linux/mm/memory.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/memory.c
   3 *
   4 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   5 */
   6
   7/*
   8 * demand-loading started 01.12.91 - seems it is high on the list of
   9 * things wanted, and it should be easy to implement. - Linus
  10 */
  11
  12/*
  13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  14 * pages started 02.12.91, seems to work. - Linus.
  15 *
  16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  17 * would have taken more than the 6M I have free, but it worked well as
  18 * far as I could see.
  19 *
  20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  21 */
  22
  23/*
  24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  25 * thought has to go into this. Oh, well..
  26 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  27 *              Found it. Everything seems to work now.
  28 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  29 */
  30
  31/*
  32 * 05.04.94  -  Multi-page memory management added for v1.1.
  33 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  34 *
  35 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  36 *              (Gerhard.Wichert@pdb.siemens.de)
  37 *
  38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  39 */
  40
  41#include <linux/kernel_stat.h>
  42#include <linux/mm.h>
  43#include <linux/hugetlb.h>
  44#include <linux/mman.h>
  45#include <linux/swap.h>
  46#include <linux/highmem.h>
  47#include <linux/pagemap.h>
  48#include <linux/ksm.h>
  49#include <linux/rmap.h>
  50#include <linux/export.h>
  51#include <linux/delayacct.h>
  52#include <linux/init.h>
  53#include <linux/writeback.h>
  54#include <linux/memcontrol.h>
  55#include <linux/mmu_notifier.h>
  56#include <linux/kallsyms.h>
  57#include <linux/swapops.h>
  58#include <linux/elf.h>
  59#include <linux/gfp.h>
  60
  61#include <asm/io.h>
  62#include <asm/pgalloc.h>
  63#include <asm/uaccess.h>
  64#include <asm/tlb.h>
  65#include <asm/tlbflush.h>
  66#include <asm/pgtable.h>
  67
  68#include "internal.h"
  69
  70#ifndef CONFIG_NEED_MULTIPLE_NODES
  71/* use the per-pgdat data instead for discontigmem - mbligh */
  72unsigned long max_mapnr;
  73struct page *mem_map;
  74
  75EXPORT_SYMBOL(max_mapnr);
  76EXPORT_SYMBOL(mem_map);
  77#endif
  78
  79unsigned long num_physpages;
  80/*
  81 * A number of key systems in x86 including ioremap() rely on the assumption
  82 * that high_memory defines the upper bound on direct map memory, then end
  83 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
  84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
  85 * and ZONE_HIGHMEM.
  86 */
  87void * high_memory;
  88
  89EXPORT_SYMBOL(num_physpages);
  90EXPORT_SYMBOL(high_memory);
  91
  92/*
  93 * Randomize the address space (stacks, mmaps, brk, etc.).
  94 *
  95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
  96 *   as ancient (libc5 based) binaries can segfault. )
  97 */
  98int randomize_va_space __read_mostly =
  99#ifdef CONFIG_COMPAT_BRK
 100                                        1;
 101#else
 102                                        2;
 103#endif
 104
 105static int __init disable_randmaps(char *s)
 106{
 107        randomize_va_space = 0;
 108        return 1;
 109}
 110__setup("norandmaps", disable_randmaps);
 111
 112unsigned long zero_pfn __read_mostly;
 113unsigned long highest_memmap_pfn __read_mostly;
 114
 115/*
 116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 117 */
 118static int __init init_zero_pfn(void)
 119{
 120        zero_pfn = page_to_pfn(ZERO_PAGE(0));
 121        return 0;
 122}
 123core_initcall(init_zero_pfn);
 124
 125
 126#if defined(SPLIT_RSS_COUNTING)
 127
 128void sync_mm_rss(struct mm_struct *mm)
 129{
 130        int i;
 131
 132        for (i = 0; i < NR_MM_COUNTERS; i++) {
 133                if (current->rss_stat.count[i]) {
 134                        add_mm_counter(mm, i, current->rss_stat.count[i]);
 135                        current->rss_stat.count[i] = 0;
 136                }
 137        }
 138        current->rss_stat.events = 0;
 139}
 140
 141static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 142{
 143        struct task_struct *task = current;
 144
 145        if (likely(task->mm == mm))
 146                task->rss_stat.count[member] += val;
 147        else
 148                add_mm_counter(mm, member, val);
 149}
 150#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 151#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 152
 153/* sync counter once per 64 page faults */
 154#define TASK_RSS_EVENTS_THRESH  (64)
 155static void check_sync_rss_stat(struct task_struct *task)
 156{
 157        if (unlikely(task != current))
 158                return;
 159        if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 160                sync_mm_rss(task->mm);
 161}
 162#else /* SPLIT_RSS_COUNTING */
 163
 164#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 165#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 166
 167static void check_sync_rss_stat(struct task_struct *task)
 168{
 169}
 170
 171#endif /* SPLIT_RSS_COUNTING */
 172
 173#ifdef HAVE_GENERIC_MMU_GATHER
 174
 175static int tlb_next_batch(struct mmu_gather *tlb)
 176{
 177        struct mmu_gather_batch *batch;
 178
 179        batch = tlb->active;
 180        if (batch->next) {
 181                tlb->active = batch->next;
 182                return 1;
 183        }
 184
 185        batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
 186        if (!batch)
 187                return 0;
 188
 189        batch->next = NULL;
 190        batch->nr   = 0;
 191        batch->max  = MAX_GATHER_BATCH;
 192
 193        tlb->active->next = batch;
 194        tlb->active = batch;
 195
 196        return 1;
 197}
 198
 199/* tlb_gather_mmu
 200 *      Called to initialize an (on-stack) mmu_gather structure for page-table
 201 *      tear-down from @mm. The @fullmm argument is used when @mm is without
 202 *      users and we're going to destroy the full address space (exit/execve).
 203 */
 204void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
 205{
 206        tlb->mm = mm;
 207
 208        tlb->fullmm     = fullmm;
 209        tlb->need_flush = 0;
 210        tlb->fast_mode  = (num_possible_cpus() == 1);
 211        tlb->local.next = NULL;
 212        tlb->local.nr   = 0;
 213        tlb->local.max  = ARRAY_SIZE(tlb->__pages);
 214        tlb->active     = &tlb->local;
 215
 216#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 217        tlb->batch = NULL;
 218#endif
 219}
 220
 221void tlb_flush_mmu(struct mmu_gather *tlb)
 222{
 223        struct mmu_gather_batch *batch;
 224
 225        if (!tlb->need_flush)
 226                return;
 227        tlb->need_flush = 0;
 228        tlb_flush(tlb);
 229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 230        tlb_table_flush(tlb);
 231#endif
 232
 233        if (tlb_fast_mode(tlb))
 234                return;
 235
 236        for (batch = &tlb->local; batch; batch = batch->next) {
 237                free_pages_and_swap_cache(batch->pages, batch->nr);
 238                batch->nr = 0;
 239        }
 240        tlb->active = &tlb->local;
 241}
 242
 243/* tlb_finish_mmu
 244 *      Called at the end of the shootdown operation to free up any resources
 245 *      that were required.
 246 */
 247void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
 248{
 249        struct mmu_gather_batch *batch, *next;
 250
 251        tlb_flush_mmu(tlb);
 252
 253        /* keep the page table cache within bounds */
 254        check_pgt_cache();
 255
 256        for (batch = tlb->local.next; batch; batch = next) {
 257                next = batch->next;
 258                free_pages((unsigned long)batch, 0);
 259        }
 260        tlb->local.next = NULL;
 261}
 262
 263/* __tlb_remove_page
 264 *      Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
 265 *      handling the additional races in SMP caused by other CPUs caching valid
 266 *      mappings in their TLBs. Returns the number of free page slots left.
 267 *      When out of page slots we must call tlb_flush_mmu().
 268 */
 269int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
 270{
 271        struct mmu_gather_batch *batch;
 272
 273        VM_BUG_ON(!tlb->need_flush);
 274
 275        if (tlb_fast_mode(tlb)) {
 276                free_page_and_swap_cache(page);
 277                return 1; /* avoid calling tlb_flush_mmu() */
 278        }
 279
 280        batch = tlb->active;
 281        batch->pages[batch->nr++] = page;
 282        if (batch->nr == batch->max) {
 283                if (!tlb_next_batch(tlb))
 284                        return 0;
 285                batch = tlb->active;
 286        }
 287        VM_BUG_ON(batch->nr > batch->max);
 288
 289        return batch->max - batch->nr;
 290}
 291
 292#endif /* HAVE_GENERIC_MMU_GATHER */
 293
 294#ifdef CONFIG_HAVE_RCU_TABLE_FREE
 295
 296/*
 297 * See the comment near struct mmu_table_batch.
 298 */
 299
 300static void tlb_remove_table_smp_sync(void *arg)
 301{
 302        /* Simply deliver the interrupt */
 303}
 304
 305static void tlb_remove_table_one(void *table)
 306{
 307        /*
 308         * This isn't an RCU grace period and hence the page-tables cannot be
 309         * assumed to be actually RCU-freed.
 310         *
 311         * It is however sufficient for software page-table walkers that rely on
 312         * IRQ disabling. See the comment near struct mmu_table_batch.
 313         */
 314        smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 315        __tlb_remove_table(table);
 316}
 317
 318static void tlb_remove_table_rcu(struct rcu_head *head)
 319{
 320        struct mmu_table_batch *batch;
 321        int i;
 322
 323        batch = container_of(head, struct mmu_table_batch, rcu);
 324
 325        for (i = 0; i < batch->nr; i++)
 326                __tlb_remove_table(batch->tables[i]);
 327
 328        free_page((unsigned long)batch);
 329}
 330
 331void tlb_table_flush(struct mmu_gather *tlb)
 332{
 333        struct mmu_table_batch **batch = &tlb->batch;
 334
 335        if (*batch) {
 336                call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 337                *batch = NULL;
 338        }
 339}
 340
 341void tlb_remove_table(struct mmu_gather *tlb, void *table)
 342{
 343        struct mmu_table_batch **batch = &tlb->batch;
 344
 345        tlb->need_flush = 1;
 346
 347        /*
 348         * When there's less then two users of this mm there cannot be a
 349         * concurrent page-table walk.
 350         */
 351        if (atomic_read(&tlb->mm->mm_users) < 2) {
 352                __tlb_remove_table(table);
 353                return;
 354        }
 355
 356        if (*batch == NULL) {
 357                *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 358                if (*batch == NULL) {
 359                        tlb_remove_table_one(table);
 360                        return;
 361                }
 362                (*batch)->nr = 0;
 363        }
 364        (*batch)->tables[(*batch)->nr++] = table;
 365        if ((*batch)->nr == MAX_TABLE_BATCH)
 366                tlb_table_flush(tlb);
 367}
 368
 369#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
 370
 371/*
 372 * If a p?d_bad entry is found while walking page tables, report
 373 * the error, before resetting entry to p?d_none.  Usually (but
 374 * very seldom) called out from the p?d_none_or_clear_bad macros.
 375 */
 376
 377void pgd_clear_bad(pgd_t *pgd)
 378{
 379        pgd_ERROR(*pgd);
 380        pgd_clear(pgd);
 381}
 382
 383void pud_clear_bad(pud_t *pud)
 384{
 385        pud_ERROR(*pud);
 386        pud_clear(pud);
 387}
 388
 389void pmd_clear_bad(pmd_t *pmd)
 390{
 391        pmd_ERROR(*pmd);
 392        pmd_clear(pmd);
 393}
 394
 395/*
 396 * Note: this doesn't free the actual pages themselves. That
 397 * has been handled earlier when unmapping all the memory regions.
 398 */
 399static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 400                           unsigned long addr)
 401{
 402        pgtable_t token = pmd_pgtable(*pmd);
 403        pmd_clear(pmd);
 404        pte_free_tlb(tlb, token, addr);
 405        tlb->mm->nr_ptes--;
 406}
 407
 408static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 409                                unsigned long addr, unsigned long end,
 410                                unsigned long floor, unsigned long ceiling)
 411{
 412        pmd_t *pmd;
 413        unsigned long next;
 414        unsigned long start;
 415
 416        start = addr;
 417        pmd = pmd_offset(pud, addr);
 418        do {
 419                next = pmd_addr_end(addr, end);
 420                if (pmd_none_or_clear_bad(pmd))
 421                        continue;
 422                free_pte_range(tlb, pmd, addr);
 423        } while (pmd++, addr = next, addr != end);
 424
 425        start &= PUD_MASK;
 426        if (start < floor)
 427                return;
 428        if (ceiling) {
 429                ceiling &= PUD_MASK;
 430                if (!ceiling)
 431                        return;
 432        }
 433        if (end - 1 > ceiling - 1)
 434                return;
 435
 436        pmd = pmd_offset(pud, start);
 437        pud_clear(pud);
 438        pmd_free_tlb(tlb, pmd, start);
 439}
 440
 441static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
 442                                unsigned long addr, unsigned long end,
 443                                unsigned long floor, unsigned long ceiling)
 444{
 445        pud_t *pud;
 446        unsigned long next;
 447        unsigned long start;
 448
 449        start = addr;
 450        pud = pud_offset(pgd, addr);
 451        do {
 452                next = pud_addr_end(addr, end);
 453                if (pud_none_or_clear_bad(pud))
 454                        continue;
 455                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 456        } while (pud++, addr = next, addr != end);
 457
 458        start &= PGDIR_MASK;
 459        if (start < floor)
 460                return;
 461        if (ceiling) {
 462                ceiling &= PGDIR_MASK;
 463                if (!ceiling)
 464                        return;
 465        }
 466        if (end - 1 > ceiling - 1)
 467                return;
 468
 469        pud = pud_offset(pgd, start);
 470        pgd_clear(pgd);
 471        pud_free_tlb(tlb, pud, start);
 472}
 473
 474/*
 475 * This function frees user-level page tables of a process.
 476 *
 477 * Must be called with pagetable lock held.
 478 */
 479void free_pgd_range(struct mmu_gather *tlb,
 480                        unsigned long addr, unsigned long end,
 481                        unsigned long floor, unsigned long ceiling)
 482{
 483        pgd_t *pgd;
 484        unsigned long next;
 485
 486        /*
 487         * The next few lines have given us lots of grief...
 488         *
 489         * Why are we testing PMD* at this top level?  Because often
 490         * there will be no work to do at all, and we'd prefer not to
 491         * go all the way down to the bottom just to discover that.
 492         *
 493         * Why all these "- 1"s?  Because 0 represents both the bottom
 494         * of the address space and the top of it (using -1 for the
 495         * top wouldn't help much: the masks would do the wrong thing).
 496         * The rule is that addr 0 and floor 0 refer to the bottom of
 497         * the address space, but end 0 and ceiling 0 refer to the top
 498         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 499         * that end 0 case should be mythical).
 500         *
 501         * Wherever addr is brought up or ceiling brought down, we must
 502         * be careful to reject "the opposite 0" before it confuses the
 503         * subsequent tests.  But what about where end is brought down
 504         * by PMD_SIZE below? no, end can't go down to 0 there.
 505         *
 506         * Whereas we round start (addr) and ceiling down, by different
 507         * masks at different levels, in order to test whether a table
 508         * now has no other vmas using it, so can be freed, we don't
 509         * bother to round floor or end up - the tests don't need that.
 510         */
 511
 512        addr &= PMD_MASK;
 513        if (addr < floor) {
 514                addr += PMD_SIZE;
 515                if (!addr)
 516                        return;
 517        }
 518        if (ceiling) {
 519                ceiling &= PMD_MASK;
 520                if (!ceiling)
 521                        return;
 522        }
 523        if (end - 1 > ceiling - 1)
 524                end -= PMD_SIZE;
 525        if (addr > end - 1)
 526                return;
 527
 528        pgd = pgd_offset(tlb->mm, addr);
 529        do {
 530                next = pgd_addr_end(addr, end);
 531                if (pgd_none_or_clear_bad(pgd))
 532                        continue;
 533                free_pud_range(tlb, pgd, addr, next, floor, ceiling);
 534        } while (pgd++, addr = next, addr != end);
 535}
 536
 537void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 538                unsigned long floor, unsigned long ceiling)
 539{
 540        while (vma) {
 541                struct vm_area_struct *next = vma->vm_next;
 542                unsigned long addr = vma->vm_start;
 543
 544                /*
 545                 * Hide vma from rmap and truncate_pagecache before freeing
 546                 * pgtables
 547                 */
 548                unlink_anon_vmas(vma);
 549                unlink_file_vma(vma);
 550
 551                if (is_vm_hugetlb_page(vma)) {
 552                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 553                                floor, next? next->vm_start: ceiling);
 554                } else {
 555                        /*
 556                         * Optimization: gather nearby vmas into one call down
 557                         */
 558                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 559                               && !is_vm_hugetlb_page(next)) {
 560                                vma = next;
 561                                next = vma->vm_next;
 562                                unlink_anon_vmas(vma);
 563                                unlink_file_vma(vma);
 564                        }
 565                        free_pgd_range(tlb, addr, vma->vm_end,
 566                                floor, next? next->vm_start: ceiling);
 567                }
 568                vma = next;
 569        }
 570}
 571
 572int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
 573                pmd_t *pmd, unsigned long address)
 574{
 575        pgtable_t new = pte_alloc_one(mm, address);
 576        int wait_split_huge_page;
 577        if (!new)
 578                return -ENOMEM;
 579
 580        /*
 581         * Ensure all pte setup (eg. pte page lock and page clearing) are
 582         * visible before the pte is made visible to other CPUs by being
 583         * put into page tables.
 584         *
 585         * The other side of the story is the pointer chasing in the page
 586         * table walking code (when walking the page table without locking;
 587         * ie. most of the time). Fortunately, these data accesses consist
 588         * of a chain of data-dependent loads, meaning most CPUs (alpha
 589         * being the notable exception) will already guarantee loads are
 590         * seen in-order. See the alpha page table accessors for the
 591         * smp_read_barrier_depends() barriers in page table walking code.
 592         */
 593        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 594
 595        spin_lock(&mm->page_table_lock);
 596        wait_split_huge_page = 0;
 597        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 598                mm->nr_ptes++;
 599                pmd_populate(mm, pmd, new);
 600                new = NULL;
 601        } else if (unlikely(pmd_trans_splitting(*pmd)))
 602                wait_split_huge_page = 1;
 603        spin_unlock(&mm->page_table_lock);
 604        if (new)
 605                pte_free(mm, new);
 606        if (wait_split_huge_page)
 607                wait_split_huge_page(vma->anon_vma, pmd);
 608        return 0;
 609}
 610
 611int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
 612{
 613        pte_t *new = pte_alloc_one_kernel(&init_mm, address);
 614        if (!new)
 615                return -ENOMEM;
 616
 617        smp_wmb(); /* See comment in __pte_alloc */
 618
 619        spin_lock(&init_mm.page_table_lock);
 620        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 621                pmd_populate_kernel(&init_mm, pmd, new);
 622                new = NULL;
 623        } else
 624                VM_BUG_ON(pmd_trans_splitting(*pmd));
 625        spin_unlock(&init_mm.page_table_lock);
 626        if (new)
 627                pte_free_kernel(&init_mm, new);
 628        return 0;
 629}
 630
 631static inline void init_rss_vec(int *rss)
 632{
 633        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 634}
 635
 636static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 637{
 638        int i;
 639
 640        if (current->mm == mm)
 641                sync_mm_rss(mm);
 642        for (i = 0; i < NR_MM_COUNTERS; i++)
 643                if (rss[i])
 644                        add_mm_counter(mm, i, rss[i]);
 645}
 646
 647/*
 648 * This function is called to print an error when a bad pte
 649 * is found. For example, we might have a PFN-mapped pte in
 650 * a region that doesn't allow it.
 651 *
 652 * The calling function must still handle the error.
 653 */
 654static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 655                          pte_t pte, struct page *page)
 656{
 657        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 658        pud_t *pud = pud_offset(pgd, addr);
 659        pmd_t *pmd = pmd_offset(pud, addr);
 660        struct address_space *mapping;
 661        pgoff_t index;
 662        static unsigned long resume;
 663        static unsigned long nr_shown;
 664        static unsigned long nr_unshown;
 665
 666        /*
 667         * Allow a burst of 60 reports, then keep quiet for that minute;
 668         * or allow a steady drip of one report per second.
 669         */
 670        if (nr_shown == 60) {
 671                if (time_before(jiffies, resume)) {
 672                        nr_unshown++;
 673                        return;
 674                }
 675                if (nr_unshown) {
 676                        printk(KERN_ALERT
 677                                "BUG: Bad page map: %lu messages suppressed\n",
 678                                nr_unshown);
 679                        nr_unshown = 0;
 680                }
 681                nr_shown = 0;
 682        }
 683        if (nr_shown++ == 0)
 684                resume = jiffies + 60 * HZ;
 685
 686        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 687        index = linear_page_index(vma, addr);
 688
 689        printk(KERN_ALERT
 690                "BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 691                current->comm,
 692                (long long)pte_val(pte), (long long)pmd_val(*pmd));
 693        if (page)
 694                dump_page(page);
 695        printk(KERN_ALERT
 696                "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
 697                (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 698        /*
 699         * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
 700         */
 701        if (vma->vm_ops)
 702                print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
 703                                (unsigned long)vma->vm_ops->fault);
 704        if (vma->vm_file && vma->vm_file->f_op)
 705                print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
 706                                (unsigned long)vma->vm_file->f_op->mmap);
 707        dump_stack();
 708        add_taint(TAINT_BAD_PAGE);
 709}
 710
 711static inline int is_cow_mapping(vm_flags_t flags)
 712{
 713        return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 714}
 715
 716#ifndef is_zero_pfn
 717static inline int is_zero_pfn(unsigned long pfn)
 718{
 719        return pfn == zero_pfn;
 720}
 721#endif
 722
 723#ifndef my_zero_pfn
 724static inline unsigned long my_zero_pfn(unsigned long addr)
 725{
 726        return zero_pfn;
 727}
 728#endif
 729
 730/*
 731 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 732 *
 733 * "Special" mappings do not wish to be associated with a "struct page" (either
 734 * it doesn't exist, or it exists but they don't want to touch it). In this
 735 * case, NULL is returned here. "Normal" mappings do have a struct page.
 736 *
 737 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 738 * pte bit, in which case this function is trivial. Secondly, an architecture
 739 * may not have a spare pte bit, which requires a more complicated scheme,
 740 * described below.
 741 *
 742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 743 * special mapping (even if there are underlying and valid "struct pages").
 744 * COWed pages of a VM_PFNMAP are always normal.
 745 *
 746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 749 * mapping will always honor the rule
 750 *
 751 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 752 *
 753 * And for normal mappings this is false.
 754 *
 755 * This restricts such mappings to be a linear translation from virtual address
 756 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 757 * as the vma is not a COW mapping; in that case, we know that all ptes are
 758 * special (because none can have been COWed).
 759 *
 760 *
 761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 762 *
 763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 764 * page" backing, however the difference is that _all_ pages with a struct
 765 * page (that is, those where pfn_valid is true) are refcounted and considered
 766 * normal pages by the VM. The disadvantage is that pages are refcounted
 767 * (which can be slower and simply not an option for some PFNMAP users). The
 768 * advantage is that we don't have to follow the strict linearity rule of
 769 * PFNMAP mappings in order to support COWable mappings.
 770 *
 771 */
 772#ifdef __HAVE_ARCH_PTE_SPECIAL
 773# define HAVE_PTE_SPECIAL 1
 774#else
 775# define HAVE_PTE_SPECIAL 0
 776#endif
 777struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 778                                pte_t pte)
 779{
 780        unsigned long pfn = pte_pfn(pte);
 781
 782        if (HAVE_PTE_SPECIAL) {
 783                if (likely(!pte_special(pte)))
 784                        goto check_pfn;
 785                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 786                        return NULL;
 787                if (!is_zero_pfn(pfn))
 788                        print_bad_pte(vma, addr, pte, NULL);
 789                return NULL;
 790        }
 791
 792        /* !HAVE_PTE_SPECIAL case follows: */
 793
 794        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 795                if (vma->vm_flags & VM_MIXEDMAP) {
 796                        if (!pfn_valid(pfn))
 797                                return NULL;
 798                        goto out;
 799                } else {
 800                        unsigned long off;
 801                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 802                        if (pfn == vma->vm_pgoff + off)
 803                                return NULL;
 804                        if (!is_cow_mapping(vma->vm_flags))
 805                                return NULL;
 806                }
 807        }
 808
 809        if (is_zero_pfn(pfn))
 810                return NULL;
 811check_pfn:
 812        if (unlikely(pfn > highest_memmap_pfn)) {
 813                print_bad_pte(vma, addr, pte, NULL);
 814                return NULL;
 815        }
 816
 817        /*
 818         * NOTE! We still have PageReserved() pages in the page tables.
 819         * eg. VDSO mappings can cause them to exist.
 820         */
 821out:
 822        return pfn_to_page(pfn);
 823}
 824
 825/*
 826 * copy one vm_area from one task to the other. Assumes the page tables
 827 * already present in the new task to be cleared in the whole range
 828 * covered by this vma.
 829 */
 830
 831static inline unsigned long
 832copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 833                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
 834                unsigned long addr, int *rss)
 835{
 836        unsigned long vm_flags = vma->vm_flags;
 837        pte_t pte = *src_pte;
 838        struct page *page;
 839
 840        /* pte contains position in swap or file, so copy. */
 841        if (unlikely(!pte_present(pte))) {
 842                if (!pte_file(pte)) {
 843                        swp_entry_t entry = pte_to_swp_entry(pte);
 844
 845                        if (swap_duplicate(entry) < 0)
 846                                return entry.val;
 847
 848                        /* make sure dst_mm is on swapoff's mmlist. */
 849                        if (unlikely(list_empty(&dst_mm->mmlist))) {
 850                                spin_lock(&mmlist_lock);
 851                                if (list_empty(&dst_mm->mmlist))
 852                                        list_add(&dst_mm->mmlist,
 853                                                 &src_mm->mmlist);
 854                                spin_unlock(&mmlist_lock);
 855                        }
 856                        if (likely(!non_swap_entry(entry)))
 857                                rss[MM_SWAPENTS]++;
 858                        else if (is_migration_entry(entry)) {
 859                                page = migration_entry_to_page(entry);
 860
 861                                if (PageAnon(page))
 862                                        rss[MM_ANONPAGES]++;
 863                                else
 864                                        rss[MM_FILEPAGES]++;
 865
 866                                if (is_write_migration_entry(entry) &&
 867                                    is_cow_mapping(vm_flags)) {
 868                                        /*
 869                                         * COW mappings require pages in both
 870                                         * parent and child to be set to read.
 871                                         */
 872                                        make_migration_entry_read(&entry);
 873                                        pte = swp_entry_to_pte(entry);
 874                                        set_pte_at(src_mm, addr, src_pte, pte);
 875                                }
 876                        }
 877                }
 878                goto out_set_pte;
 879        }
 880
 881        /*
 882         * If it's a COW mapping, write protect it both
 883         * in the parent and the child
 884         */
 885        if (is_cow_mapping(vm_flags)) {
 886                ptep_set_wrprotect(src_mm, addr, src_pte);
 887                pte = pte_wrprotect(pte);
 888        }
 889
 890        /*
 891         * If it's a shared mapping, mark it clean in
 892         * the child
 893         */
 894        if (vm_flags & VM_SHARED)
 895                pte = pte_mkclean(pte);
 896        pte = pte_mkold(pte);
 897
 898        page = vm_normal_page(vma, addr, pte);
 899        if (page) {
 900                get_page(page);
 901                page_dup_rmap(page);
 902                if (PageAnon(page))
 903                        rss[MM_ANONPAGES]++;
 904                else
 905                        rss[MM_FILEPAGES]++;
 906        }
 907
 908out_set_pte:
 909        set_pte_at(dst_mm, addr, dst_pte, pte);
 910        return 0;
 911}
 912
 913int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 914                   pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
 915                   unsigned long addr, unsigned long end)
 916{
 917        pte_t *orig_src_pte, *orig_dst_pte;
 918        pte_t *src_pte, *dst_pte;
 919        spinlock_t *src_ptl, *dst_ptl;
 920        int progress = 0;
 921        int rss[NR_MM_COUNTERS];
 922        swp_entry_t entry = (swp_entry_t){0};
 923
 924again:
 925        init_rss_vec(rss);
 926
 927        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
 928        if (!dst_pte)
 929                return -ENOMEM;
 930        src_pte = pte_offset_map(src_pmd, addr);
 931        src_ptl = pte_lockptr(src_mm, src_pmd);
 932        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
 933        orig_src_pte = src_pte;
 934        orig_dst_pte = dst_pte;
 935        arch_enter_lazy_mmu_mode();
 936
 937        do {
 938                /*
 939                 * We are holding two locks at this point - either of them
 940                 * could generate latencies in another task on another CPU.
 941                 */
 942                if (progress >= 32) {
 943                        progress = 0;
 944                        if (need_resched() ||
 945                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
 946                                break;
 947                }
 948                if (pte_none(*src_pte)) {
 949                        progress++;
 950                        continue;
 951                }
 952                entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
 953                                                        vma, addr, rss);
 954                if (entry.val)
 955                        break;
 956                progress += 8;
 957        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
 958
 959        arch_leave_lazy_mmu_mode();
 960        spin_unlock(src_ptl);
 961        pte_unmap(orig_src_pte);
 962        add_mm_rss_vec(dst_mm, rss);
 963        pte_unmap_unlock(orig_dst_pte, dst_ptl);
 964        cond_resched();
 965
 966        if (entry.val) {
 967                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
 968                        return -ENOMEM;
 969                progress = 0;
 970        }
 971        if (addr != end)
 972                goto again;
 973        return 0;
 974}
 975
 976static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 977                pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
 978                unsigned long addr, unsigned long end)
 979{
 980        pmd_t *src_pmd, *dst_pmd;
 981        unsigned long next;
 982
 983        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
 984        if (!dst_pmd)
 985                return -ENOMEM;
 986        src_pmd = pmd_offset(src_pud, addr);
 987        do {
 988                next = pmd_addr_end(addr, end);
 989                if (pmd_trans_huge(*src_pmd)) {
 990                        int err;
 991                        VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
 992                        err = copy_huge_pmd(dst_mm, src_mm,
 993                                            dst_pmd, src_pmd, addr, vma);
 994                        if (err == -ENOMEM)
 995                                return -ENOMEM;
 996                        if (!err)
 997                                continue;
 998                        /* fall through */
 999                }
1000                if (pmd_none_or_clear_bad(src_pmd))
1001                        continue;
1002                if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1003                                                vma, addr, next))
1004                        return -ENOMEM;
1005        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1006        return 0;
1007}
1008
1009static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010                pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1011                unsigned long addr, unsigned long end)
1012{
1013        pud_t *src_pud, *dst_pud;
1014        unsigned long next;
1015
1016        dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1017        if (!dst_pud)
1018                return -ENOMEM;
1019        src_pud = pud_offset(src_pgd, addr);
1020        do {
1021                next = pud_addr_end(addr, end);
1022                if (pud_none_or_clear_bad(src_pud))
1023                        continue;
1024                if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1025                                                vma, addr, next))
1026                        return -ENOMEM;
1027        } while (dst_pud++, src_pud++, addr = next, addr != end);
1028        return 0;
1029}
1030
1031int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1032                struct vm_area_struct *vma)
1033{
1034        pgd_t *src_pgd, *dst_pgd;
1035        unsigned long next;
1036        unsigned long addr = vma->vm_start;
1037        unsigned long end = vma->vm_end;
1038        int ret;
1039
1040        /*
1041         * Don't copy ptes where a page fault will fill them correctly.
1042         * Fork becomes much lighter when there are big shared or private
1043         * readonly mappings. The tradeoff is that copy_page_range is more
1044         * efficient than faulting.
1045         */
1046        if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1047                if (!vma->anon_vma)
1048                        return 0;
1049        }
1050
1051        if (is_vm_hugetlb_page(vma))
1052                return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1053
1054        if (unlikely(is_pfn_mapping(vma))) {
1055                /*
1056                 * We do not free on error cases below as remove_vma
1057                 * gets called on error from higher level routine
1058                 */
1059                ret = track_pfn_vma_copy(vma);
1060                if (ret)
1061                        return ret;
1062        }
1063
1064        /*
1065         * We need to invalidate the secondary MMU mappings only when
1066         * there could be a permission downgrade on the ptes of the
1067         * parent mm. And a permission downgrade will only happen if
1068         * is_cow_mapping() returns true.
1069         */
1070        if (is_cow_mapping(vma->vm_flags))
1071                mmu_notifier_invalidate_range_start(src_mm, addr, end);
1072
1073        ret = 0;
1074        dst_pgd = pgd_offset(dst_mm, addr);
1075        src_pgd = pgd_offset(src_mm, addr);
1076        do {
1077                next = pgd_addr_end(addr, end);
1078                if (pgd_none_or_clear_bad(src_pgd))
1079                        continue;
1080                if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1081                                            vma, addr, next))) {
1082                        ret = -ENOMEM;
1083                        break;
1084                }
1085        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1086
1087        if (is_cow_mapping(vma->vm_flags))
1088                mmu_notifier_invalidate_range_end(src_mm,
1089                                                  vma->vm_start, end);
1090        return ret;
1091}
1092
1093static unsigned long zap_pte_range(struct mmu_gather *tlb,
1094                                struct vm_area_struct *vma, pmd_t *pmd,
1095                                unsigned long addr, unsigned long end,
1096                                struct zap_details *details)
1097{
1098        struct mm_struct *mm = tlb->mm;
1099        int force_flush = 0;
1100        int rss[NR_MM_COUNTERS];
1101        spinlock_t *ptl;
1102        pte_t *start_pte;
1103        pte_t *pte;
1104
1105again:
1106        init_rss_vec(rss);
1107        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108        pte = start_pte;
1109        arch_enter_lazy_mmu_mode();
1110        do {
1111                pte_t ptent = *pte;
1112                if (pte_none(ptent)) {
1113                        continue;
1114                }
1115
1116                if (pte_present(ptent)) {
1117                        struct page *page;
1118
1119                        page = vm_normal_page(vma, addr, ptent);
1120                        if (unlikely(details) && page) {
1121                                /*
1122                                 * unmap_shared_mapping_pages() wants to
1123                                 * invalidate cache without truncating:
1124                                 * unmap shared but keep private pages.
1125                                 */
1126                                if (details->check_mapping &&
1127                                    details->check_mapping != page->mapping)
1128                                        continue;
1129                                /*
1130                                 * Each page->index must be checked when
1131                                 * invalidating or truncating nonlinear.
1132                                 */
1133                                if (details->nonlinear_vma &&
1134                                    (page->index < details->first_index ||
1135                                     page->index > details->last_index))
1136                                        continue;
1137                        }
1138                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1139                                                        tlb->fullmm);
1140                        tlb_remove_tlb_entry(tlb, pte, addr);
1141                        if (unlikely(!page))
1142                                continue;
1143                        if (unlikely(details) && details->nonlinear_vma
1144                            && linear_page_index(details->nonlinear_vma,
1145                                                addr) != page->index)
1146                                set_pte_at(mm, addr, pte,
1147                                           pgoff_to_pte(page->index));
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(!VM_SequentialReadHint(vma)))
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                force_flush = 0;
1207                tlb_flush_mmu(tlb);
1208                if (addr != end)
1209                        goto again;
1210        }
1211
1212        return addr;
1213}
1214
1215static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1216                                struct vm_area_struct *vma, pud_t *pud,
1217                                unsigned long addr, unsigned long end,
1218                                struct zap_details *details)
1219{
1220        pmd_t *pmd;
1221        unsigned long next;
1222
1223        pmd = pmd_offset(pud, addr);
1224        do {
1225                next = pmd_addr_end(addr, end);
1226                if (pmd_trans_huge(*pmd)) {
1227                        if (next - addr != HPAGE_PMD_SIZE) {
1228#ifdef CONFIG_DEBUG_VM
1229                                if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1230                                        pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1231                                                __func__, addr, end,
1232                                                vma->vm_start,
1233                                                vma->vm_end);
1234                                        BUG();
1235                                }
1236#endif
1237                                split_huge_page_pmd(vma->vm_mm, pmd);
1238                        } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1239                                goto next;
1240                        /* fall through */
1241                }
1242                /*
1243                 * Here there can be other concurrent MADV_DONTNEED or
1244                 * trans huge page faults running, and if the pmd is
1245                 * none or trans huge it can change under us. This is
1246                 * because MADV_DONTNEED holds the mmap_sem in read
1247                 * mode.
1248                 */
1249                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1250                        goto next;
1251                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1252next:
1253                cond_resched();
1254        } while (pmd++, addr = next, addr != end);
1255
1256        return addr;
1257}
1258
1259static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1260                                struct vm_area_struct *vma, pgd_t *pgd,
1261                                unsigned long addr, unsigned long end,
1262                                struct zap_details *details)
1263{
1264        pud_t *pud;
1265        unsigned long next;
1266
1267        pud = pud_offset(pgd, addr);
1268        do {
1269                next = pud_addr_end(addr, end);
1270                if (pud_none_or_clear_bad(pud))
1271                        continue;
1272                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1273        } while (pud++, addr = next, addr != end);
1274
1275        return addr;
1276}
1277
1278static void unmap_page_range(struct mmu_gather *tlb,
1279                             struct vm_area_struct *vma,
1280                             unsigned long addr, unsigned long end,
1281                             struct zap_details *details)
1282{
1283        pgd_t *pgd;
1284        unsigned long next;
1285
1286        if (details && !details->check_mapping && !details->nonlinear_vma)
1287                details = NULL;
1288
1289        BUG_ON(addr >= end);
1290        mem_cgroup_uncharge_start();
1291        tlb_start_vma(tlb, vma);
1292        pgd = pgd_offset(vma->vm_mm, addr);
1293        do {
1294                next = pgd_addr_end(addr, end);
1295                if (pgd_none_or_clear_bad(pgd))
1296                        continue;
1297                next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1298        } while (pgd++, addr = next, addr != end);
1299        tlb_end_vma(tlb, vma);
1300        mem_cgroup_uncharge_end();
1301}
1302
1303
1304static void unmap_single_vma(struct mmu_gather *tlb,
1305                struct vm_area_struct *vma, unsigned long start_addr,
1306                unsigned long end_addr,
1307                struct zap_details *details)
1308{
1309        unsigned long start = max(vma->vm_start, start_addr);
1310        unsigned long end;
1311
1312        if (start >= vma->vm_end)
1313                return;
1314        end = min(vma->vm_end, end_addr);
1315        if (end <= vma->vm_start)
1316                return;
1317
1318        if (vma->vm_file)
1319                uprobe_munmap(vma, start, end);
1320
1321        if (unlikely(is_pfn_mapping(vma)))
1322                untrack_pfn_vma(vma, 0, 0);
1323
1324        if (start != end) {
1325                if (unlikely(is_vm_hugetlb_page(vma))) {
1326                        /*
1327                         * It is undesirable to test vma->vm_file as it
1328                         * should be non-null for valid hugetlb area.
1329                         * However, vm_file will be NULL in the error
1330                         * cleanup path of do_mmap_pgoff. When
1331                         * hugetlbfs ->mmap method fails,
1332                         * do_mmap_pgoff() nullifies vma->vm_file
1333                         * before calling this function to clean up.
1334                         * Since no pte has actually been setup, it is
1335                         * safe to do nothing in this case.
1336                         */
1337                        if (vma->vm_file)
1338                                unmap_hugepage_range(vma, start, end, NULL);
1339                } else
1340                        unmap_page_range(tlb, vma, start, end, details);
1341        }
1342}
1343
1344/**
1345 * unmap_vmas - unmap a range of memory covered by a list of vma's
1346 * @tlb: address of the caller's struct mmu_gather
1347 * @vma: the starting vma
1348 * @start_addr: virtual address at which to start unmapping
1349 * @end_addr: virtual address at which to end unmapping
1350 *
1351 * Unmap all pages in the vma list.
1352 *
1353 * Only addresses between `start' and `end' will be unmapped.
1354 *
1355 * The VMA list must be sorted in ascending virtual address order.
1356 *
1357 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358 * range after unmap_vmas() returns.  So the only responsibility here is to
1359 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360 * drops the lock and schedules.
1361 */
1362void unmap_vmas(struct mmu_gather *tlb,
1363                struct vm_area_struct *vma, unsigned long start_addr,
1364                unsigned long end_addr)
1365{
1366        struct mm_struct *mm = vma->vm_mm;
1367
1368        mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1369        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1370                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1371        mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @start: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of nonlinear truncation or shared cache invalidation
1380 *
1381 * Caller must protect the VMA list
1382 */
1383void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1384                unsigned long size, struct zap_details *details)
1385{
1386        struct mm_struct *mm = vma->vm_mm;
1387        struct mmu_gather tlb;
1388        unsigned long end = start + size;
1389
1390        lru_add_drain();
1391        tlb_gather_mmu(&tlb, mm, 0);
1392        update_hiwater_rss(mm);
1393        mmu_notifier_invalidate_range_start(mm, start, end);
1394        for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1395                unmap_single_vma(&tlb, vma, start, end, details);
1396        mmu_notifier_invalidate_range_end(mm, start, end);
1397        tlb_finish_mmu(&tlb, start, end);
1398}
1399
1400/**
1401 * zap_page_range_single - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of nonlinear truncation or shared cache invalidation
1406 *
1407 * The range must fit into one VMA.
1408 */
1409static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410                unsigned long size, struct zap_details *details)
1411{
1412        struct mm_struct *mm = vma->vm_mm;
1413        struct mmu_gather tlb;
1414        unsigned long end = address + size;
1415
1416        lru_add_drain();
1417        tlb_gather_mmu(&tlb, mm, 0);
1418        update_hiwater_rss(mm);
1419        mmu_notifier_invalidate_range_start(mm, address, end);
1420        unmap_single_vma(&tlb, vma, address, end, details);
1421        mmu_notifier_invalidate_range_end(mm, address, end);
1422        tlb_finish_mmu(&tlb, address, end);
1423}
1424
1425/**
1426 * zap_vma_ptes - remove ptes mapping the vma
1427 * @vma: vm_area_struct holding ptes to be zapped
1428 * @address: starting address of pages to zap
1429 * @size: number of bytes to zap
1430 *
1431 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432 *
1433 * The entire address range must be fully contained within the vma.
1434 *
1435 * Returns 0 if successful.
1436 */
1437int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438                unsigned long size)
1439{
1440        if (address < vma->vm_start || address + size > vma->vm_end ||
1441                        !(vma->vm_flags & VM_PFNMAP))
1442                return -1;
1443        zap_page_range_single(vma, address, size, NULL);
1444        return 0;
1445}
1446EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447
1448/**
1449 * follow_page - look up a page descriptor from a user-virtual address
1450 * @vma: vm_area_struct mapping @address
1451 * @address: virtual address to look up
1452 * @flags: flags modifying lookup behaviour
1453 *
1454 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1455 *
1456 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1457 * an error pointer if there is a mapping to something not represented
1458 * by a page descriptor (see also vm_normal_page()).
1459 */
1460struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1461                        unsigned int flags)
1462{
1463        pgd_t *pgd;
1464        pud_t *pud;
1465        pmd_t *pmd;
1466        pte_t *ptep, pte;
1467        spinlock_t *ptl;
1468        struct page *page;
1469        struct mm_struct *mm = vma->vm_mm;
1470
1471        page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1472        if (!IS_ERR(page)) {
1473                BUG_ON(flags & FOLL_GET);
1474                goto out;
1475        }
1476
1477        page = NULL;
1478        pgd = pgd_offset(mm, address);
1479        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1480                goto no_page_table;
1481
1482        pud = pud_offset(pgd, address);
1483        if (pud_none(*pud))
1484                goto no_page_table;
1485        if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1486                BUG_ON(flags & FOLL_GET);
1487                page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1488                goto out;
1489        }
1490        if (unlikely(pud_bad(*pud)))
1491                goto no_page_table;
1492
1493        pmd = pmd_offset(pud, address);
1494        if (pmd_none(*pmd))
1495                goto no_page_table;
1496        if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1497                BUG_ON(flags & FOLL_GET);
1498                page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1499                goto out;
1500        }
1501        if (pmd_trans_huge(*pmd)) {
1502                if (flags & FOLL_SPLIT) {
1503                        split_huge_page_pmd(mm, pmd);
1504                        goto split_fallthrough;
1505                }
1506                spin_lock(&mm->page_table_lock);
1507                if (likely(pmd_trans_huge(*pmd))) {
1508                        if (unlikely(pmd_trans_splitting(*pmd))) {
1509                                spin_unlock(&mm->page_table_lock);
1510                                wait_split_huge_page(vma->anon_vma, pmd);
1511                        } else {
1512                                page = follow_trans_huge_pmd(mm, address,
1513                                                             pmd, flags);
1514                                spin_unlock(&mm->page_table_lock);
1515                                goto out;
1516                        }
1517                } else
1518                        spin_unlock(&mm->page_table_lock);
1519                /* fall through */
1520        }
1521split_fallthrough:
1522        if (unlikely(pmd_bad(*pmd)))
1523                goto no_page_table;
1524
1525        ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1526
1527        pte = *ptep;
1528        if (!pte_present(pte))
1529                goto no_page;
1530        if ((flags & FOLL_WRITE) && !pte_write(pte))
1531                goto unlock;
1532
1533        page = vm_normal_page(vma, address, pte);
1534        if (unlikely(!page)) {
1535                if ((flags & FOLL_DUMP) ||
1536                    !is_zero_pfn(pte_pfn(pte)))
1537                        goto bad_page;
1538                page = pte_page(pte);
1539        }
1540
1541        if (flags & FOLL_GET)
1542                get_page_foll(page);
1543        if (flags & FOLL_TOUCH) {
1544                if ((flags & FOLL_WRITE) &&
1545                    !pte_dirty(pte) && !PageDirty(page))
1546                        set_page_dirty(page);
1547                /*
1548                 * pte_mkyoung() would be more correct here, but atomic care
1549                 * is needed to avoid losing the dirty bit: it is easier to use
1550                 * mark_page_accessed().
1551                 */
1552                mark_page_accessed(page);
1553        }
1554        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1555                /*
1556                 * The preliminary mapping check is mainly to avoid the
1557                 * pointless overhead of lock_page on the ZERO_PAGE
1558                 * which might bounce very badly if there is contention.
1559                 *
1560                 * If the page is already locked, we don't need to
1561                 * handle it now - vmscan will handle it later if and
1562                 * when it attempts to reclaim the page.
1563                 */
1564                if (page->mapping && trylock_page(page)) {
1565                        lru_add_drain();  /* push cached pages to LRU */
1566                        /*
1567                         * Because we lock page here and migration is
1568                         * blocked by the pte's page reference, we need
1569                         * only check for file-cache page truncation.
1570                         */
1571                        if (page->mapping)
1572                                mlock_vma_page(page);
1573                        unlock_page(page);
1574                }
1575        }
1576unlock:
1577        pte_unmap_unlock(ptep, ptl);
1578out:
1579        return page;
1580
1581bad_page:
1582        pte_unmap_unlock(ptep, ptl);
1583        return ERR_PTR(-EFAULT);
1584
1585no_page:
1586        pte_unmap_unlock(ptep, ptl);
1587        if (!pte_none(pte))
1588                return page;
1589
1590no_page_table:
1591        /*
1592         * When core dumping an enormous anonymous area that nobody
1593         * has touched so far, we don't want to allocate unnecessary pages or
1594         * page tables.  Return error instead of NULL to skip handle_mm_fault,
1595         * then get_dump_page() will return NULL to leave a hole in the dump.
1596         * But we can only make this optimization where a hole would surely
1597         * be zero-filled if handle_mm_fault() actually did handle it.
1598         */
1599        if ((flags & FOLL_DUMP) &&
1600            (!vma->vm_ops || !vma->vm_ops->fault))
1601                return ERR_PTR(-EFAULT);
1602        return page;
1603}
1604
1605static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1606{
1607        return stack_guard_page_start(vma, addr) ||
1608               stack_guard_page_end(vma, addr+PAGE_SIZE);
1609}
1610
1611/**
1612 * __get_user_pages() - pin user pages in memory
1613 * @tsk:        task_struct of target task
1614 * @mm:         mm_struct of target mm
1615 * @start:      starting user address
1616 * @nr_pages:   number of pages from start to pin
1617 * @gup_flags:  flags modifying pin behaviour
1618 * @pages:      array that receives pointers to the pages pinned.
1619 *              Should be at least nr_pages long. Or NULL, if caller
1620 *              only intends to ensure the pages are faulted in.
1621 * @vmas:       array of pointers to vmas corresponding to each page.
1622 *              Or NULL if the caller does not require them.
1623 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1624 *
1625 * Returns number of pages pinned. This may be fewer than the number
1626 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1627 * were pinned, returns -errno. Each page returned must be released
1628 * with a put_page() call when it is finished with. vmas will only
1629 * remain valid while mmap_sem is held.
1630 *
1631 * Must be called with mmap_sem held for read or write.
1632 *
1633 * __get_user_pages walks a process's page tables and takes a reference to
1634 * each struct page that each user address corresponds to at a given
1635 * instant. That is, it takes the page that would be accessed if a user
1636 * thread accesses the given user virtual address at that instant.
1637 *
1638 * This does not guarantee that the page exists in the user mappings when
1639 * __get_user_pages returns, and there may even be a completely different
1640 * page there in some cases (eg. if mmapped pagecache has been invalidated
1641 * and subsequently re faulted). However it does guarantee that the page
1642 * won't be freed completely. And mostly callers simply care that the page
1643 * contains data that was valid *at some point in time*. Typically, an IO
1644 * or similar operation cannot guarantee anything stronger anyway because
1645 * locks can't be held over the syscall boundary.
1646 *
1647 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1648 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1649 * appropriate) must be called after the page is finished with, and
1650 * before put_page is called.
1651 *
1652 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1653 * or mmap_sem contention, and if waiting is needed to pin all pages,
1654 * *@nonblocking will be set to 0.
1655 *
1656 * In most cases, get_user_pages or get_user_pages_fast should be used
1657 * instead of __get_user_pages. __get_user_pages should be used only if
1658 * you need some special @gup_flags.
1659 */
1660int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1661                     unsigned long start, int nr_pages, unsigned int gup_flags,
1662                     struct page **pages, struct vm_area_struct **vmas,
1663                     int *nonblocking)
1664{
1665        int i;
1666        unsigned long vm_flags;
1667
1668        if (nr_pages <= 0)
1669                return 0;
1670
1671        VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1672
1673        /* 
1674         * Require read or write permissions.
1675         * If FOLL_FORCE is set, we only require the "MAY" flags.
1676         */
1677        vm_flags  = (gup_flags & FOLL_WRITE) ?
1678                        (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1679        vm_flags &= (gup_flags & FOLL_FORCE) ?
1680                        (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1681        i = 0;
1682
1683        do {
1684                struct vm_area_struct *vma;
1685
1686                vma = find_extend_vma(mm, start);
1687                if (!vma && in_gate_area(mm, start)) {
1688                        unsigned long pg = start & PAGE_MASK;
1689                        pgd_t *pgd;
1690                        pud_t *pud;
1691                        pmd_t *pmd;
1692                        pte_t *pte;
1693
1694                        /* user gate pages are read-only */
1695                        if (gup_flags & FOLL_WRITE)
1696                                return i ? : -EFAULT;
1697                        if (pg > TASK_SIZE)
1698                                pgd = pgd_offset_k(pg);
1699                        else
1700                                pgd = pgd_offset_gate(mm, pg);
1701                        BUG_ON(pgd_none(*pgd));
1702                        pud = pud_offset(pgd, pg);
1703                        BUG_ON(pud_none(*pud));
1704                        pmd = pmd_offset(pud, pg);
1705                        if (pmd_none(*pmd))
1706                                return i ? : -EFAULT;
1707                        VM_BUG_ON(pmd_trans_huge(*pmd));
1708                        pte = pte_offset_map(pmd, pg);
1709                        if (pte_none(*pte)) {
1710                                pte_unmap(pte);
1711                                return i ? : -EFAULT;
1712                        }
1713                        vma = get_gate_vma(mm);
1714                        if (pages) {
1715                                struct page *page;
1716
1717                                page = vm_normal_page(vma, start, *pte);
1718                                if (!page) {
1719                                        if (!(gup_flags & FOLL_DUMP) &&
1720                                             is_zero_pfn(pte_pfn(*pte)))
1721                                                page = pte_page(*pte);
1722                                        else {
1723                                                pte_unmap(pte);
1724                                                return i ? : -EFAULT;
1725                                        }
1726                                }
1727                                pages[i] = page;
1728                                get_page(page);
1729                        }
1730                        pte_unmap(pte);
1731                        goto next_page;
1732                }
1733
1734                if (!vma ||
1735                    (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1736                    !(vm_flags & vma->vm_flags))
1737                        return i ? : -EFAULT;
1738
1739                if (is_vm_hugetlb_page(vma)) {
1740                        i = follow_hugetlb_page(mm, vma, pages, vmas,
1741                                        &start, &nr_pages, i, gup_flags);
1742                        continue;
1743                }
1744
1745                do {
1746                        struct page *page;
1747                        unsigned int foll_flags = gup_flags;
1748
1749                        /*
1750                         * If we have a pending SIGKILL, don't keep faulting
1751                         * pages and potentially allocating memory.
1752                         */
1753                        if (unlikely(fatal_signal_pending(current)))
1754                                return i ? i : -ERESTARTSYS;
1755
1756                        cond_resched();
1757                        while (!(page = follow_page(vma, start, foll_flags))) {
1758                                int ret;
1759                                unsigned int fault_flags = 0;
1760
1761                                /* For mlock, just skip the stack guard page. */
1762                                if (foll_flags & FOLL_MLOCK) {
1763                                        if (stack_guard_page(vma, start))
1764                                                goto next_page;
1765                                }
1766                                if (foll_flags & FOLL_WRITE)
1767                                        fault_flags |= FAULT_FLAG_WRITE;
1768                                if (nonblocking)
1769                                        fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1770                                if (foll_flags & FOLL_NOWAIT)
1771                                        fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1772
1773                                ret = handle_mm_fault(mm, vma, start,
1774                                                        fault_flags);
1775
1776                                if (ret & VM_FAULT_ERROR) {
1777                                        if (ret & VM_FAULT_OOM)
1778                                                return i ? i : -ENOMEM;
1779                                        if (ret & (VM_FAULT_HWPOISON |
1780                                                   VM_FAULT_HWPOISON_LARGE)) {
1781                                                if (i)
1782                                                        return i;
1783                                                else if (gup_flags & FOLL_HWPOISON)
1784                                                        return -EHWPOISON;
1785                                                else
1786                                                        return -EFAULT;
1787                                        }
1788                                        if (ret & VM_FAULT_SIGBUS)
1789                                                return i ? i : -EFAULT;
1790                                        BUG();
1791                                }
1792
1793                                if (tsk) {
1794                                        if (ret & VM_FAULT_MAJOR)
1795                                                tsk->maj_flt++;
1796                                        else
1797                                                tsk->min_flt++;
1798                                }
1799
1800                                if (ret & VM_FAULT_RETRY) {
1801                                        if (nonblocking)
1802                                                *nonblocking = 0;
1803                                        return i;
1804                                }
1805
1806                                /*
1807                                 * The VM_FAULT_WRITE bit tells us that
1808                                 * do_wp_page has broken COW when necessary,
1809                                 * even if maybe_mkwrite decided not to set
1810                                 * pte_write. We can thus safely do subsequent
1811                                 * page lookups as if they were reads. But only
1812                                 * do so when looping for pte_write is futile:
1813                                 * in some cases userspace may also be wanting
1814                                 * to write to the gotten user page, which a
1815                                 * read fault here might prevent (a readonly
1816                                 * page might get reCOWed by userspace write).
1817                                 */
1818                                if ((ret & VM_FAULT_WRITE) &&
1819                                    !(vma->vm_flags & VM_WRITE))
1820                                        foll_flags &= ~FOLL_WRITE;
1821
1822                                cond_resched();
1823                        }
1824                        if (IS_ERR(page))
1825                                return i ? i : PTR_ERR(page);
1826                        if (pages) {
1827                                pages[i] = page;
1828
1829                                flush_anon_page(vma, page, start);
1830                                flush_dcache_page(page);
1831                        }
1832next_page:
1833                        if (vmas)
1834                                vmas[i] = vma;
1835                        i++;
1836                        start += PAGE_SIZE;
1837                        nr_pages--;
1838                } while (nr_pages && start < vma->vm_end);
1839        } while (nr_pages);
1840        return i;
1841}
1842EXPORT_SYMBOL(__get_user_pages);
1843
1844/*
1845 * fixup_user_fault() - manually resolve a user page fault
1846 * @tsk:        the task_struct to use for page fault accounting, or
1847 *              NULL if faults are not to be recorded.
1848 * @mm:         mm_struct of target mm
1849 * @address:    user address
1850 * @fault_flags:flags to pass down to handle_mm_fault()
1851 *
1852 * This is meant to be called in the specific scenario where for locking reasons
1853 * we try to access user memory in atomic context (within a pagefault_disable()
1854 * section), this returns -EFAULT, and we want to resolve the user fault before
1855 * trying again.
1856 *
1857 * Typically this is meant to be used by the futex code.
1858 *
1859 * The main difference with get_user_pages() is that this function will
1860 * unconditionally call handle_mm_fault() which will in turn perform all the
1861 * necessary SW fixup of the dirty and young bits in the PTE, while
1862 * handle_mm_fault() only guarantees to update these in the struct page.
1863 *
1864 * This is important for some architectures where those bits also gate the
1865 * access permission to the page because they are maintained in software.  On
1866 * such architectures, gup() will not be enough to make a subsequent access
1867 * succeed.
1868 *
1869 * This should be called with the mm_sem held for read.
1870 */
1871int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1872                     unsigned long address, unsigned int fault_flags)
1873{
1874        struct vm_area_struct *vma;
1875        int ret;
1876
1877        vma = find_extend_vma(mm, address);
1878        if (!vma || address < vma->vm_start)
1879                return -EFAULT;
1880
1881        ret = handle_mm_fault(mm, vma, address, fault_flags);
1882        if (ret & VM_FAULT_ERROR) {
1883                if (ret & VM_FAULT_OOM)
1884                        return -ENOMEM;
1885                if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1886                        return -EHWPOISON;
1887                if (ret & VM_FAULT_SIGBUS)
1888                        return -EFAULT;
1889                BUG();
1890        }
1891        if (tsk) {
1892                if (ret & VM_FAULT_MAJOR)
1893                        tsk->maj_flt++;
1894                else
1895                        tsk->min_flt++;
1896        }
1897        return 0;
1898}
1899
1900/*
1901 * get_user_pages() - pin user pages in memory
1902 * @tsk:        the task_struct to use for page fault accounting, or
1903 *              NULL if faults are not to be recorded.
1904 * @mm:         mm_struct of target mm
1905 * @start:      starting user address
1906 * @nr_pages:   number of pages from start to pin
1907 * @write:      whether pages will be written to by the caller
1908 * @force:      whether to force write access even if user mapping is
1909 *              readonly. This will result in the page being COWed even
1910 *              in MAP_SHARED mappings. You do not want this.
1911 * @pages:      array that receives pointers to the pages pinned.
1912 *              Should be at least nr_pages long. Or NULL, if caller
1913 *              only intends to ensure the pages are faulted in.
1914 * @vmas:       array of pointers to vmas corresponding to each page.
1915 *              Or NULL if the caller does not require them.
1916 *
1917 * Returns number of pages pinned. This may be fewer than the number
1918 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1919 * were pinned, returns -errno. Each page returned must be released
1920 * with a put_page() call when it is finished with. vmas will only
1921 * remain valid while mmap_sem is held.
1922 *
1923 * Must be called with mmap_sem held for read or write.
1924 *
1925 * get_user_pages walks a process's page tables and takes a reference to
1926 * each struct page that each user address corresponds to at a given
1927 * instant. That is, it takes the page that would be accessed if a user
1928 * thread accesses the given user virtual address at that instant.
1929 *
1930 * This does not guarantee that the page exists in the user mappings when
1931 * get_user_pages returns, and there may even be a completely different
1932 * page there in some cases (eg. if mmapped pagecache has been invalidated
1933 * and subsequently re faulted). However it does guarantee that the page
1934 * won't be freed completely. And mostly callers simply care that the page
1935 * contains data that was valid *at some point in time*. Typically, an IO
1936 * or similar operation cannot guarantee anything stronger anyway because
1937 * locks can't be held over the syscall boundary.
1938 *
1939 * If write=0, the page must not be written to. If the page is written to,
1940 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1941 * after the page is finished with, and before put_page is called.
1942 *
1943 * get_user_pages is typically used for fewer-copy IO operations, to get a
1944 * handle on the memory by some means other than accesses via the user virtual
1945 * addresses. The pages may be submitted for DMA to devices or accessed via
1946 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1947 * use the correct cache flushing APIs.
1948 *
1949 * See also get_user_pages_fast, for performance critical applications.
1950 */
1951int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1952                unsigned long start, int nr_pages, int write, int force,
1953                struct page **pages, struct vm_area_struct **vmas)
1954{
1955        int flags = FOLL_TOUCH;
1956
1957        if (pages)
1958                flags |= FOLL_GET;
1959        if (write)
1960                flags |= FOLL_WRITE;
1961        if (force)
1962                flags |= FOLL_FORCE;
1963
1964        return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1965                                NULL);
1966}
1967EXPORT_SYMBOL(get_user_pages);
1968
1969/**
1970 * get_dump_page() - pin user page in memory while writing it to core dump
1971 * @addr: user address
1972 *
1973 * Returns struct page pointer of user page pinned for dump,
1974 * to be freed afterwards by page_cache_release() or put_page().
1975 *
1976 * Returns NULL on any kind of failure - a hole must then be inserted into
1977 * the corefile, to preserve alignment with its headers; and also returns
1978 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1979 * allowing a hole to be left in the corefile to save diskspace.
1980 *
1981 * Called without mmap_sem, but after all other threads have been killed.
1982 */
1983#ifdef CONFIG_ELF_CORE
1984struct page *get_dump_page(unsigned long addr)
1985{
1986        struct vm_area_struct *vma;
1987        struct page *page;
1988
1989        if (__get_user_pages(current, current->mm, addr, 1,
1990                             FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1991                             NULL) < 1)
1992                return NULL;
1993        flush_cache_page(vma, addr, page_to_pfn(page));
1994        return page;
1995}
1996#endif /* CONFIG_ELF_CORE */
1997
1998pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1999                        spinlock_t **ptl)
2000{
2001        pgd_t * pgd = pgd_offset(mm, addr);
2002        pud_t * pud = pud_alloc(mm, pgd, addr);
2003        if (pud) {
2004                pmd_t * pmd = pmd_alloc(mm, pud, addr);
2005                if (pmd) {
2006                        VM_BUG_ON(pmd_trans_huge(*pmd));
2007                        return pte_alloc_map_lock(mm, pmd, addr, ptl);
2008                }
2009        }
2010        return NULL;
2011}
2012
2013/*
2014 * This is the old fallback for page remapping.
2015 *
2016 * For historical reasons, it only allows reserved pages. Only
2017 * old drivers should use this, and they needed to mark their
2018 * pages reserved for the old functions anyway.
2019 */
2020static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2021                        struct page *page, pgprot_t prot)
2022{
2023        struct mm_struct *mm = vma->vm_mm;
2024        int retval;
2025        pte_t *pte;
2026        spinlock_t *ptl;
2027
2028        retval = -EINVAL;
2029        if (PageAnon(page))
2030                goto out;
2031        retval = -ENOMEM;
2032        flush_dcache_page(page);
2033        pte = get_locked_pte(mm, addr, &ptl);
2034        if (!pte)
2035                goto out;
2036        retval = -EBUSY;
2037        if (!pte_none(*pte))
2038                goto out_unlock;
2039
2040        /* Ok, finally just insert the thing.. */
2041        get_page(page);
2042        inc_mm_counter_fast(mm, MM_FILEPAGES);
2043        page_add_file_rmap(page);
2044        set_pte_at(mm, addr, pte, mk_pte(page, prot));
2045
2046        retval = 0;
2047        pte_unmap_unlock(pte, ptl);
2048        return retval;
2049out_unlock:
2050        pte_unmap_unlock(pte, ptl);
2051out:
2052        return retval;
2053}
2054
2055/**
2056 * vm_insert_page - insert single page into user vma
2057 * @vma: user vma to map to
2058 * @addr: target user address of this page
2059 * @page: source kernel page
2060 *
2061 * This allows drivers to insert individual pages they've allocated
2062 * into a user vma.
2063 *
2064 * The page has to be a nice clean _individual_ kernel allocation.
2065 * If you allocate a compound page, you need to have marked it as
2066 * such (__GFP_COMP), or manually just split the page up yourself
2067 * (see split_page()).
2068 *
2069 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2070 * took an arbitrary page protection parameter. This doesn't allow
2071 * that. Your vma protection will have to be set up correctly, which
2072 * means that if you want a shared writable mapping, you'd better
2073 * ask for a shared writable mapping!
2074 *
2075 * The page does not need to be reserved.
2076 */
2077int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2078                        struct page *page)
2079{
2080        if (addr < vma->vm_start || addr >= vma->vm_end)
2081                return -EFAULT;
2082        if (!page_count(page))
2083                return -EINVAL;
2084        vma->vm_flags |= VM_INSERTPAGE;
2085        return insert_page(vma, addr, page, vma->vm_page_prot);
2086}
2087EXPORT_SYMBOL(vm_insert_page);
2088
2089static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2090                        unsigned long pfn, pgprot_t prot)
2091{
2092        struct mm_struct *mm = vma->vm_mm;
2093        int retval;
2094        pte_t *pte, entry;
2095        spinlock_t *ptl;
2096
2097        retval = -ENOMEM;
2098        pte = get_locked_pte(mm, addr, &ptl);
2099        if (!pte)
2100                goto out;
2101        retval = -EBUSY;
2102        if (!pte_none(*pte))
2103                goto out_unlock;
2104
2105        /* Ok, finally just insert the thing.. */
2106        entry = pte_mkspecial(pfn_pte(pfn, prot));
2107        set_pte_at(mm, addr, pte, entry);
2108        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2109
2110        retval = 0;
2111out_unlock:
2112        pte_unmap_unlock(pte, ptl);
2113out:
2114        return retval;
2115}
2116
2117/**
2118 * vm_insert_pfn - insert single pfn into user vma
2119 * @vma: user vma to map to
2120 * @addr: target user address of this page
2121 * @pfn: source kernel pfn
2122 *
2123 * Similar to vm_inert_page, this allows drivers to insert individual pages
2124 * they've allocated into a user vma. Same comments apply.
2125 *
2126 * This function should only be called from a vm_ops->fault handler, and
2127 * in that case the handler should return NULL.
2128 *
2129 * vma cannot be a COW mapping.
2130 *
2131 * As this is called only for pages that do not currently exist, we
2132 * do not need to flush old virtual caches or the TLB.
2133 */
2134int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2135                        unsigned long pfn)
2136{
2137        int ret;
2138        pgprot_t pgprot = vma->vm_page_prot;
2139        /*
2140         * Technically, architectures with pte_special can avoid all these
2141         * restrictions (same for remap_pfn_range).  However we would like
2142         * consistency in testing and feature parity among all, so we should
2143         * try to keep these invariants in place for everybody.
2144         */
2145        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2146        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2147                                                (VM_PFNMAP|VM_MIXEDMAP));
2148        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2149        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2150
2151        if (addr < vma->vm_start || addr >= vma->vm_end)
2152                return -EFAULT;
2153        if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2154                return -EINVAL;
2155
2156        ret = insert_pfn(vma, addr, pfn, pgprot);
2157
2158        if (ret)
2159                untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2160
2161        return ret;
2162}
2163EXPORT_SYMBOL(vm_insert_pfn);
2164
2165int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2166                        unsigned long pfn)
2167{
2168        BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2169
2170        if (addr < vma->vm_start || addr >= vma->vm_end)
2171                return -EFAULT;
2172
2173        /*
2174         * If we don't have pte special, then we have to use the pfn_valid()
2175         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2176         * refcount the page if pfn_valid is true (hence insert_page rather
2177         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2178         * without pte special, it would there be refcounted as a normal page.
2179         */
2180        if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2181                struct page *page;
2182
2183                page = pfn_to_page(pfn);
2184                return insert_page(vma, addr, page, vma->vm_page_prot);
2185        }
2186        return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2187}
2188EXPORT_SYMBOL(vm_insert_mixed);
2189
2190/*
2191 * maps a range of physical memory into the requested pages. the old
2192 * mappings are removed. any references to nonexistent pages results
2193 * in null mappings (currently treated as "copy-on-access")
2194 */
2195static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2196                        unsigned long addr, unsigned long end,
2197                        unsigned long pfn, pgprot_t prot)
2198{
2199        pte_t *pte;
2200        spinlock_t *ptl;
2201
2202        pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2203        if (!pte)
2204                return -ENOMEM;
2205        arch_enter_lazy_mmu_mode();
2206        do {
2207                BUG_ON(!pte_none(*pte));
2208                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2209                pfn++;
2210        } while (pte++, addr += PAGE_SIZE, addr != end);
2211        arch_leave_lazy_mmu_mode();
2212        pte_unmap_unlock(pte - 1, ptl);
2213        return 0;
2214}
2215
2216static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2217                        unsigned long addr, unsigned long end,
2218                        unsigned long pfn, pgprot_t prot)
2219{
2220        pmd_t *pmd;
2221        unsigned long next;
2222
2223        pfn -= addr >> PAGE_SHIFT;
2224        pmd = pmd_alloc(mm, pud, addr);
2225        if (!pmd)
2226                return -ENOMEM;
2227        VM_BUG_ON(pmd_trans_huge(*pmd));
2228        do {
2229                next = pmd_addr_end(addr, end);
2230                if (remap_pte_range(mm, pmd, addr, next,
2231                                pfn + (addr >> PAGE_SHIFT), prot))
2232                        return -ENOMEM;
2233        } while (pmd++, addr = next, addr != end);
2234        return 0;
2235}
2236
2237static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2238                        unsigned long addr, unsigned long end,
2239                        unsigned long pfn, pgprot_t prot)
2240{
2241        pud_t *pud;
2242        unsigned long next;
2243
2244        pfn -= addr >> PAGE_SHIFT;
2245        pud = pud_alloc(mm, pgd, addr);
2246        if (!pud)
2247                return -ENOMEM;
2248        do {
2249                next = pud_addr_end(addr, end);
2250                if (remap_pmd_range(mm, pud, addr, next,
2251                                pfn + (addr >> PAGE_SHIFT), prot))
2252                        return -ENOMEM;
2253        } while (pud++, addr = next, addr != end);
2254        return 0;
2255}
2256
2257/**
2258 * remap_pfn_range - remap kernel memory to userspace
2259 * @vma: user vma to map to
2260 * @addr: target user address to start at
2261 * @pfn: physical address of kernel memory
2262 * @size: size of map area
2263 * @prot: page protection flags for this mapping
2264 *
2265 *  Note: this is only safe if the mm semaphore is held when called.
2266 */
2267int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2268                    unsigned long pfn, unsigned long size, pgprot_t prot)
2269{
2270        pgd_t *pgd;
2271        unsigned long next;
2272        unsigned long end = addr + PAGE_ALIGN(size);
2273        struct mm_struct *mm = vma->vm_mm;
2274        int err;
2275
2276        /*
2277         * Physically remapped pages are special. Tell the
2278         * rest of the world about it:
2279         *   VM_IO tells people not to look at these pages
2280         *      (accesses can have side effects).
2281         *   VM_RESERVED is specified all over the place, because
2282         *      in 2.4 it kept swapout's vma scan off this vma; but
2283         *      in 2.6 the LRU scan won't even find its pages, so this
2284         *      flag means no more than count its pages in reserved_vm,
2285         *      and omit it from core dump, even when VM_IO turned off.
2286         *   VM_PFNMAP tells the core MM that the base pages are just
2287         *      raw PFN mappings, and do not have a "struct page" associated
2288         *      with them.
2289         *
2290         * There's a horrible special case to handle copy-on-write
2291         * behaviour that some programs depend on. We mark the "original"
2292         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2293         */
2294        if (addr == vma->vm_start && end == vma->vm_end) {
2295                vma->vm_pgoff = pfn;
2296                vma->vm_flags |= VM_PFN_AT_MMAP;
2297        } else if (is_cow_mapping(vma->vm_flags))
2298                return -EINVAL;
2299
2300        vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2301
2302        err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2303        if (err) {
2304                /*
2305                 * To indicate that track_pfn related cleanup is not
2306                 * needed from higher level routine calling unmap_vmas
2307                 */
2308                vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2309                vma->vm_flags &= ~VM_PFN_AT_MMAP;
2310                return -EINVAL;
2311        }
2312
2313        BUG_ON(addr >= end);
2314        pfn -= addr >> PAGE_SHIFT;
2315        pgd = pgd_offset(mm, addr);
2316        flush_cache_range(vma, addr, end);
2317        do {
2318                next = pgd_addr_end(addr, end);
2319                err = remap_pud_range(mm, pgd, addr, next,
2320                                pfn + (addr >> PAGE_SHIFT), prot);
2321                if (err)
2322                        break;
2323        } while (pgd++, addr = next, addr != end);
2324
2325        if (err)
2326                untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2327
2328        return err;
2329}
2330EXPORT_SYMBOL(remap_pfn_range);
2331
2332static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2333                                     unsigned long addr, unsigned long end,
2334                                     pte_fn_t fn, void *data)
2335{
2336        pte_t *pte;
2337        int err;
2338        pgtable_t token;
2339        spinlock_t *uninitialized_var(ptl);
2340
2341        pte = (mm == &init_mm) ?
2342                pte_alloc_kernel(pmd, addr) :
2343                pte_alloc_map_lock(mm, pmd, addr, &ptl);
2344        if (!pte)
2345                return -ENOMEM;
2346
2347        BUG_ON(pmd_huge(*pmd));
2348
2349        arch_enter_lazy_mmu_mode();
2350
2351        token = pmd_pgtable(*pmd);
2352
2353        do {
2354                err = fn(pte++, token, addr, data);
2355                if (err)
2356                        break;
2357        } while (addr += PAGE_SIZE, addr != end);
2358
2359        arch_leave_lazy_mmu_mode();
2360
2361        if (mm != &init_mm)
2362                pte_unmap_unlock(pte-1, ptl);
2363        return err;
2364}
2365
2366static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2367                                     unsigned long addr, unsigned long end,
2368                                     pte_fn_t fn, void *data)
2369{
2370        pmd_t *pmd;
2371        unsigned long next;
2372        int err;
2373
2374        BUG_ON(pud_huge(*pud));
2375
2376        pmd = pmd_alloc(mm, pud, addr);
2377        if (!pmd)
2378                return -ENOMEM;
2379        do {
2380                next = pmd_addr_end(addr, end);
2381                err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2382                if (err)
2383                        break;
2384        } while (pmd++, addr = next, addr != end);
2385        return err;
2386}
2387
2388static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2389                                     unsigned long addr, unsigned long end,
2390                                     pte_fn_t fn, void *data)
2391{
2392        pud_t *pud;
2393        unsigned long next;
2394        int err;
2395
2396        pud = pud_alloc(mm, pgd, addr);
2397        if (!pud)
2398                return -ENOMEM;
2399        do {
2400                next = pud_addr_end(addr, end);
2401                err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2402                if (err)
2403                        break;
2404        } while (pud++, addr = next, addr != end);
2405        return err;
2406}
2407
2408/*
2409 * Scan a region of virtual memory, filling in page tables as necessary
2410 * and calling a provided function on each leaf page table.
2411 */
2412int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2413                        unsigned long size, pte_fn_t fn, void *data)
2414{
2415        pgd_t *pgd;
2416        unsigned long next;
2417        unsigned long end = addr + size;
2418        int err;
2419
2420        BUG_ON(addr >= end);
2421        pgd = pgd_offset(mm, addr);
2422        do {
2423                next = pgd_addr_end(addr, end);
2424                err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2425                if (err)
2426                        break;
2427        } while (pgd++, addr = next, addr != end);
2428
2429        return err;
2430}
2431EXPORT_SYMBOL_GPL(apply_to_page_range);
2432
2433/*
2434 * handle_pte_fault chooses page fault handler according to an entry
2435 * which was read non-atomically.  Before making any commitment, on
2436 * those architectures or configurations (e.g. i386 with PAE) which
2437 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2438 * must check under lock before unmapping the pte and proceeding
2439 * (but do_wp_page is only called after already making such a check;
2440 * and do_anonymous_page can safely check later on).
2441 */
2442static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2443                                pte_t *page_table, pte_t orig_pte)
2444{
2445        int same = 1;
2446#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2447        if (sizeof(pte_t) > sizeof(unsigned long)) {
2448                spinlock_t *ptl = pte_lockptr(mm, pmd);
2449                spin_lock(ptl);
2450                same = pte_same(*page_table, orig_pte);
2451                spin_unlock(ptl);
2452        }
2453#endif
2454        pte_unmap(page_table);
2455        return same;
2456}
2457
2458static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2459{
2460        /*
2461         * If the source page was a PFN mapping, we don't have
2462         * a "struct page" for it. We do a best-effort copy by
2463         * just copying from the original user address. If that
2464         * fails, we just zero-fill it. Live with it.
2465         */
2466        if (unlikely(!src)) {
2467                void *kaddr = kmap_atomic(dst);
2468                void __user *uaddr = (void __user *)(va & PAGE_MASK);
2469
2470                /*
2471                 * This really shouldn't fail, because the page is there
2472                 * in the page tables. But it might just be unreadable,
2473                 * in which case we just give up and fill the result with
2474                 * zeroes.
2475                 */
2476                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2477                        clear_page(kaddr);
2478                kunmap_atomic(kaddr);
2479                flush_dcache_page(dst);
2480        } else
2481                copy_user_highpage(dst, src, va, vma);
2482}
2483
2484/*
2485 * This routine handles present pages, when users try to write
2486 * to a shared page. It is done by copying the page to a new address
2487 * and decrementing the shared-page counter for the old page.
2488 *
2489 * Note that this routine assumes that the protection checks have been
2490 * done by the caller (the low-level page fault routine in most cases).
2491 * Thus we can safely just mark it writable once we've done any necessary
2492 * COW.
2493 *
2494 * We also mark the page dirty at this point even though the page will
2495 * change only once the write actually happens. This avoids a few races,
2496 * and potentially makes it more efficient.
2497 *
2498 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2499 * but allow concurrent faults), with pte both mapped and locked.
2500 * We return with mmap_sem still held, but pte unmapped and unlocked.
2501 */
2502static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2503                unsigned long address, pte_t *page_table, pmd_t *pmd,
2504                spinlock_t *ptl, pte_t orig_pte)
2505        __releases(ptl)
2506{
2507        struct page *old_page, *new_page;
2508        pte_t entry;
2509        int ret = 0;
2510        int page_mkwrite = 0;
2511        struct page *dirty_page = NULL;
2512
2513        old_page = vm_normal_page(vma, address, orig_pte);
2514        if (!old_page) {
2515                /*
2516                 * VM_MIXEDMAP !pfn_valid() case
2517                 *
2518                 * We should not cow pages in a shared writeable mapping.
2519                 * Just mark the pages writable as we can't do any dirty
2520                 * accounting on raw pfn maps.
2521                 */
2522                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2523                                     (VM_WRITE|VM_SHARED))
2524                        goto reuse;
2525                goto gotten;
2526        }
2527
2528        /*
2529         * Take out anonymous pages first, anonymous shared vmas are
2530         * not dirty accountable.
2531         */
2532        if (PageAnon(old_page) && !PageKsm(old_page)) {
2533                if (!trylock_page(old_page)) {
2534                        page_cache_get(old_page);
2535                        pte_unmap_unlock(page_table, ptl);
2536                        lock_page(old_page);
2537                        page_table = pte_offset_map_lock(mm, pmd, address,
2538                                                         &ptl);
2539                        if (!pte_same(*page_table, orig_pte)) {
2540                                unlock_page(old_page);
2541                                goto unlock;
2542                        }
2543                        page_cache_release(old_page);
2544                }
2545                if (reuse_swap_page(old_page)) {
2546                        /*
2547                         * The page is all ours.  Move it to our anon_vma so
2548                         * the rmap code will not search our parent or siblings.
2549                         * Protected against the rmap code by the page lock.
2550                         */
2551                        page_move_anon_rmap(old_page, vma, address);
2552                        unlock_page(old_page);
2553                        goto reuse;
2554                }
2555                unlock_page(old_page);
2556        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2557                                        (VM_WRITE|VM_SHARED))) {
2558                /*
2559                 * Only catch write-faults on shared writable pages,
2560                 * read-only shared pages can get COWed by
2561                 * get_user_pages(.write=1, .force=1).
2562                 */
2563                if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2564                        struct vm_fault vmf;
2565                        int tmp;
2566
2567                        vmf.virtual_address = (void __user *)(address &
2568                                                                PAGE_MASK);
2569                        vmf.pgoff = old_page->index;
2570                        vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2571                        vmf.page = old_page;
2572
2573                        /*
2574                         * Notify the address space that the page is about to
2575                         * become writable so that it can prohibit this or wait
2576                         * for the page to get into an appropriate state.
2577                         *
2578                         * We do this without the lock held, so that it can
2579                         * sleep if it needs to.
2580                         */
2581                        page_cache_get(old_page);
2582                        pte_unmap_unlock(page_table, ptl);
2583
2584                        tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2585                        if (unlikely(tmp &
2586                                        (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2587                                ret = tmp;
2588                                goto unwritable_page;
2589                        }
2590                        if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2591                                lock_page(old_page);
2592                                if (!old_page->mapping) {
2593                                        ret = 0; /* retry the fault */
2594                                        unlock_page(old_page);
2595                                        goto unwritable_page;
2596                                }
2597                        } else
2598                                VM_BUG_ON(!PageLocked(old_page));
2599
2600                        /*
2601                         * Since we dropped the lock we need to revalidate
2602                         * the PTE as someone else may have changed it.  If
2603                         * they did, we just return, as we can count on the
2604                         * MMU to tell us if they didn't also make it writable.
2605                         */
2606                        page_table = pte_offset_map_lock(mm, pmd, address,
2607                                                         &ptl);
2608                        if (!pte_same(*page_table, orig_pte)) {
2609                                unlock_page(old_page);
2610                                goto unlock;
2611                        }
2612
2613                        page_mkwrite = 1;
2614                }
2615                dirty_page = old_page;
2616                get_page(dirty_page);
2617
2618reuse:
2619                flush_cache_page(vma, address, pte_pfn(orig_pte));
2620                entry = pte_mkyoung(orig_pte);
2621                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2622                if (ptep_set_access_flags(vma, address, page_table, entry,1))
2623                        update_mmu_cache(vma, address, page_table);
2624                pte_unmap_unlock(page_table, ptl);
2625                ret |= VM_FAULT_WRITE;
2626
2627                if (!dirty_page)
2628                        return ret;
2629
2630                /*
2631                 * Yes, Virginia, this is actually required to prevent a race
2632                 * with clear_page_dirty_for_io() from clearing the page dirty
2633                 * bit after it clear all dirty ptes, but before a racing
2634                 * do_wp_page installs a dirty pte.
2635                 *
2636                 * __do_fault is protected similarly.
2637                 */
2638                if (!page_mkwrite) {
2639                        wait_on_page_locked(dirty_page);
2640                        set_page_dirty_balance(dirty_page, page_mkwrite);
2641                }
2642                put_page(dirty_page);
2643                if (page_mkwrite) {
2644                        struct address_space *mapping = dirty_page->mapping;
2645
2646                        set_page_dirty(dirty_page);
2647                        unlock_page(dirty_page);
2648                        page_cache_release(dirty_page);
2649                        if (mapping)    {
2650                                /*
2651                                 * Some device drivers do not set page.mapping
2652                                 * but still dirty their pages
2653                                 */
2654                                balance_dirty_pages_ratelimited(mapping);
2655                        }
2656                }
2657
2658                /* file_update_time outside page_lock */
2659                if (vma->vm_file)
2660                        file_update_time(vma->vm_file);
2661
2662                return ret;
2663        }
2664
2665        /*
2666         * Ok, we need to copy. Oh, well..
2667         */
2668        page_cache_get(old_page);
2669gotten:
2670        pte_unmap_unlock(page_table, ptl);
2671
2672        if (unlikely(anon_vma_prepare(vma)))
2673                goto oom;
2674
2675        if (is_zero_pfn(pte_pfn(orig_pte))) {
2676                new_page = alloc_zeroed_user_highpage_movable(vma, address);
2677                if (!new_page)
2678                        goto oom;
2679        } else {
2680                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2681                if (!new_page)
2682                        goto oom;
2683                cow_user_page(new_page, old_page, address, vma);
2684        }
2685        __SetPageUptodate(new_page);
2686
2687        if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2688                goto oom_free_new;
2689
2690        /*
2691         * Re-check the pte - we dropped the lock
2692         */
2693        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2694        if (likely(pte_same(*page_table, orig_pte))) {
2695                if (old_page) {
2696                        if (!PageAnon(old_page)) {
2697                                dec_mm_counter_fast(mm, MM_FILEPAGES);
2698                                inc_mm_counter_fast(mm, MM_ANONPAGES);
2699                        }
2700                } else
2701                        inc_mm_counter_fast(mm, MM_ANONPAGES);
2702                flush_cache_page(vma, address, pte_pfn(orig_pte));
2703                entry = mk_pte(new_page, vma->vm_page_prot);
2704                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2705                /*
2706                 * Clear the pte entry and flush it first, before updating the
2707                 * pte with the new entry. This will avoid a race condition
2708                 * seen in the presence of one thread doing SMC and another
2709                 * thread doing COW.
2710                 */
2711                ptep_clear_flush(vma, address, page_table);
2712                page_add_new_anon_rmap(new_page, vma, address);
2713                /*
2714                 * We call the notify macro here because, when using secondary
2715                 * mmu page tables (such as kvm shadow page tables), we want the
2716                 * new page to be mapped directly into the secondary page table.
2717                 */
2718                set_pte_at_notify(mm, address, page_table, entry);
2719                update_mmu_cache(vma, address, page_table);
2720                if (old_page) {
2721                        /*
2722                         * Only after switching the pte to the new page may
2723                         * we remove the mapcount here. Otherwise another
2724                         * process may come and find the rmap count decremented
2725                         * before the pte is switched to the new page, and
2726                         * "reuse" the old page writing into it while our pte
2727                         * here still points into it and can be read by other
2728                         * threads.
2729                         *
2730                         * The critical issue is to order this
2731                         * page_remove_rmap with the ptp_clear_flush above.
2732                         * Those stores are ordered by (if nothing else,)
2733                         * the barrier present in the atomic_add_negative
2734                         * in page_remove_rmap.
2735                         *
2736                         * Then the TLB flush in ptep_clear_flush ensures that
2737                         * no process can access the old page before the
2738                         * decremented mapcount is visible. And the old page
2739                         * cannot be reused until after the decremented
2740                         * mapcount is visible. So transitively, TLBs to
2741                         * old page will be flushed before it can be reused.
2742                         */
2743                        page_remove_rmap(old_page);
2744                }
2745
2746                /* Free the old page.. */
2747                new_page = old_page;
2748                ret |= VM_FAULT_WRITE;
2749        } else
2750                mem_cgroup_uncharge_page(new_page);
2751
2752        if (new_page)
2753                page_cache_release(new_page);
2754unlock:
2755        pte_unmap_unlock(page_table, ptl);
2756        if (old_page) {
2757                /*
2758                 * Don't let another task, with possibly unlocked vma,
2759                 * keep the mlocked page.
2760                 */
2761                if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2762                        lock_page(old_page);    /* LRU manipulation */
2763                        munlock_vma_page(old_page);
2764                        unlock_page(old_page);
2765                }
2766                page_cache_release(old_page);
2767        }
2768        return ret;
2769oom_free_new:
2770        page_cache_release(new_page);
2771oom:
2772        if (old_page) {
2773                if (page_mkwrite) {
2774                        unlock_page(old_page);
2775                        page_cache_release(old_page);
2776                }
2777                page_cache_release(old_page);
2778        }
2779        return VM_FAULT_OOM;
2780
2781unwritable_page:
2782        page_cache_release(old_page);
2783        return ret;
2784}
2785
2786static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2787                unsigned long start_addr, unsigned long end_addr,
2788                struct zap_details *details)
2789{
2790        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2791}
2792
2793static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2794                                            struct zap_details *details)
2795{
2796        struct vm_area_struct *vma;
2797        struct prio_tree_iter iter;
2798        pgoff_t vba, vea, zba, zea;
2799
2800        vma_prio_tree_foreach(vma, &iter, root,
2801                        details->first_index, details->last_index) {
2802
2803                vba = vma->vm_pgoff;
2804                vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2805                /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2806                zba = details->first_index;
2807                if (zba < vba)
2808                        zba = vba;
2809                zea = details->last_index;
2810                if (zea > vea)
2811                        zea = vea;
2812
2813                unmap_mapping_range_vma(vma,
2814                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2815                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2816                                details);
2817        }
2818}
2819
2820static inline void unmap_mapping_range_list(struct list_head *head,
2821                                            struct zap_details *details)
2822{
2823        struct vm_area_struct *vma;
2824
2825        /*
2826         * In nonlinear VMAs there is no correspondence between virtual address
2827         * offset and file offset.  So we must perform an exhaustive search
2828         * across *all* the pages in each nonlinear VMA, not just the pages
2829         * whose virtual address lies outside the file truncation point.
2830         */
2831        list_for_each_entry(vma, head, shared.vm_set.list) {
2832                details->nonlinear_vma = vma;
2833                unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2834        }
2835}
2836
2837/**
2838 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2839 * @mapping: the address space containing mmaps to be unmapped.
2840 * @holebegin: byte in first page to unmap, relative to the start of
2841 * the underlying file.  This will be rounded down to a PAGE_SIZE
2842 * boundary.  Note that this is different from truncate_pagecache(), which
2843 * must keep the partial page.  In contrast, we must get rid of
2844 * partial pages.
2845 * @holelen: size of prospective hole in bytes.  This will be rounded
2846 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2847 * end of the file.
2848 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2849 * but 0 when invalidating pagecache, don't throw away private data.
2850 */
2851void unmap_mapping_range(struct address_space *mapping,
2852                loff_t const holebegin, loff_t const holelen, int even_cows)
2853{
2854        struct zap_details details;
2855        pgoff_t hba = holebegin >> PAGE_SHIFT;
2856        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857
2858        /* Check for overflow. */
2859        if (sizeof(holelen) > sizeof(hlen)) {
2860                long long holeend =
2861                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2862                if (holeend & ~(long long)ULONG_MAX)
2863                        hlen = ULONG_MAX - hba + 1;
2864        }
2865
2866        details.check_mapping = even_cows? NULL: mapping;
2867        details.nonlinear_vma = NULL;
2868        details.first_index = hba;
2869        details.last_index = hba + hlen - 1;
2870        if (details.last_index < details.first_index)
2871                details.last_index = ULONG_MAX;
2872
2873
2874        mutex_lock(&mapping->i_mmap_mutex);
2875        if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2876                unmap_mapping_range_tree(&mapping->i_mmap, &details);
2877        if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2878                unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2879        mutex_unlock(&mapping->i_mmap_mutex);
2880}
2881EXPORT_SYMBOL(unmap_mapping_range);
2882
2883/*
2884 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2885 * but allow concurrent faults), and pte mapped but not yet locked.
2886 * We return with mmap_sem still held, but pte unmapped and unlocked.
2887 */
2888static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2889                unsigned long address, pte_t *page_table, pmd_t *pmd,
2890                unsigned int flags, pte_t orig_pte)
2891{
2892        spinlock_t *ptl;
2893        struct page *page, *swapcache = NULL;
2894        swp_entry_t entry;
2895        pte_t pte;
2896        int locked;
2897        struct mem_cgroup *ptr;
2898        int exclusive = 0;
2899        int ret = 0;
2900
2901        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2902                goto out;
2903
2904        entry = pte_to_swp_entry(orig_pte);
2905        if (unlikely(non_swap_entry(entry))) {
2906                if (is_migration_entry(entry)) {
2907                        migration_entry_wait(mm, pmd, address);
2908                } else if (is_hwpoison_entry(entry)) {
2909                        ret = VM_FAULT_HWPOISON;
2910                } else {
2911                        print_bad_pte(vma, address, orig_pte, NULL);
2912                        ret = VM_FAULT_SIGBUS;
2913                }
2914                goto out;
2915        }
2916        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2917        page = lookup_swap_cache(entry);
2918        if (!page) {
2919                page = swapin_readahead(entry,
2920                                        GFP_HIGHUSER_MOVABLE, vma, address);
2921                if (!page) {
2922                        /*
2923                         * Back out if somebody else faulted in this pte
2924                         * while we released the pte lock.
2925                         */
2926                        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2927                        if (likely(pte_same(*page_table, orig_pte)))
2928                                ret = VM_FAULT_OOM;
2929                        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2930                        goto unlock;
2931                }
2932
2933                /* Had to read the page from swap area: Major fault */
2934                ret = VM_FAULT_MAJOR;
2935                count_vm_event(PGMAJFAULT);
2936                mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2937        } else if (PageHWPoison(page)) {
2938                /*
2939                 * hwpoisoned dirty swapcache pages are kept for killing
2940                 * owner processes (which may be unknown at hwpoison time)
2941                 */
2942                ret = VM_FAULT_HWPOISON;
2943                delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2944                goto out_release;
2945        }
2946
2947        locked = lock_page_or_retry(page, mm, flags);
2948
2949        delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2950        if (!locked) {
2951                ret |= VM_FAULT_RETRY;
2952                goto out_release;
2953        }
2954
2955        /*
2956         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2957         * release the swapcache from under us.  The page pin, and pte_same
2958         * test below, are not enough to exclude that.  Even if it is still
2959         * swapcache, we need to check that the page's swap has not changed.
2960         */
2961        if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2962                goto out_page;
2963
2964        if (ksm_might_need_to_copy(page, vma, address)) {
2965                swapcache = page;
2966                page = ksm_does_need_to_copy(page, vma, address);
2967
2968                if (unlikely(!page)) {
2969                        ret = VM_FAULT_OOM;
2970                        page = swapcache;
2971                        swapcache = NULL;
2972                        goto out_page;
2973                }
2974        }
2975
2976        if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2977                ret = VM_FAULT_OOM;
2978                goto out_page;
2979        }
2980
2981        /*
2982         * Back out if somebody else already faulted in this pte.
2983         */
2984        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2985        if (unlikely(!pte_same(*page_table, orig_pte)))
2986                goto out_nomap;
2987
2988        if (unlikely(!PageUptodate(page))) {
2989                ret = VM_FAULT_SIGBUS;
2990                goto out_nomap;
2991        }
2992
2993        /*
2994         * The page isn't present yet, go ahead with the fault.
2995         *
2996         * Be careful about the sequence of operations here.
2997         * To get its accounting right, reuse_swap_page() must be called
2998         * while the page is counted on swap but not yet in mapcount i.e.
2999         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3000         * must be called after the swap_free(), or it will never succeed.
3001         * Because delete_from_swap_page() may be called by reuse_swap_page(),
3002         * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3003         * in page->private. In this case, a record in swap_cgroup  is silently
3004         * discarded at swap_free().
3005         */
3006
3007        inc_mm_counter_fast(mm, MM_ANONPAGES);
3008        dec_mm_counter_fast(mm, MM_SWAPENTS);
3009        pte = mk_pte(page, vma->vm_page_prot);
3010        if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3011                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3012                flags &= ~FAULT_FLAG_WRITE;
3013                ret |= VM_FAULT_WRITE;
3014                exclusive = 1;
3015        }
3016        flush_icache_page(vma, page);
3017        set_pte_at(mm, address, page_table, pte);
3018        do_page_add_anon_rmap(page, vma, address, exclusive);
3019        /* It's better to call commit-charge after rmap is established */
3020        mem_cgroup_commit_charge_swapin(page, ptr);
3021
3022        swap_free(entry);
3023        if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3024                try_to_free_swap(page);
3025        unlock_page(page);
3026        if (swapcache) {
3027                /*
3028                 * Hold the lock to avoid the swap entry to be reused
3029                 * until we take the PT lock for the pte_same() check
3030                 * (to avoid false positives from pte_same). For
3031                 * further safety release the lock after the swap_free
3032                 * so that the swap count won't change under a
3033                 * parallel locked swapcache.
3034                 */
3035                unlock_page(swapcache);
3036                page_cache_release(swapcache);
3037        }
3038
3039        if (flags & FAULT_FLAG_WRITE) {
3040                ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3041                if (ret & VM_FAULT_ERROR)
3042                        ret &= VM_FAULT_ERROR;
3043                goto out;
3044        }
3045
3046        /* No need to invalidate - it was non-present before */
3047        update_mmu_cache(vma, address, page_table);
3048unlock:
3049        pte_unmap_unlock(page_table, ptl);
3050out:
3051        return ret;
3052out_nomap:
3053        mem_cgroup_cancel_charge_swapin(ptr);
3054        pte_unmap_unlock(page_table, ptl);
3055out_page:
3056        unlock_page(page);
3057out_release:
3058        page_cache_release(page);
3059        if (swapcache) {
3060                unlock_page(swapcache);
3061                page_cache_release(swapcache);
3062        }
3063        return ret;
3064}
3065
3066/*
3067 * This is like a special single-page "expand_{down|up}wards()",
3068 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3069 * doesn't hit another vma.
3070 */
3071static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3072{
3073        address &= PAGE_MASK;
3074        if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3075                struct vm_area_struct *prev = vma->vm_prev;
3076
3077                /*
3078                 * Is there a mapping abutting this one below?
3079                 *
3080                 * That's only ok if it's the same stack mapping
3081                 * that has gotten split..
3082                 */
3083                if (prev && prev->vm_end == address)
3084                        return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3085
3086                expand_downwards(vma, address - PAGE_SIZE);
3087        }
3088        if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3089                struct vm_area_struct *next = vma->vm_next;
3090
3091                /* As VM_GROWSDOWN but s/below/above/ */
3092                if (next && next->vm_start == address + PAGE_SIZE)
3093                        return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3094
3095                expand_upwards(vma, address + PAGE_SIZE);
3096        }
3097        return 0;
3098}
3099
3100/*
3101 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3102 * but allow concurrent faults), and pte mapped but not yet locked.
3103 * We return with mmap_sem still held, but pte unmapped and unlocked.
3104 */
3105static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3106                unsigned long address, pte_t *page_table, pmd_t *pmd,
3107                unsigned int flags)
3108{
3109        struct page *page;
3110        spinlock_t *ptl;
3111        pte_t entry;
3112
3113        pte_unmap(page_table);
3114
3115        /* Check if we need to add a guard page to the stack */
3116        if (check_stack_guard_page(vma, address) < 0)
3117                return VM_FAULT_SIGBUS;
3118
3119        /* Use the zero-page for reads */
3120        if (!(flags & FAULT_FLAG_WRITE)) {
3121                entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3122                                                vma->vm_page_prot));
3123                page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3124                if (!pte_none(*page_table))
3125                        goto unlock;
3126                goto setpte;
3127        }
3128
3129        /* Allocate our own private page. */
3130        if (unlikely(anon_vma_prepare(vma)))
3131                goto oom;
3132        page = alloc_zeroed_user_highpage_movable(vma, address);
3133        if (!page)
3134                goto oom;
3135        __SetPageUptodate(page);
3136
3137        if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3138                goto oom_free_page;
3139
3140        entry = mk_pte(page, vma->vm_page_prot);
3141        if (vma->vm_flags & VM_WRITE)
3142                entry = pte_mkwrite(pte_mkdirty(entry));
3143
3144        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3145        if (!pte_none(*page_table))
3146                goto release;
3147
3148        inc_mm_counter_fast(mm, MM_ANONPAGES);
3149        page_add_new_anon_rmap(page, vma, address);
3150setpte:
3151        set_pte_at(mm, address, page_table, entry);
3152
3153        /* No need to invalidate - it was non-present before */
3154        update_mmu_cache(vma, address, page_table);
3155unlock:
3156        pte_unmap_unlock(page_table, ptl);
3157        return 0;
3158release:
3159        mem_cgroup_uncharge_page(page);
3160        page_cache_release(page);
3161        goto unlock;
3162oom_free_page:
3163        page_cache_release(page);
3164oom:
3165        return VM_FAULT_OOM;
3166}
3167
3168/*
3169 * __do_fault() tries to create a new page mapping. It aggressively
3170 * tries to share with existing pages, but makes a separate copy if
3171 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3172 * the next page fault.
3173 *
3174 * As this is called only for pages that do not currently exist, we
3175 * do not need to flush old virtual caches or the TLB.
3176 *
3177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3178 * but allow concurrent faults), and pte neither mapped nor locked.
3179 * We return with mmap_sem still held, but pte unmapped and unlocked.
3180 */
3181static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3182                unsigned long address, pmd_t *pmd,
3183                pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3184{
3185        pte_t *page_table;
3186        spinlock_t *ptl;
3187        struct page *page;
3188        struct page *cow_page;
3189        pte_t entry;
3190        int anon = 0;
3191        struct page *dirty_page = NULL;
3192        struct vm_fault vmf;
3193        int ret;
3194        int page_mkwrite = 0;
3195
3196        /*
3197         * If we do COW later, allocate page befor taking lock_page()
3198         * on the file cache page. This will reduce lock holding time.
3199         */
3200        if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3201
3202                if (unlikely(anon_vma_prepare(vma)))
3203                        return VM_FAULT_OOM;
3204
3205                cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3206                if (!cow_page)
3207                        return VM_FAULT_OOM;
3208
3209                if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3210                        page_cache_release(cow_page);
3211                        return VM_FAULT_OOM;
3212                }
3213        } else
3214                cow_page = NULL;
3215
3216        vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3217        vmf.pgoff = pgoff;
3218        vmf.flags = flags;
3219        vmf.page = NULL;
3220
3221        ret = vma->vm_ops->fault(vma, &vmf);
3222        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3223                            VM_FAULT_RETRY)))
3224                goto uncharge_out;
3225
3226        if (unlikely(PageHWPoison(vmf.page))) {
3227                if (ret & VM_FAULT_LOCKED)
3228                        unlock_page(vmf.page);
3229                ret = VM_FAULT_HWPOISON;
3230                goto uncharge_out;
3231        }
3232
3233        /*
3234         * For consistency in subsequent calls, make the faulted page always
3235         * locked.
3236         */
3237        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3238                lock_page(vmf.page);
3239        else
3240                VM_BUG_ON(!PageLocked(vmf.page));
3241
3242        /*
3243         * Should we do an early C-O-W break?
3244         */
3245        page = vmf.page;
3246        if (flags & FAULT_FLAG_WRITE) {
3247                if (!(vma->vm_flags & VM_SHARED)) {
3248                        page = cow_page;
3249                        anon = 1;
3250                        copy_user_highpage(page, vmf.page, address, vma);
3251                        __SetPageUptodate(page);
3252                } else {
3253                        /*
3254                         * If the page will be shareable, see if the backing
3255                         * address space wants to know that the page is about
3256                         * to become writable
3257                         */
3258                        if (vma->vm_ops->page_mkwrite) {
3259                                int tmp;
3260
3261                                unlock_page(page);
3262                                vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3263                                tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3264                                if (unlikely(tmp &
3265                                          (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3266                                        ret = tmp;
3267                                        goto unwritable_page;
3268                                }
3269                                if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3270                                        lock_page(page);
3271                                        if (!page->mapping) {
3272                                                ret = 0; /* retry the fault */
3273                                                unlock_page(page);
3274                                                goto unwritable_page;
3275                                        }
3276                                } else
3277                                        VM_BUG_ON(!PageLocked(page));
3278                                page_mkwrite = 1;
3279                        }
3280                }
3281
3282        }
3283
3284        page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3285
3286        /*
3287         * This silly early PAGE_DIRTY setting removes a race
3288         * due to the bad i386 page protection. But it's valid
3289         * for other architectures too.
3290         *
3291         * Note that if FAULT_FLAG_WRITE is set, we either now have
3292         * an exclusive copy of the page, or this is a shared mapping,
3293         * so we can make it writable and dirty to avoid having to
3294         * handle that later.
3295         */
3296        /* Only go through if we didn't race with anybody else... */
3297        if (likely(pte_same(*page_table, orig_pte))) {
3298                flush_icache_page(vma, page);
3299                entry = mk_pte(page, vma->vm_page_prot);
3300                if (flags & FAULT_FLAG_WRITE)
3301                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3302                if (anon) {
3303                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3304                        page_add_new_anon_rmap(page, vma, address);
3305                } else {
3306                        inc_mm_counter_fast(mm, MM_FILEPAGES);
3307                        page_add_file_rmap(page);
3308                        if (flags & FAULT_FLAG_WRITE) {
3309                                dirty_page = page;
3310                                get_page(dirty_page);
3311                        }
3312                }
3313                set_pte_at(mm, address, page_table, entry);
3314
3315                /* no need to invalidate: a not-present page won't be cached */
3316                update_mmu_cache(vma, address, page_table);
3317        } else {
3318                if (cow_page)
3319                        mem_cgroup_uncharge_page(cow_page);
3320                if (anon)
3321                        page_cache_release(page);
3322                else
3323                        anon = 1; /* no anon but release faulted_page */
3324        }
3325
3326        pte_unmap_unlock(page_table, ptl);
3327
3328        if (dirty_page) {
3329                struct address_space *mapping = page->mapping;
3330
3331                if (set_page_dirty(dirty_page))
3332                        page_mkwrite = 1;
3333                unlock_page(dirty_page);
3334                put_page(dirty_page);
3335                if (page_mkwrite && mapping) {
3336                        /*
3337                         * Some device drivers do not set page.mapping but still
3338                         * dirty their pages
3339                         */
3340                        balance_dirty_pages_ratelimited(mapping);
3341                }
3342
3343                /* file_update_time outside page_lock */
3344                if (vma->vm_file)
3345                        file_update_time(vma->vm_file);
3346        } else {
3347                unlock_page(vmf.page);
3348                if (anon)
3349                        page_cache_release(vmf.page);
3350        }
3351
3352        return ret;
3353
3354unwritable_page:
3355        page_cache_release(page);
3356        return ret;
3357uncharge_out:
3358        /* fs's fault handler get error */
3359        if (cow_page) {
3360                mem_cgroup_uncharge_page(cow_page);
3361                page_cache_release(cow_page);
3362        }
3363        return ret;
3364}
3365
3366static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3367                unsigned long address, pte_t *page_table, pmd_t *pmd,
3368                unsigned int flags, pte_t orig_pte)
3369{
3370        pgoff_t pgoff = (((address & PAGE_MASK)
3371                        - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3372
3373        pte_unmap(page_table);
3374        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3375}
3376
3377/*
3378 * Fault of a previously existing named mapping. Repopulate the pte
3379 * from the encoded file_pte if possible. This enables swappable
3380 * nonlinear vmas.
3381 *
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
3385 */
3386static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3387                unsigned long address, pte_t *page_table, pmd_t *pmd,
3388                unsigned int flags, pte_t orig_pte)
3389{
3390        pgoff_t pgoff;
3391
3392        flags |= FAULT_FLAG_NONLINEAR;
3393
3394        if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3395                return 0;
3396
3397        if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3398                /*
3399                 * Page table corrupted: show pte and kill process.
3400                 */
3401                print_bad_pte(vma, address, orig_pte, NULL);
3402                return VM_FAULT_SIGBUS;
3403        }
3404
3405        pgoff = pte_to_pgoff(orig_pte);
3406        return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3407}
3408
3409/*
3410 * These routines also need to handle stuff like marking pages dirty
3411 * and/or accessed for architectures that don't do it in hardware (most
3412 * RISC architectures).  The early dirtying is also good on the i386.
3413 *
3414 * There is also a hook called "update_mmu_cache()" that architectures
3415 * with external mmu caches can use to update those (ie the Sparc or
3416 * PowerPC hashed page tables that act as extended TLBs).
3417 *
3418 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3419 * but allow concurrent faults), and pte mapped but not yet locked.
3420 * We return with mmap_sem still held, but pte unmapped and unlocked.
3421 */
3422int handle_pte_fault(struct mm_struct *mm,
3423                     struct vm_area_struct *vma, unsigned long address,
3424                     pte_t *pte, pmd_t *pmd, unsigned int flags)
3425{
3426        pte_t entry;
3427        spinlock_t *ptl;
3428
3429        entry = *pte;
3430        if (!pte_present(entry)) {
3431                if (pte_none(entry)) {
3432                        if (vma->vm_ops) {
3433                                if (likely(vma->vm_ops->fault))
3434                                        return do_linear_fault(mm, vma, address,
3435                                                pte, pmd, flags, entry);
3436                        }
3437                        return do_anonymous_page(mm, vma, address,
3438                                                 pte, pmd, flags);
3439                }
3440                if (pte_file(entry))
3441                        return do_nonlinear_fault(mm, vma, address,
3442                                        pte, pmd, flags, entry);
3443                return do_swap_page(mm, vma, address,
3444                                        pte, pmd, flags, entry);
3445        }
3446
3447        ptl = pte_lockptr(mm, pmd);
3448        spin_lock(ptl);
3449        if (unlikely(!pte_same(*pte, entry)))
3450                goto unlock;
3451        if (flags & FAULT_FLAG_WRITE) {
3452                if (!pte_write(entry))
3453                        return do_wp_page(mm, vma, address,
3454                                        pte, pmd, ptl, entry);
3455                entry = pte_mkdirty(entry);
3456        }
3457        entry = pte_mkyoung(entry);
3458        if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3459                update_mmu_cache(vma, address, pte);
3460        } else {
3461                /*
3462                 * This is needed only for protection faults but the arch code
3463                 * is not yet telling us if this is a protection fault or not.
3464                 * This still avoids useless tlb flushes for .text page faults
3465                 * with threads.
3466                 */
3467                if (flags & FAULT_FLAG_WRITE)
3468                        flush_tlb_fix_spurious_fault(vma, address);
3469        }
3470unlock:
3471        pte_unmap_unlock(pte, ptl);
3472        return 0;
3473}
3474
3475/*
3476 * By the time we get here, we already hold the mm semaphore
3477 */
3478int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3479                unsigned long address, unsigned int flags)
3480{
3481        pgd_t *pgd;
3482        pud_t *pud;
3483        pmd_t *pmd;
3484        pte_t *pte;
3485
3486        __set_current_state(TASK_RUNNING);
3487
3488        count_vm_event(PGFAULT);
3489        mem_cgroup_count_vm_event(mm, PGFAULT);
3490
3491        /* do counter updates before entering really critical section. */
3492        check_sync_rss_stat(current);
3493
3494        if (unlikely(is_vm_hugetlb_page(vma)))
3495                return hugetlb_fault(mm, vma, address, flags);
3496
3497retry:
3498        pgd = pgd_offset(mm, address);
3499        pud = pud_alloc(mm, pgd, address);
3500        if (!pud)
3501                return VM_FAULT_OOM;
3502        pmd = pmd_alloc(mm, pud, address);
3503        if (!pmd)
3504                return VM_FAULT_OOM;
3505        if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3506                if (!vma->vm_ops)
3507                        return do_huge_pmd_anonymous_page(mm, vma, address,
3508                                                          pmd, flags);
3509        } else {
3510                pmd_t orig_pmd = *pmd;
3511                int ret;
3512
3513                barrier();
3514                if (pmd_trans_huge(orig_pmd)) {
3515                        if (flags & FAULT_FLAG_WRITE &&
3516                            !pmd_write(orig_pmd) &&
3517                            !pmd_trans_splitting(orig_pmd)) {
3518                                ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3519                                                          orig_pmd);
3520                                /*
3521                                 * If COW results in an oom, the huge pmd will
3522                                 * have been split, so retry the fault on the
3523                                 * pte for a smaller charge.
3524                                 */
3525                                if (unlikely(ret & VM_FAULT_OOM))
3526                                        goto retry;
3527                                return ret;
3528                        }
3529                        return 0;
3530                }
3531        }
3532
3533        /*
3534         * Use __pte_alloc instead of pte_alloc_map, because we can't
3535         * run pte_offset_map on the pmd, if an huge pmd could
3536         * materialize from under us from a different thread.
3537         */
3538        if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3539                return VM_FAULT_OOM;
3540        /* if an huge pmd materialized from under us just retry later */
3541        if (unlikely(pmd_trans_huge(*pmd)))
3542                return 0;
3543        /*
3544         * A regular pmd is established and it can't morph into a huge pmd
3545         * from under us anymore at this point because we hold the mmap_sem
3546         * read mode and khugepaged takes it in write mode. So now it's
3547         * safe to run pte_offset_map().
3548         */
3549        pte = pte_offset_map(pmd, address);
3550
3551        return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3552}
3553
3554#ifndef __PAGETABLE_PUD_FOLDED
3555/*
3556 * Allocate page upper directory.
3557 * We've already handled the fast-path in-line.
3558 */
3559int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3560{
3561        pud_t *new = pud_alloc_one(mm, address);
3562        if (!new)
3563                return -ENOMEM;
3564
3565        smp_wmb(); /* See comment in __pte_alloc */
3566
3567        spin_lock(&mm->page_table_lock);
3568        if (pgd_present(*pgd))          /* Another has populated it */
3569                pud_free(mm, new);
3570        else
3571                pgd_populate(mm, pgd, new);
3572        spin_unlock(&mm->page_table_lock);
3573        return 0;
3574}
3575#endif /* __PAGETABLE_PUD_FOLDED */
3576
3577#ifndef __PAGETABLE_PMD_FOLDED
3578/*
3579 * Allocate page middle directory.
3580 * We've already handled the fast-path in-line.
3581 */
3582int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3583{
3584        pmd_t *new = pmd_alloc_one(mm, address);
3585        if (!new)
3586                return -ENOMEM;
3587
3588        smp_wmb(); /* See comment in __pte_alloc */
3589
3590        spin_lock(&mm->page_table_lock);
3591#ifndef __ARCH_HAS_4LEVEL_HACK
3592        if (pud_present(*pud))          /* Another has populated it */
3593                pmd_free(mm, new);
3594        else
3595                pud_populate(mm, pud, new);
3596#else
3597        if (pgd_present(*pud))          /* Another has populated it */
3598                pmd_free(mm, new);
3599        else
3600                pgd_populate(mm, pud, new);
3601#endif /* __ARCH_HAS_4LEVEL_HACK */
3602        spin_unlock(&mm->page_table_lock);
3603        return 0;
3604}
3605#endif /* __PAGETABLE_PMD_FOLDED */
3606
3607int make_pages_present(unsigned long addr, unsigned long end)
3608{
3609        int ret, len, write;
3610        struct vm_area_struct * vma;
3611
3612        vma = find_vma(current->mm, addr);
3613        if (!vma)
3614                return -ENOMEM;
3615        /*
3616         * We want to touch writable mappings with a write fault in order
3617         * to break COW, except for shared mappings because these don't COW
3618         * and we would not want to dirty them for nothing.
3619         */
3620        write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3621        BUG_ON(addr >= end);
3622        BUG_ON(end > vma->vm_end);
3623        len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3624        ret = get_user_pages(current, current->mm, addr,
3625                        len, write, 0, NULL, NULL);
3626        if (ret < 0)
3627                return ret;
3628        return ret == len ? 0 : -EFAULT;
3629}
3630
3631#if !defined(__HAVE_ARCH_GATE_AREA)
3632
3633#if defined(AT_SYSINFO_EHDR)
3634static struct vm_area_struct gate_vma;
3635
3636static int __init gate_vma_init(void)
3637{
3638        gate_vma.vm_mm = NULL;
3639        gate_vma.vm_start = FIXADDR_USER_START;
3640        gate_vma.vm_end = FIXADDR_USER_END;
3641        gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3642        gate_vma.vm_page_prot = __P101;
3643
3644        return 0;
3645}
3646__initcall(gate_vma_init);
3647#endif
3648
3649struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3650{
3651#ifdef AT_SYSINFO_EHDR
3652        return &gate_vma;
3653#else
3654        return NULL;
3655#endif
3656}
3657
3658int in_gate_area_no_mm(unsigned long addr)
3659{
3660#ifdef AT_SYSINFO_EHDR
3661        if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3662                return 1;
3663#endif
3664        return 0;
3665}
3666
3667#endif  /* __HAVE_ARCH_GATE_AREA */
3668
3669static int __follow_pte(struct mm_struct *mm, unsigned long address,
3670                pte_t **ptepp, spinlock_t **ptlp)
3671{
3672        pgd_t *pgd;
3673        pud_t *pud;
3674        pmd_t *pmd;
3675        pte_t *ptep;
3676
3677        pgd = pgd_offset(mm, address);
3678        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3679                goto out;
3680
3681        pud = pud_offset(pgd, address);
3682        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3683                goto out;
3684
3685        pmd = pmd_offset(pud, address);
3686        VM_BUG_ON(pmd_trans_huge(*pmd));
3687        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3688                goto out;
3689
3690        /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3691        if (pmd_huge(*pmd))
3692                goto out;
3693
3694        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3695        if (!ptep)
3696                goto out;
3697        if (!pte_present(*ptep))
3698                goto unlock;
3699        *ptepp = ptep;
3700        return 0;
3701unlock:
3702        pte_unmap_unlock(ptep, *ptlp);
3703out:
3704        return -EINVAL;
3705}
3706
3707static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3708                             pte_t **ptepp, spinlock_t **ptlp)
3709{
3710        int res;
3711
3712        /* (void) is needed to make gcc happy */
3713        (void) __cond_lock(*ptlp,
3714                           !(res = __follow_pte(mm, address, ptepp, ptlp)));
3715        return res;
3716}
3717
3718/**
3719 * follow_pfn - look up PFN at a user virtual address
3720 * @vma: memory mapping
3721 * @address: user virtual address
3722 * @pfn: location to store found PFN
3723 *
3724 * Only IO mappings and raw PFN mappings are allowed.
3725 *
3726 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3727 */
3728int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3729        unsigned long *pfn)
3730{
3731        int ret = -EINVAL;
3732        spinlock_t *ptl;
3733        pte_t *ptep;
3734
3735        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3736                return ret;
3737
3738        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3739        if (ret)
3740                return ret;
3741        *pfn = pte_pfn(*ptep);
3742        pte_unmap_unlock(ptep, ptl);
3743        return 0;
3744}
3745EXPORT_SYMBOL(follow_pfn);
3746
3747#ifdef CONFIG_HAVE_IOREMAP_PROT
3748int follow_phys(struct vm_area_struct *vma,
3749                unsigned long address, unsigned int flags,
3750                unsigned long *prot, resource_size_t *phys)
3751{
3752        int ret = -EINVAL;
3753        pte_t *ptep, pte;
3754        spinlock_t *ptl;
3755
3756        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3757                goto out;
3758
3759        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3760                goto out;
3761        pte = *ptep;
3762
3763        if ((flags & FOLL_WRITE) && !pte_write(pte))
3764                goto unlock;
3765
3766        *prot = pgprot_val(pte_pgprot(pte));
3767        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3768
3769        ret = 0;
3770unlock:
3771        pte_unmap_unlock(ptep, ptl);
3772out:
3773        return ret;
3774}
3775
3776int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3777                        void *buf, int len, int write)
3778{
3779        resource_size_t phys_addr;
3780        unsigned long prot = 0;
3781        void __iomem *maddr;
3782        int offset = addr & (PAGE_SIZE-1);
3783
3784        if (follow_phys(vma, addr, write, &prot, &phys_addr))
3785                return -EINVAL;
3786
3787        maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3788        if (write)
3789                memcpy_toio(maddr + offset, buf, len);
3790        else
3791                memcpy_fromio(buf, maddr + offset, len);
3792        iounmap(maddr);
3793
3794        return len;
3795}
3796#endif
3797
3798/*
3799 * Access another process' address space as given in mm.  If non-NULL, use the
3800 * given task for page fault accounting.
3801 */
3802static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3803                unsigned long addr, void *buf, int len, int write)
3804{
3805        struct vm_area_struct *vma;
3806        void *old_buf = buf;
3807
3808        down_read(&mm->mmap_sem);
3809        /* ignore errors, just check how much was successfully transferred */
3810        while (len) {
3811                int bytes, ret, offset;
3812                void *maddr;
3813                struct page *page = NULL;
3814
3815                ret = get_user_pages(tsk, mm, addr, 1,
3816                                write, 1, &page, &vma);
3817                if (ret <= 0) {
3818                        /*
3819                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
3820                         * we can access using slightly different code.
3821                         */
3822#ifdef CONFIG_HAVE_IOREMAP_PROT
3823                        vma = find_vma(mm, addr);
3824                        if (!vma || vma->vm_start > addr)
3825                                break;
3826                        if (vma->vm_ops && vma->vm_ops->access)
3827                                ret = vma->vm_ops->access(vma, addr, buf,
3828                                                          len, write);
3829                        if (ret <= 0)
3830#endif
3831                                break;
3832                        bytes = ret;
3833                } else {
3834                        bytes = len;
3835                        offset = addr & (PAGE_SIZE-1);
3836                        if (bytes > PAGE_SIZE-offset)
3837                                bytes = PAGE_SIZE-offset;
3838
3839                        maddr = kmap(page);
3840                        if (write) {
3841                                copy_to_user_page(vma, page, addr,
3842                                                  maddr + offset, buf, bytes);
3843                                set_page_dirty_lock(page);
3844                        } else {
3845                                copy_from_user_page(vma, page, addr,
3846                                                    buf, maddr + offset, bytes);
3847                        }
3848                        kunmap(page);
3849                        page_cache_release(page);
3850                }
3851                len -= bytes;
3852                buf += bytes;
3853                addr += bytes;
3854        }
3855        up_read(&mm->mmap_sem);
3856
3857        return buf - old_buf;
3858}
3859
3860/**
3861 * access_remote_vm - access another process' address space
3862 * @mm:         the mm_struct of the target address space
3863 * @addr:       start address to access
3864 * @buf:        source or destination buffer
3865 * @len:        number of bytes to transfer
3866 * @write:      whether the access is a write
3867 *
3868 * The caller must hold a reference on @mm.
3869 */
3870int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3871                void *buf, int len, int write)
3872{
3873        return __access_remote_vm(NULL, mm, addr, buf, len, write);
3874}
3875
3876/*
3877 * Access another process' address space.
3878 * Source/target buffer must be kernel space,
3879 * Do not walk the page table directly, use get_user_pages
3880 */
3881int access_process_vm(struct task_struct *tsk, unsigned long addr,
3882                void *buf, int len, int write)
3883{
3884        struct mm_struct *mm;
3885        int ret;
3886
3887        mm = get_task_mm(tsk);
3888        if (!mm)
3889                return 0;
3890
3891        ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3892        mmput(mm);
3893
3894        return ret;
3895}
3896
3897/*
3898 * Print the name of a VMA.
3899 */
3900void print_vma_addr(char *prefix, unsigned long ip)
3901{
3902        struct mm_struct *mm = current->mm;
3903        struct vm_area_struct *vma;
3904
3905        /*
3906         * Do not print if we are in atomic
3907         * contexts (in exception stacks, etc.):
3908         */
3909        if (preempt_count())
3910                return;
3911
3912        down_read(&mm->mmap_sem);
3913        vma = find_vma(mm, ip);
3914        if (vma && vma->vm_file) {
3915                struct file *f = vma->vm_file;
3916                char *buf = (char *)__get_free_page(GFP_KERNEL);
3917                if (buf) {
3918                        char *p, *s;
3919
3920                        p = d_path(&f->f_path, buf, PAGE_SIZE);
3921                        if (IS_ERR(p))
3922                                p = "?";
3923                        s = strrchr(p, '/');
3924                        if (s)
3925                                p = s+1;
3926                        printk("%s%s[%lx+%lx]", prefix, p,
3927                                        vma->vm_start,
3928                                        vma->vm_end - vma->vm_start);
3929                        free_page((unsigned long)buf);
3930                }
3931        }
3932        up_read(&current->mm->mmap_sem);
3933}
3934
3935#ifdef CONFIG_PROVE_LOCKING
3936void might_fault(void)
3937{
3938        /*
3939         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3940         * holding the mmap_sem, this is safe because kernel memory doesn't
3941         * get paged out, therefore we'll never actually fault, and the
3942         * below annotations will generate false positives.
3943         */
3944        if (segment_eq(get_fs(), KERNEL_DS))
3945                return;
3946
3947        might_sleep();
3948        /*
3949         * it would be nicer only to annotate paths which are not under
3950         * pagefault_disable, however that requires a larger audit and
3951         * providing helpers like get_user_atomic.
3952         */
3953        if (!in_atomic() && current->mm)
3954                might_lock_read(&current->mm->mmap_sem);
3955}
3956EXPORT_SYMBOL(might_fault);
3957#endif
3958
3959#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3960static void clear_gigantic_page(struct page *page,
3961                                unsigned long addr,
3962                                unsigned int pages_per_huge_page)
3963{
3964        int i;
3965        struct page *p = page;
3966
3967        might_sleep();
3968        for (i = 0; i < pages_per_huge_page;
3969             i++, p = mem_map_next(p, page, i)) {
3970                cond_resched();
3971                clear_user_highpage(p, addr + i * PAGE_SIZE);
3972        }
3973}
3974void clear_huge_page(struct page *page,
3975                     unsigned long addr, unsigned int pages_per_huge_page)
3976{
3977        int i;
3978
3979        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3980                clear_gigantic_page(page, addr, pages_per_huge_page);
3981                return;
3982        }
3983
3984        might_sleep();
3985        for (i = 0; i < pages_per_huge_page; i++) {
3986                cond_resched();
3987                clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3988        }
3989}
3990
3991static void copy_user_gigantic_page(struct page *dst, struct page *src,
3992                                    unsigned long addr,
3993                                    struct vm_area_struct *vma,
3994                                    unsigned int pages_per_huge_page)
3995{
3996        int i;
3997        struct page *dst_base = dst;
3998        struct page *src_base = src;
3999
4000        for (i = 0; i < pages_per_huge_page; ) {
4001                cond_resched();
4002                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4003
4004                i++;
4005                dst = mem_map_next(dst, dst_base, i);
4006                src = mem_map_next(src, src_base, i);
4007        }
4008}
4009
4010void copy_user_huge_page(struct page *dst, struct page *src,
4011                         unsigned long addr, struct vm_area_struct *vma,
4012                         unsigned int pages_per_huge_page)
4013{
4014        int i;
4015
4016        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4017                copy_user_gigantic_page(dst, src, addr, vma,
4018                                        pages_per_huge_page);
4019                return;
4020        }
4021
4022        might_sleep();
4023        for (i = 0; i < pages_per_huge_page; i++) {
4024                cond_resched();
4025                copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4026        }
4027}
4028#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4029
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.