Back to home page

LXR

 
 

    


0001 /*
0002  *  linux/mm/memory.c
0003  *
0004  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
0005  */
0006 
0007 /*
0008  * demand-loading started 01.12.91 - seems it is high on the list of
0009  * things wanted, and it should be easy to implement. - Linus
0010  */
0011 
0012 /*
0013  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
0014  * pages started 02.12.91, seems to work. - Linus.
0015  *
0016  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
0017  * would have taken more than the 6M I have free, but it worked well as
0018  * far as I could see.
0019  *
0020  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
0021  */
0022 
0023 /*
0024  * Real VM (paging to/from disk) started 18.12.91. Much more work and
0025  * thought has to go into this. Oh, well..
0026  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
0027  *      Found it. Everything seems to work now.
0028  * 20.12.91  -  Ok, making the swap-device changeable like the root.
0029  */
0030 
0031 /*
0032  * 05.04.94  -  Multi-page memory management added for v1.1.
0033  *      Idea by Alex Bligh (alex@cconcepts.co.uk)
0034  *
0035  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
0036  *      (Gerhard.Wichert@pdb.siemens.de)
0037  *
0038  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
0039  */
0040 
0041 #include <linux/kernel_stat.h>
0042 #include <linux/mm.h>
0043 #include <linux/hugetlb.h>
0044 #include <linux/mman.h>
0045 #include <linux/swap.h>
0046 #include <linux/highmem.h>
0047 #include <linux/pagemap.h>
0048 #include <linux/ksm.h>
0049 #include <linux/rmap.h>
0050 #include <linux/export.h>
0051 #include <linux/delayacct.h>
0052 #include <linux/init.h>
0053 #include <linux/pfn_t.h>
0054 #include <linux/writeback.h>
0055 #include <linux/memcontrol.h>
0056 #include <linux/mmu_notifier.h>
0057 #include <linux/kallsyms.h>
0058 #include <linux/swapops.h>
0059 #include <linux/elf.h>
0060 #include <linux/gfp.h>
0061 #include <linux/migrate.h>
0062 #include <linux/string.h>
0063 #include <linux/dma-debug.h>
0064 #include <linux/debugfs.h>
0065 #include <linux/userfaultfd_k.h>
0066 #include <linux/dax.h>
0067 
0068 #include <asm/io.h>
0069 #include <asm/mmu_context.h>
0070 #include <asm/pgalloc.h>
0071 #include <linux/uaccess.h>
0072 #include <asm/tlb.h>
0073 #include <asm/tlbflush.h>
0074 #include <asm/pgtable.h>
0075 
0076 #include "internal.h"
0077 
0078 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
0079 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
0080 #endif
0081 
0082 #ifndef CONFIG_NEED_MULTIPLE_NODES
0083 /* use the per-pgdat data instead for discontigmem - mbligh */
0084 unsigned long max_mapnr;
0085 struct page *mem_map;
0086 
0087 EXPORT_SYMBOL(max_mapnr);
0088 EXPORT_SYMBOL(mem_map);
0089 #endif
0090 
0091 /*
0092  * A number of key systems in x86 including ioremap() rely on the assumption
0093  * that high_memory defines the upper bound on direct map memory, then end
0094  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
0095  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
0096  * and ZONE_HIGHMEM.
0097  */
0098 void * high_memory;
0099 
0100 EXPORT_SYMBOL(high_memory);
0101 
0102 /*
0103  * Randomize the address space (stacks, mmaps, brk, etc.).
0104  *
0105  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
0106  *   as ancient (libc5 based) binaries can segfault. )
0107  */
0108 int randomize_va_space __read_mostly =
0109 #ifdef CONFIG_COMPAT_BRK
0110                     1;
0111 #else
0112                     2;
0113 #endif
0114 
0115 static int __init disable_randmaps(char *s)
0116 {
0117     randomize_va_space = 0;
0118     return 1;
0119 }
0120 __setup("norandmaps", disable_randmaps);
0121 
0122 unsigned long zero_pfn __read_mostly;
0123 unsigned long highest_memmap_pfn __read_mostly;
0124 
0125 EXPORT_SYMBOL(zero_pfn);
0126 
0127 /*
0128  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
0129  */
0130 static int __init init_zero_pfn(void)
0131 {
0132     zero_pfn = page_to_pfn(ZERO_PAGE(0));
0133     return 0;
0134 }
0135 core_initcall(init_zero_pfn);
0136 
0137 
0138 #if defined(SPLIT_RSS_COUNTING)
0139 
0140 void sync_mm_rss(struct mm_struct *mm)
0141 {
0142     int i;
0143 
0144     for (i = 0; i < NR_MM_COUNTERS; i++) {
0145         if (current->rss_stat.count[i]) {
0146             add_mm_counter(mm, i, current->rss_stat.count[i]);
0147             current->rss_stat.count[i] = 0;
0148         }
0149     }
0150     current->rss_stat.events = 0;
0151 }
0152 
0153 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
0154 {
0155     struct task_struct *task = current;
0156 
0157     if (likely(task->mm == mm))
0158         task->rss_stat.count[member] += val;
0159     else
0160         add_mm_counter(mm, member, val);
0161 }
0162 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
0163 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
0164 
0165 /* sync counter once per 64 page faults */
0166 #define TASK_RSS_EVENTS_THRESH  (64)
0167 static void check_sync_rss_stat(struct task_struct *task)
0168 {
0169     if (unlikely(task != current))
0170         return;
0171     if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
0172         sync_mm_rss(task->mm);
0173 }
0174 #else /* SPLIT_RSS_COUNTING */
0175 
0176 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
0177 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
0178 
0179 static void check_sync_rss_stat(struct task_struct *task)
0180 {
0181 }
0182 
0183 #endif /* SPLIT_RSS_COUNTING */
0184 
0185 #ifdef HAVE_GENERIC_MMU_GATHER
0186 
0187 static bool tlb_next_batch(struct mmu_gather *tlb)
0188 {
0189     struct mmu_gather_batch *batch;
0190 
0191     batch = tlb->active;
0192     if (batch->next) {
0193         tlb->active = batch->next;
0194         return true;
0195     }
0196 
0197     if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
0198         return false;
0199 
0200     batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
0201     if (!batch)
0202         return false;
0203 
0204     tlb->batch_count++;
0205     batch->next = NULL;
0206     batch->nr   = 0;
0207     batch->max  = MAX_GATHER_BATCH;
0208 
0209     tlb->active->next = batch;
0210     tlb->active = batch;
0211 
0212     return true;
0213 }
0214 
0215 /* tlb_gather_mmu
0216  *  Called to initialize an (on-stack) mmu_gather structure for page-table
0217  *  tear-down from @mm. The @fullmm argument is used when @mm is without
0218  *  users and we're going to destroy the full address space (exit/execve).
0219  */
0220 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
0221 {
0222     tlb->mm = mm;
0223 
0224     /* Is it from 0 to ~0? */
0225     tlb->fullmm     = !(start | (end+1));
0226     tlb->need_flush_all = 0;
0227     tlb->local.next = NULL;
0228     tlb->local.nr   = 0;
0229     tlb->local.max  = ARRAY_SIZE(tlb->__pages);
0230     tlb->active     = &tlb->local;
0231     tlb->batch_count = 0;
0232 
0233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
0234     tlb->batch = NULL;
0235 #endif
0236     tlb->page_size = 0;
0237 
0238     __tlb_reset_range(tlb);
0239 }
0240 
0241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
0242 {
0243     if (!tlb->end)
0244         return;
0245 
0246     tlb_flush(tlb);
0247     mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
0248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
0249     tlb_table_flush(tlb);
0250 #endif
0251     __tlb_reset_range(tlb);
0252 }
0253 
0254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
0255 {
0256     struct mmu_gather_batch *batch;
0257 
0258     for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
0259         free_pages_and_swap_cache(batch->pages, batch->nr);
0260         batch->nr = 0;
0261     }
0262     tlb->active = &tlb->local;
0263 }
0264 
0265 void tlb_flush_mmu(struct mmu_gather *tlb)
0266 {
0267     tlb_flush_mmu_tlbonly(tlb);
0268     tlb_flush_mmu_free(tlb);
0269 }
0270 
0271 /* tlb_finish_mmu
0272  *  Called at the end of the shootdown operation to free up any resources
0273  *  that were required.
0274  */
0275 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
0276 {
0277     struct mmu_gather_batch *batch, *next;
0278 
0279     tlb_flush_mmu(tlb);
0280 
0281     /* keep the page table cache within bounds */
0282     check_pgt_cache();
0283 
0284     for (batch = tlb->local.next; batch; batch = next) {
0285         next = batch->next;
0286         free_pages((unsigned long)batch, 0);
0287     }
0288     tlb->local.next = NULL;
0289 }
0290 
0291 /* __tlb_remove_page
0292  *  Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
0293  *  handling the additional races in SMP caused by other CPUs caching valid
0294  *  mappings in their TLBs. Returns the number of free page slots left.
0295  *  When out of page slots we must call tlb_flush_mmu().
0296  *returns true if the caller should flush.
0297  */
0298 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
0299 {
0300     struct mmu_gather_batch *batch;
0301 
0302     VM_BUG_ON(!tlb->end);
0303     VM_WARN_ON(tlb->page_size != page_size);
0304 
0305     batch = tlb->active;
0306     /*
0307      * Add the page and check if we are full. If so
0308      * force a flush.
0309      */
0310     batch->pages[batch->nr++] = page;
0311     if (batch->nr == batch->max) {
0312         if (!tlb_next_batch(tlb))
0313             return true;
0314         batch = tlb->active;
0315     }
0316     VM_BUG_ON_PAGE(batch->nr > batch->max, page);
0317 
0318     return false;
0319 }
0320 
0321 #endif /* HAVE_GENERIC_MMU_GATHER */
0322 
0323 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
0324 
0325 /*
0326  * See the comment near struct mmu_table_batch.
0327  */
0328 
0329 static void tlb_remove_table_smp_sync(void *arg)
0330 {
0331     /* Simply deliver the interrupt */
0332 }
0333 
0334 static void tlb_remove_table_one(void *table)
0335 {
0336     /*
0337      * This isn't an RCU grace period and hence the page-tables cannot be
0338      * assumed to be actually RCU-freed.
0339      *
0340      * It is however sufficient for software page-table walkers that rely on
0341      * IRQ disabling. See the comment near struct mmu_table_batch.
0342      */
0343     smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
0344     __tlb_remove_table(table);
0345 }
0346 
0347 static void tlb_remove_table_rcu(struct rcu_head *head)
0348 {
0349     struct mmu_table_batch *batch;
0350     int i;
0351 
0352     batch = container_of(head, struct mmu_table_batch, rcu);
0353 
0354     for (i = 0; i < batch->nr; i++)
0355         __tlb_remove_table(batch->tables[i]);
0356 
0357     free_page((unsigned long)batch);
0358 }
0359 
0360 void tlb_table_flush(struct mmu_gather *tlb)
0361 {
0362     struct mmu_table_batch **batch = &tlb->batch;
0363 
0364     if (*batch) {
0365         call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
0366         *batch = NULL;
0367     }
0368 }
0369 
0370 void tlb_remove_table(struct mmu_gather *tlb, void *table)
0371 {
0372     struct mmu_table_batch **batch = &tlb->batch;
0373 
0374     /*
0375      * When there's less then two users of this mm there cannot be a
0376      * concurrent page-table walk.
0377      */
0378     if (atomic_read(&tlb->mm->mm_users) < 2) {
0379         __tlb_remove_table(table);
0380         return;
0381     }
0382 
0383     if (*batch == NULL) {
0384         *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
0385         if (*batch == NULL) {
0386             tlb_remove_table_one(table);
0387             return;
0388         }
0389         (*batch)->nr = 0;
0390     }
0391     (*batch)->tables[(*batch)->nr++] = table;
0392     if ((*batch)->nr == MAX_TABLE_BATCH)
0393         tlb_table_flush(tlb);
0394 }
0395 
0396 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
0397 
0398 /*
0399  * Note: this doesn't free the actual pages themselves. That
0400  * has been handled earlier when unmapping all the memory regions.
0401  */
0402 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
0403                unsigned long addr)
0404 {
0405     pgtable_t token = pmd_pgtable(*pmd);
0406     pmd_clear(pmd);
0407     pte_free_tlb(tlb, token, addr);
0408     atomic_long_dec(&tlb->mm->nr_ptes);
0409 }
0410 
0411 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
0412                 unsigned long addr, unsigned long end,
0413                 unsigned long floor, unsigned long ceiling)
0414 {
0415     pmd_t *pmd;
0416     unsigned long next;
0417     unsigned long start;
0418 
0419     start = addr;
0420     pmd = pmd_offset(pud, addr);
0421     do {
0422         next = pmd_addr_end(addr, end);
0423         if (pmd_none_or_clear_bad(pmd))
0424             continue;
0425         free_pte_range(tlb, pmd, addr);
0426     } while (pmd++, addr = next, addr != end);
0427 
0428     start &= PUD_MASK;
0429     if (start < floor)
0430         return;
0431     if (ceiling) {
0432         ceiling &= PUD_MASK;
0433         if (!ceiling)
0434             return;
0435     }
0436     if (end - 1 > ceiling - 1)
0437         return;
0438 
0439     pmd = pmd_offset(pud, start);
0440     pud_clear(pud);
0441     pmd_free_tlb(tlb, pmd, start);
0442     mm_dec_nr_pmds(tlb->mm);
0443 }
0444 
0445 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
0446                 unsigned long addr, unsigned long end,
0447                 unsigned long floor, unsigned long ceiling)
0448 {
0449     pud_t *pud;
0450     unsigned long next;
0451     unsigned long start;
0452 
0453     start = addr;
0454     pud = pud_offset(pgd, addr);
0455     do {
0456         next = pud_addr_end(addr, end);
0457         if (pud_none_or_clear_bad(pud))
0458             continue;
0459         free_pmd_range(tlb, pud, addr, next, floor, ceiling);
0460     } while (pud++, addr = next, addr != end);
0461 
0462     start &= PGDIR_MASK;
0463     if (start < floor)
0464         return;
0465     if (ceiling) {
0466         ceiling &= PGDIR_MASK;
0467         if (!ceiling)
0468             return;
0469     }
0470     if (end - 1 > ceiling - 1)
0471         return;
0472 
0473     pud = pud_offset(pgd, start);
0474     pgd_clear(pgd);
0475     pud_free_tlb(tlb, pud, start);
0476 }
0477 
0478 /*
0479  * This function frees user-level page tables of a process.
0480  */
0481 void free_pgd_range(struct mmu_gather *tlb,
0482             unsigned long addr, unsigned long end,
0483             unsigned long floor, unsigned long ceiling)
0484 {
0485     pgd_t *pgd;
0486     unsigned long next;
0487 
0488     /*
0489      * The next few lines have given us lots of grief...
0490      *
0491      * Why are we testing PMD* at this top level?  Because often
0492      * there will be no work to do at all, and we'd prefer not to
0493      * go all the way down to the bottom just to discover that.
0494      *
0495      * Why all these "- 1"s?  Because 0 represents both the bottom
0496      * of the address space and the top of it (using -1 for the
0497      * top wouldn't help much: the masks would do the wrong thing).
0498      * The rule is that addr 0 and floor 0 refer to the bottom of
0499      * the address space, but end 0 and ceiling 0 refer to the top
0500      * Comparisons need to use "end - 1" and "ceiling - 1" (though
0501      * that end 0 case should be mythical).
0502      *
0503      * Wherever addr is brought up or ceiling brought down, we must
0504      * be careful to reject "the opposite 0" before it confuses the
0505      * subsequent tests.  But what about where end is brought down
0506      * by PMD_SIZE below? no, end can't go down to 0 there.
0507      *
0508      * Whereas we round start (addr) and ceiling down, by different
0509      * masks at different levels, in order to test whether a table
0510      * now has no other vmas using it, so can be freed, we don't
0511      * bother to round floor or end up - the tests don't need that.
0512      */
0513 
0514     addr &= PMD_MASK;
0515     if (addr < floor) {
0516         addr += PMD_SIZE;
0517         if (!addr)
0518             return;
0519     }
0520     if (ceiling) {
0521         ceiling &= PMD_MASK;
0522         if (!ceiling)
0523             return;
0524     }
0525     if (end - 1 > ceiling - 1)
0526         end -= PMD_SIZE;
0527     if (addr > end - 1)
0528         return;
0529     /*
0530      * We add page table cache pages with PAGE_SIZE,
0531      * (see pte_free_tlb()), flush the tlb if we need
0532      */
0533     tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
0534     pgd = pgd_offset(tlb->mm, addr);
0535     do {
0536         next = pgd_addr_end(addr, end);
0537         if (pgd_none_or_clear_bad(pgd))
0538             continue;
0539         free_pud_range(tlb, pgd, addr, next, floor, ceiling);
0540     } while (pgd++, addr = next, addr != end);
0541 }
0542 
0543 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
0544         unsigned long floor, unsigned long ceiling)
0545 {
0546     while (vma) {
0547         struct vm_area_struct *next = vma->vm_next;
0548         unsigned long addr = vma->vm_start;
0549 
0550         /*
0551          * Hide vma from rmap and truncate_pagecache before freeing
0552          * pgtables
0553          */
0554         unlink_anon_vmas(vma);
0555         unlink_file_vma(vma);
0556 
0557         if (is_vm_hugetlb_page(vma)) {
0558             hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
0559                 floor, next? next->vm_start: ceiling);
0560         } else {
0561             /*
0562              * Optimization: gather nearby vmas into one call down
0563              */
0564             while (next && next->vm_start <= vma->vm_end + PMD_SIZE
0565                    && !is_vm_hugetlb_page(next)) {
0566                 vma = next;
0567                 next = vma->vm_next;
0568                 unlink_anon_vmas(vma);
0569                 unlink_file_vma(vma);
0570             }
0571             free_pgd_range(tlb, addr, vma->vm_end,
0572                 floor, next? next->vm_start: ceiling);
0573         }
0574         vma = next;
0575     }
0576 }
0577 
0578 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
0579 {
0580     spinlock_t *ptl;
0581     pgtable_t new = pte_alloc_one(mm, address);
0582     if (!new)
0583         return -ENOMEM;
0584 
0585     /*
0586      * Ensure all pte setup (eg. pte page lock and page clearing) are
0587      * visible before the pte is made visible to other CPUs by being
0588      * put into page tables.
0589      *
0590      * The other side of the story is the pointer chasing in the page
0591      * table walking code (when walking the page table without locking;
0592      * ie. most of the time). Fortunately, these data accesses consist
0593      * of a chain of data-dependent loads, meaning most CPUs (alpha
0594      * being the notable exception) will already guarantee loads are
0595      * seen in-order. See the alpha page table accessors for the
0596      * smp_read_barrier_depends() barriers in page table walking code.
0597      */
0598     smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
0599 
0600     ptl = pmd_lock(mm, pmd);
0601     if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
0602         atomic_long_inc(&mm->nr_ptes);
0603         pmd_populate(mm, pmd, new);
0604         new = NULL;
0605     }
0606     spin_unlock(ptl);
0607     if (new)
0608         pte_free(mm, new);
0609     return 0;
0610 }
0611 
0612 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
0613 {
0614     pte_t *new = pte_alloc_one_kernel(&init_mm, address);
0615     if (!new)
0616         return -ENOMEM;
0617 
0618     smp_wmb(); /* See comment in __pte_alloc */
0619 
0620     spin_lock(&init_mm.page_table_lock);
0621     if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
0622         pmd_populate_kernel(&init_mm, pmd, new);
0623         new = NULL;
0624     }
0625     spin_unlock(&init_mm.page_table_lock);
0626     if (new)
0627         pte_free_kernel(&init_mm, new);
0628     return 0;
0629 }
0630 
0631 static inline void init_rss_vec(int *rss)
0632 {
0633     memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
0634 }
0635 
0636 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
0637 {
0638     int i;
0639 
0640     if (current->mm == mm)
0641         sync_mm_rss(mm);
0642     for (i = 0; i < NR_MM_COUNTERS; i++)
0643         if (rss[i])
0644             add_mm_counter(mm, i, rss[i]);
0645 }
0646 
0647 /*
0648  * This function is called to print an error when a bad pte
0649  * is found. For example, we might have a PFN-mapped pte in
0650  * a region that doesn't allow it.
0651  *
0652  * The calling function must still handle the error.
0653  */
0654 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
0655               pte_t pte, struct page *page)
0656 {
0657     pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
0658     pud_t *pud = pud_offset(pgd, addr);
0659     pmd_t *pmd = pmd_offset(pud, addr);
0660     struct address_space *mapping;
0661     pgoff_t index;
0662     static unsigned long resume;
0663     static unsigned long nr_shown;
0664     static unsigned long nr_unshown;
0665 
0666     /*
0667      * Allow a burst of 60 reports, then keep quiet for that minute;
0668      * or allow a steady drip of one report per second.
0669      */
0670     if (nr_shown == 60) {
0671         if (time_before(jiffies, resume)) {
0672             nr_unshown++;
0673             return;
0674         }
0675         if (nr_unshown) {
0676             pr_alert("BUG: Bad page map: %lu messages suppressed\n",
0677                  nr_unshown);
0678             nr_unshown = 0;
0679         }
0680         nr_shown = 0;
0681     }
0682     if (nr_shown++ == 0)
0683         resume = jiffies + 60 * HZ;
0684 
0685     mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
0686     index = linear_page_index(vma, addr);
0687 
0688     pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
0689          current->comm,
0690          (long long)pte_val(pte), (long long)pmd_val(*pmd));
0691     if (page)
0692         dump_page(page, "bad pte");
0693     pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
0694          (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
0695     /*
0696      * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
0697      */
0698     pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
0699          vma->vm_file,
0700          vma->vm_ops ? vma->vm_ops->fault : NULL,
0701          vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
0702          mapping ? mapping->a_ops->readpage : NULL);
0703     dump_stack();
0704     add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
0705 }
0706 
0707 /*
0708  * vm_normal_page -- This function gets the "struct page" associated with a pte.
0709  *
0710  * "Special" mappings do not wish to be associated with a "struct page" (either
0711  * it doesn't exist, or it exists but they don't want to touch it). In this
0712  * case, NULL is returned here. "Normal" mappings do have a struct page.
0713  *
0714  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
0715  * pte bit, in which case this function is trivial. Secondly, an architecture
0716  * may not have a spare pte bit, which requires a more complicated scheme,
0717  * described below.
0718  *
0719  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
0720  * special mapping (even if there are underlying and valid "struct pages").
0721  * COWed pages of a VM_PFNMAP are always normal.
0722  *
0723  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
0724  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
0725  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
0726  * mapping will always honor the rule
0727  *
0728  *  pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
0729  *
0730  * And for normal mappings this is false.
0731  *
0732  * This restricts such mappings to be a linear translation from virtual address
0733  * to pfn. To get around this restriction, we allow arbitrary mappings so long
0734  * as the vma is not a COW mapping; in that case, we know that all ptes are
0735  * special (because none can have been COWed).
0736  *
0737  *
0738  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
0739  *
0740  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
0741  * page" backing, however the difference is that _all_ pages with a struct
0742  * page (that is, those where pfn_valid is true) are refcounted and considered
0743  * normal pages by the VM. The disadvantage is that pages are refcounted
0744  * (which can be slower and simply not an option for some PFNMAP users). The
0745  * advantage is that we don't have to follow the strict linearity rule of
0746  * PFNMAP mappings in order to support COWable mappings.
0747  *
0748  */
0749 #ifdef __HAVE_ARCH_PTE_SPECIAL
0750 # define HAVE_PTE_SPECIAL 1
0751 #else
0752 # define HAVE_PTE_SPECIAL 0
0753 #endif
0754 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
0755                 pte_t pte)
0756 {
0757     unsigned long pfn = pte_pfn(pte);
0758 
0759     if (HAVE_PTE_SPECIAL) {
0760         if (likely(!pte_special(pte)))
0761             goto check_pfn;
0762         if (vma->vm_ops && vma->vm_ops->find_special_page)
0763             return vma->vm_ops->find_special_page(vma, addr);
0764         if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
0765             return NULL;
0766         if (!is_zero_pfn(pfn))
0767             print_bad_pte(vma, addr, pte, NULL);
0768         return NULL;
0769     }
0770 
0771     /* !HAVE_PTE_SPECIAL case follows: */
0772 
0773     if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
0774         if (vma->vm_flags & VM_MIXEDMAP) {
0775             if (!pfn_valid(pfn))
0776                 return NULL;
0777             goto out;
0778         } else {
0779             unsigned long off;
0780             off = (addr - vma->vm_start) >> PAGE_SHIFT;
0781             if (pfn == vma->vm_pgoff + off)
0782                 return NULL;
0783             if (!is_cow_mapping(vma->vm_flags))
0784                 return NULL;
0785         }
0786     }
0787 
0788     if (is_zero_pfn(pfn))
0789         return NULL;
0790 check_pfn:
0791     if (unlikely(pfn > highest_memmap_pfn)) {
0792         print_bad_pte(vma, addr, pte, NULL);
0793         return NULL;
0794     }
0795 
0796     /*
0797      * NOTE! We still have PageReserved() pages in the page tables.
0798      * eg. VDSO mappings can cause them to exist.
0799      */
0800 out:
0801     return pfn_to_page(pfn);
0802 }
0803 
0804 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
0805 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
0806                 pmd_t pmd)
0807 {
0808     unsigned long pfn = pmd_pfn(pmd);
0809 
0810     /*
0811      * There is no pmd_special() but there may be special pmds, e.g.
0812      * in a direct-access (dax) mapping, so let's just replicate the
0813      * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
0814      */
0815     if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
0816         if (vma->vm_flags & VM_MIXEDMAP) {
0817             if (!pfn_valid(pfn))
0818                 return NULL;
0819             goto out;
0820         } else {
0821             unsigned long off;
0822             off = (addr - vma->vm_start) >> PAGE_SHIFT;
0823             if (pfn == vma->vm_pgoff + off)
0824                 return NULL;
0825             if (!is_cow_mapping(vma->vm_flags))
0826                 return NULL;
0827         }
0828     }
0829 
0830     if (is_zero_pfn(pfn))
0831         return NULL;
0832     if (unlikely(pfn > highest_memmap_pfn))
0833         return NULL;
0834 
0835     /*
0836      * NOTE! We still have PageReserved() pages in the page tables.
0837      * eg. VDSO mappings can cause them to exist.
0838      */
0839 out:
0840     return pfn_to_page(pfn);
0841 }
0842 #endif
0843 
0844 /*
0845  * copy one vm_area from one task to the other. Assumes the page tables
0846  * already present in the new task to be cleared in the whole range
0847  * covered by this vma.
0848  */
0849 
0850 static inline unsigned long
0851 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
0852         pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
0853         unsigned long addr, int *rss)
0854 {
0855     unsigned long vm_flags = vma->vm_flags;
0856     pte_t pte = *src_pte;
0857     struct page *page;
0858 
0859     /* pte contains position in swap or file, so copy. */
0860     if (unlikely(!pte_present(pte))) {
0861         swp_entry_t entry = pte_to_swp_entry(pte);
0862 
0863         if (likely(!non_swap_entry(entry))) {
0864             if (swap_duplicate(entry) < 0)
0865                 return entry.val;
0866 
0867             /* make sure dst_mm is on swapoff's mmlist. */
0868             if (unlikely(list_empty(&dst_mm->mmlist))) {
0869                 spin_lock(&mmlist_lock);
0870                 if (list_empty(&dst_mm->mmlist))
0871                     list_add(&dst_mm->mmlist,
0872                             &src_mm->mmlist);
0873                 spin_unlock(&mmlist_lock);
0874             }
0875             rss[MM_SWAPENTS]++;
0876         } else if (is_migration_entry(entry)) {
0877             page = migration_entry_to_page(entry);
0878 
0879             rss[mm_counter(page)]++;
0880 
0881             if (is_write_migration_entry(entry) &&
0882                     is_cow_mapping(vm_flags)) {
0883                 /*
0884                  * COW mappings require pages in both
0885                  * parent and child to be set to read.
0886                  */
0887                 make_migration_entry_read(&entry);
0888                 pte = swp_entry_to_pte(entry);
0889                 if (pte_swp_soft_dirty(*src_pte))
0890                     pte = pte_swp_mksoft_dirty(pte);
0891                 set_pte_at(src_mm, addr, src_pte, pte);
0892             }
0893         }
0894         goto out_set_pte;
0895     }
0896 
0897     /*
0898      * If it's a COW mapping, write protect it both
0899      * in the parent and the child
0900      */
0901     if (is_cow_mapping(vm_flags)) {
0902         ptep_set_wrprotect(src_mm, addr, src_pte);
0903         pte = pte_wrprotect(pte);
0904     }
0905 
0906     /*
0907      * If it's a shared mapping, mark it clean in
0908      * the child
0909      */
0910     if (vm_flags & VM_SHARED)
0911         pte = pte_mkclean(pte);
0912     pte = pte_mkold(pte);
0913 
0914     page = vm_normal_page(vma, addr, pte);
0915     if (page) {
0916         get_page(page);
0917         page_dup_rmap(page, false);
0918         rss[mm_counter(page)]++;
0919     }
0920 
0921 out_set_pte:
0922     set_pte_at(dst_mm, addr, dst_pte, pte);
0923     return 0;
0924 }
0925 
0926 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
0927            pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
0928            unsigned long addr, unsigned long end)
0929 {
0930     pte_t *orig_src_pte, *orig_dst_pte;
0931     pte_t *src_pte, *dst_pte;
0932     spinlock_t *src_ptl, *dst_ptl;
0933     int progress = 0;
0934     int rss[NR_MM_COUNTERS];
0935     swp_entry_t entry = (swp_entry_t){0};
0936 
0937 again:
0938     init_rss_vec(rss);
0939 
0940     dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
0941     if (!dst_pte)
0942         return -ENOMEM;
0943     src_pte = pte_offset_map(src_pmd, addr);
0944     src_ptl = pte_lockptr(src_mm, src_pmd);
0945     spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
0946     orig_src_pte = src_pte;
0947     orig_dst_pte = dst_pte;
0948     arch_enter_lazy_mmu_mode();
0949 
0950     do {
0951         /*
0952          * We are holding two locks at this point - either of them
0953          * could generate latencies in another task on another CPU.
0954          */
0955         if (progress >= 32) {
0956             progress = 0;
0957             if (need_resched() ||
0958                 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
0959                 break;
0960         }
0961         if (pte_none(*src_pte)) {
0962             progress++;
0963             continue;
0964         }
0965         entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
0966                             vma, addr, rss);
0967         if (entry.val)
0968             break;
0969         progress += 8;
0970     } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
0971 
0972     arch_leave_lazy_mmu_mode();
0973     spin_unlock(src_ptl);
0974     pte_unmap(orig_src_pte);
0975     add_mm_rss_vec(dst_mm, rss);
0976     pte_unmap_unlock(orig_dst_pte, dst_ptl);
0977     cond_resched();
0978 
0979     if (entry.val) {
0980         if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
0981             return -ENOMEM;
0982         progress = 0;
0983     }
0984     if (addr != end)
0985         goto again;
0986     return 0;
0987 }
0988 
0989 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
0990         pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
0991         unsigned long addr, unsigned long end)
0992 {
0993     pmd_t *src_pmd, *dst_pmd;
0994     unsigned long next;
0995 
0996     dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
0997     if (!dst_pmd)
0998         return -ENOMEM;
0999     src_pmd = pmd_offset(src_pud, addr);
1000     do {
1001         next = pmd_addr_end(addr, end);
1002         if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1003             int err;
1004             VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1005             err = copy_huge_pmd(dst_mm, src_mm,
1006                         dst_pmd, src_pmd, addr, vma);
1007             if (err == -ENOMEM)
1008                 return -ENOMEM;
1009             if (!err)
1010                 continue;
1011             /* fall through */
1012         }
1013         if (pmd_none_or_clear_bad(src_pmd))
1014             continue;
1015         if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016                         vma, addr, next))
1017             return -ENOMEM;
1018     } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1019     return 0;
1020 }
1021 
1022 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023         pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1024         unsigned long addr, unsigned long end)
1025 {
1026     pud_t *src_pud, *dst_pud;
1027     unsigned long next;
1028 
1029     dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030     if (!dst_pud)
1031         return -ENOMEM;
1032     src_pud = pud_offset(src_pgd, addr);
1033     do {
1034         next = pud_addr_end(addr, end);
1035         if (pud_none_or_clear_bad(src_pud))
1036             continue;
1037         if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038                         vma, addr, next))
1039             return -ENOMEM;
1040     } while (dst_pud++, src_pud++, addr = next, addr != end);
1041     return 0;
1042 }
1043 
1044 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045         struct vm_area_struct *vma)
1046 {
1047     pgd_t *src_pgd, *dst_pgd;
1048     unsigned long next;
1049     unsigned long addr = vma->vm_start;
1050     unsigned long end = vma->vm_end;
1051     unsigned long mmun_start;   /* For mmu_notifiers */
1052     unsigned long mmun_end;     /* For mmu_notifiers */
1053     bool is_cow;
1054     int ret;
1055 
1056     /*
1057      * Don't copy ptes where a page fault will fill them correctly.
1058      * Fork becomes much lighter when there are big shared or private
1059      * readonly mappings. The tradeoff is that copy_page_range is more
1060      * efficient than faulting.
1061      */
1062     if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1063             !vma->anon_vma)
1064         return 0;
1065 
1066     if (is_vm_hugetlb_page(vma))
1067         return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1068 
1069     if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1070         /*
1071          * We do not free on error cases below as remove_vma
1072          * gets called on error from higher level routine
1073          */
1074         ret = track_pfn_copy(vma);
1075         if (ret)
1076             return ret;
1077     }
1078 
1079     /*
1080      * We need to invalidate the secondary MMU mappings only when
1081      * there could be a permission downgrade on the ptes of the
1082      * parent mm. And a permission downgrade will only happen if
1083      * is_cow_mapping() returns true.
1084      */
1085     is_cow = is_cow_mapping(vma->vm_flags);
1086     mmun_start = addr;
1087     mmun_end   = end;
1088     if (is_cow)
1089         mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1090                             mmun_end);
1091 
1092     ret = 0;
1093     dst_pgd = pgd_offset(dst_mm, addr);
1094     src_pgd = pgd_offset(src_mm, addr);
1095     do {
1096         next = pgd_addr_end(addr, end);
1097         if (pgd_none_or_clear_bad(src_pgd))
1098             continue;
1099         if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100                         vma, addr, next))) {
1101             ret = -ENOMEM;
1102             break;
1103         }
1104     } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1105 
1106     if (is_cow)
1107         mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1108     return ret;
1109 }
1110 
1111 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1112                 struct vm_area_struct *vma, pmd_t *pmd,
1113                 unsigned long addr, unsigned long end,
1114                 struct zap_details *details)
1115 {
1116     struct mm_struct *mm = tlb->mm;
1117     int force_flush = 0;
1118     int rss[NR_MM_COUNTERS];
1119     spinlock_t *ptl;
1120     pte_t *start_pte;
1121     pte_t *pte;
1122     swp_entry_t entry;
1123 
1124     tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1125 again:
1126     init_rss_vec(rss);
1127     start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1128     pte = start_pte;
1129     arch_enter_lazy_mmu_mode();
1130     do {
1131         pte_t ptent = *pte;
1132         if (pte_none(ptent)) {
1133             continue;
1134         }
1135 
1136         if (pte_present(ptent)) {
1137             struct page *page;
1138 
1139             page = vm_normal_page(vma, addr, ptent);
1140             if (unlikely(details) && page) {
1141                 /*
1142                  * unmap_shared_mapping_pages() wants to
1143                  * invalidate cache without truncating:
1144                  * unmap shared but keep private pages.
1145                  */
1146                 if (details->check_mapping &&
1147                     details->check_mapping != page_rmapping(page))
1148                     continue;
1149             }
1150             ptent = ptep_get_and_clear_full(mm, addr, pte,
1151                             tlb->fullmm);
1152             tlb_remove_tlb_entry(tlb, pte, addr);
1153             if (unlikely(!page))
1154                 continue;
1155 
1156             if (!PageAnon(page)) {
1157                 if (pte_dirty(ptent)) {
1158                     /*
1159                      * oom_reaper cannot tear down dirty
1160                      * pages
1161                      */
1162                     if (unlikely(details && details->ignore_dirty))
1163                         continue;
1164                     force_flush = 1;
1165                     set_page_dirty(page);
1166                 }
1167                 if (pte_young(ptent) &&
1168                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1169                     mark_page_accessed(page);
1170             }
1171             rss[mm_counter(page)]--;
1172             page_remove_rmap(page, false);
1173             if (unlikely(page_mapcount(page) < 0))
1174                 print_bad_pte(vma, addr, ptent, page);
1175             if (unlikely(__tlb_remove_page(tlb, page))) {
1176                 force_flush = 1;
1177                 addr += PAGE_SIZE;
1178                 break;
1179             }
1180             continue;
1181         }
1182         /* only check swap_entries if explicitly asked for in details */
1183         if (unlikely(details && !details->check_swap_entries))
1184             continue;
1185 
1186         entry = pte_to_swp_entry(ptent);
1187         if (!non_swap_entry(entry))
1188             rss[MM_SWAPENTS]--;
1189         else if (is_migration_entry(entry)) {
1190             struct page *page;
1191 
1192             page = migration_entry_to_page(entry);
1193             rss[mm_counter(page)]--;
1194         }
1195         if (unlikely(!free_swap_and_cache(entry)))
1196             print_bad_pte(vma, addr, ptent, NULL);
1197         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198     } while (pte++, addr += PAGE_SIZE, addr != end);
1199 
1200     add_mm_rss_vec(mm, rss);
1201     arch_leave_lazy_mmu_mode();
1202 
1203     /* Do the actual TLB flush before dropping ptl */
1204     if (force_flush)
1205         tlb_flush_mmu_tlbonly(tlb);
1206     pte_unmap_unlock(start_pte, ptl);
1207 
1208     /*
1209      * If we forced a TLB flush (either due to running out of
1210      * batch buffers or because we needed to flush dirty TLB
1211      * entries before releasing the ptl), free the batched
1212      * memory too. Restart if we didn't do everything.
1213      */
1214     if (force_flush) {
1215         force_flush = 0;
1216         tlb_flush_mmu_free(tlb);
1217         if (addr != end)
1218             goto again;
1219     }
1220 
1221     return addr;
1222 }
1223 
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225                 struct vm_area_struct *vma, pud_t *pud,
1226                 unsigned long addr, unsigned long end,
1227                 struct zap_details *details)
1228 {
1229     pmd_t *pmd;
1230     unsigned long next;
1231 
1232     pmd = pmd_offset(pud, addr);
1233     do {
1234         next = pmd_addr_end(addr, end);
1235         if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1236             if (next - addr != HPAGE_PMD_SIZE) {
1237                 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1238                     !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1239                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1240             } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1241                 goto next;
1242             /* fall through */
1243         }
1244         /*
1245          * Here there can be other concurrent MADV_DONTNEED or
1246          * trans huge page faults running, and if the pmd is
1247          * none or trans huge it can change under us. This is
1248          * because MADV_DONTNEED holds the mmap_sem in read
1249          * mode.
1250          */
1251         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1252             goto next;
1253         next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1254 next:
1255         cond_resched();
1256     } while (pmd++, addr = next, addr != end);
1257 
1258     return addr;
1259 }
1260 
1261 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1262                 struct vm_area_struct *vma, pgd_t *pgd,
1263                 unsigned long addr, unsigned long end,
1264                 struct zap_details *details)
1265 {
1266     pud_t *pud;
1267     unsigned long next;
1268 
1269     pud = pud_offset(pgd, addr);
1270     do {
1271         next = pud_addr_end(addr, end);
1272         if (pud_none_or_clear_bad(pud))
1273             continue;
1274         next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1275     } while (pud++, addr = next, addr != end);
1276 
1277     return addr;
1278 }
1279 
1280 void unmap_page_range(struct mmu_gather *tlb,
1281                  struct vm_area_struct *vma,
1282                  unsigned long addr, unsigned long end,
1283                  struct zap_details *details)
1284 {
1285     pgd_t *pgd;
1286     unsigned long next;
1287 
1288     BUG_ON(addr >= end);
1289     tlb_start_vma(tlb, vma);
1290     pgd = pgd_offset(vma->vm_mm, addr);
1291     do {
1292         next = pgd_addr_end(addr, end);
1293         if (pgd_none_or_clear_bad(pgd))
1294             continue;
1295         next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1296     } while (pgd++, addr = next, addr != end);
1297     tlb_end_vma(tlb, vma);
1298 }
1299 
1300 
1301 static void unmap_single_vma(struct mmu_gather *tlb,
1302         struct vm_area_struct *vma, unsigned long start_addr,
1303         unsigned long end_addr,
1304         struct zap_details *details)
1305 {
1306     unsigned long start = max(vma->vm_start, start_addr);
1307     unsigned long end;
1308 
1309     if (start >= vma->vm_end)
1310         return;
1311     end = min(vma->vm_end, end_addr);
1312     if (end <= vma->vm_start)
1313         return;
1314 
1315     if (vma->vm_file)
1316         uprobe_munmap(vma, start, end);
1317 
1318     if (unlikely(vma->vm_flags & VM_PFNMAP))
1319         untrack_pfn(vma, 0, 0);
1320 
1321     if (start != end) {
1322         if (unlikely(is_vm_hugetlb_page(vma))) {
1323             /*
1324              * It is undesirable to test vma->vm_file as it
1325              * should be non-null for valid hugetlb area.
1326              * However, vm_file will be NULL in the error
1327              * cleanup path of mmap_region. When
1328              * hugetlbfs ->mmap method fails,
1329              * mmap_region() nullifies vma->vm_file
1330              * before calling this function to clean up.
1331              * Since no pte has actually been setup, it is
1332              * safe to do nothing in this case.
1333              */
1334             if (vma->vm_file) {
1335                 i_mmap_lock_write(vma->vm_file->f_mapping);
1336                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1337                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1338             }
1339         } else
1340             unmap_page_range(tlb, vma, start, end, details);
1341     }
1342 }
1343 
1344 /**
1345  * unmap_vmas - unmap a range of memory covered by a list of vma's
1346  * @tlb: address of the caller's struct mmu_gather
1347  * @vma: the starting vma
1348  * @start_addr: virtual address at which to start unmapping
1349  * @end_addr: virtual address at which to end unmapping
1350  *
1351  * Unmap all pages in the vma list.
1352  *
1353  * Only addresses between `start' and `end' will be unmapped.
1354  *
1355  * The VMA list must be sorted in ascending virtual address order.
1356  *
1357  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358  * range after unmap_vmas() returns.  So the only responsibility here is to
1359  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360  * drops the lock and schedules.
1361  */
1362 void unmap_vmas(struct mmu_gather *tlb,
1363         struct vm_area_struct *vma, unsigned long start_addr,
1364         unsigned long end_addr)
1365 {
1366     struct mm_struct *mm = vma->vm_mm;
1367 
1368     mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1369     for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1370         unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1371     mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372 }
1373 
1374 /**
1375  * zap_page_range - remove user pages in a given range
1376  * @vma: vm_area_struct holding the applicable pages
1377  * @start: starting address of pages to zap
1378  * @size: number of bytes to zap
1379  * @details: details of shared cache invalidation
1380  *
1381  * Caller must protect the VMA list
1382  */
1383 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1384         unsigned long size, struct zap_details *details)
1385 {
1386     struct mm_struct *mm = vma->vm_mm;
1387     struct mmu_gather tlb;
1388     unsigned long end = start + size;
1389 
1390     lru_add_drain();
1391     tlb_gather_mmu(&tlb, mm, start, end);
1392     update_hiwater_rss(mm);
1393     mmu_notifier_invalidate_range_start(mm, start, end);
1394     for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1395         unmap_single_vma(&tlb, vma, start, end, details);
1396     mmu_notifier_invalidate_range_end(mm, start, end);
1397     tlb_finish_mmu(&tlb, start, end);
1398 }
1399 
1400 /**
1401  * zap_page_range_single - remove user pages in a given range
1402  * @vma: vm_area_struct holding the applicable pages
1403  * @address: starting address of pages to zap
1404  * @size: number of bytes to zap
1405  * @details: details of shared cache invalidation
1406  *
1407  * The range must fit into one VMA.
1408  */
1409 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410         unsigned long size, struct zap_details *details)
1411 {
1412     struct mm_struct *mm = vma->vm_mm;
1413     struct mmu_gather tlb;
1414     unsigned long end = address + size;
1415 
1416     lru_add_drain();
1417     tlb_gather_mmu(&tlb, mm, address, end);
1418     update_hiwater_rss(mm);
1419     mmu_notifier_invalidate_range_start(mm, address, end);
1420     unmap_single_vma(&tlb, vma, address, end, details);
1421     mmu_notifier_invalidate_range_end(mm, address, end);
1422     tlb_finish_mmu(&tlb, address, end);
1423 }
1424 
1425 /**
1426  * zap_vma_ptes - remove ptes mapping the vma
1427  * @vma: vm_area_struct holding ptes to be zapped
1428  * @address: starting address of pages to zap
1429  * @size: number of bytes to zap
1430  *
1431  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432  *
1433  * The entire address range must be fully contained within the vma.
1434  *
1435  * Returns 0 if successful.
1436  */
1437 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438         unsigned long size)
1439 {
1440     if (address < vma->vm_start || address + size > vma->vm_end ||
1441                 !(vma->vm_flags & VM_PFNMAP))
1442         return -1;
1443     zap_page_range_single(vma, address, size, NULL);
1444     return 0;
1445 }
1446 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447 
1448 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1449             spinlock_t **ptl)
1450 {
1451     pgd_t * pgd = pgd_offset(mm, addr);
1452     pud_t * pud = pud_alloc(mm, pgd, addr);
1453     if (pud) {
1454         pmd_t * pmd = pmd_alloc(mm, pud, addr);
1455         if (pmd) {
1456             VM_BUG_ON(pmd_trans_huge(*pmd));
1457             return pte_alloc_map_lock(mm, pmd, addr, ptl);
1458         }
1459     }
1460     return NULL;
1461 }
1462 
1463 /*
1464  * This is the old fallback for page remapping.
1465  *
1466  * For historical reasons, it only allows reserved pages. Only
1467  * old drivers should use this, and they needed to mark their
1468  * pages reserved for the old functions anyway.
1469  */
1470 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1471             struct page *page, pgprot_t prot)
1472 {
1473     struct mm_struct *mm = vma->vm_mm;
1474     int retval;
1475     pte_t *pte;
1476     spinlock_t *ptl;
1477 
1478     retval = -EINVAL;
1479     if (PageAnon(page))
1480         goto out;
1481     retval = -ENOMEM;
1482     flush_dcache_page(page);
1483     pte = get_locked_pte(mm, addr, &ptl);
1484     if (!pte)
1485         goto out;
1486     retval = -EBUSY;
1487     if (!pte_none(*pte))
1488         goto out_unlock;
1489 
1490     /* Ok, finally just insert the thing.. */
1491     get_page(page);
1492     inc_mm_counter_fast(mm, mm_counter_file(page));
1493     page_add_file_rmap(page, false);
1494     set_pte_at(mm, addr, pte, mk_pte(page, prot));
1495 
1496     retval = 0;
1497     pte_unmap_unlock(pte, ptl);
1498     return retval;
1499 out_unlock:
1500     pte_unmap_unlock(pte, ptl);
1501 out:
1502     return retval;
1503 }
1504 
1505 /**
1506  * vm_insert_page - insert single page into user vma
1507  * @vma: user vma to map to
1508  * @addr: target user address of this page
1509  * @page: source kernel page
1510  *
1511  * This allows drivers to insert individual pages they've allocated
1512  * into a user vma.
1513  *
1514  * The page has to be a nice clean _individual_ kernel allocation.
1515  * If you allocate a compound page, you need to have marked it as
1516  * such (__GFP_COMP), or manually just split the page up yourself
1517  * (see split_page()).
1518  *
1519  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1520  * took an arbitrary page protection parameter. This doesn't allow
1521  * that. Your vma protection will have to be set up correctly, which
1522  * means that if you want a shared writable mapping, you'd better
1523  * ask for a shared writable mapping!
1524  *
1525  * The page does not need to be reserved.
1526  *
1527  * Usually this function is called from f_op->mmap() handler
1528  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1529  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1530  * function from other places, for example from page-fault handler.
1531  */
1532 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1533             struct page *page)
1534 {
1535     if (addr < vma->vm_start || addr >= vma->vm_end)
1536         return -EFAULT;
1537     if (!page_count(page))
1538         return -EINVAL;
1539     if (!(vma->vm_flags & VM_MIXEDMAP)) {
1540         BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1541         BUG_ON(vma->vm_flags & VM_PFNMAP);
1542         vma->vm_flags |= VM_MIXEDMAP;
1543     }
1544     return insert_page(vma, addr, page, vma->vm_page_prot);
1545 }
1546 EXPORT_SYMBOL(vm_insert_page);
1547 
1548 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1549             pfn_t pfn, pgprot_t prot)
1550 {
1551     struct mm_struct *mm = vma->vm_mm;
1552     int retval;
1553     pte_t *pte, entry;
1554     spinlock_t *ptl;
1555 
1556     retval = -ENOMEM;
1557     pte = get_locked_pte(mm, addr, &ptl);
1558     if (!pte)
1559         goto out;
1560     retval = -EBUSY;
1561     if (!pte_none(*pte))
1562         goto out_unlock;
1563 
1564     /* Ok, finally just insert the thing.. */
1565     if (pfn_t_devmap(pfn))
1566         entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1567     else
1568         entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1569     set_pte_at(mm, addr, pte, entry);
1570     update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1571 
1572     retval = 0;
1573 out_unlock:
1574     pte_unmap_unlock(pte, ptl);
1575 out:
1576     return retval;
1577 }
1578 
1579 /**
1580  * vm_insert_pfn - insert single pfn into user vma
1581  * @vma: user vma to map to
1582  * @addr: target user address of this page
1583  * @pfn: source kernel pfn
1584  *
1585  * Similar to vm_insert_page, this allows drivers to insert individual pages
1586  * they've allocated into a user vma. Same comments apply.
1587  *
1588  * This function should only be called from a vm_ops->fault handler, and
1589  * in that case the handler should return NULL.
1590  *
1591  * vma cannot be a COW mapping.
1592  *
1593  * As this is called only for pages that do not currently exist, we
1594  * do not need to flush old virtual caches or the TLB.
1595  */
1596 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1597             unsigned long pfn)
1598 {
1599     return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1600 }
1601 EXPORT_SYMBOL(vm_insert_pfn);
1602 
1603 /**
1604  * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1605  * @vma: user vma to map to
1606  * @addr: target user address of this page
1607  * @pfn: source kernel pfn
1608  * @pgprot: pgprot flags for the inserted page
1609  *
1610  * This is exactly like vm_insert_pfn, except that it allows drivers to
1611  * to override pgprot on a per-page basis.
1612  *
1613  * This only makes sense for IO mappings, and it makes no sense for
1614  * cow mappings.  In general, using multiple vmas is preferable;
1615  * vm_insert_pfn_prot should only be used if using multiple VMAs is
1616  * impractical.
1617  */
1618 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1619             unsigned long pfn, pgprot_t pgprot)
1620 {
1621     int ret;
1622     /*
1623      * Technically, architectures with pte_special can avoid all these
1624      * restrictions (same for remap_pfn_range).  However we would like
1625      * consistency in testing and feature parity among all, so we should
1626      * try to keep these invariants in place for everybody.
1627      */
1628     BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1629     BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1630                         (VM_PFNMAP|VM_MIXEDMAP));
1631     BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1632     BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1633 
1634     if (addr < vma->vm_start || addr >= vma->vm_end)
1635         return -EFAULT;
1636 
1637     track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1638 
1639     ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1640 
1641     return ret;
1642 }
1643 EXPORT_SYMBOL(vm_insert_pfn_prot);
1644 
1645 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1646             pfn_t pfn)
1647 {
1648     pgprot_t pgprot = vma->vm_page_prot;
1649 
1650     BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1651 
1652     if (addr < vma->vm_start || addr >= vma->vm_end)
1653         return -EFAULT;
1654 
1655     track_pfn_insert(vma, &pgprot, pfn);
1656 
1657     /*
1658      * If we don't have pte special, then we have to use the pfn_valid()
1659      * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1660      * refcount the page if pfn_valid is true (hence insert_page rather
1661      * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1662      * without pte special, it would there be refcounted as a normal page.
1663      */
1664     if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1665         struct page *page;
1666 
1667         /*
1668          * At this point we are committed to insert_page()
1669          * regardless of whether the caller specified flags that
1670          * result in pfn_t_has_page() == false.
1671          */
1672         page = pfn_to_page(pfn_t_to_pfn(pfn));
1673         return insert_page(vma, addr, page, pgprot);
1674     }
1675     return insert_pfn(vma, addr, pfn, pgprot);
1676 }
1677 EXPORT_SYMBOL(vm_insert_mixed);
1678 
1679 /*
1680  * maps a range of physical memory into the requested pages. the old
1681  * mappings are removed. any references to nonexistent pages results
1682  * in null mappings (currently treated as "copy-on-access")
1683  */
1684 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1685             unsigned long addr, unsigned long end,
1686             unsigned long pfn, pgprot_t prot)
1687 {
1688     pte_t *pte;
1689     spinlock_t *ptl;
1690 
1691     pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1692     if (!pte)
1693         return -ENOMEM;
1694     arch_enter_lazy_mmu_mode();
1695     do {
1696         BUG_ON(!pte_none(*pte));
1697         set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1698         pfn++;
1699     } while (pte++, addr += PAGE_SIZE, addr != end);
1700     arch_leave_lazy_mmu_mode();
1701     pte_unmap_unlock(pte - 1, ptl);
1702     return 0;
1703 }
1704 
1705 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1706             unsigned long addr, unsigned long end,
1707             unsigned long pfn, pgprot_t prot)
1708 {
1709     pmd_t *pmd;
1710     unsigned long next;
1711 
1712     pfn -= addr >> PAGE_SHIFT;
1713     pmd = pmd_alloc(mm, pud, addr);
1714     if (!pmd)
1715         return -ENOMEM;
1716     VM_BUG_ON(pmd_trans_huge(*pmd));
1717     do {
1718         next = pmd_addr_end(addr, end);
1719         if (remap_pte_range(mm, pmd, addr, next,
1720                 pfn + (addr >> PAGE_SHIFT), prot))
1721             return -ENOMEM;
1722     } while (pmd++, addr = next, addr != end);
1723     return 0;
1724 }
1725 
1726 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1727             unsigned long addr, unsigned long end,
1728             unsigned long pfn, pgprot_t prot)
1729 {
1730     pud_t *pud;
1731     unsigned long next;
1732 
1733     pfn -= addr >> PAGE_SHIFT;
1734     pud = pud_alloc(mm, pgd, addr);
1735     if (!pud)
1736         return -ENOMEM;
1737     do {
1738         next = pud_addr_end(addr, end);
1739         if (remap_pmd_range(mm, pud, addr, next,
1740                 pfn + (addr >> PAGE_SHIFT), prot))
1741             return -ENOMEM;
1742     } while (pud++, addr = next, addr != end);
1743     return 0;
1744 }
1745 
1746 /**
1747  * remap_pfn_range - remap kernel memory to userspace
1748  * @vma: user vma to map to
1749  * @addr: target user address to start at
1750  * @pfn: physical address of kernel memory
1751  * @size: size of map area
1752  * @prot: page protection flags for this mapping
1753  *
1754  *  Note: this is only safe if the mm semaphore is held when called.
1755  */
1756 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1757             unsigned long pfn, unsigned long size, pgprot_t prot)
1758 {
1759     pgd_t *pgd;
1760     unsigned long next;
1761     unsigned long end = addr + PAGE_ALIGN(size);
1762     struct mm_struct *mm = vma->vm_mm;
1763     unsigned long remap_pfn = pfn;
1764     int err;
1765 
1766     /*
1767      * Physically remapped pages are special. Tell the
1768      * rest of the world about it:
1769      *   VM_IO tells people not to look at these pages
1770      *  (accesses can have side effects).
1771      *   VM_PFNMAP tells the core MM that the base pages are just
1772      *  raw PFN mappings, and do not have a "struct page" associated
1773      *  with them.
1774      *   VM_DONTEXPAND
1775      *      Disable vma merging and expanding with mremap().
1776      *   VM_DONTDUMP
1777      *      Omit vma from core dump, even when VM_IO turned off.
1778      *
1779      * There's a horrible special case to handle copy-on-write
1780      * behaviour that some programs depend on. We mark the "original"
1781      * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1782      * See vm_normal_page() for details.
1783      */
1784     if (is_cow_mapping(vma->vm_flags)) {
1785         if (addr != vma->vm_start || end != vma->vm_end)
1786             return -EINVAL;
1787         vma->vm_pgoff = pfn;
1788     }
1789 
1790     err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1791     if (err)
1792         return -EINVAL;
1793 
1794     vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1795 
1796     BUG_ON(addr >= end);
1797     pfn -= addr >> PAGE_SHIFT;
1798     pgd = pgd_offset(mm, addr);
1799     flush_cache_range(vma, addr, end);
1800     do {
1801         next = pgd_addr_end(addr, end);
1802         err = remap_pud_range(mm, pgd, addr, next,
1803                 pfn + (addr >> PAGE_SHIFT), prot);
1804         if (err)
1805             break;
1806     } while (pgd++, addr = next, addr != end);
1807 
1808     if (err)
1809         untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1810 
1811     return err;
1812 }
1813 EXPORT_SYMBOL(remap_pfn_range);
1814 
1815 /**
1816  * vm_iomap_memory - remap memory to userspace
1817  * @vma: user vma to map to
1818  * @start: start of area
1819  * @len: size of area
1820  *
1821  * This is a simplified io_remap_pfn_range() for common driver use. The
1822  * driver just needs to give us the physical memory range to be mapped,
1823  * we'll figure out the rest from the vma information.
1824  *
1825  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1826  * whatever write-combining details or similar.
1827  */
1828 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1829 {
1830     unsigned long vm_len, pfn, pages;
1831 
1832     /* Check that the physical memory area passed in looks valid */
1833     if (start + len < start)
1834         return -EINVAL;
1835     /*
1836      * You *really* shouldn't map things that aren't page-aligned,
1837      * but we've historically allowed it because IO memory might
1838      * just have smaller alignment.
1839      */
1840     len += start & ~PAGE_MASK;
1841     pfn = start >> PAGE_SHIFT;
1842     pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1843     if (pfn + pages < pfn)
1844         return -EINVAL;
1845 
1846     /* We start the mapping 'vm_pgoff' pages into the area */
1847     if (vma->vm_pgoff > pages)
1848         return -EINVAL;
1849     pfn += vma->vm_pgoff;
1850     pages -= vma->vm_pgoff;
1851 
1852     /* Can we fit all of the mapping? */
1853     vm_len = vma->vm_end - vma->vm_start;
1854     if (vm_len >> PAGE_SHIFT > pages)
1855         return -EINVAL;
1856 
1857     /* Ok, let it rip */
1858     return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1859 }
1860 EXPORT_SYMBOL(vm_iomap_memory);
1861 
1862 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1863                      unsigned long addr, unsigned long end,
1864                      pte_fn_t fn, void *data)
1865 {
1866     pte_t *pte;
1867     int err;
1868     pgtable_t token;
1869     spinlock_t *uninitialized_var(ptl);
1870 
1871     pte = (mm == &init_mm) ?
1872         pte_alloc_kernel(pmd, addr) :
1873         pte_alloc_map_lock(mm, pmd, addr, &ptl);
1874     if (!pte)
1875         return -ENOMEM;
1876 
1877     BUG_ON(pmd_huge(*pmd));
1878 
1879     arch_enter_lazy_mmu_mode();
1880 
1881     token = pmd_pgtable(*pmd);
1882 
1883     do {
1884         err = fn(pte++, token, addr, data);
1885         if (err)
1886             break;
1887     } while (addr += PAGE_SIZE, addr != end);
1888 
1889     arch_leave_lazy_mmu_mode();
1890 
1891     if (mm != &init_mm)
1892         pte_unmap_unlock(pte-1, ptl);
1893     return err;
1894 }
1895 
1896 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1897                      unsigned long addr, unsigned long end,
1898                      pte_fn_t fn, void *data)
1899 {
1900     pmd_t *pmd;
1901     unsigned long next;
1902     int err;
1903 
1904     BUG_ON(pud_huge(*pud));
1905 
1906     pmd = pmd_alloc(mm, pud, addr);
1907     if (!pmd)
1908         return -ENOMEM;
1909     do {
1910         next = pmd_addr_end(addr, end);
1911         err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1912         if (err)
1913             break;
1914     } while (pmd++, addr = next, addr != end);
1915     return err;
1916 }
1917 
1918 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1919                      unsigned long addr, unsigned long end,
1920                      pte_fn_t fn, void *data)
1921 {
1922     pud_t *pud;
1923     unsigned long next;
1924     int err;
1925 
1926     pud = pud_alloc(mm, pgd, addr);
1927     if (!pud)
1928         return -ENOMEM;
1929     do {
1930         next = pud_addr_end(addr, end);
1931         err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1932         if (err)
1933             break;
1934     } while (pud++, addr = next, addr != end);
1935     return err;
1936 }
1937 
1938 /*
1939  * Scan a region of virtual memory, filling in page tables as necessary
1940  * and calling a provided function on each leaf page table.
1941  */
1942 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1943             unsigned long size, pte_fn_t fn, void *data)
1944 {
1945     pgd_t *pgd;
1946     unsigned long next;
1947     unsigned long end = addr + size;
1948     int err;
1949 
1950     if (WARN_ON(addr >= end))
1951         return -EINVAL;
1952 
1953     pgd = pgd_offset(mm, addr);
1954     do {
1955         next = pgd_addr_end(addr, end);
1956         err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1957         if (err)
1958             break;
1959     } while (pgd++, addr = next, addr != end);
1960 
1961     return err;
1962 }
1963 EXPORT_SYMBOL_GPL(apply_to_page_range);
1964 
1965 /*
1966  * handle_pte_fault chooses page fault handler according to an entry which was
1967  * read non-atomically.  Before making any commitment, on those architectures
1968  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1969  * parts, do_swap_page must check under lock before unmapping the pte and
1970  * proceeding (but do_wp_page is only called after already making such a check;
1971  * and do_anonymous_page can safely check later on).
1972  */
1973 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1974                 pte_t *page_table, pte_t orig_pte)
1975 {
1976     int same = 1;
1977 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1978     if (sizeof(pte_t) > sizeof(unsigned long)) {
1979         spinlock_t *ptl = pte_lockptr(mm, pmd);
1980         spin_lock(ptl);
1981         same = pte_same(*page_table, orig_pte);
1982         spin_unlock(ptl);
1983     }
1984 #endif
1985     pte_unmap(page_table);
1986     return same;
1987 }
1988 
1989 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1990 {
1991     debug_dma_assert_idle(src);
1992 
1993     /*
1994      * If the source page was a PFN mapping, we don't have
1995      * a "struct page" for it. We do a best-effort copy by
1996      * just copying from the original user address. If that
1997      * fails, we just zero-fill it. Live with it.
1998      */
1999     if (unlikely(!src)) {
2000         void *kaddr = kmap_atomic(dst);
2001         void __user *uaddr = (void __user *)(va & PAGE_MASK);
2002 
2003         /*
2004          * This really shouldn't fail, because the page is there
2005          * in the page tables. But it might just be unreadable,
2006          * in which case we just give up and fill the result with
2007          * zeroes.
2008          */
2009         if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2010             clear_page(kaddr);
2011         kunmap_atomic(kaddr);
2012         flush_dcache_page(dst);
2013     } else
2014         copy_user_highpage(dst, src, va, vma);
2015 }
2016 
2017 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2018 {
2019     struct file *vm_file = vma->vm_file;
2020 
2021     if (vm_file)
2022         return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2023 
2024     /*
2025      * Special mappings (e.g. VDSO) do not have any file so fake
2026      * a default GFP_KERNEL for them.
2027      */
2028     return GFP_KERNEL;
2029 }
2030 
2031 /*
2032  * Notify the address space that the page is about to become writable so that
2033  * it can prohibit this or wait for the page to get into an appropriate state.
2034  *
2035  * We do this without the lock held, so that it can sleep if it needs to.
2036  */
2037 static int do_page_mkwrite(struct vm_fault *vmf)
2038 {
2039     int ret;
2040     struct page *page = vmf->page;
2041     unsigned int old_flags = vmf->flags;
2042 
2043     vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2044 
2045     ret = vmf->vma->vm_ops->page_mkwrite(vmf->vma, vmf);
2046     /* Restore original flags so that caller is not surprised */
2047     vmf->flags = old_flags;
2048     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2049         return ret;
2050     if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2051         lock_page(page);
2052         if (!page->mapping) {
2053             unlock_page(page);
2054             return 0; /* retry */
2055         }
2056         ret |= VM_FAULT_LOCKED;
2057     } else
2058         VM_BUG_ON_PAGE(!PageLocked(page), page);
2059     return ret;
2060 }
2061 
2062 /*
2063  * Handle dirtying of a page in shared file mapping on a write fault.
2064  *
2065  * The function expects the page to be locked and unlocks it.
2066  */
2067 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2068                     struct page *page)
2069 {
2070     struct address_space *mapping;
2071     bool dirtied;
2072     bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2073 
2074     dirtied = set_page_dirty(page);
2075     VM_BUG_ON_PAGE(PageAnon(page), page);
2076     /*
2077      * Take a local copy of the address_space - page.mapping may be zeroed
2078      * by truncate after unlock_page().   The address_space itself remains
2079      * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2080      * release semantics to prevent the compiler from undoing this copying.
2081      */
2082     mapping = page_rmapping(page);
2083     unlock_page(page);
2084 
2085     if ((dirtied || page_mkwrite) && mapping) {
2086         /*
2087          * Some device drivers do not set page.mapping
2088          * but still dirty their pages
2089          */
2090         balance_dirty_pages_ratelimited(mapping);
2091     }
2092 
2093     if (!page_mkwrite)
2094         file_update_time(vma->vm_file);
2095 }
2096 
2097 /*
2098  * Handle write page faults for pages that can be reused in the current vma
2099  *
2100  * This can happen either due to the mapping being with the VM_SHARED flag,
2101  * or due to us being the last reference standing to the page. In either
2102  * case, all we need to do here is to mark the page as writable and update
2103  * any related book-keeping.
2104  */
2105 static inline void wp_page_reuse(struct vm_fault *vmf)
2106     __releases(vmf->ptl)
2107 {
2108     struct vm_area_struct *vma = vmf->vma;
2109     struct page *page = vmf->page;
2110     pte_t entry;
2111     /*
2112      * Clear the pages cpupid information as the existing
2113      * information potentially belongs to a now completely
2114      * unrelated process.
2115      */
2116     if (page)
2117         page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2118 
2119     flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2120     entry = pte_mkyoung(vmf->orig_pte);
2121     entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2122     if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2123         update_mmu_cache(vma, vmf->address, vmf->pte);
2124     pte_unmap_unlock(vmf->pte, vmf->ptl);
2125 }
2126 
2127 /*
2128  * Handle the case of a page which we actually need to copy to a new page.
2129  *
2130  * Called with mmap_sem locked and the old page referenced, but
2131  * without the ptl held.
2132  *
2133  * High level logic flow:
2134  *
2135  * - Allocate a page, copy the content of the old page to the new one.
2136  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2137  * - Take the PTL. If the pte changed, bail out and release the allocated page
2138  * - If the pte is still the way we remember it, update the page table and all
2139  *   relevant references. This includes dropping the reference the page-table
2140  *   held to the old page, as well as updating the rmap.
2141  * - In any case, unlock the PTL and drop the reference we took to the old page.
2142  */
2143 static int wp_page_copy(struct vm_fault *vmf)
2144 {
2145     struct vm_area_struct *vma = vmf->vma;
2146     struct mm_struct *mm = vma->vm_mm;
2147     struct page *old_page = vmf->page;
2148     struct page *new_page = NULL;
2149     pte_t entry;
2150     int page_copied = 0;
2151     const unsigned long mmun_start = vmf->address & PAGE_MASK;
2152     const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2153     struct mem_cgroup *memcg;
2154 
2155     if (unlikely(anon_vma_prepare(vma)))
2156         goto oom;
2157 
2158     if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2159         new_page = alloc_zeroed_user_highpage_movable(vma,
2160                                   vmf->address);
2161         if (!new_page)
2162             goto oom;
2163     } else {
2164         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2165                 vmf->address);
2166         if (!new_page)
2167             goto oom;
2168         cow_user_page(new_page, old_page, vmf->address, vma);
2169     }
2170 
2171     if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2172         goto oom_free_new;
2173 
2174     __SetPageUptodate(new_page);
2175 
2176     mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2177 
2178     /*
2179      * Re-check the pte - we dropped the lock
2180      */
2181     vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2182     if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2183         if (old_page) {
2184             if (!PageAnon(old_page)) {
2185                 dec_mm_counter_fast(mm,
2186                         mm_counter_file(old_page));
2187                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2188             }
2189         } else {
2190             inc_mm_counter_fast(mm, MM_ANONPAGES);
2191         }
2192         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2193         entry = mk_pte(new_page, vma->vm_page_prot);
2194         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2195         /*
2196          * Clear the pte entry and flush it first, before updating the
2197          * pte with the new entry. This will avoid a race condition
2198          * seen in the presence of one thread doing SMC and another
2199          * thread doing COW.
2200          */
2201         ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2202         page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2203         mem_cgroup_commit_charge(new_page, memcg, false, false);
2204         lru_cache_add_active_or_unevictable(new_page, vma);
2205         /*
2206          * We call the notify macro here because, when using secondary
2207          * mmu page tables (such as kvm shadow page tables), we want the
2208          * new page to be mapped directly into the secondary page table.
2209          */
2210         set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2211         update_mmu_cache(vma, vmf->address, vmf->pte);
2212         if (old_page) {
2213             /*
2214              * Only after switching the pte to the new page may
2215              * we remove the mapcount here. Otherwise another
2216              * process may come and find the rmap count decremented
2217              * before the pte is switched to the new page, and
2218              * "reuse" the old page writing into it while our pte
2219              * here still points into it and can be read by other
2220              * threads.
2221              *
2222              * The critical issue is to order this
2223              * page_remove_rmap with the ptp_clear_flush above.
2224              * Those stores are ordered by (if nothing else,)
2225              * the barrier present in the atomic_add_negative
2226              * in page_remove_rmap.
2227              *
2228              * Then the TLB flush in ptep_clear_flush ensures that
2229              * no process can access the old page before the
2230              * decremented mapcount is visible. And the old page
2231              * cannot be reused until after the decremented
2232              * mapcount is visible. So transitively, TLBs to
2233              * old page will be flushed before it can be reused.
2234              */
2235             page_remove_rmap(old_page, false);
2236         }
2237 
2238         /* Free the old page.. */
2239         new_page = old_page;
2240         page_copied = 1;
2241     } else {
2242         mem_cgroup_cancel_charge(new_page, memcg, false);
2243     }
2244 
2245     if (new_page)
2246         put_page(new_page);
2247 
2248     pte_unmap_unlock(vmf->pte, vmf->ptl);
2249     mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2250     if (old_page) {
2251         /*
2252          * Don't let another task, with possibly unlocked vma,
2253          * keep the mlocked page.
2254          */
2255         if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2256             lock_page(old_page);    /* LRU manipulation */
2257             if (PageMlocked(old_page))
2258                 munlock_vma_page(old_page);
2259             unlock_page(old_page);
2260         }
2261         put_page(old_page);
2262     }
2263     return page_copied ? VM_FAULT_WRITE : 0;
2264 oom_free_new:
2265     put_page(new_page);
2266 oom:
2267     if (old_page)
2268         put_page(old_page);
2269     return VM_FAULT_OOM;
2270 }
2271 
2272 /**
2273  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2274  *            writeable once the page is prepared
2275  *
2276  * @vmf: structure describing the fault
2277  *
2278  * This function handles all that is needed to finish a write page fault in a
2279  * shared mapping due to PTE being read-only once the mapped page is prepared.
2280  * It handles locking of PTE and modifying it. The function returns
2281  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2282  * lock.
2283  *
2284  * The function expects the page to be locked or other protection against
2285  * concurrent faults / writeback (such as DAX radix tree locks).
2286  */
2287 int finish_mkwrite_fault(struct vm_fault *vmf)
2288 {
2289     WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2290     vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2291                        &vmf->ptl);
2292     /*
2293      * We might have raced with another page fault while we released the
2294      * pte_offset_map_lock.
2295      */
2296     if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2297         pte_unmap_unlock(vmf->pte, vmf->ptl);
2298         return VM_FAULT_NOPAGE;
2299     }
2300     wp_page_reuse(vmf);
2301     return 0;
2302 }
2303 
2304 /*
2305  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2306  * mapping
2307  */
2308 static int wp_pfn_shared(struct vm_fault *vmf)
2309 {
2310     struct vm_area_struct *vma = vmf->vma;
2311 
2312     if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2313         int ret;
2314 
2315         pte_unmap_unlock(vmf->pte, vmf->ptl);
2316         vmf->flags |= FAULT_FLAG_MKWRITE;
2317         ret = vma->vm_ops->pfn_mkwrite(vma, vmf);
2318         if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2319             return ret;
2320         return finish_mkwrite_fault(vmf);
2321     }
2322     wp_page_reuse(vmf);
2323     return VM_FAULT_WRITE;
2324 }
2325 
2326 static int wp_page_shared(struct vm_fault *vmf)
2327     __releases(vmf->ptl)
2328 {
2329     struct vm_area_struct *vma = vmf->vma;
2330 
2331     get_page(vmf->page);
2332 
2333     if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2334         int tmp;
2335 
2336         pte_unmap_unlock(vmf->pte, vmf->ptl);
2337         tmp = do_page_mkwrite(vmf);
2338         if (unlikely(!tmp || (tmp &
2339                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2340             put_page(vmf->page);
2341             return tmp;
2342         }
2343         tmp = finish_mkwrite_fault(vmf);
2344         if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2345             unlock_page(vmf->page);
2346             put_page(vmf->page);
2347             return tmp;
2348         }
2349     } else {
2350         wp_page_reuse(vmf);
2351         lock_page(vmf->page);
2352     }
2353     fault_dirty_shared_page(vma, vmf->page);
2354     put_page(vmf->page);
2355 
2356     return VM_FAULT_WRITE;
2357 }
2358 
2359 /*
2360  * This routine handles present pages, when users try to write
2361  * to a shared page. It is done by copying the page to a new address
2362  * and decrementing the shared-page counter for the old page.
2363  *
2364  * Note that this routine assumes that the protection checks have been
2365  * done by the caller (the low-level page fault routine in most cases).
2366  * Thus we can safely just mark it writable once we've done any necessary
2367  * COW.
2368  *
2369  * We also mark the page dirty at this point even though the page will
2370  * change only once the write actually happens. This avoids a few races,
2371  * and potentially makes it more efficient.
2372  *
2373  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2374  * but allow concurrent faults), with pte both mapped and locked.
2375  * We return with mmap_sem still held, but pte unmapped and unlocked.
2376  */
2377 static int do_wp_page(struct vm_fault *vmf)
2378     __releases(vmf->ptl)
2379 {
2380     struct vm_area_struct *vma = vmf->vma;
2381 
2382     vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2383     if (!vmf->page) {
2384         /*
2385          * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2386          * VM_PFNMAP VMA.
2387          *
2388          * We should not cow pages in a shared writeable mapping.
2389          * Just mark the pages writable and/or call ops->pfn_mkwrite.
2390          */
2391         if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2392                      (VM_WRITE|VM_SHARED))
2393             return wp_pfn_shared(vmf);
2394 
2395         pte_unmap_unlock(vmf->pte, vmf->ptl);
2396         return wp_page_copy(vmf);
2397     }
2398 
2399     /*
2400      * Take out anonymous pages first, anonymous shared vmas are
2401      * not dirty accountable.
2402      */
2403     if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2404         int total_mapcount;
2405         if (!trylock_page(vmf->page)) {
2406             get_page(vmf->page);
2407             pte_unmap_unlock(vmf->pte, vmf->ptl);
2408             lock_page(vmf->page);
2409             vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2410                     vmf->address, &vmf->ptl);
2411             if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2412                 unlock_page(vmf->page);
2413                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414                 put_page(vmf->page);
2415                 return 0;
2416             }
2417             put_page(vmf->page);
2418         }
2419         if (reuse_swap_page(vmf->page, &total_mapcount)) {
2420             if (total_mapcount == 1) {
2421                 /*
2422                  * The page is all ours. Move it to
2423                  * our anon_vma so the rmap code will
2424                  * not search our parent or siblings.
2425                  * Protected against the rmap code by
2426                  * the page lock.
2427                  */
2428                 page_move_anon_rmap(vmf->page, vma);
2429             }
2430             unlock_page(vmf->page);
2431             wp_page_reuse(vmf);
2432             return VM_FAULT_WRITE;
2433         }
2434         unlock_page(vmf->page);
2435     } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2436                     (VM_WRITE|VM_SHARED))) {
2437         return wp_page_shared(vmf);
2438     }
2439 
2440     /*
2441      * Ok, we need to copy. Oh, well..
2442      */
2443     get_page(vmf->page);
2444 
2445     pte_unmap_unlock(vmf->pte, vmf->ptl);
2446     return wp_page_copy(vmf);
2447 }
2448 
2449 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2450         unsigned long start_addr, unsigned long end_addr,
2451         struct zap_details *details)
2452 {
2453     zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2454 }
2455 
2456 static inline void unmap_mapping_range_tree(struct rb_root *root,
2457                         struct zap_details *details)
2458 {
2459     struct vm_area_struct *vma;
2460     pgoff_t vba, vea, zba, zea;
2461 
2462     vma_interval_tree_foreach(vma, root,
2463             details->first_index, details->last_index) {
2464 
2465         vba = vma->vm_pgoff;
2466         vea = vba + vma_pages(vma) - 1;
2467         zba = details->first_index;
2468         if (zba < vba)
2469             zba = vba;
2470         zea = details->last_index;
2471         if (zea > vea)
2472             zea = vea;
2473 
2474         unmap_mapping_range_vma(vma,
2475             ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2476             ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2477                 details);
2478     }
2479 }
2480 
2481 /**
2482  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2483  * address_space corresponding to the specified page range in the underlying
2484  * file.
2485  *
2486  * @mapping: the address space containing mmaps to be unmapped.
2487  * @holebegin: byte in first page to unmap, relative to the start of
2488  * the underlying file.  This will be rounded down to a PAGE_SIZE
2489  * boundary.  Note that this is different from truncate_pagecache(), which
2490  * must keep the partial page.  In contrast, we must get rid of
2491  * partial pages.
2492  * @holelen: size of prospective hole in bytes.  This will be rounded
2493  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2494  * end of the file.
2495  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2496  * but 0 when invalidating pagecache, don't throw away private data.
2497  */
2498 void unmap_mapping_range(struct address_space *mapping,
2499         loff_t const holebegin, loff_t const holelen, int even_cows)
2500 {
2501     struct zap_details details = { };
2502     pgoff_t hba = holebegin >> PAGE_SHIFT;
2503     pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2504 
2505     /* Check for overflow. */
2506     if (sizeof(holelen) > sizeof(hlen)) {
2507         long long holeend =
2508             (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2509         if (holeend & ~(long long)ULONG_MAX)
2510             hlen = ULONG_MAX - hba + 1;
2511     }
2512 
2513     details.check_mapping = even_cows? NULL: mapping;
2514     details.first_index = hba;
2515     details.last_index = hba + hlen - 1;
2516     if (details.last_index < details.first_index)
2517         details.last_index = ULONG_MAX;
2518 
2519     i_mmap_lock_write(mapping);
2520     if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2521         unmap_mapping_range_tree(&mapping->i_mmap, &details);
2522     i_mmap_unlock_write(mapping);
2523 }
2524 EXPORT_SYMBOL(unmap_mapping_range);
2525 
2526 /*
2527  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2528  * but allow concurrent faults), and pte mapped but not yet locked.
2529  * We return with pte unmapped and unlocked.
2530  *
2531  * We return with the mmap_sem locked or unlocked in the same cases
2532  * as does filemap_fault().
2533  */
2534 int do_swap_page(struct vm_fault *vmf)
2535 {
2536     struct vm_area_struct *vma = vmf->vma;
2537     struct page *page, *swapcache;
2538     struct mem_cgroup *memcg;
2539     swp_entry_t entry;
2540     pte_t pte;
2541     int locked;
2542     int exclusive = 0;
2543     int ret = 0;
2544 
2545     if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2546         goto out;
2547 
2548     entry = pte_to_swp_entry(vmf->orig_pte);
2549     if (unlikely(non_swap_entry(entry))) {
2550         if (is_migration_entry(entry)) {
2551             migration_entry_wait(vma->vm_mm, vmf->pmd,
2552                          vmf->address);
2553         } else if (is_hwpoison_entry(entry)) {
2554             ret = VM_FAULT_HWPOISON;
2555         } else {
2556             print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2557             ret = VM_FAULT_SIGBUS;
2558         }
2559         goto out;
2560     }
2561     delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2562     page = lookup_swap_cache(entry);
2563     if (!page) {
2564         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2565                     vmf->address);
2566         if (!page) {
2567             /*
2568              * Back out if somebody else faulted in this pte
2569              * while we released the pte lock.
2570              */
2571             vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2572                     vmf->address, &vmf->ptl);
2573             if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2574                 ret = VM_FAULT_OOM;
2575             delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2576             goto unlock;
2577         }
2578 
2579         /* Had to read the page from swap area: Major fault */
2580         ret = VM_FAULT_MAJOR;
2581         count_vm_event(PGMAJFAULT);
2582         mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2583     } else if (PageHWPoison(page)) {
2584         /*
2585          * hwpoisoned dirty swapcache pages are kept for killing
2586          * owner processes (which may be unknown at hwpoison time)
2587          */
2588         ret = VM_FAULT_HWPOISON;
2589         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2590         swapcache = page;
2591         goto out_release;
2592     }
2593 
2594     swapcache = page;
2595     locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2596 
2597     delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2598     if (!locked) {
2599         ret |= VM_FAULT_RETRY;
2600         goto out_release;
2601     }
2602 
2603     /*
2604      * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2605      * release the swapcache from under us.  The page pin, and pte_same
2606      * test below, are not enough to exclude that.  Even if it is still
2607      * swapcache, we need to check that the page's swap has not changed.
2608      */
2609     if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2610         goto out_page;
2611 
2612     page = ksm_might_need_to_copy(page, vma, vmf->address);
2613     if (unlikely(!page)) {
2614         ret = VM_FAULT_OOM;
2615         page = swapcache;
2616         goto out_page;
2617     }
2618 
2619     if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2620                 &memcg, false)) {
2621         ret = VM_FAULT_OOM;
2622         goto out_page;
2623     }
2624 
2625     /*
2626      * Back out if somebody else already faulted in this pte.
2627      */
2628     vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2629             &vmf->ptl);
2630     if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2631         goto out_nomap;
2632 
2633     if (unlikely(!PageUptodate(page))) {
2634         ret = VM_FAULT_SIGBUS;
2635         goto out_nomap;
2636     }
2637 
2638     /*
2639      * The page isn't present yet, go ahead with the fault.
2640      *
2641      * Be careful about the sequence of operations here.
2642      * To get its accounting right, reuse_swap_page() must be called
2643      * while the page is counted on swap but not yet in mapcount i.e.
2644      * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2645      * must be called after the swap_free(), or it will never succeed.
2646      */
2647 
2648     inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2649     dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2650     pte = mk_pte(page, vma->vm_page_prot);
2651     if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2652         pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2653         vmf->flags &= ~FAULT_FLAG_WRITE;
2654         ret |= VM_FAULT_WRITE;
2655         exclusive = RMAP_EXCLUSIVE;
2656     }
2657     flush_icache_page(vma, page);
2658     if (pte_swp_soft_dirty(vmf->orig_pte))
2659         pte = pte_mksoft_dirty(pte);
2660     set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2661     vmf->orig_pte = pte;
2662     if (page == swapcache) {
2663         do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2664         mem_cgroup_commit_charge(page, memcg, true, false);
2665         activate_page(page);
2666     } else { /* ksm created a completely new copy */
2667         page_add_new_anon_rmap(page, vma, vmf->address, false);
2668         mem_cgroup_commit_charge(page, memcg, false, false);
2669         lru_cache_add_active_or_unevictable(page, vma);
2670     }
2671 
2672     swap_free(entry);
2673     if (mem_cgroup_swap_full(page) ||
2674         (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2675         try_to_free_swap(page);
2676     unlock_page(page);
2677     if (page != swapcache) {
2678         /*
2679          * Hold the lock to avoid the swap entry to be reused
2680          * until we take the PT lock for the pte_same() check
2681          * (to avoid false positives from pte_same). For
2682          * further safety release the lock after the swap_free
2683          * so that the swap count won't change under a
2684          * parallel locked swapcache.
2685          */
2686         unlock_page(swapcache);
2687         put_page(swapcache);
2688     }
2689 
2690     if (vmf->flags & FAULT_FLAG_WRITE) {
2691         ret |= do_wp_page(vmf);
2692         if (ret & VM_FAULT_ERROR)
2693             ret &= VM_FAULT_ERROR;
2694         goto out;
2695     }
2696 
2697     /* No need to invalidate - it was non-present before */
2698     update_mmu_cache(vma, vmf->address, vmf->pte);
2699 unlock:
2700     pte_unmap_unlock(vmf->pte, vmf->ptl);
2701 out:
2702     return ret;
2703 out_nomap:
2704     mem_cgroup_cancel_charge(page, memcg, false);
2705     pte_unmap_unlock(vmf->pte, vmf->ptl);
2706 out_page:
2707     unlock_page(page);
2708 out_release:
2709     put_page(page);
2710     if (page != swapcache) {
2711         unlock_page(swapcache);
2712         put_page(swapcache);
2713     }
2714     return ret;
2715 }
2716 
2717 /*
2718  * This is like a special single-page "expand_{down|up}wards()",
2719  * except we must first make sure that 'address{-|+}PAGE_SIZE'
2720  * doesn't hit another vma.
2721  */
2722 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2723 {
2724     address &= PAGE_MASK;
2725     if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2726         struct vm_area_struct *prev = vma->vm_prev;
2727 
2728         /*
2729          * Is there a mapping abutting this one below?
2730          *
2731          * That's only ok if it's the same stack mapping
2732          * that has gotten split..
2733          */
2734         if (prev && prev->vm_end == address)
2735             return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2736 
2737         return expand_downwards(vma, address - PAGE_SIZE);
2738     }
2739     if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2740         struct vm_area_struct *next = vma->vm_next;
2741 
2742         /* As VM_GROWSDOWN but s/below/above/ */
2743         if (next && next->vm_start == address + PAGE_SIZE)
2744             return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2745 
2746         return expand_upwards(vma, address + PAGE_SIZE);
2747     }
2748     return 0;
2749 }
2750 
2751 /*
2752  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2753  * but allow concurrent faults), and pte mapped but not yet locked.
2754  * We return with mmap_sem still held, but pte unmapped and unlocked.
2755  */
2756 static int do_anonymous_page(struct vm_fault *vmf)
2757 {
2758     struct vm_area_struct *vma = vmf->vma;
2759     struct mem_cgroup *memcg;
2760     struct page *page;
2761     pte_t entry;
2762 
2763     /* File mapping without ->vm_ops ? */
2764     if (vma->vm_flags & VM_SHARED)
2765         return VM_FAULT_SIGBUS;
2766 
2767     /* Check if we need to add a guard page to the stack */
2768     if (check_stack_guard_page(vma, vmf->address) < 0)
2769         return VM_FAULT_SIGSEGV;
2770 
2771     /*
2772      * Use pte_alloc() instead of pte_alloc_map().  We can't run
2773      * pte_offset_map() on pmds where a huge pmd might be created
2774      * from a different thread.
2775      *
2776      * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2777      * parallel threads are excluded by other means.
2778      *
2779      * Here we only have down_read(mmap_sem).
2780      */
2781     if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2782         return VM_FAULT_OOM;
2783 
2784     /* See the comment in pte_alloc_one_map() */
2785     if (unlikely(pmd_trans_unstable(vmf->pmd)))
2786         return 0;
2787 
2788     /* Use the zero-page for reads */
2789     if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2790             !mm_forbids_zeropage(vma->vm_mm)) {
2791         entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2792                         vma->vm_page_prot));
2793         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2794                 vmf->address, &vmf->ptl);
2795         if (!pte_none(*vmf->pte))
2796             goto unlock;
2797         /* Deliver the page fault to userland, check inside PT lock */
2798         if (userfaultfd_missing(vma)) {
2799             pte_unmap_unlock(vmf->pte, vmf->ptl);
2800             return handle_userfault(vmf, VM_UFFD_MISSING);
2801         }
2802         goto setpte;
2803     }
2804 
2805     /* Allocate our own private page. */
2806     if (unlikely(anon_vma_prepare(vma)))
2807         goto oom;
2808     page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2809     if (!page)
2810         goto oom;
2811 
2812     if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2813         goto oom_free_page;
2814 
2815     /*
2816      * The memory barrier inside __SetPageUptodate makes sure that
2817      * preceeding stores to the page contents become visible before
2818      * the set_pte_at() write.
2819      */
2820     __SetPageUptodate(page);
2821 
2822     entry = mk_pte(page, vma->vm_page_prot);
2823     if (vma->vm_flags & VM_WRITE)
2824         entry = pte_mkwrite(pte_mkdirty(entry));
2825 
2826     vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2827             &vmf->ptl);
2828     if (!pte_none(*vmf->pte))
2829         goto release;
2830 
2831     /* Deliver the page fault to userland, check inside PT lock */
2832     if (userfaultfd_missing(vma)) {
2833         pte_unmap_unlock(vmf->pte, vmf->ptl);
2834         mem_cgroup_cancel_charge(page, memcg, false);
2835         put_page(page);
2836         return handle_userfault(vmf, VM_UFFD_MISSING);
2837     }
2838 
2839     inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2840     page_add_new_anon_rmap(page, vma, vmf->address, false);
2841     mem_cgroup_commit_charge(page, memcg, false, false);
2842     lru_cache_add_active_or_unevictable(page, vma);
2843 setpte:
2844     set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2845 
2846     /* No need to invalidate - it was non-present before */
2847     update_mmu_cache(vma, vmf->address, vmf->pte);
2848 unlock:
2849     pte_unmap_unlock(vmf->pte, vmf->ptl);
2850     return 0;
2851 release:
2852     mem_cgroup_cancel_charge(page, memcg, false);
2853     put_page(page);
2854     goto unlock;
2855 oom_free_page:
2856     put_page(page);
2857 oom:
2858     return VM_FAULT_OOM;
2859 }
2860 
2861 /*
2862  * The mmap_sem must have been held on entry, and may have been
2863  * released depending on flags and vma->vm_ops->fault() return value.
2864  * See filemap_fault() and __lock_page_retry().
2865  */
2866 static int __do_fault(struct vm_fault *vmf)
2867 {
2868     struct vm_area_struct *vma = vmf->vma;
2869     int ret;
2870 
2871     ret = vma->vm_ops->fault(vma, vmf);
2872     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2873                 VM_FAULT_DONE_COW)))
2874         return ret;
2875 
2876     if (unlikely(PageHWPoison(vmf->page))) {
2877         if (ret & VM_FAULT_LOCKED)
2878             unlock_page(vmf->page);
2879         put_page(vmf->page);
2880         vmf->page = NULL;
2881         return VM_FAULT_HWPOISON;
2882     }
2883 
2884     if (unlikely(!(ret & VM_FAULT_LOCKED)))
2885         lock_page(vmf->page);
2886     else
2887         VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2888 
2889     return ret;
2890 }
2891 
2892 static int pte_alloc_one_map(struct vm_fault *vmf)
2893 {
2894     struct vm_area_struct *vma = vmf->vma;
2895 
2896     if (!pmd_none(*vmf->pmd))
2897         goto map_pte;
2898     if (vmf->prealloc_pte) {
2899         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2900         if (unlikely(!pmd_none(*vmf->pmd))) {
2901             spin_unlock(vmf->ptl);
2902             goto map_pte;
2903         }
2904 
2905         atomic_long_inc(&vma->vm_mm->nr_ptes);
2906         pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2907         spin_unlock(vmf->ptl);
2908         vmf->prealloc_pte = 0;
2909     } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
2910         return VM_FAULT_OOM;
2911     }
2912 map_pte:
2913     /*
2914      * If a huge pmd materialized under us just retry later.  Use
2915      * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2916      * didn't become pmd_trans_huge under us and then back to pmd_none, as
2917      * a result of MADV_DONTNEED running immediately after a huge pmd fault
2918      * in a different thread of this mm, in turn leading to a misleading
2919      * pmd_trans_huge() retval.  All we have to ensure is that it is a
2920      * regular pmd that we can walk with pte_offset_map() and we can do that
2921      * through an atomic read in C, which is what pmd_trans_unstable()
2922      * provides.
2923      */
2924     if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
2925         return VM_FAULT_NOPAGE;
2926 
2927     vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2928             &vmf->ptl);
2929     return 0;
2930 }
2931 
2932 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2933 
2934 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2935 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2936         unsigned long haddr)
2937 {
2938     if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2939             (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2940         return false;
2941     if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2942         return false;
2943     return true;
2944 }
2945 
2946 static void deposit_prealloc_pte(struct vm_fault *vmf)
2947 {
2948     struct vm_area_struct *vma = vmf->vma;
2949 
2950     pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2951     /*
2952      * We are going to consume the prealloc table,
2953      * count that as nr_ptes.
2954      */
2955     atomic_long_inc(&vma->vm_mm->nr_ptes);
2956     vmf->prealloc_pte = 0;
2957 }
2958 
2959 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
2960 {
2961     struct vm_area_struct *vma = vmf->vma;
2962     bool write = vmf->flags & FAULT_FLAG_WRITE;
2963     unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2964     pmd_t entry;
2965     int i, ret;
2966 
2967     if (!transhuge_vma_suitable(vma, haddr))
2968         return VM_FAULT_FALLBACK;
2969 
2970     ret = VM_FAULT_FALLBACK;
2971     page = compound_head(page);
2972 
2973     /*
2974      * Archs like ppc64 need additonal space to store information
2975      * related to pte entry. Use the preallocated table for that.
2976      */
2977     if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
2978         vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
2979         if (!vmf->prealloc_pte)
2980             return VM_FAULT_OOM;
2981         smp_wmb(); /* See comment in __pte_alloc() */
2982     }
2983 
2984     vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2985     if (unlikely(!pmd_none(*vmf->pmd)))
2986         goto out;
2987 
2988     for (i = 0; i < HPAGE_PMD_NR; i++)
2989         flush_icache_page(vma, page + i);
2990 
2991     entry = mk_huge_pmd(page, vma->vm_page_prot);
2992     if (write)
2993         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2994 
2995     add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2996     page_add_file_rmap(page, true);
2997     /*
2998      * deposit and withdraw with pmd lock held
2999      */
3000     if (arch_needs_pgtable_deposit())
3001         deposit_prealloc_pte(vmf);
3002 
3003     set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3004 
3005     update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3006 
3007     /* fault is handled */
3008     ret = 0;
3009     count_vm_event(THP_FILE_MAPPED);
3010 out:
3011     spin_unlock(vmf->ptl);
3012     return ret;
3013 }
3014 #else
3015 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3016 {
3017     BUILD_BUG();
3018     return 0;
3019 }
3020 #endif
3021 
3022 /**
3023  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3024  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3025  *
3026  * @vmf: fault environment
3027  * @memcg: memcg to charge page (only for private mappings)
3028  * @page: page to map
3029  *
3030  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3031  * return.
3032  *
3033  * Target users are page handler itself and implementations of
3034  * vm_ops->map_pages.
3035  */
3036 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3037         struct page *page)
3038 {
3039     struct vm_area_struct *vma = vmf->vma;
3040     bool write = vmf->flags & FAULT_FLAG_WRITE;
3041     pte_t entry;
3042     int ret;
3043 
3044     if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3045             IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3046         /* THP on COW? */
3047         VM_BUG_ON_PAGE(memcg, page);
3048 
3049         ret = do_set_pmd(vmf, page);
3050         if (ret != VM_FAULT_FALLBACK)
3051             return ret;
3052     }
3053 
3054     if (!vmf->pte) {
3055         ret = pte_alloc_one_map(vmf);
3056         if (ret)
3057             return ret;
3058     }
3059 
3060     /* Re-check under ptl */
3061     if (unlikely(!pte_none(*vmf->pte)))
3062         return VM_FAULT_NOPAGE;
3063 
3064     flush_icache_page(vma, page);
3065     entry = mk_pte(page, vma->vm_page_prot);
3066     if (write)
3067         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3068     /* copy-on-write page */
3069     if (write && !(vma->vm_flags & VM_SHARED)) {
3070         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3071         page_add_new_anon_rmap(page, vma, vmf->address, false);
3072         mem_cgroup_commit_charge(page, memcg, false, false);
3073         lru_cache_add_active_or_unevictable(page, vma);
3074     } else {
3075         inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3076         page_add_file_rmap(page, false);
3077     }
3078     set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3079 
3080     /* no need to invalidate: a not-present page won't be cached */
3081     update_mmu_cache(vma, vmf->address, vmf->pte);
3082 
3083     return 0;
3084 }
3085 
3086 
3087 /**
3088  * finish_fault - finish page fault once we have prepared the page to fault
3089  *
3090  * @vmf: structure describing the fault
3091  *
3092  * This function handles all that is needed to finish a page fault once the
3093  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3094  * given page, adds reverse page mapping, handles memcg charges and LRU
3095  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3096  * error.
3097  *
3098  * The function expects the page to be locked and on success it consumes a
3099  * reference of a page being mapped (for the PTE which maps it).
3100  */
3101 int finish_fault(struct vm_fault *vmf)
3102 {
3103     struct page *page;
3104     int ret;
3105 
3106     /* Did we COW the page? */
3107     if ((vmf->flags & FAULT_FLAG_WRITE) &&
3108         !(vmf->vma->vm_flags & VM_SHARED))
3109         page = vmf->cow_page;
3110     else
3111         page = vmf->page;
3112     ret = alloc_set_pte(vmf, vmf->memcg, page);
3113     if (vmf->pte)
3114         pte_unmap_unlock(vmf->pte, vmf->ptl);
3115     return ret;
3116 }
3117 
3118 static unsigned long fault_around_bytes __read_mostly =
3119     rounddown_pow_of_two(65536);
3120 
3121 #ifdef CONFIG_DEBUG_FS
3122 static int fault_around_bytes_get(void *data, u64 *val)
3123 {
3124     *val = fault_around_bytes;
3125     return 0;
3126 }
3127 
3128 /*
3129  * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3130  * rounded down to nearest page order. It's what do_fault_around() expects to
3131  * see.
3132  */
3133 static int fault_around_bytes_set(void *data, u64 val)
3134 {
3135     if (val / PAGE_SIZE > PTRS_PER_PTE)
3136         return -EINVAL;
3137     if (val > PAGE_SIZE)
3138         fault_around_bytes = rounddown_pow_of_two(val);
3139     else
3140         fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3141     return 0;
3142 }
3143 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3144         fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3145 
3146 static int __init fault_around_debugfs(void)
3147 {
3148     void *ret;
3149 
3150     ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3151             &fault_around_bytes_fops);
3152     if (!ret)
3153         pr_warn("Failed to create fault_around_bytes in debugfs");
3154     return 0;
3155 }
3156 late_initcall(fault_around_debugfs);
3157 #endif
3158 
3159 /*
3160  * do_fault_around() tries to map few pages around the fault address. The hope
3161  * is that the pages will be needed soon and this will lower the number of
3162  * faults to handle.
3163  *
3164  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3165  * not ready to be mapped: not up-to-date, locked, etc.
3166  *
3167  * This function is called with the page table lock taken. In the split ptlock
3168  * case the page table lock only protects only those entries which belong to
3169  * the page table corresponding to the fault address.
3170  *
3171  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3172  * only once.
3173  *
3174  * fault_around_pages() defines how many pages we'll try to map.
3175  * do_fault_around() expects it to return a power of two less than or equal to
3176  * PTRS_PER_PTE.
3177  *
3178  * The virtual address of the area that we map is naturally aligned to the
3179  * fault_around_pages() value (and therefore to page order).  This way it's
3180  * easier to guarantee that we don't cross page table boundaries.
3181  */
3182 static int do_fault_around(struct vm_fault *vmf)
3183 {
3184     unsigned long address = vmf->address, nr_pages, mask;
3185     pgoff_t start_pgoff = vmf->pgoff;
3186     pgoff_t end_pgoff;
3187     int off, ret = 0;
3188 
3189     nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3190     mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3191 
3192     vmf->address = max(address & mask, vmf->vma->vm_start);
3193     off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3194     start_pgoff -= off;
3195 
3196     /*
3197      *  end_pgoff is either end of page table or end of vma
3198      *  or fault_around_pages() from start_pgoff, depending what is nearest.
3199      */
3200     end_pgoff = start_pgoff -
3201         ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3202         PTRS_PER_PTE - 1;
3203     end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3204             start_pgoff + nr_pages - 1);
3205 
3206     if (pmd_none(*vmf->pmd)) {
3207         vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3208                           vmf->address);
3209         if (!vmf->prealloc_pte)
3210             goto out;
3211         smp_wmb(); /* See comment in __pte_alloc() */
3212     }
3213 
3214     vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3215 
3216     /* Huge page is mapped? Page fault is solved */
3217     if (pmd_trans_huge(*vmf->pmd)) {
3218         ret = VM_FAULT_NOPAGE;
3219         goto out;
3220     }
3221 
3222     /* ->map_pages() haven't done anything useful. Cold page cache? */
3223     if (!vmf->pte)
3224         goto out;
3225 
3226     /* check if the page fault is solved */
3227     vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3228     if (!pte_none(*vmf->pte))
3229         ret = VM_FAULT_NOPAGE;
3230     pte_unmap_unlock(vmf->pte, vmf->ptl);
3231 out:
3232     vmf->address = address;
3233     vmf->pte = NULL;
3234     return ret;
3235 }
3236 
3237 static int do_read_fault(struct vm_fault *vmf)
3238 {
3239     struct vm_area_struct *vma = vmf->vma;
3240     int ret = 0;
3241 
3242     /*
3243      * Let's call ->map_pages() first and use ->fault() as fallback
3244      * if page by the offset is not ready to be mapped (cold cache or
3245      * something).
3246      */
3247     if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3248         ret = do_fault_around(vmf);
3249         if (ret)
3250             return ret;
3251     }
3252 
3253     ret = __do_fault(vmf);
3254     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3255         return ret;
3256 
3257     ret |= finish_fault(vmf);
3258     unlock_page(vmf->page);
3259     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3260         put_page(vmf->page);
3261     return ret;
3262 }
3263 
3264 static int do_cow_fault(struct vm_fault *vmf)
3265 {
3266     struct vm_area_struct *vma = vmf->vma;
3267     int ret;
3268 
3269     if (unlikely(anon_vma_prepare(vma)))
3270         return VM_FAULT_OOM;
3271 
3272     vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3273     if (!vmf->cow_page)
3274         return VM_FAULT_OOM;
3275 
3276     if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3277                 &vmf->memcg, false)) {
3278         put_page(vmf->cow_page);
3279         return VM_FAULT_OOM;
3280     }
3281 
3282     ret = __do_fault(vmf);
3283     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3284         goto uncharge_out;
3285     if (ret & VM_FAULT_DONE_COW)
3286         return ret;
3287 
3288     copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3289     __SetPageUptodate(vmf->cow_page);
3290 
3291     ret |= finish_fault(vmf);
3292     unlock_page(vmf->page);
3293     put_page(vmf->page);
3294     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3295         goto uncharge_out;
3296     return ret;
3297 uncharge_out:
3298     mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3299     put_page(vmf->cow_page);
3300     return ret;
3301 }
3302 
3303 static int do_shared_fault(struct vm_fault *vmf)
3304 {
3305     struct vm_area_struct *vma = vmf->vma;
3306     int ret, tmp;
3307 
3308     ret = __do_fault(vmf);
3309     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3310         return ret;
3311 
3312     /*
3313      * Check if the backing address space wants to know that the page is
3314      * about to become writable
3315      */
3316     if (vma->vm_ops->page_mkwrite) {
3317         unlock_page(vmf->page);
3318         tmp = do_page_mkwrite(vmf);
3319         if (unlikely(!tmp ||
3320                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3321             put_page(vmf->page);
3322             return tmp;
3323         }
3324     }
3325 
3326     ret |= finish_fault(vmf);
3327     if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3328                     VM_FAULT_RETRY))) {
3329         unlock_page(vmf->page);
3330         put_page(vmf->page);
3331         return ret;
3332     }
3333 
3334     fault_dirty_shared_page(vma, vmf->page);
3335     return ret;
3336 }
3337 
3338 /*
3339  * We enter with non-exclusive mmap_sem (to exclude vma changes,
3340  * but allow concurrent faults).
3341  * The mmap_sem may have been released depending on flags and our
3342  * return value.  See filemap_fault() and __lock_page_or_retry().
3343  */
3344 static int do_fault(struct vm_fault *vmf)
3345 {
3346     struct vm_area_struct *vma = vmf->vma;
3347     int ret;
3348 
3349     /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3350     if (!vma->vm_ops->fault)
3351         ret = VM_FAULT_SIGBUS;
3352     else if (!(vmf->flags & FAULT_FLAG_WRITE))
3353         ret = do_read_fault(vmf);
3354     else if (!(vma->vm_flags & VM_SHARED))
3355         ret = do_cow_fault(vmf);
3356     else
3357         ret = do_shared_fault(vmf);
3358 
3359     /* preallocated pagetable is unused: free it */
3360     if (vmf->prealloc_pte) {
3361         pte_free(vma->vm_mm, vmf->prealloc_pte);
3362         vmf->prealloc_pte = 0;
3363     }
3364     return ret;
3365 }
3366 
3367 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3368                 unsigned long addr, int page_nid,
3369                 int *flags)
3370 {
3371     get_page(page);
3372 
3373     count_vm_numa_event(NUMA_HINT_FAULTS);
3374     if (page_nid == numa_node_id()) {
3375         count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3376         *flags |= TNF_FAULT_LOCAL;
3377     }
3378 
3379     return mpol_misplaced(page, vma, addr);
3380 }
3381 
3382 static int do_numa_page(struct vm_fault *vmf)
3383 {
3384     struct vm_area_struct *vma = vmf->vma;
3385     struct page *page = NULL;
3386     int page_nid = -1;
3387     int last_cpupid;
3388     int target_nid;
3389     bool migrated = false;
3390     pte_t pte = vmf->orig_pte;
3391     bool was_writable = pte_write(pte);
3392     int flags = 0;
3393 
3394     /*
3395     * The "pte" at this point cannot be used safely without
3396     * validation through pte_unmap_same(). It's of NUMA type but
3397     * the pfn may be screwed if the read is non atomic.
3398     *
3399     * We can safely just do a "set_pte_at()", because the old
3400     * page table entry is not accessible, so there would be no
3401     * concurrent hardware modifications to the PTE.
3402     */
3403     vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3404     spin_lock(vmf->ptl);
3405     if (unlikely(!pte_same(*vmf->pte, pte))) {
3406         pte_unmap_unlock(vmf->pte, vmf->ptl);
3407         goto out;
3408     }
3409 
3410     /* Make it present again */
3411     pte = pte_modify(pte, vma->vm_page_prot);
3412     pte = pte_mkyoung(pte);
3413     if (was_writable)
3414         pte = pte_mkwrite(pte);
3415     set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3416     update_mmu_cache(vma, vmf->address, vmf->pte);
3417 
3418     page = vm_normal_page(vma, vmf->address, pte);
3419     if (!page) {
3420         pte_unmap_unlock(vmf->pte, vmf->ptl);
3421         return 0;
3422     }
3423 
3424     /* TODO: handle PTE-mapped THP */
3425     if (PageCompound(page)) {
3426         pte_unmap_unlock(vmf->pte, vmf->ptl);
3427         return 0;
3428     }
3429 
3430     /*
3431      * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3432      * much anyway since they can be in shared cache state. This misses
3433      * the case where a mapping is writable but the process never writes
3434      * to it but pte_write gets cleared during protection updates and
3435      * pte_dirty has unpredictable behaviour between PTE scan updates,
3436      * background writeback, dirty balancing and application behaviour.
3437      */
3438     if (!pte_write(pte))
3439         flags |= TNF_NO_GROUP;
3440 
3441     /*
3442      * Flag if the page is shared between multiple address spaces. This
3443      * is later used when determining whether to group tasks together
3444      */
3445     if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3446         flags |= TNF_SHARED;
3447 
3448     last_cpupid = page_cpupid_last(page);
3449     page_nid = page_to_nid(page);
3450     target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3451             &flags);
3452     pte_unmap_unlock(vmf->pte, vmf->ptl);
3453     if (target_nid == -1) {
3454         put_page(page);
3455         goto out;
3456     }
3457 
3458     /* Migrate to the requested node */
3459     migrated = migrate_misplaced_page(page, vma, target_nid);
3460     if (migrated) {
3461         page_nid = target_nid;
3462         flags |= TNF_MIGRATED;
3463     } else
3464         flags |= TNF_MIGRATE_FAIL;
3465 
3466 out:
3467     if (page_nid != -1)
3468         task_numa_fault(last_cpupid, page_nid, 1, flags);
3469     return 0;
3470 }
3471 
3472 static int create_huge_pmd(struct vm_fault *vmf)
3473 {
3474     struct vm_area_struct *vma = vmf->vma;
3475     if (vma_is_anonymous(vma))
3476         return do_huge_pmd_anonymous_page(vmf);
3477     if (vma->vm_ops->pmd_fault)
3478         return vma->vm_ops->pmd_fault(vma, vmf->address, vmf->pmd,
3479                 vmf->flags);
3480     return VM_FAULT_FALLBACK;
3481 }
3482 
3483 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3484 {
3485     if (vma_is_anonymous(vmf->vma))
3486         return do_huge_pmd_wp_page(vmf, orig_pmd);
3487     if (vmf->vma->vm_ops->pmd_fault)
3488         return vmf->vma->vm_ops->pmd_fault(vmf->vma, vmf->address,
3489                            vmf->pmd, vmf->flags);
3490 
3491     /* COW handled on pte level: split pmd */
3492     VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3493     __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3494 
3495     return VM_FAULT_FALLBACK;
3496 }
3497 
3498 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3499 {
3500     return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3501 }
3502 
3503 /*
3504  * These routines also need to handle stuff like marking pages dirty
3505  * and/or accessed for architectures that don't do it in hardware (most
3506  * RISC architectures).  The early dirtying is also good on the i386.
3507  *
3508  * There is also a hook called "update_mmu_cache()" that architectures
3509  * with external mmu caches can use to update those (ie the Sparc or
3510  * PowerPC hashed page tables that act as extended TLBs).
3511  *
3512  * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3513  * concurrent faults).
3514  *
3515  * The mmap_sem may have been released depending on flags and our return value.
3516  * See filemap_fault() and __lock_page_or_retry().
3517  */
3518 static int handle_pte_fault(struct vm_fault *vmf)
3519 {
3520     pte_t entry;
3521 
3522     if (unlikely(pmd_none(*vmf->pmd))) {
3523         /*
3524          * Leave __pte_alloc() until later: because vm_ops->fault may
3525          * want to allocate huge page, and if we expose page table
3526          * for an instant, it will be difficult to retract from
3527          * concurrent faults and from rmap lookups.
3528          */
3529         vmf->pte = NULL;
3530     } else {
3531         /* See comment in pte_alloc_one_map() */
3532         if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3533             return 0;
3534         /*
3535          * A regular pmd is established and it can't morph into a huge
3536          * pmd from under us anymore at this point because we hold the
3537          * mmap_sem read mode and khugepaged takes it in write mode.
3538          * So now it's safe to run pte_offset_map().
3539          */
3540         vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3541         vmf->orig_pte = *vmf->pte;
3542 
3543         /*
3544          * some architectures can have larger ptes than wordsize,
3545          * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3546          * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3547          * atomic accesses.  The code below just needs a consistent
3548          * view for the ifs and we later double check anyway with the
3549          * ptl lock held. So here a barrier will do.
3550          */
3551         barrier();
3552         if (pte_none(vmf->orig_pte)) {
3553             pte_unmap(vmf->pte);
3554             vmf->pte = NULL;
3555         }
3556     }
3557 
3558     if (!vmf->pte) {
3559         if (vma_is_anonymous(vmf->vma))
3560             return do_anonymous_page(vmf);
3561         else
3562             return do_fault(vmf);
3563     }
3564 
3565     if (!pte_present(vmf->orig_pte))
3566         return do_swap_page(vmf);
3567 
3568     if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3569         return do_numa_page(vmf);
3570 
3571     vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3572     spin_lock(vmf->ptl);
3573     entry = vmf->orig_pte;
3574     if (unlikely(!pte_same(*vmf->pte, entry)))
3575         goto unlock;
3576     if (vmf->flags & FAULT_FLAG_WRITE) {
3577         if (!pte_write(entry))
3578             return do_wp_page(vmf);
3579         entry = pte_mkdirty(entry);
3580     }
3581     entry = pte_mkyoung(entry);
3582     if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3583                 vmf->flags & FAULT_FLAG_WRITE)) {
3584         update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3585     } else {
3586         /*
3587          * This is needed only for protection faults but the arch code
3588          * is not yet telling us if this is a protection fault or not.
3589          * This still avoids useless tlb flushes for .text page faults
3590          * with threads.
3591          */
3592         if (vmf->flags & FAULT_FLAG_WRITE)
3593             flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3594     }
3595 unlock:
3596     pte_unmap_unlock(vmf->pte, vmf->ptl);
3597     return 0;
3598 }
3599 
3600 /*
3601  * By the time we get here, we already hold the mm semaphore
3602  *
3603  * The mmap_sem may have been released depending on flags and our
3604  * return value.  See filemap_fault() and __lock_page_or_retry().
3605  */
3606 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3607         unsigned int flags)
3608 {
3609     struct vm_fault vmf = {
3610         .vma = vma,
3611         .address = address & PAGE_MASK,
3612         .flags = flags,
3613         .pgoff = linear_page_index(vma, address),
3614         .gfp_mask = __get_fault_gfp_mask(vma),
3615     };
3616     struct mm_struct *mm = vma->vm_mm;
3617     pgd_t *pgd;
3618     pud_t *pud;
3619 
3620     pgd = pgd_offset(mm, address);
3621     pud = pud_alloc(mm, pgd, address);
3622     if (!pud)
3623         return VM_FAULT_OOM;
3624     vmf.pmd = pmd_alloc(mm, pud, address);
3625     if (!vmf.pmd)
3626         return VM_FAULT_OOM;
3627     if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3628         int ret = create_huge_pmd(&vmf);
3629         if (!(ret & VM_FAULT_FALLBACK))
3630             return ret;
3631     } else {
3632         pmd_t orig_pmd = *vmf.pmd;
3633         int ret;
3634 
3635         barrier();
3636         if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3637             if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3638                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3639 
3640             if ((vmf.flags & FAULT_FLAG_WRITE) &&
3641                     !pmd_write(orig_pmd)) {
3642                 ret = wp_huge_pmd(&vmf, orig_pmd);
3643                 if (!(ret & VM_FAULT_FALLBACK))
3644                     return ret;
3645             } else {
3646                 huge_pmd_set_accessed(&vmf, orig_pmd);
3647                 return 0;
3648             }
3649         }
3650     }
3651 
3652     return handle_pte_fault(&vmf);
3653 }
3654 
3655 /*
3656  * By the time we get here, we already hold the mm semaphore
3657  *
3658  * The mmap_sem may have been released depending on flags and our
3659  * return value.  See filemap_fault() and __lock_page_or_retry().
3660  */
3661 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3662         unsigned int flags)
3663 {
3664     int ret;
3665 
3666     __set_current_state(TASK_RUNNING);
3667 
3668     count_vm_event(PGFAULT);
3669     mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3670 
3671     /* do counter updates before entering really critical section. */
3672     check_sync_rss_stat(current);
3673 
3674     /*
3675      * Enable the memcg OOM handling for faults triggered in user
3676      * space.  Kernel faults are handled more gracefully.
3677      */
3678     if (flags & FAULT_FLAG_USER)
3679         mem_cgroup_oom_enable();
3680 
3681     if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3682                         flags & FAULT_FLAG_INSTRUCTION,
3683                         flags & FAULT_FLAG_REMOTE))
3684         return VM_FAULT_SIGSEGV;
3685 
3686     if (unlikely(is_vm_hugetlb_page(vma)))
3687         ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3688     else
3689         ret = __handle_mm_fault(vma, address, flags);
3690 
3691     if (flags & FAULT_FLAG_USER) {
3692         mem_cgroup_oom_disable();
3693                 /*
3694                  * The task may have entered a memcg OOM situation but
3695                  * if the allocation error was handled gracefully (no
3696                  * VM_FAULT_OOM), there is no need to kill anything.
3697                  * Just clean up the OOM state peacefully.
3698                  */
3699                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3700                         mem_cgroup_oom_synchronize(false);
3701     }
3702 
3703     /*
3704      * This mm has been already reaped by the oom reaper and so the
3705      * refault cannot be trusted in general. Anonymous refaults would
3706      * lose data and give a zero page instead e.g. This is especially
3707      * problem for use_mm() because regular tasks will just die and
3708      * the corrupted data will not be visible anywhere while kthread
3709      * will outlive the oom victim and potentially propagate the date
3710      * further.
3711      */
3712     if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3713                 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3714         ret = VM_FAULT_SIGBUS;
3715 
3716     return ret;
3717 }
3718 EXPORT_SYMBOL_GPL(handle_mm_fault);
3719 
3720 #ifndef __PAGETABLE_PUD_FOLDED
3721 /*
3722  * Allocate page upper directory.
3723  * We've already handled the fast-path in-line.
3724  */
3725 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3726 {
3727     pud_t *new = pud_alloc_one(mm, address);
3728     if (!new)
3729         return -ENOMEM;
3730 
3731     smp_wmb(); /* See comment in __pte_alloc */
3732 
3733     spin_lock(&mm->page_table_lock);
3734     if (pgd_present(*pgd))      /* Another has populated it */
3735         pud_free(mm, new);
3736     else
3737         pgd_populate(mm, pgd, new);
3738     spin_unlock(&mm->page_table_lock);
3739     return 0;
3740 }
3741 #endif /* __PAGETABLE_PUD_FOLDED */
3742 
3743 #ifndef __PAGETABLE_PMD_FOLDED
3744 /*
3745  * Allocate page middle directory.
3746  * We've already handled the fast-path in-line.
3747  */
3748 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3749 {
3750     pmd_t *new = pmd_alloc_one(mm, address);
3751     if (!new)
3752         return -ENOMEM;
3753 
3754     smp_wmb(); /* See comment in __pte_alloc */
3755 
3756     spin_lock(&mm->page_table_lock);
3757 #ifndef __ARCH_HAS_4LEVEL_HACK
3758     if (!pud_present(*pud)) {
3759         mm_inc_nr_pmds(mm);
3760         pud_populate(mm, pud, new);
3761     } else  /* Another has populated it */
3762         pmd_free(mm, new);
3763 #else
3764     if (!pgd_present(*pud)) {
3765         mm_inc_nr_pmds(mm);
3766         pgd_populate(mm, pud, new);
3767     } else /* Another has populated it */
3768         pmd_free(mm, new);
3769 #endif /* __ARCH_HAS_4LEVEL_HACK */
3770     spin_unlock(&mm->page_table_lock);
3771     return 0;
3772 }
3773 #endif /* __PAGETABLE_PMD_FOLDED */
3774 
3775 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3776         pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3777 {
3778     pgd_t *pgd;
3779     pud_t *pud;
3780     pmd_t *pmd;
3781     pte_t *ptep;
3782 
3783     pgd = pgd_offset(mm, address);
3784     if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3785         goto out;
3786 
3787     pud = pud_offset(pgd, address);
3788     if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3789         goto out;
3790 
3791     pmd = pmd_offset(pud, address);
3792     VM_BUG_ON(pmd_trans_huge(*pmd));
3793 
3794     if (pmd_huge(*pmd)) {
3795         if (!pmdpp)
3796             goto out;
3797 
3798         *ptlp = pmd_lock(mm, pmd);
3799         if (pmd_huge(*pmd)) {
3800             *pmdpp = pmd;
3801             return 0;
3802         }
3803         spin_unlock(*ptlp);
3804     }
3805 
3806     if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3807         goto out;
3808 
3809     ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3810     if (!ptep)
3811         goto out;
3812     if (!pte_present(*ptep))
3813         goto unlock;
3814     *ptepp = ptep;
3815     return 0;
3816 unlock:
3817     pte_unmap_unlock(ptep, *ptlp);
3818 out:
3819     return -EINVAL;
3820 }
3821 
3822 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3823                  pte_t **ptepp, spinlock_t **ptlp)
3824 {
3825     int res;
3826 
3827     /* (void) is needed to make gcc happy */
3828     (void) __cond_lock(*ptlp,
3829                !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
3830                        ptlp)));
3831     return res;
3832 }
3833 
3834 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3835                  pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3836 {
3837     int res;
3838 
3839     /* (void) is needed to make gcc happy */
3840     (void) __cond_lock(*ptlp,
3841                !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
3842                        ptlp)));
3843     return res;
3844 }
3845 EXPORT_SYMBOL(follow_pte_pmd);
3846 
3847 /**
3848  * follow_pfn - look up PFN at a user virtual address
3849  * @vma: memory mapping
3850  * @address: user virtual address
3851  * @pfn: location to store found PFN
3852  *
3853  * Only IO mappings and raw PFN mappings are allowed.
3854  *
3855  * Returns zero and the pfn at @pfn on success, -ve otherwise.
3856  */
3857 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3858     unsigned long *pfn)
3859 {
3860     int ret = -EINVAL;
3861     spinlock_t *ptl;
3862     pte_t *ptep;
3863 
3864     if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3865         return ret;
3866 
3867     ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3868     if (ret)
3869         return ret;
3870     *pfn = pte_pfn(*ptep);
3871     pte_unmap_unlock(ptep, ptl);
3872     return 0;
3873 }
3874 EXPORT_SYMBOL(follow_pfn);
3875 
3876 #ifdef CONFIG_HAVE_IOREMAP_PROT
3877 int follow_phys(struct vm_area_struct *vma,
3878         unsigned long address, unsigned int flags,
3879         unsigned long *prot, resource_size_t *phys)
3880 {
3881     int ret = -EINVAL;
3882     pte_t *ptep, pte;
3883     spinlock_t *ptl;
3884 
3885     if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3886         goto out;
3887 
3888     if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3889         goto out;
3890     pte = *ptep;
3891 
3892     if ((flags & FOLL_WRITE) && !pte_write(pte))
3893         goto unlock;
3894 
3895     *prot = pgprot_val(pte_pgprot(pte));
3896     *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3897 
3898     ret = 0;
3899 unlock:
3900     pte_unmap_unlock(ptep, ptl);
3901 out:
3902     return ret;
3903 }
3904 
3905 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3906             void *buf, int len, int write)
3907 {
3908     resource_size_t phys_addr;
3909     unsigned long prot = 0;
3910     void __iomem *maddr;
3911     int offset = addr & (PAGE_SIZE-1);
3912 
3913     if (follow_phys(vma, addr, write, &prot, &phys_addr))
3914         return -EINVAL;
3915 
3916     maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3917     if (write)
3918         memcpy_toio(maddr + offset, buf, len);
3919     else
3920         memcpy_fromio(buf, maddr + offset, len);
3921     iounmap(maddr);
3922 
3923     return len;
3924 }
3925 EXPORT_SYMBOL_GPL(generic_access_phys);
3926 #endif
3927 
3928 /*
3929  * Access another process' address space as given in mm.  If non-NULL, use the
3930  * given task for page fault accounting.
3931  */
3932 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3933         unsigned long addr, void *buf, int len, unsigned int gup_flags)
3934 {
3935     struct vm_area_struct *vma;
3936     void *old_buf = buf;
3937     int write = gup_flags & FOLL_WRITE;
3938 
3939     down_read(&mm->mmap_sem);
3940     /* ignore errors, just check how much was successfully transferred */
3941     while (len) {
3942         int bytes, ret, offset;
3943         void *maddr;
3944         struct page *page = NULL;
3945 
3946         ret = get_user_pages_remote(tsk, mm, addr, 1,
3947                 gup_flags, &page, &vma, NULL);
3948         if (ret <= 0) {
3949 #ifndef CONFIG_HAVE_IOREMAP_PROT
3950             break;
3951 #else
3952             /*
3953              * Check if this is a VM_IO | VM_PFNMAP VMA, which
3954              * we can access using slightly different code.
3955              */
3956             vma = find_vma(mm, addr);
3957             if (!vma || vma->vm_start > addr)
3958                 break;
3959             if (vma->vm_ops && vma->vm_ops->access)
3960                 ret = vma->vm_ops->access(vma, addr, buf,
3961                               len, write);
3962             if (ret <= 0)
3963                 break;
3964             bytes = ret;
3965 #endif
3966         } else {
3967             bytes = len;
3968             offset = addr & (PAGE_SIZE-1);
3969             if (bytes > PAGE_SIZE-offset)
3970                 bytes = PAGE_SIZE-offset;
3971 
3972             maddr = kmap(page);
3973             if (write) {
3974                 copy_to_user_page(vma, page, addr,
3975                           maddr + offset, buf, bytes);
3976                 set_page_dirty_lock(page);
3977             } else {
3978                 copy_from_user_page(vma, page, addr,
3979                             buf, maddr + offset, bytes);
3980             }
3981             kunmap(page);
3982             put_page(page);
3983         }
3984         len -= bytes;
3985         buf += bytes;
3986         addr += bytes;
3987     }
3988     up_read(&mm->mmap_sem);
3989 
3990     return buf - old_buf;
3991 }
3992 
3993 /**
3994  * access_remote_vm - access another process' address space
3995  * @mm:     the mm_struct of the target address space
3996  * @addr:   start address to access
3997  * @buf:    source or destination buffer
3998  * @len:    number of bytes to transfer
3999  * @gup_flags:  flags modifying lookup behaviour
4000  *
4001  * The caller must hold a reference on @mm.
4002  */
4003 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4004         void *buf, int len, unsigned int gup_flags)
4005 {
4006     return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4007 }
4008 
4009 /*
4010  * Access another process' address space.
4011  * Source/target buffer must be kernel space,
4012  * Do not walk the page table directly, use get_user_pages
4013  */
4014 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4015         void *buf, int len, unsigned int gup_flags)
4016 {
4017     struct mm_struct *mm;
4018     int ret;
4019 
4020     mm = get_task_mm(tsk);
4021     if (!mm)
4022         return 0;
4023 
4024     ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4025 
4026     mmput(mm);
4027 
4028     return ret;
4029 }
4030 EXPORT_SYMBOL_GPL(access_process_vm);
4031 
4032 /*
4033  * Print the name of a VMA.
4034  */
4035 void print_vma_addr(char *prefix, unsigned long ip)
4036 {
4037     struct mm_struct *mm = current->mm;
4038     struct vm_area_struct *vma;
4039 
4040     /*
4041      * Do not print if we are in atomic
4042      * contexts (in exception stacks, etc.):
4043      */
4044     if (preempt_count())
4045         return;
4046 
4047     down_read(&mm->mmap_sem);
4048     vma = find_vma(mm, ip);
4049     if (vma && vma->vm_file) {
4050         struct file *f = vma->vm_file;
4051         char *buf = (char *)__get_free_page(GFP_KERNEL);
4052         if (buf) {
4053             char *p;
4054 
4055             p = file_path(f, buf, PAGE_SIZE);
4056             if (IS_ERR(p))
4057                 p = "?";
4058             printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4059                     vma->vm_start,
4060                     vma->vm_end - vma->vm_start);
4061             free_page((unsigned long)buf);
4062         }
4063     }
4064     up_read(&mm->mmap_sem);
4065 }
4066 
4067 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4068 void __might_fault(const char *file, int line)
4069 {
4070     /*
4071      * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4072      * holding the mmap_sem, this is safe because kernel memory doesn't
4073      * get paged out, therefore we'll never actually fault, and the
4074      * below annotations will generate false positives.
4075      */
4076     if (segment_eq(get_fs(), KERNEL_DS))
4077         return;
4078     if (pagefault_disabled())
4079         return;
4080     __might_sleep(file, line, 0);
4081 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4082     if (current->mm)
4083         might_lock_read(&current->mm->mmap_sem);
4084 #endif
4085 }
4086 EXPORT_SYMBOL(__might_fault);
4087 #endif
4088 
4089 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4090 static void clear_gigantic_page(struct page *page,
4091                 unsigned long addr,
4092                 unsigned int pages_per_huge_page)
4093 {
4094     int i;
4095     struct page *p = page;
4096 
4097     might_sleep();
4098     for (i = 0; i < pages_per_huge_page;
4099          i++, p = mem_map_next(p, page, i)) {
4100         cond_resched();
4101         clear_user_highpage(p, addr + i * PAGE_SIZE);
4102     }
4103 }
4104 void clear_huge_page(struct page *page,
4105              unsigned long addr, unsigned int pages_per_huge_page)
4106 {
4107     int i;
4108 
4109     if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4110         clear_gigantic_page(page, addr, pages_per_huge_page);
4111         return;
4112     }
4113 
4114     might_sleep();
4115     for (i = 0; i < pages_per_huge_page; i++) {
4116         cond_resched();
4117         clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4118     }
4119 }
4120 
4121 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4122                     unsigned long addr,
4123                     struct vm_area_struct *vma,
4124                     unsigned int pages_per_huge_page)
4125 {
4126     int i;
4127     struct page *dst_base = dst;
4128     struct page *src_base = src;
4129 
4130     for (i = 0; i < pages_per_huge_page; ) {
4131         cond_resched();
4132         copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4133 
4134         i++;
4135         dst = mem_map_next(dst, dst_base, i);
4136         src = mem_map_next(src, src_base, i);
4137     }
4138 }
4139 
4140 void copy_user_huge_page(struct page *dst, struct page *src,
4141              unsigned long addr, struct vm_area_struct *vma,
4142              unsigned int pages_per_huge_page)
4143 {
4144     int i;
4145 
4146     if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4147         copy_user_gigantic_page(dst, src, addr, vma,
4148                     pages_per_huge_page);
4149         return;
4150     }
4151 
4152     might_sleep();
4153     for (i = 0; i < pages_per_huge_page; i++) {
4154         cond_resched();
4155         copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4156     }
4157 }
4158 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4159 
4160 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4161 
4162 static struct kmem_cache *page_ptl_cachep;
4163 
4164 void __init ptlock_cache_init(void)
4165 {
4166     page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4167             SLAB_PANIC, NULL);
4168 }
4169 
4170 bool ptlock_alloc(struct page *page)
4171 {
4172     spinlock_t *ptl;
4173 
4174     ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4175     if (!ptl)
4176         return false;
4177     page->ptl = ptl;
4178     return true;
4179 }
4180 
4181 void ptlock_free(struct page *page)
4182 {
4183     kmem_cache_free(page_ptl_cachep, page->ptl);
4184 }
4185 #endif