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