struct zap_details *details)
1277{
1278        pud_t *pud;
1279        unsigned long next;
1280
1281        pud = pud_offset(pgd, addr);
1282        do {
1283                next = pud_addr_end(addr, end);
1284                if (pud_none_or_clear_bad(pud))
1285                        continue;
1286                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1287        } while (pud++, addr = next, addr != end);
1288
1289        return addr;
1290}
1291
1292static void unmap_page_range(struct mmu_gather *tlb,
1293                             struct vm_area_struct *vma,
1294                             unsigned long addr, unsigned long end,
1295                             struct zap_details *details)
1296{
1297        pgd_t *pgd;
1298        unsigned long next;
1299
1300        if (details && !details->check_mapping && !details->nonlinear_vma)
1301                details = NULL;
1302
1303        BUG_ON(addr >= end);
1304        mem_cgroup_uncharge_start();
1305        tlb_start_vma(tlb, vma);
1306        pgd = pgd_offset(vma->vm_mm, addr);
1307        do {
1308                next = pgd_addr_end(addr, end);
1309                if (pgd_none_or_clear_bad(pgd))
1310                        continue;
1311                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1312        } while (pgd++, addr = next, addr != end);
1313        tlb_end_vma(tlb, vma);
1314        mem_cgroup_uncharge_end();
1315}
1316
1317
1318static void unmap_single_vma(struct mmu_gather *tlb,
1319                struct vm_area_struct *vma, unsigned long start_addr,
1320                unsigned long end_addr,
1321                struct zap_details *details)
1322{
1323        unsigned long start = max(vma->vm_start, start_addr);
1324        unsigned long end;
1325
1326        if (start >= vma->vm_end)
1327                return;
1328        end = min(vma->vm_end, end_addr);
1329        if (end <= vma->vm_start)
1330                return;
1331
1332        if (vma->vm_file)
1333                uprobe_munmap(vma, start, end);
1334
1335        if (unlikely(vma->vm_flags & VM_PFNMAP))
1336                untrack_pfn(vma, 0, 0);
1337
1338        if (start != end) {
1339                if (unlikely(is_vm_hugetlb_page(vma))) {
1340                        /*
1341                         * It is undesirable to test vma->vm_file as it
1342                         * should be non-null for valid hugetlb area.
1343                         * However, vm_file will be NULL in the error
1344                         * cleanup path of do_mmap_pgoff. When
1345                         * hugetlbfs ->mmap method fails,
1346                         * do_mmap_pgoff() nullifies vma->vm_file
1347                         * before calling this function to clean up.
1348                         * Since no pte has actually been setup, it is
1349                         * safe to do nothing in this case.
1350                         */
1351                        if (vma->vm_file) {
1352                                mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1353                                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1354                                mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1355                        }
1356                } else
1357                        unmap_page_range(tlb, vma, start, end, details);
1358        }
1359}
1360
1361/**
1362 * unmap_vmas - unmap a range of memory covered by a list of vma's
1363 * @tlb: address of the caller's struct mmu_gather
1364 * @vma: the starting vma
1365 * @start_addr: virtual address at which to start unmapping
1366 * @end_addr: virtual address at which to end unmapping
1367 *
1368 * Unmap all pages in the vma list.
1369 *
1370 * Only addresses between `start' and `end' will be unmapped.
1371 *
1372 * The VMA list must be sorted in ascending virtual address order.
1373 *
1374 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1375 * range after unmap_vmas() returns.  So the only responsibility here is to
1376 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1377 * drops the lock and schedules.
1378 */
1379void unmap_vmas(struct mmu_gather *tlb,
1380                struct vm_area_struct *vma, unsigned long start_addr,
1381                unsigned long end_addr)
1382{
1383        struct mm_struct *mm = vma->vm_mm;
1384
1385        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1386        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1387                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1388        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1389}
1390
1391/**
1392 * zap_page_range - remove user pages in a given range
1393 * @vma: vm_area_struct holding the applicable pages
1394 * @start: starting address of pages to zap
1395 * @size: number of bytes to zap
1396 * @details: details of nonlinear truncation or shared cache invalidation
1397 *
1398 * Caller must protect the VMA list
1399 */
1400void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1401                unsigned long size, struct zap_details *details)
1402{
1403        struct mm_struct *mm = vma->vm_mm;
1404        struct mmu_gather tlb;
1405        unsigned long end = start + size;
1406
1407        lru_add_drain();
1408        tlb_gather_mmu(&tlb, mm, 0);
1409        update_hiwater_rss(mm);
1410        mmu_notifier_invalidate_range_start(mm, start, end);
1411        for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1412                unmap_single_vma(&tlb, vma, start, end, details);
1413        mmu_notifier_invalidate_range_end(mm, start, end);
1414        tlb_finish_mmu(&tlb, start, end);
1415}
1416
1417/**
1418 * zap_page_range_single - remove user pages in a given range
1419 * @vma: vm_area_struct holding the applicable pages
1420 * @address: starting address of pages to zap
1421 * @size: number of bytes to zap
1422 * @details: details of nonlinear truncation or shared cache invalidation
1423 *
1424 * The range must fit into one VMA.
1425 */
1426static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1427                unsigned long size, struct zap_details *details)
1428{
1429        struct mm_struct *mm = vma->vm_mm;
1430        struct mmu_gather tlb;
1431        unsigned long end = address + size;
1432
1433        lru_add_drain();
1434        tlb_gather_mmu(&tlb, mm, 0);
1435        update_hiwater_rss(mm);
1436        mmu_notifier_invalidate_range_start(mm, address, end);
1437        unmap_single_vma(&tlb, vma, address, end, details);
1438        mmu_notifier_invalidate_range_end(mm, address, end);
1439        tlb_finish_mmu(&tlb, address, end);
1440}
1441
1442/**
1443 * zap_vma_ptes - remove ptes mapping the vma
1444 * @vma: vm_area_struct holding ptes to be zapped
1445 * @address: starting address of pages to zap
1446 * @size: number of bytes to zap
1447 *
1448 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1449 *
1450 * The entire address range must be fully contained within the vma.
1451 *
1452 * Returns 0 if successful.
1453 */
1454int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1455                unsigned long size)
1456{
1457        if (address < vma->vm_start || address + size > vma->vm_end ||
1458                        !(vma->vm_flags & VM_PFNMAP))
1459                return -1;
1460        zap_page_range_single(vma, address, size, NULL);
1461        return 0;
1462}
1463EXPORT_SYMBOL_GPL(zap_vma_ptes);
1464
1465/**
1466 * follow_page_mask - look up a page descriptor from a user-virtual address
1467 * @vma: vm_area_struct mapping @address
1468 * @address: virtual address to look up
1469 * @flags: flags modifying lookup behaviour
1470 * @page_mask: on output, *page_mask is set according to the size of the page
1471 *
1472 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1473 *
1474 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1475 * an error pointer if there is a mapping to something not represented
1476 * by a page descriptor (see also vm_normal_page()).
1477 */
1478struct page *follow_page_mask(struct vm_area_struct *vma,
1479                              unsigned long address, unsigned int flags,
1480                              unsigned int *page_mask)
1481{
1482        pgd_t *pgd;
1483        pud_t *pud;
1484        pmd_t *pmd;
1485        pte_t *ptep, pte;
1486        spinlock_t *ptl;
1487        struct page *page;
1488        struct mm_struct *mm = vma->vm_mm;
1489
1490        *page_mask = 0;
1491
1492        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1493        if (!IS_ERR(page)) {
1494                BUG_ON(flags & FOLL_GET);
1495                goto out;
1496        }
1497
1498        page = NULL;
1499        pgd = pgd_offset(mm, address);
1500        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1501                goto no_page_table;
1502
1503        pud = pud_offset(pgd, address);
1504        if (pud_none(*pud))
1505                goto no_page_table;
1506        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1507                BUG_ON(flags & FOLL_GET);
1508                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1509                goto out;
1510        }
1511        if (unlikely(pud_bad(*pud)))
1512                goto no_page_table;
1513
1514        pmd = pmd_offset(pud, address);
1515        if (pmd_none(*pmd))
1516                goto no_page_table;
1517        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1518                BUG_ON(flags & FOLL_GET);
1519                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1520                goto out;
1521        }
1522        if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1523                goto no_page_table;
1524        if (pmd_trans_huge(*pmd)) {
1525                if (flags & FOLL_SPLIT) {
1526                        split_huge_page_pmd(vma, address, pmd);
1527                        goto split_fallthrough;
1528                }
1529                spin_lock(&mm->page_table_lock);
1530                if (likely(pmd_trans_huge(*pmd))) {
1531                        if (unlikely(pmd_trans_splitting(*pmd))) {
1532                                spin_unlock(&mm->page_table_lock);
1533                                wait_split_huge_page(vma->anon_vma, pmd);
1534                        } else {
1535                                page = follow_trans_huge_pmd(vma, address,
1536                                                             pmd, flags);
1537                                spin_unlock(&mm->page_table_lock);
1538                                *page_mask = HPAGE_PMD_NR - 1;
1539                                goto out;
1540                        }
1541                } else
1542                        spin_unlock(&mm->page_table_lock);
1543                /* fall through */
1544        }
1545split_fallthrough:
1546        if (unlikely(pmd_bad(*pmd)))
1547                goto no_page_table;
1548
1549        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1550
1551        pte = *ptep;
1552        if (!pte_present(pte)) {
1553                swp_entry_t entry;
1554                /*
1555                 * KSM's break_ksm() relies upon recognizing a ksm page
1556                 * even while it is being migrated, so for that case we
1557                 * need migration_entry_wait().
1558                 */
1559                if (likely(!(flags & FOLL_MIGRATION)))
1560                        goto no_page;
1561                if (pte_none(pte) || pte_file(pte))
1562                        goto no_page;
1563                entry = pte_to_swp_entry(pte);
1564                if (!is_migration_entry(entry))
1565                        goto no_page;
1566                pte_unmap_unlock(ptep, ptl);
1567                migration_entry_wait(mm, pmd, address);
1568                goto split_fallthrough;
1569        }
1570        if ((flags & FOLL_NUMA) && pte_numa(pte))
1571                goto no_page;
1572        if ((flags & FOLL_WRITE) && !pte_write(pte))
1573                goto unlock;
1574
1575        page = vm_normal_page(vma, address, pte);
1576        if (unlikely(!page)) {
1577                if ((flags & FOLL_DUMP) ||
1578                    !is_zero_pfn(pte_pfn(pte)))
1579                        goto bad_page;
1580                page = pte_page(pte);
1581        }
1582
1583        if (flags & FOLL_GET)
1584                get_page_foll(page);
1585        if (flags & FOLL_TOUCH) {
1586                if ((flags & FOLL_WRITE) &&
1587                    !pte_dirty(pte) && !PageDirty(page))
1588                        set_page_dirty(page);
1589                /*
1590                 * pte_mkyoung() would be more correct here, but atomic care
1591                 * is needed to avoid losing the dirty bit: it is easier to use
1592                 * mark_page_accessed().
1593                 */
1594                mark_page_accessed(page);
1595        }
1596        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1597                /*
1598                 * The preliminary mapping check is mainly to avoid the
1599                 * pointless overhead of lock_page on the ZERO_PAGE
1600                 * which might bounce very badly if there is contention.
1601                 *
1602                 * If the page is already locked, we don't need to
1603                 * handle it now - vmscan will handle it later if and
1604                 * when it attempts to reclaim the page.
1605                 */
1606                if (page->mapping && trylock_page(page)) {
1607                        lru_add_drain();  /* push cached pages to LRU */
1608                        /*
1609                         * Because we lock page here, and migration is
1610                         * blocked by the pte's page reference, and we
1611                         * know the page is still mapped, we don't even
1612                         * need to check for file-cache page truncation.
1613                         */
1614                        mlock_vma_page(page);
1615                        unlock_page(page);
1616                }
1617        }
1618unlock:
1619        pte_unmap_unlock(ptep, ptl);
1620out:
1621        return page;
1622
1623bad_page:
1624        pte_unmap_unlock(ptep, ptl);
1625        return ERR_PTR(-EFAULT);
1626
1627no_page:
1628        pte_unmap_unlock(ptep, ptl);
1629        if (!pte_none(pte))
1630                return page;
1631
1632no_page_table:
1633        /*
1634         * When core dumping an enormous anonymous area that nobody
1635         * has touched so far, we don't want to allocate unnecessary pages or
1636         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1637         * then get_dump_page() will return NULL to leave a hole in the dump.
1638         * But we can only make this optimization where a hole would surely
1639         * be zero-filled if handle_mm_fault() actually did handle it.
1640         */
1641        if ((flags & FOLL_DUMP) &&
1642            (!vma->vm_ops || !vma->vm_ops->fault))
1643                return ERR_PTR(-EFAULT);
1644        return page;
1645}
1646
1647static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1648{
1649        return stack_guard_page_start(vma, addr) ||
1650               stack_guard_page_end(vma, addr+PAGE_SIZE);
1651}
1652
1653/**
1654 * __get_user_pages() - pin user pages in memory
1655 * @tsk:        task_struct of target task
1656 * @mm:         mm_struct of target mm
1657 * @start:      starting user address
1658 * @nr_pages:   number of pages from start to pin
1659 * @gup_flags:  flags modifying pin behaviour
1660 * @pages:      array that receives pointers to the pages pinned.
1661 *              Should be at least nr_pages long. Or NULL, if caller
1662 *              only intends to ensure the pages are faulted in.
1663 * @vmas:       array of pointers to vmas corresponding to each page.
1664 *              Or NULL if the caller does not require them.
1665 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1666 *
1667 * Returns number of pages pinned. This may be fewer than the number
1668 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1669 * were pinned, returns -errno. Each page returned must be released
1670 * with a put_page() call when it is finished with. vmas will only
1671 * remain valid while mmap_sem is held.
1672 *
1673 * Must be called with mmap_sem held for read or write.
1674 *
1675 * __get_user_pages walks a process's page tables and takes a reference to
1676 * each struct page that each user address corresponds to at a given
1677 * instant. That is, it takes the page that would be accessed if a user
1678 * thread accesses the given user virtual address at that instant.
1679 *
1680 * This does not guarantee that the page exists in the user mappings when
1681 * __get_user_pages returns, and there may even be a completely different
1682 * page there in some cases (eg. if mmapped pagecache has been invalidated
1683 * and subsequently re faulted). However it does guarantee that the page
1684 * won't be freed completely. And mostly callers simply care that the page
1685 * contains data that was valid *at some point in time*. Typically, an IO
1686 * or similar operation cannot guarantee anything stronger anyway because
1687 * locks can't be held over the syscall boundary.
1688 *
1689 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1690 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1691 * appropriate) must be called after the page is finished with, and
1692 * before put_page is called.
1693 *
1694 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1695 * or mmap_sem contention, and if waiting is needed to pin all pages,
1696 * *@nonblocking will be set to 0.
1697 *
1698 * In most cases, get_user_pages or get_user_pages_fast should be used
1699 * instead of __get_user_pages. __get_user_pages should be used only if
1700 * you need some special @gup_flags.
1701 */
1702long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1703                unsigned long start, unsigned long nr_pages,
1704                unsigned int gup_flags, struct page **pages,
1705                struct vm_area_struct **vmas, int *nonblocking)
1706{
1707        long i;
1708        unsigned long vm_flags;
1709        unsigned int page_mask;
1710
1711        if (!nr_pages)
1712                return 0;
1713
1714        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1715
1716        /* 
1717         * Require read or write permissions.
1718         * If FOLL_FORCE is set, we only require the "MAY" flags.
1719         */
1720        vm_flags  = (gup_flags & FOLL_WRITE) ?
1721                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1722        vm_flags &= (gup_flags & FOLL_FORCE) ?
1723                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1724
1725        /*
1726         * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1727         * would be called on PROT_NONE ranges. We must never invoke
1728         * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1729         * page faults would unprotect the PROT_NONE ranges if
1730         * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1731         * bitflag. So to avoid that, don't set FOLL_NUMA if
1732         * FOLL_FORCE is set.
1733         */
1734        if (!(gup_flags & FOLL_FORCE))
1735                gup_flags |= FOLL_NUMA;
1736
1737        i = 0;
1738
1739        do {
1740                struct vm_area_struct *vma;
1741
1742                vma = find_extend_vma(mm, start);
1743                if (!vma && in_gate_area(mm, start)) {
1744                        unsigned long pg = start & PAGE_MASK;
1745                        pgd_t *pgd;
1746                        pud_t *pud;
1747                        pmd_t *pmd;
1748                        pte_t *pte;
1749
1750                        /* user gate pages are read-only */
1751                        if (gup_flags & FOLL_WRITE)
1752                                return i ? : -EFAULT;
1753                        if (pg > TASK_SIZE)
1754                                pgd = pgd_offset_k(pg);
1755                        else
1756                                pgd = pgd_offset_gate(mm, pg);
1757                        BUG_ON(pgd_none(*pgd));
1758                        pud = pud_offset(pgd, pg);
1759                        BUG_ON(pud_none(*pud));
1760                        pmd = pmd_offset(pud, pg);
1761                        if (pmd_none(*pmd))
1762                                return i ? : -EFAULT;
1763                        VM_BUG_ON(pmd_trans_huge(*pmd));
1764                        pte = pte_offset_map(pmd, pg);
1765                        if (pte_none(*pte)) {
1766                                pte_unmap(pte);
1767                                return i ? : -EFAULT;
1768                        }
1769                        vma = get_gate_vma(mm);
1770                        if (pages) {
1771                                struct page *page;
1772
1773                                page = vm_normal_page(vma, start, *pte);
1774                                if (!page) {
1775                                        if (!(gup_flags & FOLL_DUMP) &&
1776                                             is_zero_pfn(pte_pfn(*pte)))
1777                                                page = pte_page(*pte);
1778                                        else {
1779                                                pte_unmap(pte);
1780                                                return i ? : -EFAULT;
1781                                        }
1782                                }
1783                                pages[i] = page;
1784                                get_page(page);
1785                        }
1786                        pte_unmap(pte);
1787                        page_mask = 0;
1788                        goto next_page;
1789                }
1790
1791                if (!vma ||
1792                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1793                    !(vm_flags & vma->vm_flags))
1794                        return i ? : -EFAULT;
1795
1796                if (is_vm_hugetlb_page(vma)) {
1797                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1798                                        &start, &nr_pages, i, gup_flags);
1799                        continue;
1800                }
1801
1802                do {
1803                        struct page *page;
1804                        unsigned int foll_flags = gup_flags;
1805                        unsigned int page_increm;
1806
1807                        /*
1808                         * If we have a pending SIGKILL, don't keep faulting
1809                         * pages and potentially allocating memory.
1810                         */
1811                        if (unlikely(fatal_signal_pending(current)))
1812                                return i ? i : -ERESTARTSYS;
1813
1814                        cond_resched();
1815                        while (!(page = follow_page_mask(vma, start,
1816                                                foll_flags, &page_mask))) {
1817                                int ret;
1818                                unsigned int fault_flags = 0;
1819
1820                                /* For mlock, just skip the stack guard page. */
1821                                if (foll_flags & FOLL_MLOCK) {
1822                                        if (stack_guard_page(vma, start))
1823                                                goto next_page;
1824                                }
1825                                if (foll_flags & FOLL_WRITE)
1826                                        fault_flags |= FAULT_FLAG_WRITE;
1827                                if (nonblocking)
1828                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1829                                if (foll_flags & FOLL_NOWAIT)
1830                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1831
1832                                ret = handle_mm_fault(mm, vma, start,
1833                                                        fault_flags);
1834
1835                                if (ret & VM_FAULT_ERROR) {
1836                                        if (ret & VM_FAULT_OOM)
1837                                                return i ? i : -ENOMEM;
1838                                        if (ret & (VM_FAULT_HWPOISON |
1839                                                   VM_FAULT_HWPOISON_LARGE)) {
1840                                                if (i)
1841                                                        return i;
1842                                                else if (gup_flags & FOLL_HWPOISON)
1843                                                        return -EHWPOISON;
1844                                                else
1845                                                        return -EFAULT;
1846                                        }
1847                                        if (ret & VM_FAULT_SIGBUS)
1848                                                return i ? i : -EFAULT;
1849                                        BUG();
1850                                }
1851
1852                                if (tsk) {
1853                                        if (ret & VM_FAULT_MAJOR)
1854                                                tsk->maj_flt++;
1855                                        else
1856                                                tsk->min_flt++;
1857                                }
1858
1859                                if (ret & VM_FAULT_RETRY) {
1860                                        if (nonblocking)
1861                                                *nonblocking = 0;
1862                                        return i;
1863                                }
1864
1865                                /*
1866                                 * The VM_FAULT_WRITE bit tells us that
1867                                 * do_wp_page has broken COW when necessary,
1868                                 * even if maybe_mkwrite decided not to set
1869                                 * pte_write. We can thus safely do subsequent
1870                                 * page lookups as if they were reads. But only
1871                                 * do so when looping for pte_write is futile:
1872                                 * in some cases userspace may also be wanting
1873                                 * to write to the gotten user page, which a
1874                                 * read fault here might prevent (a readonly
1875                                 * page might get reCOWed by userspace write).
1876                                 */
1877                                if ((ret & VM_FAULT_WRITE) &&
1878                                    !(vma->vm_flags & VM_WRITE))
1879                                        foll_flags &= ~FOLL_WRITE;
1880
1881                                cond_resched();
1882                        }
1883                        if (IS_ERR(page))
1884                                return i ? i : PTR_ERR(page);
1885                        if (pages) {
1886                                pages[i] = page;
1887
1888                                flush_anon_page(vma, page, start);
1889                                flush_dcache_page(page);
1890                                page_mask = 0;
1891                        }
1892next_page:
1893                        if (vmas) {
1894                                vmas[i] = vma;
1895                                page_mask = 0;
1896                        }
1897                        page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1898                        if (page_increm > nr_pages)
1899                                page_increm = nr_pages;
1900                        i += page_increm;
1901                        start += page_increm * PAGE_SIZE;
1902                        nr_pages -= page_increm;
1903                } while (nr_pages && start < vma->vm_end);
1904        } while (nr_pages);
1905        return i;
1906}
1907EXPORT_SYMBOL(__get_user_pages);
1908
1909/*
1910 * fixup_user_fault() - manually resolve a user page fault
1911 * @tsk:        the task_struct to use for page fault accounting, or
1912 *              NULL if faults are not to be recorded.
1913 * @mm:         mm_struct of target mm
1914 * @address:    user address
1915 * @fault_flags:flags to pass down to handle_mm_fault()
1916 *
1917 * This is meant to be called in the specific scenario where for locking reasons
1918 * we try to access user memory in atomic context (within a pagefault_disable()
1919 * section), this returns -EFAULT, and we want to resolve the user fault before
1920 * trying again.
1921 *
1922 * Typically this is meant to be used by the futex code.
1923 *
1924 * The main difference with get_user_pages() is that this function will
1925 * unconditionally call handle_mm_fault() which will in turn perform all the
1926 * necessary SW fixup of the dirty and young bits in the PTE, while
1927 * handle_mm_fault() only guarantees to update these in the struct page.
1928 *
1929 * This is important for some architectures where those bits also gate the
1930 * access permission to the page because they are maintained in software.  On
1931 * such architectures, gup() will not be enough to make a subsequent access
1932 * succeed.
1933 *
1934 * This should be called with the mm_sem held for read.
1935 */
1936int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1937                     unsigned long address, unsigned int fault_flags)
1938{
1939        struct vm_area_struct *vma;
1940        int ret;
1941
1942        vma = find_extend_vma(mm, address);
1943        if (!vma || address < vma->vm_start)
1944                return -EFAULT;
1945
1946        ret = handle_mm_fault(mm, vma, address, fault_flags);
1947        if (ret & VM_FAULT_ERROR) {
1948                if (ret & VM_FAULT_OOM)
1949                        return -ENOMEM;
1950                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1951                        return -EHWPOISON;
1952                if (ret & VM_FAULT_SIGBUS)
1953                        return -EFAULT;
1954                BUG();
1955        }
1956        if (tsk) {
1957                if (ret & VM_FAULT_MAJOR)
1958                        tsk->maj_flt++;
1959                else
1960                        tsk->min_flt++;
1961        }
1962        return 0;
1963}
1964
1965/*
1966 * get_user_pages() - pin user pages in memory
1967 * @tsk:        the task_struct to use for page fault accounting, or
1968 *              NULL if faults are not to be recorded.
1969 * @mm:         mm_struct of target mm
1970 * @start:      starting user address
1971 * @nr_pages:   number of pages from start to pin
1972 * @write:      whether pages will be written to by the caller
1973 * @force:      whether to force write access even if user mapping is
1974 *              readonly. This will result in the page being COWed even
1975 *              in MAP_SHARED mappings. You do not want this.
1976 * @pages:      array that receives pointers to the pages pinned.
1977 *              Should be at least nr_pages long. Or NULL, if caller
1978 *              only intends to ensure the pages are faulted in.
1979 * @vmas:       array of pointers to vmas corresponding to each page.
1980 *              Or NULL if the caller does not require them.
1981 *
1982 * Returns number of pages pinned. This may be fewer than the number
1983 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1984 * were pinned, returns -errno. Each page returned must be released
1985 * with a put_page() call when it is finished with. vmas will only
1986 * remain valid while mmap_sem is held.
1987 *
1988 * Must be called with mmap_sem held for read or write.
1989 *
1990 * get_user_pages walks a process's page tables and takes a reference to
1991 * each struct page that each user address corresponds to at a given
1992 * instant. That is, it takes the page that would be accessed if a user
1993 * thread accesses the given user virtual address at that instant.
1994 *
1995 * This does not guarantee that the page exists in the user mappings when
1996 * get_user_pages returns, and there may even be a completely different
1997 * page there in some cases (eg. if mmapped pagecache has been invalidated
1998 * and subsequently re faulted). However it does guarantee that the page
1999 * won't be freed completely. And mostly callers simply care that the page
2000 * contains data that was valid *at some point in time*. Typically, an IO
2001 * or similar operation cannot guarantee anything stronger anyway because
2002 * locks can't be held over the syscall boundary.
2003 *
2004 * If write=0, the page must not be written to. If the page is written to,
2005 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2006 * after the page is finished with, and before put_page is called.
2007 *
2008 * get_user_pages is typically used for fewer-copy IO operations, to get a
2009 * handle on the memory by some means other than accesses via the user virtual
2010 * addresses. The pages may be submitted for DMA to devices or accessed via
2011 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2012 * use the correct cache flushing APIs.
2013 *
2014 * See also get_user_pages_fast, for performance critical applications.
2015 */
2016long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2017                unsigned long start, unsigned long nr_pages, int write,
2018                int force, struct page **pages, struct vm_area_struct **vmas)
2019{
2020        int flags = FOLL_TOUCH;
2021
2022        if (pages)
2023                flags |= FOLL_GET;
2024        if (write)
2025                flags |= FOLL_WRITE;
2026        if (force)
2027                flags |= FOLL_FORCE;
2028
2029        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2030                                NULL);
2031}
2032EXPORT_SYMBOL(get_user_pages);
2033
2034/**
2035 * get_dump_page() - pin user page in memory while writing it to core dump
2036 * @addr: user address
2037 *
2038 * Returns struct page pointer of user page pinned for dump,
2039 * to be freed afterwards by page_cache_release() or put_page().
2040 *
2041 * Returns NULL on any kind of failure - a hole must then be inserted into
2042 * the corefile, to preserve alignment with its headers; and also returns
2043 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2044 * allowing a hole to be left in the corefile to save diskspace.
2045 *
2046 * Called without mmap_sem, but after all other threads have been killed.
2047 */
2048#ifdef CONFIG_ELF_CORE
2049struct page *get_dump_page(unsigned long addr)
2050{
2051        struct vm_area_struct *vma;
2052        struct page *page;
2053
2054        if (__get_user_pages(current, current->mm, addr, 1,
2055                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2056                             NULL) < 1)
2057                return NULL;
2058        flush_cache_page(vma, addr, page_to_pfn(page));
2059        return page;
2060}
2061#endif /* CONFIG_ELF_CORE */
2062
2063pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2064                        spinlock_t **ptl)
2065{
2066        pgd_t * pgd = pgd_offset(mm, addr);
2067        pud_t * pud = pud_alloc(mm, pgd, addr);
2068        if (pud) {
2069                pmd_t * pmd = pmd_alloc(mm, pud, addr);
2070                if (pmd) {
2071                        VM_BUG_ON(pmd_trans_huge(*pmd));
2072                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
2073                }
2074        }
2075        return NULL;
2076}
2077
2078/*
2079 * This is the old fallback for page remapping.
2080 *
2081 * For historical reasons, it only allows reserved pages. Only
2082 * old drivers should use this, and they needed to mark their
2083 * pages reserved for the old functions anyway.
2084 */
2085static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2086                        struct page *page, pgprot_t prot)
2087{
2088        struct mm_struct *mm = vma->vm_mm;
2089        int retval;
2090        pte_t *pte;
2091        spinlock_t *ptl;
2092
2093        retval = -EINVAL;
2094        if (PageAnon(page))
2095                goto out;
2096        retval = -ENOMEM;
2097        flush_dcache_page(page);
2098        pte = get_locked_pte(mm, addr, &ptl);
2099        if (!pte)
2100                goto out;
2101        retval = -EBUSY;
2102        if (!pte_none(*pte))
2103                goto out_unlock;
2104
2105        /* Ok, finally just insert the thing.. */
2106        get_page(page);
2107        inc_mm_counter_fast(mm, MM_FILEPAGES);
2108        page_add_file_rmap(page);
2109        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2110
2111        retval = 0;
2112        pte_unmap_unlock(pte, ptl);
2113        return retval;
2114out_unlock:
2115        pte_unmap_unlock(pte, ptl);
2116out:
2117        return retval;
2118}
2119
2120/**
2121 * vm_insert_page - insert single page into user vma
2122 * @vma: user vma to map to
2123 * @addr: target user address of this page
2124 * @page: source kernel page
2125 *
2126 * This allows drivers to insert individual pages they've allocated
2127 * into a user vma.
2128 *
2129 * The page has to be a nice clean _individual_ kernel allocation.
2130 * If you allocate a compound page, you need to have marked it as
2131 * such (__GFP_COMP), or manually just split the page up yourself
2132 * (see split_page()).
2133 *
2134 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2135 * took an arbitrary page protection parameter. This doesn't allow
2136 * that. Your vma protection will have to be set up correctly, which
2137 * means that if you want a shared writable mapping, you'd better
2138 * ask for a shared writable mapping!
2139 *
2140 * The page does not need to be reserved.
2141 *
2142 * Usually this function is called from f_op->mmap() handler
2143 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2144 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2145 * function from other places, for example from page-fault handler.
2146 */
2147int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2148                        struct page *page)
2149{
2150        if (addr < vma->vm_start || addr >= vma->vm_end)
2151                return -EFAULT;
2152        if (!page_count(page))
2153                return -EINVAL;
2154        if (!(vma->vm_flags & VM_MIXEDMAP)) {
2155                BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2156                BUG_ON(vma->vm_flags & VM_PFNMAP);
2157                vma->vm_flags |= VM_MIXEDMAP;
2158        }
2159        return insert_page(vma, addr, page, vma->vm_page_prot);
2160}
2161EXPORT_SYMBOL(vm_insert_page);
2162
2163static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2164                        unsigned long pfn, pgprot_t prot)
2165{
2166        struct mm_struct *mm = vma->vm_mm;
2167        int retval;
2168        pte_t *pte, entry;
2169        spinlock_t *ptl;
2170
2171        retval = -ENOMEM;
2172        pte = get_locked_pte(mm, addr, &ptl);
2173        if (!pte)
2174                goto out;
2175        retval = -EBUSY;
2176        if (!pte_none(*pte))
2177                goto out_unlock;
2178
2179        /* Ok, finally just insert the thing.. */
2180        entry = pte_mkspecial(pfn_pte(pfn, prot));
2181        set_pte_at(mm, addr, pte, entry);
2182        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2183
2184        retval = 0;
2185out_unlock:
2186        pte_unmap_unlock(pte, ptl);
2187out:
2188        return retval;
2189}
2190
2191/**
2192 * vm_insert_pfn - insert single pfn into user vma
2193 * @vma: user vma to map to
2194 * @addr: target user address of this page
2195 * @pfn: source kernel pfn
2196 *
2197 * Similar to vm_insert_page, this allows drivers to insert individual pages
2198 * they've allocated into a user vma. Same comments apply.
2199 *
2200 * This function should only be called from a vm_ops->fault handler, and
2201 * in that case the handler should return NULL.
2202 *
2203 * vma cannot be a COW mapping.
2204 *
2205 * As this is called only for pages that do not currently exist, we
2206 * do not need to flush old virtual caches or the TLB.
2207 */
2208int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2209                        unsigned long pfn)
2210{
2211        int ret;
2212        pgprot_t pgprot = vma->vm_page_prot;
2213        /*
2214         * Technically, architectures with pte_special can avoid all these
2215         * restrictions (same for remap_pfn_range).  However we would like
2216         * consistency in testing and feature parity among all, so we should
2217         * try to keep these invariants in place for everybody.
2218         */
2219        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2220        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2221                                                (VM_PFNMAP|VM_MIXEDMAP));
2222        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2223        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2224
2225        if (addr < vma->vm_start || addr >= vma->vm_end)
2226                return -EFAULT;
2227        if (track_pfn_insert(vma, &pgprot, pfn))
2228                return -EINVAL;
2229
2230        ret = insert_pfn(vma, addr, pfn, pgprot);
2231
2232        return ret;
2233}
2234EXPORT_SYMBOL(vm_insert_pfn);
2235
2236int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2237                        unsigned long pfn)
2238{
2239        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2240
2241        if (addr < vma->vm_start || addr >= vma->vm_end)
2242                return -EFAULT;
2243
2244        /*
2245         * If we don't have pte special, then we have to use the pfn_valid()
2246         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2247         * refcount the page if pfn_valid is true (hence insert_page rather
2248         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2249         * without pte special, it would there be refcounted as a normal page.
2250         */
2251        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2252                struct page *page;
2253
2254                page = pfn_to_page(pfn);
2255                return insert_page(vma, addr, page, vma->vm_page_prot);
2256        }
2257        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2258}
2259EXPORT_SYMBOL(vm_insert_mixed);
2260
2261/*
2262 * maps a range of physical memory into the requested pages. the old
2263 * mappings are removed. any references to nonexistent pages results
2264 * in null mappings (currently treated as "copy-on-access")
2265 */
2266static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2267                        unsigned long addr, unsigned long end,
2268                        unsigned long pfn, pgprot_t prot)
2269{
2270        pte_t *pte;
2271        spinlock_t *ptl;
2272
2273        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2274        if (!pte)
2275                return -ENOMEM;
2276        arch_enter_lazy_mmu_mode();
2277        do {
2278                BUG_ON(!pte_none(*pte));
2279                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2280                pfn++;
2281        } while (pte++, addr += PAGE_SIZE, addr != end);
2282        arch_leave_lazy_mmu_mode();
2283        pte_unmap_unlock(pte - 1, ptl);
2284        return 0;
2285}
2286
2287static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2288                        unsigned long addr, unsigned long end,
2289                        unsigned long pfn, pgprot_t prot)
2290{
2291        pmd_t *pmd;
2292        unsigned long next;
2293
2294        pfn -= addr >> PAGE_SHIFT;
2295        pmd = pmd_alloc(mm, pud, addr);
2296        if (!pmd)
2297                return -ENOMEM;
2298        VM_BUG_ON(pmd_trans_huge(*pmd));
2299        do {
2300                next = pmd_addr_end(addr, end);
2301                if (remap_pte_range(mm, pmd, addr, next,
2302                                pfn + (addr >> PAGE_SHIFT), prot))
2303                        return -ENOMEM;
2304        } while (pmd++, addr = next, addr != end);
2305        return 0;
2306}
2307
2308static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2309                        unsigned long addr, unsigned long end,
2310                        unsigned long pfn, pgprot_t prot)
2311{
2312        pud_t *pud;
2313        unsigned long next;
2314
2315        pfn -= addr >> PAGE_SHIFT;
2316        pud = pud_alloc(mm, pgd, addr);
2317        if (!pud)
2318                return -ENOMEM;
2319        do {
2320                next = pud_addr_end(addr, end);
2321                if (remap_pmd_range(mm, pud, addr, next,
2322                                pfn + (addr >> PAGE_SHIFT), prot))
2323                        return -ENOMEM;
2324        } while (pud++, addr = next, addr != end);
2325        return 0;
2326}
2327
2328/**
2329 * remap_pfn_range - remap kernel memory to userspace
2330 * @vma: user vma to map to
2331 * @addr: target user address to start at
2332 * @pfn: physical address of kernel memory
2333 * @size: size of map area
2334 * @prot: page protection flags for this mapping
2335 *
2336 *  Note: this is only safe if the mm semaphore is held when called.
2337 */
2338int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2339                    unsigned long pfn, unsigned long size, pgprot_t prot)
2340{
2341        pgd_t *pgd;
2342        unsigned long next;
2343        unsigned long end = addr + PAGE_ALIGN(size);
2344        struct mm_struct *mm = vma->vm_mm;
2345        int err;
2346
2347        /*
2348         * Physically remapped pages are special. Tell the
2349         * rest of the world about it:
2350         *   VM_IO tells people not to look at these pages
2351         *      (accesses can have side effects).
2352         *   VM_PFNMAP tells the core MM that the base pages are just
2353         *      raw PFN mappings, and do not have a "struct page" associated
2354         *      with them.
2355         *   VM_DONTEXPAND
2356         *      Disable vma merging and expanding with mremap().
2357         *   VM_DONTDUMP
2358         *      Omit vma from core dump, even when VM_IO turned off.
2359         *
2360         * There's a horrible special case to handle copy-on-write
2361         * behaviour that some programs depend on. We mark the "original"
2362         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2363         * See vm_normal_page() for details.
2364         */
2365        if (is_cow_mapping(vma->vm_flags)) {
2366                if (addr != vma->vm_start || end != vma->vm_end)
2367                        return -EINVAL;
2368                vma->vm_pgoff = pfn;
2369        }
2370
2371        err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2372        if (err)
2373                return -EINVAL;
2374
2375        vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2376
2377        BUG_ON(addr >= end);
2378        pfn -= addr >> PAGE_SHIFT;
2379        pgd = pgd_offset(mm, addr);
2380        flush_cache_range(vma, addr, end);
2381        do {
2382                next = pgd_addr_end(addr, end);
2383                err = remap_pud_range(mm, pgd, addr, next,
2384                                pfn + (addr >> PAGE_SHIFT), prot);
2385                if (err)
2386                        break;
2387        } while (pgd++, addr = next, addr != end);
2388
2389        if (err)
2390                untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2391
2392        return err;
2393}
2394EXPORT_SYMBOL(remap_pfn_range);
2395
2396/**
2397 * vm_iomap_memory - remap memory to userspace
2398 * @vma: user vma to map to
2399 * @start: start of area
2400 * @len: size of area
2401 *
2402 * This is a simplified io_remap_pfn_range() for common driver use. The
2403 * driver just needs to give us the physical memory range to be mapped,
2404 * we'll figure out the rest from the vma information.
2405 *
2406 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2407 * whatever write-combining details or similar.
2408 */
2409int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2410{
2411        unsigned long vm_len, pfn, pages;
2412
2413        /* Check that the physical memory area passed in looks valid */
2414        if (start + len < start)
2415                return -EINVAL;
2416        /*
2417         * You *really* shouldn't map things that aren't page-aligned,
2418         * but we've historically allowed it because IO memory might
2419         * just have smaller alignment.
2420         */
2421        len += start & ~PAGE_MASK;
2422        pfn = start >> PAGE_SHIFT;
2423        pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2424        if (pfn + pages < pfn)
2425                return -EINVAL;
2426
2427        /* We start the mapping 'vm_pgoff' pages into the area */
2428        if (vma->vm_pgoff > pages)
2429                return -EINVAL;
2430        pfn += vma->vm_pgoff;
2431        pages -= vma->vm_pgoff;
2432
2433        /* Can we fit all of the mapping? */
2434        vm_len = vma->vm_end - vma->vm_start;
2435        if (vm_len >> PAGE_SHIFT > pages)
2436                return -EINVAL;
2437
2438        /* Ok, let it rip */
2439        return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2440}
2441EXPORT_SYMBOL(vm_iomap_memory);
2442
2443static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2444                                     unsigned long addr, unsigned long end,
2445                                     pte_fn_t fn, void *data)
2446{
2447        pte_t *pte;
2448        int err;
2449        pgtable_t token;
2450        spinlock_t *uninitialized_var(ptl);
2451
2452        pte = (mm == &init_mm) ?
2453                pte_alloc_kernel(pmd, addr) :
2454                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2455        if (!pte)
2456                return -ENOMEM;
2457
2458        BUG_ON(pmd_huge(*pmd));
2459
2460        arch_enter_lazy_mmu_mode();
2461
2462        token = pmd_pgtable(*pmd);
2463
2464        do {
2465                err = fn(pte++, token, addr, data);
2466                if (err)
2467                        break;
2468        } while (addr += PAGE_SIZE, addr != end);
2469
2470        arch_leave_lazy_mmu_mode();
2471
2472        if (mm != &init_mm)
2473                pte_unmap_unlock(pte-1, ptl);
2474        return err;
2475}
2476
2477static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2478                                     unsigned long addr, unsigned long end,
2479                                     pte_fn_t fn, void *data)
2480{
2481        pmd_t *pmd;
2482        unsigned long next;
2483        int err;
2484
2485        BUG_ON(pud_huge(*pud));
2486
2487        pmd = pmd_alloc(mm, pud, addr);
2488        if (!pmd)
2489                return -ENOMEM;
2490        do {
2491                next = pmd_addr_end(addr, end);
2492                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2493                if (err)
2494                        break;
2495        } while (pmd++, addr = next, addr != end);
2496        return err;
2497}
2498
2499static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2500                                     unsigned long addr, unsigned long end,
2501                                     pte_fn_t fn, void *data)
2502{
2503        pud_t *pud;
2504        unsigned long next;
2505        int err;
2506
2507        pud = pud_alloc(mm, pgd, addr);
2508        if (!pud)
2509                return -ENOMEM;
2510        do {
2511                next = pud_addr_end(addr, end);
2512                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2513                if (err)
2514                        break;
2515        } while (pud++, addr = next, addr != end);
2516        return err;
2517}
2518
2519/*
2520 * Scan a region of virtual memory, filling in page tables as necessary
2521 * and calling a provided function on each leaf page table.
2522 */
2523int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2524                        unsigned long size, pte_fn_t fn, void *data)
2525{
2526        pgd_t *pgd;
2527        unsigned long next;
2528        unsigned long end = addr + size;
2529        int err;
2530
2531        BUG_ON(addr >= end);
2532        pgd = pgd_offset(mm, addr);
2533        do {
2534                next = pgd_addr_end(addr, end);
2535                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2536                if (err)
2537                        break;
2538        } while (pgd++, addr = next, addr != end);
2539
2540        return err;
2541}
2542EXPORT_SYMBOL_GPL(apply_to_page_range);
2543
2544/*
2545 * handle_pte_fault chooses page fault handler according to an entry
2546 * which was read non-atomically.  Before making any commitment, on
2547 * those architectures or configurations (e.g. i386 with PAE) which
2548 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2549 * must check under lock before unmapping the pte and proceeding
2550 * (but do_wp_page is only called after already making such a check;
2551 * and do_anonymous_page can safely check later on).
2552 */
2553static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2554                                pte_t *page_table, pte_t orig_pte)
2555{
2556        int same = 1;
2557#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2558        if (sizeof(pte_t) > sizeof(unsigned long)) {
2559                spinlock_t *ptl = pte_lockptr(mm, pmd);
2560                spin_lock(ptl);
2561                same = pte_same(*page_table, orig_pte);
2562                spin_unlock(ptl);
2563        }
2564#endif
2565        pte_unmap(page_table);
2566        return same;
2567}
2568
2569static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2570{
2571        /*
2572         * If the source page was a PFN mapping, we don't have
2573         * a "struct page" for it. We do a best-effort copy by
2574         * just copying from the original user address. If that
2575         * fails, we just zero-fill it. Live with it.
2576         */
2577        if (unlikely(!src)) {
2578                void *kaddr = kmap_atomic(dst);
2579                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2580
2581                /*
2582                 * This really shouldn't fail, because the page is there
2583                 * in the page tables. But it might just be unreadable,
2584                 * in which case we just give up and fill the result with
2585                 * zeroes.
2586                 */
2587                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2588                        clear_page(kaddr);
2589                kunmap_atomic(kaddr);
2590                flush_dcache_page(dst);
2591        } else
2592                copy_user_highpage(dst, src, va, vma);
2593}
2594
2595/*
2596 * This routine handles present pages, when users try to write
2597 * to a shared page. It is done by copying the page to a new address
2598 * and decrementing the shared-page counter for the old page.
2599 *
2600 * Note that this routine assumes that the protection checks have been
2601 * done by the caller (the low-level page fault routine in most cases).
2602 * Thus we can safely just mark it writable once we've done any necessary
2603 * COW.
2604 *
2605 * We also mark the page dirty at this point even though the page will
2606 * change only once the write actually happens. This avoids a few races,
2607 * and potentially makes it more efficient.
2608 *
2609 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2610 * but allow concurrent faults), with pte both mapped and locked.
2611 * We return with mmap_sem still held, but pte unmapped and unlocked.
2612 */
2613static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2614                unsigned long address, pte_t *page_table, pmd_t *pmd,
2615                spinlock_t *ptl, pte_t orig_pte)
2616        __releases(ptl)
2617{
2618        struct page *old_page, *new_page = NULL;
2619        pte_t entry;
2620        int ret = 0;
2621        int page_mkwrite = 0;
2622        struct page *dirty_page = NULL;
2623        unsigned long mmun_start = 0;   /* For mmu_notifiers */
2624        unsigned long mmun_end = 0;     /* For mmu_notifiers */
2625
2626        old_page = vm_normal_page(vma, address, orig_pte);
2627        if (!old_page) {
2628                /*
2629                 * VM_MIXEDMAP !pfn_valid() case
2630                 *
2631                 * We should not cow pages in a shared writeable mapping.
2632                 * Just mark the pages writable as we can't do any dirty
2633                 * accounting on raw pfn maps.
2634                 */
2635                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2636                                     (VM_WRITE|VM_SHARED))
2637                        goto reuse;
2638                goto gotten;
2639        }
2640
2641        /*
2642         * Take out anonymous pages first, anonymous shared vmas are
2643         * not dirty accountable.
2644         */
2645        if (PageAnon(old_page) && !PageKsm(old_page)) {
2646                if (!trylock_page(old_page)) {
2647                        page_cache_get(old_page);
2648                        pte_unmap_unlock(page_table, ptl);
2649                        lock_page(old_page);
2650                        page_table = pte_offset_map_lock(mm, pmd, address,
2651                                                         &ptl);
2652                        if (!pte_same(*page_table, orig_pte)) {
2653                                unlock_page(old_page);
2654                                goto unlock;
2655                        }
2656                        page_cache_release(old_page);
2657                }
2658                if (reuse_swap_page(old_page)) {
2659                        /*
2660                         * The page is all ours.  Move it to our anon_vma so
2661                         * the rmap code will not search our parent or siblings.
2662                         * Protected against the rmap code by the page lock.
2663                         */
2664                        page_move_anon_rmap(old_page, vma, address);
2665                        unlock_page(old_page);
2666                        goto reuse;
2667                }
2668                unlock_page(old_page);
2669        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2670                                        (VM_WRITE|VM_SHARED))) {
2671                /*
2672                 * Only catch write-faults on shared writable pages,
2673                 * read-only shared pages can get COWed by
2674                 * get_user_pages(.write=1, .force=1).
2675                 */
2676                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2677                        struct vm_fault vmf;
2678                        int tmp;
2679
2680                        vmf.virtual_address = (void __user *)(address &
2681                                                                PAGE_MASK);
2682                        vmf.pgoff = old_page->index;
2683                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2684                        vmf.page = old_page;
2685
2686                        /*
2687                         * Notify the address space that the page is about to
2688                         * become writable so that it can prohibit this or wait
2689                         * for the page to get into an appropriate state.
2690                         *
2691                         * We do this without the lock held, so that it can
2692                         * sleep if it needs to.
2693                         */
2694                        page_cache_get(old_page);
2695                        pte_unmap_unlock(page_table, ptl);
2696
2697                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2698                        if (unlikely(tmp &
2699                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2700                                ret = tmp;
2701                                goto unwritable_page;
2702                        }
2703                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2704                                lock_page(old_page);
2705                                if (!old_page->mapping) {
2706                                        ret = 0; /* retry the fault */
2707                                        unlock_page(old_page);
2708                                        goto unwritable_page;
2709                                }
2710                        } else
2711                                VM_BUG_ON(!PageLocked(old_page));
2712
2713                        /*
2714                         * Since we dropped the lock we need to revalidate
2715                         * the PTE as someone else may have changed it.  If
2716                         * they did, we just return, as we can count on the
2717                         * MMU to tell us if they didn't also make it writable.
2718                         */
2719                        page_table = pte_offset_map_lock(mm, pmd, address,
2720                                                         &ptl);
2721                        if (!pte_same(*page_table, orig_pte)) {
2722                                unlock_page(old_page);
2723                                goto unlock;
2724                        }
2725
2726                        page_mkwrite = 1;
2727                }
2728                dirty_page = old_page;
2729                get_page(dirty_page);
2730
2731reuse:
2732                flush_cache_page(vma, address, pte_pfn(orig_pte));
2733                entry = pte_mkyoung(orig_pte);
2734                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2735                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2736                        update_mmu_cache(vma, address, page_table);
2737                pte_unmap_unlock(page_table, ptl);
2738                ret |= VM_FAULT_WRITE;
2739
2740                if (!dirty_page)
2741                        return ret;
2742
2743                /*
2744                 * Yes, Virginia, this is actually required to prevent a race
2745                 * with clear_page_dirty_for_io() from clearing the page dirty
2746                 * bit after it clear all dirty ptes, but before a racing
2747                 * do_wp_page installs a dirty pte.
2748                 *
2749                 * __do_fault is protected similarly.
2750                 */
2751                if (!page_mkwrite) {
2752                        wait_on_page_locked(dirty_page);
2753                        set_page_dirty_balance(dirty_page, page_mkwrite);
2754                        /* file_update_time outside page_lock */
2755                        if (vma->vm_file)
2756                                file_update_time(vma->vm_file);
2757                }
2758                put_page(dirty_page);
2759                if (page_mkwrite) {
2760                        struct address_space *mapping = dirty_page->mapping;
2761
2762                        set_page_dirty(dirty_page);
2763                        unlock_page(dirty_page);
2764                        page_cache_release(dirty_page);
2765                        if (mapping)    {
2766                                /*
2767                                 * Some device drivers do not set page.mapping
2768                                 * but still dirty their pages
2769                                 */
2770                                balance_dirty_pages_ratelimited(mapping);
2771                        }
2772                }
2773
2774                return ret;
2775        }
2776
2777        /*
2778         * Ok, we need to copy. Oh, well..
2779         */
2780        page_cache_get(old_page);
2781gotten:
2782        pte_unmap_unlock(page_table, ptl);
2783
2784        if (unlikely(anon_vma_prepare(vma)))
2785                goto oom;
2786
2787        if (is_zero_pfn(pte_pfn(orig_pte))) {
2788                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2789                if (!new_page)
2790                        goto oom;
2791        } else {
2792                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2793                if (!new_page)
2794                        goto oom;
2795                cow_user_page(new_page, old_page, address, vma);
2796        }
2797        __SetPageUptodate(new_page);
2798
2799        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2800                goto oom_free_new;
2801
2802        mmun_start  = address & PAGE_MASK;
2803        mmun_end    = mmun_start + PAGE_SIZE;
2804        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2805
2806        /*
2807         * Re-check the pte - we dropped the lock
2808         */
2809        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2810        if (likely(pte_same(*page_table, orig_pte))) {
2811                if (old_page) {
2812                        if (!PageAnon(old_page)) {
2813                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2814                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2815                        }
2816                } else
2817                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2818                flush_cache_page(vma, address, pte_pfn(orig_pte));
2819                entry = mk_pte(new_page, vma->vm_page_prot);
2820                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2821                /*
2822                 * Clear the pte entry and flush it first, before updating the
2823                 * pte with the new entry. This will avoid a race condition
2824                 * seen in the presence of one thread doing SMC and another
2825                 * thread doing COW.
2826                 */
2827                ptep_clear_flush(vma, address, page_table);
2828                page_add_new_anon_rmap(new_page, vma, address);
2829                /*
2830                 * We call the notify macro here because, when using secondary
2831                 * mmu page tables (such as kvm shadow page tables), we want the
2832                 * new page to be mapped directly into the secondary page table.
2833                 */
2834                set_pte_at_notify(mm, address, page_table, entry);
2835                update_mmu_cache(vma, address, page_table);
2836                if (old_page) {
2837                        /*
2838                         * Only after switching the pte to the new page may
2839                         * we remove the mapcount here. Otherwise another
2840                         * process may come and find the rmap count decremented
2841                         * before the pte is switched to the new page, and
2842                         * "reuse" the old page writing into it while our pte
2843                         * here still points into it and can be read by other
2844                         * threads.
2845                         *
2846                         * The critical issue is to order this
2847                         * page_remove_rmap with the ptp_clear_flush above.
2848                         * Those stores are ordered by (if nothing else,)
2849                         * the barrier present in the atomic_add_negative
2850                         * in page_remove_rmap.
2851                         *
2852                         * Then the TLB flush in ptep_clear_flush ensures that
2853                         * no process can access the old page before the
2854                         * decremented mapcount is visible. And the old page
2855                         * cannot be reused until after the decremented
2856                         * mapcount is visible. So transitively, TLBs to
2857                         * old page will be flushed before it can be reused.
2858                         */
2859                        page_remove_rmap(old_page);
2860                }
2861
2862                /* Free the old page.. */
2863                new_page = old_page;
2864                ret |= VM_FAULT_WRITE;
2865        } else
2866                mem_cgroup_uncharge_page(new_page);
2867
2868        if (new_page)
2869                page_cache_release(new_page);
2870unlock:
2871        pte_unmap_unlock(page_table, ptl);
2872        if (mmun_end > mmun_start)
2873                mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2874        if (old_page) {
2875                /*
2876                 * Don't let another task, with possibly unlocked vma,
2877                 * keep the mlocked page.
2878                 */
2879                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2880                        lock_page(old_page);    /* LRU manipulation */
2881                        munlock_vma_page(old_page);
2882                        unlock_page(old_page);
2883                }
2884                page_cache_release(old_page);
2885        }
2886        return ret;
2887oom_free_new:
2888        page_cache_release(new_page);
2889oom:
2890        if (old_page)
2891                page_cache_release(old_page);
2892        return VM_FAULT_OOM;
2893
2894unwritable_page:
2895        page_cache_release(old_page);
2896        return ret;
2897}
2898
2899static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2900                unsigned long start_addr, unsigned long end_addr,
2901                struct zap_details *details)
2902{
2903        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2904}
2905
2906static inline void unmap_mapping_range_tree(struct rb_root *root,
2907                                            struct zap_details *details)
2908{
2909        struct vm_area_struct *vma;
2910        pgoff_t vba, vea, zba, zea;
2911
2912        vma_interval_tree_foreach(vma, root,
2913                        details->first_index, details->last_index) {
2914
2915                vba = vma->vm_pgoff;
2916                vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2917                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2918                zba = details->first_index;
2919                if (zba < vba)
2920                        zba = vba;
2921                zea = details->last_index;
2922                if (zea > vea)
2923                        zea = vea;
2924
2925                unmap_mapping_range_vma(vma,
2926                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2927                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2928                                details);
2929        }
2930}
2931
2932static inline void unmap_mapping_range_list(struct list_head *head,
2933                                            struct zap_details *details)
2934{
2935        struct vm_area_struct *vma;
2936
2937        /*
2938         * In nonlinear VMAs there is no correspondence between virtual address
2939         * offset and file offset.  So we must perform an exhaustive search
2940         * across *all* the pages in each nonlinear VMA, not just the pages
2941         * whose virtual address lies outside the file truncation point.
2942         */
2943        list_for_each_entry(vma, head, shared.nonlinear) {
2944                details->nonlinear_vma = vma;
2945                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2946        }
2947}
2948
2949/**
2950 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2951 * @mapping: the address space containing mmaps to be unmapped.
2952 * @holebegin: byte in first page to unmap, relative to the start of
2953 * the underlying file.  This will be rounded down to a PAGE_SIZE
2954 * boundary.  Note that this is different from truncate_pagecache(), which
2955 * must keep the partial page.  In contrast, we must get rid of
2956 * partial pages.
2957 * @holelen: size of prospective hole in bytes.  This will be rounded
2958 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2959 * end of the file.
2960 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2961 * but 0 when invalidating pagecache, don't throw away private data.
2962 */
2963void unmap_mapping_range(struct address_space *mapping,
2964                loff_t const holebegin, loff_t const holelen, int even_cows)
2965{
2966        struct zap_details details;
2967        pgoff_t hba = holebegin >> PAGE_SHIFT;
2968        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2969
2970        /* Check for overflow. */
2971        if (sizeof(holelen) > sizeof(hlen)) {
2972                long long holeend =
2973                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2974                if (holeend & ~(long long)ULONG_MAX)
2975                        hlen = ULONG_MAX - hba + 1;
2976        }
2977
2978        details.check_mapping = even_cows? NULL: mapping;
2979        details.nonlinear_vma = NULL;
2980        details.first_index = hba;
2981        details.last_index = hba + hlen - 1;
2982        if (details.last_index < details.first_index)
2983                details.last_index = ULONG_MAX;
2984
2985
2986        mutex_lock(&mapping->i_mmap_mutex);
2987        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2988                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2989        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2990                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2991        mutex_unlock(&mapping->i_mmap_mutex);
2992}
2993EXPORT_SYMBOL(unmap_mapping_range);
2994
2995/*
2996 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2997 * but allow concurrent faults), and pte mapped but not yet locked.
2998 * We return with mmap_sem still held, but pte unmapped and unlocked.
2999 */
3000static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3001                unsigned long address, pte_t *page_table, pmd_t *pmd,
3002                unsigned int flags, pte_t orig_pte)
3003{
3004        spinlock_t *ptl;
3005        struct page *page, *swapcache;
3006        swp_entry_t entry;
3007        pte_t pte;
3008        int locked;
3009        struct mem_cgroup *ptr;
3010        int exclusive = 0;
3011        int ret = 0;
3012
3013        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3014                goto out;
3015
3016        entry = pte_to_swp_entry(orig_pte);
3017        if (unlikely(non_swap_entry(entry))) {
3018                if (is_migration_entry(entry)) {
3019                        migration_entry_wait(mm, pmd, address);
3020                } else if (is_hwpoison_entry(entry)) {
3021                        ret = VM_FAULT_HWPOISON;
3022                } else {
3023                        print_bad_pte(vma, address, orig_pte, NULL);
3024                        ret = VM_FAULT_SIGBUS;
3025                }
3026                goto out;
3027        }
3028        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3029        page = lookup_swap_cache(entry);
3030        if (!page) {
3031                page = swapin_readahead(entry,
3032                                        GFP_HIGHUSER_MOVABLE, vma, address);
3033                if (!page) {
3034                        /*
3035                         * Back out if somebody else faulted in this pte
3036                         * while we released the pte lock.
3037                         */
3038                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3039                        if (likely(pte_same(*page_table, orig_pte)))
3040                                ret = VM_FAULT_OOM;
3041                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3042                        goto unlock;
3043                }
3044
3045                /* Had to read the page from swap area: Major fault */
3046                ret = VM_FAULT_MAJOR;
3047                count_vm_event(PGMAJFAULT);
3048                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3049        } else if (PageHWPoison(page)) {
3050                /*
3051                 * hwpoisoned dirty swapcache pages are kept for killing
3052                 * owner processes (which may be unknown at hwpoison time)
3053                 */
3054                ret = VM_FAULT_HWPOISON;
3055                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3056                swapcache = page;
3057                goto out_release;
3058        }
3059
3060        swapcache = page;
3061        locked = lock_page_or_retry(page, mm, flags);
3062
3063        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3064        if (!locked) {
3065                ret |= VM_FAULT_RETRY;
3066                goto out_release;
3067        }
3068
3069        /*
3070         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3071         * release the swapcache from under us.  The page pin, and pte_same
3072         * test below, are not enough to exclude that.  Even if it is still
3073         * swapcache, we need to check that the page's swap has not changed.
3074         */
3075        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3076                goto out_page;
3077
3078        page = ksm_might_need_to_copy(page, vma, address);
3079        if (unlikely(!page)) {
3080                ret = VM_FAULT_OOM;
3081                page = swapcache;
3082                goto out_page;
3083        }
3084
3085        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3086                ret = VM_FAULT_OOM;
3087                goto out_page;
3088        }
3089
3090        /*
3091         * Back out if somebody else already faulted in this pte.
3092         */
3093        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3094        if (unlikely(!pte_same(*page_table, orig_pte)))
3095                goto out_nomap;
3096
3097        if (unlikely(!PageUptodate(page))) {
3098                ret = VM_FAULT_SIGBUS;
3099                goto out_nomap;
3100        }
3101
3102        /*
3103         * The page isn't present yet, go ahead with the fault.
3104         *
3105         * Be careful about the sequence of operations here.
3106         * To get its accounting right, reuse_swap_page() must be called
3107         * while the page is counted on swap but not yet in mapcount i.e.
3108         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3109         * must be called after the swap_free(), or it will never succeed.
3110         * Because delete_from_swap_page() may be called by reuse_swap_page(),
3111         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3112         * in page->private. In this case, a record in swap_cgroup  is silently
3113         * discarded at swap_free().
3114         */
3115
3116        inc_mm_counter_fast(mm, MM_ANONPAGES);
3117        dec_mm_counter_fast(mm, MM_SWAPENTS);
3118        pte = mk_pte(page, vma->vm_page_prot);
3119        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3120                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3121                flags &= ~FAULT_FLAG_WRITE;
3122                ret |= VM_FAULT_WRITE;
3123                exclusive = 1;
3124        }
3125        flush_icache_page(vma, page);
3126        set_pte_at(mm, address, page_table, pte);
3127        if (page == swapcache)
3128                do_page_add_anon_rmap(page, vma, address, exclusive);
3129        else /* ksm created a completely new copy */
3130                page_add_new_anon_rmap(page, vma, address);
3131        /* It's better to call commit-charge after rmap is established */
3132        mem_cgroup_commit_charge_swapin(page, ptr);
3133
3134        swap_free(entry);
3135        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3136                try_to_free_swap(page);
3137        unlock_page(page);
3138        if (page != swapcache) {
3139                /*
3140                 * Hold the lock to avoid the swap entry to be reused
3141                 * until we take the PT lock for the pte_same() check
3142                 * (to avoid false positives from pte_same). For
3143                 * further safety release the lock after the swap_free
3144                 * so that the swap count won't change under a
3145                 * parallel locked swapcache.
3146                 */
3147                unlock_page(swapcache);
3148                page_cache_release(swapcache);
3149        }
3150
3151        if (flags & FAULT_FLAG_WRITE) {
3152                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3153                if (ret & VM_FAULT_ERROR)
3154                        ret &= VM_FAULT_ERROR;
3155                goto out;
3156        }
3157
3158        /* No need to invalidate - it was non-present before */
3159        update_mmu_cache(vma, address, page_table);
3160unlock:
3161        pte_unmap_unlock(page_table, ptl);
3162out:
3163        return ret;
3164out_nomap:
3165        mem_cgroup_cancel_charge_swapin(ptr);
3166        pte_unmap_unlock(page_table, ptl);
3167out_page:
3168        unlock_page(page);
3169out_release:
3170        page_cache_release(page);
3171        if (page != swapcache) {
3172                unlock_page(swapcache);
3173                page_cache_release(swapcache);
3174        }
3175        return ret;
3176}
3177
3178/*
3179 * This is like a special single-page "expand_{down|up}wards()",
3180 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3181 * doesn't hit another vma.
3182 */
3183static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3184{
3185        address &= PAGE_MASK;
3186        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3187                struct vm_area_struct *prev = vma->vm_prev;
3188
3189                /*
3190                 * Is there a mapping abutting this one below?
3191                 *
3192                 * That's only ok if it's the same stack mapping
3193                 * that has gotten split..
3194                 */
3195                if (prev && prev->vm_end == address)
3196                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3197
3198                expand_downwards(vma, address - PAGE_SIZE);
3199        }
3200        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3201                struct vm_area_struct *next = vma->vm_next;
3202
3203                /* As VM_GROWSDOWN but s/below/above/ */
3204                if (next && next->vm_start == address + PAGE_SIZE)
3205                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3206
3207                expand_upwards(vma, address + PAGE_SIZE);
3208        }
3209        return 0;
3210}
3211
3212/*
3213 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3214 * but allow concurrent faults), and pte mapped but not yet locked.
3215 * We return with mmap_sem still held, but pte unmapped and unlocked.
3216 */
3217static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3218                unsigned long address, pte_t *page_table, pmd_t *pmd,
3219                unsigned int flags)
3220{
3221        struct page *page;
3222        spinlock_t *ptl;
3223        pte_t entry;
3224
3225        pte_unmap(page_table);
3226
3227        /* Check if we need to add a guard page to the stack */
3228        if (check_stack_guard_page(vma, address) < 0)
3229                return VM_FAULT_SIGBUS;
3230
3231        /* Use the zero-page for reads */
3232        if (!(flags & FAULT_FLAG_WRITE)) {
3233                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3234                                                vma->vm_page_prot));
3235                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3236                if (!pte_none(*page_table))
3237                        goto unlock;
3238                goto setpte;
3239        }
3240
3241        /* Allocate our own private page. */
3242        if (unlikely(anon_vma_prepare(vma)))
3243                goto oom;
3244        page = alloc_zeroed_user_highpage_movable(vma, address);
3245        if (!page)
3246                goto oom;
3247        __SetPageUptodate(page);
3248
3249        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3250                goto oom_free_page;
3251
3252        entry = mk_pte(page, vma->vm_page_prot);
3253        if (vma->vm_flags & VM_WRITE)
3254                entry = pte_mkwrite(pte_mkdirty(entry));
3255
3256        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3257        if (!pte_none(*page_table))
3258                goto release;
3259
3260        inc_mm_counter_fast(mm, MM_ANONPAGES);
3261        page_add_new_anon_rmap(page, vma, address);
3262setpte:
3263        set_pte_at(mm, address, page_table, entry);
3264
3265        /* No need to invalidate - it was non-present before */
3266        update_mmu_cache(vma, address, page_table);
3267unlock:
3268        pte_unmap_unlock(page_table, ptl);
3269        return 0;
3270release:
3271        mem_cgroup_uncharge_page(page);
3272        page_cache_release(page);
3273        goto unlock;
3274oom_free_page:
3275        page_cache_release(page);
3276oom:
3277        return VM_FAULT_OOM;
3278}
3279
3280/*
3281 * __do_fault() tries to create a new page mapping. It aggressively
3282 * tries to share with existing pages, but makes a separate copy if
3283 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3284 * the next page fault.
3285 *
3286 * As this is called only for pages that do not currently exist, we
3287 * do not need to flush old virtual caches or the TLB.
3288 *
3289 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3290 * but allow concurrent faults), and pte neither mapped nor locked.
3291 * We return with mmap_sem still held, but pte unmapped and unlocked.
3292 */
3293static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3294                unsigned long address, pmd_t *pmd,
3295                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3296{
3297        pte_t *page_table;
3298        spinlock_t *ptl;
3299        struct page *page;
3300        struct page *cow_page;
3301        pte_t entry;
3302        int anon = 0;
3303        struct page *dirty_page = NULL;
3304        struct vm_fault vmf;
3305        int ret;
3306        int page_mkwrite = 0;
3307
3308        /*
3309         * If we do COW later, allocate page befor taking lock_page()
3310         * on the file cache page. This will reduce lock holding time.
3311         */
3312        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3313
3314                if (unlikely(anon_vma_prepare(vma)))
3315                        return VM_FAULT_OOM;
3316
3317                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3318                if (!cow_page)
3319                        return VM_FAULT_OOM;
3320
3321                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3322                        page_cache_release(cow_page);
3323                        return VM_FAULT_OOM;
3324                }
3325        } else
3326                cow_page = NULL;
3327
3328        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3329        vmf.pgoff = pgoff;
3330        vmf.flags = flags;
3331        vmf.page = NULL;
3332
3333        ret = vma->vm_ops->fault(vma, &vmf);
3334        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3335                            VM_FAULT_RETRY)))
3336                goto uncharge_out;
3337
3338        if (unlikely(PageHWPoison(vmf.page))) {
3339                if (ret & VM_FAULT_LOCKED)
3340                        unlock_page(vmf.page);
3341                ret = VM_FAULT_HWPOISON;
3342                goto uncharge_out;
3343        }
3344
3345        /*
3346         * For consistency in subsequent calls, make the faulted page always
3347         * locked.
3348         */
3349        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3350                lock_page(vmf.page);
3351        else
3352                VM_BUG_ON(!PageLocked(vmf.page));
3353
3354        /*
3355         * Should we do an early C-O-W break?
3356         */
3357        page = vmf.page;
3358        if (flags & FAULT_FLAG_WRITE) {
3359                if (!(vma->vm_flags & VM_SHARED)) {
3360                        page = cow_page;
3361                        anon = 1;
3362                        copy_user_highpage(page, vmf.page, address, vma);
3363                        __SetPageUptodate(page);
3364                } else {
3365                        /*
3366                         * If the page will be shareable, see if the backing
3367                         * address space wants to know that the page is about
3368                         * to become writable
3369                         */
3370                        if (vma->vm_ops->page_mkwrite) {
3371                                int tmp;
3372
3373                                unlock_page(page);
3374                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3375                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3376                                if (unlikely(tmp &
3377                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3378                                        ret = tmp;
3379                                        goto unwritable_page;
3380                                }
3381                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3382                                        lock_page(page);
3383                                        if (!page->mapping) {
3384                                                ret = 0; /* retry the fault */
3385                                                unlock_page(page);
3386                                                goto unwritable_page;
3387                                        }
3388                                } else
3389                                        VM_BUG_ON(!PageLocked(page));
3390                                page_mkwrite = 1;
3391                        }
3392                }
3393
3394        }
3395
3396        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3397
3398        /*
3399         * This silly early PAGE_DIRTY setting removes a race
3400         * due to the bad i386 page protection. But it's valid
3401         * for other architectures too.
3402         *
3403         * Note that if FAULT_FLAG_WRITE is set, we either now have
3404         * an exclusive copy of the page, or this is a shared mapping,
3405         * so we can make it writable and dirty to avoid having to
3406         * handle that later.
3407         */
3408        /* Only go through if we didn't race with anybody else... */
3409        if (likely(pte_same(*page_table, orig_pte))) {
3410                flush_icache_page(vma, page);
3411                entry = mk_pte(page, vma->vm_page_prot);
3412                if (flags & FAULT_FLAG_WRITE)
3413                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3414                if (anon) {
3415                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3416                        page_add_new_anon_rmap(page, vma, address);
3417                } else {
3418                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3419                        page_add_file_rmap(page);
3420                        if (flags & FAULT_FLAG_WRITE) {
3421                                dirty_page = page;
3422                                get_page(dirty_page);
3423                        }
3424                }
3425                set_pte_at(mm, address, page_table, entry);
3426
3427                /* no need to invalidate: a not-present page won't be cached */
3428                update_mmu_cache(vma, address, page_table);
3429        } else {
3430                if (cow_page)
3431                        mem_cgroup_uncharge_page(cow_page);
3432                if (anon)
3433                        page_cache_release(page);
3434                else
3435                        anon = 1; /* no anon but release faulted_page */
3436        }
3437
3438        pte_unmap_unlock(page_table, ptl);
3439
3440        if (dirty_page) {
3441                struct address_space *mapping = page->mapping;
3442                int dirtied = 0;
3443
3444                if (set_page_dirty(dirty_page))
3445                        dirtied = 1;
3446                unlock_page(dirty_page);
3447                put_page(dirty_page);
3448                if ((dirtied || page_mkwrite) && mapping) {
3449                        /*
3450                         * Some device drivers do not set page.mapping but still
3451                         * dirty their pages
3452                         */
3453                        balance_dirty_pages_ratelimited(mapping);
3454                }
3455
3456                /* file_update_time outside page_lock */
3457                if (vma->vm_file && !page_mkwrite)
3458                        file_update_time(vma->vm_file);
3459        } else {
3460                unlock_page(vmf.page);
3461                if (anon)
3462                        page_cache_release(vmf.page);
3463        }
3464
3465        return ret;
3466
3467unwritable_page:
3468        page_cache_release(page);
3469        return ret;
3470uncharge_out:
3471        /* fs's fault handler get error */
3472        if (cow_page) {
3473                mem_cgroup_uncharge_page(cow_page);
3474                page_cache_release(cow_page);
3475        }
3476        return ret;
3477}
3478
3479static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3480                unsigned long address, pte_t *page_table, pmd_t *pmd,
3481                unsigned int flags, pte_t orig_pte)
3482{
3483        pgoff_t pgoff = (((address & PAGE_MASK)
3484                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3485
3486        pte_unmap(page_table);
3487        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3488}
3489
3490/*
3491 * Fault of a previously existing named mapping. Repopulate the pte
3492 * from the encoded file_pte if possible. This enables swappable
3493 * nonlinear vmas.
3494 *
3495 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3496 * but allow concurrent faults), and pte mapped but not yet locked.
3497 * We return with mmap_sem still held, but pte unmapped and unlocked.
3498 */
3499static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3500                unsigned long address, pte_t *page_table, pmd_t *pmd,
3501                unsigned int flags, pte_t orig_pte)
3502{
3503        pgoff_t pgoff;
3504
3505        flags |= FAULT_FLAG_NONLINEAR;
3506
3507        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3508                return 0;
3509
3510        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3511                /*
3512                 * Page table corrupted: show pte and kill process.
3513                 */
3514                print_bad_pte(vma, address, orig_pte, NULL);
3515                return VM_FAULT_SIGBUS;
3516        }
3517
3518        pgoff = pte_to_pgoff(orig_pte);
3519        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3520}
3521
3522int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3523                                unsigned long addr, int current_nid)
3524{
3525        get_page(page);
3526
3527        count_vm_numa_event(NUMA_HINT_FAULTS);
3528        if (current_nid == numa_node_id())
3529                count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3530
3531        return mpol_misplaced(page, vma, addr);
3532}
3533
3534int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3535                   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3536{
3537        struct page *page = NULL;
3538        spinlock_t *ptl;
3539        int current_nid = -1;
3540        int target_nid;
3541        bool migrated = false;
3542
3543        /*
3544        * The "pte" at this point cannot be used safely without
3545        * validation through pte_unmap_same(). It's of NUMA type but
3546        * the pfn may be screwed if the read is non atomic.
3547        *
3548        * ptep_modify_prot_start is not called as this is clearing
3549        * the _PAGE_NUMA bit and it is not really expected that there
3550        * would be concurrent hardware modifications to the PTE.
3551        */
3552        ptl = pte_lockptr(mm, pmd);
3553        spin_lock(ptl);
3554        if (unlikely(!pte_same(*ptep, pte))) {
3555                pte_unmap_unlock(ptep, ptl);
3556                goto out;
3557        }
3558
3559        pte = pte_mknonnuma(pte);
3560        set_pte_at(mm, addr, ptep, pte);
3561        update_mmu_cache(vma, addr, ptep);
3562
3563        page = vm_normal_page(vma, addr, pte);
3564        if (!page) {
3565                pte_unmap_unlock(ptep, ptl);
3566                return 0;
3567        }
3568
3569        current_nid = page_to_nid(page);
3570        target_nid = numa_migrate_prep(page, vma, addr, current_nid);
3571        pte_unmap_unlock(ptep, ptl);
3572        if (target_nid == -1) {
3573                /*
3574                 * Account for the fault against the current node if it not
3575                 * being replaced regardless of where the page is located.
3576                 */
3577                current_nid = numa_node_id();
3578                put_page(page);
3579                goto out;
3580        }
3581
3582        /* Migrate to the requested node */
3583        migrated = migrate_misplaced_page(page, target_nid);
3584        if (migrated)
3585                current_nid = target_nid;
3586
3587out:
3588        if (current_nid != -1)
3589                task_numa_fault(current_nid, 1, migrated);
3590        return 0;
3591}
3592
3593/* NUMA hinting page fault entry point for regular pmds */
3594#ifdef CONFIG_NUMA_BALANCING
3595static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3596                     unsigned long addr, pmd_t *pmdp)
3597{
3598        pmd_t pmd;
3599        pte_t *pte, *orig_pte;
3600        unsigned long _addr = addr & PMD_MASK;
3601        unsigned long offset;
3602        spinlock_t *ptl;
3603        bool numa = false;
3604        int local_nid = numa_node_id();
3605
3606        spin_lock(&mm->page_table_lock);
3607        pmd = *pmdp;
3608        if (pmd_numa(pmd)) {
3609                set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3610                numa = true;
3611        }
3612        spin_unlock(&mm->page_table_lock);
3613
3614        if (!numa)
3615                return 0;
3616
3617        /* we're in a page fault so some vma must be in the range */
3618        BUG_ON(!vma);
3619        BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3620        offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3621        VM_BUG_ON(offset >= PMD_SIZE);
3622        orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3623        pte += offset >> PAGE_SHIFT;
3624        for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3625                pte_t pteval = *pte;
3626                struct page *page;
3627                int curr_nid = local_nid;
3628                int target_nid;
3629                bool migrated;
3630                if (!pte_present(pteval))
3631                        continue;
3632                if (!pte_numa(pteval))
3633                        continue;
3634                if (addr >= vma->vm_end) {
3635                        vma = find_vma(mm, addr);
3636                        /* there's a pte present so there must be a vma */
3637                        BUG_ON(!vma);
3638                        BUG_ON(addr < vma->vm_start);
3639                }
3640                if (pte_numa(pteval)) {
3641                        pteval = pte_mknonnuma(pteval);
3642                        set_pte_at(mm, addr, pte, pteval);
3643                }
3644                page = vm_normal_page(vma, addr, pteval);
3645                if (unlikely(!page))
3646                        continue;
3647                /* only check non-shared pages */
3648                if (unlikely(page_mapcount(page) != 1))
3649                        continue;
3650
3651                /*
3652                 * Note that the NUMA fault is later accounted to either
3653                 * the node that is currently running or where the page is
3654                 * migrated to.
3655                 */
3656                curr_nid = local_nid;
3657                target_nid = numa_migrate_prep(page, vma, addr,
3658                                               page_to_nid(page));
3659                if (target_nid == -1) {
3660                        put_page(page);
3661                        continue;
3662                }
3663
3664                /* Migrate to the requested node */
3665                pte_unmap_unlock(pte, ptl);
3666                migrated = migrate_misplaced_page(page, target_nid);
3667                if (migrated)
3668                        curr_nid = target_nid;
3669                task_numa_fault(curr_nid, 1, migrated);
3670
3671                pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3672        }
3673        pte_unmap_unlock(orig_pte, ptl);
3674
3675        return 0;
3676}
3677#else
3678static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3679                     unsigned long addr, pmd_t *pmdp)
3680{
3681        BUG();
3682        return 0;
3683}
3684#endif /* CONFIG_NUMA_BALANCING */
3685
3686/*
3687 * These routines also need to handle stuff like marking pages dirty
3688 * and/or accessed for architectures that don't do it in hardware (most
3689 * RISC architectures).  The early dirtying is also good on the i386.
3690 *
3691 * There is also a hook called "update_mmu_cache()" that architectures
3692 * with external mmu caches can use to update those (ie the Sparc or
3693 * PowerPC hashed page tables that act as extended TLBs).
3694 *
3695 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3696 * but allow concurrent faults), and pte mapped but not yet locked.
3697 * We return with mmap_sem still held, but pte unmapped and unlocked.
3698 */
3699int handle_pte_fault(struct mm_struct *mm,
3700                     struct vm_area_struct *vma, unsigned long address,
3701                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3702{
3703        pte_t entry;
3704        spinlock_t *ptl;
3705
3706        entry = *pte;
3707        if (!pte_present(entry)) {
3708                if (pte_none(entry)) {
3709                        if (vma->vm_ops) {
3710                                if (likely(vma->vm_ops->fault))
3711                                        return do_linear_fault(mm, vma, address,
3712                                                pte, pmd, flags, entry);
3713                        }
3714                        return do_anonymous_page(mm, vma, address,
3715                                                 pte, pmd, flags);
3716                }
3717                if (pte_file(entry))
3718                        return do_nonlinear_fault(mm, vma, address,
3719                                        pte, pmd, flags, entry);
3720                return do_swap_page(mm, vma, address,
3721                                        pte, pmd, flags, entry);
3722        }
3723
3724        if (pte_numa(entry))
3725                return do_numa_page(mm, vma, address, entry, pte, pmd);
3726
3727        ptl = pte_lockptr(mm, pmd);
3728        spin_lock(ptl);
3729        if (unlikely(!pte_same(*pte, entry)))
3730                goto unlock;
3731        if (flags & FAULT_FLAG_WRITE) {
3732                if (!pte_write(entry))
3733                        return do_wp_page(mm, vma, address,
3734                                        pte, pmd, ptl, entry);
3735                entry = pte_mkdirty(entry);
3736        }
3737        entry = pte_mkyoung(entry);
3738        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3739                update_mmu_cache(vma, address, pte);
3740        } else {
3741                /*
3742                 * This is needed only for protection faults but the arch code
3743                 * is not yet telling us if this is a protection fault or not.
3744                 * This still avoids useless tlb flushes for .text page faults
3745                 * with threads.
3746                 */
3747                if (flags & FAULT_FLAG_WRITE)
3748                        flush_tlb_fix_spurious_fault(vma, address);
3749        }
3750unlock:
3751        pte_unmap_unlock(pte, ptl);
3752        return 0;
3753}
3754
3755/*
3756 * By the time we get here, we already hold the mm semaphore
3757 */
3758int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3759                unsigned long address, unsigned int flags)
3760{
3761        pgd_t *pgd;
3762        pud_t *pud;
3763        pmd_t *pmd;
3764        pte_t *pte;
3765
3766        __set_current_state(TASK_RUNNING);
3767
3768        count_vm_event(PGFAULT);
3769        mem_cgroup_count_vm_event(mm, PGFAULT);
3770
3771        /* do counter updates before entering really critical section. */
3772        check_sync_rss_stat(current);
3773
3774        if (unlikely(is_vm_hugetlb_page(vma)))
3775                return hugetlb_fault(mm, vma, address, flags);
3776
3777retry:
3778        pgd = pgd_offset(mm, address);
3779        pud = pud_alloc(mm, pgd, address);
3780        if (!pud)
3781                return VM_FAULT_OOM;
3782        pmd = pmd_alloc(mm, pud, address);
3783        if (!pmd)
3784                return VM_FAULT_OOM;
3785        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3786                if (!vma->vm_ops)
3787                        return do_huge_pmd_anonymous_page(mm, vma, address,
3788                                                          pmd, flags);
3789        } else {
3790                pmd_t orig_pmd = *pmd;
3791                int ret;
3792
3793                barrier();
3794                if (pmd_trans_huge(orig_pmd)) {
3795                        unsigned int dirty = flags & FAULT_FLAG_WRITE;
3796
3797                        /*
3798                         * If the pmd is splitting, return and retry the
3799                         * the fault.  Alternative: wait until the split
3800                         * is done, and goto retry.
3801                         */
3802                        if (pmd_trans_splitting(orig_pmd))
3803                                return 0;
3804
3805                        if (pmd_numa(orig_pmd))
3806                                return do_huge_pmd_numa_page(mm, vma, address,
3807                                                             orig_pmd, pmd);
3808
3809                        if (dirty && !pmd_write(orig_pmd)) {
3810                                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3811                                                          orig_pmd);
3812                                /*
3813                                 * If COW results in an oom, the huge pmd will
3814                                 * have been split, so retry the fault on the
3815                                 * pte for a smaller charge.
3816                                 */
3817                                if (unlikely(ret & VM_FAULT_OOM))
3818                                        goto retry;
3819                                return ret;
3820                        } else {
3821                                huge_pmd_set_accessed(mm, vma, address, pmd,
3822                                                      orig_pmd, dirty);
3823                        }
3824
3825                        return 0;
3826                }
3827        }
3828
3829        if (pmd_numa(*pmd))
3830                return do_pmd_numa_page(mm, vma, address, pmd);
3831
3832        /*
3833         * Use __pte_alloc instead of pte_alloc_map, because we can't
3834         * run pte_offset_map on the pmd, if an huge pmd could
3835         * materialize from under us from a different thread.
3836         */
3837        if (unlikely(pmd_none(*pmd)) &&
3838            unlikely(__pte_alloc(mm, vma, pmd, address)))
3839                return VM_FAULT_OOM;
3840        /* if an huge pmd materialized from under us just retry later */
3841        if (unlikely(pmd_trans_huge(*pmd)))
3842                return 0;
3843        /*
3844         * A regular pmd is established and it can't morph into a huge pmd
3845         * from under us anymore at this point because we hold the mmap_sem
3846         * read mode and khugepaged takes it in write mode. So now it's
3847         * safe to run pte_offset_map().
3848         */
3849        pte = pte_offset_map(pmd, address);
3850
3851        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3852}
3853
3854#ifndef __PAGETABLE_PUD_FOLDED
3855/*
3856 * Allocate page upper directory.
3857 * We've already handled the fast-path in-line.
3858 */
3859int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3860{
3861        pud_t *new = pud_alloc_one(mm, address);
3862        if (!new)
3863                return -ENOMEM;
3864
3865        smp_wmb(); /* See comment in __pte_alloc */
3866
3867        spin_lock(&mm->page_table_lock);
3868        if (pgd_present(*pgd))          /* Another has populated it */
3869                pud_free(mm, new);
3870        else
3871                pgd_populate(mm, pgd, new);
3872        spin_unlock(&mm->page_table_lock);
3873        return 0;
3874}
3875#endif /* __PAGETABLE_PUD_FOLDED */
3876
3877#ifndef __PAGETABLE_PMD_FOLDED
3878/*
3879 * Allocate page middle directory.
3880 * We've already handled the fast-path in-line.
3881 */
3882int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3883{
3884        pmd_t *new = pmd_alloc_one(mm, address);
3885        if (!new)
3886                return -ENOMEM;
3887
3888        smp_wmb(); /* See comment in __pte_alloc */
3889
3890        spin_lock(&mm->page_table_lock);
3891#ifndef __ARCH_HAS_4LEVEL_HACK
3892        if (pud_present(*pud))          /* Another has populated it */
3893                pmd_free(mm, new);
3894        else
3895                pud_populate(mm, pud, new);
3896#else
3897        if (pgd_present(*pud))          /* Another has populated it */
3898                pmd_free(mm, new);
3899        else
3900                pgd_populate(mm, pud, new);
3901#endif /* __ARCH_HAS_4LEVEL_HACK */
3902        spin_unlock(&mm->page_table_lock);
3903        return 0;
3904}
3905#endif /* __PAGETABLE_PMD_FOLDED */
3906
3907#if !defined(__HAVE_ARCH_GATE_AREA)
3908
3909#if defined(AT_SYSINFO_EHDR)
3910static struct vm_area_struct gate_vma;
3911
3912static int __init gate_vma_init(void)
3913{
3914        gate_vma.vm_mm = NULL;
3915        gate_vma.vm_start = FIXADDR_USER_START;
3916        gate_vma.vm_end = FIXADDR_USER_END;
3917        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3918        gate_vma.vm_page_prot = __P101;
3919
3920        return 0;
3921}
3922__initcall(gate_vma_init);
3923#endif
3924
3925struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3926{
3927#ifdef AT_SYSINFO_EHDR
3928        return &gate_vma;
3929#else
3930        return NULL;
3931#endif
3932}
3933
3934int in_gate_area_no_mm(unsigned long addr)
3935{
3936#ifdef AT_SYSINFO_EHDR
3937        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3938                return 1;
3939#endif
3940        return 0;
3941}
3942
3943#endif  /* __HAVE_ARCH_GATE_AREA */
3944
3945static int __follow_pte(struct mm_struct *mm, unsigned long address,
3946                pte_t **ptepp, spinlock_t **ptlp)
3947{
3948        pgd_t *pgd;
3949        pud_t *pud;
3950        pmd_t *pmd;
3951        pte_t *ptep;
3952
3953        pgd = pgd_offset(mm, address);
3954        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3955                goto out;
3956
3957        pud = pud_offset(pgd, address);
3958        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3959                goto out;
3960
3961        pmd = pmd_offset(pud, address);
3962        VM_BUG_ON(pmd_trans_huge(*pmd));
3963        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3964                goto out;
3965
3966        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3967        if (pmd_huge(*pmd))
3968                goto out;
3969
3970        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3971        if (!ptep)
3972                goto out;
3973        if (!pte_present(*ptep))
3974                goto unlock;
3975        *ptepp = ptep;
3976        return 0;
3977unlock:
3978        pte_unmap_unlock(ptep, *ptlp);
3979out:
3980        return -EINVAL;
3981}
3982
3983static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3984                             pte_t **ptepp, spinlock_t **ptlp)
3985{
3986        int res;
3987
3988        /* (void) is needed to make gcc happy */
3989        (void) __cond_lock(*ptlp,
3990                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
3991        return res;
3992}
3993
3994/**
3995 * follow_pfn - look up PFN at a user virtual address
3996 * @vma: memory mapping
3997 * @address: user virtual address
3998 * @pfn: location to store found PFN
3999 *
4000 * Only IO mappings and raw PFN mappings are allowed.
4001 *
4002 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4003 */
4004int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4005        unsigned long *pfn)
4006{
4007        int ret = -EINVAL;
4008        spinlock_t *ptl;
4009        pte_t *ptep;
4010
4011        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4012                return ret;
4013
4014        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4015        if (ret)
4016                return ret;
4017        *pfn = pte_pfn(*ptep);
4018        pte_unmap_unlock(ptep, ptl);
4019        return 0;
4020}
4021EXPORT_SYMBOL(follow_pfn);
4022
4023#ifdef CONFIG_HAVE_IOREMAP_PROT
4024int follow_phys(struct vm_area_struct *vma,
4025                unsigned long address, unsigned int flags,
4026                unsigned long *prot, resource_size_t *phys)
4027{
4028        int ret = -EINVAL;
4029        pte_t *ptep, pte;
4030        spinlock_t *ptl;
4031
4032        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4033                goto out;
4034
4035        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4036                goto out;
4037        pte = *ptep;
4038
4039        if ((flags & FOLL_WRITE) && !pte_write(pte))
4040                goto unlock;
4041
4042        *prot = pgprot_val(pte_pgprot(pte));
4043        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4044
4045        ret = 0;
4046unlock:
4047        pte_unmap_unlock(ptep, ptl);
4048out:
4049        return ret;
4050}
4051
4052int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4053                        void *buf, int len, int write)
4054{
4055        resource_size_t phys_addr;
4056        unsigned long prot = 0;
4057        void __iomem *maddr;
4058        int offset = addr & (PAGE_SIZE-1);
4059
4060        if (follow_phys(vma, addr, write, &prot, &phys_addr))
4061                return -EINVAL;
4062
4063        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4064        if (write)
4065                memcpy_toio(maddr + offset, buf, len);
4066        else
4067                memcpy_fromio(buf, maddr + offset, len);
4068        iounmap(maddr);
4069
4070        return len;
4071}
4072#endif
4073
4074/*
4075 * Access another process' address space as given in mm.  If non-NULL, use the
4076 * given task for page fault accounting.
4077 */
4078static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4079                unsigned long addr, void *buf, int len, int write)
4080{
4081        struct vm_area_struct *vma;
4082        void *old_buf = buf;
4083
4084        down_read(&mm->mmap_sem);
4085        /* ignore errors, just check how much was successfully transferred */
4086        while (len) {
4087                int bytes, ret, offset;
4088                void *maddr;
4089                struct page *page = NULL;
4090
4091                ret = get_user_pages(tsk, mm, addr, 1,
4092                                write, 1, &page, &vma);
4093                if (ret <= 0) {
4094                        /*
4095                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
4096                         * we can access using slightly different code.
4097                         */
4098#ifdef CONFIG_HAVE_IOREMAP_PROT
4099                        vma = find_vma(mm, addr);
4100                        if (!vma || vma->vm_start > addr)
4101                                break;
4102                        if (vma->vm_ops && vma->vm_ops->access)
4103                                ret = vma->vm_ops->access(vma, addr, buf,
4104                                                          len, write);
4105                        if (ret <= 0)
4106#endif
4107                                break;
4108                        bytes = ret;
4109                } else {
4110                        bytes = len;
4111                        offset = addr & (PAGE_SIZE-1);
4112                        if (bytes > PAGE_SIZE-offset)
4113                                bytes = PAGE_SIZE-offset;
4114
4115                        maddr = kmap(page);
4116                        if (write) {
4117                                copy_to_user_page(vma, page, addr,
4118                                                  maddr + offset, buf, bytes);
4119                                set_page_dirty_lock(page);
4120                        } else {
4121                                copy_from_user_page(vma, page, addr,
4122                                                    buf, maddr + offset, bytes);
4123                        }
4124                        kunmap(page);
4125                        page_cache_release(page);
4126                }
4127                len -= bytes;
4128                buf += bytes;
4129                addr += bytes;
4130        }
4131        up_read(&mm->mmap_sem);
4132
4133        return buf - old_buf;
4134}
4135
4136/**
4137 * access_remote_vm - access another process' address space
4138 * @mm:         the mm_struct of the target address space
4139 * @addr:       start address to access
4140 * @buf:        source or destination buffer
4141 * @len:        number of bytes to transfer
4142 * @write:      whether the access is a write
4143 *
4144 * The caller must hold a reference on @mm.
4145 */
4146int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4147                void *buf, int len, int write)
4148{
4149        return __access_remote_vm(NULL, mm, addr, buf, len, write);
4150}
4151
4152/*
4153 * Access another process' address space.
4154 * Source/target buffer must be kernel space,
4155 * Do not walk the page table directly, use get_user_pages
4156 */
4157int access_process_vm(struct task_struct *tsk, unsigned long addr,
4158                void *buf, int len, int write)
4159{
4160        struct mm_struct *mm;
4161        int ret;
4162
4163        mm = get_task_mm(tsk);
4164        if (!mm)
4165                return 0;
4166
4167        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4168        mmput(mm);
4169
4170        return ret;
4171}
4172
4173/*
4174 * Print the name of a VMA.
4175 */
4176void print_vma_addr(char *prefix, unsigned long ip)
4177{
4178        struct mm_struct *mm = current->mm;
4179        struct vm_area_struct *vma;
4180
4181        /*
4182         * Do not print if we are in atomic
4183         * contexts (in exception stacks, etc.):
4184         */
4185        if (preempt_count())
4186                return;
4187
4188        down_read(&mm->mmap_sem);
4189        vma = find_vma(mm, ip);
4190        if (vma && vma->vm_file) {
4191                struct file *f = vma->vm_file;
4192                char *buf = (char *)__get_free_page(GFP_KERNEL);
4193                if (buf) {
4194                        char *p;
4195
4196                        p = d_path(&f->f_path, buf, PAGE_SIZE);
4197                        if (IS_ERR(p))
4198                                p = "?";
4199                        printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4200                                        vma->vm_start,
4201                                        vma->vm_end - vma->vm_start);
4202                        free_page((unsigned long)buf);
4203                }
4204        }
4205        up_read(&mm->mmap_sem);
4206}
4207
4208#ifdef CONFIG_PROVE_LOCKING
4209void might_fault(void)
4210{
4211        /*
4212         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4213         * holding the mmap_sem, this is safe because kernel memory doesn't
4214         * get paged out, therefore we'll never actually fault, and the
4215         * below annotations will generate false positives.
4216         */
4217        if (segment_eq(get_fs(), KERNEL_DS))
4218                return;
4219
4220        might_sleep();
4221        /*
4222         * it would be nicer only to annotate paths which are not under
4223         * pagefault_disable, however that requires a larger audit and
4224         * providing helpers like get_user_atomic.
4225         */
4226        if (!in_atomic() && current->mm)
4227                might_lock_read(&current->mm->mmap_sem);
4228}
4229EXPORT_SYMBOL(might_fault);
4230#endif
4231
4232#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4233static void clear_gigantic_page(struct page *page,
4234                                unsigned long addr,
4235                                unsigned int pages_per_huge_page)
4236{
4237        int i;
4238        struct page *p = page;
4239
4240        might_sleep();
4241        for (i = 0; i < pages_per_huge_page;
4242             i++, p = mem_map_next(p, page, i)) {
4243                cond_resched();
4244                clear_user_highpage(p, addr + i * PAGE_SIZE);
4245        }
4246}
4247void clear_huge_page(struct page *page,
4248                     unsigned long addr, unsigned int pages_per_huge_page)
4249{
4250        int i;
4251
4252        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4253                clear_gigantic_page(page, addr, pages_per_huge_page);
4254                return;
4255        }
4256
4257        might_sleep();
4258        for (i = 0; i < pages_per_huge_page; i++) {
4259                cond_resched();
4260                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4261        }
4262}
4263
4264static void copy_user_gigantic_page(struct page *dst, struct page *src,
4265                                    unsigned long addr,
4266                                    struct vm_area_struct *vma,
4267                                    unsigned int pages_per_huge_page)
4268{
4269        int i;
4270        struct page *dst_base = dst;
4271        struct page *src_base = src;
4272
4273        for (i = 0; i < pages_per_huge_page; ) {
4274                cond_resched();
4275                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4276
4277                i++;
4278                dst = mem_map_next(dst, dst_base, i);
4279                src = mem_map_next(src, src_base, i);
4280        }
4281}
4282
4283void copy_user_huge_page(struct page *dst, struct page *src,
4284                         unsigned long addr, struct vm_area_struct *vma,
4285                         unsigned int pages_per_huge_page)
4286{
4287        int i;
4288
4289        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4290                copy_user_gigantic_page(dst, src, addr, vma,
4291                                        pages_per_huge_page);
4292                return;
4293        }
4294
4295        might_sleep();
4296        for (i = 0; i < pages_per_huge_page; i++) {
4297                cond_resched();
4298                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4299        }
4300}
4301#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4302
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.