linux/mm/memory.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/memory.c
   4 *
   5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
   6 */
   7
   8/*
   9 * demand-loading started 01.12.91 - seems it is high on the list of
  10 * things wanted, and it should be easy to implement. - Linus
  11 */
  12
  13/*
  14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
  15 * pages started 02.12.91, seems to work. - Linus.
  16 *
  17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
  18 * would have taken more than the 6M I have free, but it worked well as
  19 * far as I could see.
  20 *
  21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
  22 */
  23
  24/*
  25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
  26 * thought has to go into this. Oh, well..
  27 * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
  28 *              Found it. Everything seems to work now.
  29 * 20.12.91  -  Ok, making the swap-device changeable like the root.
  30 */
  31
  32/*
  33 * 05.04.94  -  Multi-page memory management added for v1.1.
  34 *              Idea by Alex Bligh (alex@cconcepts.co.uk)
  35 *
  36 * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
  37 *              (Gerhard.Wichert@pdb.siemens.de)
  38 *
  39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
  40 */
  41
  42#include <linux/kernel_stat.h>
  43#include <linux/mm.h>
  44#include <linux/sched/mm.h>
  45#include <linux/sched/coredump.h>
  46#include <linux/sched/numa_balancing.h>
  47#include <linux/sched/task.h>
  48#include <linux/hugetlb.h>
  49#include <linux/mman.h>
  50#include <linux/swap.h>
  51#include <linux/highmem.h>
  52#include <linux/pagemap.h>
  53#include <linux/memremap.h>
  54#include <linux/ksm.h>
  55#include <linux/rmap.h>
  56#include <linux/export.h>
  57#include <linux/delayacct.h>
  58#include <linux/init.h>
  59#include <linux/pfn_t.h>
  60#include <linux/writeback.h>
  61#include <linux/memcontrol.h>
  62#include <linux/mmu_notifier.h>
  63#include <linux/swapops.h>
  64#include <linux/elf.h>
  65#include <linux/gfp.h>
  66#include <linux/migrate.h>
  67#include <linux/string.h>
  68#include <linux/debugfs.h>
  69#include <linux/userfaultfd_k.h>
  70#include <linux/dax.h>
  71#include <linux/oom.h>
  72#include <linux/numa.h>
  73#include <linux/perf_event.h>
  74#include <linux/ptrace.h>
  75#include <linux/vmalloc.h>
  76
  77#include <trace/events/kmem.h>
  78
  79#include <asm/io.h>
  80#include <asm/mmu_context.h>
  81#include <asm/pgalloc.h>
  82#include <linux/uaccess.h>
  83#include <asm/tlb.h>
  84#include <asm/tlbflush.h>
  85
  86#include "pgalloc-track.h"
  87#include "internal.h"
  88
  89#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
  90#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
  91#endif
  92
  93#ifndef CONFIG_NUMA
  94unsigned long max_mapnr;
  95EXPORT_SYMBOL(max_mapnr);
  96
  97struct page *mem_map;
  98EXPORT_SYMBOL(mem_map);
  99#endif
 100
 101/*
 102 * A number of key systems in x86 including ioremap() rely on the assumption
 103 * that high_memory defines the upper bound on direct map memory, then end
 104 * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
 105 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
 106 * and ZONE_HIGHMEM.
 107 */
 108void *high_memory;
 109EXPORT_SYMBOL(high_memory);
 110
 111/*
 112 * Randomize the address space (stacks, mmaps, brk, etc.).
 113 *
 114 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
 115 *   as ancient (libc5 based) binaries can segfault. )
 116 */
 117int randomize_va_space __read_mostly =
 118#ifdef CONFIG_COMPAT_BRK
 119                                        1;
 120#else
 121                                        2;
 122#endif
 123
 124#ifndef arch_faults_on_old_pte
 125static inline bool arch_faults_on_old_pte(void)
 126{
 127        /*
 128         * Those arches which don't have hw access flag feature need to
 129         * implement their own helper. By default, "true" means pagefault
 130         * will be hit on old pte.
 131         */
 132        return true;
 133}
 134#endif
 135
 136#ifndef arch_wants_old_prefaulted_pte
 137static inline bool arch_wants_old_prefaulted_pte(void)
 138{
 139        /*
 140         * Transitioning a PTE from 'old' to 'young' can be expensive on
 141         * some architectures, even if it's performed in hardware. By
 142         * default, "false" means prefaulted entries will be 'young'.
 143         */
 144        return false;
 145}
 146#endif
 147
 148static int __init disable_randmaps(char *s)
 149{
 150        randomize_va_space = 0;
 151        return 1;
 152}
 153__setup("norandmaps", disable_randmaps);
 154
 155unsigned long zero_pfn __read_mostly;
 156EXPORT_SYMBOL(zero_pfn);
 157
 158unsigned long highest_memmap_pfn __read_mostly;
 159
 160/*
 161 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
 162 */
 163static int __init init_zero_pfn(void)
 164{
 165        zero_pfn = page_to_pfn(ZERO_PAGE(0));
 166        return 0;
 167}
 168early_initcall(init_zero_pfn);
 169
 170void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
 171{
 172        trace_rss_stat(mm, member, count);
 173}
 174
 175#if defined(SPLIT_RSS_COUNTING)
 176
 177void sync_mm_rss(struct mm_struct *mm)
 178{
 179        int i;
 180
 181        for (i = 0; i < NR_MM_COUNTERS; i++) {
 182                if (current->rss_stat.count[i]) {
 183                        add_mm_counter(mm, i, current->rss_stat.count[i]);
 184                        current->rss_stat.count[i] = 0;
 185                }
 186        }
 187        current->rss_stat.events = 0;
 188}
 189
 190static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
 191{
 192        struct task_struct *task = current;
 193
 194        if (likely(task->mm == mm))
 195                task->rss_stat.count[member] += val;
 196        else
 197                add_mm_counter(mm, member, val);
 198}
 199#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
 200#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
 201
 202/* sync counter once per 64 page faults */
 203#define TASK_RSS_EVENTS_THRESH  (64)
 204static void check_sync_rss_stat(struct task_struct *task)
 205{
 206        if (unlikely(task != current))
 207                return;
 208        if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
 209                sync_mm_rss(task->mm);
 210}
 211#else /* SPLIT_RSS_COUNTING */
 212
 213#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
 214#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
 215
 216static void check_sync_rss_stat(struct task_struct *task)
 217{
 218}
 219
 220#endif /* SPLIT_RSS_COUNTING */
 221
 222/*
 223 * Note: this doesn't free the actual pages themselves. That
 224 * has been handled earlier when unmapping all the memory regions.
 225 */
 226static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
 227                           unsigned long addr)
 228{
 229        pgtable_t token = pmd_pgtable(*pmd);
 230        pmd_clear(pmd);
 231        pte_free_tlb(tlb, token, addr);
 232        mm_dec_nr_ptes(tlb->mm);
 233}
 234
 235static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
 236                                unsigned long addr, unsigned long end,
 237                                unsigned long floor, unsigned long ceiling)
 238{
 239        pmd_t *pmd;
 240        unsigned long next;
 241        unsigned long start;
 242
 243        start = addr;
 244        pmd = pmd_offset(pud, addr);
 245        do {
 246                next = pmd_addr_end(addr, end);
 247                if (pmd_none_or_clear_bad(pmd))
 248                        continue;
 249                free_pte_range(tlb, pmd, addr);
 250        } while (pmd++, addr = next, addr != end);
 251
 252        start &= PUD_MASK;
 253        if (start < floor)
 254                return;
 255        if (ceiling) {
 256                ceiling &= PUD_MASK;
 257                if (!ceiling)
 258                        return;
 259        }
 260        if (end - 1 > ceiling - 1)
 261                return;
 262
 263        pmd = pmd_offset(pud, start);
 264        pud_clear(pud);
 265        pmd_free_tlb(tlb, pmd, start);
 266        mm_dec_nr_pmds(tlb->mm);
 267}
 268
 269static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
 270                                unsigned long addr, unsigned long end,
 271                                unsigned long floor, unsigned long ceiling)
 272{
 273        pud_t *pud;
 274        unsigned long next;
 275        unsigned long start;
 276
 277        start = addr;
 278        pud = pud_offset(p4d, addr);
 279        do {
 280                next = pud_addr_end(addr, end);
 281                if (pud_none_or_clear_bad(pud))
 282                        continue;
 283                free_pmd_range(tlb, pud, addr, next, floor, ceiling);
 284        } while (pud++, addr = next, addr != end);
 285
 286        start &= P4D_MASK;
 287        if (start < floor)
 288                return;
 289        if (ceiling) {
 290                ceiling &= P4D_MASK;
 291                if (!ceiling)
 292                        return;
 293        }
 294        if (end - 1 > ceiling - 1)
 295                return;
 296
 297        pud = pud_offset(p4d, start);
 298        p4d_clear(p4d);
 299        pud_free_tlb(tlb, pud, start);
 300        mm_dec_nr_puds(tlb->mm);
 301}
 302
 303static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
 304                                unsigned long addr, unsigned long end,
 305                                unsigned long floor, unsigned long ceiling)
 306{
 307        p4d_t *p4d;
 308        unsigned long next;
 309        unsigned long start;
 310
 311        start = addr;
 312        p4d = p4d_offset(pgd, addr);
 313        do {
 314                next = p4d_addr_end(addr, end);
 315                if (p4d_none_or_clear_bad(p4d))
 316                        continue;
 317                free_pud_range(tlb, p4d, addr, next, floor, ceiling);
 318        } while (p4d++, addr = next, addr != end);
 319
 320        start &= PGDIR_MASK;
 321        if (start < floor)
 322                return;
 323        if (ceiling) {
 324                ceiling &= PGDIR_MASK;
 325                if (!ceiling)
 326                        return;
 327        }
 328        if (end - 1 > ceiling - 1)
 329                return;
 330
 331        p4d = p4d_offset(pgd, start);
 332        pgd_clear(pgd);
 333        p4d_free_tlb(tlb, p4d, start);
 334}
 335
 336/*
 337 * This function frees user-level page tables of a process.
 338 */
 339void free_pgd_range(struct mmu_gather *tlb,
 340                        unsigned long addr, unsigned long end,
 341                        unsigned long floor, unsigned long ceiling)
 342{
 343        pgd_t *pgd;
 344        unsigned long next;
 345
 346        /*
 347         * The next few lines have given us lots of grief...
 348         *
 349         * Why are we testing PMD* at this top level?  Because often
 350         * there will be no work to do at all, and we'd prefer not to
 351         * go all the way down to the bottom just to discover that.
 352         *
 353         * Why all these "- 1"s?  Because 0 represents both the bottom
 354         * of the address space and the top of it (using -1 for the
 355         * top wouldn't help much: the masks would do the wrong thing).
 356         * The rule is that addr 0 and floor 0 refer to the bottom of
 357         * the address space, but end 0 and ceiling 0 refer to the top
 358         * Comparisons need to use "end - 1" and "ceiling - 1" (though
 359         * that end 0 case should be mythical).
 360         *
 361         * Wherever addr is brought up or ceiling brought down, we must
 362         * be careful to reject "the opposite 0" before it confuses the
 363         * subsequent tests.  But what about where end is brought down
 364         * by PMD_SIZE below? no, end can't go down to 0 there.
 365         *
 366         * Whereas we round start (addr) and ceiling down, by different
 367         * masks at different levels, in order to test whether a table
 368         * now has no other vmas using it, so can be freed, we don't
 369         * bother to round floor or end up - the tests don't need that.
 370         */
 371
 372        addr &= PMD_MASK;
 373        if (addr < floor) {
 374                addr += PMD_SIZE;
 375                if (!addr)
 376                        return;
 377        }
 378        if (ceiling) {
 379                ceiling &= PMD_MASK;
 380                if (!ceiling)
 381                        return;
 382        }
 383        if (end - 1 > ceiling - 1)
 384                end -= PMD_SIZE;
 385        if (addr > end - 1)
 386                return;
 387        /*
 388         * We add page table cache pages with PAGE_SIZE,
 389         * (see pte_free_tlb()), flush the tlb if we need
 390         */
 391        tlb_change_page_size(tlb, PAGE_SIZE);
 392        pgd = pgd_offset(tlb->mm, addr);
 393        do {
 394                next = pgd_addr_end(addr, end);
 395                if (pgd_none_or_clear_bad(pgd))
 396                        continue;
 397                free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
 398        } while (pgd++, addr = next, addr != end);
 399}
 400
 401void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
 402                unsigned long floor, unsigned long ceiling)
 403{
 404        while (vma) {
 405                struct vm_area_struct *next = vma->vm_next;
 406                unsigned long addr = vma->vm_start;
 407
 408                /*
 409                 * Hide vma from rmap and truncate_pagecache before freeing
 410                 * pgtables
 411                 */
 412                unlink_anon_vmas(vma);
 413                unlink_file_vma(vma);
 414
 415                if (is_vm_hugetlb_page(vma)) {
 416                        hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
 417                                floor, next ? next->vm_start : ceiling);
 418                } else {
 419                        /*
 420                         * Optimization: gather nearby vmas into one call down
 421                         */
 422                        while (next && next->vm_start <= vma->vm_end + PMD_SIZE
 423                               && !is_vm_hugetlb_page(next)) {
 424                                vma = next;
 425                                next = vma->vm_next;
 426                                unlink_anon_vmas(vma);
 427                                unlink_file_vma(vma);
 428                        }
 429                        free_pgd_range(tlb, addr, vma->vm_end,
 430                                floor, next ? next->vm_start : ceiling);
 431                }
 432                vma = next;
 433        }
 434}
 435
 436int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
 437{
 438        spinlock_t *ptl;
 439        pgtable_t new = pte_alloc_one(mm);
 440        if (!new)
 441                return -ENOMEM;
 442
 443        /*
 444         * Ensure all pte setup (eg. pte page lock and page clearing) are
 445         * visible before the pte is made visible to other CPUs by being
 446         * put into page tables.
 447         *
 448         * The other side of the story is the pointer chasing in the page
 449         * table walking code (when walking the page table without locking;
 450         * ie. most of the time). Fortunately, these data accesses consist
 451         * of a chain of data-dependent loads, meaning most CPUs (alpha
 452         * being the notable exception) will already guarantee loads are
 453         * seen in-order. See the alpha page table accessors for the
 454         * smp_rmb() barriers in page table walking code.
 455         */
 456        smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
 457
 458        ptl = pmd_lock(mm, pmd);
 459        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 460                mm_inc_nr_ptes(mm);
 461                pmd_populate(mm, pmd, new);
 462                new = NULL;
 463        }
 464        spin_unlock(ptl);
 465        if (new)
 466                pte_free(mm, new);
 467        return 0;
 468}
 469
 470int __pte_alloc_kernel(pmd_t *pmd)
 471{
 472        pte_t *new = pte_alloc_one_kernel(&init_mm);
 473        if (!new)
 474                return -ENOMEM;
 475
 476        smp_wmb(); /* See comment in __pte_alloc */
 477
 478        spin_lock(&init_mm.page_table_lock);
 479        if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
 480                pmd_populate_kernel(&init_mm, pmd, new);
 481                new = NULL;
 482        }
 483        spin_unlock(&init_mm.page_table_lock);
 484        if (new)
 485                pte_free_kernel(&init_mm, new);
 486        return 0;
 487}
 488
 489static inline void init_rss_vec(int *rss)
 490{
 491        memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
 492}
 493
 494static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
 495{
 496        int i;
 497
 498        if (current->mm == mm)
 499                sync_mm_rss(mm);
 500        for (i = 0; i < NR_MM_COUNTERS; i++)
 501                if (rss[i])
 502                        add_mm_counter(mm, i, rss[i]);
 503}
 504
 505/*
 506 * This function is called to print an error when a bad pte
 507 * is found. For example, we might have a PFN-mapped pte in
 508 * a region that doesn't allow it.
 509 *
 510 * The calling function must still handle the error.
 511 */
 512static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
 513                          pte_t pte, struct page *page)
 514{
 515        pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
 516        p4d_t *p4d = p4d_offset(pgd, addr);
 517        pud_t *pud = pud_offset(p4d, addr);
 518        pmd_t *pmd = pmd_offset(pud, addr);
 519        struct address_space *mapping;
 520        pgoff_t index;
 521        static unsigned long resume;
 522        static unsigned long nr_shown;
 523        static unsigned long nr_unshown;
 524
 525        /*
 526         * Allow a burst of 60 reports, then keep quiet for that minute;
 527         * or allow a steady drip of one report per second.
 528         */
 529        if (nr_shown == 60) {
 530                if (time_before(jiffies, resume)) {
 531                        nr_unshown++;
 532                        return;
 533                }
 534                if (nr_unshown) {
 535                        pr_alert("BUG: Bad page map: %lu messages suppressed\n",
 536                                 nr_unshown);
 537                        nr_unshown = 0;
 538                }
 539                nr_shown = 0;
 540        }
 541        if (nr_shown++ == 0)
 542                resume = jiffies + 60 * HZ;
 543
 544        mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
 545        index = linear_page_index(vma, addr);
 546
 547        pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
 548                 current->comm,
 549                 (long long)pte_val(pte), (long long)pmd_val(*pmd));
 550        if (page)
 551                dump_page(page, "bad pte");
 552        pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
 553                 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
 554        pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
 555                 vma->vm_file,
 556                 vma->vm_ops ? vma->vm_ops->fault : NULL,
 557                 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
 558                 mapping ? mapping->a_ops->readpage : NULL);
 559        dump_stack();
 560        add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
 561}
 562
 563/*
 564 * vm_normal_page -- This function gets the "struct page" associated with a pte.
 565 *
 566 * "Special" mappings do not wish to be associated with a "struct page" (either
 567 * it doesn't exist, or it exists but they don't want to touch it). In this
 568 * case, NULL is returned here. "Normal" mappings do have a struct page.
 569 *
 570 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
 571 * pte bit, in which case this function is trivial. Secondly, an architecture
 572 * may not have a spare pte bit, which requires a more complicated scheme,
 573 * described below.
 574 *
 575 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
 576 * special mapping (even if there are underlying and valid "struct pages").
 577 * COWed pages of a VM_PFNMAP are always normal.
 578 *
 579 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
 580 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
 581 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
 582 * mapping will always honor the rule
 583 *
 584 *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
 585 *
 586 * And for normal mappings this is false.
 587 *
 588 * This restricts such mappings to be a linear translation from virtual address
 589 * to pfn. To get around this restriction, we allow arbitrary mappings so long
 590 * as the vma is not a COW mapping; in that case, we know that all ptes are
 591 * special (because none can have been COWed).
 592 *
 593 *
 594 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
 595 *
 596 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
 597 * page" backing, however the difference is that _all_ pages with a struct
 598 * page (that is, those where pfn_valid is true) are refcounted and considered
 599 * normal pages by the VM. The disadvantage is that pages are refcounted
 600 * (which can be slower and simply not an option for some PFNMAP users). The
 601 * advantage is that we don't have to follow the strict linearity rule of
 602 * PFNMAP mappings in order to support COWable mappings.
 603 *
 604 */
 605struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
 606                            pte_t pte)
 607{
 608        unsigned long pfn = pte_pfn(pte);
 609
 610        if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
 611                if (likely(!pte_special(pte)))
 612                        goto check_pfn;
 613                if (vma->vm_ops && vma->vm_ops->find_special_page)
 614                        return vma->vm_ops->find_special_page(vma, addr);
 615                if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
 616                        return NULL;
 617                if (is_zero_pfn(pfn))
 618                        return NULL;
 619                if (pte_devmap(pte))
 620                        return NULL;
 621
 622                print_bad_pte(vma, addr, pte, NULL);
 623                return NULL;
 624        }
 625
 626        /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
 627
 628        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 629                if (vma->vm_flags & VM_MIXEDMAP) {
 630                        if (!pfn_valid(pfn))
 631                                return NULL;
 632                        goto out;
 633                } else {
 634                        unsigned long off;
 635                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 636                        if (pfn == vma->vm_pgoff + off)
 637                                return NULL;
 638                        if (!is_cow_mapping(vma->vm_flags))
 639                                return NULL;
 640                }
 641        }
 642
 643        if (is_zero_pfn(pfn))
 644                return NULL;
 645
 646check_pfn:
 647        if (unlikely(pfn > highest_memmap_pfn)) {
 648                print_bad_pte(vma, addr, pte, NULL);
 649                return NULL;
 650        }
 651
 652        /*
 653         * NOTE! We still have PageReserved() pages in the page tables.
 654         * eg. VDSO mappings can cause them to exist.
 655         */
 656out:
 657        return pfn_to_page(pfn);
 658}
 659
 660#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 661struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
 662                                pmd_t pmd)
 663{
 664        unsigned long pfn = pmd_pfn(pmd);
 665
 666        /*
 667         * There is no pmd_special() but there may be special pmds, e.g.
 668         * in a direct-access (dax) mapping, so let's just replicate the
 669         * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
 670         */
 671        if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
 672                if (vma->vm_flags & VM_MIXEDMAP) {
 673                        if (!pfn_valid(pfn))
 674                                return NULL;
 675                        goto out;
 676                } else {
 677                        unsigned long off;
 678                        off = (addr - vma->vm_start) >> PAGE_SHIFT;
 679                        if (pfn == vma->vm_pgoff + off)
 680                                return NULL;
 681                        if (!is_cow_mapping(vma->vm_flags))
 682                                return NULL;
 683                }
 684        }
 685
 686        if (pmd_devmap(pmd))
 687                return NULL;
 688        if (is_huge_zero_pmd(pmd))
 689                return NULL;
 690        if (unlikely(pfn > highest_memmap_pfn))
 691                return NULL;
 692
 693        /*
 694         * NOTE! We still have PageReserved() pages in the page tables.
 695         * eg. VDSO mappings can cause them to exist.
 696         */
 697out:
 698        return pfn_to_page(pfn);
 699}
 700#endif
 701
 702static void restore_exclusive_pte(struct vm_area_struct *vma,
 703                                  struct page *page, unsigned long address,
 704                                  pte_t *ptep)
 705{
 706        pte_t pte;
 707        swp_entry_t entry;
 708
 709        pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
 710        if (pte_swp_soft_dirty(*ptep))
 711                pte = pte_mksoft_dirty(pte);
 712
 713        entry = pte_to_swp_entry(*ptep);
 714        if (pte_swp_uffd_wp(*ptep))
 715                pte = pte_mkuffd_wp(pte);
 716        else if (is_writable_device_exclusive_entry(entry))
 717                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
 718
 719        set_pte_at(vma->vm_mm, address, ptep, pte);
 720
 721        /*
 722         * No need to take a page reference as one was already
 723         * created when the swap entry was made.
 724         */
 725        if (PageAnon(page))
 726                page_add_anon_rmap(page, vma, address, false);
 727        else
 728                /*
 729                 * Currently device exclusive access only supports anonymous
 730                 * memory so the entry shouldn't point to a filebacked page.
 731                 */
 732                WARN_ON_ONCE(!PageAnon(page));
 733
 734        if (vma->vm_flags & VM_LOCKED)
 735                mlock_vma_page(page);
 736
 737        /*
 738         * No need to invalidate - it was non-present before. However
 739         * secondary CPUs may have mappings that need invalidating.
 740         */
 741        update_mmu_cache(vma, address, ptep);
 742}
 743
 744/*
 745 * Tries to restore an exclusive pte if the page lock can be acquired without
 746 * sleeping.
 747 */
 748static int
 749try_restore_exclusive_pte(pte_t *src_pte, struct vm_area_struct *vma,
 750                        unsigned long addr)
 751{
 752        swp_entry_t entry = pte_to_swp_entry(*src_pte);
 753        struct page *page = pfn_swap_entry_to_page(entry);
 754
 755        if (trylock_page(page)) {
 756                restore_exclusive_pte(vma, page, addr, src_pte);
 757                unlock_page(page);
 758                return 0;
 759        }
 760
 761        return -EBUSY;
 762}
 763
 764/*
 765 * copy one vm_area from one task to the other. Assumes the page tables
 766 * already present in the new task to be cleared in the whole range
 767 * covered by this vma.
 768 */
 769
 770static unsigned long
 771copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 772                pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *dst_vma,
 773                struct vm_area_struct *src_vma, unsigned long addr, int *rss)
 774{
 775        unsigned long vm_flags = dst_vma->vm_flags;
 776        pte_t pte = *src_pte;
 777        struct page *page;
 778        swp_entry_t entry = pte_to_swp_entry(pte);
 779
 780        if (likely(!non_swap_entry(entry))) {
 781                if (swap_duplicate(entry) < 0)
 782                        return -EIO;
 783
 784                /* make sure dst_mm is on swapoff's mmlist. */
 785                if (unlikely(list_empty(&dst_mm->mmlist))) {
 786                        spin_lock(&mmlist_lock);
 787                        if (list_empty(&dst_mm->mmlist))
 788                                list_add(&dst_mm->mmlist,
 789                                                &src_mm->mmlist);
 790                        spin_unlock(&mmlist_lock);
 791                }
 792                rss[MM_SWAPENTS]++;
 793        } else if (is_migration_entry(entry)) {
 794                page = pfn_swap_entry_to_page(entry);
 795
 796                rss[mm_counter(page)]++;
 797
 798                if (is_writable_migration_entry(entry) &&
 799                                is_cow_mapping(vm_flags)) {
 800                        /*
 801                         * COW mappings require pages in both
 802                         * parent and child to be set to read.
 803                         */
 804                        entry = make_readable_migration_entry(
 805                                                        swp_offset(entry));
 806                        pte = swp_entry_to_pte(entry);
 807                        if (pte_swp_soft_dirty(*src_pte))
 808                                pte = pte_swp_mksoft_dirty(pte);
 809                        if (pte_swp_uffd_wp(*src_pte))
 810                                pte = pte_swp_mkuffd_wp(pte);
 811                        set_pte_at(src_mm, addr, src_pte, pte);
 812                }
 813        } else if (is_device_private_entry(entry)) {
 814                page = pfn_swap_entry_to_page(entry);
 815
 816                /*
 817                 * Update rss count even for unaddressable pages, as
 818                 * they should treated just like normal pages in this
 819                 * respect.
 820                 *
 821                 * We will likely want to have some new rss counters
 822                 * for unaddressable pages, at some point. But for now
 823                 * keep things as they are.
 824                 */
 825                get_page(page);
 826                rss[mm_counter(page)]++;
 827                page_dup_rmap(page, false);
 828
 829                /*
 830                 * We do not preserve soft-dirty information, because so
 831                 * far, checkpoint/restore is the only feature that
 832                 * requires that. And checkpoint/restore does not work
 833                 * when a device driver is involved (you cannot easily
 834                 * save and restore device driver state).
 835                 */
 836                if (is_writable_device_private_entry(entry) &&
 837                    is_cow_mapping(vm_flags)) {
 838                        entry = make_readable_device_private_entry(
 839                                                        swp_offset(entry));
 840                        pte = swp_entry_to_pte(entry);
 841                        if (pte_swp_uffd_wp(*src_pte))
 842                                pte = pte_swp_mkuffd_wp(pte);
 843                        set_pte_at(src_mm, addr, src_pte, pte);
 844                }
 845        } else if (is_device_exclusive_entry(entry)) {
 846                /*
 847                 * Make device exclusive entries present by restoring the
 848                 * original entry then copying as for a present pte. Device
 849                 * exclusive entries currently only support private writable
 850                 * (ie. COW) mappings.
 851                 */
 852                VM_BUG_ON(!is_cow_mapping(src_vma->vm_flags));
 853                if (try_restore_exclusive_pte(src_pte, src_vma, addr))
 854                        return -EBUSY;
 855                return -ENOENT;
 856        }
 857        if (!userfaultfd_wp(dst_vma))
 858                pte = pte_swp_clear_uffd_wp(pte);
 859        set_pte_at(dst_mm, addr, dst_pte, pte);
 860        return 0;
 861}
 862
 863/*
 864 * Copy a present and normal page if necessary.
 865 *
 866 * NOTE! The usual case is that this doesn't need to do
 867 * anything, and can just return a positive value. That
 868 * will let the caller know that it can just increase
 869 * the page refcount and re-use the pte the traditional
 870 * way.
 871 *
 872 * But _if_ we need to copy it because it needs to be
 873 * pinned in the parent (and the child should get its own
 874 * copy rather than just a reference to the same page),
 875 * we'll do that here and return zero to let the caller
 876 * know we're done.
 877 *
 878 * And if we need a pre-allocated page but don't yet have
 879 * one, return a negative error to let the preallocation
 880 * code know so that it can do so outside the page table
 881 * lock.
 882 */
 883static inline int
 884copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
 885                  pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
 886                  struct page **prealloc, pte_t pte, struct page *page)
 887{
 888        struct page *new_page;
 889
 890        /*
 891         * What we want to do is to check whether this page may
 892         * have been pinned by the parent process.  If so,
 893         * instead of wrprotect the pte on both sides, we copy
 894         * the page immediately so that we'll always guarantee
 895         * the pinned page won't be randomly replaced in the
 896         * future.
 897         *
 898         * The page pinning checks are just "has this mm ever
 899         * seen pinning", along with the (inexact) check of
 900         * the page count. That might give false positives for
 901         * for pinning, but it will work correctly.
 902         */
 903        if (likely(!page_needs_cow_for_dma(src_vma, page)))
 904                return 1;
 905
 906        new_page = *prealloc;
 907        if (!new_page)
 908                return -EAGAIN;
 909
 910        /*
 911         * We have a prealloc page, all good!  Take it
 912         * over and copy the page & arm it.
 913         */
 914        *prealloc = NULL;
 915        copy_user_highpage(new_page, page, addr, src_vma);
 916        __SetPageUptodate(new_page);
 917        page_add_new_anon_rmap(new_page, dst_vma, addr, false);
 918        lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
 919        rss[mm_counter(new_page)]++;
 920
 921        /* All done, just insert the new page copy in the child */
 922        pte = mk_pte(new_page, dst_vma->vm_page_prot);
 923        pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
 924        if (userfaultfd_pte_wp(dst_vma, *src_pte))
 925                /* Uffd-wp needs to be delivered to dest pte as well */
 926                pte = pte_wrprotect(pte_mkuffd_wp(pte));
 927        set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
 928        return 0;
 929}
 930
 931/*
 932 * Copy one pte.  Returns 0 if succeeded, or -EAGAIN if one preallocated page
 933 * is required to copy this pte.
 934 */
 935static inline int
 936copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
 937                 pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
 938                 struct page **prealloc)
 939{
 940        struct mm_struct *src_mm = src_vma->vm_mm;
 941        unsigned long vm_flags = src_vma->vm_flags;
 942        pte_t pte = *src_pte;
 943        struct page *page;
 944
 945        page = vm_normal_page(src_vma, addr, pte);
 946        if (page) {
 947                int retval;
 948
 949                retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
 950                                           addr, rss, prealloc, pte, page);
 951                if (retval <= 0)
 952                        return retval;
 953
 954                get_page(page);
 955                page_dup_rmap(page, false);
 956                rss[mm_counter(page)]++;
 957        }
 958
 959        /*
 960         * If it's a COW mapping, write protect it both
 961         * in the parent and the child
 962         */
 963        if (is_cow_mapping(vm_flags) && pte_write(pte)) {
 964                ptep_set_wrprotect(src_mm, addr, src_pte);
 965                pte = pte_wrprotect(pte);
 966        }
 967
 968        /*
 969         * If it's a shared mapping, mark it clean in
 970         * the child
 971         */
 972        if (vm_flags & VM_SHARED)
 973                pte = pte_mkclean(pte);
 974        pte = pte_mkold(pte);
 975
 976        if (!userfaultfd_wp(dst_vma))
 977                pte = pte_clear_uffd_wp(pte);
 978
 979        set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
 980        return 0;
 981}
 982
 983static inline struct page *
 984page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
 985                   unsigned long addr)
 986{
 987        struct page *new_page;
 988
 989        new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
 990        if (!new_page)
 991                return NULL;
 992
 993        if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
 994                put_page(new_page);
 995                return NULL;
 996        }
 997        cgroup_throttle_swaprate(new_page, GFP_KERNEL);
 998
 999        return new_page;
1000}
1001
1002static int
1003copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1004               pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1005               unsigned long end)
1006{
1007        struct mm_struct *dst_mm = dst_vma->vm_mm;
1008        struct mm_struct *src_mm = src_vma->vm_mm;
1009        pte_t *orig_src_pte, *orig_dst_pte;
1010        pte_t *src_pte, *dst_pte;
1011        spinlock_t *src_ptl, *dst_ptl;
1012        int progress, ret = 0;
1013        int rss[NR_MM_COUNTERS];
1014        swp_entry_t entry = (swp_entry_t){0};
1015        struct page *prealloc = NULL;
1016
1017again:
1018        progress = 0;
1019        init_rss_vec(rss);
1020
1021        dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1022        if (!dst_pte) {
1023                ret = -ENOMEM;
1024                goto out;
1025        }
1026        src_pte = pte_offset_map(src_pmd, addr);
1027        src_ptl = pte_lockptr(src_mm, src_pmd);
1028        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1029        orig_src_pte = src_pte;
1030        orig_dst_pte = dst_pte;
1031        arch_enter_lazy_mmu_mode();
1032
1033        do {
1034                /*
1035                 * We are holding two locks at this point - either of them
1036                 * could generate latencies in another task on another CPU.
1037                 */
1038                if (progress >= 32) {
1039                        progress = 0;
1040                        if (need_resched() ||
1041                            spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
1042                                break;
1043                }
1044                if (pte_none(*src_pte)) {
1045                        progress++;
1046                        continue;
1047                }
1048                if (unlikely(!pte_present(*src_pte))) {
1049                        ret = copy_nonpresent_pte(dst_mm, src_mm,
1050                                                  dst_pte, src_pte,
1051                                                  dst_vma, src_vma,
1052                                                  addr, rss);
1053                        if (ret == -EIO) {
1054                                entry = pte_to_swp_entry(*src_pte);
1055                                break;
1056                        } else if (ret == -EBUSY) {
1057                                break;
1058                        } else if (!ret) {
1059                                progress += 8;
1060                                continue;
1061                        }
1062
1063                        /*
1064                         * Device exclusive entry restored, continue by copying
1065                         * the now present pte.
1066                         */
1067                        WARN_ON_ONCE(ret != -ENOENT);
1068                }
1069                /* copy_present_pte() will clear `*prealloc' if consumed */
1070                ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
1071                                       addr, rss, &prealloc);
1072                /*
1073                 * If we need a pre-allocated page for this pte, drop the
1074                 * locks, allocate, and try again.
1075                 */
1076                if (unlikely(ret == -EAGAIN))
1077                        break;
1078                if (unlikely(prealloc)) {
1079                        /*
1080                         * pre-alloc page cannot be reused by next time so as
1081                         * to strictly follow mempolicy (e.g., alloc_page_vma()
1082                         * will allocate page according to address).  This
1083                         * could only happen if one pinned pte changed.
1084                         */
1085                        put_page(prealloc);
1086                        prealloc = NULL;
1087                }
1088                progress += 8;
1089        } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1090
1091        arch_leave_lazy_mmu_mode();
1092        spin_unlock(src_ptl);
1093        pte_unmap(orig_src_pte);
1094        add_mm_rss_vec(dst_mm, rss);
1095        pte_unmap_unlock(orig_dst_pte, dst_ptl);
1096        cond_resched();
1097
1098        if (ret == -EIO) {
1099                VM_WARN_ON_ONCE(!entry.val);
1100                if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1101                        ret = -ENOMEM;
1102                        goto out;
1103                }
1104                entry.val = 0;
1105        } else if (ret == -EBUSY) {
1106                goto out;
1107        } else if (ret ==  -EAGAIN) {
1108                prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1109                if (!prealloc)
1110                        return -ENOMEM;
1111        } else if (ret) {
1112                VM_WARN_ON_ONCE(1);
1113        }
1114
1115        /* We've captured and resolved the error. Reset, try again. */
1116        ret = 0;
1117
1118        if (addr != end)
1119                goto again;
1120out:
1121        if (unlikely(prealloc))
1122                put_page(prealloc);
1123        return ret;
1124}
1125
1126static inline int
1127copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1128               pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1129               unsigned long end)
1130{
1131        struct mm_struct *dst_mm = dst_vma->vm_mm;
1132        struct mm_struct *src_mm = src_vma->vm_mm;
1133        pmd_t *src_pmd, *dst_pmd;
1134        unsigned long next;
1135
1136        dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1137        if (!dst_pmd)
1138                return -ENOMEM;
1139        src_pmd = pmd_offset(src_pud, addr);
1140        do {
1141                next = pmd_addr_end(addr, end);
1142                if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1143                        || pmd_devmap(*src_pmd)) {
1144                        int err;
1145                        VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1146                        err = copy_huge_pmd(dst_mm, src_mm, dst_pmd, src_pmd,
1147                                            addr, dst_vma, src_vma);
1148                        if (err == -ENOMEM)
1149                                return -ENOMEM;
1150                        if (!err)
1151                                continue;
1152                        /* fall through */
1153                }
1154                if (pmd_none_or_clear_bad(src_pmd))
1155                        continue;
1156                if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1157                                   addr, next))
1158                        return -ENOMEM;
1159        } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1160        return 0;
1161}
1162
1163static inline int
1164copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1165               p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1166               unsigned long end)
1167{
1168        struct mm_struct *dst_mm = dst_vma->vm_mm;
1169        struct mm_struct *src_mm = src_vma->vm_mm;
1170        pud_t *src_pud, *dst_pud;
1171        unsigned long next;
1172
1173        dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1174        if (!dst_pud)
1175                return -ENOMEM;
1176        src_pud = pud_offset(src_p4d, addr);
1177        do {
1178                next = pud_addr_end(addr, end);
1179                if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1180                        int err;
1181
1182                        VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1183                        err = copy_huge_pud(dst_mm, src_mm,
1184                                            dst_pud, src_pud, addr, src_vma);
1185                        if (err == -ENOMEM)
1186                                return -ENOMEM;
1187                        if (!err)
1188                                continue;
1189                        /* fall through */
1190                }
1191                if (pud_none_or_clear_bad(src_pud))
1192                        continue;
1193                if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1194                                   addr, next))
1195                        return -ENOMEM;
1196        } while (dst_pud++, src_pud++, addr = next, addr != end);
1197        return 0;
1198}
1199
1200static inline int
1201copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1202               pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1203               unsigned long end)
1204{
1205        struct mm_struct *dst_mm = dst_vma->vm_mm;
1206        p4d_t *src_p4d, *dst_p4d;
1207        unsigned long next;
1208
1209        dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1210        if (!dst_p4d)
1211                return -ENOMEM;
1212        src_p4d = p4d_offset(src_pgd, addr);
1213        do {
1214                next = p4d_addr_end(addr, end);
1215                if (p4d_none_or_clear_bad(src_p4d))
1216                        continue;
1217                if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1218                                   addr, next))
1219                        return -ENOMEM;
1220        } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1221        return 0;
1222}
1223
1224int
1225copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1226{
1227        pgd_t *src_pgd, *dst_pgd;
1228        unsigned long next;
1229        unsigned long addr = src_vma->vm_start;
1230        unsigned long end = src_vma->vm_end;
1231        struct mm_struct *dst_mm = dst_vma->vm_mm;
1232        struct mm_struct *src_mm = src_vma->vm_mm;
1233        struct mmu_notifier_range range;
1234        bool is_cow;
1235        int ret;
1236
1237        /*
1238         * Don't copy ptes where a page fault will fill them correctly.
1239         * Fork becomes much lighter when there are big shared or private
1240         * readonly mappings. The tradeoff is that copy_page_range is more
1241         * efficient than faulting.
1242         */
1243        if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1244            !src_vma->anon_vma)
1245                return 0;
1246
1247        if (is_vm_hugetlb_page(src_vma))
1248                return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1249
1250        if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1251                /*
1252                 * We do not free on error cases below as remove_vma
1253                 * gets called on error from higher level routine
1254                 */
1255                ret = track_pfn_copy(src_vma);
1256                if (ret)
1257                        return ret;
1258        }
1259
1260        /*
1261         * We need to invalidate the secondary MMU mappings only when
1262         * there could be a permission downgrade on the ptes of the
1263         * parent mm. And a permission downgrade will only happen if
1264         * is_cow_mapping() returns true.
1265         */
1266        is_cow = is_cow_mapping(src_vma->vm_flags);
1267
1268        if (is_cow) {
1269                mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1270                                        0, src_vma, src_mm, addr, end);
1271                mmu_notifier_invalidate_range_start(&range);
1272                /*
1273                 * Disabling preemption is not needed for the write side, as
1274                 * the read side doesn't spin, but goes to the mmap_lock.
1275                 *
1276                 * Use the raw variant of the seqcount_t write API to avoid
1277                 * lockdep complaining about preemptibility.
1278                 */
1279                mmap_assert_write_locked(src_mm);
1280                raw_write_seqcount_begin(&src_mm->write_protect_seq);
1281        }
1282
1283        ret = 0;
1284        dst_pgd = pgd_offset(dst_mm, addr);
1285        src_pgd = pgd_offset(src_mm, addr);
1286        do {
1287                next = pgd_addr_end(addr, end);
1288                if (pgd_none_or_clear_bad(src_pgd))
1289                        continue;
1290                if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1291                                            addr, next))) {
1292                        ret = -ENOMEM;
1293                        break;
1294                }
1295        } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1296
1297        if (is_cow) {
1298                raw_write_seqcount_end(&src_mm->write_protect_seq);
1299                mmu_notifier_invalidate_range_end(&range);
1300        }
1301        return ret;
1302}
1303
1304static unsigned long zap_pte_range(struct mmu_gather *tlb,
1305                                struct vm_area_struct *vma, pmd_t *pmd,
1306                                unsigned long addr, unsigned long end,
1307                                struct zap_details *details)
1308{
1309        struct mm_struct *mm = tlb->mm;
1310        int force_flush = 0;
1311        int rss[NR_MM_COUNTERS];
1312        spinlock_t *ptl;
1313        pte_t *start_pte;
1314        pte_t *pte;
1315        swp_entry_t entry;
1316
1317        tlb_change_page_size(tlb, PAGE_SIZE);
1318again:
1319        init_rss_vec(rss);
1320        start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1321        pte = start_pte;
1322        flush_tlb_batched_pending(mm);
1323        arch_enter_lazy_mmu_mode();
1324        do {
1325                pte_t ptent = *pte;
1326                if (pte_none(ptent))
1327                        continue;
1328
1329                if (need_resched())
1330                        break;
1331
1332                if (pte_present(ptent)) {
1333                        struct page *page;
1334
1335                        page = vm_normal_page(vma, addr, ptent);
1336                        if (unlikely(details) && page) {
1337                                /*
1338                                 * unmap_shared_mapping_pages() wants to
1339                                 * invalidate cache without truncating:
1340                                 * unmap shared but keep private pages.
1341                                 */
1342                                if (details->check_mapping &&
1343                                    details->check_mapping != page_rmapping(page))
1344                                        continue;
1345                        }
1346                        ptent = ptep_get_and_clear_full(mm, addr, pte,
1347                                                        tlb->fullmm);
1348                        tlb_remove_tlb_entry(tlb, pte, addr);
1349                        if (unlikely(!page))
1350                                continue;
1351
1352                        if (!PageAnon(page)) {
1353                                if (pte_dirty(ptent)) {
1354                                        force_flush = 1;
1355                                        set_page_dirty(page);
1356                                }
1357                                if (pte_young(ptent) &&
1358                                    likely(!(vma->vm_flags & VM_SEQ_READ)))
1359                                        mark_page_accessed(page);
1360                        }
1361                        rss[mm_counter(page)]--;
1362                        page_remove_rmap(page, false);
1363                        if (unlikely(page_mapcount(page) < 0))
1364                                print_bad_pte(vma, addr, ptent, page);
1365                        if (unlikely(__tlb_remove_page(tlb, page))) {
1366                                force_flush = 1;
1367                                addr += PAGE_SIZE;
1368                                break;
1369                        }
1370                        continue;
1371                }
1372
1373                entry = pte_to_swp_entry(ptent);
1374                if (is_device_private_entry(entry) ||
1375                    is_device_exclusive_entry(entry)) {
1376                        struct page *page = pfn_swap_entry_to_page(entry);
1377
1378                        if (unlikely(details && details->check_mapping)) {
1379                                /*
1380                                 * unmap_shared_mapping_pages() wants to
1381                                 * invalidate cache without truncating:
1382                                 * unmap shared but keep private pages.
1383                                 */
1384                                if (details->check_mapping !=
1385                                    page_rmapping(page))
1386                                        continue;
1387                        }
1388
1389                        pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1390                        rss[mm_counter(page)]--;
1391
1392                        if (is_device_private_entry(entry))
1393                                page_remove_rmap(page, false);
1394
1395                        put_page(page);
1396                        continue;
1397                }
1398
1399                /* If details->check_mapping, we leave swap entries. */
1400                if (unlikely(details))
1401                        continue;
1402
1403                if (!non_swap_entry(entry))
1404                        rss[MM_SWAPENTS]--;
1405                else if (is_migration_entry(entry)) {
1406                        struct page *page;
1407
1408                        page = pfn_swap_entry_to_page(entry);
1409                        rss[mm_counter(page)]--;
1410                }
1411                if (unlikely(!free_swap_and_cache(entry)))
1412                        print_bad_pte(vma, addr, ptent, NULL);
1413                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1414        } while (pte++, addr += PAGE_SIZE, addr != end);
1415
1416        add_mm_rss_vec(mm, rss);
1417        arch_leave_lazy_mmu_mode();
1418
1419        /* Do the actual TLB flush before dropping ptl */
1420        if (force_flush)
1421                tlb_flush_mmu_tlbonly(tlb);
1422        pte_unmap_unlock(start_pte, ptl);
1423
1424        /*
1425         * If we forced a TLB flush (either due to running out of
1426         * batch buffers or because we needed to flush dirty TLB
1427         * entries before releasing the ptl), free the batched
1428         * memory too. Restart if we didn't do everything.
1429         */
1430        if (force_flush) {
1431                force_flush = 0;
1432                tlb_flush_mmu(tlb);
1433        }
1434
1435        if (addr != end) {
1436                cond_resched();
1437                goto again;
1438        }
1439
1440        return addr;
1441}
1442
1443static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1444                                struct vm_area_struct *vma, pud_t *pud,
1445                                unsigned long addr, unsigned long end,
1446                                struct zap_details *details)
1447{
1448        pmd_t *pmd;
1449        unsigned long next;
1450
1451        pmd = pmd_offset(pud, addr);
1452        do {
1453                next = pmd_addr_end(addr, end);
1454                if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1455                        if (next - addr != HPAGE_PMD_SIZE)
1456                                __split_huge_pmd(vma, pmd, addr, false, NULL);
1457                        else if (zap_huge_pmd(tlb, vma, pmd, addr))
1458                                goto next;
1459                        /* fall through */
1460                } else if (details && details->single_page &&
1461                           PageTransCompound(details->single_page) &&
1462                           next - addr == HPAGE_PMD_SIZE && pmd_none(*pmd)) {
1463                        spinlock_t *ptl = pmd_lock(tlb->mm, pmd);
1464                        /*
1465                         * Take and drop THP pmd lock so that we cannot return
1466                         * prematurely, while zap_huge_pmd() has cleared *pmd,
1467                         * but not yet decremented compound_mapcount().
1468                         */
1469                        spin_unlock(ptl);
1470                }
1471
1472                /*
1473                 * Here there can be other concurrent MADV_DONTNEED or
1474                 * trans huge page faults running, and if the pmd is
1475                 * none or trans huge it can change under us. This is
1476                 * because MADV_DONTNEED holds the mmap_lock in read
1477                 * mode.
1478                 */
1479                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1480                        goto next;
1481                next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1482next:
1483                cond_resched();
1484        } while (pmd++, addr = next, addr != end);
1485
1486        return addr;
1487}
1488
1489static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1490                                struct vm_area_struct *vma, p4d_t *p4d,
1491                                unsigned long addr, unsigned long end,
1492                                struct zap_details *details)
1493{
1494        pud_t *pud;
1495        unsigned long next;
1496
1497        pud = pud_offset(p4d, addr);
1498        do {
1499                next = pud_addr_end(addr, end);
1500                if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1501                        if (next - addr != HPAGE_PUD_SIZE) {
1502                                mmap_assert_locked(tlb->mm);
1503                                split_huge_pud(vma, pud, addr);
1504                        } else if (zap_huge_pud(tlb, vma, pud, addr))
1505                                goto next;
1506                        /* fall through */
1507                }
1508                if (pud_none_or_clear_bad(pud))
1509                        continue;
1510                next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1511next:
1512                cond_resched();
1513        } while (pud++, addr = next, addr != end);
1514
1515        return addr;
1516}
1517
1518static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1519                                struct vm_area_struct *vma, pgd_t *pgd,
1520                                unsigned long addr, unsigned long end,
1521                                struct zap_details *details)
1522{
1523        p4d_t *p4d;
1524        unsigned long next;
1525
1526        p4d = p4d_offset(pgd, addr);
1527        do {
1528                next = p4d_addr_end(addr, end);
1529                if (p4d_none_or_clear_bad(p4d))
1530                        continue;
1531                next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1532        } while (p4d++, addr = next, addr != end);
1533
1534        return addr;
1535}
1536
1537void unmap_page_range(struct mmu_gather *tlb,
1538                             struct vm_area_struct *vma,
1539                             unsigned long addr, unsigned long end,
1540                             struct zap_details *details)
1541{
1542        pgd_t *pgd;
1543        unsigned long next;
1544
1545        BUG_ON(addr >= end);
1546        tlb_start_vma(tlb, vma);
1547        pgd = pgd_offset(vma->vm_mm, addr);
1548        do {
1549                next = pgd_addr_end(addr, end);
1550                if (pgd_none_or_clear_bad(pgd))
1551                        continue;
1552                next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1553        } while (pgd++, addr = next, addr != end);
1554        tlb_end_vma(tlb, vma);
1555}
1556
1557
1558static void unmap_single_vma(struct mmu_gather *tlb,
1559                struct vm_area_struct *vma, unsigned long start_addr,
1560                unsigned long end_addr,
1561                struct zap_details *details)
1562{
1563        unsigned long start = max(vma->vm_start, start_addr);
1564        unsigned long end;
1565
1566        if (start >= vma->vm_end)
1567                return;
1568        end = min(vma->vm_end, end_addr);
1569        if (end <= vma->vm_start)
1570                return;
1571
1572        if (vma->vm_file)
1573                uprobe_munmap(vma, start, end);
1574
1575        if (unlikely(vma->vm_flags & VM_PFNMAP))
1576                untrack_pfn(vma, 0, 0);
1577
1578        if (start != end) {
1579                if (unlikely(is_vm_hugetlb_page(vma))) {
1580                        /*
1581                         * It is undesirable to test vma->vm_file as it
1582                         * should be non-null for valid hugetlb area.
1583                         * However, vm_file will be NULL in the error
1584                         * cleanup path of mmap_region. When
1585                         * hugetlbfs ->mmap method fails,
1586                         * mmap_region() nullifies vma->vm_file
1587                         * before calling this function to clean up.
1588                         * Since no pte has actually been setup, it is
1589                         * safe to do nothing in this case.
1590                         */
1591                        if (vma->vm_file) {
1592                                i_mmap_lock_write(vma->vm_file->f_mapping);
1593                                __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1594                                i_mmap_unlock_write(vma->vm_file->f_mapping);
1595                        }
1596                } else
1597                        unmap_page_range(tlb, vma, start, end, details);
1598        }
1599}
1600
1601/**
1602 * unmap_vmas - unmap a range of memory covered by a list of vma's
1603 * @tlb: address of the caller's struct mmu_gather
1604 * @vma: the starting vma
1605 * @start_addr: virtual address at which to start unmapping
1606 * @end_addr: virtual address at which to end unmapping
1607 *
1608 * Unmap all pages in the vma list.
1609 *
1610 * Only addresses between `start' and `end' will be unmapped.
1611 *
1612 * The VMA list must be sorted in ascending virtual address order.
1613 *
1614 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1615 * range after unmap_vmas() returns.  So the only responsibility here is to
1616 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1617 * drops the lock and schedules.
1618 */
1619void unmap_vmas(struct mmu_gather *tlb,
1620                struct vm_area_struct *vma, unsigned long start_addr,
1621                unsigned long end_addr)
1622{
1623        struct mmu_notifier_range range;
1624
1625        mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1626                                start_addr, end_addr);
1627        mmu_notifier_invalidate_range_start(&range);
1628        for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1629                unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1630        mmu_notifier_invalidate_range_end(&range);
1631}
1632
1633/**
1634 * zap_page_range - remove user pages in a given range
1635 * @vma: vm_area_struct holding the applicable pages
1636 * @start: starting address of pages to zap
1637 * @size: number of bytes to zap
1638 *
1639 * Caller must protect the VMA list
1640 */
1641void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1642                unsigned long size)
1643{
1644        struct mmu_notifier_range range;
1645        struct mmu_gather tlb;
1646
1647        lru_add_drain();
1648        mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1649                                start, start + size);
1650        tlb_gather_mmu(&tlb, vma->vm_mm);
1651        update_hiwater_rss(vma->vm_mm);
1652        mmu_notifier_invalidate_range_start(&range);
1653        for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1654                unmap_single_vma(&tlb, vma, start, range.end, NULL);
1655        mmu_notifier_invalidate_range_end(&range);
1656        tlb_finish_mmu(&tlb);
1657}
1658
1659/**
1660 * zap_page_range_single - remove user pages in a given range
1661 * @vma: vm_area_struct holding the applicable pages
1662 * @address: starting address of pages to zap
1663 * @size: number of bytes to zap
1664 * @details: details of shared cache invalidation
1665 *
1666 * The range must fit into one VMA.
1667 */
1668static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1669                unsigned long size, struct zap_details *details)
1670{
1671        struct mmu_notifier_range range;
1672        struct mmu_gather tlb;
1673
1674        lru_add_drain();
1675        mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1676                                address, address + size);
1677        tlb_gather_mmu(&tlb, vma->vm_mm);
1678        update_hiwater_rss(vma->vm_mm);
1679        mmu_notifier_invalidate_range_start(&range);
1680        unmap_single_vma(&tlb, vma, address, range.end, details);
1681        mmu_notifier_invalidate_range_end(&range);
1682        tlb_finish_mmu(&tlb);
1683}
1684
1685/**
1686 * zap_vma_ptes - remove ptes mapping the vma
1687 * @vma: vm_area_struct holding ptes to be zapped
1688 * @address: starting address of pages to zap
1689 * @size: number of bytes to zap
1690 *
1691 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1692 *
1693 * The entire address range must be fully contained within the vma.
1694 *
1695 */
1696void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1697                unsigned long size)
1698{
1699        if (address < vma->vm_start || address + size > vma->vm_end ||
1700                        !(vma->vm_flags & VM_PFNMAP))
1701                return;
1702
1703        zap_page_range_single(vma, address, size, NULL);
1704}
1705EXPORT_SYMBOL_GPL(zap_vma_ptes);
1706
1707static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1708{
1709        pgd_t *pgd;
1710        p4d_t *p4d;
1711        pud_t *pud;
1712        pmd_t *pmd;
1713
1714        pgd = pgd_offset(mm, addr);
1715        p4d = p4d_alloc(mm, pgd, addr);
1716        if (!p4d)
1717                return NULL;
1718        pud = pud_alloc(mm, p4d, addr);
1719        if (!pud)
1720                return NULL;
1721        pmd = pmd_alloc(mm, pud, addr);
1722        if (!pmd)
1723                return NULL;
1724
1725        VM_BUG_ON(pmd_trans_huge(*pmd));
1726        return pmd;
1727}
1728
1729pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1730                        spinlock_t **ptl)
1731{
1732        pmd_t *pmd = walk_to_pmd(mm, addr);
1733
1734        if (!pmd)
1735                return NULL;
1736        return pte_alloc_map_lock(mm, pmd, addr, ptl);
1737}
1738
1739static int validate_page_before_insert(struct page *page)
1740{
1741        if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1742                return -EINVAL;
1743        flush_dcache_page(page);
1744        return 0;
1745}
1746
1747static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1748                        unsigned long addr, struct page *page, pgprot_t prot)
1749{
1750        if (!pte_none(*pte))
1751                return -EBUSY;
1752        /* Ok, finally just insert the thing.. */
1753        get_page(page);
1754        inc_mm_counter_fast(mm, mm_counter_file(page));
1755        page_add_file_rmap(page, false);
1756        set_pte_at(mm, addr, pte, mk_pte(page, prot));
1757        return 0;
1758}
1759
1760/*
1761 * This is the old fallback for page remapping.
1762 *
1763 * For historical reasons, it only allows reserved pages. Only
1764 * old drivers should use this, and they needed to mark their
1765 * pages reserved for the old functions anyway.
1766 */
1767static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1768                        struct page *page, pgprot_t prot)
1769{
1770        struct mm_struct *mm = vma->vm_mm;
1771        int retval;
1772        pte_t *pte;
1773        spinlock_t *ptl;
1774
1775        retval = validate_page_before_insert(page);
1776        if (retval)
1777                goto out;
1778        retval = -ENOMEM;
1779        pte = get_locked_pte(mm, addr, &ptl);
1780        if (!pte)
1781                goto out;
1782        retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1783        pte_unmap_unlock(pte, ptl);
1784out:
1785        return retval;
1786}
1787
1788#ifdef pte_index
1789static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1790                        unsigned long addr, struct page *page, pgprot_t prot)
1791{
1792        int err;
1793
1794        if (!page_count(page))
1795                return -EINVAL;
1796        err = validate_page_before_insert(page);
1797        if (err)
1798                return err;
1799        return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1800}
1801
1802/* insert_pages() amortizes the cost of spinlock operations
1803 * when inserting pages in a loop. Arch *must* define pte_index.
1804 */
1805static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1806                        struct page **pages, unsigned long *num, pgprot_t prot)
1807{
1808        pmd_t *pmd = NULL;
1809        pte_t *start_pte, *pte;
1810        spinlock_t *pte_lock;
1811        struct mm_struct *const mm = vma->vm_mm;
1812        unsigned long curr_page_idx = 0;
1813        unsigned long remaining_pages_total = *num;
1814        unsigned long pages_to_write_in_pmd;
1815        int ret;
1816more:
1817        ret = -EFAULT;
1818        pmd = walk_to_pmd(mm, addr);
1819        if (!pmd)
1820                goto out;
1821
1822        pages_to_write_in_pmd = min_t(unsigned long,
1823                remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1824
1825        /* Allocate the PTE if necessary; takes PMD lock once only. */
1826        ret = -ENOMEM;
1827        if (pte_alloc(mm, pmd))
1828                goto out;
1829
1830        while (pages_to_write_in_pmd) {
1831                int pte_idx = 0;
1832                const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1833
1834                start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1835                for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1836                        int err = insert_page_in_batch_locked(mm, pte,
1837                                addr, pages[curr_page_idx], prot);
1838                        if (unlikely(err)) {
1839                                pte_unmap_unlock(start_pte, pte_lock);
1840                                ret = err;
1841                                remaining_pages_total -= pte_idx;
1842                                goto out;
1843                        }
1844                        addr += PAGE_SIZE;
1845                        ++curr_page_idx;
1846                }
1847                pte_unmap_unlock(start_pte, pte_lock);
1848                pages_to_write_in_pmd -= batch_size;
1849                remaining_pages_total -= batch_size;
1850        }
1851        if (remaining_pages_total)
1852                goto more;
1853        ret = 0;
1854out:
1855        *num = remaining_pages_total;
1856        return ret;
1857}
1858#endif  /* ifdef pte_index */
1859
1860/**
1861 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1862 * @vma: user vma to map to
1863 * @addr: target start user address of these pages
1864 * @pages: source kernel pages
1865 * @num: in: number of pages to map. out: number of pages that were *not*
1866 * mapped. (0 means all pages were successfully mapped).
1867 *
1868 * Preferred over vm_insert_page() when inserting multiple pages.
1869 *
1870 * In case of error, we may have mapped a subset of the provided
1871 * pages. It is the caller's responsibility to account for this case.
1872 *
1873 * The same restrictions apply as in vm_insert_page().
1874 */
1875int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1876                        struct page **pages, unsigned long *num)
1877{
1878#ifdef pte_index
1879        const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1880
1881        if (addr < vma->vm_start || end_addr >= vma->vm_end)
1882                return -EFAULT;
1883        if (!(vma->vm_flags & VM_MIXEDMAP)) {
1884                BUG_ON(mmap_read_trylock(vma->vm_mm));
1885                BUG_ON(vma->vm_flags & VM_PFNMAP);
1886                vma->vm_flags |= VM_MIXEDMAP;
1887        }
1888        /* Defer page refcount checking till we're about to map that page. */
1889        return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1890#else
1891        unsigned long idx = 0, pgcount = *num;
1892        int err = -EINVAL;
1893
1894        for (; idx < pgcount; ++idx) {
1895                err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1896                if (err)
1897                        break;
1898        }
1899        *num = pgcount - idx;
1900        return err;
1901#endif  /* ifdef pte_index */
1902}
1903EXPORT_SYMBOL(vm_insert_pages);
1904
1905/**
1906 * vm_insert_page - insert single page into user vma
1907 * @vma: user vma to map to
1908 * @addr: target user address of this page
1909 * @page: source kernel page
1910 *
1911 * This allows drivers to insert individual pages they've allocated
1912 * into a user vma.
1913 *
1914 * The page has to be a nice clean _individual_ kernel allocation.
1915 * If you allocate a compound page, you need to have marked it as
1916 * such (__GFP_COMP), or manually just split the page up yourself
1917 * (see split_page()).
1918 *
1919 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1920 * took an arbitrary page protection parameter. This doesn't allow
1921 * that. Your vma protection will have to be set up correctly, which
1922 * means that if you want a shared writable mapping, you'd better
1923 * ask for a shared writable mapping!
1924 *
1925 * The page does not need to be reserved.
1926 *
1927 * Usually this function is called from f_op->mmap() handler
1928 * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1929 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1930 * function from other places, for example from page-fault handler.
1931 *
1932 * Return: %0 on success, negative error code otherwise.
1933 */
1934int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1935                        struct page *page)
1936{
1937        if (addr < vma->vm_start || addr >= vma->vm_end)
1938                return -EFAULT;
1939        if (!page_count(page))
1940                return -EINVAL;
1941        if (!(vma->vm_flags & VM_MIXEDMAP)) {
1942                BUG_ON(mmap_read_trylock(vma->vm_mm));
1943                BUG_ON(vma->vm_flags & VM_PFNMAP);
1944                vma->vm_flags |= VM_MIXEDMAP;
1945        }
1946        return insert_page(vma, addr, page, vma->vm_page_prot);
1947}
1948EXPORT_SYMBOL(vm_insert_page);
1949
1950/*
1951 * __vm_map_pages - maps range of kernel pages into user vma
1952 * @vma: user vma to map to
1953 * @pages: pointer to array of source kernel pages
1954 * @num: number of pages in page array
1955 * @offset: user's requested vm_pgoff
1956 *
1957 * This allows drivers to map range of kernel pages into a user vma.
1958 *
1959 * Return: 0 on success and error code otherwise.
1960 */
1961static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1962                                unsigned long num, unsigned long offset)
1963{
1964        unsigned long count = vma_pages(vma);
1965        unsigned long uaddr = vma->vm_start;
1966        int ret, i;
1967
1968        /* Fail if the user requested offset is beyond the end of the object */
1969        if (offset >= num)
1970                return -ENXIO;
1971
1972        /* Fail if the user requested size exceeds available object size */
1973        if (count > num - offset)
1974                return -ENXIO;
1975
1976        for (i = 0; i < count; i++) {
1977                ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1978                if (ret < 0)
1979                        return ret;
1980                uaddr += PAGE_SIZE;
1981        }
1982
1983        return 0;
1984}
1985
1986/**
1987 * vm_map_pages - maps range of kernel pages starts with non zero offset
1988 * @vma: user vma to map to
1989 * @pages: pointer to array of source kernel pages
1990 * @num: number of pages in page array
1991 *
1992 * Maps an object consisting of @num pages, catering for the user's
1993 * requested vm_pgoff
1994 *
1995 * If we fail to insert any page into the vma, the function will return
1996 * immediately leaving any previously inserted pages present.  Callers
1997 * from the mmap handler may immediately return the error as their caller
1998 * will destroy the vma, removing any successfully inserted pages. Other
1999 * callers should make their own arrangements for calling unmap_region().
2000 *
2001 * Context: Process context. Called by mmap handlers.
2002 * Return: 0 on success and error code otherwise.
2003 */
2004int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
2005                                unsigned long num)
2006{
2007        return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
2008}
2009EXPORT_SYMBOL(vm_map_pages);
2010
2011/**
2012 * vm_map_pages_zero - map range of kernel pages starts with zero offset
2013 * @vma: user vma to map to
2014 * @pages: pointer to array of source kernel pages
2015 * @num: number of pages in page array
2016 *
2017 * Similar to vm_map_pages(), except that it explicitly sets the offset
2018 * to 0. This function is intended for the drivers that did not consider
2019 * vm_pgoff.
2020 *
2021 * Context: Process context. Called by mmap handlers.
2022 * Return: 0 on success and error code otherwise.
2023 */
2024int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
2025                                unsigned long num)
2026{
2027        return __vm_map_pages(vma, pages, num, 0);
2028}
2029EXPORT_SYMBOL(vm_map_pages_zero);
2030
2031static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2032                        pfn_t pfn, pgprot_t prot, bool mkwrite)
2033{
2034        struct mm_struct *mm = vma->vm_mm;
2035        pte_t *pte, entry;
2036        spinlock_t *ptl;
2037
2038        pte = get_locked_pte(mm, addr, &ptl);
2039        if (!pte)
2040                return VM_FAULT_OOM;
2041        if (!pte_none(*pte)) {
2042                if (mkwrite) {
2043                        /*
2044                         * For read faults on private mappings the PFN passed
2045                         * in may not match the PFN we have mapped if the
2046                         * mapped PFN is a writeable COW page.  In the mkwrite
2047                         * case we are creating a writable PTE for a shared
2048                         * mapping and we expect the PFNs to match. If they
2049                         * don't match, we are likely racing with block
2050                         * allocation and mapping invalidation so just skip the
2051                         * update.
2052                         */
2053                        if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
2054                                WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
2055                                goto out_unlock;
2056                        }
2057                        entry = pte_mkyoung(*pte);
2058                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2059                        if (ptep_set_access_flags(vma, addr, pte, entry, 1))
2060                                update_mmu_cache(vma, addr, pte);
2061                }
2062                goto out_unlock;
2063        }
2064
2065        /* Ok, finally just insert the thing.. */
2066        if (pfn_t_devmap(pfn))
2067                entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
2068        else
2069                entry = pte_mkspecial(pfn_t_pte(pfn, prot));
2070
2071        if (mkwrite) {
2072                entry = pte_mkyoung(entry);
2073                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2074        }
2075
2076        set_pte_at(mm, addr, pte, entry);
2077        update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2078
2079out_unlock:
2080        pte_unmap_unlock(pte, ptl);
2081        return VM_FAULT_NOPAGE;
2082}
2083
2084/**
2085 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
2086 * @vma: user vma to map to
2087 * @addr: target user address of this page
2088 * @pfn: source kernel pfn
2089 * @pgprot: pgprot flags for the inserted page
2090 *
2091 * This is exactly like vmf_insert_pfn(), except that it allows drivers
2092 * to override pgprot on a per-page basis.
2093 *
2094 * This only makes sense for IO mappings, and it makes no sense for
2095 * COW mappings.  In general, using multiple vmas is preferable;
2096 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
2097 * impractical.
2098 *
2099 * See vmf_insert_mixed_prot() for a discussion of the implication of using
2100 * a value of @pgprot different from that of @vma->vm_page_prot.
2101 *
2102 * Context: Process context.  May allocate using %GFP_KERNEL.
2103 * Return: vm_fault_t value.
2104 */
2105vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
2106                        unsigned long pfn, pgprot_t pgprot)
2107{
2108        /*
2109         * Technically, architectures with pte_special can avoid all these
2110         * restrictions (same for remap_pfn_range).  However we would like
2111         * consistency in testing and feature parity among all, so we should
2112         * try to keep these invariants in place for everybody.
2113         */
2114        BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2115        BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2116                                                (VM_PFNMAP|VM_MIXEDMAP));
2117        BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2118        BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2119
2120        if (addr < vma->vm_start || addr >= vma->vm_end)
2121                return VM_FAULT_SIGBUS;
2122
2123        if (!pfn_modify_allowed(pfn, pgprot))
2124                return VM_FAULT_SIGBUS;
2125
2126        track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2127
2128        return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2129                        false);
2130}
2131EXPORT_SYMBOL(vmf_insert_pfn_prot);
2132
2133/**
2134 * vmf_insert_pfn - insert single pfn into user vma
2135 * @vma: user vma to map to
2136 * @addr: target user address of this page
2137 * @pfn: source kernel pfn
2138 *
2139 * Similar to vm_insert_page, this allows drivers to insert individual pages
2140 * they've allocated into a user vma. Same comments apply.
2141 *
2142 * This function should only be called from a vm_ops->fault handler, and
2143 * in that case the handler should return the result of this function.
2144 *
2145 * vma cannot be a COW mapping.
2146 *
2147 * As this is called only for pages that do not currently exist, we
2148 * do not need to flush old virtual caches or the TLB.
2149 *
2150 * Context: Process context.  May allocate using %GFP_KERNEL.
2151 * Return: vm_fault_t value.
2152 */
2153vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2154                        unsigned long pfn)
2155{
2156        return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2157}
2158EXPORT_SYMBOL(vmf_insert_pfn);
2159
2160static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2161{
2162        /* these checks mirror the abort conditions in vm_normal_page */
2163        if (vma->vm_flags & VM_MIXEDMAP)
2164                return true;
2165        if (pfn_t_devmap(pfn))
2166                return true;
2167        if (pfn_t_special(pfn))
2168                return true;
2169        if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2170                return true;
2171        return false;
2172}
2173
2174static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2175                unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2176                bool mkwrite)
2177{
2178        int err;
2179
2180        BUG_ON(!vm_mixed_ok(vma, pfn));
2181
2182        if (addr < vma->vm_start || addr >= vma->vm_end)
2183                return VM_FAULT_SIGBUS;
2184
2185        track_pfn_insert(vma, &pgprot, pfn);
2186
2187        if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2188                return VM_FAULT_SIGBUS;
2189
2190        /*
2191         * If we don't have pte special, then we have to use the pfn_valid()
2192         * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2193         * refcount the page if pfn_valid is true (hence insert_page rather
2194         * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2195         * without pte special, it would there be refcounted as a normal page.
2196         */
2197        if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2198            !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2199                struct page *page;
2200
2201                /*
2202                 * At this point we are committed to insert_page()
2203                 * regardless of whether the caller specified flags that
2204                 * result in pfn_t_has_page() == false.
2205                 */
2206                page = pfn_to_page(pfn_t_to_pfn(pfn));
2207                err = insert_page(vma, addr, page, pgprot);
2208        } else {
2209                return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2210        }
2211
2212        if (err == -ENOMEM)
2213                return VM_FAULT_OOM;
2214        if (err < 0 && err != -EBUSY)
2215                return VM_FAULT_SIGBUS;
2216
2217        return VM_FAULT_NOPAGE;
2218}
2219
2220/**
2221 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2222 * @vma: user vma to map to
2223 * @addr: target user address of this page
2224 * @pfn: source kernel pfn
2225 * @pgprot: pgprot flags for the inserted page
2226 *
2227 * This is exactly like vmf_insert_mixed(), except that it allows drivers
2228 * to override pgprot on a per-page basis.
2229 *
2230 * Typically this function should be used by drivers to set caching- and
2231 * encryption bits different than those of @vma->vm_page_prot, because
2232 * the caching- or encryption mode may not be known at mmap() time.
2233 * This is ok as long as @vma->vm_page_prot is not used by the core vm
2234 * to set caching and encryption bits for those vmas (except for COW pages).
2235 * This is ensured by core vm only modifying these page table entries using
2236 * functions that don't touch caching- or encryption bits, using pte_modify()
2237 * if needed. (See for example mprotect()).
2238 * Also when new page-table entries are created, this is only done using the
2239 * fault() callback, and never using the value of vma->vm_page_prot,
2240 * except for page-table entries that point to anonymous pages as the result
2241 * of COW.
2242 *
2243 * Context: Process context.  May allocate using %GFP_KERNEL.
2244 * Return: vm_fault_t value.
2245 */
2246vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2247                                 pfn_t pfn, pgprot_t pgprot)
2248{
2249        return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2250}
2251EXPORT_SYMBOL(vmf_insert_mixed_prot);
2252
2253vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2254                pfn_t pfn)
2255{
2256        return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2257}
2258EXPORT_SYMBOL(vmf_insert_mixed);
2259
2260/*
2261 *  If the insertion of PTE failed because someone else already added a
2262 *  different entry in the mean time, we treat that as success as we assume
2263 *  the same entry was actually inserted.
2264 */
2265vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2266                unsigned long addr, pfn_t pfn)
2267{
2268        return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2269}
2270EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2271
2272/*
2273 * maps a range of physical memory into the requested pages. the old
2274 * mappings are removed. any references to nonexistent pages results
2275 * in null mappings (currently treated as "copy-on-access")
2276 */
2277static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2278                        unsigned long addr, unsigned long end,
2279                        unsigned long pfn, pgprot_t prot)
2280{
2281        pte_t *pte, *mapped_pte;
2282        spinlock_t *ptl;
2283        int err = 0;
2284
2285        mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2286        if (!pte)
2287                return -ENOMEM;
2288        arch_enter_lazy_mmu_mode();
2289        do {
2290                BUG_ON(!pte_none(*pte));
2291                if (!pfn_modify_allowed(pfn, prot)) {
2292                        err = -EACCES;
2293                        break;
2294                }
2295                set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2296                pfn++;
2297        } while (pte++, addr += PAGE_SIZE, addr != end);
2298        arch_leave_lazy_mmu_mode();
2299        pte_unmap_unlock(mapped_pte, ptl);
2300        return err;
2301}
2302
2303static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2304                        unsigned long addr, unsigned long end,
2305                        unsigned long pfn, pgprot_t prot)
2306{
2307        pmd_t *pmd;
2308        unsigned long next;
2309        int err;
2310
2311        pfn -= addr >> PAGE_SHIFT;
2312        pmd = pmd_alloc(mm, pud, addr);
2313        if (!pmd)
2314                return -ENOMEM;
2315        VM_BUG_ON(pmd_trans_huge(*pmd));
2316        do {
2317                next = pmd_addr_end(addr, end);
2318                err = remap_pte_range(mm, pmd, addr, next,
2319                                pfn + (addr >> PAGE_SHIFT), prot);
2320                if (err)
2321                        return err;
2322        } while (pmd++, addr = next, addr != end);
2323        return 0;
2324}
2325
2326static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2327                        unsigned long addr, unsigned long end,
2328                        unsigned long pfn, pgprot_t prot)
2329{
2330        pud_t *pud;
2331        unsigned long next;
2332        int err;
2333
2334        pfn -= addr >> PAGE_SHIFT;
2335        pud = pud_alloc(mm, p4d, addr);
2336        if (!pud)
2337                return -ENOMEM;
2338        do {
2339                next = pud_addr_end(addr, end);
2340                err = remap_pmd_range(mm, pud, addr, next,
2341                                pfn + (addr >> PAGE_SHIFT), prot);
2342                if (err)
2343                        return err;
2344        } while (pud++, addr = next, addr != end);
2345        return 0;
2346}
2347
2348static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2349                        unsigned long addr, unsigned long end,
2350                        unsigned long pfn, pgprot_t prot)
2351{
2352        p4d_t *p4d;
2353        unsigned long next;
2354        int err;
2355
2356        pfn -= addr >> PAGE_SHIFT;
2357        p4d = p4d_alloc(mm, pgd, addr);
2358        if (!p4d)
2359                return -ENOMEM;
2360        do {
2361                next = p4d_addr_end(addr, end);
2362                err = remap_pud_range(mm, p4d, addr, next,
2363                                pfn + (addr >> PAGE_SHIFT), prot);
2364                if (err)
2365                        return err;
2366        } while (p4d++, addr = next, addr != end);
2367        return 0;
2368}
2369
2370/*
2371 * Variant of remap_pfn_range that does not call track_pfn_remap.  The caller
2372 * must have pre-validated the caching bits of the pgprot_t.
2373 */
2374int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
2375                unsigned long pfn, unsigned long size, pgprot_t prot)
2376{
2377        pgd_t *pgd;
2378        unsigned long next;
2379        unsigned long end = addr + PAGE_ALIGN(size);
2380        struct mm_struct *mm = vma->vm_mm;
2381        int err;
2382
2383        if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2384                return -EINVAL;
2385
2386        /*
2387         * Physically remapped pages are special. Tell the
2388         * rest of the world about it:
2389         *   VM_IO tells people not to look at these pages
2390         *      (accesses can have side effects).
2391         *   VM_PFNMAP tells the core MM that the base pages are just
2392         *      raw PFN mappings, and do not have a "struct page" associated
2393         *      with them.
2394         *   VM_DONTEXPAND
2395         *      Disable vma merging and expanding with mremap().
2396         *   VM_DONTDUMP
2397         *      Omit vma from core dump, even when VM_IO turned off.
2398         *
2399         * There's a horrible special case to handle copy-on-write
2400         * behaviour that some programs depend on. We mark the "original"
2401         * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2402         * See vm_normal_page() for details.
2403         */
2404        if (is_cow_mapping(vma->vm_flags)) {
2405                if (addr != vma->vm_start || end != vma->vm_end)
2406                        return -EINVAL;
2407                vma->vm_pgoff = pfn;
2408        }
2409
2410        vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2411
2412        BUG_ON(addr >= end);
2413        pfn -= addr >> PAGE_SHIFT;
2414        pgd = pgd_offset(mm, addr);
2415        flush_cache_range(vma, addr, end);
2416        do {
2417                next = pgd_addr_end(addr, end);
2418                err = remap_p4d_range(mm, pgd, addr, next,
2419                                pfn + (addr >> PAGE_SHIFT), prot);
2420                if (err)
2421                        return err;
2422        } while (pgd++, addr = next, addr != end);
2423
2424        return 0;
2425}
2426
2427/**
2428 * remap_pfn_range - remap kernel memory to userspace
2429 * @vma: user vma to map to
2430 * @addr: target page aligned user address to start at
2431 * @pfn: page frame number of kernel physical memory address
2432 * @size: size of mapping area
2433 * @prot: page protection flags for this mapping
2434 *
2435 * Note: this is only safe if the mm semaphore is held when called.
2436 *
2437 * Return: %0 on success, negative error code otherwise.
2438 */
2439int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2440                    unsigned long pfn, unsigned long size, pgprot_t prot)
2441{
2442        int err;
2443
2444        err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2445        if (err)
2446                return -EINVAL;
2447
2448        err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
2449        if (err)
2450                untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2451        return err;
2452}
2453EXPORT_SYMBOL(remap_pfn_range);
2454
2455/**
2456 * vm_iomap_memory - remap memory to userspace
2457 * @vma: user vma to map to
2458 * @start: start of the physical memory to be mapped
2459 * @len: size of area
2460 *
2461 * This is a simplified io_remap_pfn_range() for common driver use. The
2462 * driver just needs to give us the physical memory range to be mapped,
2463 * we'll figure out the rest from the vma information.
2464 *
2465 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2466 * whatever write-combining details or similar.
2467 *
2468 * Return: %0 on success, negative error code otherwise.
2469 */
2470int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2471{
2472        unsigned long vm_len, pfn, pages;
2473
2474        /* Check that the physical memory area passed in looks valid */
2475        if (start + len < start)
2476                return -EINVAL;
2477        /*
2478         * You *really* shouldn't map things that aren't page-aligned,
2479         * but we've historically allowed it because IO memory might
2480         * just have smaller alignment.
2481         */
2482        len += start & ~PAGE_MASK;
2483        pfn = start >> PAGE_SHIFT;
2484        pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2485        if (pfn + pages < pfn)
2486                return -EINVAL;
2487
2488        /* We start the mapping 'vm_pgoff' pages into the area */
2489        if (vma->vm_pgoff > pages)
2490                return -EINVAL;
2491        pfn += vma->vm_pgoff;
2492        pages -= vma->vm_pgoff;
2493
2494        /* Can we fit all of the mapping? */
2495        vm_len = vma->vm_end - vma->vm_start;
2496        if (vm_len >> PAGE_SHIFT > pages)
2497                return -EINVAL;
2498
2499        /* Ok, let it rip */
2500        return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2501}
2502EXPORT_SYMBOL(vm_iomap_memory);
2503
2504static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2505                                     unsigned long addr, unsigned long end,
2506                                     pte_fn_t fn, void *data, bool create,
2507                                     pgtbl_mod_mask *mask)
2508{
2509        pte_t *pte, *mapped_pte;
2510        int err = 0;
2511        spinlock_t *ptl;
2512
2513        if (create) {
2514                mapped_pte = pte = (mm == &init_mm) ?
2515                        pte_alloc_kernel_track(pmd, addr, mask) :
2516                        pte_alloc_map_lock(mm, pmd, addr, &ptl);
2517                if (!pte)
2518                        return -ENOMEM;
2519        } else {
2520                mapped_pte = pte = (mm == &init_mm) ?
2521                        pte_offset_kernel(pmd, addr) :
2522                        pte_offset_map_lock(mm, pmd, addr, &ptl);
2523        }
2524
2525        BUG_ON(pmd_huge(*pmd));
2526
2527        arch_enter_lazy_mmu_mode();
2528
2529        if (fn) {
2530                do {
2531                        if (create || !pte_none(*pte)) {
2532                                err = fn(pte++, addr, data);
2533                                if (err)
2534                                        break;
2535                        }
2536                } while (addr += PAGE_SIZE, addr != end);
2537        }
2538        *mask |= PGTBL_PTE_MODIFIED;
2539
2540        arch_leave_lazy_mmu_mode();
2541
2542        if (mm != &init_mm)
2543                pte_unmap_unlock(mapped_pte, ptl);
2544        return err;
2545}
2546
2547static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2548                                     unsigned long addr, unsigned long end,
2549                                     pte_fn_t fn, void *data, bool create,
2550                                     pgtbl_mod_mask *mask)
2551{
2552        pmd_t *pmd;
2553        unsigned long next;
2554        int err = 0;
2555
2556        BUG_ON(pud_huge(*pud));
2557
2558        if (create) {
2559                pmd = pmd_alloc_track(mm, pud, addr, mask);
2560                if (!pmd)
2561                        return -ENOMEM;
2562        } else {
2563                pmd = pmd_offset(pud, addr);
2564        }
2565        do {
2566                next = pmd_addr_end(addr, end);
2567                if (pmd_none(*pmd) && !create)
2568                        continue;
2569                if (WARN_ON_ONCE(pmd_leaf(*pmd)))
2570                        return -EINVAL;
2571                if (!pmd_none(*pmd) && WARN_ON_ONCE(pmd_bad(*pmd))) {
2572                        if (!create)
2573                                continue;
2574                        pmd_clear_bad(pmd);
2575                }
2576                err = apply_to_pte_range(mm, pmd, addr, next,
2577                                         fn, data, create, mask);
2578                if (err)
2579                        break;
2580        } while (pmd++, addr = next, addr != end);
2581
2582        return err;
2583}
2584
2585static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2586                                     unsigned long addr, unsigned long end,
2587                                     pte_fn_t fn, void *data, bool create,
2588                                     pgtbl_mod_mask *mask)
2589{
2590        pud_t *pud;
2591        unsigned long next;
2592        int err = 0;
2593
2594        if (create) {
2595                pud = pud_alloc_track(mm, p4d, addr, mask);
2596                if (!pud)
2597                        return -ENOMEM;
2598        } else {
2599                pud = pud_offset(p4d, addr);
2600        }
2601        do {
2602                next = pud_addr_end(addr, end);
2603                if (pud_none(*pud) && !create)
2604                        continue;
2605                if (WARN_ON_ONCE(pud_leaf(*pud)))
2606                        return -EINVAL;
2607                if (!pud_none(*pud) && WARN_ON_ONCE(pud_bad(*pud))) {
2608                        if (!create)
2609                                continue;
2610                        pud_clear_bad(pud);
2611                }
2612                err = apply_to_pmd_range(mm, pud, addr, next,
2613                                         fn, data, create, mask);
2614                if (err)
2615                        break;
2616        } while (pud++, addr = next, addr != end);
2617
2618        return err;
2619}
2620
2621static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2622                                     unsigned long addr, unsigned long end,
2623                                     pte_fn_t fn, void *data, bool create,
2624                                     pgtbl_mod_mask *mask)
2625{
2626        p4d_t *p4d;
2627        unsigned long next;
2628        int err = 0;
2629
2630        if (create) {
2631                p4d = p4d_alloc_track(mm, pgd, addr, mask);
2632                if (!p4d)
2633                        return -ENOMEM;
2634        } else {
2635                p4d = p4d_offset(pgd, addr);
2636        }
2637        do {
2638                next = p4d_addr_end(addr, end);
2639                if (p4d_none(*p4d) && !create)
2640                        continue;
2641                if (WARN_ON_ONCE(p4d_leaf(*p4d)))
2642                        return -EINVAL;
2643                if (!p4d_none(*p4d) && WARN_ON_ONCE(p4d_bad(*p4d))) {
2644                        if (!create)
2645                                continue;
2646                        p4d_clear_bad(p4d);
2647                }
2648                err = apply_to_pud_range(mm, p4d, addr, next,
2649                                         fn, data, create, mask);
2650                if (err)
2651                        break;
2652        } while (p4d++, addr = next, addr != end);
2653
2654        return err;
2655}
2656
2657static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2658                                 unsigned long size, pte_fn_t fn,
2659                                 void *data, bool create)
2660{
2661        pgd_t *pgd;
2662        unsigned long start = addr, next;
2663        unsigned long end = addr + size;
2664        pgtbl_mod_mask mask = 0;
2665        int err = 0;
2666
2667        if (WARN_ON(addr >= end))
2668                return -EINVAL;
2669
2670        pgd = pgd_offset(mm, addr);
2671        do {
2672                next = pgd_addr_end(addr, end);
2673                if (pgd_none(*pgd) && !create)
2674                        continue;
2675                if (WARN_ON_ONCE(pgd_leaf(*pgd)))
2676                        return -EINVAL;
2677                if (!pgd_none(*pgd) && WARN_ON_ONCE(pgd_bad(*pgd))) {
2678                        if (!create)
2679                                continue;
2680                        pgd_clear_bad(pgd);
2681                }
2682                err = apply_to_p4d_range(mm, pgd, addr, next,
2683                                         fn, data, create, &mask);
2684                if (err)
2685                        break;
2686        } while (pgd++, addr = next, addr != end);
2687
2688        if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2689                arch_sync_kernel_mappings(start, start + size);
2690
2691        return err;
2692}
2693
2694/*
2695 * Scan a region of virtual memory, filling in page tables as necessary
2696 * and calling a provided function on each leaf page table.
2697 */
2698int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2699                        unsigned long size, pte_fn_t fn, void *data)
2700{
2701        return __apply_to_page_range(mm, addr, size, fn, data, true);
2702}
2703EXPORT_SYMBOL_GPL(apply_to_page_range);
2704
2705/*
2706 * Scan a region of virtual memory, calling a provided function on
2707 * each leaf page table where it exists.
2708 *
2709 * Unlike apply_to_page_range, this does _not_ fill in page tables
2710 * where they are absent.
2711 */
2712int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2713                                 unsigned long size, pte_fn_t fn, void *data)
2714{
2715        return __apply_to_page_range(mm, addr, size, fn, data, false);
2716}
2717EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2718
2719/*
2720 * handle_pte_fault chooses page fault handler according to an entry which was
2721 * read non-atomically.  Before making any commitment, on those architectures
2722 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2723 * parts, do_swap_page must check under lock before unmapping the pte and
2724 * proceeding (but do_wp_page is only called after already making such a check;
2725 * and do_anonymous_page can safely check later on).
2726 */
2727static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2728                                pte_t *page_table, pte_t orig_pte)
2729{
2730        int same = 1;
2731#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2732        if (sizeof(pte_t) > sizeof(unsigned long)) {
2733                spinlock_t *ptl = pte_lockptr(mm, pmd);
2734                spin_lock(ptl);
2735                same = pte_same(*page_table, orig_pte);
2736                spin_unlock(ptl);
2737        }
2738#endif
2739        pte_unmap(page_table);
2740        return same;
2741}
2742
2743static inline bool cow_user_page(struct page *dst, struct page *src,
2744                                 struct vm_fault *vmf)
2745{
2746        bool ret;
2747        void *kaddr;
2748        void __user *uaddr;
2749        bool locked = false;
2750        struct vm_area_struct *vma = vmf->vma;
2751        struct mm_struct *mm = vma->vm_mm;
2752        unsigned long addr = vmf->address;
2753
2754        if (likely(src)) {
2755                copy_user_highpage(dst, src, addr, vma);
2756                return true;
2757        }
2758
2759        /*
2760         * If the source page was a PFN mapping, we don't have
2761         * a "struct page" for it. We do a best-effort copy by
2762         * just copying from the original user address. If that
2763         * fails, we just zero-fill it. Live with it.
2764         */
2765        kaddr = kmap_atomic(dst);
2766        uaddr = (void __user *)(addr & PAGE_MASK);
2767
2768        /*
2769         * On architectures with software "accessed" bits, we would
2770         * take a double page fault, so mark it accessed here.
2771         */
2772        if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2773                pte_t entry;
2774
2775                vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2776                locked = true;
2777                if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2778                        /*
2779                         * Other thread has already handled the fault
2780                         * and update local tlb only
2781                         */
2782                        update_mmu_tlb(vma, addr, vmf->pte);
2783                        ret = false;
2784                        goto pte_unlock;
2785                }
2786
2787                entry = pte_mkyoung(vmf->orig_pte);
2788                if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2789                        update_mmu_cache(vma, addr, vmf->pte);
2790        }
2791
2792        /*
2793         * This really shouldn't fail, because the page is there
2794         * in the page tables. But it might just be unreadable,
2795         * in which case we just give up and fill the result with
2796         * zeroes.
2797         */
2798        if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2799                if (locked)
2800                        goto warn;
2801
2802                /* Re-validate under PTL if the page is still mapped */
2803                vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2804                locked = true;
2805                if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2806                        /* The PTE changed under us, update local tlb */
2807                        update_mmu_tlb(vma, addr, vmf->pte);
2808                        ret = false;
2809                        goto pte_unlock;
2810                }
2811
2812                /*
2813                 * The same page can be mapped back since last copy attempt.
2814                 * Try to copy again under PTL.
2815                 */
2816                if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2817                        /*
2818                         * Give a warn in case there can be some obscure
2819                         * use-case
2820                         */
2821warn:
2822                        WARN_ON_ONCE(1);
2823                        clear_page(kaddr);
2824                }
2825        }
2826
2827        ret = true;
2828
2829pte_unlock:
2830        if (locked)
2831                pte_unmap_unlock(vmf->pte, vmf->ptl);
2832        kunmap_atomic(kaddr);
2833        flush_dcache_page(dst);
2834
2835        return ret;
2836}
2837
2838static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2839{
2840        struct file *vm_file = vma->vm_file;
2841
2842        if (vm_file)
2843                return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2844
2845        /*
2846         * Special mappings (e.g. VDSO) do not have any file so fake
2847         * a default GFP_KERNEL for them.
2848         */
2849        return GFP_KERNEL;
2850}
2851
2852/*
2853 * Notify the address space that the page is about to become writable so that
2854 * it can prohibit this or wait for the page to get into an appropriate state.
2855 *
2856 * We do this without the lock held, so that it can sleep if it needs to.
2857 */
2858static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2859{
2860        vm_fault_t ret;
2861        struct page *page = vmf->page;
2862        unsigned int old_flags = vmf->flags;
2863
2864        vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2865
2866        if (vmf->vma->vm_file &&
2867            IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2868                return VM_FAULT_SIGBUS;
2869
2870        ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2871        /* Restore original flags so that caller is not surprised */
2872        vmf->flags = old_flags;
2873        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2874                return ret;
2875        if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2876                lock_page(page);
2877                if (!page->mapping) {
2878                        unlock_page(page);
2879                        return 0; /* retry */
2880                }
2881                ret |= VM_FAULT_LOCKED;
2882        } else
2883                VM_BUG_ON_PAGE(!PageLocked(page), page);
2884        return ret;
2885}
2886
2887/*
2888 * Handle dirtying of a page in shared file mapping on a write fault.
2889 *
2890 * The function expects the page to be locked and unlocks it.
2891 */
2892static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2893{
2894        struct vm_area_struct *vma = vmf->vma;
2895        struct address_space *mapping;
2896        struct page *page = vmf->page;
2897        bool dirtied;
2898        bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2899
2900        dirtied = set_page_dirty(page);
2901        VM_BUG_ON_PAGE(PageAnon(page), page);
2902        /*
2903         * Take a local copy of the address_space - page.mapping may be zeroed
2904         * by truncate after unlock_page().   The address_space itself remains
2905         * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2906         * release semantics to prevent the compiler from undoing this copying.
2907         */
2908        mapping = page_rmapping(page);
2909        unlock_page(page);
2910
2911        if (!page_mkwrite)
2912                file_update_time(vma->vm_file);
2913
2914        /*
2915         * Throttle page dirtying rate down to writeback speed.
2916         *
2917         * mapping may be NULL here because some device drivers do not
2918         * set page.mapping but still dirty their pages
2919         *
2920         * Drop the mmap_lock before waiting on IO, if we can. The file
2921         * is pinning the mapping, as per above.
2922         */
2923        if ((dirtied || page_mkwrite) && mapping) {
2924                struct file *fpin;
2925
2926                fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2927                balance_dirty_pages_ratelimited(mapping);
2928                if (fpin) {
2929                        fput(fpin);
2930                        return VM_FAULT_RETRY;
2931                }
2932        }
2933
2934        return 0;
2935}
2936
2937/*
2938 * Handle write page faults for pages that can be reused in the current vma
2939 *
2940 * This can happen either due to the mapping being with the VM_SHARED flag,
2941 * or due to us being the last reference standing to the page. In either
2942 * case, all we need to do here is to mark the page as writable and update
2943 * any related book-keeping.
2944 */
2945static inline void wp_page_reuse(struct vm_fault *vmf)
2946        __releases(vmf->ptl)
2947{
2948        struct vm_area_struct *vma = vmf->vma;
2949        struct page *page = vmf->page;
2950        pte_t entry;
2951        /*
2952         * Clear the pages cpupid information as the existing
2953         * information potentially belongs to a now completely
2954         * unrelated process.
2955         */
2956        if (page)
2957                page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2958
2959        flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2960        entry = pte_mkyoung(vmf->orig_pte);
2961        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2962        if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2963                update_mmu_cache(vma, vmf->address, vmf->pte);
2964        pte_unmap_unlock(vmf->pte, vmf->ptl);
2965        count_vm_event(PGREUSE);
2966}
2967
2968/*
2969 * Handle the case of a page which we actually need to copy to a new page.
2970 *
2971 * Called with mmap_lock locked and the old page referenced, but
2972 * without the ptl held.
2973 *
2974 * High level logic flow:
2975 *
2976 * - Allocate a page, copy the content of the old page to the new one.
2977 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2978 * - Take the PTL. If the pte changed, bail out and release the allocated page
2979 * - If the pte is still the way we remember it, update the page table and all
2980 *   relevant references. This includes dropping the reference the page-table
2981 *   held to the old page, as well as updating the rmap.
2982 * - In any case, unlock the PTL and drop the reference we took to the old page.
2983 */
2984static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2985{
2986        struct vm_area_struct *vma = vmf->vma;
2987        struct mm_struct *mm = vma->vm_mm;
2988        struct page *old_page = vmf->page;
2989        struct page *new_page = NULL;
2990        pte_t entry;
2991        int page_copied = 0;
2992        struct mmu_notifier_range range;
2993
2994        if (unlikely(anon_vma_prepare(vma)))
2995                goto oom;
2996
2997        if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2998                new_page = alloc_zeroed_user_highpage_movable(vma,
2999                                                              vmf->address);
3000                if (!new_page)
3001                        goto oom;
3002        } else {
3003                new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3004                                vmf->address);
3005                if (!new_page)
3006                        goto oom;
3007
3008                if (!cow_user_page(new_page, old_page, vmf)) {
3009                        /*
3010                         * COW failed, if the fault was solved by other,
3011                         * it's fine. If not, userspace would re-fault on
3012                         * the same address and we will handle the fault
3013                         * from the second attempt.
3014                         */
3015                        put_page(new_page);
3016                        if (old_page)
3017                                put_page(old_page);
3018                        return 0;
3019                }
3020        }
3021
3022        if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
3023                goto oom_free_new;
3024        cgroup_throttle_swaprate(new_page, GFP_KERNEL);
3025
3026        __SetPageUptodate(new_page);
3027
3028        mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
3029                                vmf->address & PAGE_MASK,
3030                                (vmf->address & PAGE_MASK) + PAGE_SIZE);
3031        mmu_notifier_invalidate_range_start(&range);
3032
3033        /*
3034         * Re-check the pte - we dropped the lock
3035         */
3036        vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
3037        if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
3038                if (old_page) {
3039                        if (!PageAnon(old_page)) {
3040                                dec_mm_counter_fast(mm,
3041                                                mm_counter_file(old_page));
3042                                inc_mm_counter_fast(mm, MM_ANONPAGES);
3043                        }
3044                } else {
3045                        inc_mm_counter_fast(mm, MM_ANONPAGES);
3046                }
3047                flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
3048                entry = mk_pte(new_page, vma->vm_page_prot);
3049                entry = pte_sw_mkyoung(entry);
3050                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3051
3052                /*
3053                 * Clear the pte entry and flush it first, before updating the
3054                 * pte with the new entry, to keep TLBs on different CPUs in
3055                 * sync. This code used to set the new PTE then flush TLBs, but
3056                 * that left a window where the new PTE could be loaded into
3057                 * some TLBs while the old PTE remains in others.
3058                 */
3059                ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
3060                page_add_new_anon_rmap(new_page, vma, vmf->address, false);
3061                lru_cache_add_inactive_or_unevictable(new_page, vma);
3062                /*
3063                 * We call the notify macro here because, when using secondary
3064                 * mmu page tables (such as kvm shadow page tables), we want the
3065                 * new page to be mapped directly into the secondary page table.
3066                 */
3067                set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
3068                update_mmu_cache(vma, vmf->address, vmf->pte);
3069                if (old_page) {
3070                        /*
3071                         * Only after switching the pte to the new page may
3072                         * we remove the mapcount here. Otherwise another
3073                         * process may come and find the rmap count decremented
3074                         * before the pte is switched to the new page, and
3075                         * "reuse" the old page writing into it while our pte
3076                         * here still points into it and can be read by other
3077                         * threads.
3078                         *
3079                         * The critical issue is to order this
3080                         * page_remove_rmap with the ptp_clear_flush above.
3081                         * Those stores are ordered by (if nothing else,)
3082                         * the barrier present in the atomic_add_negative
3083                         * in page_remove_rmap.
3084                         *
3085                         * Then the TLB flush in ptep_clear_flush ensures that
3086                         * no process can access the old page before the
3087                         * decremented mapcount is visible. And the old page
3088                         * cannot be reused until after the decremented
3089                         * mapcount is visible. So transitively, TLBs to
3090                         * old page will be flushed before it can be reused.
3091                         */
3092                        page_remove_rmap(old_page, false);
3093                }
3094
3095                /* Free the old page.. */
3096                new_page = old_page;
3097                page_copied = 1;
3098        } else {
3099                update_mmu_tlb(vma, vmf->address, vmf->pte);
3100        }
3101
3102        if (new_page)
3103                put_page(new_page);
3104
3105        pte_unmap_unlock(vmf->pte, vmf->ptl);
3106        /*
3107         * No need to double call mmu_notifier->invalidate_range() callback as
3108         * the above ptep_clear_flush_notify() did already call it.
3109         */
3110        mmu_notifier_invalidate_range_only_end(&range);
3111        if (old_page) {
3112                /*
3113                 * Don't let another task, with possibly unlocked vma,
3114                 * keep the mlocked page.
3115                 */
3116                if (page_copied && (vma->vm_flags & VM_LOCKED)) {
3117                        lock_page(old_page);    /* LRU manipulation */
3118                        if (PageMlocked(old_page))
3119                                munlock_vma_page(old_page);
3120                        unlock_page(old_page);
3121                }
3122                if (page_copied)
3123                        free_swap_cache(old_page);
3124                put_page(old_page);
3125        }
3126        return page_copied ? VM_FAULT_WRITE : 0;
3127oom_free_new:
3128        put_page(new_page);
3129oom:
3130        if (old_page)
3131                put_page(old_page);
3132        return VM_FAULT_OOM;
3133}
3134
3135/**
3136 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
3137 *                        writeable once the page is prepared
3138 *
3139 * @vmf: structure describing the fault
3140 *
3141 * This function handles all that is needed to finish a write page fault in a
3142 * shared mapping due to PTE being read-only once the mapped page is prepared.
3143 * It handles locking of PTE and modifying it.
3144 *
3145 * The function expects the page to be locked or other protection against
3146 * concurrent faults / writeback (such as DAX radix tree locks).
3147 *
3148 * Return: %0 on success, %VM_FAULT_NOPAGE when PTE got changed before
3149 * we acquired PTE lock.
3150 */
3151vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
3152{
3153        WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3154        vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3155                                       &vmf->ptl);
3156        /*
3157         * We might have raced with another page fault while we released the
3158         * pte_offset_map_lock.
3159         */
3160        if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3161                update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3162                pte_unmap_unlock(vmf->pte, vmf->ptl);
3163                return VM_FAULT_NOPAGE;
3164        }
3165        wp_page_reuse(vmf);
3166        return 0;
3167}
3168
3169/*
3170 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3171 * mapping
3172 */
3173static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3174{
3175        struct vm_area_struct *vma = vmf->vma;
3176
3177        if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3178                vm_fault_t ret;
3179
3180                pte_unmap_unlock(vmf->pte, vmf->ptl);
3181                vmf->flags |= FAULT_FLAG_MKWRITE;
3182                ret = vma->vm_ops->pfn_mkwrite(vmf);
3183                if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3184                        return ret;
3185                return finish_mkwrite_fault(vmf);
3186        }
3187        wp_page_reuse(vmf);
3188        return VM_FAULT_WRITE;
3189}
3190
3191static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3192        __releases(vmf->ptl)
3193{
3194        struct vm_area_struct *vma = vmf->vma;
3195        vm_fault_t ret = VM_FAULT_WRITE;
3196
3197        get_page(vmf->page);
3198
3199        if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3200                vm_fault_t tmp;
3201
3202                pte_unmap_unlock(vmf->pte, vmf->ptl);
3203                tmp = do_page_mkwrite(vmf);
3204                if (unlikely(!tmp || (tmp &
3205                                      (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3206                        put_page(vmf->page);
3207                        return tmp;
3208                }
3209                tmp = finish_mkwrite_fault(vmf);
3210                if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3211                        unlock_page(vmf->page);
3212                        put_page(vmf->page);
3213                        return tmp;
3214                }
3215        } else {
3216                wp_page_reuse(vmf);
3217                lock_page(vmf->page);
3218        }
3219        ret |= fault_dirty_shared_page(vmf);
3220        put_page(vmf->page);
3221
3222        return ret;
3223}
3224
3225/*
3226 * This routine handles present pages, when users try to write
3227 * to a shared page. It is done by copying the page to a new address
3228 * and decrementing the shared-page counter for the old page.
3229 *
3230 * Note that this routine assumes that the protection checks have been
3231 * done by the caller (the low-level page fault routine in most cases).
3232 * Thus we can safely just mark it writable once we've done any necessary
3233 * COW.
3234 *
3235 * We also mark the page dirty at this point even though the page will
3236 * change only once the write actually happens. This avoids a few races,
3237 * and potentially makes it more efficient.
3238 *
3239 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3240 * but allow concurrent faults), with pte both mapped and locked.
3241 * We return with mmap_lock still held, but pte unmapped and unlocked.
3242 */
3243static vm_fault_t do_wp_page(struct vm_fault *vmf)
3244        __releases(vmf->ptl)
3245{
3246        struct vm_area_struct *vma = vmf->vma;
3247
3248        if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3249                pte_unmap_unlock(vmf->pte, vmf->ptl);
3250                return handle_userfault(vmf, VM_UFFD_WP);
3251        }
3252
3253        /*
3254         * Userfaultfd write-protect can defer flushes. Ensure the TLB
3255         * is flushed in this case before copying.
3256         */
3257        if (unlikely(userfaultfd_wp(vmf->vma) &&
3258                     mm_tlb_flush_pending(vmf->vma->vm_mm)))
3259                flush_tlb_page(vmf->vma, vmf->address);
3260
3261        vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3262        if (!vmf->page) {
3263                /*
3264                 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3265                 * VM_PFNMAP VMA.
3266                 *
3267                 * We should not cow pages in a shared writeable mapping.
3268                 * Just mark the pages writable and/or call ops->pfn_mkwrite.
3269                 */
3270                if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3271                                     (VM_WRITE|VM_SHARED))
3272                        return wp_pfn_shared(vmf);
3273
3274                pte_unmap_unlock(vmf->pte, vmf->ptl);
3275                return wp_page_copy(vmf);
3276        }
3277
3278        /*
3279         * Take out anonymous pages first, anonymous shared vmas are
3280         * not dirty accountable.
3281         */
3282        if (PageAnon(vmf->page)) {
3283                struct page *page = vmf->page;
3284
3285                /* PageKsm() doesn't necessarily raise the page refcount */
3286                if (PageKsm(page) || page_count(page) != 1)
3287                        goto copy;
3288                if (!trylock_page(page))
3289                        goto copy;
3290                if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3291                        unlock_page(page);
3292                        goto copy;
3293                }
3294                /*
3295                 * Ok, we've got the only map reference, and the only
3296                 * page count reference, and the page is locked,
3297                 * it's dark out, and we're wearing sunglasses. Hit it.
3298                 */
3299                unlock_page(page);
3300                wp_page_reuse(vmf);
3301                return VM_FAULT_WRITE;
3302        } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3303                                        (VM_WRITE|VM_SHARED))) {
3304                return wp_page_shared(vmf);
3305        }
3306copy:
3307        /*
3308         * Ok, we need to copy. Oh, well..
3309         */
3310        get_page(vmf->page);
3311
3312        pte_unmap_unlock(vmf->pte, vmf->ptl);
3313        return wp_page_copy(vmf);
3314}
3315
3316static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3317                unsigned long start_addr, unsigned long end_addr,
3318                struct zap_details *details)
3319{
3320        zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3321}
3322
3323static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3324                                            struct zap_details *details)
3325{
3326        struct vm_area_struct *vma;
3327        pgoff_t vba, vea, zba, zea;
3328
3329        vma_interval_tree_foreach(vma, root,
3330                        details->first_index, details->last_index) {
3331
3332                vba = vma->vm_pgoff;
3333                vea = vba + vma_pages(vma) - 1;
3334                zba = details->first_index;
3335                if (zba < vba)
3336                        zba = vba;
3337                zea = details->last_index;
3338                if (zea > vea)
3339                        zea = vea;
3340
3341                unmap_mapping_range_vma(vma,
3342                        ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3343                        ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3344                                details);
3345        }
3346}
3347
3348/**
3349 * unmap_mapping_page() - Unmap single page from processes.
3350 * @page: The locked page to be unmapped.
3351 *
3352 * Unmap this page from any userspace process which still has it mmaped.
3353 * Typically, for efficiency, the range of nearby pages has already been
3354 * unmapped by unmap_mapping_pages() or unmap_mapping_range().  But once
3355 * truncation or invalidation holds the lock on a page, it may find that
3356 * the page has been remapped again: and then uses unmap_mapping_page()
3357 * to unmap it finally.
3358 */
3359void unmap_mapping_page(struct page *page)
3360{
3361        struct address_space *mapping = page->mapping;
3362        struct zap_details details = { };
3363
3364        VM_BUG_ON(!PageLocked(page));
3365        VM_BUG_ON(PageTail(page));
3366
3367        details.check_mapping = mapping;
3368        details.first_index = page->index;
3369        details.last_index = page->index + thp_nr_pages(page) - 1;
3370        details.single_page = page;
3371
3372        i_mmap_lock_write(mapping);
3373        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3374                unmap_mapping_range_tree(&mapping->i_mmap, &details);
3375        i_mmap_unlock_write(mapping);
3376}
3377
3378/**
3379 * unmap_mapping_pages() - Unmap pages from processes.
3380 * @mapping: The address space containing pages to be unmapped.
3381 * @start: Index of first page to be unmapped.
3382 * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
3383 * @even_cows: Whether to unmap even private COWed pages.
3384 *
3385 * Unmap the pages in this address space from any userspace process which
3386 * has them mmaped.  Generally, you want to remove COWed pages as well when
3387 * a file is being truncated, but not when invalidating pages from the page
3388 * cache.
3389 */
3390void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3391                pgoff_t nr, bool even_cows)
3392{
3393        struct zap_details details = { };
3394
3395        details.check_mapping = even_cows ? NULL : mapping;
3396        details.first_index = start;
3397        details.last_index = start + nr - 1;
3398        if (details.last_index < details.first_index)
3399                details.last_index = ULONG_MAX;
3400
3401        i_mmap_lock_write(mapping);
3402        if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3403                unmap_mapping_range_tree(&mapping->i_mmap, &details);
3404        i_mmap_unlock_write(mapping);
3405}
3406EXPORT_SYMBOL_GPL(unmap_mapping_pages);
3407
3408/**
3409 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3410 * address_space corresponding to the specified byte range in the underlying
3411 * file.
3412 *
3413 * @mapping: the address space containing mmaps to be unmapped.
3414 * @holebegin: byte in first page to unmap, relative to the start of
3415 * the underlying file.  This will be rounded down to a PAGE_SIZE
3416 * boundary.  Note that this is different from truncate_pagecache(), which
3417 * must keep the partial page.  In contrast, we must get rid of
3418 * partial pages.
3419 * @holelen: size of prospective hole in bytes.  This will be rounded
3420 * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
3421 * end of the file.
3422 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3423 * but 0 when invalidating pagecache, don't throw away private data.
3424 */
3425void unmap_mapping_range(struct address_space *mapping,
3426                loff_t const holebegin, loff_t const holelen, int even_cows)
3427{
3428        pgoff_t hba = holebegin >> PAGE_SHIFT;
3429        pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3430
3431        /* Check for overflow. */
3432        if (sizeof(holelen) > sizeof(hlen)) {
3433                long long holeend =
3434                        (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3435                if (holeend & ~(long long)ULONG_MAX)
3436                        hlen = ULONG_MAX - hba + 1;
3437        }
3438
3439        unmap_mapping_pages(mapping, hba, hlen, even_cows);
3440}
3441EXPORT_SYMBOL(unmap_mapping_range);
3442
3443/*
3444 * Restore a potential device exclusive pte to a working pte entry
3445 */
3446static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
3447{
3448        struct page *page = vmf->page;
3449        struct vm_area_struct *vma = vmf->vma;
3450        struct mmu_notifier_range range;
3451
3452        if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
3453                return VM_FAULT_RETRY;
3454        mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
3455                                vma->vm_mm, vmf->address & PAGE_MASK,
3456                                (vmf->address & PAGE_MASK) + PAGE_SIZE, NULL);
3457        mmu_notifier_invalidate_range_start(&range);
3458
3459        vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3460                                &vmf->ptl);
3461        if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3462                restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
3463
3464        pte_unmap_unlock(vmf->pte, vmf->ptl);
3465        unlock_page(page);
3466
3467        mmu_notifier_invalidate_range_end(&range);
3468        return 0;
3469}
3470
3471/*
3472 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3473 * but allow concurrent faults), and pte mapped but not yet locked.
3474 * We return with pte unmapped and unlocked.
3475 *
3476 * We return with the mmap_lock locked or unlocked in the same cases
3477 * as does filemap_fault().
3478 */
3479vm_fault_t do_swap_page(struct vm_fault *vmf)
3480{
3481        struct vm_area_struct *vma = vmf->vma;
3482        struct page *page = NULL, *swapcache;
3483        struct swap_info_struct *si = NULL;
3484        swp_entry_t entry;
3485        pte_t pte;
3486        int locked;
3487        int exclusive = 0;
3488        vm_fault_t ret = 0;
3489        void *shadow = NULL;
3490
3491        if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3492                goto out;
3493
3494        entry = pte_to_swp_entry(vmf->orig_pte);
3495        if (unlikely(non_swap_entry(entry))) {
3496                if (is_migration_entry(entry)) {
3497                        migration_entry_wait(vma->vm_mm, vmf->pmd,
3498                                             vmf->address);
3499                } else if (is_device_exclusive_entry(entry)) {
3500                        vmf->page = pfn_swap_entry_to_page(entry);
3501                        ret = remove_device_exclusive_entry(vmf);
3502                } else if (is_device_private_entry(entry)) {
3503                        vmf->page = pfn_swap_entry_to_page(entry);
3504                        ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3505                } else if (is_hwpoison_entry(entry)) {
3506                        ret = VM_FAULT_HWPOISON;
3507                } else {
3508                        print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3509                        ret = VM_FAULT_SIGBUS;
3510                }
3511                goto out;
3512        }
3513
3514        /* Prevent swapoff from happening to us. */
3515        si = get_swap_device(entry);
3516        if (unlikely(!si))
3517                goto out;
3518
3519        delayacct_set_flag(current, DELAYACCT_PF_SWAPIN);
3520        page = lookup_swap_cache(entry, vma, vmf->address);
3521        swapcache = page;
3522
3523        if (!page) {
3524                if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3525                    __swap_count(entry) == 1) {
3526                        /* skip swapcache */
3527                        page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3528                                                        vmf->address);
3529                        if (page) {
3530                                __SetPageLocked(page);
3531                                __SetPageSwapBacked(page);
3532
3533                                if (mem_cgroup_swapin_charge_page(page,
3534                                        vma->vm_mm, GFP_KERNEL, entry)) {
3535                                        ret = VM_FAULT_OOM;
3536                                        goto out_page;
3537                                }
3538                                mem_cgroup_swapin_uncharge_swap(entry);
3539
3540                                shadow = get_shadow_from_swap_cache(entry);
3541                                if (shadow)
3542                                        workingset_refault(page, shadow);
3543
3544                                lru_cache_add(page);
3545
3546                                /* To provide entry to swap_readpage() */
3547                                set_page_private(page, entry.val);
3548                                swap_readpage(page, true);
3549                                set_page_private(page, 0);
3550                        }
3551                } else {
3552                        page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3553                                                vmf);
3554                        swapcache = page;
3555                }
3556
3557                if (!page) {
3558                        /*
3559                         * Back out if somebody else faulted in this pte
3560                         * while we released the pte lock.
3561                         */
3562                        vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3563                                        vmf->address, &vmf->ptl);
3564                        if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3565                                ret = VM_FAULT_OOM;
3566                        delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3567                        goto unlock;
3568                }
3569
3570                /* Had to read the page from swap area: Major fault */
3571                ret = VM_FAULT_MAJOR;
3572                count_vm_event(PGMAJFAULT);
3573                count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3574        } else if (PageHWPoison(page)) {
3575                /*
3576                 * hwpoisoned dirty swapcache pages are kept for killing
3577                 * owner processes (which may be unknown at hwpoison time)
3578                 */
3579                ret = VM_FAULT_HWPOISON;
3580                delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3581                goto out_release;
3582        }
3583
3584        locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3585
3586        delayacct_clear_flag(current, DELAYACCT_PF_SWAPIN);
3587        if (!locked) {
3588                ret |= VM_FAULT_RETRY;
3589                goto out_release;
3590        }
3591
3592        /*
3593         * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3594         * release the swapcache from under us.  The page pin, and pte_same
3595         * test below, are not enough to exclude that.  Even if it is still
3596         * swapcache, we need to check that the page's swap has not changed.
3597         */
3598        if (unlikely((!PageSwapCache(page) ||
3599                        page_private(page) != entry.val)) && swapcache)
3600                goto out_page;
3601
3602        page = ksm_might_need_to_copy(page, vma, vmf->address);
3603        if (unlikely(!page)) {
3604                ret = VM_FAULT_OOM;
3605                page = swapcache;
3606                goto out_page;
3607        }
3608
3609        cgroup_throttle_swaprate(page, GFP_KERNEL);
3610
3611        /*
3612         * Back out if somebody else already faulted in this pte.
3613         */
3614        vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3615                        &vmf->ptl);
3616        if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3617                goto out_nomap;
3618
3619        if (unlikely(!PageUptodate(page))) {
3620                ret = VM_FAULT_SIGBUS;
3621                goto out_nomap;
3622        }
3623
3624        /*
3625         * The page isn't present yet, go ahead with the fault.
3626         *
3627         * Be careful about the sequence of operations here.
3628         * To get its accounting right, reuse_swap_page() must be called
3629         * while the page is counted on swap but not yet in mapcount i.e.
3630         * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3631         * must be called after the swap_free(), or it will never succeed.
3632         */
3633
3634        inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3635        dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3636        pte = mk_pte(page, vma->vm_page_prot);
3637        if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3638                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3639                vmf->flags &= ~FAULT_FLAG_WRITE;
3640                ret |= VM_FAULT_WRITE;
3641                exclusive = RMAP_EXCLUSIVE;
3642        }
3643        flush_icache_page(vma, page);
3644        if (pte_swp_soft_dirty(vmf->orig_pte))
3645                pte = pte_mksoft_dirty(pte);
3646        if (pte_swp_uffd_wp(vmf->orig_pte)) {
3647                pte = pte_mkuffd_wp(pte);
3648                pte = pte_wrprotect(pte);
3649        }
3650        set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3651        arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3652        vmf->orig_pte = pte;
3653
3654        /* ksm created a completely new copy */
3655        if (unlikely(page != swapcache && swapcache)) {
3656                page_add_new_anon_rmap(page, vma, vmf->address, false);
3657                lru_cache_add_inactive_or_unevictable(page, vma);
3658        } else {
3659                do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3660        }
3661
3662        swap_free(entry);
3663        if (mem_cgroup_swap_full(page) ||
3664            (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3665                try_to_free_swap(page);
3666        unlock_page(page);
3667        if (page != swapcache && swapcache) {
3668                /*
3669                 * Hold the lock to avoid the swap entry to be reused
3670                 * until we take the PT lock for the pte_same() check
3671                 * (to avoid false positives from pte_same). For
3672                 * further safety release the lock after the swap_free
3673                 * so that the swap count won't change under a
3674                 * parallel locked swapcache.
3675                 */
3676                unlock_page(swapcache);
3677                put_page(swapcache);
3678        }
3679
3680        if (vmf->flags & FAULT_FLAG_WRITE) {
3681                ret |= do_wp_page(vmf);
3682                if (ret & VM_FAULT_ERROR)
3683                        ret &= VM_FAULT_ERROR;
3684                goto out;
3685        }
3686
3687        /* No need to invalidate - it was non-present before */
3688        update_mmu_cache(vma, vmf->address, vmf->pte);
3689unlock:
3690        pte_unmap_unlock(vmf->pte, vmf->ptl);
3691out:
3692        if (si)
3693                put_swap_device(si);
3694        return ret;
3695out_nomap:
3696        pte_unmap_unlock(vmf->pte, vmf->ptl);
3697out_page:
3698        unlock_page(page);
3699out_release:
3700        put_page(page);
3701        if (page != swapcache && swapcache) {
3702                unlock_page(swapcache);
3703                put_page(swapcache);
3704        }
3705        if (si)
3706                put_swap_device(si);
3707        return ret;
3708}
3709
3710/*
3711 * We enter with non-exclusive mmap_lock (to exclude vma changes,
3712 * but allow concurrent faults), and pte mapped but not yet locked.
3713 * We return with mmap_lock still held, but pte unmapped and unlocked.
3714 */
3715static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3716{
3717        struct vm_area_struct *vma = vmf->vma;
3718        struct page *page;
3719        vm_fault_t ret = 0;
3720        pte_t entry;
3721
3722        /* File mapping without ->vm_ops ? */
3723        if (vma->vm_flags & VM_SHARED)
3724                return VM_FAULT_SIGBUS;
3725
3726        /*
3727         * Use pte_alloc() instead of pte_alloc_map().  We can't run
3728         * pte_offset_map() on pmds where a huge pmd might be created
3729         * from a different thread.
3730         *
3731         * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3732         * parallel threads are excluded by other means.
3733         *
3734         * Here we only have mmap_read_lock(mm).
3735         */
3736        if (pte_alloc(vma->vm_mm, vmf->pmd))
3737                return VM_FAULT_OOM;
3738
3739        /* See comment in handle_pte_fault() */
3740        if (unlikely(pmd_trans_unstable(vmf->pmd)))
3741                return 0;
3742
3743        /* Use the zero-page for reads */
3744        if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3745                        !mm_forbids_zeropage(vma->vm_mm)) {
3746                entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3747                                                vma->vm_page_prot));
3748                vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3749                                vmf->address, &vmf->ptl);
3750                if (!pte_none(*vmf->pte)) {
3751                        update_mmu_tlb(vma, vmf->address, vmf->pte);
3752                        goto unlock;
3753                }
3754                ret = check_stable_address_space(vma->vm_mm);
3755                if (ret)
3756                        goto unlock;
3757                /* Deliver the page fault to userland, check inside PT lock */
3758                if (userfaultfd_missing(vma)) {
3759                        pte_unmap_unlock(vmf->pte, vmf->ptl);
3760                        return handle_userfault(vmf, VM_UFFD_MISSING);
3761                }
3762                goto setpte;
3763        }
3764
3765        /* Allocate our own private page. */
3766        if (unlikely(anon_vma_prepare(vma)))
3767                goto oom;
3768        page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3769        if (!page)
3770                goto oom;
3771
3772        if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3773                goto oom_free_page;
3774        cgroup_throttle_swaprate(page, GFP_KERNEL);
3775
3776        /*
3777         * The memory barrier inside __SetPageUptodate makes sure that
3778         * preceding stores to the page contents become visible before
3779         * the set_pte_at() write.
3780         */
3781        __SetPageUptodate(page);
3782
3783        entry = mk_pte(page, vma->vm_page_prot);
3784        entry = pte_sw_mkyoung(entry);
3785        if (vma->vm_flags & VM_WRITE)
3786                entry = pte_mkwrite(pte_mkdirty(entry));
3787
3788        vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3789                        &vmf->ptl);
3790        if (!pte_none(*vmf->pte)) {
3791                update_mmu_cache(vma, vmf->address, vmf->pte);
3792                goto release;
3793        }
3794
3795        ret = check_stable_address_space(vma->vm_mm);
3796        if (ret)
3797                goto release;
3798
3799        /* Deliver the page fault to userland, check inside PT lock */
3800        if (userfaultfd_missing(vma)) {
3801                pte_unmap_unlock(vmf->pte, vmf->ptl);
3802                put_page(page);
3803                return handle_userfault(vmf, VM_UFFD_MISSING);
3804        }
3805
3806        inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3807        page_add_new_anon_rmap(page, vma, vmf->address, false);
3808        lru_cache_add_inactive_or_unevictable(page, vma);
3809setpte:
3810        set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3811
3812        /* No need to invalidate - it was non-present before */
3813        update_mmu_cache(vma, vmf->address, vmf->pte);
3814unlock:
3815        pte_unmap_unlock(vmf->pte, vmf->ptl);
3816        return ret;
3817release:
3818        put_page(page);
3819        goto unlock;
3820oom_free_page:
3821        put_page(page);
3822oom:
3823        return VM_FAULT_OOM;
3824}
3825
3826/*
3827 * The mmap_lock must have been held on entry, and may have been
3828 * released depending on flags and vma->vm_ops->fault() return value.
3829 * See filemap_fault() and __lock_page_retry().
3830 */
3831static vm_fault_t __do_fault(struct vm_fault *vmf)
3832{
3833        struct vm_area_struct *vma = vmf->vma;
3834        vm_fault_t ret;
3835
3836        /*
3837         * Preallocate pte before we take page_lock because this might lead to
3838         * deadlocks for memcg reclaim which waits for pages under writeback:
3839         *                              lock_page(A)
3840         *                              SetPageWriteback(A)
3841         *                              unlock_page(A)
3842         * lock_page(B)
3843         *                              lock_page(B)
3844         * pte_alloc_one
3845         *   shrink_page_list
3846         *     wait_on_page_writeback(A)
3847         *                              SetPageWriteback(B)
3848         *                              unlock_page(B)
3849         *                              # flush A, B to clear the writeback
3850         */
3851        if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3852                vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3853                if (!vmf->prealloc_pte)
3854                        return VM_FAULT_OOM;
3855                smp_wmb(); /* See comment in __pte_alloc() */
3856        }
3857
3858        ret = vma->vm_ops->fault(vmf);
3859        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3860                            VM_FAULT_DONE_COW)))
3861                return ret;
3862
3863        if (unlikely(PageHWPoison(vmf->page))) {
3864                if (ret & VM_FAULT_LOCKED)
3865                        unlock_page(vmf->page);
3866                put_page(vmf->page);
3867                vmf->page = NULL;
3868                return VM_FAULT_HWPOISON;
3869        }
3870
3871        if (unlikely(!(ret & VM_FAULT_LOCKED)))
3872                lock_page(vmf->page);
3873        else
3874                VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3875
3876        return ret;
3877}
3878
3879#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3880static void deposit_prealloc_pte(struct vm_fault *vmf)
3881{
3882        struct vm_area_struct *vma = vmf->vma;
3883
3884        pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3885        /*
3886         * We are going to consume the prealloc table,
3887         * count that as nr_ptes.
3888         */
3889        mm_inc_nr_ptes(vma->vm_mm);
3890        vmf->prealloc_pte = NULL;
3891}
3892
3893vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3894{
3895        struct vm_area_struct *vma = vmf->vma;
3896        bool write = vmf->flags & FAULT_FLAG_WRITE;
3897        unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3898        pmd_t entry;
3899        int i;
3900        vm_fault_t ret = VM_FAULT_FALLBACK;
3901
3902        if (!transhuge_vma_suitable(vma, haddr))
3903                return ret;
3904
3905        page = compound_head(page);
3906        if (compound_order(page) != HPAGE_PMD_ORDER)
3907                return ret;
3908
3909        /*
3910         * Just backoff if any subpage of a THP is corrupted otherwise
3911         * the corrupted page may mapped by PMD silently to escape the
3912         * check.  This kind of THP just can be PTE mapped.  Access to
3913         * the corrupted subpage should trigger SIGBUS as expected.
3914         */
3915        if (unlikely(PageHasHWPoisoned(page)))
3916                return ret;
3917
3918        /*
3919         * Archs like ppc64 need additional space to store information
3920         * related to pte entry. Use the preallocated table for that.
3921         */
3922        if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3923                vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3924                if (!vmf->prealloc_pte)
3925                        return VM_FAULT_OOM;
3926                smp_wmb(); /* See comment in __pte_alloc() */
3927        }
3928
3929        vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3930        if (unlikely(!pmd_none(*vmf->pmd)))
3931                goto out;
3932
3933        for (i = 0; i < HPAGE_PMD_NR; i++)
3934                flush_icache_page(vma, page + i);
3935
3936        entry = mk_huge_pmd(page, vma->vm_page_prot);
3937        if (write)
3938                entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3939
3940        add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3941        page_add_file_rmap(page, true);
3942        /*
3943         * deposit and withdraw with pmd lock held
3944         */
3945        if (arch_needs_pgtable_deposit())
3946                deposit_prealloc_pte(vmf);
3947
3948        set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3949
3950        update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3951
3952        /* fault is handled */
3953        ret = 0;
3954        count_vm_event(THP_FILE_MAPPED);
3955out:
3956        spin_unlock(vmf->ptl);
3957        return ret;
3958}
3959#else
3960vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3961{
3962        return VM_FAULT_FALLBACK;
3963}
3964#endif
3965
3966void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3967{
3968        struct vm_area_struct *vma = vmf->vma;
3969        bool write = vmf->flags & FAULT_FLAG_WRITE;
3970        bool prefault = vmf->address != addr;
3971        pte_t entry;
3972
3973        flush_icache_page(vma, page);
3974        entry = mk_pte(page, vma->vm_page_prot);
3975
3976        if (prefault && arch_wants_old_prefaulted_pte())
3977                entry = pte_mkold(entry);
3978        else
3979                entry = pte_sw_mkyoung(entry);
3980
3981        if (write)
3982                entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3983        /* copy-on-write page */
3984        if (write && !(vma->vm_flags & VM_SHARED)) {
3985                inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3986                page_add_new_anon_rmap(page, vma, addr, false);
3987                lru_cache_add_inactive_or_unevictable(page, vma);
3988        } else {
3989                inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3990                page_add_file_rmap(page, false);
3991        }
3992        set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3993}
3994
3995/**
3996 * finish_fault - finish page fault once we have prepared the page to fault
3997 *
3998 * @vmf: structure describing the fault
3999 *
4000 * This function handles all that is needed to finish a page fault once the
4001 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
4002 * given page, adds reverse page mapping, handles memcg charges and LRU
4003 * addition.
4004 *
4005 * The function expects the page to be locked and on success it consumes a
4006 * reference of a page being mapped (for the PTE which maps it).
4007 *
4008 * Return: %0 on success, %VM_FAULT_ code in case of error.
4009 */
4010vm_fault_t finish_fault(struct vm_fault *vmf)
4011{
4012        struct vm_area_struct *vma = vmf->vma;
4013        struct page *page;
4014        vm_fault_t ret;
4015
4016        /* Did we COW the page? */
4017        if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
4018                page = vmf->cow_page;
4019        else
4020                page = vmf->page;
4021
4022        /*
4023         * check even for read faults because we might have lost our CoWed
4024         * page
4025         */
4026        if (!(vma->vm_flags & VM_SHARED)) {
4027                ret = check_stable_address_space(vma->vm_mm);
4028                if (ret)
4029                        return ret;
4030        }
4031
4032        if (pmd_none(*vmf->pmd)) {
4033                if (PageTransCompound(page)) {
4034                        ret = do_set_pmd(vmf, page);
4035                        if (ret != VM_FAULT_FALLBACK)
4036                                return ret;
4037                }
4038
4039                if (vmf->prealloc_pte) {
4040                        vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
4041                        if (likely(pmd_none(*vmf->pmd))) {
4042                                mm_inc_nr_ptes(vma->vm_mm);
4043                                pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
4044                                vmf->prealloc_pte = NULL;
4045                        }
4046                        spin_unlock(vmf->ptl);
4047                } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
4048                        return VM_FAULT_OOM;
4049                }
4050        }
4051
4052        /* See comment in handle_pte_fault() */
4053        if (pmd_devmap_trans_unstable(vmf->pmd))
4054                return 0;
4055
4056        vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
4057                                      vmf->address, &vmf->ptl);
4058        ret = 0;
4059        /* Re-check under ptl */
4060        if (likely(pte_none(*vmf->pte)))
4061                do_set_pte(vmf, page, vmf->address);
4062        else
4063                ret = VM_FAULT_NOPAGE;
4064
4065        update_mmu_tlb(vma, vmf->address, vmf->pte);
4066        pte_unmap_unlock(vmf->pte, vmf->ptl);
4067        return ret;
4068}
4069
4070static unsigned long fault_around_bytes __read_mostly =
4071        rounddown_pow_of_two(65536);
4072
4073#ifdef CONFIG_DEBUG_FS
4074static int fault_around_bytes_get(void *data, u64 *val)
4075{
4076        *val = fault_around_bytes;
4077        return 0;
4078}
4079
4080/*
4081 * fault_around_bytes must be rounded down to the nearest page order as it's
4082 * what do_fault_around() expects to see.
4083 */
4084static int fault_around_bytes_set(void *data, u64 val)
4085{
4086        if (val / PAGE_SIZE > PTRS_PER_PTE)
4087                return -EINVAL;
4088        if (val > PAGE_SIZE)
4089                fault_around_bytes = rounddown_pow_of_two(val);
4090        else
4091                fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
4092        return 0;
4093}
4094DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
4095                fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
4096
4097static int __init fault_around_debugfs(void)
4098{
4099        debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
4100                                   &fault_around_bytes_fops);
4101        return 0;
4102}
4103late_initcall(fault_around_debugfs);
4104#endif
4105
4106/*
4107 * do_fault_around() tries to map few pages around the fault address. The hope
4108 * is that the pages will be needed soon and this will lower the number of
4109 * faults to handle.
4110 *
4111 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
4112 * not ready to be mapped: not up-to-date, locked, etc.
4113 *
4114 * This function is called with the page table lock taken. In the split ptlock
4115 * case the page table lock only protects only those entries which belong to
4116 * the page table corresponding to the fault address.
4117 *
4118 * This function doesn't cross the VMA boundaries, in order to call map_pages()
4119 * only once.
4120 *
4121 * fault_around_bytes defines how many bytes we'll try to map.
4122 * do_fault_around() expects it to be set to a power of two less than or equal
4123 * to PTRS_PER_PTE.
4124 *
4125 * The virtual address of the area that we map is naturally aligned to
4126 * fault_around_bytes rounded down to the machine page size
4127 * (and therefore to page order).  This way it's easier to guarantee
4128 * that we don't cross page table boundaries.
4129 */
4130static vm_fault_t do_fault_around(struct vm_fault *vmf)
4131{
4132        unsigned long address = vmf->address, nr_pages, mask;
4133        pgoff_t start_pgoff = vmf->pgoff;
4134        pgoff_t end_pgoff;
4135        int off;
4136
4137        nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
4138        mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
4139
4140        address = max(address & mask, vmf->vma->vm_start);
4141        off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
4142        start_pgoff -= off;
4143
4144        /*
4145         *  end_pgoff is either the end of the page table, the end of
4146         *  the vma or nr_pages from start_pgoff, depending what is nearest.
4147         */
4148        end_pgoff = start_pgoff -
4149                ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
4150                PTRS_PER_PTE - 1;
4151        end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
4152                        start_pgoff + nr_pages - 1);
4153
4154        if (pmd_none(*vmf->pmd)) {
4155                vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
4156                if (!vmf->prealloc_pte)
4157                        return VM_FAULT_OOM;
4158                smp_wmb(); /* See comment in __pte_alloc() */
4159        }
4160
4161        return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
4162}
4163
4164static vm_fault_t do_read_fault(struct vm_fault *vmf)
4165{
4166        struct vm_area_struct *vma = vmf->vma;
4167        vm_fault_t ret = 0;
4168
4169        /*
4170         * Let's call ->map_pages() first and use ->fault() as fallback
4171         * if page by the offset is not ready to be mapped (cold cache or
4172         * something).
4173         */
4174        if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
4175                if (likely(!userfaultfd_minor(vmf->vma))) {
4176                        ret = do_fault_around(vmf);
4177                        if (ret)
4178                                return ret;
4179                }
4180        }
4181
4182        ret = __do_fault(vmf);
4183        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4184                return ret;
4185
4186        ret |= finish_fault(vmf);
4187        unlock_page(vmf->page);
4188        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4189                put_page(vmf->page);
4190        return ret;
4191}
4192
4193static vm_fault_t do_cow_fault(struct vm_fault *vmf)
4194{
4195        struct vm_area_struct *vma = vmf->vma;
4196        vm_fault_t ret;
4197
4198        if (unlikely(anon_vma_prepare(vma)))
4199                return VM_FAULT_OOM;
4200
4201        vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
4202        if (!vmf->cow_page)
4203                return VM_FAULT_OOM;
4204
4205        if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
4206                put_page(vmf->cow_page);
4207                return VM_FAULT_OOM;
4208        }
4209        cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
4210
4211        ret = __do_fault(vmf);
4212        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4213                goto uncharge_out;
4214        if (ret & VM_FAULT_DONE_COW)
4215                return ret;
4216
4217        copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
4218        __SetPageUptodate(vmf->cow_page);
4219
4220        ret |= finish_fault(vmf);
4221        unlock_page(vmf->page);
4222        put_page(vmf->page);
4223        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4224                goto uncharge_out;
4225        return ret;
4226uncharge_out:
4227        put_page(vmf->cow_page);
4228        return ret;
4229}
4230
4231static vm_fault_t do_shared_fault(struct vm_fault *vmf)
4232{
4233        struct vm_area_struct *vma = vmf->vma;
4234        vm_fault_t ret, tmp;
4235
4236        ret = __do_fault(vmf);
4237        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
4238                return ret;
4239
4240        /*
4241         * Check if the backing address space wants to know that the page is
4242         * about to become writable
4243         */
4244        if (vma->vm_ops->page_mkwrite) {
4245                unlock_page(vmf->page);
4246                tmp = do_page_mkwrite(vmf);
4247                if (unlikely(!tmp ||
4248                                (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4249                        put_page(vmf->page);
4250                        return tmp;
4251                }
4252        }
4253
4254        ret |= finish_fault(vmf);
4255        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4256                                        VM_FAULT_RETRY))) {
4257                unlock_page(vmf->page);
4258                put_page(vmf->page);
4259                return ret;
4260        }
4261
4262        ret |= fault_dirty_shared_page(vmf);
4263        return ret;
4264}
4265
4266/*
4267 * We enter with non-exclusive mmap_lock (to exclude vma changes,
4268 * but allow concurrent faults).
4269 * The mmap_lock may have been released depending on flags and our
4270 * return value.  See filemap_fault() and __lock_page_or_retry().
4271 * If mmap_lock is released, vma may become invalid (for example
4272 * by other thread calling munmap()).
4273 */
4274static vm_fault_t do_fault(struct vm_fault *vmf)
4275{
4276        struct vm_area_struct *vma = vmf->vma;
4277        struct mm_struct *vm_mm = vma->vm_mm;
4278        vm_fault_t ret;
4279
4280        /*
4281         * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4282         */
4283        if (!vma->vm_ops->fault) {
4284                /*
4285                 * If we find a migration pmd entry or a none pmd entry, which
4286                 * should never happen, return SIGBUS
4287                 */
4288                if (unlikely(!pmd_present(*vmf->pmd)))
4289                        ret = VM_FAULT_SIGBUS;
4290                else {
4291                        vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4292                                                       vmf->pmd,
4293                                                       vmf->address,
4294                                                       &vmf->ptl);
4295                        /*
4296                         * Make sure this is not a temporary clearing of pte
4297                         * by holding ptl and checking again. A R/M/W update
4298                         * of pte involves: take ptl, clearing the pte so that
4299                         * we don't have concurrent modification by hardware
4300                         * followed by an update.
4301                         */
4302                        if (unlikely(pte_none(*vmf->pte)))
4303                                ret = VM_FAULT_SIGBUS;
4304                        else
4305                                ret = VM_FAULT_NOPAGE;
4306
4307                        pte_unmap_unlock(vmf->pte, vmf->ptl);
4308                }
4309        } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4310                ret = do_read_fault(vmf);
4311        else if (!(vma->vm_flags & VM_SHARED))
4312                ret = do_cow_fault(vmf);
4313        else
4314                ret = do_shared_fault(vmf);
4315
4316        /* preallocated pagetable is unused: free it */
4317        if (vmf->prealloc_pte) {
4318                pte_free(vm_mm, vmf->prealloc_pte);
4319                vmf->prealloc_pte = NULL;
4320        }
4321        return ret;
4322}
4323
4324int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4325                      unsigned long addr, int page_nid, int *flags)
4326{
4327        get_page(page);
4328
4329        count_vm_numa_event(NUMA_HINT_FAULTS);
4330        if (page_nid == numa_node_id()) {
4331                count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4332                *flags |= TNF_FAULT_LOCAL;
4333        }
4334
4335        return mpol_misplaced(page, vma, addr);
4336}
4337
4338static vm_fault_t do_numa_page(struct vm_fault *vmf)
4339{
4340        struct vm_area_struct *vma = vmf->vma;
4341        struct page *page = NULL;
4342        int page_nid = NUMA_NO_NODE;
4343        int last_cpupid;
4344        int target_nid;
4345        pte_t pte, old_pte;
4346        bool was_writable = pte_savedwrite(vmf->orig_pte);
4347        int flags = 0;
4348
4349        /*
4350         * The "pte" at this point cannot be used safely without
4351         * validation through pte_unmap_same(). It's of NUMA type but
4352         * the pfn may be screwed if the read is non atomic.
4353         */
4354        vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4355        spin_lock(vmf->ptl);
4356        if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4357                pte_unmap_unlock(vmf->pte, vmf->ptl);
4358                goto out;
4359        }
4360
4361        /* Get the normal PTE  */
4362        old_pte = ptep_get(vmf->pte);
4363        pte = pte_modify(old_pte, vma->vm_page_prot);
4364
4365        page = vm_normal_page(vma, vmf->address, pte);
4366        if (!page)
4367                goto out_map;
4368
4369        /* TODO: handle PTE-mapped THP */
4370        if (PageCompound(page))
4371                goto out_map;
4372
4373        /*
4374         * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4375         * much anyway since they can be in shared cache state. This misses
4376         * the case where a mapping is writable but the process never writes
4377         * to it but pte_write gets cleared during protection updates and
4378         * pte_dirty has unpredictable behaviour between PTE scan updates,
4379         * background writeback, dirty balancing and application behaviour.
4380         */
4381        if (!was_writable)
4382                flags |= TNF_NO_GROUP;
4383
4384        /*
4385         * Flag if the page is shared between multiple address spaces. This
4386         * is later used when determining whether to group tasks together
4387         */
4388        if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4389                flags |= TNF_SHARED;
4390
4391        last_cpupid = page_cpupid_last(page);
4392        page_nid = page_to_nid(page);
4393        target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4394                        &flags);
4395        if (target_nid == NUMA_NO_NODE) {
4396                put_page(page);
4397                goto out_map;
4398        }
4399        pte_unmap_unlock(vmf->pte, vmf->ptl);
4400
4401        /* Migrate to the requested node */
4402        if (migrate_misplaced_page(page, vma, target_nid)) {
4403                page_nid = target_nid;
4404                flags |= TNF_MIGRATED;
4405        } else {
4406                flags |= TNF_MIGRATE_FAIL;
4407                vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4408                spin_lock(vmf->ptl);
4409                if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4410                        pte_unmap_unlock(vmf->pte, vmf->ptl);
4411                        goto out;
4412                }
4413                goto out_map;
4414        }
4415
4416out:
4417        if (page_nid != NUMA_NO_NODE)
4418                task_numa_fault(last_cpupid, page_nid, 1, flags);
4419        return 0;
4420out_map:
4421        /*
4422         * Make it present again, depending on how arch implements
4423         * non-accessible ptes, some can allow access by kernel mode.
4424         */
4425        old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4426        pte = pte_modify(old_pte, vma->vm_page_prot);
4427        pte = pte_mkyoung(pte);
4428        if (was_writable)
4429                pte = pte_mkwrite(pte);
4430        ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4431        update_mmu_cache(vma, vmf->address, vmf->pte);
4432        pte_unmap_unlock(vmf->pte, vmf->ptl);
4433        goto out;
4434}
4435
4436static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4437{
4438        if (vma_is_anonymous(vmf->vma))
4439                return do_huge_pmd_anonymous_page(vmf);
4440        if (vmf->vma->vm_ops->huge_fault)
4441                return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4442        return VM_FAULT_FALLBACK;
4443}
4444
4445/* `inline' is required to avoid gcc 4.1.2 build error */
4446static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf)
4447{
4448        if (vma_is_anonymous(vmf->vma)) {
4449                if (userfaultfd_huge_pmd_wp(vmf->vma, vmf->orig_pmd))
4450                        return handle_userfault(vmf, VM_UFFD_WP);
4451                return do_huge_pmd_wp_page(vmf);
4452        }
4453        if (vmf->vma->vm_ops->huge_fault) {
4454                vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4455
4456                if (!(ret & VM_FAULT_FALLBACK))
4457                        return ret;
4458        }
4459
4460        /* COW or write-notify handled on pte level: split pmd. */
4461        __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4462
4463        return VM_FAULT_FALLBACK;
4464}
4465
4466static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4467{
4468#if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&                     \
4469        defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4470        /* No support for anonymous transparent PUD pages yet */
4471        if (vma_is_anonymous(vmf->vma))
4472                goto split;
4473        if (vmf->vma->vm_ops->huge_fault) {
4474                vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4475
4476                if (!(ret & VM_FAULT_FALLBACK))
4477                        return ret;
4478        }
4479split:
4480        /* COW or write-notify not handled on PUD level: split pud.*/
4481        __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4482#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4483        return VM_FAULT_FALLBACK;
4484}
4485
4486static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4487{
4488#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4489        /* No support for anonymous transparent PUD pages yet */
4490        if (vma_is_anonymous(vmf->vma))
4491                return VM_FAULT_FALLBACK;
4492        if (vmf->vma->vm_ops->huge_fault)
4493                return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4494#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4495        return VM_FAULT_FALLBACK;
4496}
4497
4498/*
4499 * These routines also need to handle stuff like marking pages dirty
4500 * and/or accessed for architectures that don't do it in hardware (most
4501 * RISC architectures).  The early dirtying is also good on the i386.
4502 *
4503 * There is also a hook called "update_mmu_cache()" that architectures
4504 * with external mmu caches can use to update those (ie the Sparc or
4505 * PowerPC hashed page tables that act as extended TLBs).
4506 *
4507 * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4508 * concurrent faults).
4509 *
4510 * The mmap_lock may have been released depending on flags and our return value.
4511 * See filemap_fault() and __lock_page_or_retry().
4512 */
4513static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4514{
4515        pte_t entry;
4516
4517        if (unlikely(pmd_none(*vmf->pmd))) {
4518                /*
4519                 * Leave __pte_alloc() until later: because vm_ops->fault may
4520                 * want to allocate huge page, and if we expose page table
4521                 * for an instant, it will be difficult to retract from
4522                 * concurrent faults and from rmap lookups.
4523                 */
4524                vmf->pte = NULL;
4525        } else {
4526                /*
4527                 * If a huge pmd materialized under us just retry later.  Use
4528                 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4529                 * of pmd_trans_huge() to ensure the pmd didn't become
4530                 * pmd_trans_huge under us and then back to pmd_none, as a
4531                 * result of MADV_DONTNEED running immediately after a huge pmd
4532                 * fault in a different thread of this mm, in turn leading to a
4533                 * misleading pmd_trans_huge() retval. All we have to ensure is
4534                 * that it is a regular pmd that we can walk with
4535                 * pte_offset_map() and we can do that through an atomic read
4536                 * in C, which is what pmd_trans_unstable() provides.
4537                 */
4538                if (pmd_devmap_trans_unstable(vmf->pmd))
4539                        return 0;
4540                /*
4541                 * A regular pmd is established and it can't morph into a huge
4542                 * pmd from under us anymore at this point because we hold the
4543                 * mmap_lock read mode and khugepaged takes it in write mode.
4544                 * So now it's safe to run pte_offset_map().
4545                 */
4546                vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4547                vmf->orig_pte = *vmf->pte;
4548
4549                /*
4550                 * some architectures can have larger ptes than wordsize,
4551                 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4552                 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4553                 * accesses.  The code below just needs a consistent view
4554                 * for the ifs and we later double check anyway with the
4555                 * ptl lock held. So here a barrier will do.
4556                 */
4557                barrier();
4558                if (pte_none(vmf->orig_pte)) {
4559                        pte_unmap(vmf->pte);
4560                        vmf->pte = NULL;
4561                }
4562        }
4563
4564        if (!vmf->pte) {
4565                if (vma_is_anonymous(vmf->vma))
4566                        return do_anonymous_page(vmf);
4567                else
4568                        return do_fault(vmf);
4569        }
4570
4571        if (!pte_present(vmf->orig_pte))
4572                return do_swap_page(vmf);
4573
4574        if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4575                return do_numa_page(vmf);
4576
4577        vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4578        spin_lock(vmf->ptl);
4579        entry = vmf->orig_pte;
4580        if (unlikely(!pte_same(*vmf->pte, entry))) {
4581                update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4582                goto unlock;
4583        }
4584        if (vmf->flags & FAULT_FLAG_WRITE) {
4585                if (!pte_write(entry))
4586                        return do_wp_page(vmf);
4587                entry = pte_mkdirty(entry);
4588        }
4589        entry = pte_mkyoung(entry);
4590        if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4591                                vmf->flags & FAULT_FLAG_WRITE)) {
4592                update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4593        } else {
4594                /* Skip spurious TLB flush for retried page fault */
4595                if (vmf->flags & FAULT_FLAG_TRIED)
4596                        goto unlock;
4597                /*
4598                 * This is needed only for protection faults but the arch code
4599                 * is not yet telling us if this is a protection fault or not.
4600                 * This still avoids useless tlb flushes for .text page faults
4601                 * with threads.
4602                 */
4603                if (vmf->flags & FAULT_FLAG_WRITE)
4604                        flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4605        }
4606unlock:
4607        pte_unmap_unlock(vmf->pte, vmf->ptl);
4608        return 0;
4609}
4610
4611/*
4612 * By the time we get here, we already hold the mm semaphore
4613 *
4614 * The mmap_lock may have been released depending on flags and our
4615 * return value.  See filemap_fault() and __lock_page_or_retry().
4616 */
4617static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4618                unsigned long address, unsigned int flags)
4619{
4620        struct vm_fault vmf = {
4621                .vma = vma,
4622                .address = address & PAGE_MASK,
4623                .flags = flags,
4624                .pgoff = linear_page_index(vma, address),
4625                .gfp_mask = __get_fault_gfp_mask(vma),
4626        };
4627        unsigned int dirty = flags & FAULT_FLAG_WRITE;
4628        struct mm_struct *mm = vma->vm_mm;
4629        pgd_t *pgd;
4630        p4d_t *p4d;
4631        vm_fault_t ret;
4632
4633        pgd = pgd_offset(mm, address);
4634        p4d = p4d_alloc(mm, pgd, address);
4635        if (!p4d)
4636                return VM_FAULT_OOM;
4637
4638        vmf.pud = pud_alloc(mm, p4d, address);
4639        if (!vmf.pud)
4640                return VM_FAULT_OOM;
4641retry_pud:
4642        if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4643                ret = create_huge_pud(&vmf);
4644                if (!(ret & VM_FAULT_FALLBACK))
4645                        return ret;
4646        } else {
4647                pud_t orig_pud = *vmf.pud;
4648
4649                barrier();
4650                if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4651
4652                        /* NUMA case for anonymous PUDs would go here */
4653
4654                        if (dirty && !pud_write(orig_pud)) {
4655                                ret = wp_huge_pud(&vmf, orig_pud);
4656                                if (!(ret & VM_FAULT_FALLBACK))
4657                                        return ret;
4658                        } else {
4659                                huge_pud_set_accessed(&vmf, orig_pud);
4660                                return 0;
4661                        }
4662                }
4663        }
4664
4665        vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4666        if (!vmf.pmd)
4667                return VM_FAULT_OOM;
4668
4669        /* Huge pud page fault raced with pmd_alloc? */
4670        if (pud_trans_unstable(vmf.pud))
4671                goto retry_pud;
4672
4673        if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4674                ret = create_huge_pmd(&vmf);
4675                if (!(ret & VM_FAULT_FALLBACK))
4676                        return ret;
4677        } else {
4678                vmf.orig_pmd = *vmf.pmd;
4679
4680                barrier();
4681                if (unlikely(is_swap_pmd(vmf.orig_pmd))) {
4682                        VM_BUG_ON(thp_migration_supported() &&
4683                                          !is_pmd_migration_entry(vmf.orig_pmd));
4684                        if (is_pmd_migration_entry(vmf.orig_pmd))
4685                                pmd_migration_entry_wait(mm, vmf.pmd);
4686                        return 0;
4687                }
4688                if (pmd_trans_huge(vmf.orig_pmd) || pmd_devmap(vmf.orig_pmd)) {
4689                        if (pmd_protnone(vmf.orig_pmd) && vma_is_accessible(vma))
4690                                return do_huge_pmd_numa_page(&vmf);
4691
4692                        if (dirty && !pmd_write(vmf.orig_pmd)) {
4693                                ret = wp_huge_pmd(&vmf);
4694                                if (!(ret & VM_FAULT_FALLBACK))
4695                                        return ret;
4696                        } else {
4697                                huge_pmd_set_accessed(&vmf);
4698                                return 0;
4699                        }
4700                }
4701        }
4702
4703        return handle_pte_fault(&vmf);
4704}
4705
4706/**
4707 * mm_account_fault - Do page fault accounting
4708 *
4709 * @regs: the pt_regs struct pointer.  When set to NULL, will skip accounting
4710 *        of perf event counters, but we'll still do the per-task accounting to
4711 *        the task who triggered this page fault.
4712 * @address: the faulted address.
4713 * @flags: the fault flags.
4714 * @ret: the fault retcode.
4715 *
4716 * This will take care of most of the page fault accounting.  Meanwhile, it
4717 * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4718 * updates.  However, note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4719 * still be in per-arch page fault handlers at the entry of page fault.
4720 */
4721static inline void mm_account_fault(struct pt_regs *regs,
4722                                    unsigned long address, unsigned int flags,
4723                                    vm_fault_t ret)
4724{
4725        bool major;
4726
4727        /*
4728         * We don't do accounting for some specific faults:
4729         *
4730         * - Unsuccessful faults (e.g. when the address wasn't valid).  That
4731         *   includes arch_vma_access_permitted() failing before reaching here.
4732         *   So this is not a "this many hardware page faults" counter.  We
4733         *   should use the hw profiling for that.
4734         *
4735         * - Incomplete faults (VM_FAULT_RETRY).  They will only be counted
4736         *   once they're completed.
4737         */
4738        if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4739                return;
4740
4741        /*
4742         * We define the fault as a major fault when the final successful fault
4743         * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4744         * handle it immediately previously).
4745         */
4746        major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4747
4748        if (major)
4749                current->maj_flt++;
4750        else
4751                current->min_flt++;
4752
4753        /*
4754         * If the fault is done for GUP, regs will be NULL.  We only do the
4755         * accounting for the per thread fault counters who triggered the
4756         * fault, and we skip the perf event updates.
4757         */
4758        if (!regs)
4759                return;
4760
4761        if (major)
4762                perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4763        else
4764                perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4765}
4766
4767/*
4768 * By the time we get here, we already hold the mm semaphore
4769 *
4770 * The mmap_lock may have been released depending on flags and our
4771 * return value.  See filemap_fault() and __lock_page_or_retry().
4772 */
4773vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4774                           unsigned int flags, struct pt_regs *regs)
4775{
4776        vm_fault_t ret;
4777
4778        __set_current_state(TASK_RUNNING);
4779
4780        count_vm_event(PGFAULT);
4781        count_memcg_event_mm(vma->vm_mm, PGFAULT);
4782
4783        /* do counter updates before entering really critical section. */
4784        check_sync_rss_stat(current);
4785
4786        if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4787                                            flags & FAULT_FLAG_INSTRUCTION,
4788                                            flags & FAULT_FLAG_REMOTE))
4789                return VM_FAULT_SIGSEGV;
4790
4791        /*
4792         * Enable the memcg OOM handling for faults triggered in user
4793         * space.  Kernel faults are handled more gracefully.
4794         */
4795        if (flags & FAULT_FLAG_USER)
4796                mem_cgroup_enter_user_fault();
4797
4798        if (unlikely(is_vm_hugetlb_page(vma)))
4799                ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4800        else
4801                ret = __handle_mm_fault(vma, address, flags);
4802
4803        if (flags & FAULT_FLAG_USER) {
4804                mem_cgroup_exit_user_fault();
4805                /*
4806                 * The task may have entered a memcg OOM situation but
4807                 * if the allocation error was handled gracefully (no
4808                 * VM_FAULT_OOM), there is no need to kill anything.
4809                 * Just clean up the OOM state peacefully.
4810                 */
4811                if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4812                        mem_cgroup_oom_synchronize(false);
4813        }
4814
4815        mm_account_fault(regs, address, flags, ret);
4816
4817        return ret;
4818}
4819EXPORT_SYMBOL_GPL(handle_mm_fault);
4820
4821#ifndef __PAGETABLE_P4D_FOLDED
4822/*
4823 * Allocate p4d page table.
4824 * We've already handled the fast-path in-line.
4825 */
4826int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4827{
4828        p4d_t *new = p4d_alloc_one(mm, address);
4829        if (!new)
4830                return -ENOMEM;
4831
4832        smp_wmb(); /* See comment in __pte_alloc */
4833
4834        spin_lock(&mm->page_table_lock);
4835        if (pgd_present(*pgd))          /* Another has populated it */
4836                p4d_free(mm, new);
4837        else
4838                pgd_populate(mm, pgd, new);
4839        spin_unlock(&mm->page_table_lock);
4840        return 0;
4841}
4842#endif /* __PAGETABLE_P4D_FOLDED */
4843
4844#ifndef __PAGETABLE_PUD_FOLDED
4845/*
4846 * Allocate page upper directory.
4847 * We've already handled the fast-path in-line.
4848 */
4849int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4850{
4851        pud_t *new = pud_alloc_one(mm, address);
4852        if (!new)
4853                return -ENOMEM;
4854
4855        smp_wmb(); /* See comment in __pte_alloc */
4856
4857        spin_lock(&mm->page_table_lock);
4858        if (!p4d_present(*p4d)) {
4859                mm_inc_nr_puds(mm);
4860                p4d_populate(mm, p4d, new);
4861        } else  /* Another has populated it */
4862                pud_free(mm, new);
4863        spin_unlock(&mm->page_table_lock);
4864        return 0;
4865}
4866#endif /* __PAGETABLE_PUD_FOLDED */
4867
4868#ifndef __PAGETABLE_PMD_FOLDED
4869/*
4870 * Allocate page middle directory.
4871 * We've already handled the fast-path in-line.
4872 */
4873int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4874{
4875        spinlock_t *ptl;
4876        pmd_t *new = pmd_alloc_one(mm, address);
4877        if (!new)
4878                return -ENOMEM;
4879
4880        smp_wmb(); /* See comment in __pte_alloc */
4881
4882        ptl = pud_lock(mm, pud);
4883        if (!pud_present(*pud)) {
4884                mm_inc_nr_pmds(mm);
4885                pud_populate(mm, pud, new);
4886        } else  /* Another has populated it */
4887                pmd_free(mm, new);
4888        spin_unlock(ptl);
4889        return 0;
4890}
4891#endif /* __PAGETABLE_PMD_FOLDED */
4892
4893int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4894                          struct mmu_notifier_range *range, pte_t **ptepp,
4895                          pmd_t **pmdpp, spinlock_t **ptlp)
4896{
4897        pgd_t *pgd;
4898        p4d_t *p4d;
4899        pud_t *pud;
4900        pmd_t *pmd;
4901        pte_t *ptep;
4902
4903        pgd = pgd_offset(mm, address);
4904        if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4905                goto out;
4906
4907        p4d = p4d_offset(pgd, address);
4908        if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4909                goto out;
4910
4911        pud = pud_offset(p4d, address);
4912        if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4913                goto out;
4914
4915        pmd = pmd_offset(pud, address);
4916        VM_BUG_ON(pmd_trans_huge(*pmd));
4917
4918        if (pmd_huge(*pmd)) {
4919                if (!pmdpp)
4920                        goto out;
4921
4922                if (range) {
4923                        mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4924                                                NULL, mm, address & PMD_MASK,
4925                                                (address & PMD_MASK) + PMD_SIZE);
4926                        mmu_notifier_invalidate_range_start(range);
4927                }
4928                *ptlp = pmd_lock(mm, pmd);
4929                if (pmd_huge(*pmd)) {
4930                        *pmdpp = pmd;
4931                        return 0;
4932                }
4933                spin_unlock(*ptlp);
4934                if (range)
4935                        mmu_notifier_invalidate_range_end(range);
4936        }
4937
4938        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4939                goto out;
4940
4941        if (range) {
4942                mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4943                                        address & PAGE_MASK,
4944                                        (address & PAGE_MASK) + PAGE_SIZE);
4945                mmu_notifier_invalidate_range_start(range);
4946        }
4947        ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4948        if (!pte_present(*ptep))
4949                goto unlock;
4950        *ptepp = ptep;
4951        return 0;
4952unlock:
4953        pte_unmap_unlock(ptep, *ptlp);
4954        if (range)
4955                mmu_notifier_invalidate_range_end(range);
4956out:
4957        return -EINVAL;
4958}
4959
4960/**
4961 * follow_pte - look up PTE at a user virtual address
4962 * @mm: the mm_struct of the target address space
4963 * @address: user virtual address
4964 * @ptepp: location to store found PTE
4965 * @ptlp: location to store the lock for the PTE
4966 *
4967 * On a successful return, the pointer to the PTE is stored in @ptepp;
4968 * the corresponding lock is taken and its location is stored in @ptlp.
4969 * The contents of the PTE are only stable until @ptlp is released;
4970 * any further use, if any, must be protected against invalidation
4971 * with MMU notifiers.
4972 *
4973 * Only IO mappings and raw PFN mappings are allowed.  The mmap semaphore
4974 * should be taken for read.
4975 *
4976 * KVM uses this function.  While it is arguably less bad than ``follow_pfn``,
4977 * it is not a good general-purpose API.
4978 *
4979 * Return: zero on success, -ve otherwise.
4980 */
4981int follow_pte(struct mm_struct *mm, unsigned long address,
4982               pte_t **ptepp, spinlock_t **ptlp)
4983{
4984        return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4985}
4986EXPORT_SYMBOL_GPL(follow_pte);
4987
4988/**
4989 * follow_pfn - look up PFN at a user virtual address
4990 * @vma: memory mapping
4991 * @address: user virtual address
4992 * @pfn: location to store found PFN
4993 *
4994 * Only IO mappings and raw PFN mappings are allowed.
4995 *
4996 * This function does not allow the caller to read the permissions
4997 * of the PTE.  Do not use it.
4998 *
4999 * Return: zero and the pfn at @pfn on success, -ve otherwise.
5000 */
5001int follow_pfn(struct vm_area_struct *vma, unsigned long address,
5002        unsigned long *pfn)
5003{
5004        int ret = -EINVAL;
5005        spinlock_t *ptl;
5006        pte_t *ptep;
5007
5008        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5009                return ret;
5010
5011        ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
5012        if (ret)
5013                return ret;
5014        *pfn = pte_pfn(*ptep);
5015        pte_unmap_unlock(ptep, ptl);
5016        return 0;
5017}
5018EXPORT_SYMBOL(follow_pfn);
5019
5020#ifdef CONFIG_HAVE_IOREMAP_PROT
5021int follow_phys(struct vm_area_struct *vma,
5022                unsigned long address, unsigned int flags,
5023                unsigned long *prot, resource_size_t *phys)
5024{
5025        int ret = -EINVAL;
5026        pte_t *ptep, pte;
5027        spinlock_t *ptl;
5028
5029        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5030                goto out;
5031
5032        if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
5033                goto out;
5034        pte = *ptep;
5035
5036        if ((flags & FOLL_WRITE) && !pte_write(pte))
5037                goto unlock;
5038
5039        *prot = pgprot_val(pte_pgprot(pte));
5040        *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5041
5042        ret = 0;
5043unlock:
5044        pte_unmap_unlock(ptep, ptl);
5045out:
5046        return ret;
5047}
5048
5049/**
5050 * generic_access_phys - generic implementation for iomem mmap access
5051 * @vma: the vma to access
5052 * @addr: userspace address, not relative offset within @vma
5053 * @buf: buffer to read/write
5054 * @len: length of transfer
5055 * @write: set to FOLL_WRITE when writing, otherwise reading
5056 *
5057 * This is a generic implementation for &vm_operations_struct.access for an
5058 * iomem mapping. This callback is used by access_process_vm() when the @vma is
5059 * not page based.
5060 */
5061int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
5062                        void *buf, int len, int write)
5063{
5064        resource_size_t phys_addr;
5065        unsigned long prot = 0;
5066        void __iomem *maddr;
5067        pte_t *ptep, pte;
5068        spinlock_t *ptl;
5069        int offset = offset_in_page(addr);
5070        int ret = -EINVAL;
5071
5072        if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
5073                return -EINVAL;
5074
5075retry:
5076        if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5077                return -EINVAL;
5078        pte = *ptep;
5079        pte_unmap_unlock(ptep, ptl);
5080
5081        prot = pgprot_val(pte_pgprot(pte));
5082        phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
5083
5084        if ((write & FOLL_WRITE) && !pte_write(pte))
5085                return -EINVAL;
5086
5087        maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
5088        if (!maddr)
5089                return -ENOMEM;
5090
5091        if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
5092                goto out_unmap;
5093
5094        if (!pte_same(pte, *ptep)) {
5095                pte_unmap_unlock(ptep, ptl);
5096                iounmap(maddr);
5097
5098                goto retry;
5099        }
5100
5101        if (write)
5102                memcpy_toio(maddr + offset, buf, len);
5103        else
5104                memcpy_fromio(buf, maddr + offset, len);
5105        ret = len;
5106        pte_unmap_unlock(ptep, ptl);
5107out_unmap:
5108        iounmap(maddr);
5109
5110        return ret;
5111}
5112EXPORT_SYMBOL_GPL(generic_access_phys);
5113#endif
5114
5115/*
5116 * Access another process' address space as given in mm.
5117 */
5118int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
5119                       int len, unsigned int gup_flags)
5120{
5121        struct vm_area_struct *vma;
5122        void *old_buf = buf;
5123        int write = gup_flags & FOLL_WRITE;
5124
5125        if (mmap_read_lock_killable(mm))
5126                return 0;
5127
5128        /* ignore errors, just check how much was successfully transferred */
5129        while (len) {
5130                int bytes, ret, offset;
5131                void *maddr;
5132                struct page *page = NULL;
5133
5134                ret = get_user_pages_remote(mm, addr, 1,
5135                                gup_flags, &page, &vma, NULL);
5136                if (ret <= 0) {
5137#ifndef CONFIG_HAVE_IOREMAP_PROT
5138                        break;
5139#else
5140                        /*
5141                         * Check if this is a VM_IO | VM_PFNMAP VMA, which
5142                         * we can access using slightly different code.
5143                         */
5144                        vma = vma_lookup(mm, addr);
5145                        if (!vma)
5146                                break;
5147                        if (vma->vm_ops && vma->vm_ops->access)
5148                                ret = vma->vm_ops->access(vma, addr, buf,
5149                                                          len, write);
5150                        if (ret <= 0)
5151                                break;
5152                        bytes = ret;
5153#endif
5154                } else {
5155                        bytes = len;
5156                        offset = addr & (PAGE_SIZE-1);
5157                        if (bytes > PAGE_SIZE-offset)
5158                                bytes = PAGE_SIZE-offset;
5159
5160                        maddr = kmap(page);
5161                        if (write) {
5162                                copy_to_user_page(vma, page, addr,
5163                                                  maddr + offset, buf, bytes);
5164                                set_page_dirty_lock(page);
5165                        } else {
5166                                copy_from_user_page(vma, page, addr,
5167                                                    buf, maddr + offset, bytes);
5168                        }
5169                        kunmap(page);
5170                        put_page(page);
5171                }
5172                len -= bytes;
5173                buf += bytes;
5174                addr += bytes;
5175        }
5176        mmap_read_unlock(mm);
5177
5178        return buf - old_buf;
5179}
5180
5181/**
5182 * access_remote_vm - access another process' address space
5183 * @mm:         the mm_struct of the target address space
5184 * @addr:       start address to access
5185 * @buf:        source or destination buffer
5186 * @len:        number of bytes to transfer
5187 * @gup_flags:  flags modifying lookup behaviour
5188 *
5189 * The caller must hold a reference on @mm.
5190 *
5191 * Return: number of bytes copied from source to destination.
5192 */
5193int access_remote_vm(struct mm_struct *mm, unsigned long addr,
5194                void *buf, int len, unsigned int gup_flags)
5195{
5196        return __access_remote_vm(mm, addr, buf, len, gup_flags);
5197}
5198
5199/*
5200 * Access another process' address space.
5201 * Source/target buffer must be kernel space,
5202 * Do not walk the page table directly, use get_user_pages
5203 */
5204int access_process_vm(struct task_struct *tsk, unsigned long addr,
5205                void *buf, int len, unsigned int gup_flags)
5206{
5207        struct mm_struct *mm;
5208        int ret;
5209
5210        mm = get_task_mm(tsk);
5211        if (!mm)
5212                return 0;
5213
5214        ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
5215
5216        mmput(mm);
5217
5218        return ret;
5219}
5220EXPORT_SYMBOL_GPL(access_process_vm);
5221
5222/*
5223 * Print the name of a VMA.
5224 */
5225void print_vma_addr(char *prefix, unsigned long ip)
5226{
5227        struct mm_struct *mm = current->mm;
5228        struct vm_area_struct *vma;
5229
5230        /*
5231         * we might be running from an atomic context so we cannot sleep
5232         */
5233        if (!mmap_read_trylock(mm))
5234                return;
5235
5236        vma = find_vma(mm, ip);
5237        if (vma && vma->vm_file) {
5238                struct file *f = vma->vm_file;
5239                char *buf = (char *)__get_free_page(GFP_NOWAIT);
5240                if (buf) {
5241                        char *p;
5242
5243                        p = file_path(f, buf, PAGE_SIZE);
5244                        if (IS_ERR(p))
5245                                p = "?";
5246                        printk("%s%s[%lx+%lx]", prefix, kbasename(p),
5247                                        vma->vm_start,
5248                                        vma->vm_end - vma->vm_start);
5249                        free_page((unsigned long)buf);
5250                }
5251        }
5252        mmap_read_unlock(mm);
5253}
5254
5255#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5256void __might_fault(const char *file, int line)
5257{
5258        /*
5259         * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5260         * holding the mmap_lock, this is safe because kernel memory doesn't
5261         * get paged out, therefore we'll never actually fault, and the
5262         * below annotations will generate false positives.
5263         */
5264        if (uaccess_kernel())
5265                return;
5266        if (pagefault_disabled())
5267                return;
5268        __might_sleep(file, line, 0);
5269#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5270        if (current->mm)
5271                might_lock_read(&current->mm->mmap_lock);
5272#endif
5273}
5274EXPORT_SYMBOL(__might_fault);
5275#endif
5276
5277#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5278/*
5279 * Process all subpages of the specified huge page with the specified
5280 * operation.  The target subpage will be processed last to keep its
5281 * cache lines hot.
5282 */
5283static inline void process_huge_page(
5284        unsigned long addr_hint, unsigned int pages_per_huge_page,
5285        void (*process_subpage)(unsigned long addr, int idx, void *arg),
5286        void *arg)
5287{
5288        int i, n, base, l;
5289        unsigned long addr = addr_hint &
5290                ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5291
5292        /* Process target subpage last to keep its cache lines hot */
5293        might_sleep();
5294        n = (addr_hint - addr) / PAGE_SIZE;
5295        if (2 * n <= pages_per_huge_page) {
5296                /* If target subpage in first half of huge page */
5297                base = 0;
5298                l = n;
5299                /* Process subpages at the end of huge page */
5300                for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5301                        cond_resched();
5302                        process_subpage(addr + i * PAGE_SIZE, i, arg);
5303                }
5304        } else {
5305                /* If target subpage in second half of huge page */
5306                base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5307                l = pages_per_huge_page - n;
5308                /* Process subpages at the begin of huge page */
5309                for (i = 0; i < base; i++) {
5310                        cond_resched();
5311                        process_subpage(addr + i * PAGE_SIZE, i, arg);
5312                }
5313        }
5314        /*
5315         * Process remaining subpages in left-right-left-right pattern
5316         * towards the target subpage
5317         */
5318        for (i = 0; i < l; i++) {
5319                int left_idx = base + i;
5320                int right_idx = base + 2 * l - 1 - i;
5321
5322                cond_resched();
5323                process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5324                cond_resched();
5325                process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5326        }
5327}
5328
5329static void clear_gigantic_page(struct page *page,
5330                                unsigned long addr,
5331                                unsigned int pages_per_huge_page)
5332{
5333        int i;
5334        struct page *p = page;
5335
5336        might_sleep();
5337        for (i = 0; i < pages_per_huge_page;
5338             i++, p = mem_map_next(p, page, i)) {
5339                cond_resched();
5340                clear_user_highpage(p, addr + i * PAGE_SIZE);
5341        }
5342}
5343
5344static void clear_subpage(unsigned long addr, int idx, void *arg)
5345{
5346        struct page *page = arg;
5347
5348        clear_user_highpage(page + idx, addr);
5349}
5350
5351void clear_huge_page(struct page *page,
5352                     unsigned long addr_hint, unsigned int pages_per_huge_page)
5353{
5354        unsigned long addr = addr_hint &
5355                ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5356
5357        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5358                clear_gigantic_page(page, addr, pages_per_huge_page);
5359                return;
5360        }
5361
5362        process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5363}
5364
5365static void copy_user_gigantic_page(struct page *dst, struct page *src,
5366                                    unsigned long addr,
5367                                    struct vm_area_struct *vma,
5368                                    unsigned int pages_per_huge_page)
5369{
5370        int i;
5371        struct page *dst_base = dst;
5372        struct page *src_base = src;
5373
5374        for (i = 0; i < pages_per_huge_page; ) {
5375                cond_resched();
5376                copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5377
5378                i++;
5379                dst = mem_map_next(dst, dst_base, i);
5380                src = mem_map_next(src, src_base, i);
5381        }
5382}
5383
5384struct copy_subpage_arg {
5385        struct page *dst;
5386        struct page *src;
5387        struct vm_area_struct *vma;
5388};
5389
5390static void copy_subpage(unsigned long addr, int idx, void *arg)
5391{
5392        struct copy_subpage_arg *copy_arg = arg;
5393
5394        copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5395                           addr, copy_arg->vma);
5396}
5397
5398void copy_user_huge_page(struct page *dst, struct page *src,
5399                         unsigned long addr_hint, struct vm_area_struct *vma,
5400                         unsigned int pages_per_huge_page)
5401{
5402        unsigned long addr = addr_hint &
5403                ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5404        struct copy_subpage_arg arg = {
5405                .dst = dst,
5406                .src = src,
5407                .vma = vma,
5408        };
5409
5410        if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5411                copy_user_gigantic_page(dst, src, addr, vma,
5412                                        pages_per_huge_page);
5413                return;
5414        }
5415
5416        process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5417}
5418
5419long copy_huge_page_from_user(struct page *dst_page,
5420                                const void __user *usr_src,
5421                                unsigned int pages_per_huge_page,
5422                                bool allow_pagefault)
5423{
5424        void *src = (void *)usr_src;
5425        void *page_kaddr;
5426        unsigned long i, rc = 0;
5427        unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5428        struct page *subpage = dst_page;
5429
5430        for (i = 0; i < pages_per_huge_page;
5431             i++, subpage = mem_map_next(subpage, dst_page, i)) {
5432                if (allow_pagefault)
5433                        page_kaddr = kmap(subpage);
5434                else
5435                        page_kaddr = kmap_atomic(subpage);
5436                rc = copy_from_user(page_kaddr,
5437                                (const void __user *)(src + i * PAGE_SIZE),
5438                                PAGE_SIZE);
5439                if (allow_pagefault)
5440                        kunmap(subpage);
5441                else
5442                        kunmap_atomic(page_kaddr);
5443
5444                ret_val -= (PAGE_SIZE - rc);
5445                if (rc)
5446                        break;
5447
5448                cond_resched();
5449        }
5450        return ret_val;
5451}
5452#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5453
5454#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5455
5456static struct kmem_cache *page_ptl_cachep;
5457
5458void __init ptlock_cache_init(void)
5459{
5460        page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5461                        SLAB_PANIC, NULL);
5462}
5463
5464bool ptlock_alloc(struct page *page)
5465{
5466        spinlock_t *ptl;
5467
5468        ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5469        if (!ptl)
5470                return false;
5471        page->ptl = ptl;
5472        return true;
5473}
5474
5475void ptlock_free(struct page *page)
5476{
5477        kmem_cache_free(page_ptl_cachep, page->ptl);
5478}
5479#endif
5480