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