Back to home page

LXR

 
 

    


0001 #include <linux/kernel.h>
0002 #include <linux/errno.h>
0003 #include <linux/err.h>
0004 #include <linux/spinlock.h>
0005 
0006 #include <linux/mm.h>
0007 #include <linux/memremap.h>
0008 #include <linux/pagemap.h>
0009 #include <linux/rmap.h>
0010 #include <linux/swap.h>
0011 #include <linux/swapops.h>
0012 
0013 #include <linux/sched.h>
0014 #include <linux/rwsem.h>
0015 #include <linux/hugetlb.h>
0016 
0017 #include <asm/mmu_context.h>
0018 #include <asm/pgtable.h>
0019 #include <asm/tlbflush.h>
0020 
0021 #include "internal.h"
0022 
0023 static struct page *no_page_table(struct vm_area_struct *vma,
0024         unsigned int flags)
0025 {
0026     /*
0027      * When core dumping an enormous anonymous area that nobody
0028      * has touched so far, we don't want to allocate unnecessary pages or
0029      * page tables.  Return error instead of NULL to skip handle_mm_fault,
0030      * then get_dump_page() will return NULL to leave a hole in the dump.
0031      * But we can only make this optimization where a hole would surely
0032      * be zero-filled if handle_mm_fault() actually did handle it.
0033      */
0034     if ((flags & FOLL_DUMP) && (!vma->vm_ops || !vma->vm_ops->fault))
0035         return ERR_PTR(-EFAULT);
0036     return NULL;
0037 }
0038 
0039 static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
0040         pte_t *pte, unsigned int flags)
0041 {
0042     /* No page to get reference */
0043     if (flags & FOLL_GET)
0044         return -EFAULT;
0045 
0046     if (flags & FOLL_TOUCH) {
0047         pte_t entry = *pte;
0048 
0049         if (flags & FOLL_WRITE)
0050             entry = pte_mkdirty(entry);
0051         entry = pte_mkyoung(entry);
0052 
0053         if (!pte_same(*pte, entry)) {
0054             set_pte_at(vma->vm_mm, address, pte, entry);
0055             update_mmu_cache(vma, address, pte);
0056         }
0057     }
0058 
0059     /* Proper page table entry exists, but no corresponding struct page */
0060     return -EEXIST;
0061 }
0062 
0063 /*
0064  * FOLL_FORCE can write to even unwritable pte's, but only
0065  * after we've gone through a COW cycle and they are dirty.
0066  */
0067 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
0068 {
0069     return pte_write(pte) ||
0070         ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
0071 }
0072 
0073 static struct page *follow_page_pte(struct vm_area_struct *vma,
0074         unsigned long address, pmd_t *pmd, unsigned int flags)
0075 {
0076     struct mm_struct *mm = vma->vm_mm;
0077     struct dev_pagemap *pgmap = NULL;
0078     struct page *page;
0079     spinlock_t *ptl;
0080     pte_t *ptep, pte;
0081 
0082 retry:
0083     if (unlikely(pmd_bad(*pmd)))
0084         return no_page_table(vma, flags);
0085 
0086     ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
0087     pte = *ptep;
0088     if (!pte_present(pte)) {
0089         swp_entry_t entry;
0090         /*
0091          * KSM's break_ksm() relies upon recognizing a ksm page
0092          * even while it is being migrated, so for that case we
0093          * need migration_entry_wait().
0094          */
0095         if (likely(!(flags & FOLL_MIGRATION)))
0096             goto no_page;
0097         if (pte_none(pte))
0098             goto no_page;
0099         entry = pte_to_swp_entry(pte);
0100         if (!is_migration_entry(entry))
0101             goto no_page;
0102         pte_unmap_unlock(ptep, ptl);
0103         migration_entry_wait(mm, pmd, address);
0104         goto retry;
0105     }
0106     if ((flags & FOLL_NUMA) && pte_protnone(pte))
0107         goto no_page;
0108     if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags)) {
0109         pte_unmap_unlock(ptep, ptl);
0110         return NULL;
0111     }
0112 
0113     page = vm_normal_page(vma, address, pte);
0114     if (!page && pte_devmap(pte) && (flags & FOLL_GET)) {
0115         /*
0116          * Only return device mapping pages in the FOLL_GET case since
0117          * they are only valid while holding the pgmap reference.
0118          */
0119         pgmap = get_dev_pagemap(pte_pfn(pte), NULL);
0120         if (pgmap)
0121             page = pte_page(pte);
0122         else
0123             goto no_page;
0124     } else if (unlikely(!page)) {
0125         if (flags & FOLL_DUMP) {
0126             /* Avoid special (like zero) pages in core dumps */
0127             page = ERR_PTR(-EFAULT);
0128             goto out;
0129         }
0130 
0131         if (is_zero_pfn(pte_pfn(pte))) {
0132             page = pte_page(pte);
0133         } else {
0134             int ret;
0135 
0136             ret = follow_pfn_pte(vma, address, ptep, flags);
0137             page = ERR_PTR(ret);
0138             goto out;
0139         }
0140     }
0141 
0142     if (flags & FOLL_SPLIT && PageTransCompound(page)) {
0143         int ret;
0144         get_page(page);
0145         pte_unmap_unlock(ptep, ptl);
0146         lock_page(page);
0147         ret = split_huge_page(page);
0148         unlock_page(page);
0149         put_page(page);
0150         if (ret)
0151             return ERR_PTR(ret);
0152         goto retry;
0153     }
0154 
0155     if (flags & FOLL_GET) {
0156         get_page(page);
0157 
0158         /* drop the pgmap reference now that we hold the page */
0159         if (pgmap) {
0160             put_dev_pagemap(pgmap);
0161             pgmap = NULL;
0162         }
0163     }
0164     if (flags & FOLL_TOUCH) {
0165         if ((flags & FOLL_WRITE) &&
0166             !pte_dirty(pte) && !PageDirty(page))
0167             set_page_dirty(page);
0168         /*
0169          * pte_mkyoung() would be more correct here, but atomic care
0170          * is needed to avoid losing the dirty bit: it is easier to use
0171          * mark_page_accessed().
0172          */
0173         mark_page_accessed(page);
0174     }
0175     if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
0176         /* Do not mlock pte-mapped THP */
0177         if (PageTransCompound(page))
0178             goto out;
0179 
0180         /*
0181          * The preliminary mapping check is mainly to avoid the
0182          * pointless overhead of lock_page on the ZERO_PAGE
0183          * which might bounce very badly if there is contention.
0184          *
0185          * If the page is already locked, we don't need to
0186          * handle it now - vmscan will handle it later if and
0187          * when it attempts to reclaim the page.
0188          */
0189         if (page->mapping && trylock_page(page)) {
0190             lru_add_drain();  /* push cached pages to LRU */
0191             /*
0192              * Because we lock page here, and migration is
0193              * blocked by the pte's page reference, and we
0194              * know the page is still mapped, we don't even
0195              * need to check for file-cache page truncation.
0196              */
0197             mlock_vma_page(page);
0198             unlock_page(page);
0199         }
0200     }
0201 out:
0202     pte_unmap_unlock(ptep, ptl);
0203     return page;
0204 no_page:
0205     pte_unmap_unlock(ptep, ptl);
0206     if (!pte_none(pte))
0207         return NULL;
0208     return no_page_table(vma, flags);
0209 }
0210 
0211 /**
0212  * follow_page_mask - look up a page descriptor from a user-virtual address
0213  * @vma: vm_area_struct mapping @address
0214  * @address: virtual address to look up
0215  * @flags: flags modifying lookup behaviour
0216  * @page_mask: on output, *page_mask is set according to the size of the page
0217  *
0218  * @flags can have FOLL_ flags set, defined in <linux/mm.h>
0219  *
0220  * Returns the mapped (struct page *), %NULL if no mapping exists, or
0221  * an error pointer if there is a mapping to something not represented
0222  * by a page descriptor (see also vm_normal_page()).
0223  */
0224 struct page *follow_page_mask(struct vm_area_struct *vma,
0225                   unsigned long address, unsigned int flags,
0226                   unsigned int *page_mask)
0227 {
0228     pgd_t *pgd;
0229     pud_t *pud;
0230     pmd_t *pmd;
0231     spinlock_t *ptl;
0232     struct page *page;
0233     struct mm_struct *mm = vma->vm_mm;
0234 
0235     *page_mask = 0;
0236 
0237     page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
0238     if (!IS_ERR(page)) {
0239         BUG_ON(flags & FOLL_GET);
0240         return page;
0241     }
0242 
0243     pgd = pgd_offset(mm, address);
0244     if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
0245         return no_page_table(vma, flags);
0246 
0247     pud = pud_offset(pgd, address);
0248     if (pud_none(*pud))
0249         return no_page_table(vma, flags);
0250     if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
0251         page = follow_huge_pud(mm, address, pud, flags);
0252         if (page)
0253             return page;
0254         return no_page_table(vma, flags);
0255     }
0256     if (unlikely(pud_bad(*pud)))
0257         return no_page_table(vma, flags);
0258 
0259     pmd = pmd_offset(pud, address);
0260     if (pmd_none(*pmd))
0261         return no_page_table(vma, flags);
0262     if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
0263         page = follow_huge_pmd(mm, address, pmd, flags);
0264         if (page)
0265             return page;
0266         return no_page_table(vma, flags);
0267     }
0268     if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
0269         return no_page_table(vma, flags);
0270     if (pmd_devmap(*pmd)) {
0271         ptl = pmd_lock(mm, pmd);
0272         page = follow_devmap_pmd(vma, address, pmd, flags);
0273         spin_unlock(ptl);
0274         if (page)
0275             return page;
0276     }
0277     if (likely(!pmd_trans_huge(*pmd)))
0278         return follow_page_pte(vma, address, pmd, flags);
0279 
0280     ptl = pmd_lock(mm, pmd);
0281     if (unlikely(!pmd_trans_huge(*pmd))) {
0282         spin_unlock(ptl);
0283         return follow_page_pte(vma, address, pmd, flags);
0284     }
0285     if (flags & FOLL_SPLIT) {
0286         int ret;
0287         page = pmd_page(*pmd);
0288         if (is_huge_zero_page(page)) {
0289             spin_unlock(ptl);
0290             ret = 0;
0291             split_huge_pmd(vma, pmd, address);
0292             if (pmd_trans_unstable(pmd))
0293                 ret = -EBUSY;
0294         } else {
0295             get_page(page);
0296             spin_unlock(ptl);
0297             lock_page(page);
0298             ret = split_huge_page(page);
0299             unlock_page(page);
0300             put_page(page);
0301             if (pmd_none(*pmd))
0302                 return no_page_table(vma, flags);
0303         }
0304 
0305         return ret ? ERR_PTR(ret) :
0306             follow_page_pte(vma, address, pmd, flags);
0307     }
0308 
0309     page = follow_trans_huge_pmd(vma, address, pmd, flags);
0310     spin_unlock(ptl);
0311     *page_mask = HPAGE_PMD_NR - 1;
0312     return page;
0313 }
0314 
0315 static int get_gate_page(struct mm_struct *mm, unsigned long address,
0316         unsigned int gup_flags, struct vm_area_struct **vma,
0317         struct page **page)
0318 {
0319     pgd_t *pgd;
0320     pud_t *pud;
0321     pmd_t *pmd;
0322     pte_t *pte;
0323     int ret = -EFAULT;
0324 
0325     /* user gate pages are read-only */
0326     if (gup_flags & FOLL_WRITE)
0327         return -EFAULT;
0328     if (address > TASK_SIZE)
0329         pgd = pgd_offset_k(address);
0330     else
0331         pgd = pgd_offset_gate(mm, address);
0332     BUG_ON(pgd_none(*pgd));
0333     pud = pud_offset(pgd, address);
0334     BUG_ON(pud_none(*pud));
0335     pmd = pmd_offset(pud, address);
0336     if (pmd_none(*pmd))
0337         return -EFAULT;
0338     VM_BUG_ON(pmd_trans_huge(*pmd));
0339     pte = pte_offset_map(pmd, address);
0340     if (pte_none(*pte))
0341         goto unmap;
0342     *vma = get_gate_vma(mm);
0343     if (!page)
0344         goto out;
0345     *page = vm_normal_page(*vma, address, *pte);
0346     if (!*page) {
0347         if ((gup_flags & FOLL_DUMP) || !is_zero_pfn(pte_pfn(*pte)))
0348             goto unmap;
0349         *page = pte_page(*pte);
0350     }
0351     get_page(*page);
0352 out:
0353     ret = 0;
0354 unmap:
0355     pte_unmap(pte);
0356     return ret;
0357 }
0358 
0359 /*
0360  * mmap_sem must be held on entry.  If @nonblocking != NULL and
0361  * *@flags does not include FOLL_NOWAIT, the mmap_sem may be released.
0362  * If it is, *@nonblocking will be set to 0 and -EBUSY returned.
0363  */
0364 static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
0365         unsigned long address, unsigned int *flags, int *nonblocking)
0366 {
0367     unsigned int fault_flags = 0;
0368     int ret;
0369 
0370     /* mlock all present pages, but do not fault in new pages */
0371     if ((*flags & (FOLL_POPULATE | FOLL_MLOCK)) == FOLL_MLOCK)
0372         return -ENOENT;
0373     /* For mm_populate(), just skip the stack guard page. */
0374     if ((*flags & FOLL_POPULATE) &&
0375             (stack_guard_page_start(vma, address) ||
0376              stack_guard_page_end(vma, address + PAGE_SIZE)))
0377         return -ENOENT;
0378     if (*flags & FOLL_WRITE)
0379         fault_flags |= FAULT_FLAG_WRITE;
0380     if (*flags & FOLL_REMOTE)
0381         fault_flags |= FAULT_FLAG_REMOTE;
0382     if (nonblocking)
0383         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
0384     if (*flags & FOLL_NOWAIT)
0385         fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
0386     if (*flags & FOLL_TRIED) {
0387         VM_WARN_ON_ONCE(fault_flags & FAULT_FLAG_ALLOW_RETRY);
0388         fault_flags |= FAULT_FLAG_TRIED;
0389     }
0390 
0391     ret = handle_mm_fault(vma, address, fault_flags);
0392     if (ret & VM_FAULT_ERROR) {
0393         if (ret & VM_FAULT_OOM)
0394             return -ENOMEM;
0395         if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
0396             return *flags & FOLL_HWPOISON ? -EHWPOISON : -EFAULT;
0397         if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
0398             return -EFAULT;
0399         BUG();
0400     }
0401 
0402     if (tsk) {
0403         if (ret & VM_FAULT_MAJOR)
0404             tsk->maj_flt++;
0405         else
0406             tsk->min_flt++;
0407     }
0408 
0409     if (ret & VM_FAULT_RETRY) {
0410         if (nonblocking)
0411             *nonblocking = 0;
0412         return -EBUSY;
0413     }
0414 
0415     /*
0416      * The VM_FAULT_WRITE bit tells us that do_wp_page has broken COW when
0417      * necessary, even if maybe_mkwrite decided not to set pte_write. We
0418      * can thus safely do subsequent page lookups as if they were reads.
0419      * But only do so when looping for pte_write is futile: in some cases
0420      * userspace may also be wanting to write to the gotten user page,
0421      * which a read fault here might prevent (a readonly page might get
0422      * reCOWed by userspace write).
0423      */
0424     if ((ret & VM_FAULT_WRITE) && !(vma->vm_flags & VM_WRITE))
0425             *flags |= FOLL_COW;
0426     return 0;
0427 }
0428 
0429 static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags)
0430 {
0431     vm_flags_t vm_flags = vma->vm_flags;
0432     int write = (gup_flags & FOLL_WRITE);
0433     int foreign = (gup_flags & FOLL_REMOTE);
0434 
0435     if (vm_flags & (VM_IO | VM_PFNMAP))
0436         return -EFAULT;
0437 
0438     if (write) {
0439         if (!(vm_flags & VM_WRITE)) {
0440             if (!(gup_flags & FOLL_FORCE))
0441                 return -EFAULT;
0442             /*
0443              * We used to let the write,force case do COW in a
0444              * VM_MAYWRITE VM_SHARED !VM_WRITE vma, so ptrace could
0445              * set a breakpoint in a read-only mapping of an
0446              * executable, without corrupting the file (yet only
0447              * when that file had been opened for writing!).
0448              * Anon pages in shared mappings are surprising: now
0449              * just reject it.
0450              */
0451             if (!is_cow_mapping(vm_flags))
0452                 return -EFAULT;
0453         }
0454     } else if (!(vm_flags & VM_READ)) {
0455         if (!(gup_flags & FOLL_FORCE))
0456             return -EFAULT;
0457         /*
0458          * Is there actually any vma we can reach here which does not
0459          * have VM_MAYREAD set?
0460          */
0461         if (!(vm_flags & VM_MAYREAD))
0462             return -EFAULT;
0463     }
0464     /*
0465      * gups are always data accesses, not instruction
0466      * fetches, so execute=false here
0467      */
0468     if (!arch_vma_access_permitted(vma, write, false, foreign))
0469         return -EFAULT;
0470     return 0;
0471 }
0472 
0473 /**
0474  * __get_user_pages() - pin user pages in memory
0475  * @tsk:    task_struct of target task
0476  * @mm:     mm_struct of target mm
0477  * @start:  starting user address
0478  * @nr_pages:   number of pages from start to pin
0479  * @gup_flags:  flags modifying pin behaviour
0480  * @pages:  array that receives pointers to the pages pinned.
0481  *      Should be at least nr_pages long. Or NULL, if caller
0482  *      only intends to ensure the pages are faulted in.
0483  * @vmas:   array of pointers to vmas corresponding to each page.
0484  *      Or NULL if the caller does not require them.
0485  * @nonblocking: whether waiting for disk IO or mmap_sem contention
0486  *
0487  * Returns number of pages pinned. This may be fewer than the number
0488  * requested. If nr_pages is 0 or negative, returns 0. If no pages
0489  * were pinned, returns -errno. Each page returned must be released
0490  * with a put_page() call when it is finished with. vmas will only
0491  * remain valid while mmap_sem is held.
0492  *
0493  * Must be called with mmap_sem held.  It may be released.  See below.
0494  *
0495  * __get_user_pages walks a process's page tables and takes a reference to
0496  * each struct page that each user address corresponds to at a given
0497  * instant. That is, it takes the page that would be accessed if a user
0498  * thread accesses the given user virtual address at that instant.
0499  *
0500  * This does not guarantee that the page exists in the user mappings when
0501  * __get_user_pages returns, and there may even be a completely different
0502  * page there in some cases (eg. if mmapped pagecache has been invalidated
0503  * and subsequently re faulted). However it does guarantee that the page
0504  * won't be freed completely. And mostly callers simply care that the page
0505  * contains data that was valid *at some point in time*. Typically, an IO
0506  * or similar operation cannot guarantee anything stronger anyway because
0507  * locks can't be held over the syscall boundary.
0508  *
0509  * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
0510  * the page is written to, set_page_dirty (or set_page_dirty_lock, as
0511  * appropriate) must be called after the page is finished with, and
0512  * before put_page is called.
0513  *
0514  * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
0515  * or mmap_sem contention, and if waiting is needed to pin all pages,
0516  * *@nonblocking will be set to 0.  Further, if @gup_flags does not
0517  * include FOLL_NOWAIT, the mmap_sem will be released via up_read() in
0518  * this case.
0519  *
0520  * A caller using such a combination of @nonblocking and @gup_flags
0521  * must therefore hold the mmap_sem for reading only, and recognize
0522  * when it's been released.  Otherwise, it must be held for either
0523  * reading or writing and will not be released.
0524  *
0525  * In most cases, get_user_pages or get_user_pages_fast should be used
0526  * instead of __get_user_pages. __get_user_pages should be used only if
0527  * you need some special @gup_flags.
0528  */
0529 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
0530         unsigned long start, unsigned long nr_pages,
0531         unsigned int gup_flags, struct page **pages,
0532         struct vm_area_struct **vmas, int *nonblocking)
0533 {
0534     long i = 0;
0535     unsigned int page_mask;
0536     struct vm_area_struct *vma = NULL;
0537 
0538     if (!nr_pages)
0539         return 0;
0540 
0541     VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
0542 
0543     /*
0544      * If FOLL_FORCE is set then do not force a full fault as the hinting
0545      * fault information is unrelated to the reference behaviour of a task
0546      * using the address space
0547      */
0548     if (!(gup_flags & FOLL_FORCE))
0549         gup_flags |= FOLL_NUMA;
0550 
0551     do {
0552         struct page *page;
0553         unsigned int foll_flags = gup_flags;
0554         unsigned int page_increm;
0555 
0556         /* first iteration or cross vma bound */
0557         if (!vma || start >= vma->vm_end) {
0558             vma = find_extend_vma(mm, start);
0559             if (!vma && in_gate_area(mm, start)) {
0560                 int ret;
0561                 ret = get_gate_page(mm, start & PAGE_MASK,
0562                         gup_flags, &vma,
0563                         pages ? &pages[i] : NULL);
0564                 if (ret)
0565                     return i ? : ret;
0566                 page_mask = 0;
0567                 goto next_page;
0568             }
0569 
0570             if (!vma || check_vma_flags(vma, gup_flags))
0571                 return i ? : -EFAULT;
0572             if (is_vm_hugetlb_page(vma)) {
0573                 i = follow_hugetlb_page(mm, vma, pages, vmas,
0574                         &start, &nr_pages, i,
0575                         gup_flags);
0576                 continue;
0577             }
0578         }
0579 retry:
0580         /*
0581          * If we have a pending SIGKILL, don't keep faulting pages and
0582          * potentially allocating memory.
0583          */
0584         if (unlikely(fatal_signal_pending(current)))
0585             return i ? i : -ERESTARTSYS;
0586         cond_resched();
0587         page = follow_page_mask(vma, start, foll_flags, &page_mask);
0588         if (!page) {
0589             int ret;
0590             ret = faultin_page(tsk, vma, start, &foll_flags,
0591                     nonblocking);
0592             switch (ret) {
0593             case 0:
0594                 goto retry;
0595             case -EFAULT:
0596             case -ENOMEM:
0597             case -EHWPOISON:
0598                 return i ? i : ret;
0599             case -EBUSY:
0600                 return i;
0601             case -ENOENT:
0602                 goto next_page;
0603             }
0604             BUG();
0605         } else if (PTR_ERR(page) == -EEXIST) {
0606             /*
0607              * Proper page table entry exists, but no corresponding
0608              * struct page.
0609              */
0610             goto next_page;
0611         } else if (IS_ERR(page)) {
0612             return i ? i : PTR_ERR(page);
0613         }
0614         if (pages) {
0615             pages[i] = page;
0616             flush_anon_page(vma, page, start);
0617             flush_dcache_page(page);
0618             page_mask = 0;
0619         }
0620 next_page:
0621         if (vmas) {
0622             vmas[i] = vma;
0623             page_mask = 0;
0624         }
0625         page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
0626         if (page_increm > nr_pages)
0627             page_increm = nr_pages;
0628         i += page_increm;
0629         start += page_increm * PAGE_SIZE;
0630         nr_pages -= page_increm;
0631     } while (nr_pages);
0632     return i;
0633 }
0634 
0635 static bool vma_permits_fault(struct vm_area_struct *vma,
0636                   unsigned int fault_flags)
0637 {
0638     bool write   = !!(fault_flags & FAULT_FLAG_WRITE);
0639     bool foreign = !!(fault_flags & FAULT_FLAG_REMOTE);
0640     vm_flags_t vm_flags = write ? VM_WRITE : VM_READ;
0641 
0642     if (!(vm_flags & vma->vm_flags))
0643         return false;
0644 
0645     /*
0646      * The architecture might have a hardware protection
0647      * mechanism other than read/write that can deny access.
0648      *
0649      * gup always represents data access, not instruction
0650      * fetches, so execute=false here:
0651      */
0652     if (!arch_vma_access_permitted(vma, write, false, foreign))
0653         return false;
0654 
0655     return true;
0656 }
0657 
0658 /*
0659  * fixup_user_fault() - manually resolve a user page fault
0660  * @tsk:    the task_struct to use for page fault accounting, or
0661  *      NULL if faults are not to be recorded.
0662  * @mm:     mm_struct of target mm
0663  * @address:    user address
0664  * @fault_flags:flags to pass down to handle_mm_fault()
0665  * @unlocked:   did we unlock the mmap_sem while retrying, maybe NULL if caller
0666  *      does not allow retry
0667  *
0668  * This is meant to be called in the specific scenario where for locking reasons
0669  * we try to access user memory in atomic context (within a pagefault_disable()
0670  * section), this returns -EFAULT, and we want to resolve the user fault before
0671  * trying again.
0672  *
0673  * Typically this is meant to be used by the futex code.
0674  *
0675  * The main difference with get_user_pages() is that this function will
0676  * unconditionally call handle_mm_fault() which will in turn perform all the
0677  * necessary SW fixup of the dirty and young bits in the PTE, while
0678  * get_user_pages() only guarantees to update these in the struct page.
0679  *
0680  * This is important for some architectures where those bits also gate the
0681  * access permission to the page because they are maintained in software.  On
0682  * such architectures, gup() will not be enough to make a subsequent access
0683  * succeed.
0684  *
0685  * This function will not return with an unlocked mmap_sem. So it has not the
0686  * same semantics wrt the @mm->mmap_sem as does filemap_fault().
0687  */
0688 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
0689              unsigned long address, unsigned int fault_flags,
0690              bool *unlocked)
0691 {
0692     struct vm_area_struct *vma;
0693     int ret, major = 0;
0694 
0695     if (unlocked)
0696         fault_flags |= FAULT_FLAG_ALLOW_RETRY;
0697 
0698 retry:
0699     vma = find_extend_vma(mm, address);
0700     if (!vma || address < vma->vm_start)
0701         return -EFAULT;
0702 
0703     if (!vma_permits_fault(vma, fault_flags))
0704         return -EFAULT;
0705 
0706     ret = handle_mm_fault(vma, address, fault_flags);
0707     major |= ret & VM_FAULT_MAJOR;
0708     if (ret & VM_FAULT_ERROR) {
0709         if (ret & VM_FAULT_OOM)
0710             return -ENOMEM;
0711         if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
0712             return -EHWPOISON;
0713         if (ret & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV))
0714             return -EFAULT;
0715         BUG();
0716     }
0717 
0718     if (ret & VM_FAULT_RETRY) {
0719         down_read(&mm->mmap_sem);
0720         if (!(fault_flags & FAULT_FLAG_TRIED)) {
0721             *unlocked = true;
0722             fault_flags &= ~FAULT_FLAG_ALLOW_RETRY;
0723             fault_flags |= FAULT_FLAG_TRIED;
0724             goto retry;
0725         }
0726     }
0727 
0728     if (tsk) {
0729         if (major)
0730             tsk->maj_flt++;
0731         else
0732             tsk->min_flt++;
0733     }
0734     return 0;
0735 }
0736 EXPORT_SYMBOL_GPL(fixup_user_fault);
0737 
0738 static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
0739                         struct mm_struct *mm,
0740                         unsigned long start,
0741                         unsigned long nr_pages,
0742                         struct page **pages,
0743                         struct vm_area_struct **vmas,
0744                         int *locked, bool notify_drop,
0745                         unsigned int flags)
0746 {
0747     long ret, pages_done;
0748     bool lock_dropped;
0749 
0750     if (locked) {
0751         /* if VM_FAULT_RETRY can be returned, vmas become invalid */
0752         BUG_ON(vmas);
0753         /* check caller initialized locked */
0754         BUG_ON(*locked != 1);
0755     }
0756 
0757     if (pages)
0758         flags |= FOLL_GET;
0759 
0760     pages_done = 0;
0761     lock_dropped = false;
0762     for (;;) {
0763         ret = __get_user_pages(tsk, mm, start, nr_pages, flags, pages,
0764                        vmas, locked);
0765         if (!locked)
0766             /* VM_FAULT_RETRY couldn't trigger, bypass */
0767             return ret;
0768 
0769         /* VM_FAULT_RETRY cannot return errors */
0770         if (!*locked) {
0771             BUG_ON(ret < 0);
0772             BUG_ON(ret >= nr_pages);
0773         }
0774 
0775         if (!pages)
0776             /* If it's a prefault don't insist harder */
0777             return ret;
0778 
0779         if (ret > 0) {
0780             nr_pages -= ret;
0781             pages_done += ret;
0782             if (!nr_pages)
0783                 break;
0784         }
0785         if (*locked) {
0786             /* VM_FAULT_RETRY didn't trigger */
0787             if (!pages_done)
0788                 pages_done = ret;
0789             break;
0790         }
0791         /* VM_FAULT_RETRY triggered, so seek to the faulting offset */
0792         pages += ret;
0793         start += ret << PAGE_SHIFT;
0794 
0795         /*
0796          * Repeat on the address that fired VM_FAULT_RETRY
0797          * without FAULT_FLAG_ALLOW_RETRY but with
0798          * FAULT_FLAG_TRIED.
0799          */
0800         *locked = 1;
0801         lock_dropped = true;
0802         down_read(&mm->mmap_sem);
0803         ret = __get_user_pages(tsk, mm, start, 1, flags | FOLL_TRIED,
0804                        pages, NULL, NULL);
0805         if (ret != 1) {
0806             BUG_ON(ret > 1);
0807             if (!pages_done)
0808                 pages_done = ret;
0809             break;
0810         }
0811         nr_pages--;
0812         pages_done++;
0813         if (!nr_pages)
0814             break;
0815         pages++;
0816         start += PAGE_SIZE;
0817     }
0818     if (notify_drop && lock_dropped && *locked) {
0819         /*
0820          * We must let the caller know we temporarily dropped the lock
0821          * and so the critical section protected by it was lost.
0822          */
0823         up_read(&mm->mmap_sem);
0824         *locked = 0;
0825     }
0826     return pages_done;
0827 }
0828 
0829 /*
0830  * We can leverage the VM_FAULT_RETRY functionality in the page fault
0831  * paths better by using either get_user_pages_locked() or
0832  * get_user_pages_unlocked().
0833  *
0834  * get_user_pages_locked() is suitable to replace the form:
0835  *
0836  *      down_read(&mm->mmap_sem);
0837  *      do_something()
0838  *      get_user_pages(tsk, mm, ..., pages, NULL);
0839  *      up_read(&mm->mmap_sem);
0840  *
0841  *  to:
0842  *
0843  *      int locked = 1;
0844  *      down_read(&mm->mmap_sem);
0845  *      do_something()
0846  *      get_user_pages_locked(tsk, mm, ..., pages, &locked);
0847  *      if (locked)
0848  *          up_read(&mm->mmap_sem);
0849  */
0850 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
0851                unsigned int gup_flags, struct page **pages,
0852                int *locked)
0853 {
0854     return __get_user_pages_locked(current, current->mm, start, nr_pages,
0855                        pages, NULL, locked, true,
0856                        gup_flags | FOLL_TOUCH);
0857 }
0858 EXPORT_SYMBOL(get_user_pages_locked);
0859 
0860 /*
0861  * Same as get_user_pages_unlocked(...., FOLL_TOUCH) but it allows for
0862  * tsk, mm to be specified.
0863  *
0864  * NOTE: here FOLL_TOUCH is not set implicitly and must be set by the
0865  * caller if required (just like with __get_user_pages). "FOLL_GET"
0866  * is set implicitly if "pages" is non-NULL.
0867  */
0868 static __always_inline long __get_user_pages_unlocked(struct task_struct *tsk,
0869         struct mm_struct *mm, unsigned long start,
0870         unsigned long nr_pages, struct page **pages,
0871         unsigned int gup_flags)
0872 {
0873     long ret;
0874     int locked = 1;
0875 
0876     down_read(&mm->mmap_sem);
0877     ret = __get_user_pages_locked(tsk, mm, start, nr_pages, pages, NULL,
0878                       &locked, false, gup_flags);
0879     if (locked)
0880         up_read(&mm->mmap_sem);
0881     return ret;
0882 }
0883 
0884 /*
0885  * get_user_pages_unlocked() is suitable to replace the form:
0886  *
0887  *      down_read(&mm->mmap_sem);
0888  *      get_user_pages(tsk, mm, ..., pages, NULL);
0889  *      up_read(&mm->mmap_sem);
0890  *
0891  *  with:
0892  *
0893  *      get_user_pages_unlocked(tsk, mm, ..., pages);
0894  *
0895  * It is functionally equivalent to get_user_pages_fast so
0896  * get_user_pages_fast should be used instead if specific gup_flags
0897  * (e.g. FOLL_FORCE) are not required.
0898  */
0899 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
0900                  struct page **pages, unsigned int gup_flags)
0901 {
0902     return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
0903                      pages, gup_flags | FOLL_TOUCH);
0904 }
0905 EXPORT_SYMBOL(get_user_pages_unlocked);
0906 
0907 /*
0908  * get_user_pages_remote() - pin user pages in memory
0909  * @tsk:    the task_struct to use for page fault accounting, or
0910  *      NULL if faults are not to be recorded.
0911  * @mm:     mm_struct of target mm
0912  * @start:  starting user address
0913  * @nr_pages:   number of pages from start to pin
0914  * @gup_flags:  flags modifying lookup behaviour
0915  * @pages:  array that receives pointers to the pages pinned.
0916  *      Should be at least nr_pages long. Or NULL, if caller
0917  *      only intends to ensure the pages are faulted in.
0918  * @vmas:   array of pointers to vmas corresponding to each page.
0919  *      Or NULL if the caller does not require them.
0920  * @locked: pointer to lock flag indicating whether lock is held and
0921  *      subsequently whether VM_FAULT_RETRY functionality can be
0922  *      utilised. Lock must initially be held.
0923  *
0924  * Returns number of pages pinned. This may be fewer than the number
0925  * requested. If nr_pages is 0 or negative, returns 0. If no pages
0926  * were pinned, returns -errno. Each page returned must be released
0927  * with a put_page() call when it is finished with. vmas will only
0928  * remain valid while mmap_sem is held.
0929  *
0930  * Must be called with mmap_sem held for read or write.
0931  *
0932  * get_user_pages walks a process's page tables and takes a reference to
0933  * each struct page that each user address corresponds to at a given
0934  * instant. That is, it takes the page that would be accessed if a user
0935  * thread accesses the given user virtual address at that instant.
0936  *
0937  * This does not guarantee that the page exists in the user mappings when
0938  * get_user_pages returns, and there may even be a completely different
0939  * page there in some cases (eg. if mmapped pagecache has been invalidated
0940  * and subsequently re faulted). However it does guarantee that the page
0941  * won't be freed completely. And mostly callers simply care that the page
0942  * contains data that was valid *at some point in time*. Typically, an IO
0943  * or similar operation cannot guarantee anything stronger anyway because
0944  * locks can't be held over the syscall boundary.
0945  *
0946  * If gup_flags & FOLL_WRITE == 0, the page must not be written to. If the page
0947  * is written to, set_page_dirty (or set_page_dirty_lock, as appropriate) must
0948  * be called after the page is finished with, and before put_page is called.
0949  *
0950  * get_user_pages is typically used for fewer-copy IO operations, to get a
0951  * handle on the memory by some means other than accesses via the user virtual
0952  * addresses. The pages may be submitted for DMA to devices or accessed via
0953  * their kernel linear mapping (via the kmap APIs). Care should be taken to
0954  * use the correct cache flushing APIs.
0955  *
0956  * See also get_user_pages_fast, for performance critical applications.
0957  *
0958  * get_user_pages should be phased out in favor of
0959  * get_user_pages_locked|unlocked or get_user_pages_fast. Nothing
0960  * should use get_user_pages because it cannot pass
0961  * FAULT_FLAG_ALLOW_RETRY to handle_mm_fault.
0962  */
0963 long get_user_pages_remote(struct task_struct *tsk, struct mm_struct *mm,
0964         unsigned long start, unsigned long nr_pages,
0965         unsigned int gup_flags, struct page **pages,
0966         struct vm_area_struct **vmas, int *locked)
0967 {
0968     return __get_user_pages_locked(tsk, mm, start, nr_pages, pages, vmas,
0969                        locked, true,
0970                        gup_flags | FOLL_TOUCH | FOLL_REMOTE);
0971 }
0972 EXPORT_SYMBOL(get_user_pages_remote);
0973 
0974 /*
0975  * This is the same as get_user_pages_remote(), just with a
0976  * less-flexible calling convention where we assume that the task
0977  * and mm being operated on are the current task's and don't allow
0978  * passing of a locked parameter.  We also obviously don't pass
0979  * FOLL_REMOTE in here.
0980  */
0981 long get_user_pages(unsigned long start, unsigned long nr_pages,
0982         unsigned int gup_flags, struct page **pages,
0983         struct vm_area_struct **vmas)
0984 {
0985     return __get_user_pages_locked(current, current->mm, start, nr_pages,
0986                        pages, vmas, NULL, false,
0987                        gup_flags | FOLL_TOUCH);
0988 }
0989 EXPORT_SYMBOL(get_user_pages);
0990 
0991 /**
0992  * populate_vma_page_range() -  populate a range of pages in the vma.
0993  * @vma:   target vma
0994  * @start: start address
0995  * @end:   end address
0996  * @nonblocking:
0997  *
0998  * This takes care of mlocking the pages too if VM_LOCKED is set.
0999  *
1000  * return 0 on success, negative error code on error.
1001  *
1002  * vma->vm_mm->mmap_sem must be held.
1003  *
1004  * If @nonblocking is NULL, it may be held for read or write and will
1005  * be unperturbed.
1006  *
1007  * If @nonblocking is non-NULL, it must held for read only and may be
1008  * released.  If it's released, *@nonblocking will be set to 0.
1009  */
1010 long populate_vma_page_range(struct vm_area_struct *vma,
1011         unsigned long start, unsigned long end, int *nonblocking)
1012 {
1013     struct mm_struct *mm = vma->vm_mm;
1014     unsigned long nr_pages = (end - start) / PAGE_SIZE;
1015     int gup_flags;
1016 
1017     VM_BUG_ON(start & ~PAGE_MASK);
1018     VM_BUG_ON(end   & ~PAGE_MASK);
1019     VM_BUG_ON_VMA(start < vma->vm_start, vma);
1020     VM_BUG_ON_VMA(end   > vma->vm_end, vma);
1021     VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
1022 
1023     gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
1024     if (vma->vm_flags & VM_LOCKONFAULT)
1025         gup_flags &= ~FOLL_POPULATE;
1026     /*
1027      * We want to touch writable mappings with a write fault in order
1028      * to break COW, except for shared mappings because these don't COW
1029      * and we would not want to dirty them for nothing.
1030      */
1031     if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
1032         gup_flags |= FOLL_WRITE;
1033 
1034     /*
1035      * We want mlock to succeed for regions that have any permissions
1036      * other than PROT_NONE.
1037      */
1038     if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
1039         gup_flags |= FOLL_FORCE;
1040 
1041     /*
1042      * We made sure addr is within a VMA, so the following will
1043      * not result in a stack expansion that recurses back here.
1044      */
1045     return __get_user_pages(current, mm, start, nr_pages, gup_flags,
1046                 NULL, NULL, nonblocking);
1047 }
1048 
1049 /*
1050  * __mm_populate - populate and/or mlock pages within a range of address space.
1051  *
1052  * This is used to implement mlock() and the MAP_POPULATE / MAP_LOCKED mmap
1053  * flags. VMAs must be already marked with the desired vm_flags, and
1054  * mmap_sem must not be held.
1055  */
1056 int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
1057 {
1058     struct mm_struct *mm = current->mm;
1059     unsigned long end, nstart, nend;
1060     struct vm_area_struct *vma = NULL;
1061     int locked = 0;
1062     long ret = 0;
1063 
1064     VM_BUG_ON(start & ~PAGE_MASK);
1065     VM_BUG_ON(len != PAGE_ALIGN(len));
1066     end = start + len;
1067 
1068     for (nstart = start; nstart < end; nstart = nend) {
1069         /*
1070          * We want to fault in pages for [nstart; end) address range.
1071          * Find first corresponding VMA.
1072          */
1073         if (!locked) {
1074             locked = 1;
1075             down_read(&mm->mmap_sem);
1076             vma = find_vma(mm, nstart);
1077         } else if (nstart >= vma->vm_end)
1078             vma = vma->vm_next;
1079         if (!vma || vma->vm_start >= end)
1080             break;
1081         /*
1082          * Set [nstart; nend) to intersection of desired address
1083          * range with the first VMA. Also, skip undesirable VMA types.
1084          */
1085         nend = min(end, vma->vm_end);
1086         if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1087             continue;
1088         if (nstart < vma->vm_start)
1089             nstart = vma->vm_start;
1090         /*
1091          * Now fault in a range of pages. populate_vma_page_range()
1092          * double checks the vma flags, so that it won't mlock pages
1093          * if the vma was already munlocked.
1094          */
1095         ret = populate_vma_page_range(vma, nstart, nend, &locked);
1096         if (ret < 0) {
1097             if (ignore_errors) {
1098                 ret = 0;
1099                 continue;   /* continue at next VMA */
1100             }
1101             break;
1102         }
1103         nend = nstart + ret * PAGE_SIZE;
1104         ret = 0;
1105     }
1106     if (locked)
1107         up_read(&mm->mmap_sem);
1108     return ret; /* 0 or negative error code */
1109 }
1110 
1111 /**
1112  * get_dump_page() - pin user page in memory while writing it to core dump
1113  * @addr: user address
1114  *
1115  * Returns struct page pointer of user page pinned for dump,
1116  * to be freed afterwards by put_page().
1117  *
1118  * Returns NULL on any kind of failure - a hole must then be inserted into
1119  * the corefile, to preserve alignment with its headers; and also returns
1120  * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1121  * allowing a hole to be left in the corefile to save diskspace.
1122  *
1123  * Called without mmap_sem, but after all other threads have been killed.
1124  */
1125 #ifdef CONFIG_ELF_CORE
1126 struct page *get_dump_page(unsigned long addr)
1127 {
1128     struct vm_area_struct *vma;
1129     struct page *page;
1130 
1131     if (__get_user_pages(current, current->mm, addr, 1,
1132                  FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1133                  NULL) < 1)
1134         return NULL;
1135     flush_cache_page(vma, addr, page_to_pfn(page));
1136     return page;
1137 }
1138 #endif /* CONFIG_ELF_CORE */
1139 
1140 /*
1141  * Generic RCU Fast GUP
1142  *
1143  * get_user_pages_fast attempts to pin user pages by walking the page
1144  * tables directly and avoids taking locks. Thus the walker needs to be
1145  * protected from page table pages being freed from under it, and should
1146  * block any THP splits.
1147  *
1148  * One way to achieve this is to have the walker disable interrupts, and
1149  * rely on IPIs from the TLB flushing code blocking before the page table
1150  * pages are freed. This is unsuitable for architectures that do not need
1151  * to broadcast an IPI when invalidating TLBs.
1152  *
1153  * Another way to achieve this is to batch up page table containing pages
1154  * belonging to more than one mm_user, then rcu_sched a callback to free those
1155  * pages. Disabling interrupts will allow the fast_gup walker to both block
1156  * the rcu_sched callback, and an IPI that we broadcast for splitting THPs
1157  * (which is a relatively rare event). The code below adopts this strategy.
1158  *
1159  * Before activating this code, please be aware that the following assumptions
1160  * are currently made:
1161  *
1162  *  *) HAVE_RCU_TABLE_FREE is enabled, and tlb_remove_table is used to free
1163  *      pages containing page tables.
1164  *
1165  *  *) ptes can be read atomically by the architecture.
1166  *
1167  *  *) access_ok is sufficient to validate userspace address ranges.
1168  *
1169  * The last two assumptions can be relaxed by the addition of helper functions.
1170  *
1171  * This code is based heavily on the PowerPC implementation by Nick Piggin.
1172  */
1173 #ifdef CONFIG_HAVE_GENERIC_RCU_GUP
1174 
1175 #ifdef __HAVE_ARCH_PTE_SPECIAL
1176 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1177              int write, struct page **pages, int *nr)
1178 {
1179     pte_t *ptep, *ptem;
1180     int ret = 0;
1181 
1182     ptem = ptep = pte_offset_map(&pmd, addr);
1183     do {
1184         /*
1185          * In the line below we are assuming that the pte can be read
1186          * atomically. If this is not the case for your architecture,
1187          * please wrap this in a helper function!
1188          *
1189          * for an example see gup_get_pte in arch/x86/mm/gup.c
1190          */
1191         pte_t pte = READ_ONCE(*ptep);
1192         struct page *head, *page;
1193 
1194         /*
1195          * Similar to the PMD case below, NUMA hinting must take slow
1196          * path using the pte_protnone check.
1197          */
1198         if (!pte_present(pte) || pte_special(pte) ||
1199             pte_protnone(pte) || (write && !pte_write(pte)))
1200             goto pte_unmap;
1201 
1202         if (!arch_pte_access_permitted(pte, write))
1203             goto pte_unmap;
1204 
1205         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1206         page = pte_page(pte);
1207         head = compound_head(page);
1208 
1209         if (!page_cache_get_speculative(head))
1210             goto pte_unmap;
1211 
1212         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1213             put_page(head);
1214             goto pte_unmap;
1215         }
1216 
1217         VM_BUG_ON_PAGE(compound_head(page) != head, page);
1218         pages[*nr] = page;
1219         (*nr)++;
1220 
1221     } while (ptep++, addr += PAGE_SIZE, addr != end);
1222 
1223     ret = 1;
1224 
1225 pte_unmap:
1226     pte_unmap(ptem);
1227     return ret;
1228 }
1229 #else
1230 
1231 /*
1232  * If we can't determine whether or not a pte is special, then fail immediately
1233  * for ptes. Note, we can still pin HugeTLB and THP as these are guaranteed not
1234  * to be special.
1235  *
1236  * For a futex to be placed on a THP tail page, get_futex_key requires a
1237  * __get_user_pages_fast implementation that can pin pages. Thus it's still
1238  * useful to have gup_huge_pmd even if we can't operate on ptes.
1239  */
1240 static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
1241              int write, struct page **pages, int *nr)
1242 {
1243     return 0;
1244 }
1245 #endif /* __HAVE_ARCH_PTE_SPECIAL */
1246 
1247 static int gup_huge_pmd(pmd_t orig, pmd_t *pmdp, unsigned long addr,
1248         unsigned long end, int write, struct page **pages, int *nr)
1249 {
1250     struct page *head, *page;
1251     int refs;
1252 
1253     if (write && !pmd_write(orig))
1254         return 0;
1255 
1256     refs = 0;
1257     head = pmd_page(orig);
1258     page = head + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
1259     do {
1260         VM_BUG_ON_PAGE(compound_head(page) != head, page);
1261         pages[*nr] = page;
1262         (*nr)++;
1263         page++;
1264         refs++;
1265     } while (addr += PAGE_SIZE, addr != end);
1266 
1267     if (!page_cache_add_speculative(head, refs)) {
1268         *nr -= refs;
1269         return 0;
1270     }
1271 
1272     if (unlikely(pmd_val(orig) != pmd_val(*pmdp))) {
1273         *nr -= refs;
1274         while (refs--)
1275             put_page(head);
1276         return 0;
1277     }
1278 
1279     return 1;
1280 }
1281 
1282 static int gup_huge_pud(pud_t orig, pud_t *pudp, unsigned long addr,
1283         unsigned long end, int write, struct page **pages, int *nr)
1284 {
1285     struct page *head, *page;
1286     int refs;
1287 
1288     if (write && !pud_write(orig))
1289         return 0;
1290 
1291     refs = 0;
1292     head = pud_page(orig);
1293     page = head + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
1294     do {
1295         VM_BUG_ON_PAGE(compound_head(page) != head, page);
1296         pages[*nr] = page;
1297         (*nr)++;
1298         page++;
1299         refs++;
1300     } while (addr += PAGE_SIZE, addr != end);
1301 
1302     if (!page_cache_add_speculative(head, refs)) {
1303         *nr -= refs;
1304         return 0;
1305     }
1306 
1307     if (unlikely(pud_val(orig) != pud_val(*pudp))) {
1308         *nr -= refs;
1309         while (refs--)
1310             put_page(head);
1311         return 0;
1312     }
1313 
1314     return 1;
1315 }
1316 
1317 static int gup_huge_pgd(pgd_t orig, pgd_t *pgdp, unsigned long addr,
1318             unsigned long end, int write,
1319             struct page **pages, int *nr)
1320 {
1321     int refs;
1322     struct page *head, *page;
1323 
1324     if (write && !pgd_write(orig))
1325         return 0;
1326 
1327     refs = 0;
1328     head = pgd_page(orig);
1329     page = head + ((addr & ~PGDIR_MASK) >> PAGE_SHIFT);
1330     do {
1331         VM_BUG_ON_PAGE(compound_head(page) != head, page);
1332         pages[*nr] = page;
1333         (*nr)++;
1334         page++;
1335         refs++;
1336     } while (addr += PAGE_SIZE, addr != end);
1337 
1338     if (!page_cache_add_speculative(head, refs)) {
1339         *nr -= refs;
1340         return 0;
1341     }
1342 
1343     if (unlikely(pgd_val(orig) != pgd_val(*pgdp))) {
1344         *nr -= refs;
1345         while (refs--)
1346             put_page(head);
1347         return 0;
1348     }
1349 
1350     return 1;
1351 }
1352 
1353 static int gup_pmd_range(pud_t pud, unsigned long addr, unsigned long end,
1354         int write, struct page **pages, int *nr)
1355 {
1356     unsigned long next;
1357     pmd_t *pmdp;
1358 
1359     pmdp = pmd_offset(&pud, addr);
1360     do {
1361         pmd_t pmd = READ_ONCE(*pmdp);
1362 
1363         next = pmd_addr_end(addr, end);
1364         if (pmd_none(pmd))
1365             return 0;
1366 
1367         if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd))) {
1368             /*
1369              * NUMA hinting faults need to be handled in the GUP
1370              * slowpath for accounting purposes and so that they
1371              * can be serialised against THP migration.
1372              */
1373             if (pmd_protnone(pmd))
1374                 return 0;
1375 
1376             if (!gup_huge_pmd(pmd, pmdp, addr, next, write,
1377                 pages, nr))
1378                 return 0;
1379 
1380         } else if (unlikely(is_hugepd(__hugepd(pmd_val(pmd))))) {
1381             /*
1382              * architecture have different format for hugetlbfs
1383              * pmd format and THP pmd format
1384              */
1385             if (!gup_huge_pd(__hugepd(pmd_val(pmd)), addr,
1386                      PMD_SHIFT, next, write, pages, nr))
1387                 return 0;
1388         } else if (!gup_pte_range(pmd, addr, next, write, pages, nr))
1389                 return 0;
1390     } while (pmdp++, addr = next, addr != end);
1391 
1392     return 1;
1393 }
1394 
1395 static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
1396              int write, struct page **pages, int *nr)
1397 {
1398     unsigned long next;
1399     pud_t *pudp;
1400 
1401     pudp = pud_offset(&pgd, addr);
1402     do {
1403         pud_t pud = READ_ONCE(*pudp);
1404 
1405         next = pud_addr_end(addr, end);
1406         if (pud_none(pud))
1407             return 0;
1408         if (unlikely(pud_huge(pud))) {
1409             if (!gup_huge_pud(pud, pudp, addr, next, write,
1410                       pages, nr))
1411                 return 0;
1412         } else if (unlikely(is_hugepd(__hugepd(pud_val(pud))))) {
1413             if (!gup_huge_pd(__hugepd(pud_val(pud)), addr,
1414                      PUD_SHIFT, next, write, pages, nr))
1415                 return 0;
1416         } else if (!gup_pmd_range(pud, addr, next, write, pages, nr))
1417             return 0;
1418     } while (pudp++, addr = next, addr != end);
1419 
1420     return 1;
1421 }
1422 
1423 /*
1424  * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
1425  * the regular GUP. It will only return non-negative values.
1426  */
1427 int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
1428               struct page **pages)
1429 {
1430     struct mm_struct *mm = current->mm;
1431     unsigned long addr, len, end;
1432     unsigned long next, flags;
1433     pgd_t *pgdp;
1434     int nr = 0;
1435 
1436     start &= PAGE_MASK;
1437     addr = start;
1438     len = (unsigned long) nr_pages << PAGE_SHIFT;
1439     end = start + len;
1440 
1441     if (unlikely(!access_ok(write ? VERIFY_WRITE : VERIFY_READ,
1442                     start, len)))
1443         return 0;
1444 
1445     /*
1446      * Disable interrupts.  We use the nested form as we can already have
1447      * interrupts disabled by get_futex_key.
1448      *
1449      * With interrupts disabled, we block page table pages from being
1450      * freed from under us. See mmu_gather_tlb in asm-generic/tlb.h
1451      * for more details.
1452      *
1453      * We do not adopt an rcu_read_lock(.) here as we also want to
1454      * block IPIs that come from THPs splitting.
1455      */
1456 
1457     local_irq_save(flags);
1458     pgdp = pgd_offset(mm, addr);
1459     do {
1460         pgd_t pgd = READ_ONCE(*pgdp);
1461 
1462         next = pgd_addr_end(addr, end);
1463         if (pgd_none(pgd))
1464             break;
1465         if (unlikely(pgd_huge(pgd))) {
1466             if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
1467                       pages, &nr))
1468                 break;
1469         } else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
1470             if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
1471                      PGDIR_SHIFT, next, write, pages, &nr))
1472                 break;
1473         } else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
1474             break;
1475     } while (pgdp++, addr = next, addr != end);
1476     local_irq_restore(flags);
1477 
1478     return nr;
1479 }
1480 
1481 /**
1482  * get_user_pages_fast() - pin user pages in memory
1483  * @start:  starting user address
1484  * @nr_pages:   number of pages from start to pin
1485  * @write:  whether pages will be written to
1486  * @pages:  array that receives pointers to the pages pinned.
1487  *      Should be at least nr_pages long.
1488  *
1489  * Attempt to pin user pages in memory without taking mm->mmap_sem.
1490  * If not successful, it will fall back to taking the lock and
1491  * calling get_user_pages().
1492  *
1493  * Returns number of pages pinned. This may be fewer than the number
1494  * requested. If nr_pages is 0 or negative, returns 0. If no pages
1495  * were pinned, returns -errno.
1496  */
1497 int get_user_pages_fast(unsigned long start, int nr_pages, int write,
1498             struct page **pages)
1499 {
1500     int nr, ret;
1501 
1502     start &= PAGE_MASK;
1503     nr = __get_user_pages_fast(start, nr_pages, write, pages);
1504     ret = nr;
1505 
1506     if (nr < nr_pages) {
1507         /* Try to get the remaining pages with get_user_pages */
1508         start += nr << PAGE_SHIFT;
1509         pages += nr;
1510 
1511         ret = get_user_pages_unlocked(start, nr_pages - nr, pages,
1512                 write ? FOLL_WRITE : 0);
1513 
1514         /* Have to be a bit careful with return values */
1515         if (nr > 0) {
1516             if (ret < 0)
1517                 ret = nr;
1518             else
1519                 ret += nr;
1520         }
1521     }
1522 
1523     return ret;
1524 }
1525 
1526 #endif /* CONFIG_HAVE_GENERIC_RCU_GUP */