linux/mm/memory.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *              Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *              (Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/hugetlb.h>
  44#include <linux/mman.h>
  45#include <linux/swap.h>
  46#include <linux/highmem.h>
  47#include <linux/pagemap.h>
  48#include <linux/ksm.h>
  49#include <linux/rmap.h>
  50#include <linux/export.h>
  51#include <linux/delayacct.h>
  52#include <linux/init.h>
  53#include <linux/writeback.h>
  54#include <linux/memcontrol.h>
  55#include <linux/mmu_notifier.h>
  56#include <linux/kallsyms.h>
  57#include <linux/swapops.h>
  58#include <linux/elf.h>
  59#include <linux/gfp.h>
  60#include <linux/migrate.h>
  61#include <linux/string.h>
  62
  63#include <asm/io.h>
  64#include <asm/pgalloc.h>
  65#include <asm/uaccess.h>
  66#include <asm/tlb.h>
  67#include <asm/tlbflush.h>
  68#include <asm/pgtable.h>
  69
  70#include "internal.h"
  71
  72#ifdef LAST_NID_NOT_IN_PAGE_FLAGS
  73#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
  74#endif
  75
  76#ifndef CONFIG_NEED_MULTIPLE_NODES
  77/* use the per-pgdat data instead for discontigmem - mbligh */
  78unsigned long max_mapnr;
  79struct page *mem_map;
  80
  81EXPORT_SYMBOL(max_mapnr);
  82EXPORT_SYMBOL(mem_map);
  83#endif
  84
  85/*
  86 * A number of key systems in x86 including ioremap() rely on the assumption
  87 * that high_memory defines the upper bound on direct map memory, then end
  88 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  90 * and ZONE_HIGHMEM.
  91 */
  92void * high_memory;
  93
  94EXPORT_SYMBOL(high_memory);
  95
  96/*
  97 * Randomize the address space (stacks, mmaps, brk, etc.).
  98 *
  99 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 100 *   as ancient (libc5 based) binaries can segfault. )
 101 */
 102int randomize_va_space __read_mostly =
 103#ifdef CONFIG_COMPAT_BRK
 104                                        1;
 105#else
 106                                        2;
 107#endif
 108
 109static int __init disable_randmaps(char *s)
 110{
 111        randomize_va_space = 0;
 112        return 1;
 113}
 114__setup("norandmaps", disable_randmaps);
 115
 116unsigned long zero_pfn __read_mostly;
 117unsigned long highest_memmap_pfn __read_mostly;
 118
 119/*
 120 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 121 */
 122static int __init init_zero_pfn(void)
 123{
 124        zero_pfn = page_to_pfn(ZERO_PAGE(0));
 125        return 0;
 126}
 127core_initcall(init_zero_pfn);
 128
 129
 130#if defined(SPLIT_RSS_COUNTING)
 131
 132void sync_mm_rss(struct mm_struct *mm)
 133{
 134        int i;
 135
 136        for (i = 0; i < NR_MM_COUNTERS; i++) {
 137                if (current->rss_stat.count[i]) {
 138                        add_mm_counter(mm, i, current->rss_stat.count[i]);
 139                        current->rss_stat.count[i] = 0;
 140                }
 141        }
 142        current->rss_stat.events = 0;
 143}
 144
 145static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 146{
 147        struct task_struct *task = current;
 148
 149        if (likely(task->mm == mm))
 150                task->rss_stat.count[member] += val;
 151        else
 152                add_mm_counter(mm, member, val);
 153}
 154#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 155#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 156
 157/* sync counter once per 64 page faults */
 158#define TASK_RSS_EVENTS_THRESH  (64)
 159static void check_sync_rss_stat(struct task_struct *task)
 160{
 161        if (unlikely(task != current))
 162                return;
 163        if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 164                sync_mm_rss(task->mm);
 165}
 166#else /* SPLIT_RSS_COUNTING */
 167
 168#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 169#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 170
 171static void check_sync_rss_stat(struct task_struct *task)
 172{
 173}
 174
 175#endif /* SPLIT_RSS_COUNTING */
 176
 177#ifdef HAVE_GENERIC_MMU_GATHER
 178
 179static int tlb_next_batch(struct mmu_gather *tlb)
 180{
 181        struct mmu_gather_batch *batch;
 182
 183        batch = tlb->active;
 184        if (batch->next) {
 185                tlb->active = batch->next;
 186                return 1;
 187        }
 188
 189        if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
 190                return 0;
 191
 192        batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 193        if (!batch)
 194                return 0;
 195
 196        tlb->batch_count++;
 197        batch->next = NULL;
 198        batch->nr   = 0;
 199        batch->max  = MAX_GATHER_BATCH;
 200
 201        tlb->active->next = batch;
 202        tlb->active = batch;
 203
 204        return 1;
 205}
 206
 207/* tlb_gather_mmu
 208 *      Called to initialize an (on-stack) mmu_gather structure for page-table
 209 *      tear-down from @mm. The @fullmm argument is used when @mm is without
 210 *      users and we're going to destroy the full address space (exit/execve).
 211 */
 212void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
 213{
 214        tlb->mm = mm;
 215
 216        /* Is it from 0 to ~0? */
 217        tlb->fullmm     = !(start | (end+1));
 218        tlb->need_flush_all = 0;
 219        tlb->start      = start;
 220        tlb->end        = end;
 221        tlb->need_flush = 0;
 222        tlb->local.next = NULL;
 223        tlb->local.nr   = 0;
 224        tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 225        tlb->active     = &tlb->local;
 226        tlb->batch_count = 0;
 227
 228#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 229        tlb->batch = NULL;
 230#endif
 231}
 232
 233void tlb_flush_mmu(struct mmu_gather *tlb)
 234{
 235        struct mmu_gather_batch *batch;
 236
 237        if (!tlb->need_flush)
 238                return;
 239        tlb->need_flush = 0;
 240        tlb_flush(tlb);
 241#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 242        tlb_table_flush(tlb);
 243#endif
 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_flush_mmu(tlb);
 261
 262        /* keep the page table cache within bounds */
 263        check_pgt_cache();
 264
 265        for (batch = tlb->local.next; batch; batch = next) {
 266                next = batch->next;
 267                free_pages((unsigned long)batch, 0);
 268        }
 269        tlb->local.next = NULL;
 270}
 271
 272/* __tlb_remove_page
 273 *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 274 *      handling the additional races in SMP caused by other CPUs caching valid
 275 *      mappings in their TLBs. Returns the number of free page slots left.
 276 *      When out of page slots we must call tlb_flush_mmu().
 277 */
 278int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 279{
 280        struct mmu_gather_batch *batch;
 281
 282        VM_BUG_ON(!tlb->need_flush);
 283
 284        batch = tlb->active;
 285        batch->pages[batch->nr++] = page;
 286        if (batch->nr == batch->max) {
 287                if (!tlb_next_batch(tlb))
 288                        return 0;
 289                batch = tlb->active;
 290        }
 291        VM_BUG_ON(batch->nr > batch->max);
 292
 293        return batch->max - batch->nr;
 294}
 295
 296#endif /* HAVE_GENERIC_MMU_GATHER */
 297
 298#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 299
 300/*
 301 * See the comment near struct mmu_table_batch.
 302 */
 303
 304static void tlb_remove_table_smp_sync(void *arg)
 305{
 306        /* Simply deliver the interrupt */
 307}
 308
 309static void tlb_remove_table_one(void *table)
 310{
 311        /*
 312         * This isn't an RCU grace period and hence the page-tables cannot be
 313         * assumed to be actually RCU-freed.
 314         *
 315         * It is however sufficient for software page-table walkers that rely on
 316         * IRQ disabling. See the comment near struct mmu_table_batch.
 317         */
 318        smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 319        __tlb_remove_table(table);
 320}
 321
 322static void tlb_remove_table_rcu(struct rcu_head *head)
 323{
 324        struct mmu_table_batch *batch;
 325        int i;
 326
 327        batch = container_of(head, struct mmu_table_batch, rcu);
 328
 329        for (i = 0; i < batch->nr; i++)
 330                __tlb_remove_table(batch->tables[i]);
 331
 332        free_page((unsigned long)batch);
 333}
 334
 335void tlb_table_flush(struct mmu_gather *tlb)
 336{
 337        struct mmu_table_batch **batch = &tlb->batch;
 338
 339        if (*batch) {
 340                call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 341                *batch = NULL;
 342        }
 343}
 344
 345void tlb_remove_table(struct mmu_gather *tlb, void *table)
 346{
 347        struct mmu_table_batch **batch = &tlb->batch;
 348
 349        tlb->need_flush = 1;
 350
 351        /*
 352         * When there's less then two users of this mm there cannot be a
 353         * concurrent page-table walk.
 354         */
 355        if (atomic_read(&tlb->mm->mm_users) < 2) {
 356                __tlb_remove_table(table);
 357                return;
 358        }
 359
 360        if (*batch == NULL) {
 361                *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 362                if (*batch == NULL) {
 363                        tlb_remove_table_one(table);
 364                        return;
 365                }
 366                (*batch)->nr = 0;
 367        }
 368        (*batch)->tables[(*batch)->nr++] = table;
 369        if ((*batch)->nr == MAX_TABLE_BATCH)
 370                tlb_table_flush(tlb);
 371}
 372
 373#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 374
 375/*
 376 * If a p?d_bad entry is found while walking page tables, report
 377 * the error, before resetting entry to p?d_none.  Usually (but
 378 * very seldom) called out from the p?d_none_or_clear_bad macros.
 379 */
 380
 381void pgd_clear_bad(pgd_t *pgd)
 382{
 383        pgd_ERROR(*pgd);
 384        pgd_clear(pgd);
 385}
 386
 387void pud_clear_bad(pud_t *pud)
 388{
 389        pud_ERROR(*pud);
 390        pud_clear(pud);
 391}
 392
 393void pmd_clear_bad(pmd_t *pmd)
 394{
 395        pmd_ERROR(*pmd);
 396        pmd_clear(pmd);
 397}
 398
 399/*
 400 * Note: this doesn't free the actual pages themselves. That
 401 * has been handled earlier when unmapping all the memory regions.
 402 */
 403static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 404                           unsigned long addr)
 405{
 406        pgtable_t token = pmd_pgtable(*pmd);
 407        pmd_clear(pmd);
 408        pte_free_tlb(tlb, token, addr);
 409        tlb->mm->nr_ptes--;
 410}
 411
 412static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 413                                unsigned long addr, unsigned long end,
 414                                unsigned long floor, unsigned long ceiling)
 415{
 416        pmd_t *pmd;
 417        unsigned long next;
 418        unsigned long start;
 419
 420        start = addr;
 421        pmd = pmd_offset(pud, addr);
 422        do {
 423                next = pmd_addr_end(addr, end);
 424                if (pmd_none_or_clear_bad(pmd))
 425                        continue;
 426                free_pte_range(tlb, pmd, addr);
 427        } while (pmd++, addr = next, addr != end);
 428
 429        start &= PUD_MASK;
 430        if (start < floor)
 431                return;
 432        if (ceiling) {
 433                ceiling &= PUD_MASK;
 434                if (!ceiling)
 435                        return;
 436        }
 437        if (end - 1 > ceiling - 1)
 438                return;
 439
 440        pmd = pmd_offset(pud, start);
 441        pud_clear(pud);
 442        pmd_free_tlb(tlb, pmd, start);
 443}
 444
 445static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 446                                unsigned long addr, unsigned long end,
 447                                unsigned long floor, unsigned long ceiling)
 448{
 449        pud_t *pud;
 450        unsigned long next;
 451        unsigned long start;
 452
 453        start = addr;
 454        pud = pud_offset(pgd, addr);
 455        do {
 456                next = pud_addr_end(addr, end);
 457                if (pud_none_or_clear_bad(pud))
 458                        continue;
 459                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 460        } while (pud++, addr = next, addr != end);
 461
 462        start &= PGDIR_MASK;
 463        if (start < floor)
 464                return;
 465        if (ceiling) {
 466                ceiling &= PGDIR_MASK;
 467                if (!ceiling)
 468                        return;
 469        }
 470        if (end - 1 > ceiling - 1)
 471                return;
 472
 473        pud = pud_offset(pgd, start);
 474        pgd_clear(pgd);
 475        pud_free_tlb(tlb, pud, start);
 476}
 477
 478/*
 479 * This function frees user-level page tables of a process.
 480 *
 481 * Must be called with pagetable lock held.
 482 */
 483void free_pgd_range(struct mmu_gather *tlb,
 484                        unsigned long addr, unsigned long end,
 485                        unsigned long floor, unsigned long ceiling)
 486{
 487        pgd_t *pgd;
 488        unsigned long next;
 489
 490        /*
 491         * The next few lines have given us lots of grief...
 492         *
 493         * Why are we testing PMD* at this top level?  Because often
 494         * there will be no work to do at all, and we'd prefer not to
 495         * go all the way down to the bottom just to discover that.
 496         *
 497         * Why all these "- 1"s?  Because 0 represents both the bottom
 498         * of the address space and the top of it (using -1 for the
 499         * top wouldn't help much: the masks would do the wrong thing).
 500         * The rule is that addr 0 and floor 0 refer to the bottom of
 501         * the address space, but end 0 and ceiling 0 refer to the top
 502         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 503         * that end 0 case should be mythical).
 504         *
 505         * Wherever addr is brought up or ceiling brought down, we must
 506         * be careful to reject "the opposite 0" before it confuses the
 507         * subsequent tests.  But what about where end is brought down
 508         * by PMD_SIZE below? no, end can't go down to 0 there.
 509         *
 510         * Whereas we round start (addr) and ceiling down, by different
 511         * masks at different levels, in order to test whether a table
 512         * now has no other vmas using it, so can be freed, we don't
 513         * bother to round floor or end up - the tests don't need that.
 514         */
 515
 516        addr &= PMD_MASK;
 517        if (addr < floor) {
 518                addr += PMD_SIZE;
 519                if (!addr)
 520                        return;
 521        }
 522        if (ceiling) {
 523                ceiling &= PMD_MASK;
 524                if (!ceiling)
 525                        return;
 526        }
 527        if (end - 1 > ceiling - 1)
 528                end -= PMD_SIZE;
 529        if (addr > end - 1)
 530                return;
 531
 532        pgd = pgd_offset(tlb->mm, addr);
 533        do {
 534                next = pgd_addr_end(addr, end);
 535                if (pgd_none_or_clear_bad(pgd))
 536                        continue;
 537                free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 538        } while (pgd++, addr = next, addr != end);
 539}
 540
 541void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 542                unsigned long floor, unsigned long ceiling)
 543{
 544        while (vma) {
 545                struct vm_area_struct *next = vma->vm_next;
 546                unsigned long addr = vma->vm_start;
 547
 548                /*
 549                 * Hide vma from rmap and truncate_pagecache before freeing
 550                 * pgtables
 551                 */
 552                unlink_anon_vmas(vma);
 553                unlink_file_vma(vma);
 554
 555                if (is_vm_hugetlb_page(vma)) {
 556                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 557                                floor, next? next->vm_start: ceiling);
 558                } else {
 559                        /*
 560                         * Optimization: gather nearby vmas into one call down
 561                         */
 562                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 563                               && !is_vm_hugetlb_page(next)) {
 564                                vma = next;
 565                                next = vma->vm_next;
 566                                unlink_anon_vmas(vma);
 567                                unlink_file_vma(vma);
 568                        }
 569                        free_pgd_range(tlb, addr, vma->vm_end,
 570                                floor, next? next->vm_start: ceiling);
 571                }
 572                vma = next;
 573        }
 574}
 575
 576int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 577                pmd_t *pmd, unsigned long address)
 578{
 579        pgtable_t new = pte_alloc_one(mm, address);
 580        int wait_split_huge_page;
 581        if (!new)
 582                return -ENOMEM;
 583
 584        /*
 585         * Ensure all pte setup (eg. pte page lock and page clearing) are
 586         * visible before the pte is made visible to other CPUs by being
 587         * put into page tables.
 588         *
 589         * The other side of the story is the pointer chasing in the page
 590         * table walking code (when walking the page table without locking;
 591         * ie. most of the time). Fortunately, these data accesses consist
 592         * of a chain of data-dependent loads, meaning most CPUs (alpha
 593         * being the notable exception) will already guarantee loads are
 594         * seen in-order. See the alpha page table accessors for the
 595         * smp_read_barrier_depends() barriers in page table walking code.
 596         */
 597        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 598
 599        spin_lock(&mm->page_table_lock);
 600        wait_split_huge_page = 0;
 601        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 602                mm->nr_ptes++;
 603                pmd_populate(mm, pmd, new);
 604                new = NULL;
 605        } else if (unlikely(pmd_trans_splitting(*pmd)))
 606                wait_split_huge_page = 1;
 607        spin_unlock(&mm->page_table_lock);
 608        if (new)
 609                pte_free(mm, new);
 610        if (wait_split_huge_page)
 611                wait_split_huge_page(vma->anon_vma, pmd);
 612        return 0;
 613}
 614
 615int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 616{
 617        pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 618        if (!new)
 619                return -ENOMEM;
 620
 621        smp_wmb(); /* See comment in __pte_alloc */
 622
 623        spin_lock(&init_mm.page_table_lock);
 624        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 625                pmd_populate_kernel(&init_mm, pmd, new);
 626                new = NULL;
 627        } else
 628                VM_BUG_ON(pmd_trans_splitting(*pmd));
 629        spin_unlock(&init_mm.page_table_lock);
 630        if (new)
 631                pte_free_kernel(&init_mm, new);
 632        return 0;
 633}
 634
 635static inline void init_rss_vec(int *rss)
 636{
 637        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 638}
 639
 640static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 641{
 642        int i;
 643
 644        if (current->mm == mm)
 645                sync_mm_rss(mm);
 646        for (i = 0; i < NR_MM_COUNTERS; i++)
 647                if (rss[i])
 648                        add_mm_counter(mm, i, rss[i]);
 649}
 650
 651/*
 652 * This function is called to print an error when a bad pte
 653 * is found. For example, we might have a PFN-mapped pte in
 654 * a region that doesn't allow it.
 655 *
 656 * The calling function must still handle the error.
 657 */
 658static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 659                          pte_t pte, struct page *page)
 660{
 661        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 662        pud_t *pud = pud_offset(pgd, addr);
 663        pmd_t *pmd = pmd_offset(pud, addr);
 664        struct address_space *mapping;
 665        pgoff_t index;
 666        static unsigned long resume;
 667        static unsigned long nr_shown;
 668        static unsigned long nr_unshown;
 669
 670        /*
 671         * Allow a burst of 60 reports, then keep quiet for that minute;
 672         * or allow a steady drip of one report per second.
 673         */
 674        if (nr_shown == 60) {
 675                if (time_before(jiffies, resume)) {
 676                        nr_unshown++;
 677                        return;
 678                }
 679                if (nr_unshown) {
 680                        printk(KERN_ALERT
 681                                "BUG: Bad page map: %lu messages suppressed\n",
 682                                nr_unshown);
 683                        nr_unshown = 0;
 684                }
 685                nr_shown = 0;
 686        }
 687        if (nr_shown++ == 0)
 688                resume = jiffies + 60 * HZ;
 689
 690        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 691        index = linear_page_index(vma, addr);
 692
 693        printk(KERN_ALERT
 694                "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 695                current->comm,
 696                (long long)pte_val(pte), (long long)pmd_val(*pmd));
 697        if (page)
 698                dump_page(page);
 699        printk(KERN_ALERT
 700                "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 701                (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 702        /*
 703         * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 704         */
 705        if (vma->vm_ops)
 706                printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
 707                       vma->vm_ops->fault);
 708        if (vma->vm_file && vma->vm_file->f_op)
 709                printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
 710                       vma->vm_file->f_op->mmap);
 711        dump_stack();
 712        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 713}
 714
 715static inline bool is_cow_mapping(vm_flags_t flags)
 716{
 717        return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 718}
 719
 720/*
 721 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 722 *
 723 * "Special" mappings do not wish to be associated with a "struct page" (either
 724 * it doesn't exist, or it exists but they don't want to touch it). In this
 725 * case, NULL is returned here. "Normal" mappings do have a struct page.
 726 *
 727 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 728 * pte bit, in which case this function is trivial. Secondly, an architecture
 729 * may not have a spare pte bit, which requires a more complicated scheme,
 730 * described below.
 731 *
 732 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 733 * special mapping (even if there are underlying and valid "struct pages").
 734 * COWed pages of a VM_PFNMAP are always normal.
 735 *
 736 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 737 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 738 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 739 * mapping will always honor the rule
 740 *
 741 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 742 *
 743 * And for normal mappings this is false.
 744 *
 745 * This restricts such mappings to be a linear translation from virtual address
 746 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 747 * as the vma is not a COW mapping; in that case, we know that all ptes are
 748 * special (because none can have been COWed).
 749 *
 750 *
 751 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 752 *
 753 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 754 * page" backing, however the difference is that _all_ pages with a struct
 755 * page (that is, those where pfn_valid is true) are refcounted and considered
 756 * normal pages by the VM. The disadvantage is that pages are refcounted
 757 * (which can be slower and simply not an option for some PFNMAP users). The
 758 * advantage is that we don't have to follow the strict linearity rule of
 759 * PFNMAP mappings in order to support COWable mappings.
 760 *
 761 */
 762#ifdef __HAVE_ARCH_PTE_SPECIAL
 763# define HAVE_PTE_SPECIAL 1
 764#else
 765# define HAVE_PTE_SPECIAL 0
 766#endif
 767struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 768                                pte_t pte)
 769{
 770        unsigned long pfn = pte_pfn(pte);
 771
 772        if (HAVE_PTE_SPECIAL) {
 773                if (likely(!pte_special(pte)))
 774                        goto check_pfn;
 775                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 776                        return NULL;
 777                if (!is_zero_pfn(pfn))
 778                        print_bad_pte(vma, addr, pte, NULL);
 779                return NULL;
 780        }
 781
 782        /* !HAVE_PTE_SPECIAL case follows: */
 783
 784        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 785                if (vma->vm_flags & VM_MIXEDMAP) {
 786                        if (!pfn_valid(pfn))
 787                                return NULL;
 788                        goto out;
 789                } else {
 790                        unsigned long off;
 791                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 792                        if (pfn == vma->vm_pgoff + off)
 793                                return NULL;
 794                        if (!is_cow_mapping(vma->vm_flags))
 795                                return NULL;
 796                }
 797        }
 798
 799        if (is_zero_pfn(pfn))
 800                return NULL;
 801check_pfn:
 802        if (unlikely(pfn > highest_memmap_pfn)) {
 803                print_bad_pte(vma, addr, pte, NULL);
 804                return NULL;
 805        }
 806
 807        /*
 808         * NOTE! We still have PageReserved() pages in the page tables.
 809         * eg. VDSO mappings can cause them to exist.
 810         */
 811out:
 812        return pfn_to_page(pfn);
 813}
 814
 815/*
 816 * copy one vm_area from one task to the other. Assumes the page tables
 817 * already present in the new task to be cleared in the whole range
 818 * covered by this vma.
 819 */
 820
 821static inline unsigned long
 822copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 823                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 824                unsigned long addr, int *rss)
 825{
 826        unsigned long vm_flags = vma->vm_flags;
 827        pte_t pte = *src_pte;
 828        struct page *page;
 829
 830        /* pte contains position in swap or file, so copy. */
 831        if (unlikely(!pte_present(pte))) {
 832                if (!pte_file(pte)) {
 833                        swp_entry_t entry = pte_to_swp_entry(pte);
 834
 835                        if (swap_duplicate(entry) < 0)
 836                                return entry.val;
 837
 838                        /* make sure dst_mm is on swapoff's mmlist. */
 839                        if (unlikely(list_empty(&dst_mm->mmlist))) {
 840                                spin_lock(&mmlist_lock);
 841                                if (list_empty(&dst_mm->mmlist))
 842                                        list_add(&dst_mm->mmlist,
 843                                                 &src_mm->mmlist);
 844                                spin_unlock(&mmlist_lock);
 845                        }
 846                        if (likely(!non_swap_entry(entry)))
 847                                rss[MM_SWAPENTS]++;
 848                        else if (is_migration_entry(entry)) {
 849                                page = migration_entry_to_page(entry);
 850
 851                                if (PageAnon(page))
 852                                        rss[MM_ANONPAGES]++;
 853                                else
 854                                        rss[MM_FILEPAGES]++;
 855
 856                                if (is_write_migration_entry(entry) &&
 857                                    is_cow_mapping(vm_flags)) {
 858                                        /*
 859                                         * COW mappings require pages in both
 860                                         * parent and child to be set to read.
 861                                         */
 862                                        make_migration_entry_read(&entry);
 863                                        pte = swp_entry_to_pte(entry);
 864                                        set_pte_at(src_mm, addr, src_pte, pte);
 865                                }
 866                        }
 867                }
 868                goto out_set_pte;
 869        }
 870
 871        /*
 872         * If it's a COW mapping, write protect it both
 873         * in the parent and the child
 874         */
 875        if (is_cow_mapping(vm_flags)) {
 876                ptep_set_wrprotect(src_mm, addr, src_pte);
 877                pte = pte_wrprotect(pte);
 878        }
 879
 880        /*
 881         * If it's a shared mapping, mark it clean in
 882         * the child
 883         */
 884        if (vm_flags & VM_SHARED)
 885                pte = pte_mkclean(pte);
 886        pte = pte_mkold(pte);
 887
 888        page = vm_normal_page(vma, addr, pte);
 889        if (page) {
 890                get_page(page);
 891                page_dup_rmap(page);
 892                if (PageAnon(page))
 893                        rss[MM_ANONPAGES]++;
 894                else
 895                        rss[MM_FILEPAGES]++;
 896        }
 897
 898out_set_pte:
 899        set_pte_at(dst_mm, addr, dst_pte, pte);
 900        return 0;
 901}
 902
 903int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 904                   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 905                   unsigned long addr, unsigned long end)
 906{
 907        pte_t *orig_src_pte, *orig_dst_pte;
 908        pte_t *src_pte, *dst_pte;
 909        spinlock_t *src_ptl, *dst_ptl;
 910        int progress = 0;
 911        int rss[NR_MM_COUNTERS];
 912        swp_entry_t entry = (swp_entry_t){0};
 913
 914again:
 915        init_rss_vec(rss);
 916
 917        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 918        if (!dst_pte)
 919                return -ENOMEM;
 920        src_pte = pte_offset_map(src_pmd, addr);
 921        src_ptl = pte_lockptr(src_mm, src_pmd);
 922        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 923        orig_src_pte = src_pte;
 924        orig_dst_pte = dst_pte;
 925        arch_enter_lazy_mmu_mode();
 926
 927        do {
 928                /*
 929                 * We are holding two locks at this point - either of them
 930                 * could generate latencies in another task on another CPU.
 931                 */
 932                if (progress >= 32) {
 933                        progress = 0;
 934                        if (need_resched() ||
 935                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 936                                break;
 937                }
 938                if (pte_none(*src_pte)) {
 939                        progress++;
 940                        continue;
 941                }
 942                entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 943                                                        vma, addr, rss);
 944                if (entry.val)
 945                        break;
 946                progress += 8;
 947        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 948
 949        arch_leave_lazy_mmu_mode();
 950        spin_unlock(src_ptl);
 951        pte_unmap(orig_src_pte);
 952        add_mm_rss_vec(dst_mm, rss);
 953        pte_unmap_unlock(orig_dst_pte, dst_ptl);
 954        cond_resched();
 955
 956        if (entry.val) {
 957                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 958                        return -ENOMEM;
 959                progress = 0;
 960        }
 961        if (addr != end)
 962                goto again;
 963        return 0;
 964}
 965
 966static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 967                pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 968                unsigned long addr, unsigned long end)
 969{
 970        pmd_t *src_pmd, *dst_pmd;
 971        unsigned long next;
 972
 973        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 974        if (!dst_pmd)
 975                return -ENOMEM;
 976        src_pmd = pmd_offset(src_pud, addr);
 977        do {
 978                next = pmd_addr_end(addr, end);
 979                if (pmd_trans_huge(*src_pmd)) {
 980                        int err;
 981                        VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
 982                        err = copy_huge_pmd(dst_mm, src_mm,
 983                                            dst_pmd, src_pmd, addr, vma);
 984                        if (err == -ENOMEM)
 985                                return -ENOMEM;
 986                        if (!err)
 987                                continue;
 988                        /* fall through */
 989                }
 990                if (pmd_none_or_clear_bad(src_pmd))
 991                        continue;
 992                if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
 993                                                vma, addr, next))
 994                        return -ENOMEM;
 995        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
 996        return 0;
 997}
 998
 999static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1000                pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1001                unsigned long addr, unsigned long end)
1002{
1003        pud_t *src_pud, *dst_pud;
1004        unsigned long next;
1005
1006        dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1007        if (!dst_pud)
1008                return -ENOMEM;
1009        src_pud = pud_offset(src_pgd, addr);
1010        do {
1011                next = pud_addr_end(addr, end);
1012                if (pud_none_or_clear_bad(src_pud))
1013                        continue;
1014                if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1015                                                vma, addr, next))
1016                        return -ENOMEM;
1017        } while (dst_pud++, src_pud++, addr = next, addr != end);
1018        return 0;
1019}
1020
1021int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1022                struct vm_area_struct *vma)
1023{
1024        pgd_t *src_pgd, *dst_pgd;
1025        unsigned long next;
1026        unsigned long addr = vma->vm_start;
1027        unsigned long end = vma->vm_end;
1028        unsigned long mmun_start;       /* For mmu_notifiers */
1029        unsigned long mmun_end;         /* For mmu_notifiers */
1030        bool is_cow;
1031        int ret;
1032
1033        /*
1034         * Don't copy ptes where a page fault will fill them correctly.
1035         * Fork becomes much lighter when there are big shared or private
1036         * readonly mappings. The tradeoff is that copy_page_range is more
1037         * efficient than faulting.
1038         */
1039        if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1040                               VM_PFNMAP | VM_MIXEDMAP))) {
1041                if (!vma->anon_vma)
1042                        return 0;
1043        }
1044
1045        if (is_vm_hugetlb_page(vma))
1046                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1047
1048        if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1049                /*
1050                 * We do not free on error cases below as remove_vma
1051                 * gets called on error from higher level routine
1052                 */
1053                ret = track_pfn_copy(vma);
1054                if (ret)
1055                        return ret;
1056        }
1057
1058        /*
1059         * We need to invalidate the secondary MMU mappings only when
1060         * there could be a permission downgrade on the ptes of the
1061         * parent mm. And a permission downgrade will only happen if
1062         * is_cow_mapping() returns true.
1063         */
1064        is_cow = is_cow_mapping(vma->vm_flags);
1065        mmun_start = addr;
1066        mmun_end   = end;
1067        if (is_cow)
1068                mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1069                                                    mmun_end);
1070
1071        ret = 0;
1072        dst_pgd = pgd_offset(dst_mm, addr);
1073        src_pgd = pgd_offset(src_mm, addr);
1074        do {
1075                next = pgd_addr_end(addr, end);
1076                if (pgd_none_or_clear_bad(src_pgd))
1077                        continue;
1078                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1079                                            vma, addr, next))) {
1080                        ret = -ENOMEM;
1081                        break;
1082                }
1083        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1084
1085        if (is_cow)
1086                mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1087        return ret;
1088}
1089
1090static unsigned long zap_pte_range(struct mmu_gather *tlb,
1091                                struct vm_area_struct *vma, pmd_t *pmd,
1092                                unsigned long addr, unsigned long end,
1093                                struct zap_details *details)
1094{
1095        struct mm_struct *mm = tlb->mm;
1096        int force_flush = 0;
1097        int rss[NR_MM_COUNTERS];
1098        spinlock_t *ptl;
1099        pte_t *start_pte;
1100        pte_t *pte;
1101
1102again:
1103        init_rss_vec(rss);
1104        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1105        pte = start_pte;
1106        arch_enter_lazy_mmu_mode();
1107        do {
1108                pte_t ptent = *pte;
1109                if (pte_none(ptent)) {
1110                        continue;
1111                }
1112
1113                if (pte_present(ptent)) {
1114                        struct page *page;
1115
1116                        page = vm_normal_page(vma, addr, ptent);
1117                        if (unlikely(details) && page) {
1118                                /*
1119                                 * unmap_shared_mapping_pages() wants to
1120                                 * invalidate cache without truncating:
1121                                 * unmap shared but keep private pages.
1122                                 */
1123                                if (details->check_mapping &&
1124                                    details->check_mapping != page->mapping)
1125                                        continue;
1126                                /*
1127                                 * Each page->index must be checked when
1128                                 * invalidating or truncating nonlinear.
1129                                 */
1130                                if (details->nonlinear_vma &&
1131                                    (page->index < details->first_index ||
1132                                     page->index > details->last_index))
1133                                        continue;
1134                        }
1135                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1136                                                        tlb->fullmm);
1137                        tlb_remove_tlb_entry(tlb, pte, addr);
1138                        if (unlikely(!page))
1139                                continue;
1140                        if (unlikely(details) && details->nonlinear_vma
1141                            && linear_page_index(details->nonlinear_vma,
1142                                                addr) != page->index) {
1143                                pte_t ptfile = pgoff_to_pte(page->index);
1144                                if (pte_soft_dirty(ptent))
1145                                        pte_file_mksoft_dirty(ptfile);
1146                                set_pte_at(mm, addr, pte, ptfile);
1147                        }
1148                        if (PageAnon(page))
1149                                rss[MM_ANONPAGES]--;
1150                        else {
1151                                if (pte_dirty(ptent))
1152                                        set_page_dirty(page);
1153                                if (pte_young(ptent) &&
1154                                    likely(!(vma->vm_flags & VM_SEQ_READ)))
1155                                        mark_page_accessed(page);
1156                                rss[MM_FILEPAGES]--;
1157                        }
1158                        page_remove_rmap(page);
1159                        if (unlikely(page_mapcount(page) < 0))
1160                                print_bad_pte(vma, addr, ptent, page);
1161                        force_flush = !__tlb_remove_page(tlb, page);
1162                        if (force_flush)
1163                                break;
1164                        continue;
1165                }
1166                /*
1167                 * If details->check_mapping, we leave swap entries;
1168                 * if details->nonlinear_vma, we leave file entries.
1169                 */
1170                if (unlikely(details))
1171                        continue;
1172                if (pte_file(ptent)) {
1173                        if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1174                                print_bad_pte(vma, addr, ptent, NULL);
1175                } else {
1176                        swp_entry_t entry = pte_to_swp_entry(ptent);
1177
1178                        if (!non_swap_entry(entry))
1179                                rss[MM_SWAPENTS]--;
1180                        else if (is_migration_entry(entry)) {
1181                                struct page *page;
1182
1183                                page = migration_entry_to_page(entry);
1184
1185                                if (PageAnon(page))
1186                                        rss[MM_ANONPAGES]--;
1187                                else
1188                                        rss[MM_FILEPAGES]--;
1189                        }
1190                        if (unlikely(!free_swap_and_cache(entry)))
1191                                print_bad_pte(vma, addr, ptent, NULL);
1192                }
1193                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1194        } while (pte++, addr += PAGE_SIZE, addr != end);
1195
1196        add_mm_rss_vec(mm, rss);
1197        arch_leave_lazy_mmu_mode();
1198        pte_unmap_unlock(start_pte, ptl);
1199
1200        /*
1201         * mmu_gather ran out of room to batch pages, we break out of
1202         * the PTE lock to avoid doing the potential expensive TLB invalidate
1203         * and page-free while holding it.
1204         */
1205        if (force_flush) {
1206                unsigned long old_end;
1207
1208                force_flush = 0;
1209
1210                /*
1211                 * Flush the TLB just for the previous segment,
1212                 * then update the range to be the remaining
1213                 * TLB range.
1214                 */
1215                old_end = tlb->end;
1216                tlb->end = addr;
1217
1218                tlb_flush_mmu(tlb);
1219
1220                tlb->start = addr;
1221                tlb->end = old_end;
1222
1223                if (addr != end)
1224                        goto again;
1225        }
1226
1227        return addr;
1228}
1229
1230static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1231                                struct vm_area_struct *vma, pud_t *pud,
1232                                unsigned long addr, unsigned long end,
1233                                struct zap_details *details)
1234{
1235        pmd_t *pmd;
1236        unsigned long next;
1237
1238        pmd = pmd_offset(pud, addr);
1239        do {
1240                next = pmd_addr_end(addr, end);
1241                if (pmd_trans_huge(*pmd)) {
1242                        if (next - addr != HPAGE_PMD_SIZE) {
1243#ifdef CONFIG_DEBUG_VM
1244                                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1245                                        pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1246                                                __func__, addr, end,
1247                                                vma->vm_start,
1248                                                vma->vm_end);
1249                                        BUG();
1250                                }
1251#endif
1252                                split_huge_page_pmd(vma, addr, pmd);
1253                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1254                                goto next;
1255                        /* fall through */
1256                }
1257                /*
1258                 * Here there can be other concurrent MADV_DONTNEED or
1259                 * trans huge page faults running, and if the pmd is
1260                 * none or trans huge it can change under us. This is
1261                 * because MADV_DONTNEED holds the mmap_sem in read
1262                 * mode.
1263                 */
1264                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1265                        goto next;
1266                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1267next:
1268                cond_resched();
1269        } while (pmd++, addr = next, addr != end);
1270
1271        return addr;
1272}
1273
1274static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1275                                struct vm_area_struct *vma, pgd_t *pgd,
1276                                unsigned long addr, unsigned long end,
1277                                struct zap_details *details)
1278{
1279        pud_t *pud;
1280        unsigned long next;
1281
1282        pud = pud_offset(pgd, addr);
1283        do {
1284                next = pud_addr_end(addr, end);
1285                if (pud_none_or_clear_bad(pud))
1286                        continue;
1287                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1288        } while (pud++, addr = next, addr != end);
1289
1290        return addr;
1291}
1292
1293static void unmap_page_range(struct mmu_gather *tlb,
1294                             struct vm_area_struct *vma,
1295                             unsigned long addr, unsigned long end,
1296                             struct zap_details *details)
1297{
1298        pgd_t *pgd;
1299        unsigned long next;
1300
1301        if (details && !details->check_mapping && !details->nonlinear_vma)
1302                details = NULL;
1303
1304        BUG_ON(addr >= end);
1305        mem_cgroup_uncharge_start();
1306        tlb_start_vma(tlb, vma);
1307        pgd = pgd_offset(vma->vm_mm, addr);
1308        do {
1309                next = pgd_addr_end(addr, end);
1310                if (pgd_none_or_clear_bad(pgd))
1311                        continue;
1312                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1313        } while (pgd++, addr = next, addr != end);
1314        tlb_end_vma(tlb, vma);
1315        mem_cgroup_uncharge_end();
1316}
1317
1318
1319static void unmap_single_vma(struct mmu_gather *tlb,
1320                struct vm_area_struct *vma, unsigned long start_addr,
1321                unsigned long end_addr,
1322                struct zap_details *details)
1323{
1324        unsigned long start = max(vma->vm_start, start_addr);
1325        unsigned long end;
1326
1327        if (start >= vma->vm_end)
1328                return;
1329        end = min(vma->vm_end, end_addr);
1330        if (end <= vma->vm_start)
1331                return;
1332
1333        if (vma->vm_file)
1334                uprobe_munmap(vma, start, end);
1335
1336        if (unlikely(vma->vm_flags & VM_PFNMAP))
1337                untrack_pfn(vma, 0, 0);
1338
1339        if (start != end) {
1340                if (unlikely(is_vm_hugetlb_page(vma))) {
1341                        /*
1342                         * It is undesirable to test vma->vm_file as it
1343                         * should be non-null for valid hugetlb area.
1344                         * However, vm_file will be NULL in the error
1345                         * cleanup path of do_mmap_pgoff. When
1346                         * hugetlbfs ->mmap method fails,
1347                         * do_mmap_pgoff() nullifies vma->vm_file
1348                         * before calling this function to clean up.
1349                         * Since no pte has actually been setup, it is
1350                         * safe to do nothing in this case.
1351                         */
1352                        if (vma->vm_file) {
1353                                mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1354                                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1355                                mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1356                        }
1357                } else
1358                        unmap_page_range(tlb, vma, start, end, details);
1359        }
1360}
1361
1362/**
1363 * unmap_vmas - unmap a range of memory covered by a list of vma's
1364 * @tlb: address of the caller's struct mmu_gather
1365 * @vma: the starting vma
1366 * @start_addr: virtual address at which to start unmapping
1367 * @end_addr: virtual address at which to end unmapping
1368 *
1369 * Unmap all pages in the vma list.
1370 *
1371 * Only addresses between `start' and `end' will be unmapped.
1372 *
1373 * The VMA list must be sorted in ascending virtual address order.
1374 *
1375 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1376 * range after unmap_vmas() returns.  So the only responsibility here is to
1377 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1378 * drops the lock and schedules.
1379 */
1380void unmap_vmas(struct mmu_gather *tlb,
1381                struct vm_area_struct *vma, unsigned long start_addr,
1382                unsigned long end_addr)
1383{
1384        struct mm_struct *mm = vma->vm_mm;
1385
1386        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1387        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1388                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1389        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1390}
1391
1392/**
1393 * zap_page_range - remove user pages in a given range
1394 * @vma: vm_area_struct holding the applicable pages
1395 * @start: starting address of pages to zap
1396 * @size: number of bytes to zap
1397 * @details: details of nonlinear truncation or shared cache invalidation
1398 *
1399 * Caller must protect the VMA list
1400 */
1401void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1402                unsigned long size, struct zap_details *details)
1403{
1404        struct mm_struct *mm = vma->vm_mm;
1405        struct mmu_gather tlb;
1406        unsigned long end = start + size;
1407
1408        lru_add_drain();
1409        tlb_gather_mmu(&tlb, mm, start, end);
1410        update_hiwater_rss(mm);
1411        mmu_notifier_invalidate_range_start(mm, start, end);
1412        for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1413                unmap_single_vma(&tlb, vma, start, end, details);
1414        mmu_notifier_invalidate_range_end(mm, start, end);
1415        tlb_finish_mmu(&tlb, start, end);
1416}
1417
1418/**
1419 * zap_page_range_single - remove user pages in a given range
1420 * @vma: vm_area_struct holding the applicable pages
1421 * @address: starting address of pages to zap
1422 * @size: number of bytes to zap
1423 * @details: details of nonlinear truncation or shared cache invalidation
1424 *
1425 * The range must fit into one VMA.
1426 */
1427static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1428                unsigned long size, struct zap_details *details)
1429{
1430        struct mm_struct *mm = vma->vm_mm;
1431        struct mmu_gather tlb;
1432        unsigned long end = address + size;
1433
1434        lru_add_drain();
1435        tlb_gather_mmu(&tlb, mm, address, end);
1436        update_hiwater_rss(mm);
1437        mmu_notifier_invalidate_range_start(mm, address, end);
1438        unmap_single_vma(&tlb, vma, address, end, details);
1439        mmu_notifier_invalidate_range_end(mm, address, end);
1440        tlb_finish_mmu(&tlb, address, end);
1441}
1442
1443/**
1444 * zap_vma_ptes - remove ptes mapping the vma
1445 * @vma: vm_area_struct holding ptes to be zapped
1446 * @address: starting address of pages to zap
1447 * @size: number of bytes to zap
1448 *
1449 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1450 *
1451 * The entire address range must be fully contained within the vma.
1452 *
1453 * Returns 0 if successful.
1454 */
1455int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1456                unsigned long size)
1457{
1458        if (address < vma->vm_start || address + size > vma->vm_end ||
1459                        !(vma->vm_flags & VM_PFNMAP))
1460                return -1;
1461        zap_page_range_single(vma, address, size, NULL);
1462        return 0;
1463}
1464EXPORT_SYMBOL_GPL(zap_vma_ptes);
1465
1466/**
1467 * follow_page_mask - look up a page descriptor from a user-virtual address
1468 * @vma: vm_area_struct mapping @address
1469 * @address: virtual address to look up
1470 * @flags: flags modifying lookup behaviour
1471 * @page_mask: on output, *page_mask is set according to the size of the page
1472 *
1473 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1474 *
1475 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1476 * an error pointer if there is a mapping to something not represented
1477 * by a page descriptor (see also vm_normal_page()).
1478 */
1479struct page *follow_page_mask(struct vm_area_struct *vma,
1480                              unsigned long address, unsigned int flags,
1481                              unsigned int *page_mask)
1482{
1483        pgd_t *pgd;
1484        pud_t *pud;
1485        pmd_t *pmd;
1486        pte_t *ptep, pte;
1487        spinlock_t *ptl;
1488        struct page *page;
1489        struct mm_struct *mm = vma->vm_mm;
1490
1491        *page_mask = 0;
1492
1493        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1494        if (!IS_ERR(page)) {
1495                BUG_ON(flags & FOLL_GET);
1496                goto out;
1497        }
1498
1499        page = NULL;
1500        pgd = pgd_offset(mm, address);
1501        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1502                goto no_page_table;
1503
1504        pud = pud_offset(pgd, address);
1505        if (pud_none(*pud))
1506                goto no_page_table;
1507        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1508                BUG_ON(flags & FOLL_GET);
1509                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1510                goto out;
1511        }
1512        if (unlikely(pud_bad(*pud)))
1513                goto no_page_table;
1514
1515        pmd = pmd_offset(pud, address);
1516        if (pmd_none(*pmd))
1517                goto no_page_table;
1518        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1519                BUG_ON(flags & FOLL_GET);
1520                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1521                goto out;
1522        }
1523        if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1524                goto no_page_table;
1525        if (pmd_trans_huge(*pmd)) {
1526                if (flags & FOLL_SPLIT) {
1527                        split_huge_page_pmd(vma, address, pmd);
1528                        goto split_fallthrough;
1529                }
1530                spin_lock(&mm->page_table_lock);
1531                if (likely(pmd_trans_huge(*pmd))) {
1532                        if (unlikely(pmd_trans_splitting(*pmd))) {
1533                                spin_unlock(&mm->page_table_lock);
1534                                wait_split_huge_page(vma->anon_vma, pmd);
1535                        } else {
1536                                page = follow_trans_huge_pmd(vma, address,
1537                                                             pmd, flags);
1538                                spin_unlock(&mm->page_table_lock);
1539                                *page_mask = HPAGE_PMD_NR - 1;
1540                                goto out;
1541                        }
1542                } else
1543                        spin_unlock(&mm->page_table_lock);
1544                /* fall through */
1545        }
1546split_fallthrough:
1547        if (unlikely(pmd_bad(*pmd)))
1548                goto no_page_table;
1549
1550        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1551
1552        pte = *ptep;
1553        if (!pte_present(pte)) {
1554                swp_entry_t entry;
1555                /*
1556                 * KSM's break_ksm() relies upon recognizing a ksm page
1557                 * even while it is being migrated, so for that case we
1558                 * need migration_entry_wait().
1559                 */
1560                if (likely(!(flags & FOLL_MIGRATION)))
1561                        goto no_page;
1562                if (pte_none(pte) || pte_file(pte))
1563                        goto no_page;
1564                entry = pte_to_swp_entry(pte);
1565                if (!is_migration_entry(entry))
1566                        goto no_page;
1567                pte_unmap_unlock(ptep, ptl);
1568                migration_entry_wait(mm, pmd, address);
1569                goto split_fallthrough;
1570        }
1571        if ((flags & FOLL_NUMA) && pte_numa(pte))
1572                goto no_page;
1573        if ((flags & FOLL_WRITE) && !pte_write(pte))
1574                goto unlock;
1575
1576        page = vm_normal_page(vma, address, pte);
1577        if (unlikely(!page)) {
1578                if ((flags & FOLL_DUMP) ||
1579                    !is_zero_pfn(pte_pfn(pte)))
1580                        goto bad_page;
1581                page = pte_page(pte);
1582        }
1583
1584        if (flags & FOLL_GET)
1585                get_page_foll(page);
1586        if (flags & FOLL_TOUCH) {
1587                if ((flags & FOLL_WRITE) &&
1588                    !pte_dirty(pte) && !PageDirty(page))
1589                        set_page_dirty(page);
1590                /*
1591                 * pte_mkyoung() would be more correct here, but atomic care
1592                 * is needed to avoid losing the dirty bit: it is easier to use
1593                 * mark_page_accessed().
1594                 */
1595                mark_page_accessed(page);
1596        }
1597        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1598                /*
1599                 * The preliminary mapping check is mainly to avoid the
1600                 * pointless overhead of lock_page on the ZERO_PAGE
1601                 * which might bounce very badly if there is contention.
1602                 *
1603                 * If the page is already locked, we don't need to
1604                 * handle it now - vmscan will handle it later if and
1605                 * when it attempts to reclaim the page.
1606                 */
1607                if (page->mapping && trylock_page(page)) {
1608                        lru_add_drain();  /* push cached pages to LRU */
1609                        /*
1610                         * Because we lock page here, and migration is
1611                         * blocked by the pte's page reference, and we
1612                         * know the page is still mapped, we don't even
1613                         * need to check for file-cache page truncation.
1614                         */
1615                        mlock_vma_page(page);
1616                        unlock_page(page);
1617                }
1618        }
1619unlock:
1620        pte_unmap_unlock(ptep, ptl);
1621out:
1622        return page;
1623
1624bad_page:
1625        pte_unmap_unlock(ptep, ptl);
1626        return ERR_PTR(-EFAULT);
1627
1628no_page:
1629        pte_unmap_unlock(ptep, ptl);
1630        if (!pte_none(pte))
1631                return page;
1632
1633no_page_table:
1634        /*
1635         * When core dumping an enormous anonymous area that nobody
1636         * has touched so far, we don't want to allocate unnecessary pages or
1637         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1638         * then get_dump_page() will return NULL to leave a hole in the dump.
1639         * But we can only make this optimization where a hole would surely
1640         * be zero-filled if handle_mm_fault() actually did handle it.
1641         */
1642        if ((flags & FOLL_DUMP) &&
1643            (!vma->vm_ops || !vma->vm_ops->fault))
1644                return ERR_PTR(-EFAULT);
1645        return page;
1646}
1647
1648static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1649{
1650        return stack_guard_page_start(vma, addr) ||
1651               stack_guard_page_end(vma, addr+PAGE_SIZE);
1652}
1653
1654/**
1655 * __get_user_pages() - pin user pages in memory
1656 * @tsk:        task_struct of target task
1657 * @mm:         mm_struct of target mm
1658 * @start:      starting user address
1659 * @nr_pages:   number of pages from start to pin
1660 * @gup_flags:  flags modifying pin behaviour
1661 * @pages:      array that receives pointers to the pages pinned.
1662 *              Should be at least nr_pages long. Or NULL, if caller
1663 *              only intends to ensure the pages are faulted in.
1664 * @vmas:       array of pointers to vmas corresponding to each page.
1665 *              Or NULL if the caller does not require them.
1666 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1667 *
1668 * Returns number of pages pinned. This may be fewer than the number
1669 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1670 * were pinned, returns -errno. Each page returned must be released
1671 * with a put_page() call when it is finished with. vmas will only
1672 * remain valid while mmap_sem is held.
1673 *
1674 * Must be called with mmap_sem held for read or write.
1675 *
1676 * __get_user_pages walks a process's page tables and takes a reference to
1677 * each struct page that each user address corresponds to at a given
1678 * instant. That is, it takes the page that would be accessed if a user
1679 * thread accesses the given user virtual address at that instant.
1680 *
1681 * This does not guarantee that the page exists in the user mappings when
1682 * __get_user_pages returns, and there may even be a completely different
1683 * page there in some cases (eg. if mmapped pagecache has been invalidated
1684 * and subsequently re faulted). However it does guarantee that the page
1685 * won't be freed completely. And mostly callers simply care that the page
1686 * contains data that was valid *at some point in time*. Typically, an IO
1687 * or similar operation cannot guarantee anything stronger anyway because
1688 * locks can't be held over the syscall boundary.
1689 *
1690 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1691 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1692 * appropriate) must be called after the page is finished with, and
1693 * before put_page is called.
1694 *
1695 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1696 * or mmap_sem contention, and if waiting is needed to pin all pages,
1697 * *@nonblocking will be set to 0.
1698 *
1699 * In most cases, get_user_pages or get_user_pages_fast should be used
1700 * instead of __get_user_pages. __get_user_pages should be used only if
1701 * you need some special @gup_flags.
1702 */
1703long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1704                unsigned long start, unsigned long nr_pages,
1705                unsigned int gup_flags, struct page **pages,
1706                struct vm_area_struct **vmas, int *nonblocking)
1707{
1708        long i;
1709        unsigned long vm_flags;
1710        unsigned int page_mask;
1711
1712        if (!nr_pages)
1713                return 0;
1714
1715        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1716
1717        /* 
1718         * Require read or write permissions.
1719         * If FOLL_FORCE is set, we only require the "MAY" flags.
1720         */
1721        vm_flags  = (gup_flags & FOLL_WRITE) ?
1722                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1723        vm_flags &= (gup_flags & FOLL_FORCE) ?
1724                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1725
1726        /*
1727         * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1728         * would be called on PROT_NONE ranges. We must never invoke
1729         * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1730         * page faults would unprotect the PROT_NONE ranges if
1731         * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1732         * bitflag. So to avoid that, don't set FOLL_NUMA if
1733         * FOLL_FORCE is set.
1734         */
1735        if (!(gup_flags & FOLL_FORCE))
1736                gup_flags |= FOLL_NUMA;
1737
1738        i = 0;
1739
1740        do {
1741                struct vm_area_struct *vma;
1742
1743                vma = find_extend_vma(mm, start);
1744                if (!vma && in_gate_area(mm, start)) {
1745                        unsigned long pg = start & PAGE_MASK;
1746                        pgd_t *pgd;
1747                        pud_t *pud;
1748                        pmd_t *pmd;
1749                        pte_t *pte;
1750
1751                        /* user gate pages are read-only */
1752                        if (gup_flags & FOLL_WRITE)
1753                                return i ? : -EFAULT;
1754                        if (pg > TASK_SIZE)
1755                                pgd = pgd_offset_k(pg);
1756                        else
1757                                pgd = pgd_offset_gate(mm, pg);
1758                        BUG_ON(pgd_none(*pgd));
1759                        pud = pud_offset(pgd, pg);
1760                        BUG_ON(pud_none(*pud));
1761                        pmd = pmd_offset(pud, pg);
1762                        if (pmd_none(*pmd))
1763                                return i ? : -EFAULT;
1764                        VM_BUG_ON(pmd_trans_huge(*pmd));
1765                        pte = pte_offset_map(pmd, pg);
1766                        if (pte_none(*pte)) {
1767                                pte_unmap(pte);
1768                                return i ? : -EFAULT;
1769                        }
1770                        vma = get_gate_vma(mm);
1771                        if (pages) {
1772                                struct page *page;
1773
1774                                page = vm_normal_page(vma, start, *pte);
1775                                if (!page) {
1776                                        if (!(gup_flags & FOLL_DUMP) &&
1777                                             is_zero_pfn(pte_pfn(*pte)))
1778                                                page = pte_page(*pte);
1779                                        else {
1780                                                pte_unmap(pte);
1781                                                return i ? : -EFAULT;
1782                                        }
1783                                }
1784                                pages[i] = page;
1785                                get_page(page);
1786                        }
1787                        pte_unmap(pte);
1788                        page_mask = 0;
1789                        goto next_page;
1790                }
1791
1792                if (!vma ||
1793                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1794                    !(vm_flags & vma->vm_flags))
1795                        return i ? : -EFAULT;
1796
1797                if (is_vm_hugetlb_page(vma)) {
1798                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1799                                        &start, &nr_pages, i, gup_flags);
1800                        continue;
1801                }
1802
1803                do {
1804                        struct page *page;
1805                        unsigned int foll_flags = gup_flags;
1806                        unsigned int page_increm;
1807
1808                        /*
1809                         * If we have a pending SIGKILL, don't keep faulting
1810                         * pages and potentially allocating memory.
1811                         */
1812                        if (unlikely(fatal_signal_pending(current)))
1813                                return i ? i : -ERESTARTSYS;
1814
1815                        cond_resched();
1816                        while (!(page = follow_page_mask(vma, start,
1817                                                foll_flags, &page_mask))) {
1818                                int ret;
1819                                unsigned int fault_flags = 0;
1820
1821                                /* For mlock, just skip the stack guard page. */
1822                                if (foll_flags & FOLL_MLOCK) {
1823                                        if (stack_guard_page(vma, start))
1824                                                goto next_page;
1825                                }
1826                                if (foll_flags & FOLL_WRITE)
1827                                        fault_flags |= FAULT_FLAG_WRITE;
1828                                if (nonblocking)
1829                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1830                                if (foll_flags & FOLL_NOWAIT)
1831                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1832
1833                                ret = handle_mm_fault(mm, vma, start,
1834                                                        fault_flags);
1835
1836                                if (ret & VM_FAULT_ERROR) {
1837                                        if (ret & VM_FAULT_OOM)
1838                                                return i ? i : -ENOMEM;
1839                                        if (ret & (VM_FAULT_HWPOISON |
1840                                                   VM_FAULT_HWPOISON_LARGE)) {
1841                                                if (i)
1842                                                        return i;
1843                                                else if (gup_flags & FOLL_HWPOISON)
1844                                                        return -EHWPOISON;
1845                                                else
1846                                                        return -EFAULT;
1847                                        }
1848                                        if (ret & VM_FAULT_SIGBUS)
1849                                                return i ? i : -EFAULT;
1850                                        BUG();
1851                                }
1852
1853                                if (tsk) {
1854                                        if (ret & VM_FAULT_MAJOR)
1855                                                tsk->maj_flt++;
1856                                        else
1857                                                tsk->min_flt++;
1858                                }
1859
1860                                if (ret & VM_FAULT_RETRY) {
1861                                        if (nonblocking)
1862                                                *nonblocking = 0;
1863                                        return i;
1864                                }
1865
1866                                /*
1867                                 * The VM_FAULT_WRITE bit tells us that
1868                                 * do_wp_page has broken COW when necessary,
1869                                 * even if maybe_mkwrite decided not to set
1870                                 * pte_write. We can thus safely do subsequent
1871                                 * page lookups as if they were reads. But only
1872                                 * do so when looping for pte_write is futile:
1873                                 * in some cases userspace may also be wanting
1874                                 * to write to the gotten user page, which a
1875                                 * read fault here might prevent (a readonly
1876                                 * page might get reCOWed by userspace write).
1877                                 */
1878                                if ((ret & VM_FAULT_WRITE) &&
1879                                    !(vma->vm_flags & VM_WRITE))
1880                                        foll_flags &= ~FOLL_WRITE;
1881
1882                                cond_resched();
1883                        }
1884                        if (IS_ERR(page))
1885                                return i ? i : PTR_ERR(page);
1886                        if (pages) {
1887                                pages[i] = page;
1888
1889                                flush_anon_page(vma, page, start);
1890                                flush_dcache_page(page);
1891                                page_mask = 0;
1892                        }
1893next_page:
1894                        if (vmas) {
1895                                vmas[i] = vma;
1896                                page_mask = 0;
1897                        }
1898                        page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1899                        if (page_increm > nr_pages)
1900                                page_increm = nr_pages;
1901                        i += page_increm;
1902                        start += page_increm * PAGE_SIZE;
1903                        nr_pages -= page_increm;
1904                } while (nr_pages && start < vma->vm_end);
1905        } while (nr_pages);
1906        return i;
1907}
1908EXPORT_SYMBOL(__get_user_pages);
1909
1910/*
1911 * fixup_user_fault() - manually resolve a user page fault
1912 * @tsk:        the task_struct to use for page fault accounting, or
1913 *              NULL if faults are not to be recorded.
1914 * @mm:         mm_struct of target mm
1915 * @address:    user address
1916 * @fault_flags:flags to pass down to handle_mm_fault()
1917 *
1918 * This is meant to be called in the specific scenario where for locking reasons
1919 * we try to access user memory in atomic context (within a pagefault_disable()
1920 * section), this returns -EFAULT, and we want to resolve the user fault before
1921 * trying again.
1922 *
1923 * Typically this is meant to be used by the futex code.
1924 *
1925 * The main difference with get_user_pages() is that this function will
1926 * unconditionally call handle_mm_fault() which will in turn perform all the
1927 * necessary SW fixup of the dirty and young bits in the PTE, while
1928 * handle_mm_fault() only guarantees to update these in the struct page.
1929 *
1930 * This is important for some architectures where those bits also gate the
1931 * access permission to the page because they are maintained in software.  On
1932 * such architectures, gup() will not be enough to make a subsequent access
1933 * succeed.
1934 *
1935 * This should be called with the mm_sem held for read.
1936 */
1937int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1938                     unsigned long address, unsigned int fault_flags)
1939{
1940        struct vm_area_struct *vma;
1941        int ret;
1942
1943        vma = find_extend_vma(mm, address);
1944        if (!vma || address < vma->vm_start)
1945                return -EFAULT;
1946
1947        ret = handle_mm_fault(mm, vma, address, fault_flags);
1948        if (ret & VM_FAULT_ERROR) {
1949                if (ret & VM_FAULT_OOM)
1950                        return -ENOMEM;
1951                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1952                        return -EHWPOISON;
1953                if (ret & VM_FAULT_SIGBUS)
1954                        return -EFAULT;
1955                BUG();
1956        }
1957        if (tsk) {
1958                if (ret & VM_FAULT_MAJOR)
1959                        tsk->maj_flt++;
1960                else
1961                        tsk->min_flt++;
1962        }
1963        return 0;
1964}
1965
1966/*
1967 * get_user_pages() - pin user pages in memory
1968 * @tsk:        the task_struct to use for page fault accounting, or
1969 *              NULL if faults are not to be recorded.
1970 * @mm:         mm_struct of target mm
1971 * @start:      starting user address
1972 * @nr_pages:   number of pages from start to pin
1973 * @write:      whether pages will be written to by the caller
1974 * @force:      whether to force write access even if user mapping is
1975 *              readonly. This will result in the page being COWed even
1976 *              in MAP_SHARED mappings. You do not want this.
1977 * @pages:      array that receives pointers to the pages pinned.
1978 *              Should be at least nr_pages long. Or NULL, if caller
1979 *              only intends to ensure the pages are faulted in.
1980 * @vmas:       array of pointers to vmas corresponding to each page.
1981 *              Or NULL if the caller does not require them.
1982 *
1983 * Returns number of pages pinned. This may be fewer than the number
1984 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1985 * were pinned, returns -errno. Each page returned must be released
1986 * with a put_page() call when it is finished with. vmas will only
1987 * remain valid while mmap_sem is held.
1988 *
1989 * Must be called with mmap_sem held for read or write.
1990 *
1991 * get_user_pages walks a process's page tables and takes a reference to
1992 * each struct page that each user address corresponds to at a given
1993 * instant. That is, it takes the page that would be accessed if a user
1994 * thread accesses the given user virtual address at that instant.
1995 *
1996 * This does not guarantee that the page exists in the user mappings when
1997 * get_user_pages returns, and there may even be a completely different
1998 * page there in some cases (eg. if mmapped pagecache has been invalidated
1999 * and subsequently re faulted). However it does guarantee that the page
2000 * won't be freed completely. And mostly callers simply care that the page
2001 * contains data that was valid *at some point in time*. Typically, an IO
2002 * or similar operation cannot guarantee anything stronger anyway because
2003 * locks can't be held over the syscall boundary.
2004 *
2005 * If write=0, the page must not be written to. If the page is written to,
2006 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2007 * after the page is finished with, and before put_page is called.
2008 *
2009 * get_user_pages is typically used for fewer-copy IO operations, to get a
2010 * handle on the memory by some means other than accesses via the user virtual
2011 * addresses. The pages may be submitted for DMA to devices or accessed via
2012 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2013 * use the correct cache flushing APIs.
2014 *
2015 * See also get_user_pages_fast, for performance critical applications.
2016 */
2017long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2018                unsigned long start, unsigned long nr_pages, int write,
2019                int force, struct page **pages, struct vm_area_struct **vmas)
2020{
2021        int flags = FOLL_TOUCH;
2022
2023        if (pages)
2024                flags |= FOLL_GET;
2025        if (write)
2026                flags |= FOLL_WRITE;
2027        if (force)
2028                flags |= FOLL_FORCE;
2029
2030        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2031                                NULL);
2032}
2033EXPORT_SYMBOL(get_user_pages);
2034
2035/**
2036 * get_dump_page() - pin user page in memory while writing it to core dump
2037 * @addr: user address
2038 *
2039 * Returns struct page pointer of user page pinned for dump,
2040 * to be freed afterwards by page_cache_release() or put_page().
2041 *
2042 * Returns NULL on any kind of failure - a hole must then be inserted into
2043 * the corefile, to preserve alignment with its headers; and also returns
2044 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2045 * allowing a hole to be left in the corefile to save diskspace.
2046 *
2047 * Called without mmap_sem, but after all other threads have been killed.
2048 */
2049#ifdef CONFIG_ELF_CORE
2050struct page *get_dump_page(unsigned long addr)
2051{
2052        struct vm_area_struct *vma;
2053        struct page *page;
2054
2055        if (__get_user_pages(current, current->mm, addr, 1,
2056                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2057                             NULL) < 1)
2058                return NULL;
2059        flush_cache_page(vma, addr, page_to_pfn(page));
2060        return page;
2061}
2062#endif /* CONFIG_ELF_CORE */
2063
2064pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2065                        spinlock_t **ptl)
2066{
2067        pgd_t * pgd = pgd_offset(mm, addr);
2068        pud_t * pud = pud_alloc(mm, pgd, addr);
2069        if (pud) {
2070                pmd_t * pmd = pmd_alloc(mm, pud, addr);
2071                if (pmd) {
2072                        VM_BUG_ON(pmd_trans_huge(*pmd));
2073                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
2074                }
2075        }
2076        return NULL;
2077}
2078
2079/*
2080 * This is the old fallback for page remapping.
2081 *
2082 * For historical reasons, it only allows reserved pages. Only
2083 * old drivers should use this, and they needed to mark their
2084 * pages reserved for the old functions anyway.
2085 */
2086static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2087                        struct page *page, pgprot_t prot)
2088{
2089        struct mm_struct *mm = vma->vm_mm;
2090        int retval;
2091        pte_t *pte;
2092        spinlock_t *ptl;
2093
2094        retval = -EINVAL;
2095        if (PageAnon(page))
2096                goto out;
2097        retval = -ENOMEM;
2098        flush_dcache_page(page);
2099        pte = get_locked_pte(mm, addr, &ptl);
2100        if (!pte)
2101                goto out;
2102        retval = -EBUSY;
2103        if (!pte_none(*pte))
2104                goto out_unlock;
2105
2106        /* Ok, finally just insert the thing.. */
2107        get_page(page);
2108        inc_mm_counter_fast(mm, MM_FILEPAGES);
2109        page_add_file_rmap(page);
2110        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2111
2112        retval = 0;
2113        pte_unmap_unlock(pte, ptl);
2114        return retval;
2115out_unlock:
2116        pte_unmap_unlock(pte, ptl);
2117out:
2118        return retval;
2119}
2120
2121/**
2122 * vm_insert_page - insert single page into user vma
2123 * @vma: user vma to map to
2124 * @addr: target user address of this page
2125 * @page: source kernel page
2126 *
2127 * This allows drivers to insert individual pages they've allocated
2128 * into a user vma.
2129 *
2130 * The page has to be a nice clean _individual_ kernel allocation.
2131 * If you allocate a compound page, you need to have marked it as
2132 * such (__GFP_COMP), or manually just split the page up yourself
2133 * (see split_page()).
2134 *
2135 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2136 * took an arbitrary page protection parameter. This doesn't allow
2137 * that. Your vma protection will have to be set up correctly, which
2138 * means that if you want a shared writable mapping, you'd better
2139 * ask for a shared writable mapping!
2140 *
2141 * The page does not need to be reserved.
2142 *
2143 * Usually this function is called from f_op->mmap() handler
2144 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2145 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2146 * function from other places, for example from page-fault handler.
2147 */
2148int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2149                        struct page *page)
2150{
2151        if (addr < vma->vm_start || addr >= vma->vm_end)
2152                return -EFAULT;
2153        if (!page_count(page))
2154                return -EINVAL;
2155        if (!(vma->vm_flags & VM_MIXEDMAP)) {
2156                BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2157                BUG_ON(vma->vm_flags & VM_PFNMAP);
2158                vma->vm_flags |= VM_MIXEDMAP;
2159        }
2160        return insert_page(vma, addr, page, vma->vm_page_prot);
2161}
2162EXPORT_SYMBOL(vm_insert_page);
2163
2164static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2165                        unsigned long pfn, pgprot_t prot)
2166{
2167        struct mm_struct *mm = vma->vm_mm;
2168        int retval;
2169        pte_t *pte, entry;
2170        spinlock_t *ptl;
2171
2172        retval = -ENOMEM;
2173        pte = get_locked_pte(mm, addr, &ptl);
2174        if (!pte)
2175                goto out;
2176        retval = -EBUSY;
2177        if (!pte_none(*pte))
2178                goto out_unlock;
2179
2180        /* Ok, finally just insert the thing.. */
2181        entry = pte_mkspecial(pfn_pte(pfn, prot));
2182        set_pte_at(mm, addr, pte, entry);
2183        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2184
2185        retval = 0;
2186out_unlock:
2187        pte_unmap_unlock(pte, ptl);
2188out:
2189        return retval;
2190}
2191
2192/**
2193 * vm_insert_pfn - insert single pfn into user vma
2194 * @vma: user vma to map to
2195 * @addr: target user address of this page
2196 * @pfn: source kernel pfn
2197 *
2198 * Similar to vm_insert_page, this allows drivers to insert individual pages
2199 * they've allocated into a user vma. Same comments apply.
2200 *
2201 * This function should only be called from a vm_ops->fault handler, and
2202 * in that case the handler should return NULL.
2203 *
2204 * vma cannot be a COW mapping.
2205 *
2206 * As this is called only for pages that do not currently exist, we
2207 * do not need to flush old virtual caches or the TLB.
2208 */
2209int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2210                        unsigned long pfn)
2211{
2212        int ret;
2213        pgprot_t pgprot = vma->vm_page_prot;
2214        /*
2215         * Technically, architectures with pte_special can avoid all these
2216         * restrictions (same for remap_pfn_range).  However we would like
2217         * consistency in testing and feature parity among all, so we should
2218         * try to keep these invariants in place for everybody.
2219         */
2220        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2221        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2222                                                (VM_PFNMAP|VM_MIXEDMAP));
2223        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2224        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2225
2226        if (addr < vma->vm_start || addr >= vma->vm_end)
2227                return -EFAULT;
2228        if (track_pfn_insert(vma, &pgprot, pfn))
2229                return -EINVAL;
2230
2231        ret = insert_pfn(vma, addr, pfn, pgprot);
2232
2233        return ret;
2234}
2235EXPORT_SYMBOL(vm_insert_pfn);
2236
2237int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2238                        unsigned long pfn)
2239{
2240        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2241
2242        if (addr < vma->vm_start || addr >= vma->vm_end)
2243                return -EFAULT;
2244
2245        /*
2246         * If we don't have pte special, then we have to use the pfn_valid()
2247         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2248         * refcount the page if pfn_valid is true (hence insert_page rather
2249         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2250         * without pte special, it would there be refcounted as a normal page.
2251         */
2252        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2253                struct page *page;
2254
2255                page = pfn_to_page(pfn);
2256                return insert_page(vma, addr, page, vma->vm_page_prot);
2257        }
2258        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2259}
2260EXPORT_SYMBOL(vm_insert_mixed);
2261
2262/*
2263 * maps a range of physical memory into the requested pages. the old
2264 * mappings are removed. any references to nonexistent pages results
2265 * in null mappings (currently treated as "copy-on-access")
2266 */
2267static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2268                        unsigned long addr, unsigned long end,
2269                        unsigned long pfn, pgprot_t prot)
2270{
2271        pte_t *pte;
2272        spinlock_t *ptl;
2273
2274        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2275        if (!pte)
2276                return -ENOMEM;
2277        arch_enter_lazy_mmu_mode();
2278        do {
2279                BUG_ON(!pte_none(*pte));
2280                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2281                pfn++;
2282        } while (pte++, addr += PAGE_SIZE, addr != end);
2283        arch_leave_lazy_mmu_mode();
2284        pte_unmap_unlock(pte - 1, ptl);
2285        return 0;
2286}
2287
2288static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2289                        unsigned long addr, unsigned long end,
2290                        unsigned long pfn, pgprot_t prot)
2291{
2292        pmd_t *pmd;
2293        unsigned long next;
2294
2295        pfn -= addr >> PAGE_SHIFT;
2296        pmd = pmd_alloc(mm, pud, addr);
2297        if (!pmd)
2298                return -ENOMEM;
2299        VM_BUG_ON(pmd_trans_huge(*pmd));
2300        do {
2301                next = pmd_addr_end(addr, end);
2302                if (remap_pte_range(mm, pmd, addr, next,
2303                                pfn + (addr >> PAGE_SHIFT), prot))
2304                        return -ENOMEM;
2305        } while (pmd++, addr = next, addr != end);
2306        return 0;
2307}
2308
2309static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2310                        unsigned long addr, unsigned long end,
2311                        unsigned long pfn, pgprot_t prot)
2312{
2313        pud_t *pud;
2314        unsigned long next;
2315
2316        pfn -= addr >> PAGE_SHIFT;
2317        pud = pud_alloc(mm, pgd, addr);
2318        if (!pud)
2319                return -ENOMEM;
2320        do {
2321                next = pud_addr_end(addr, end);
2322                if (remap_pmd_range(mm, pud, addr, next,
2323                                pfn + (addr >> PAGE_SHIFT), prot))
2324                        return -ENOMEM;
2325        } while (pud++, addr = next, addr != end);
2326        return 0;
2327}
2328
2329/**
2330 * remap_pfn_range - remap kernel memory to userspace
2331 * @vma: user vma to map to
2332 * @addr: target user address to start at
2333 * @pfn: physical address of kernel memory
2334 * @size: size of map area
2335 * @prot: page protection flags for this mapping
2336 *
2337 *  Note: this is only safe if the mm semaphore is held when called.
2338 */
2339int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2340                    unsigned long pfn, unsigned long size, pgprot_t prot)
2341{
2342        pgd_t *pgd;
2343        unsigned long next;
2344        unsigned long end = addr + PAGE_ALIGN(size);
2345        struct mm_struct *mm = vma->vm_mm;
2346        int err;
2347
2348        /*
2349         * Physically remapped pages are special. Tell the
2350         * rest of the world about it:
2351         *   VM_IO tells people not to look at these pages
2352         *      (accesses can have side effects).
2353         *   VM_PFNMAP tells the core MM that the base pages are just
2354         *      raw PFN mappings, and do not have a "struct page" associated
2355         *      with them.
2356         *   VM_DONTEXPAND
2357         *      Disable vma merging and expanding with mremap().
2358         *   VM_DONTDUMP
2359         *      Omit vma from core dump, even when VM_IO turned off.
2360         *
2361         * There's a horrible special case to handle copy-on-write
2362         * behaviour that some programs depend on. We mark the "original"
2363         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2364         * See vm_normal_page() for details.
2365         */
2366        if (is_cow_mapping(vma->vm_flags)) {
2367                if (addr != vma->vm_start || end != vma->vm_end)
2368                        return -EINVAL;
2369                vma->vm_pgoff = pfn;
2370        }
2371
2372        err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2373        if (err)
2374                return -EINVAL;
2375
2376        vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2377
2378        BUG_ON(addr >= end);
2379        pfn -= addr >> PAGE_SHIFT;
2380        pgd = pgd_offset(mm, addr);
2381        flush_cache_range(vma, addr, end);
2382        do {
2383                next = pgd_addr_end(addr, end);
2384                err = remap_pud_range(mm, pgd, addr, next,
2385                                pfn + (addr >> PAGE_SHIFT), prot);
2386                if (err)
2387                        break;
2388        } while (pgd++, addr = next, addr != end);
2389
2390        if (err)
2391                untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2392
2393        return err;
2394}
2395EXPORT_SYMBOL(remap_pfn_range);
2396
2397/**
2398 * vm_iomap_memory - remap memory to userspace
2399 * @vma: user vma to map to
2400 * @start: start of area
2401 * @len: size of area
2402 *
2403 * This is a simplified io_remap_pfn_range() for common driver use. The
2404 * driver just needs to give us the physical memory range to be mapped,
2405 * we'll figure out the rest from the vma information.
2406 *
2407 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2408 * whatever write-combining details or similar.
2409 */
2410int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2411{
2412        unsigned long vm_len, pfn, pages;
2413
2414        /* Check that the physical memory area passed in looks valid */
2415        if (start + len < start)
2416                return -EINVAL;
2417        /*
2418         * You *really* shouldn't map things that aren't page-aligned,
2419         * but we've historically allowed it because IO memory might
2420         * just have smaller alignment.
2421         */
2422        len += start & ~PAGE_MASK;
2423        pfn = start >> PAGE_SHIFT;
2424        pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2425        if (pfn + pages < pfn)
2426                return -EINVAL;
2427
2428        /* We start the mapping 'vm_pgoff' pages into the area */
2429        if (vma->vm_pgoff > pages)
2430                return -EINVAL;
2431        pfn += vma->vm_pgoff;
2432        pages -= vma->vm_pgoff;
2433
2434        /* Can we fit all of the mapping? */
2435        vm_len = vma->vm_end - vma->vm_start;
2436        if (vm_len >> PAGE_SHIFT > pages)
2437                return -EINVAL;
2438
2439        /* Ok, let it rip */
2440        return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2441}
2442EXPORT_SYMBOL(vm_iomap_memory);
2443
2444static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2445                                     unsigned long addr, unsigned long end,
2446                                     pte_fn_t fn, void *data)
2447{
2448        pte_t *pte;
2449        int err;
2450        pgtable_t token;
2451        spinlock_t *uninitialized_var(ptl);
2452
2453        pte = (mm == &init_mm) ?
2454                pte_alloc_kernel(pmd, addr) :
2455                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2456        if (!pte)
2457                return -ENOMEM;
2458
2459        BUG_ON(pmd_huge(*pmd));
2460
2461        arch_enter_lazy_mmu_mode();
2462
2463        token = pmd_pgtable(*pmd);
2464
2465        do {
2466                err = fn(pte++, token, addr, data);
2467                if (err)
2468                        break;
2469        } while (addr += PAGE_SIZE, addr != end);
2470
2471        arch_leave_lazy_mmu_mode();
2472
2473        if (mm != &init_mm)
2474                pte_unmap_unlock(pte-1, ptl);
2475        return err;
2476}
2477
2478static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2479                                     unsigned long addr, unsigned long end,
2480                                     pte_fn_t fn, void *data)
2481{
2482        pmd_t *pmd;
2483        unsigned long next;
2484        int err;
2485
2486        BUG_ON(pud_huge(*pud));
2487
2488        pmd = pmd_alloc(mm, pud, addr);
2489        if (!pmd)
2490                return -ENOMEM;
2491        do {
2492                next = pmd_addr_end(addr, end);
2493                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2494                if (err)
2495                        break;
2496        } while (pmd++, addr = next, addr != end);
2497        return err;
2498}
2499
2500static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2501                                     unsigned long addr, unsigned long end,
2502                                     pte_fn_t fn, void *data)
2503{
2504        pud_t *pud;
2505        unsigned long next;
2506        int err;
2507
2508        pud = pud_alloc(mm, pgd, addr);
2509        if (!pud)
2510                return -ENOMEM;
2511        do {
2512                next = pud_addr_end(addr, end);
2513                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2514                if (err)
2515                        break;
2516        } while (pud++, addr = next, addr != end);
2517        return err;
2518}
2519
2520/*
2521 * Scan a region of virtual memory, filling in page tables as necessary
2522 * and calling a provided function on each leaf page table.
2523 */
2524int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2525                        unsigned long size, pte_fn_t fn, void *data)
2526{
2527        pgd_t *pgd;
2528        unsigned long next;
2529        unsigned long end = addr + size;
2530        int err;
2531
2532        BUG_ON(addr >= end);
2533        pgd = pgd_offset(mm, addr);
2534        do {
2535                next = pgd_addr_end(addr, end);
2536                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2537                if (err)
2538                        break;
2539        } while (pgd++, addr = next, addr != end);
2540
2541        return err;
2542}
2543EXPORT_SYMBOL_GPL(apply_to_page_range);
2544
2545/*
2546 * handle_pte_fault chooses page fault handler according to an entry
2547 * which was read non-atomically.  Before making any commitment, on
2548 * those architectures or configurations (e.g. i386 with PAE) which
2549 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2550 * must check under lock before unmapping the pte and proceeding
2551 * (but do_wp_page is only called after already making such a check;
2552 * and do_anonymous_page can safely check later on).
2553 */
2554static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2555                                pte_t *page_table, pte_t orig_pte)
2556{
2557        int same = 1;
2558#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2559        if (sizeof(pte_t) > sizeof(unsigned long)) {
2560                spinlock_t *ptl = pte_lockptr(mm, pmd);
2561                spin_lock(ptl);
2562                same = pte_same(*page_table, orig_pte);
2563                spin_unlock(ptl);
2564        }
2565#endif
2566        pte_unmap(page_table);
2567        return same;
2568}
2569
2570static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2571{
2572        /*
2573         * If the source page was a PFN mapping, we don't have
2574         * a "struct page" for it. We do a best-effort copy by
2575         * just copying from the original user address. If that
2576         * fails, we just zero-fill it. Live with it.
2577         */
2578        if (unlikely(!src)) {
2579                void *kaddr = kmap_atomic(dst);
2580                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2581
2582                /*
2583                 * This really shouldn't fail, because the page is there
2584                 * in the page tables. But it might just be unreadable,
2585                 * in which case we just give up and fill the result with
2586                 * zeroes.
2587                 */
2588                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2589                        clear_page(kaddr);
2590                kunmap_atomic(kaddr);
2591                flush_dcache_page(dst);
2592        } else
2593                copy_user_highpage(dst, src, va, vma);
2594}
2595
2596/*
2597 * This routine handles present pages, when users try to write
2598 * to a shared page. It is done by copying the page to a new address
2599 * and decrementing the shared-page counter for the old page.
2600 *
2601 * Note that this routine assumes that the protection checks have been
2602 * done by the caller (the low-level page fault routine in most cases).
2603 * Thus we can safely just mark it writable once we've done any necessary
2604 * COW.
2605 *
2606 * We also mark the page dirty at this point even though the page will
2607 * change only once the write actually happens. This avoids a few races,
2608 * and potentially makes it more efficient.
2609 *
2610 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2611 * but allow concurrent faults), with pte both mapped and locked.
2612 * We return with mmap_sem still held, but pte unmapped and unlocked.
2613 */
2614static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2615                unsigned long address, pte_t *page_table, pmd_t *pmd,
2616                spinlock_t *ptl, pte_t orig_pte)
2617        __releases(ptl)
2618{
2619        struct page *old_page, *new_page = NULL;
2620        pte_t entry;
2621        int ret = 0;
2622        int page_mkwrite = 0;
2623        struct page *dirty_page = NULL;
2624        unsigned long mmun_start = 0;   /* For mmu_notifiers */
2625        unsigned long mmun_end = 0;     /* For mmu_notifiers */
2626
2627        old_page = vm_normal_page(vma, address, orig_pte);
2628        if (!old_page) {
2629                /*
2630                 * VM_MIXEDMAP !pfn_valid() case
2631                 *
2632                 * We should not cow pages in a shared writeable mapping.
2633                 * Just mark the pages writable as we can't do any dirty
2634                 * accounting on raw pfn maps.
2635                 */
2636                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2637                                     (VM_WRITE|VM_SHARED))
2638                        goto reuse;
2639                goto gotten;
2640        }
2641
2642        /*
2643         * Take out anonymous pages first, anonymous shared vmas are
2644         * not dirty accountable.
2645         */
2646        if (PageAnon(old_page) && !PageKsm(old_page)) {
2647                if (!trylock_page(old_page)) {
2648                        page_cache_get(old_page);
2649                        pte_unmap_unlock(page_table, ptl);
2650                        lock_page(old_page);
2651                        page_table = pte_offset_map_lock(mm, pmd, address,
2652                                                         &ptl);
2653                        if (!pte_same(*page_table, orig_pte)) {
2654                                unlock_page(old_page);
2655                                goto unlock;
2656                        }
2657                        page_cache_release(old_page);
2658                }
2659                if (reuse_swap_page(old_page)) {
2660                        /*
2661                         * The page is all ours.  Move it to our anon_vma so
2662                         * the rmap code will not search our parent or siblings.
2663                         * Protected against the rmap code by the page lock.
2664                         */
2665                        page_move_anon_rmap(old_page, vma, address);
2666                        unlock_page(old_page);
2667                        goto reuse;
2668                }
2669                unlock_page(old_page);
2670        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2671                                        (VM_WRITE|VM_SHARED))) {
2672                /*
2673                 * Only catch write-faults on shared writable pages,
2674                 * read-only shared pages can get COWed by
2675                 * get_user_pages(.write=1, .force=1).
2676                 */
2677                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2678                        struct vm_fault vmf;
2679                        int tmp;
2680
2681                        vmf.virtual_address = (void __user *)(address &
2682                                                                PAGE_MASK);
2683                        vmf.pgoff = old_page->index;
2684                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2685                        vmf.page = old_page;
2686
2687                        /*
2688                         * Notify the address space that the page is about to
2689                         * become writable so that it can prohibit this or wait
2690                         * for the page to get into an appropriate state.
2691                         *
2692                         * We do this without the lock held, so that it can
2693                         * sleep if it needs to.
2694                         */
2695                        page_cache_get(old_page);
2696                        pte_unmap_unlock(page_table, ptl);
2697
2698                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2699                        if (unlikely(tmp &
2700                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2701                                ret = tmp;
2702                                goto unwritable_page;
2703                        }
2704                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2705                                lock_page(old_page);
2706                                if (!old_page->mapping) {
2707                                        ret = 0; /* retry the fault */
2708                                        unlock_page(old_page);
2709                                        goto unwritable_page;
2710                                }
2711                        } else
2712                                VM_BUG_ON(!PageLocked(old_page));
2713
2714                        /*
2715                         * Since we dropped the lock we need to revalidate
2716                         * the PTE as someone else may have changed it.  If
2717                         * they did, we just return, as we can count on the
2718                         * MMU to tell us if they didn't also make it writable.
2719                         */
2720                        page_table = pte_offset_map_lock(mm, pmd, address,
2721                                                         &ptl);
2722                        if (!pte_same(*page_table, orig_pte)) {
2723                                unlock_page(old_page);
2724                                goto unlock;
2725                        }
2726
2727                        page_mkwrite = 1;
2728                }
2729                dirty_page = old_page;
2730                get_page(dirty_page);
2731
2732reuse:
2733                flush_cache_page(vma, address, pte_pfn(orig_pte));
2734                entry = pte_mkyoung(orig_pte);
2735                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2736                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2737                        update_mmu_cache(vma, address, page_table);
2738                pte_unmap_unlock(page_table, ptl);
2739                ret |= VM_FAULT_WRITE;
2740
2741                if (!dirty_page)
2742                        return ret;
2743
2744                /*
2745                 * Yes, Virginia, this is actually required to prevent a race
2746                 * with clear_page_dirty_for_io() from clearing the page dirty
2747                 * bit after it clear all dirty ptes, but before a racing
2748                 * do_wp_page installs a dirty pte.
2749                 *
2750                 * __do_fault is protected similarly.
2751                 */
2752                if (!page_mkwrite) {
2753                        wait_on_page_locked(dirty_page);
2754                        set_page_dirty_balance(dirty_page, page_mkwrite);
2755                        /* file_update_time outside page_lock */
2756                        if (vma->vm_file)
2757                                file_update_time(vma->vm_file);
2758                }
2759                put_page(dirty_page);
2760                if (page_mkwrite) {
2761                        struct address_space *mapping = dirty_page->mapping;
2762
2763                        set_page_dirty(dirty_page);
2764                        unlock_page(dirty_page);
2765                        page_cache_release(dirty_page);
2766                        if (mapping)    {
2767                                /*
2768                                 * Some device drivers do not set page.mapping
2769                                 * but still dirty their pages
2770                                 */
2771                                balance_dirty_pages_ratelimited(mapping);
2772                        }
2773                }
2774
2775                return ret;
2776        }
2777
2778        /*
2779         * Ok, we need to copy. Oh, well..
2780         */
2781        page_cache_get(old_page);
2782gotten:
2783        pte_unmap_unlock(page_table, ptl);
2784
2785        if (unlikely(anon_vma_prepare(vma)))
2786                goto oom;
2787
2788        if (is_zero_pfn(pte_pfn(orig_pte))) {
2789                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2790                if (!new_page)
2791                        goto oom;
2792        } else {
2793                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2794                if (!new_page)
2795                        goto oom;
2796                cow_user_page(new_page, old_page, address, vma);
2797        }
2798        __SetPageUptodate(new_page);
2799
2800        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2801                goto oom_free_new;
2802
2803        mmun_start  = address & PAGE_MASK;
2804        mmun_end    = mmun_start + PAGE_SIZE;
2805        mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2806
2807        /*
2808         * Re-check the pte - we dropped the lock
2809         */
2810        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2811        if (likely(pte_same(*page_table, orig_pte))) {
2812                if (old_page) {
2813                        if (!PageAnon(old_page)) {
2814                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2815                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2816                        }
2817                } else
2818                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2819                flush_cache_page(vma, address, pte_pfn(orig_pte));
2820                entry = mk_pte(new_page, vma->vm_page_prot);
2821                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2822                /*
2823                 * Clear the pte entry and flush it first, before updating the
2824                 * pte with the new entry. This will avoid a race condition
2825                 * seen in the presence of one thread doing SMC and another
2826                 * thread doing COW.
2827                 */
2828                ptep_clear_flush(vma, address, page_table);
2829                page_add_new_anon_rmap(new_page, vma, address);
2830                /*
2831                 * We call the notify macro here because, when using secondary
2832                 * mmu page tables (such as kvm shadow page tables), we want the
2833                 * new page to be mapped directly into the secondary page table.
2834                 */
2835                set_pte_at_notify(mm, address, page_table, entry);
2836                update_mmu_cache(vma, address, page_table);
2837                if (old_page) {
2838                        /*
2839                         * Only after switching the pte to the new page may
2840                         * we remove the mapcount here. Otherwise another
2841                         * process may come and find the rmap count decremented
2842                         * before the pte is switched to the new page, and
2843                         * "reuse" the old page writing into it while our pte
2844                         * here still points into it and can be read by other
2845                         * threads.
2846                         *
2847                         * The critical issue is to order this
2848                         * page_remove_rmap with the ptp_clear_flush above.
2849                         * Those stores are ordered by (if nothing else,)
2850                         * the barrier present in the atomic_add_negative
2851                         * in page_remove_rmap.
2852                         *
2853                         * Then the TLB flush in ptep_clear_flush ensures that
2854                         * no process can access the old page before the
2855                         * decremented mapcount is visible. And the old page
2856                         * cannot be reused until after the decremented
2857                         * mapcount is visible. So transitively, TLBs to
2858                         * old page will be flushed before it can be reused.
2859                         */
2860                        page_remove_rmap(old_page);
2861                }
2862
2863                /* Free the old page.. */
2864                new_page = old_page;
2865                ret |= VM_FAULT_WRITE;
2866        } else
2867                mem_cgroup_uncharge_page(new_page);
2868
2869        if (new_page)
2870                page_cache_release(new_page);
2871unlock:
2872        pte_unmap_unlock(page_table, ptl);
2873        if (mmun_end > mmun_start)
2874                mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2875        if (old_page) {
2876                /*
2877                 * Don't let another task, with possibly unlocked vma,
2878                 * keep the mlocked page.
2879                 */
2880                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2881                        lock_page(old_page);    /* LRU manipulation */
2882                        munlock_vma_page(old_page);
2883                        unlock_page(old_page);
2884                }
2885                page_cache_release(old_page);
2886        }
2887        return ret;
2888oom_free_new:
2889        page_cache_release(new_page);
2890oom:
2891        if (old_page)
2892                page_cache_release(old_page);
2893        return VM_FAULT_OOM;
2894
2895unwritable_page:
2896        page_cache_release(old_page);
2897        return ret;
2898}
2899
2900static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2901                unsigned long start_addr, unsigned long end_addr,
2902                struct zap_details *details)
2903{
2904        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2905}
2906
2907static inline void unmap_mapping_range_tree(struct rb_root *root,
2908                                            struct zap_details *details)
2909{
2910        struct vm_area_struct *vma;
2911        pgoff_t vba, vea, zba, zea;
2912
2913        vma_interval_tree_foreach(vma, root,
2914                        details->first_index, details->last_index) {
2915
2916                vba = vma->vm_pgoff;
2917                vea = vba + vma_pages(vma) - 1;
2918                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2919                zba = details->first_index;
2920                if (zba < vba)
2921                        zba = vba;
2922                zea = details->last_index;
2923                if (zea > vea)
2924                        zea = vea;
2925
2926                unmap_mapping_range_vma(vma,
2927                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2928                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2929                                details);
2930        }
2931}
2932
2933static inline void unmap_mapping_range_list(struct list_head *head,
2934                                            struct zap_details *details)
2935{
2936        struct vm_area_struct *vma;
2937
2938        /*
2939         * In nonlinear VMAs there is no correspondence between virtual address
2940         * offset and file offset.  So we must perform an exhaustive search
2941         * across *all* the pages in each nonlinear VMA, not just the pages
2942         * whose virtual address lies outside the file truncation point.
2943         */
2944        list_for_each_entry(vma, head, shared.nonlinear) {
2945                details->nonlinear_vma = vma;
2946                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2947        }
2948}
2949
2950/**
2951 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2952 * @mapping: the address space containing mmaps to be unmapped.
2953 * @holebegin: byte in first page to unmap, relative to the start of
2954 * the underlying file.  This will be rounded down to a PAGE_SIZE
2955 * boundary.  Note that this is different from truncate_pagecache(), which
2956 * must keep the partial page.  In contrast, we must get rid of
2957 * partial pages.
2958 * @holelen: size of prospective hole in bytes.  This will be rounded
2959 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2960 * end of the file.
2961 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2962 * but 0 when invalidating pagecache, don't throw away private data.
2963 */
2964void unmap_mapping_range(struct address_space *mapping,
2965                loff_t const holebegin, loff_t const holelen, int even_cows)
2966{
2967        struct zap_details details;
2968        pgoff_t hba = holebegin >> PAGE_SHIFT;
2969        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2970
2971        /* Check for overflow. */
2972        if (sizeof(holelen) > sizeof(hlen)) {
2973                long long holeend =
2974                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2975                if (holeend & ~(long long)ULONG_MAX)
2976                        hlen = ULONG_MAX - hba + 1;
2977        }
2978
2979        details.check_mapping = even_cows? NULL: mapping;
2980        details.nonlinear_vma = NULL;
2981        details.first_index = hba;
2982        details.last_index = hba + hlen - 1;
2983        if (details.last_index < details.first_index)
2984                details.last_index = ULONG_MAX;
2985
2986
2987        mutex_lock(&mapping->i_mmap_mutex);
2988        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2989                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2990        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2991                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2992        mutex_unlock(&mapping->i_mmap_mutex);
2993}
2994EXPORT_SYMBOL(unmap_mapping_range);
2995
2996/*
2997 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2998 * but allow concurrent faults), and pte mapped but not yet locked.
2999 * We return with mmap_sem still held, but pte unmapped and unlocked.
3000 */
3001static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3002                unsigned long address, pte_t *page_table, pmd_t *pmd,
3003                unsigned int flags, pte_t orig_pte)
3004{
3005        spinlock_t *ptl;
3006        struct page *page, *swapcache;
3007        swp_entry_t entry;
3008        pte_t pte;
3009        int locked;
3010        struct mem_cgroup *ptr;
3011        int exclusive = 0;
3012        int ret = 0;
3013
3014        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3015                goto out;
3016
3017        entry = pte_to_swp_entry(orig_pte);
3018        if (unlikely(non_swap_entry(entry))) {
3019                if (is_migration_entry(entry)) {
3020                        migration_entry_wait(mm, pmd, address);
3021                } else if (is_hwpoison_entry(entry)) {
3022                        ret = VM_FAULT_HWPOISON;
3023                } else {
3024                        print_bad_pte(vma, address, orig_pte, NULL);
3025                        ret = VM_FAULT_SIGBUS;
3026                }
3027                goto out;
3028        }
3029        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3030        page = lookup_swap_cache(entry);
3031        if (!page) {
3032                page = swapin_readahead(entry,
3033                                        GFP_HIGHUSER_MOVABLE, vma, address);
3034                if (!page) {
3035                        /*
3036                         * Back out if somebody else faulted in this pte
3037                         * while we released the pte lock.
3038                         */
3039                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3040                        if (likely(pte_same(*page_table, orig_pte)))
3041                                ret = VM_FAULT_OOM;
3042                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3043                        goto unlock;
3044                }
3045
3046                /* Had to read the page from swap area: Major fault */
3047                ret = VM_FAULT_MAJOR;
3048                count_vm_event(PGMAJFAULT);
3049                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3050        } else if (PageHWPoison(page)) {
3051                /*
3052                 * hwpoisoned dirty swapcache pages are kept for killing
3053                 * owner processes (which may be unknown at hwpoison time)
3054                 */
3055                ret = VM_FAULT_HWPOISON;
3056                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3057                swapcache = page;
3058                goto out_release;
3059        }
3060
3061        swapcache = page;
3062        locked = lock_page_or_retry(page, mm, flags);
3063
3064        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3065        if (!locked) {
3066                ret |= VM_FAULT_RETRY;
3067                goto out_release;
3068        }
3069
3070        /*
3071         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3072         * release the swapcache from under us.  The page pin, and pte_same
3073         * test below, are not enough to exclude that.  Even if it is still
3074         * swapcache, we need to check that the page's swap has not changed.
3075         */
3076        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3077                goto out_page;
3078
3079        page = ksm_might_need_to_copy(page, vma, address);
3080        if (unlikely(!page)) {
3081                ret = VM_FAULT_OOM;
3082                page = swapcache;
3083                goto out_page;
3084        }
3085
3086        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3087                ret = VM_FAULT_OOM;
3088                goto out_page;
3089        }
3090
3091        /*
3092         * Back out if somebody else already faulted in this pte.
3093         */
3094        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3095        if (unlikely(!pte_same(*page_table, orig_pte)))
3096                goto out_nomap;
3097
3098        if (unlikely(!PageUptodate(page))) {
3099                ret = VM_FAULT_SIGBUS;
3100                goto out_nomap;
3101        }
3102
3103        /*
3104         * The page isn't present yet, go ahead with the fault.
3105         *
3106         * Be careful about the sequence of operations here.
3107         * To get its accounting right, reuse_swap_page() must be called
3108         * while the page is counted on swap but not yet in mapcount i.e.
3109         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3110         * must be called after the swap_free(), or it will never succeed.
3111         * Because delete_from_swap_page() may be called by reuse_swap_page(),
3112         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3113         * in page->private. In this case, a record in swap_cgroup  is silently
3114         * discarded at swap_free().
3115         */
3116
3117        inc_mm_counter_fast(mm, MM_ANONPAGES);
3118        dec_mm_counter_fast(mm, MM_SWAPENTS);
3119        pte = mk_pte(page, vma->vm_page_prot);
3120        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3121                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3122                flags &= ~FAULT_FLAG_WRITE;
3123                ret |= VM_FAULT_WRITE;
3124                exclusive = 1;
3125        }
3126        flush_icache_page(vma, page);
3127        if (pte_swp_soft_dirty(orig_pte))
3128                pte = pte_mksoft_dirty(pte);
3129        set_pte_at(mm, address, page_table, pte);
3130        if (page == swapcache)
3131                do_page_add_anon_rmap(page, vma, address, exclusive);
3132        else /* ksm created a completely new copy */
3133                page_add_new_anon_rmap(page, vma, address);
3134        /* It's better to call commit-charge after rmap is established */
3135        mem_cgroup_commit_charge_swapin(page, ptr);
3136
3137        swap_free(entry);
3138        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3139                try_to_free_swap(page);
3140        unlock_page(page);
3141        if (page != swapcache) {
3142                /*
3143                 * Hold the lock to avoid the swap entry to be reused
3144                 * until we take the PT lock for the pte_same() check
3145                 * (to avoid false positives from pte_same). For
3146                 * further safety release the lock after the swap_free
3147                 * so that the swap count won't change under a
3148                 * parallel locked swapcache.
3149                 */
3150                unlock_page(swapcache);
3151                page_cache_release(swapcache);
3152        }
3153
3154        if (flags & FAULT_FLAG_WRITE) {
3155                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3156                if (ret & VM_FAULT_ERROR)
3157                        ret &= VM_FAULT_ERROR;
3158                goto out;
3159        }
3160
3161        /* No need to invalidate - it was non-present before */
3162        update_mmu_cache(vma, address, page_table);
3163unlock:
3164        pte_unmap_unlock(page_table, ptl);
3165out:
3166        return ret;
3167out_nomap:
3168        mem_cgroup_cancel_charge_swapin(ptr);
3169        pte_unmap_unlock(page_table, ptl);
3170out_page:
3171        unlock_page(page);
3172out_release:
3173        page_cache_release(page);
3174        if (page != swapcache) {
3175                unlock_page(swapcache);
3176                page_cache_release(swapcache);
3177        }
3178        return ret;
3179}
3180
3181/*
3182 * This is like a special single-page "expand_{down|up}wards()",
3183 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3184 * doesn't hit another vma.
3185 */
3186static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3187{
3188        address &= PAGE_MASK;
3189        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3190                struct vm_area_struct *prev = vma->vm_prev;
3191
3192                /*
3193                 * Is there a mapping abutting this one below?
3194                 *
3195                 * That's only ok if it's the same stack mapping
3196                 * that has gotten split..
3197                 */
3198                if (prev && prev->vm_end == address)
3199                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3200
3201                expand_downwards(vma, address - PAGE_SIZE);
3202        }
3203        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3204                struct vm_area_struct *next = vma->vm_next;
3205
3206                /* As VM_GROWSDOWN but s/below/above/ */
3207                if (next && next->vm_start == address + PAGE_SIZE)
3208                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3209
3210                expand_upwards(vma, address + PAGE_SIZE);
3211        }
3212        return 0;
3213}
3214
3215/*
3216 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3217 * but allow concurrent faults), and pte mapped but not yet locked.
3218 * We return with mmap_sem still held, but pte unmapped and unlocked.
3219 */
3220static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3221                unsigned long address, pte_t *page_table, pmd_t *pmd,
3222                unsigned int flags)
3223{
3224        struct page *page;
3225        spinlock_t *ptl;
3226        pte_t entry;
3227
3228        pte_unmap(page_table);
3229
3230        /* Check if we need to add a guard page to the stack */
3231        if (check_stack_guard_page(vma, address) < 0)
3232                return VM_FAULT_SIGBUS;
3233
3234        /* Use the zero-page for reads */
3235        if (!(flags & FAULT_FLAG_WRITE)) {
3236                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3237                                                vma->vm_page_prot));
3238                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3239                if (!pte_none(*page_table))
3240                        goto unlock;
3241                goto setpte;
3242        }
3243
3244        /* Allocate our own private page. */
3245        if (unlikely(anon_vma_prepare(vma)))
3246                goto oom;
3247        page = alloc_zeroed_user_highpage_movable(vma, address);
3248        if (!page)
3249                goto oom;
3250        /*
3251         * The memory barrier inside __SetPageUptodate makes sure that
3252         * preceeding stores to the page contents become visible before
3253         * the set_pte_at() write.
3254         */
3255        __SetPageUptodate(page);
3256
3257        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3258                goto oom_free_page;
3259
3260        entry = mk_pte(page, vma->vm_page_prot);
3261        if (vma->vm_flags & VM_WRITE)
3262                entry = pte_mkwrite(pte_mkdirty(entry));
3263
3264        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3265        if (!pte_none(*page_table))
3266                goto release;
3267
3268        inc_mm_counter_fast(mm, MM_ANONPAGES);
3269        page_add_new_anon_rmap(page, vma, address);
3270setpte:
3271        set_pte_at(mm, address, page_table, entry);
3272
3273        /* No need to invalidate - it was non-present before */
3274        update_mmu_cache(vma, address, page_table);
3275unlock:
3276        pte_unmap_unlock(page_table, ptl);
3277        return 0;
3278release:
3279        mem_cgroup_uncharge_page(page);
3280        page_cache_release(page);
3281        goto unlock;
3282oom_free_page:
3283        page_cache_release(page);
3284oom:
3285        return VM_FAULT_OOM;
3286}
3287
3288/*
3289 * __do_fault() tries to create a new page mapping. It aggressively
3290 * tries to share with existing pages, but makes a separate copy if
3291 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3292 * the next page fault.
3293 *
3294 * As this is called only for pages that do not currently exist, we
3295 * do not need to flush old virtual caches or the TLB.
3296 *
3297 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3298 * but allow concurrent faults), and pte neither mapped nor locked.
3299 * We return with mmap_sem still held, but pte unmapped and unlocked.
3300 */
3301static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3302                unsigned long address, pmd_t *pmd,
3303                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3304{
3305        pte_t *page_table;
3306        spinlock_t *ptl;
3307        struct page *page;
3308        struct page *cow_page;
3309        pte_t entry;
3310        int anon = 0;
3311        struct page *dirty_page = NULL;
3312        struct vm_fault vmf;
3313        int ret;
3314        int page_mkwrite = 0;
3315
3316        /*
3317         * If we do COW later, allocate page befor taking lock_page()
3318         * on the file cache page. This will reduce lock holding time.
3319         */
3320        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3321
3322                if (unlikely(anon_vma_prepare(vma)))
3323                        return VM_FAULT_OOM;
3324
3325                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3326                if (!cow_page)
3327                        return VM_FAULT_OOM;
3328
3329                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3330                        page_cache_release(cow_page);
3331                        return VM_FAULT_OOM;
3332                }
3333        } else
3334                cow_page = NULL;
3335
3336        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3337        vmf.pgoff = pgoff;
3338        vmf.flags = flags;
3339        vmf.page = NULL;
3340
3341        ret = vma->vm_ops->fault(vma, &vmf);
3342        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3343                            VM_FAULT_RETRY)))
3344                goto uncharge_out;
3345
3346        if (unlikely(PageHWPoison(vmf.page))) {
3347                if (ret & VM_FAULT_LOCKED)
3348                        unlock_page(vmf.page);
3349                ret = VM_FAULT_HWPOISON;
3350                goto uncharge_out;
3351        }
3352
3353        /*
3354         * For consistency in subsequent calls, make the faulted page always
3355         * locked.
3356         */
3357        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358                lock_page(vmf.page);
3359        else
3360                VM_BUG_ON(!PageLocked(vmf.page));
3361
3362        /*
3363         * Should we do an early C-O-W break?
3364         */
3365        page = vmf.page;
3366        if (flags & FAULT_FLAG_WRITE) {
3367                if (!(vma->vm_flags & VM_SHARED)) {
3368                        page = cow_page;
3369                        anon = 1;
3370                        copy_user_highpage(page, vmf.page, address, vma);
3371                        __SetPageUptodate(page);
3372                } else {
3373                        /*
3374                         * If the page will be shareable, see if the backing
3375                         * address space wants to know that the page is about
3376                         * to become writable
3377                         */
3378                        if (vma->vm_ops->page_mkwrite) {
3379                                int tmp;
3380
3381                                unlock_page(page);
3382                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3383                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3384                                if (unlikely(tmp &
3385                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3386                                        ret = tmp;
3387                                        goto unwritable_page;
3388                                }
3389                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3390                                        lock_page(page);
3391                                        if (!page->mapping) {
3392                                                ret = 0; /* retry the fault */
3393                                                unlock_page(page);
3394                                                goto unwritable_page;
3395                                        }
3396                                } else
3397                                        VM_BUG_ON(!PageLocked(page));
3398                                page_mkwrite = 1;
3399                        }
3400                }
3401
3402        }
3403
3404        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3405
3406        /*
3407         * This silly early PAGE_DIRTY setting removes a race
3408         * due to the bad i386 page protection. But it's valid
3409         * for other architectures too.
3410         *
3411         * Note that if FAULT_FLAG_WRITE is set, we either now have
3412         * an exclusive copy of the page, or this is a shared mapping,
3413         * so we can make it writable and dirty to avoid having to
3414         * handle that later.
3415         */
3416        /* Only go through if we didn't race with anybody else... */
3417        if (likely(pte_same(*page_table, orig_pte))) {
3418                flush_icache_page(vma, page);
3419                entry = mk_pte(page, vma->vm_page_prot);
3420                if (flags & FAULT_FLAG_WRITE)
3421                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3422                else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3423                        pte_mksoft_dirty(entry);
3424                if (anon) {
3425                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3426                        page_add_new_anon_rmap(page, vma, address);
3427                } else {
3428                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3429                        page_add_file_rmap(page);
3430                        if (flags & FAULT_FLAG_WRITE) {
3431                                dirty_page = page;
3432                                get_page(dirty_page);
3433                        }
3434                }
3435                set_pte_at(mm, address, page_table, entry);
3436
3437                /* no need to invalidate: a not-present page won't be cached */
3438                update_mmu_cache(vma, address, page_table);
3439        } else {
3440                if (cow_page)
3441                        mem_cgroup_uncharge_page(cow_page);
3442                if (anon)
3443                        page_cache_release(page);
3444                else
3445                        anon = 1; /* no anon but release faulted_page */
3446        }
3447
3448        pte_unmap_unlock(page_table, ptl);
3449
3450        if (dirty_page) {
3451                struct address_space *mapping = page->mapping;
3452                int dirtied = 0;
3453
3454                if (set_page_dirty(dirty_page))
3455                        dirtied = 1;
3456                unlock_page(dirty_page);
3457                put_page(dirty_page);
3458                if ((dirtied || page_mkwrite) && mapping) {
3459                        /*
3460                         * Some device drivers do not set page.mapping but still
3461                         * dirty their pages
3462                         */
3463                        balance_dirty_pages_ratelimited(mapping);
3464                }
3465
3466                /* file_update_time outside page_lock */
3467                if (vma->vm_file && !page_mkwrite)
3468                        file_update_time(vma->vm_file);
3469        } else {
3470                unlock_page(vmf.page);
3471                if (anon)
3472                        page_cache_release(vmf.page);
3473        }
3474
3475        return ret;
3476
3477unwritable_page:
3478        page_cache_release(page);
3479        return ret;
3480uncharge_out:
3481        /* fs's fault handler get error */
3482        if (cow_page) {
3483                mem_cgroup_uncharge_page(cow_page);
3484                page_cache_release(cow_page);
3485        }
3486        return ret;
3487}
3488
3489static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3490                unsigned long address, pte_t *page_table, pmd_t *pmd,
3491                unsigned int flags, pte_t orig_pte)
3492{
3493        pgoff_t pgoff = (((address & PAGE_MASK)
3494                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3495
3496        pte_unmap(page_table);
3497        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3498}
3499
3500/*
3501 * Fault of a previously existing named mapping. Repopulate the pte
3502 * from the encoded file_pte if possible. This enables swappable
3503 * nonlinear vmas.
3504 *
3505 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3506 * but allow concurrent faults), and pte mapped but not yet locked.
3507 * We return with mmap_sem still held, but pte unmapped and unlocked.
3508 */
3509static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3510                unsigned long address, pte_t *page_table, pmd_t *pmd,
3511                unsigned int flags, pte_t orig_pte)
3512{
3513        pgoff_t pgoff;
3514
3515        flags |= FAULT_FLAG_NONLINEAR;
3516
3517        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3518                return 0;
3519
3520        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3521                /*
3522                 * Page table corrupted: show pte and kill process.
3523                 */
3524                print_bad_pte(vma, address, orig_pte, NULL);
3525                return VM_FAULT_SIGBUS;
3526        }
3527
3528        pgoff = pte_to_pgoff(orig_pte);
3529        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3530}
3531
3532int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3533                                unsigned long addr, int current_nid)
3534{
3535        get_page(page);
3536
3537        count_vm_numa_event(NUMA_HINT_FAULTS);
3538        if (current_nid == numa_node_id())
3539                count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3540
3541        return mpol_misplaced(page, vma, addr);
3542}
3543
3544int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3545                   unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3546{
3547        struct page *page = NULL;
3548        spinlock_t *ptl;
3549        int current_nid = -1;
3550        int target_nid;
3551        bool migrated = false;
3552
3553        /*
3554        * The "pte" at this point cannot be used safely without
3555        * validation through pte_unmap_same(). It's of NUMA type but
3556        * the pfn may be screwed if the read is non atomic.
3557        *
3558        * ptep_modify_prot_start is not called as this is clearing
3559        * the _PAGE_NUMA bit and it is not really expected that there
3560        * would be concurrent hardware modifications to the PTE.
3561        */
3562        ptl = pte_lockptr(mm, pmd);
3563        spin_lock(ptl);
3564        if (unlikely(!pte_same(*ptep, pte))) {
3565                pte_unmap_unlock(ptep, ptl);
3566                goto out;
3567        }
3568
3569        pte = pte_mknonnuma(pte);
3570        set_pte_at(mm, addr, ptep, pte);
3571        update_mmu_cache(vma, addr, ptep);
3572
3573        page = vm_normal_page(vma, addr, pte);
3574        if (!page) {
3575                pte_unmap_unlock(ptep, ptl);
3576                return 0;
3577        }
3578
3579        current_nid = page_to_nid(page);
3580        target_nid = numa_migrate_prep(page, vma, addr, current_nid);
3581        pte_unmap_unlock(ptep, ptl);
3582        if (target_nid == -1) {
3583                /*
3584                 * Account for the fault against the current node if it not
3585                 * being replaced regardless of where the page is located.
3586                 */
3587                current_nid = numa_node_id();
3588                put_page(page);
3589                goto out;
3590        }
3591
3592        /* Migrate to the requested node */
3593        migrated = migrate_misplaced_page(page, target_nid);
3594        if (migrated)
3595                current_nid = target_nid;
3596
3597out:
3598        if (current_nid != -1)
3599                task_numa_fault(current_nid, 1, migrated);
3600        return 0;
3601}
3602
3603/* NUMA hinting page fault entry point for regular pmds */
3604#ifdef CONFIG_NUMA_BALANCING
3605static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3606                     unsigned long addr, pmd_t *pmdp)
3607{
3608        pmd_t pmd;
3609        pte_t *pte, *orig_pte;
3610        unsigned long _addr = addr & PMD_MASK;
3611        unsigned long offset;
3612        spinlock_t *ptl;
3613        bool numa = false;
3614        int local_nid = numa_node_id();
3615
3616        spin_lock(&mm->page_table_lock);
3617        pmd = *pmdp;
3618        if (pmd_numa(pmd)) {
3619                set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3620                numa = true;
3621        }
3622        spin_unlock(&mm->page_table_lock);
3623
3624        if (!numa)
3625                return 0;
3626
3627        /* we're in a page fault so some vma must be in the range */
3628        BUG_ON(!vma);
3629        BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3630        offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3631        VM_BUG_ON(offset >= PMD_SIZE);
3632        orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3633        pte += offset >> PAGE_SHIFT;
3634        for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3635                pte_t pteval = *pte;
3636                struct page *page;
3637                int curr_nid = local_nid;
3638                int target_nid;
3639                bool migrated;
3640                if (!pte_present(pteval))
3641                        continue;
3642                if (!pte_numa(pteval))
3643                        continue;
3644                if (addr >= vma->vm_end) {
3645                        vma = find_vma(mm, addr);
3646                        /* there's a pte present so there must be a vma */
3647                        BUG_ON(!vma);
3648                        BUG_ON(addr < vma->vm_start);
3649                }
3650                if (pte_numa(pteval)) {
3651                        pteval = pte_mknonnuma(pteval);
3652                        set_pte_at(mm, addr, pte, pteval);
3653                }
3654                page = vm_normal_page(vma, addr, pteval);
3655                if (unlikely(!page))
3656                        continue;
3657                /* only check non-shared pages */
3658                if (unlikely(page_mapcount(page) != 1))
3659                        continue;
3660
3661                /*
3662                 * Note that the NUMA fault is later accounted to either
3663                 * the node that is currently running or where the page is
3664                 * migrated to.
3665                 */
3666                curr_nid = local_nid;
3667                target_nid = numa_migrate_prep(page, vma, addr,
3668                                               page_to_nid(page));
3669                if (target_nid == -1) {
3670                        put_page(page);
3671                        continue;
3672                }
3673
3674                /* Migrate to the requested node */
3675                pte_unmap_unlock(pte, ptl);
3676                migrated = migrate_misplaced_page(page, target_nid);
3677                if (migrated)
3678                        curr_nid = target_nid;
3679                task_numa_fault(curr_nid, 1, migrated);
3680
3681                pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3682        }
3683        pte_unmap_unlock(orig_pte, ptl);
3684
3685        return 0;
3686}
3687#else
3688static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3689                     unsigned long addr, pmd_t *pmdp)
3690{
3691        BUG();
3692        return 0;
3693}
3694#endif /* CONFIG_NUMA_BALANCING */
3695
3696/*
3697 * These routines also need to handle stuff like marking pages dirty
3698 * and/or accessed for architectures that don't do it in hardware (most
3699 * RISC architectures).  The early dirtying is also good on the i386.
3700 *
3701 * There is also a hook called "update_mmu_cache()" that architectures
3702 * with external mmu caches can use to update those (ie the Sparc or
3703 * PowerPC hashed page tables that act as extended TLBs).
3704 *
3705 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3706 * but allow concurrent faults), and pte mapped but not yet locked.
3707 * We return with mmap_sem still held, but pte unmapped and unlocked.
3708 */
3709int handle_pte_fault(struct mm_struct *mm,
3710                     struct vm_area_struct *vma, unsigned long address,
3711                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3712{
3713        pte_t entry;
3714        spinlock_t *ptl;
3715
3716        entry = *pte;
3717        if (!pte_present(entry)) {
3718                if (pte_none(entry)) {
3719                        if (vma->vm_ops) {
3720                                if (likely(vma->vm_ops->fault))
3721                                        return do_linear_fault(mm, vma, address,
3722                                                pte, pmd, flags, entry);
3723                        }
3724                        return do_anonymous_page(mm, vma, address,
3725                                                 pte, pmd, flags);
3726                }
3727                if (pte_file(entry))
3728                        return do_nonlinear_fault(mm, vma, address,
3729                                        pte, pmd, flags, entry);
3730                return do_swap_page(mm, vma, address,
3731                                        pte, pmd, flags, entry);
3732        }
3733
3734        if (pte_numa(entry))
3735                return do_numa_page(mm, vma, address, entry, pte, pmd);
3736
3737        ptl = pte_lockptr(mm, pmd);
3738        spin_lock(ptl);
3739        if (unlikely(!pte_same(*pte, entry)))
3740                goto unlock;
3741        if (flags & FAULT_FLAG_WRITE) {
3742                if (!pte_write(entry))
3743                        return do_wp_page(mm, vma, address,
3744                                        pte, pmd, ptl, entry);
3745                entry = pte_mkdirty(entry);
3746        }
3747        entry = pte_mkyoung(entry);
3748        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3749                update_mmu_cache(vma, address, pte);
3750        } else {
3751                /*
3752                 * This is needed only for protection faults but the arch code
3753                 * is not yet telling us if this is a protection fault or not.
3754                 * This still avoids useless tlb flushes for .text page faults
3755                 * with threads.
3756                 */
3757                if (flags & FAULT_FLAG_WRITE)
3758                        flush_tlb_fix_spurious_fault(vma, address);
3759        }
3760unlock:
3761        pte_unmap_unlock(pte, ptl);
3762        return 0;
3763}
3764
3765/*
3766 * By the time we get here, we already hold the mm semaphore
3767 */
3768int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3769                unsigned long address, unsigned int flags)
3770{
3771        pgd_t *pgd;
3772        pud_t *pud;
3773        pmd_t *pmd;
3774        pte_t *pte;
3775
3776        __set_current_state(TASK_RUNNING);
3777
3778        count_vm_event(PGFAULT);
3779        mem_cgroup_count_vm_event(mm, PGFAULT);
3780
3781        /* do counter updates before entering really critical section. */
3782        check_sync_rss_stat(current);
3783
3784        if (unlikely(is_vm_hugetlb_page(vma)))
3785                return hugetlb_fault(mm, vma, address, flags);
3786
3787retry:
3788        pgd = pgd_offset(mm, address);
3789        pud = pud_alloc(mm, pgd, address);
3790        if (!pud)
3791                return VM_FAULT_OOM;
3792        pmd = pmd_alloc(mm, pud, address);
3793        if (!pmd)
3794                return VM_FAULT_OOM;
3795        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3796                if (!vma->vm_ops)
3797                        return do_huge_pmd_anonymous_page(mm, vma, address,
3798                                                          pmd, flags);
3799        } else {
3800                pmd_t orig_pmd = *pmd;
3801                int ret;
3802
3803                barrier();
3804                if (pmd_trans_huge(orig_pmd)) {
3805                        unsigned int dirty = flags & FAULT_FLAG_WRITE;
3806
3807                        /*
3808                         * If the pmd is splitting, return and retry the
3809                         * the fault.  Alternative: wait until the split
3810                         * is done, and goto retry.
3811                         */
3812                        if (pmd_trans_splitting(orig_pmd))
3813                                return 0;
3814
3815                        if (pmd_numa(orig_pmd))
3816                                return do_huge_pmd_numa_page(mm, vma, address,
3817                                                             orig_pmd, pmd);
3818
3819                        if (dirty && !pmd_write(orig_pmd)) {
3820                                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3821                                                          orig_pmd);
3822                                /*
3823                                 * If COW results in an oom, the huge pmd will
3824                                 * have been split, so retry the fault on the
3825                                 * pte for a smaller charge.
3826                                 */
3827                                if (unlikely(ret & VM_FAULT_OOM))
3828                                        goto retry;
3829                                return ret;
3830                        } else {
3831                                huge_pmd_set_accessed(mm, vma, address, pmd,
3832                                                      orig_pmd, dirty);
3833                        }
3834
3835                        return 0;
3836                }
3837        }
3838
3839        if (pmd_numa(*pmd))
3840                return do_pmd_numa_page(mm, vma, address, pmd);
3841
3842        /*
3843         * Use __pte_alloc instead of pte_alloc_map, because we can't
3844         * run pte_offset_map on the pmd, if an huge pmd could
3845         * materialize from under us from a different thread.
3846         */
3847        if (unlikely(pmd_none(*pmd)) &&
3848            unlikely(__pte_alloc(mm, vma, pmd, address)))
3849                return VM_FAULT_OOM;
3850        /* if an huge pmd materialized from under us just retry later */
3851        if (unlikely(pmd_trans_huge(*pmd)))
3852                return 0;
3853        /*
3854         * A regular pmd is established and it can't morph into a huge pmd
3855         * from under us anymore at this point because we hold the mmap_sem
3856         * read mode and khugepaged takes it in write mode. So now it's
3857         * safe to run pte_offset_map().
3858         */
3859        pte = pte_offset_map(pmd, address);
3860
3861        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3862}
3863
3864#ifndef __PAGETABLE_PUD_FOLDED
3865/*
3866 * Allocate page upper directory.
3867 * We've already handled the fast-path in-line.
3868 */
3869int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3870{
3871        pud_t *new = pud_alloc_one(mm, address);
3872        if (!new)
3873                return -ENOMEM;
3874
3875        smp_wmb(); /* See comment in __pte_alloc */
3876
3877        spin_lock(&mm->page_table_lock);
3878        if (pgd_present(*pgd))          /* Another has populated it */
3879                pud_free(mm, new);
3880        else
3881                pgd_populate(mm, pgd, new);
3882        spin_unlock(&mm->page_table_lock);
3883        return 0;
3884}
3885#endif /* __PAGETABLE_PUD_FOLDED */
3886
3887#ifndef __PAGETABLE_PMD_FOLDED
3888/*
3889 * Allocate page middle directory.
3890 * We've already handled the fast-path in-line.
3891 */
3892int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3893{
3894        pmd_t *new = pmd_alloc_one(mm, address);
3895        if (!new)
3896                return -ENOMEM;
3897
3898        smp_wmb(); /* See comment in __pte_alloc */
3899
3900        spin_lock(&mm->page_table_lock);
3901#ifndef __ARCH_HAS_4LEVEL_HACK
3902        if (pud_present(*pud))          /* Another has populated it */
3903                pmd_free(mm, new);
3904        else
3905                pud_populate(mm, pud, new);
3906#else
3907        if (pgd_present(*pud))          /* Another has populated it */
3908                pmd_free(mm, new);
3909        else
3910                pgd_populate(mm, pud, new);
3911#endif /* __ARCH_HAS_4LEVEL_HACK */
3912        spin_unlock(&mm->page_table_lock);
3913        return 0;
3914}
3915#endif /* __PAGETABLE_PMD_FOLDED */
3916
3917#if !defined(__HAVE_ARCH_GATE_AREA)
3918
3919#if defined(AT_SYSINFO_EHDR)
3920static struct vm_area_struct gate_vma;
3921
3922static int __init gate_vma_init(void)
3923{
3924        gate_vma.vm_mm = NULL;
3925        gate_vma.vm_start = FIXADDR_USER_START;
3926        gate_vma.vm_end = FIXADDR_USER_END;
3927        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3928        gate_vma.vm_page_prot = __P101;
3929
3930        return 0;
3931}
3932__initcall(gate_vma_init);
3933#endif
3934
3935struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3936{
3937#ifdef AT_SYSINFO_EHDR
3938        return &gate_vma;
3939#else
3940        return NULL;
3941#endif
3942}
3943
3944int in_gate_area_no_mm(unsigned long addr)
3945{
3946#ifdef AT_SYSINFO_EHDR
3947        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3948                return 1;
3949#endif
3950        return 0;
3951}
3952
3953#endif  /* __HAVE_ARCH_GATE_AREA */
3954
3955static int __follow_pte(struct mm_struct *mm, unsigned long address,
3956                pte_t **ptepp, spinlock_t **ptlp)
3957{
3958        pgd_t *pgd;
3959        pud_t *pud;
3960        pmd_t *pmd;
3961        pte_t *ptep;
3962
3963        pgd = pgd_offset(mm, address);
3964        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3965                goto out;
3966
3967        pud = pud_offset(pgd, address);
3968        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3969                goto out;
3970
3971        pmd = pmd_offset(pud, address);
3972        VM_BUG_ON(pmd_trans_huge(*pmd));
3973        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3974                goto out;
3975
3976        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3977        if (pmd_huge(*pmd))
3978                goto out;
3979
3980        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3981        if (!ptep)
3982                goto out;
3983        if (!pte_present(*ptep))
3984                goto unlock;
3985        *ptepp = ptep;
3986        return 0;
3987unlock:
3988        pte_unmap_unlock(ptep, *ptlp);
3989out:
3990        return -EINVAL;
3991}
3992
3993static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3994                             pte_t **ptepp, spinlock_t **ptlp)
3995{
3996        int res;
3997
3998        /* (void) is needed to make gcc happy */
3999        (void) __cond_lock(*ptlp,
4000                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
4001        return res;
4002}
4003
4004/**
4005 * follow_pfn - look up PFN at a user virtual address
4006 * @vma: memory mapping
4007 * @address: user virtual address
4008 * @pfn: location to store found PFN
4009 *
4010 * Only IO mappings and raw PFN mappings are allowed.
4011 *
4012 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4013 */
4014int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4015        unsigned long *pfn)
4016{
4017        int ret = -EINVAL;
4018        spinlock_t *ptl;
4019        pte_t *ptep;
4020
4021        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4022                return ret;
4023
4024        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4025        if (ret)
4026                return ret;
4027        *pfn = pte_pfn(*ptep);
4028        pte_unmap_unlock(ptep, ptl);
4029        return 0;
4030}
4031EXPORT_SYMBOL(follow_pfn);
4032
4033#ifdef CONFIG_HAVE_IOREMAP_PROT
4034int follow_phys(struct vm_area_struct *vma,
4035                unsigned long address, unsigned int flags,
4036                unsigned long *prot, resource_size_t *phys)
4037{
4038        int ret = -EINVAL;
4039        pte_t *ptep, pte;
4040        spinlock_t *ptl;
4041
4042        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4043                goto out;
4044
4045        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4046                goto out;
4047        pte = *ptep;
4048
4049        if ((flags & FOLL_WRITE) && !pte_write(pte))
4050                goto unlock;
4051
4052        *prot = pgprot_val(pte_pgprot(pte));
4053        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4054
4055        ret = 0;
4056unlock:
4057        pte_unmap_unlock(ptep, ptl);
4058out:
4059        return ret;
4060}
4061
4062int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4063                        void *buf, int len, int write)
4064{
4065        resource_size_t phys_addr;
4066        unsigned long prot = 0;
4067        void __iomem *maddr;
4068        int offset = addr & (PAGE_SIZE-1);
4069
4070        if (follow_phys(vma, addr, write, &prot, &phys_addr))
4071                return -EINVAL;
4072
4073        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4074        if (write)
4075                memcpy_toio(maddr + offset, buf, len);
4076        else
4077                memcpy_fromio(buf, maddr + offset, len);
4078        iounmap(maddr);
4079
4080        return len;
4081}
4082#endif
4083
4084/*
4085 * Access another process' address space as given in mm.  If non-NULL, use the
4086 * given task for page fault accounting.
4087 */
4088static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4089                unsigned long addr, void *buf, int len, int write)
4090{
4091        struct vm_area_struct *vma;
4092        void *old_buf = buf;
4093
4094        down_read(&mm->mmap_sem);
4095        /* ignore errors, just check how much was successfully transferred */
4096        while (len) {
4097                int bytes, ret, offset;
4098                void *maddr;
4099                struct page *page = NULL;
4100
4101                ret = get_user_pages(tsk, mm, addr, 1,
4102                                write, 1, &page, &vma);
4103                if (ret <= 0) {
4104                        /*
4105                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
4106                         * we can access using slightly different code.
4107                         */
4108#ifdef CONFIG_HAVE_IOREMAP_PROT
4109                        vma = find_vma(mm, addr);
4110                        if (!vma || vma->vm_start > addr)
4111                                break;
4112                        if (vma->vm_ops && vma->vm_ops->access)
4113                                ret = vma->vm_ops->access(vma, addr, buf,
4114                                                          len, write);
4115                        if (ret <= 0)
4116#endif
4117                                break;
4118                        bytes = ret;
4119                } else {
4120                        bytes = len;
4121                        offset = addr & (PAGE_SIZE-1);
4122                        if (bytes > PAGE_SIZE-offset)
4123                                bytes = PAGE_SIZE-offset;
4124
4125                        maddr = kmap(page);
4126                        if (write) {
4127                                copy_to_user_page(vma, page, addr,
4128                                                  maddr + offset, buf, bytes);
4129                                set_page_dirty_lock(page);
4130                        } else {
4131                                copy_from_user_page(vma, page, addr,
4132                                                    buf, maddr + offset, bytes);
4133                        }
4134                        kunmap(page);
4135                        page_cache_release(page);
4136                }
4137                len -= bytes;
4138                buf += bytes;
4139                addr += bytes;
4140        }
4141        up_read(&mm->mmap_sem);
4142
4143        return buf - old_buf;
4144}
4145
4146/**
4147 * access_remote_vm - access another process' address space
4148 * @mm:         the mm_struct of the target address space
4149 * @addr:       start address to access
4150 * @buf:        source or destination buffer
4151 * @len:        number of bytes to transfer
4152 * @write:      whether the access is a write
4153 *
4154 * The caller must hold a reference on @mm.
4155 */
4156int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4157                void *buf, int len, int write)
4158{
4159        return __access_remote_vm(NULL, mm, addr, buf, len, write);
4160}
4161
4162/*
4163 * Access another process' address space.
4164 * Source/target buffer must be kernel space,
4165 * Do not walk the page table directly, use get_user_pages
4166 */
4167int access_process_vm(struct task_struct *tsk, unsigned long addr,
4168                void *buf, int len, int write)
4169{
4170        struct mm_struct *mm;
4171        int ret;
4172
4173        mm = get_task_mm(tsk);
4174        if (!mm)
4175                return 0;
4176
4177        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4178        mmput(mm);
4179
4180        return ret;
4181}
4182
4183/*
4184 * Print the name of a VMA.
4185 */
4186void print_vma_addr(char *prefix, unsigned long ip)
4187{
4188        struct mm_struct *mm = current->mm;
4189        struct vm_area_struct *vma;
4190
4191        /*
4192         * Do not print if we are in atomic
4193         * contexts (in exception stacks, etc.):
4194         */
4195        if (preempt_count())
4196                return;
4197
4198        down_read(&mm->mmap_sem);
4199        vma = find_vma(mm, ip);
4200        if (vma && vma->vm_file) {
4201                struct file *f = vma->vm_file;
4202                char *buf = (char *)__get_free_page(GFP_KERNEL);
4203                if (buf) {
4204                        char *p;
4205
4206                        p = d_path(&f->f_path, buf, PAGE_SIZE);
4207                        if (IS_ERR(p))
4208                                p = "?";
4209                        printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4210                                        vma->vm_start,
4211                                        vma->vm_end - vma->vm_start);
4212                        free_page((unsigned long)buf);
4213                }
4214        }
4215        up_read(&mm->mmap_sem);
4216}
4217
4218#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4219void might_fault(void)
4220{
4221        /*
4222         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4223         * holding the mmap_sem, this is safe because kernel memory doesn't
4224         * get paged out, therefore we'll never actually fault, and the
4225         * below annotations will generate false positives.
4226         */
4227        if (segment_eq(get_fs(), KERNEL_DS))
4228                return;
4229
4230        /*
4231         * it would be nicer only to annotate paths which are not under
4232         * pagefault_disable, however that requires a larger audit and
4233         * providing helpers like get_user_atomic.
4234         */
4235        if (in_atomic())
4236                return;
4237
4238        __might_sleep(__FILE__, __LINE__, 0);
4239
4240        if (current->mm)
4241                might_lock_read(&current->mm->mmap_sem);
4242}
4243EXPORT_SYMBOL(might_fault);
4244#endif
4245
4246#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4247static void clear_gigantic_page(struct page *page,
4248                                unsigned long addr,
4249                                unsigned int pages_per_huge_page)
4250{
4251        int i;
4252        struct page *p = page;
4253
4254        might_sleep();
4255        for (i = 0; i < pages_per_huge_page;
4256             i++, p = mem_map_next(p, page, i)) {
4257                cond_resched();
4258                clear_user_highpage(p, addr + i * PAGE_SIZE);
4259        }
4260}
4261void clear_huge_page(struct page *page,
4262                     unsigned long addr, unsigned int pages_per_huge_page)
4263{
4264        int i;
4265
4266        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4267                clear_gigantic_page(page, addr, pages_per_huge_page);
4268                return;
4269        }
4270
4271        might_sleep();
4272        for (i = 0; i < pages_per_huge_page; i++) {
4273                cond_resched();
4274                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4275        }
4276}
4277
4278static void copy_user_gigantic_page(struct page *dst, struct page *src,
4279                                    unsigned long addr,
4280                                    struct vm_area_struct *vma,
4281                                    unsigned int pages_per_huge_page)
4282{
4283        int i;
4284        struct page *dst_base = dst;
4285        struct page *src_base = src;
4286
4287        for (i = 0; i < pages_per_huge_page; ) {
4288                cond_resched();
4289                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4290
4291                i++;
4292                dst = mem_map_next(dst, dst_base, i);
4293                src = mem_map_next(src, src_base, i);
4294        }
4295}
4296
4297void copy_user_huge_page(struct page *dst, struct page *src,
4298                         unsigned long addr, struct vm_area_struct *vma,
4299                         unsigned int pages_per_huge_page)
4300{
4301        int i;
4302
4303        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4304                copy_user_gigantic_page(dst, src, addr, vma,
4305                                        pages_per_huge_page);
4306                return;
4307        }
4308
4309        might_sleep();
4310        for (i = 0; i < pages_per_huge_page; i++) {
4311                cond_resched();
4312                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4313        }
4314}
4315#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4316
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.