Back to home page

LXR

 
 

    


0001 /*
0002  * mm/mmap.c
0003  *
0004  * Written by obz.
0005  *
0006  * Address space accounting code    <alan@lxorguk.ukuu.org.uk>
0007  */
0008 
0009 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0010 
0011 #include <linux/kernel.h>
0012 #include <linux/slab.h>
0013 #include <linux/backing-dev.h>
0014 #include <linux/mm.h>
0015 #include <linux/vmacache.h>
0016 #include <linux/shm.h>
0017 #include <linux/mman.h>
0018 #include <linux/pagemap.h>
0019 #include <linux/swap.h>
0020 #include <linux/syscalls.h>
0021 #include <linux/capability.h>
0022 #include <linux/init.h>
0023 #include <linux/file.h>
0024 #include <linux/fs.h>
0025 #include <linux/personality.h>
0026 #include <linux/security.h>
0027 #include <linux/hugetlb.h>
0028 #include <linux/shmem_fs.h>
0029 #include <linux/profile.h>
0030 #include <linux/export.h>
0031 #include <linux/mount.h>
0032 #include <linux/mempolicy.h>
0033 #include <linux/rmap.h>
0034 #include <linux/mmu_notifier.h>
0035 #include <linux/mmdebug.h>
0036 #include <linux/perf_event.h>
0037 #include <linux/audit.h>
0038 #include <linux/khugepaged.h>
0039 #include <linux/uprobes.h>
0040 #include <linux/rbtree_augmented.h>
0041 #include <linux/notifier.h>
0042 #include <linux/memory.h>
0043 #include <linux/printk.h>
0044 #include <linux/userfaultfd_k.h>
0045 #include <linux/moduleparam.h>
0046 #include <linux/pkeys.h>
0047 
0048 #include <linux/uaccess.h>
0049 #include <asm/cacheflush.h>
0050 #include <asm/tlb.h>
0051 #include <asm/mmu_context.h>
0052 
0053 #include "internal.h"
0054 
0055 #ifndef arch_mmap_check
0056 #define arch_mmap_check(addr, len, flags)   (0)
0057 #endif
0058 
0059 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
0060 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
0061 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
0062 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
0063 #endif
0064 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
0065 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
0066 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
0067 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
0068 #endif
0069 
0070 static bool ignore_rlimit_data;
0071 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
0072 
0073 static void unmap_region(struct mm_struct *mm,
0074         struct vm_area_struct *vma, struct vm_area_struct *prev,
0075         unsigned long start, unsigned long end);
0076 
0077 /* description of effects of mapping type and prot in current implementation.
0078  * this is due to the limited x86 page protection hardware.  The expected
0079  * behavior is in parens:
0080  *
0081  * map_type prot
0082  *      PROT_NONE   PROT_READ   PROT_WRITE  PROT_EXEC
0083  * MAP_SHARED   r: (no) no  r: (yes) yes    r: (no) yes r: (no) yes
0084  *      w: (no) no  w: (no) no  w: (yes) yes    w: (no) no
0085  *      x: (no) no  x: (no) yes x: (no) yes x: (yes) yes
0086  *
0087  * MAP_PRIVATE  r: (no) no  r: (yes) yes    r: (no) yes r: (no) yes
0088  *      w: (no) no  w: (no) no  w: (copy) copy  w: (no) no
0089  *      x: (no) no  x: (no) yes x: (no) yes x: (yes) yes
0090  *
0091  * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
0092  * MAP_PRIVATE:
0093  *                              r: (no) no
0094  *                              w: (no) no
0095  *                              x: (yes) yes
0096  */
0097 pgprot_t protection_map[16] = {
0098     __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
0099     __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
0100 };
0101 
0102 pgprot_t vm_get_page_prot(unsigned long vm_flags)
0103 {
0104     return __pgprot(pgprot_val(protection_map[vm_flags &
0105                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
0106             pgprot_val(arch_vm_get_page_prot(vm_flags)));
0107 }
0108 EXPORT_SYMBOL(vm_get_page_prot);
0109 
0110 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
0111 {
0112     return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
0113 }
0114 
0115 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
0116 void vma_set_page_prot(struct vm_area_struct *vma)
0117 {
0118     unsigned long vm_flags = vma->vm_flags;
0119     pgprot_t vm_page_prot;
0120 
0121     vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
0122     if (vma_wants_writenotify(vma, vm_page_prot)) {
0123         vm_flags &= ~VM_SHARED;
0124         vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
0125     }
0126     /* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
0127     WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
0128 }
0129 
0130 /*
0131  * Requires inode->i_mapping->i_mmap_rwsem
0132  */
0133 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
0134         struct file *file, struct address_space *mapping)
0135 {
0136     if (vma->vm_flags & VM_DENYWRITE)
0137         atomic_inc(&file_inode(file)->i_writecount);
0138     if (vma->vm_flags & VM_SHARED)
0139         mapping_unmap_writable(mapping);
0140 
0141     flush_dcache_mmap_lock(mapping);
0142     vma_interval_tree_remove(vma, &mapping->i_mmap);
0143     flush_dcache_mmap_unlock(mapping);
0144 }
0145 
0146 /*
0147  * Unlink a file-based vm structure from its interval tree, to hide
0148  * vma from rmap and vmtruncate before freeing its page tables.
0149  */
0150 void unlink_file_vma(struct vm_area_struct *vma)
0151 {
0152     struct file *file = vma->vm_file;
0153 
0154     if (file) {
0155         struct address_space *mapping = file->f_mapping;
0156         i_mmap_lock_write(mapping);
0157         __remove_shared_vm_struct(vma, file, mapping);
0158         i_mmap_unlock_write(mapping);
0159     }
0160 }
0161 
0162 /*
0163  * Close a vm structure and free it, returning the next.
0164  */
0165 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
0166 {
0167     struct vm_area_struct *next = vma->vm_next;
0168 
0169     might_sleep();
0170     if (vma->vm_ops && vma->vm_ops->close)
0171         vma->vm_ops->close(vma);
0172     if (vma->vm_file)
0173         fput(vma->vm_file);
0174     mpol_put(vma_policy(vma));
0175     kmem_cache_free(vm_area_cachep, vma);
0176     return next;
0177 }
0178 
0179 static int do_brk(unsigned long addr, unsigned long len);
0180 
0181 SYSCALL_DEFINE1(brk, unsigned long, brk)
0182 {
0183     unsigned long retval;
0184     unsigned long newbrk, oldbrk;
0185     struct mm_struct *mm = current->mm;
0186     unsigned long min_brk;
0187     bool populate;
0188 
0189     if (down_write_killable(&mm->mmap_sem))
0190         return -EINTR;
0191 
0192 #ifdef CONFIG_COMPAT_BRK
0193     /*
0194      * CONFIG_COMPAT_BRK can still be overridden by setting
0195      * randomize_va_space to 2, which will still cause mm->start_brk
0196      * to be arbitrarily shifted
0197      */
0198     if (current->brk_randomized)
0199         min_brk = mm->start_brk;
0200     else
0201         min_brk = mm->end_data;
0202 #else
0203     min_brk = mm->start_brk;
0204 #endif
0205     if (brk < min_brk)
0206         goto out;
0207 
0208     /*
0209      * Check against rlimit here. If this check is done later after the test
0210      * of oldbrk with newbrk then it can escape the test and let the data
0211      * segment grow beyond its set limit the in case where the limit is
0212      * not page aligned -Ram Gupta
0213      */
0214     if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
0215                   mm->end_data, mm->start_data))
0216         goto out;
0217 
0218     newbrk = PAGE_ALIGN(brk);
0219     oldbrk = PAGE_ALIGN(mm->brk);
0220     if (oldbrk == newbrk)
0221         goto set_brk;
0222 
0223     /* Always allow shrinking brk. */
0224     if (brk <= mm->brk) {
0225         if (!do_munmap(mm, newbrk, oldbrk-newbrk))
0226             goto set_brk;
0227         goto out;
0228     }
0229 
0230     /* Check against existing mmap mappings. */
0231     if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
0232         goto out;
0233 
0234     /* Ok, looks good - let it rip. */
0235     if (do_brk(oldbrk, newbrk-oldbrk) < 0)
0236         goto out;
0237 
0238 set_brk:
0239     mm->brk = brk;
0240     populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
0241     up_write(&mm->mmap_sem);
0242     if (populate)
0243         mm_populate(oldbrk, newbrk - oldbrk);
0244     return brk;
0245 
0246 out:
0247     retval = mm->brk;
0248     up_write(&mm->mmap_sem);
0249     return retval;
0250 }
0251 
0252 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
0253 {
0254     unsigned long max, subtree_gap;
0255     max = vma->vm_start;
0256     if (vma->vm_prev)
0257         max -= vma->vm_prev->vm_end;
0258     if (vma->vm_rb.rb_left) {
0259         subtree_gap = rb_entry(vma->vm_rb.rb_left,
0260                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
0261         if (subtree_gap > max)
0262             max = subtree_gap;
0263     }
0264     if (vma->vm_rb.rb_right) {
0265         subtree_gap = rb_entry(vma->vm_rb.rb_right,
0266                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
0267         if (subtree_gap > max)
0268             max = subtree_gap;
0269     }
0270     return max;
0271 }
0272 
0273 #ifdef CONFIG_DEBUG_VM_RB
0274 static int browse_rb(struct mm_struct *mm)
0275 {
0276     struct rb_root *root = &mm->mm_rb;
0277     int i = 0, j, bug = 0;
0278     struct rb_node *nd, *pn = NULL;
0279     unsigned long prev = 0, pend = 0;
0280 
0281     for (nd = rb_first(root); nd; nd = rb_next(nd)) {
0282         struct vm_area_struct *vma;
0283         vma = rb_entry(nd, struct vm_area_struct, vm_rb);
0284         if (vma->vm_start < prev) {
0285             pr_emerg("vm_start %lx < prev %lx\n",
0286                   vma->vm_start, prev);
0287             bug = 1;
0288         }
0289         if (vma->vm_start < pend) {
0290             pr_emerg("vm_start %lx < pend %lx\n",
0291                   vma->vm_start, pend);
0292             bug = 1;
0293         }
0294         if (vma->vm_start > vma->vm_end) {
0295             pr_emerg("vm_start %lx > vm_end %lx\n",
0296                   vma->vm_start, vma->vm_end);
0297             bug = 1;
0298         }
0299         spin_lock(&mm->page_table_lock);
0300         if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
0301             pr_emerg("free gap %lx, correct %lx\n",
0302                    vma->rb_subtree_gap,
0303                    vma_compute_subtree_gap(vma));
0304             bug = 1;
0305         }
0306         spin_unlock(&mm->page_table_lock);
0307         i++;
0308         pn = nd;
0309         prev = vma->vm_start;
0310         pend = vma->vm_end;
0311     }
0312     j = 0;
0313     for (nd = pn; nd; nd = rb_prev(nd))
0314         j++;
0315     if (i != j) {
0316         pr_emerg("backwards %d, forwards %d\n", j, i);
0317         bug = 1;
0318     }
0319     return bug ? -1 : i;
0320 }
0321 
0322 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
0323 {
0324     struct rb_node *nd;
0325 
0326     for (nd = rb_first(root); nd; nd = rb_next(nd)) {
0327         struct vm_area_struct *vma;
0328         vma = rb_entry(nd, struct vm_area_struct, vm_rb);
0329         VM_BUG_ON_VMA(vma != ignore &&
0330             vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
0331             vma);
0332     }
0333 }
0334 
0335 static void validate_mm(struct mm_struct *mm)
0336 {
0337     int bug = 0;
0338     int i = 0;
0339     unsigned long highest_address = 0;
0340     struct vm_area_struct *vma = mm->mmap;
0341 
0342     while (vma) {
0343         struct anon_vma *anon_vma = vma->anon_vma;
0344         struct anon_vma_chain *avc;
0345 
0346         if (anon_vma) {
0347             anon_vma_lock_read(anon_vma);
0348             list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
0349                 anon_vma_interval_tree_verify(avc);
0350             anon_vma_unlock_read(anon_vma);
0351         }
0352 
0353         highest_address = vma->vm_end;
0354         vma = vma->vm_next;
0355         i++;
0356     }
0357     if (i != mm->map_count) {
0358         pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
0359         bug = 1;
0360     }
0361     if (highest_address != mm->highest_vm_end) {
0362         pr_emerg("mm->highest_vm_end %lx, found %lx\n",
0363               mm->highest_vm_end, highest_address);
0364         bug = 1;
0365     }
0366     i = browse_rb(mm);
0367     if (i != mm->map_count) {
0368         if (i != -1)
0369             pr_emerg("map_count %d rb %d\n", mm->map_count, i);
0370         bug = 1;
0371     }
0372     VM_BUG_ON_MM(bug, mm);
0373 }
0374 #else
0375 #define validate_mm_rb(root, ignore) do { } while (0)
0376 #define validate_mm(mm) do { } while (0)
0377 #endif
0378 
0379 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
0380              unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
0381 
0382 /*
0383  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
0384  * vma->vm_prev->vm_end values changed, without modifying the vma's position
0385  * in the rbtree.
0386  */
0387 static void vma_gap_update(struct vm_area_struct *vma)
0388 {
0389     /*
0390      * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
0391      * function that does exacltly what we want.
0392      */
0393     vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
0394 }
0395 
0396 static inline void vma_rb_insert(struct vm_area_struct *vma,
0397                  struct rb_root *root)
0398 {
0399     /* All rb_subtree_gap values must be consistent prior to insertion */
0400     validate_mm_rb(root, NULL);
0401 
0402     rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
0403 }
0404 
0405 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
0406 {
0407     /*
0408      * Note rb_erase_augmented is a fairly large inline function,
0409      * so make sure we instantiate it only once with our desired
0410      * augmented rbtree callbacks.
0411      */
0412     rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
0413 }
0414 
0415 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
0416                         struct rb_root *root,
0417                         struct vm_area_struct *ignore)
0418 {
0419     /*
0420      * All rb_subtree_gap values must be consistent prior to erase,
0421      * with the possible exception of the "next" vma being erased if
0422      * next->vm_start was reduced.
0423      */
0424     validate_mm_rb(root, ignore);
0425 
0426     __vma_rb_erase(vma, root);
0427 }
0428 
0429 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
0430                      struct rb_root *root)
0431 {
0432     /*
0433      * All rb_subtree_gap values must be consistent prior to erase,
0434      * with the possible exception of the vma being erased.
0435      */
0436     validate_mm_rb(root, vma);
0437 
0438     __vma_rb_erase(vma, root);
0439 }
0440 
0441 /*
0442  * vma has some anon_vma assigned, and is already inserted on that
0443  * anon_vma's interval trees.
0444  *
0445  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
0446  * vma must be removed from the anon_vma's interval trees using
0447  * anon_vma_interval_tree_pre_update_vma().
0448  *
0449  * After the update, the vma will be reinserted using
0450  * anon_vma_interval_tree_post_update_vma().
0451  *
0452  * The entire update must be protected by exclusive mmap_sem and by
0453  * the root anon_vma's mutex.
0454  */
0455 static inline void
0456 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
0457 {
0458     struct anon_vma_chain *avc;
0459 
0460     list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
0461         anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
0462 }
0463 
0464 static inline void
0465 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
0466 {
0467     struct anon_vma_chain *avc;
0468 
0469     list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
0470         anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
0471 }
0472 
0473 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
0474         unsigned long end, struct vm_area_struct **pprev,
0475         struct rb_node ***rb_link, struct rb_node **rb_parent)
0476 {
0477     struct rb_node **__rb_link, *__rb_parent, *rb_prev;
0478 
0479     __rb_link = &mm->mm_rb.rb_node;
0480     rb_prev = __rb_parent = NULL;
0481 
0482     while (*__rb_link) {
0483         struct vm_area_struct *vma_tmp;
0484 
0485         __rb_parent = *__rb_link;
0486         vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
0487 
0488         if (vma_tmp->vm_end > addr) {
0489             /* Fail if an existing vma overlaps the area */
0490             if (vma_tmp->vm_start < end)
0491                 return -ENOMEM;
0492             __rb_link = &__rb_parent->rb_left;
0493         } else {
0494             rb_prev = __rb_parent;
0495             __rb_link = &__rb_parent->rb_right;
0496         }
0497     }
0498 
0499     *pprev = NULL;
0500     if (rb_prev)
0501         *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
0502     *rb_link = __rb_link;
0503     *rb_parent = __rb_parent;
0504     return 0;
0505 }
0506 
0507 static unsigned long count_vma_pages_range(struct mm_struct *mm,
0508         unsigned long addr, unsigned long end)
0509 {
0510     unsigned long nr_pages = 0;
0511     struct vm_area_struct *vma;
0512 
0513     /* Find first overlaping mapping */
0514     vma = find_vma_intersection(mm, addr, end);
0515     if (!vma)
0516         return 0;
0517 
0518     nr_pages = (min(end, vma->vm_end) -
0519         max(addr, vma->vm_start)) >> PAGE_SHIFT;
0520 
0521     /* Iterate over the rest of the overlaps */
0522     for (vma = vma->vm_next; vma; vma = vma->vm_next) {
0523         unsigned long overlap_len;
0524 
0525         if (vma->vm_start > end)
0526             break;
0527 
0528         overlap_len = min(end, vma->vm_end) - vma->vm_start;
0529         nr_pages += overlap_len >> PAGE_SHIFT;
0530     }
0531 
0532     return nr_pages;
0533 }
0534 
0535 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
0536         struct rb_node **rb_link, struct rb_node *rb_parent)
0537 {
0538     /* Update tracking information for the gap following the new vma. */
0539     if (vma->vm_next)
0540         vma_gap_update(vma->vm_next);
0541     else
0542         mm->highest_vm_end = vma->vm_end;
0543 
0544     /*
0545      * vma->vm_prev wasn't known when we followed the rbtree to find the
0546      * correct insertion point for that vma. As a result, we could not
0547      * update the vma vm_rb parents rb_subtree_gap values on the way down.
0548      * So, we first insert the vma with a zero rb_subtree_gap value
0549      * (to be consistent with what we did on the way down), and then
0550      * immediately update the gap to the correct value. Finally we
0551      * rebalance the rbtree after all augmented values have been set.
0552      */
0553     rb_link_node(&vma->vm_rb, rb_parent, rb_link);
0554     vma->rb_subtree_gap = 0;
0555     vma_gap_update(vma);
0556     vma_rb_insert(vma, &mm->mm_rb);
0557 }
0558 
0559 static void __vma_link_file(struct vm_area_struct *vma)
0560 {
0561     struct file *file;
0562 
0563     file = vma->vm_file;
0564     if (file) {
0565         struct address_space *mapping = file->f_mapping;
0566 
0567         if (vma->vm_flags & VM_DENYWRITE)
0568             atomic_dec(&file_inode(file)->i_writecount);
0569         if (vma->vm_flags & VM_SHARED)
0570             atomic_inc(&mapping->i_mmap_writable);
0571 
0572         flush_dcache_mmap_lock(mapping);
0573         vma_interval_tree_insert(vma, &mapping->i_mmap);
0574         flush_dcache_mmap_unlock(mapping);
0575     }
0576 }
0577 
0578 static void
0579 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
0580     struct vm_area_struct *prev, struct rb_node **rb_link,
0581     struct rb_node *rb_parent)
0582 {
0583     __vma_link_list(mm, vma, prev, rb_parent);
0584     __vma_link_rb(mm, vma, rb_link, rb_parent);
0585 }
0586 
0587 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
0588             struct vm_area_struct *prev, struct rb_node **rb_link,
0589             struct rb_node *rb_parent)
0590 {
0591     struct address_space *mapping = NULL;
0592 
0593     if (vma->vm_file) {
0594         mapping = vma->vm_file->f_mapping;
0595         i_mmap_lock_write(mapping);
0596     }
0597 
0598     __vma_link(mm, vma, prev, rb_link, rb_parent);
0599     __vma_link_file(vma);
0600 
0601     if (mapping)
0602         i_mmap_unlock_write(mapping);
0603 
0604     mm->map_count++;
0605     validate_mm(mm);
0606 }
0607 
0608 /*
0609  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
0610  * mm's list and rbtree.  It has already been inserted into the interval tree.
0611  */
0612 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
0613 {
0614     struct vm_area_struct *prev;
0615     struct rb_node **rb_link, *rb_parent;
0616 
0617     if (find_vma_links(mm, vma->vm_start, vma->vm_end,
0618                &prev, &rb_link, &rb_parent))
0619         BUG();
0620     __vma_link(mm, vma, prev, rb_link, rb_parent);
0621     mm->map_count++;
0622 }
0623 
0624 static __always_inline void __vma_unlink_common(struct mm_struct *mm,
0625                         struct vm_area_struct *vma,
0626                         struct vm_area_struct *prev,
0627                         bool has_prev,
0628                         struct vm_area_struct *ignore)
0629 {
0630     struct vm_area_struct *next;
0631 
0632     vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
0633     next = vma->vm_next;
0634     if (has_prev)
0635         prev->vm_next = next;
0636     else {
0637         prev = vma->vm_prev;
0638         if (prev)
0639             prev->vm_next = next;
0640         else
0641             mm->mmap = next;
0642     }
0643     if (next)
0644         next->vm_prev = prev;
0645 
0646     /* Kill the cache */
0647     vmacache_invalidate(mm);
0648 }
0649 
0650 static inline void __vma_unlink_prev(struct mm_struct *mm,
0651                      struct vm_area_struct *vma,
0652                      struct vm_area_struct *prev)
0653 {
0654     __vma_unlink_common(mm, vma, prev, true, vma);
0655 }
0656 
0657 /*
0658  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
0659  * is already present in an i_mmap tree without adjusting the tree.
0660  * The following helper function should be used when such adjustments
0661  * are necessary.  The "insert" vma (if any) is to be inserted
0662  * before we drop the necessary locks.
0663  */
0664 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
0665     unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
0666     struct vm_area_struct *expand)
0667 {
0668     struct mm_struct *mm = vma->vm_mm;
0669     struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
0670     struct address_space *mapping = NULL;
0671     struct rb_root *root = NULL;
0672     struct anon_vma *anon_vma = NULL;
0673     struct file *file = vma->vm_file;
0674     bool start_changed = false, end_changed = false;
0675     long adjust_next = 0;
0676     int remove_next = 0;
0677 
0678     if (next && !insert) {
0679         struct vm_area_struct *exporter = NULL, *importer = NULL;
0680 
0681         if (end >= next->vm_end) {
0682             /*
0683              * vma expands, overlapping all the next, and
0684              * perhaps the one after too (mprotect case 6).
0685              * The only other cases that gets here are
0686              * case 1, case 7 and case 8.
0687              */
0688             if (next == expand) {
0689                 /*
0690                  * The only case where we don't expand "vma"
0691                  * and we expand "next" instead is case 8.
0692                  */
0693                 VM_WARN_ON(end != next->vm_end);
0694                 /*
0695                  * remove_next == 3 means we're
0696                  * removing "vma" and that to do so we
0697                  * swapped "vma" and "next".
0698                  */
0699                 remove_next = 3;
0700                 VM_WARN_ON(file != next->vm_file);
0701                 swap(vma, next);
0702             } else {
0703                 VM_WARN_ON(expand != vma);
0704                 /*
0705                  * case 1, 6, 7, remove_next == 2 is case 6,
0706                  * remove_next == 1 is case 1 or 7.
0707                  */
0708                 remove_next = 1 + (end > next->vm_end);
0709                 VM_WARN_ON(remove_next == 2 &&
0710                        end != next->vm_next->vm_end);
0711                 VM_WARN_ON(remove_next == 1 &&
0712                        end != next->vm_end);
0713                 /* trim end to next, for case 6 first pass */
0714                 end = next->vm_end;
0715             }
0716 
0717             exporter = next;
0718             importer = vma;
0719 
0720             /*
0721              * If next doesn't have anon_vma, import from vma after
0722              * next, if the vma overlaps with it.
0723              */
0724             if (remove_next == 2 && !next->anon_vma)
0725                 exporter = next->vm_next;
0726 
0727         } else if (end > next->vm_start) {
0728             /*
0729              * vma expands, overlapping part of the next:
0730              * mprotect case 5 shifting the boundary up.
0731              */
0732             adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
0733             exporter = next;
0734             importer = vma;
0735             VM_WARN_ON(expand != importer);
0736         } else if (end < vma->vm_end) {
0737             /*
0738              * vma shrinks, and !insert tells it's not
0739              * split_vma inserting another: so it must be
0740              * mprotect case 4 shifting the boundary down.
0741              */
0742             adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
0743             exporter = vma;
0744             importer = next;
0745             VM_WARN_ON(expand != importer);
0746         }
0747 
0748         /*
0749          * Easily overlooked: when mprotect shifts the boundary,
0750          * make sure the expanding vma has anon_vma set if the
0751          * shrinking vma had, to cover any anon pages imported.
0752          */
0753         if (exporter && exporter->anon_vma && !importer->anon_vma) {
0754             int error;
0755 
0756             importer->anon_vma = exporter->anon_vma;
0757             error = anon_vma_clone(importer, exporter);
0758             if (error)
0759                 return error;
0760         }
0761     }
0762 again:
0763     vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
0764 
0765     if (file) {
0766         mapping = file->f_mapping;
0767         root = &mapping->i_mmap;
0768         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
0769 
0770         if (adjust_next)
0771             uprobe_munmap(next, next->vm_start, next->vm_end);
0772 
0773         i_mmap_lock_write(mapping);
0774         if (insert) {
0775             /*
0776              * Put into interval tree now, so instantiated pages
0777              * are visible to arm/parisc __flush_dcache_page
0778              * throughout; but we cannot insert into address
0779              * space until vma start or end is updated.
0780              */
0781             __vma_link_file(insert);
0782         }
0783     }
0784 
0785     anon_vma = vma->anon_vma;
0786     if (!anon_vma && adjust_next)
0787         anon_vma = next->anon_vma;
0788     if (anon_vma) {
0789         VM_WARN_ON(adjust_next && next->anon_vma &&
0790                anon_vma != next->anon_vma);
0791         anon_vma_lock_write(anon_vma);
0792         anon_vma_interval_tree_pre_update_vma(vma);
0793         if (adjust_next)
0794             anon_vma_interval_tree_pre_update_vma(next);
0795     }
0796 
0797     if (root) {
0798         flush_dcache_mmap_lock(mapping);
0799         vma_interval_tree_remove(vma, root);
0800         if (adjust_next)
0801             vma_interval_tree_remove(next, root);
0802     }
0803 
0804     if (start != vma->vm_start) {
0805         vma->vm_start = start;
0806         start_changed = true;
0807     }
0808     if (end != vma->vm_end) {
0809         vma->vm_end = end;
0810         end_changed = true;
0811     }
0812     vma->vm_pgoff = pgoff;
0813     if (adjust_next) {
0814         next->vm_start += adjust_next << PAGE_SHIFT;
0815         next->vm_pgoff += adjust_next;
0816     }
0817 
0818     if (root) {
0819         if (adjust_next)
0820             vma_interval_tree_insert(next, root);
0821         vma_interval_tree_insert(vma, root);
0822         flush_dcache_mmap_unlock(mapping);
0823     }
0824 
0825     if (remove_next) {
0826         /*
0827          * vma_merge has merged next into vma, and needs
0828          * us to remove next before dropping the locks.
0829          */
0830         if (remove_next != 3)
0831             __vma_unlink_prev(mm, next, vma);
0832         else
0833             /*
0834              * vma is not before next if they've been
0835              * swapped.
0836              *
0837              * pre-swap() next->vm_start was reduced so
0838              * tell validate_mm_rb to ignore pre-swap()
0839              * "next" (which is stored in post-swap()
0840              * "vma").
0841              */
0842             __vma_unlink_common(mm, next, NULL, false, vma);
0843         if (file)
0844             __remove_shared_vm_struct(next, file, mapping);
0845     } else if (insert) {
0846         /*
0847          * split_vma has split insert from vma, and needs
0848          * us to insert it before dropping the locks
0849          * (it may either follow vma or precede it).
0850          */
0851         __insert_vm_struct(mm, insert);
0852     } else {
0853         if (start_changed)
0854             vma_gap_update(vma);
0855         if (end_changed) {
0856             if (!next)
0857                 mm->highest_vm_end = end;
0858             else if (!adjust_next)
0859                 vma_gap_update(next);
0860         }
0861     }
0862 
0863     if (anon_vma) {
0864         anon_vma_interval_tree_post_update_vma(vma);
0865         if (adjust_next)
0866             anon_vma_interval_tree_post_update_vma(next);
0867         anon_vma_unlock_write(anon_vma);
0868     }
0869     if (mapping)
0870         i_mmap_unlock_write(mapping);
0871 
0872     if (root) {
0873         uprobe_mmap(vma);
0874 
0875         if (adjust_next)
0876             uprobe_mmap(next);
0877     }
0878 
0879     if (remove_next) {
0880         if (file) {
0881             uprobe_munmap(next, next->vm_start, next->vm_end);
0882             fput(file);
0883         }
0884         if (next->anon_vma)
0885             anon_vma_merge(vma, next);
0886         mm->map_count--;
0887         mpol_put(vma_policy(next));
0888         kmem_cache_free(vm_area_cachep, next);
0889         /*
0890          * In mprotect's case 6 (see comments on vma_merge),
0891          * we must remove another next too. It would clutter
0892          * up the code too much to do both in one go.
0893          */
0894         if (remove_next != 3) {
0895             /*
0896              * If "next" was removed and vma->vm_end was
0897              * expanded (up) over it, in turn
0898              * "next->vm_prev->vm_end" changed and the
0899              * "vma->vm_next" gap must be updated.
0900              */
0901             next = vma->vm_next;
0902         } else {
0903             /*
0904              * For the scope of the comment "next" and
0905              * "vma" considered pre-swap(): if "vma" was
0906              * removed, next->vm_start was expanded (down)
0907              * over it and the "next" gap must be updated.
0908              * Because of the swap() the post-swap() "vma"
0909              * actually points to pre-swap() "next"
0910              * (post-swap() "next" as opposed is now a
0911              * dangling pointer).
0912              */
0913             next = vma;
0914         }
0915         if (remove_next == 2) {
0916             remove_next = 1;
0917             end = next->vm_end;
0918             goto again;
0919         }
0920         else if (next)
0921             vma_gap_update(next);
0922         else {
0923             /*
0924              * If remove_next == 2 we obviously can't
0925              * reach this path.
0926              *
0927              * If remove_next == 3 we can't reach this
0928              * path because pre-swap() next is always not
0929              * NULL. pre-swap() "next" is not being
0930              * removed and its next->vm_end is not altered
0931              * (and furthermore "end" already matches
0932              * next->vm_end in remove_next == 3).
0933              *
0934              * We reach this only in the remove_next == 1
0935              * case if the "next" vma that was removed was
0936              * the highest vma of the mm. However in such
0937              * case next->vm_end == "end" and the extended
0938              * "vma" has vma->vm_end == next->vm_end so
0939              * mm->highest_vm_end doesn't need any update
0940              * in remove_next == 1 case.
0941              */
0942             VM_WARN_ON(mm->highest_vm_end != end);
0943         }
0944     }
0945     if (insert && file)
0946         uprobe_mmap(insert);
0947 
0948     validate_mm(mm);
0949 
0950     return 0;
0951 }
0952 
0953 /*
0954  * If the vma has a ->close operation then the driver probably needs to release
0955  * per-vma resources, so we don't attempt to merge those.
0956  */
0957 static inline int is_mergeable_vma(struct vm_area_struct *vma,
0958                 struct file *file, unsigned long vm_flags,
0959                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
0960 {
0961     /*
0962      * VM_SOFTDIRTY should not prevent from VMA merging, if we
0963      * match the flags but dirty bit -- the caller should mark
0964      * merged VMA as dirty. If dirty bit won't be excluded from
0965      * comparison, we increase pressue on the memory system forcing
0966      * the kernel to generate new VMAs when old one could be
0967      * extended instead.
0968      */
0969     if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
0970         return 0;
0971     if (vma->vm_file != file)
0972         return 0;
0973     if (vma->vm_ops && vma->vm_ops->close)
0974         return 0;
0975     if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
0976         return 0;
0977     return 1;
0978 }
0979 
0980 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
0981                     struct anon_vma *anon_vma2,
0982                     struct vm_area_struct *vma)
0983 {
0984     /*
0985      * The list_is_singular() test is to avoid merging VMA cloned from
0986      * parents. This can improve scalability caused by anon_vma lock.
0987      */
0988     if ((!anon_vma1 || !anon_vma2) && (!vma ||
0989         list_is_singular(&vma->anon_vma_chain)))
0990         return 1;
0991     return anon_vma1 == anon_vma2;
0992 }
0993 
0994 /*
0995  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
0996  * in front of (at a lower virtual address and file offset than) the vma.
0997  *
0998  * We cannot merge two vmas if they have differently assigned (non-NULL)
0999  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1000  *
1001  * We don't check here for the merged mmap wrapping around the end of pagecache
1002  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1003  * wrap, nor mmaps which cover the final page at index -1UL.
1004  */
1005 static int
1006 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1007              struct anon_vma *anon_vma, struct file *file,
1008              pgoff_t vm_pgoff,
1009              struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1010 {
1011     if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1012         is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1013         if (vma->vm_pgoff == vm_pgoff)
1014             return 1;
1015     }
1016     return 0;
1017 }
1018 
1019 /*
1020  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1021  * beyond (at a higher virtual address and file offset than) the vma.
1022  *
1023  * We cannot merge two vmas if they have differently assigned (non-NULL)
1024  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1025  */
1026 static int
1027 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1028             struct anon_vma *anon_vma, struct file *file,
1029             pgoff_t vm_pgoff,
1030             struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1031 {
1032     if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1033         is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1034         pgoff_t vm_pglen;
1035         vm_pglen = vma_pages(vma);
1036         if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1037             return 1;
1038     }
1039     return 0;
1040 }
1041 
1042 /*
1043  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1044  * whether that can be merged with its predecessor or its successor.
1045  * Or both (it neatly fills a hole).
1046  *
1047  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1048  * certain not to be mapped by the time vma_merge is called; but when
1049  * called for mprotect, it is certain to be already mapped (either at
1050  * an offset within prev, or at the start of next), and the flags of
1051  * this area are about to be changed to vm_flags - and the no-change
1052  * case has already been eliminated.
1053  *
1054  * The following mprotect cases have to be considered, where AAAA is
1055  * the area passed down from mprotect_fixup, never extending beyond one
1056  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1057  *
1058  *     AAAA             AAAA                AAAA          AAAA
1059  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1060  *    cannot merge    might become    might become    might become
1061  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1062  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1063  *    mremap move:                                    PPPPXXXXXXXX 8
1064  *        AAAA
1065  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1066  *    might become    case 1 below    case 2 below    case 3 below
1067  *
1068  * It is important for case 8 that the the vma NNNN overlapping the
1069  * region AAAA is never going to extended over XXXX. Instead XXXX must
1070  * be extended in region AAAA and NNNN must be removed. This way in
1071  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1072  * rmap_locks, the properties of the merged vma will be already
1073  * correct for the whole merged range. Some of those properties like
1074  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1075  * be correct for the whole merged range immediately after the
1076  * rmap_locks are released. Otherwise if XXXX would be removed and
1077  * NNNN would be extended over the XXXX range, remove_migration_ptes
1078  * or other rmap walkers (if working on addresses beyond the "end"
1079  * parameter) may establish ptes with the wrong permissions of NNNN
1080  * instead of the right permissions of XXXX.
1081  */
1082 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1083             struct vm_area_struct *prev, unsigned long addr,
1084             unsigned long end, unsigned long vm_flags,
1085             struct anon_vma *anon_vma, struct file *file,
1086             pgoff_t pgoff, struct mempolicy *policy,
1087             struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1088 {
1089     pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1090     struct vm_area_struct *area, *next;
1091     int err;
1092 
1093     /*
1094      * We later require that vma->vm_flags == vm_flags,
1095      * so this tests vma->vm_flags & VM_SPECIAL, too.
1096      */
1097     if (vm_flags & VM_SPECIAL)
1098         return NULL;
1099 
1100     if (prev)
1101         next = prev->vm_next;
1102     else
1103         next = mm->mmap;
1104     area = next;
1105     if (area && area->vm_end == end)        /* cases 6, 7, 8 */
1106         next = next->vm_next;
1107 
1108     /* verify some invariant that must be enforced by the caller */
1109     VM_WARN_ON(prev && addr <= prev->vm_start);
1110     VM_WARN_ON(area && end > area->vm_end);
1111     VM_WARN_ON(addr >= end);
1112 
1113     /*
1114      * Can it merge with the predecessor?
1115      */
1116     if (prev && prev->vm_end == addr &&
1117             mpol_equal(vma_policy(prev), policy) &&
1118             can_vma_merge_after(prev, vm_flags,
1119                         anon_vma, file, pgoff,
1120                         vm_userfaultfd_ctx)) {
1121         /*
1122          * OK, it can.  Can we now merge in the successor as well?
1123          */
1124         if (next && end == next->vm_start &&
1125                 mpol_equal(policy, vma_policy(next)) &&
1126                 can_vma_merge_before(next, vm_flags,
1127                              anon_vma, file,
1128                              pgoff+pglen,
1129                              vm_userfaultfd_ctx) &&
1130                 is_mergeable_anon_vma(prev->anon_vma,
1131                               next->anon_vma, NULL)) {
1132                             /* cases 1, 6 */
1133             err = __vma_adjust(prev, prev->vm_start,
1134                      next->vm_end, prev->vm_pgoff, NULL,
1135                      prev);
1136         } else                  /* cases 2, 5, 7 */
1137             err = __vma_adjust(prev, prev->vm_start,
1138                      end, prev->vm_pgoff, NULL, prev);
1139         if (err)
1140             return NULL;
1141         khugepaged_enter_vma_merge(prev, vm_flags);
1142         return prev;
1143     }
1144 
1145     /*
1146      * Can this new request be merged in front of next?
1147      */
1148     if (next && end == next->vm_start &&
1149             mpol_equal(policy, vma_policy(next)) &&
1150             can_vma_merge_before(next, vm_flags,
1151                          anon_vma, file, pgoff+pglen,
1152                          vm_userfaultfd_ctx)) {
1153         if (prev && addr < prev->vm_end)    /* case 4 */
1154             err = __vma_adjust(prev, prev->vm_start,
1155                      addr, prev->vm_pgoff, NULL, next);
1156         else {                  /* cases 3, 8 */
1157             err = __vma_adjust(area, addr, next->vm_end,
1158                      next->vm_pgoff - pglen, NULL, next);
1159             /*
1160              * In case 3 area is already equal to next and
1161              * this is a noop, but in case 8 "area" has
1162              * been removed and next was expanded over it.
1163              */
1164             area = next;
1165         }
1166         if (err)
1167             return NULL;
1168         khugepaged_enter_vma_merge(area, vm_flags);
1169         return area;
1170     }
1171 
1172     return NULL;
1173 }
1174 
1175 /*
1176  * Rough compatbility check to quickly see if it's even worth looking
1177  * at sharing an anon_vma.
1178  *
1179  * They need to have the same vm_file, and the flags can only differ
1180  * in things that mprotect may change.
1181  *
1182  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1183  * we can merge the two vma's. For example, we refuse to merge a vma if
1184  * there is a vm_ops->close() function, because that indicates that the
1185  * driver is doing some kind of reference counting. But that doesn't
1186  * really matter for the anon_vma sharing case.
1187  */
1188 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1189 {
1190     return a->vm_end == b->vm_start &&
1191         mpol_equal(vma_policy(a), vma_policy(b)) &&
1192         a->vm_file == b->vm_file &&
1193         !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1194         b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1195 }
1196 
1197 /*
1198  * Do some basic sanity checking to see if we can re-use the anon_vma
1199  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1200  * the same as 'old', the other will be the new one that is trying
1201  * to share the anon_vma.
1202  *
1203  * NOTE! This runs with mm_sem held for reading, so it is possible that
1204  * the anon_vma of 'old' is concurrently in the process of being set up
1205  * by another page fault trying to merge _that_. But that's ok: if it
1206  * is being set up, that automatically means that it will be a singleton
1207  * acceptable for merging, so we can do all of this optimistically. But
1208  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1209  *
1210  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1211  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1212  * is to return an anon_vma that is "complex" due to having gone through
1213  * a fork).
1214  *
1215  * We also make sure that the two vma's are compatible (adjacent,
1216  * and with the same memory policies). That's all stable, even with just
1217  * a read lock on the mm_sem.
1218  */
1219 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1220 {
1221     if (anon_vma_compatible(a, b)) {
1222         struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1223 
1224         if (anon_vma && list_is_singular(&old->anon_vma_chain))
1225             return anon_vma;
1226     }
1227     return NULL;
1228 }
1229 
1230 /*
1231  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1232  * neighbouring vmas for a suitable anon_vma, before it goes off
1233  * to allocate a new anon_vma.  It checks because a repetitive
1234  * sequence of mprotects and faults may otherwise lead to distinct
1235  * anon_vmas being allocated, preventing vma merge in subsequent
1236  * mprotect.
1237  */
1238 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1239 {
1240     struct anon_vma *anon_vma;
1241     struct vm_area_struct *near;
1242 
1243     near = vma->vm_next;
1244     if (!near)
1245         goto try_prev;
1246 
1247     anon_vma = reusable_anon_vma(near, vma, near);
1248     if (anon_vma)
1249         return anon_vma;
1250 try_prev:
1251     near = vma->vm_prev;
1252     if (!near)
1253         goto none;
1254 
1255     anon_vma = reusable_anon_vma(near, near, vma);
1256     if (anon_vma)
1257         return anon_vma;
1258 none:
1259     /*
1260      * There's no absolute need to look only at touching neighbours:
1261      * we could search further afield for "compatible" anon_vmas.
1262      * But it would probably just be a waste of time searching,
1263      * or lead to too many vmas hanging off the same anon_vma.
1264      * We're trying to allow mprotect remerging later on,
1265      * not trying to minimize memory used for anon_vmas.
1266      */
1267     return NULL;
1268 }
1269 
1270 /*
1271  * If a hint addr is less than mmap_min_addr change hint to be as
1272  * low as possible but still greater than mmap_min_addr
1273  */
1274 static inline unsigned long round_hint_to_min(unsigned long hint)
1275 {
1276     hint &= PAGE_MASK;
1277     if (((void *)hint != NULL) &&
1278         (hint < mmap_min_addr))
1279         return PAGE_ALIGN(mmap_min_addr);
1280     return hint;
1281 }
1282 
1283 static inline int mlock_future_check(struct mm_struct *mm,
1284                      unsigned long flags,
1285                      unsigned long len)
1286 {
1287     unsigned long locked, lock_limit;
1288 
1289     /*  mlock MCL_FUTURE? */
1290     if (flags & VM_LOCKED) {
1291         locked = len >> PAGE_SHIFT;
1292         locked += mm->locked_vm;
1293         lock_limit = rlimit(RLIMIT_MEMLOCK);
1294         lock_limit >>= PAGE_SHIFT;
1295         if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1296             return -EAGAIN;
1297     }
1298     return 0;
1299 }
1300 
1301 /*
1302  * The caller must hold down_write(&current->mm->mmap_sem).
1303  */
1304 unsigned long do_mmap(struct file *file, unsigned long addr,
1305             unsigned long len, unsigned long prot,
1306             unsigned long flags, vm_flags_t vm_flags,
1307             unsigned long pgoff, unsigned long *populate)
1308 {
1309     struct mm_struct *mm = current->mm;
1310     int pkey = 0;
1311 
1312     *populate = 0;
1313 
1314     if (!len)
1315         return -EINVAL;
1316 
1317     /*
1318      * Does the application expect PROT_READ to imply PROT_EXEC?
1319      *
1320      * (the exception is when the underlying filesystem is noexec
1321      *  mounted, in which case we dont add PROT_EXEC.)
1322      */
1323     if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1324         if (!(file && path_noexec(&file->f_path)))
1325             prot |= PROT_EXEC;
1326 
1327     if (!(flags & MAP_FIXED))
1328         addr = round_hint_to_min(addr);
1329 
1330     /* Careful about overflows.. */
1331     len = PAGE_ALIGN(len);
1332     if (!len)
1333         return -ENOMEM;
1334 
1335     /* offset overflow? */
1336     if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1337         return -EOVERFLOW;
1338 
1339     /* Too many mappings? */
1340     if (mm->map_count > sysctl_max_map_count)
1341         return -ENOMEM;
1342 
1343     /* Obtain the address to map to. we verify (or select) it and ensure
1344      * that it represents a valid section of the address space.
1345      */
1346     addr = get_unmapped_area(file, addr, len, pgoff, flags);
1347     if (offset_in_page(addr))
1348         return addr;
1349 
1350     if (prot == PROT_EXEC) {
1351         pkey = execute_only_pkey(mm);
1352         if (pkey < 0)
1353             pkey = 0;
1354     }
1355 
1356     /* Do simple checking here so the lower-level routines won't have
1357      * to. we assume access permissions have been handled by the open
1358      * of the memory object, so we don't do any here.
1359      */
1360     vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1361             mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1362 
1363     if (flags & MAP_LOCKED)
1364         if (!can_do_mlock())
1365             return -EPERM;
1366 
1367     if (mlock_future_check(mm, vm_flags, len))
1368         return -EAGAIN;
1369 
1370     if (file) {
1371         struct inode *inode = file_inode(file);
1372 
1373         switch (flags & MAP_TYPE) {
1374         case MAP_SHARED:
1375             if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1376                 return -EACCES;
1377 
1378             /*
1379              * Make sure we don't allow writing to an append-only
1380              * file..
1381              */
1382             if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1383                 return -EACCES;
1384 
1385             /*
1386              * Make sure there are no mandatory locks on the file.
1387              */
1388             if (locks_verify_locked(file))
1389                 return -EAGAIN;
1390 
1391             vm_flags |= VM_SHARED | VM_MAYSHARE;
1392             if (!(file->f_mode & FMODE_WRITE))
1393                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1394 
1395             /* fall through */
1396         case MAP_PRIVATE:
1397             if (!(file->f_mode & FMODE_READ))
1398                 return -EACCES;
1399             if (path_noexec(&file->f_path)) {
1400                 if (vm_flags & VM_EXEC)
1401                     return -EPERM;
1402                 vm_flags &= ~VM_MAYEXEC;
1403             }
1404 
1405             if (!file->f_op->mmap)
1406                 return -ENODEV;
1407             if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1408                 return -EINVAL;
1409             break;
1410 
1411         default:
1412             return -EINVAL;
1413         }
1414     } else {
1415         switch (flags & MAP_TYPE) {
1416         case MAP_SHARED:
1417             if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1418                 return -EINVAL;
1419             /*
1420              * Ignore pgoff.
1421              */
1422             pgoff = 0;
1423             vm_flags |= VM_SHARED | VM_MAYSHARE;
1424             break;
1425         case MAP_PRIVATE:
1426             /*
1427              * Set pgoff according to addr for anon_vma.
1428              */
1429             pgoff = addr >> PAGE_SHIFT;
1430             break;
1431         default:
1432             return -EINVAL;
1433         }
1434     }
1435 
1436     /*
1437      * Set 'VM_NORESERVE' if we should not account for the
1438      * memory use of this mapping.
1439      */
1440     if (flags & MAP_NORESERVE) {
1441         /* We honor MAP_NORESERVE if allowed to overcommit */
1442         if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1443             vm_flags |= VM_NORESERVE;
1444 
1445         /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1446         if (file && is_file_hugepages(file))
1447             vm_flags |= VM_NORESERVE;
1448     }
1449 
1450     addr = mmap_region(file, addr, len, vm_flags, pgoff);
1451     if (!IS_ERR_VALUE(addr) &&
1452         ((vm_flags & VM_LOCKED) ||
1453          (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1454         *populate = len;
1455     return addr;
1456 }
1457 
1458 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1459         unsigned long, prot, unsigned long, flags,
1460         unsigned long, fd, unsigned long, pgoff)
1461 {
1462     struct file *file = NULL;
1463     unsigned long retval;
1464 
1465     if (!(flags & MAP_ANONYMOUS)) {
1466         audit_mmap_fd(fd, flags);
1467         file = fget(fd);
1468         if (!file)
1469             return -EBADF;
1470         if (is_file_hugepages(file))
1471             len = ALIGN(len, huge_page_size(hstate_file(file)));
1472         retval = -EINVAL;
1473         if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1474             goto out_fput;
1475     } else if (flags & MAP_HUGETLB) {
1476         struct user_struct *user = NULL;
1477         struct hstate *hs;
1478 
1479         hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & SHM_HUGE_MASK);
1480         if (!hs)
1481             return -EINVAL;
1482 
1483         len = ALIGN(len, huge_page_size(hs));
1484         /*
1485          * VM_NORESERVE is used because the reservations will be
1486          * taken when vm_ops->mmap() is called
1487          * A dummy user value is used because we are not locking
1488          * memory so no accounting is necessary
1489          */
1490         file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1491                 VM_NORESERVE,
1492                 &user, HUGETLB_ANONHUGE_INODE,
1493                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1494         if (IS_ERR(file))
1495             return PTR_ERR(file);
1496     }
1497 
1498     flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1499 
1500     retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1501 out_fput:
1502     if (file)
1503         fput(file);
1504     return retval;
1505 }
1506 
1507 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1508 struct mmap_arg_struct {
1509     unsigned long addr;
1510     unsigned long len;
1511     unsigned long prot;
1512     unsigned long flags;
1513     unsigned long fd;
1514     unsigned long offset;
1515 };
1516 
1517 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1518 {
1519     struct mmap_arg_struct a;
1520 
1521     if (copy_from_user(&a, arg, sizeof(a)))
1522         return -EFAULT;
1523     if (offset_in_page(a.offset))
1524         return -EINVAL;
1525 
1526     return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1527                   a.offset >> PAGE_SHIFT);
1528 }
1529 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1530 
1531 /*
1532  * Some shared mappigns will want the pages marked read-only
1533  * to track write events. If so, we'll downgrade vm_page_prot
1534  * to the private version (using protection_map[] without the
1535  * VM_SHARED bit).
1536  */
1537 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1538 {
1539     vm_flags_t vm_flags = vma->vm_flags;
1540     const struct vm_operations_struct *vm_ops = vma->vm_ops;
1541 
1542     /* If it was private or non-writable, the write bit is already clear */
1543     if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1544         return 0;
1545 
1546     /* The backer wishes to know when pages are first written to? */
1547     if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1548         return 1;
1549 
1550     /* The open routine did something to the protections that pgprot_modify
1551      * won't preserve? */
1552     if (pgprot_val(vm_page_prot) !=
1553         pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1554         return 0;
1555 
1556     /* Do we need to track softdirty? */
1557     if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1558         return 1;
1559 
1560     /* Specialty mapping? */
1561     if (vm_flags & VM_PFNMAP)
1562         return 0;
1563 
1564     /* Can the mapping track the dirty pages? */
1565     return vma->vm_file && vma->vm_file->f_mapping &&
1566         mapping_cap_account_dirty(vma->vm_file->f_mapping);
1567 }
1568 
1569 /*
1570  * We account for memory if it's a private writeable mapping,
1571  * not hugepages and VM_NORESERVE wasn't set.
1572  */
1573 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1574 {
1575     /*
1576      * hugetlb has its own accounting separate from the core VM
1577      * VM_HUGETLB may not be set yet so we cannot check for that flag.
1578      */
1579     if (file && is_file_hugepages(file))
1580         return 0;
1581 
1582     return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1583 }
1584 
1585 unsigned long mmap_region(struct file *file, unsigned long addr,
1586         unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1587 {
1588     struct mm_struct *mm = current->mm;
1589     struct vm_area_struct *vma, *prev;
1590     int error;
1591     struct rb_node **rb_link, *rb_parent;
1592     unsigned long charged = 0;
1593 
1594     /* Check against address space limit. */
1595     if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1596         unsigned long nr_pages;
1597 
1598         /*
1599          * MAP_FIXED may remove pages of mappings that intersects with
1600          * requested mapping. Account for the pages it would unmap.
1601          */
1602         nr_pages = count_vma_pages_range(mm, addr, addr + len);
1603 
1604         if (!may_expand_vm(mm, vm_flags,
1605                     (len >> PAGE_SHIFT) - nr_pages))
1606             return -ENOMEM;
1607     }
1608 
1609     /* Clear old maps */
1610     while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1611                   &rb_parent)) {
1612         if (do_munmap(mm, addr, len))
1613             return -ENOMEM;
1614     }
1615 
1616     /*
1617      * Private writable mapping: check memory availability
1618      */
1619     if (accountable_mapping(file, vm_flags)) {
1620         charged = len >> PAGE_SHIFT;
1621         if (security_vm_enough_memory_mm(mm, charged))
1622             return -ENOMEM;
1623         vm_flags |= VM_ACCOUNT;
1624     }
1625 
1626     /*
1627      * Can we just expand an old mapping?
1628      */
1629     vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1630             NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1631     if (vma)
1632         goto out;
1633 
1634     /*
1635      * Determine the object being mapped and call the appropriate
1636      * specific mapper. the address has already been validated, but
1637      * not unmapped, but the maps are removed from the list.
1638      */
1639     vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1640     if (!vma) {
1641         error = -ENOMEM;
1642         goto unacct_error;
1643     }
1644 
1645     vma->vm_mm = mm;
1646     vma->vm_start = addr;
1647     vma->vm_end = addr + len;
1648     vma->vm_flags = vm_flags;
1649     vma->vm_page_prot = vm_get_page_prot(vm_flags);
1650     vma->vm_pgoff = pgoff;
1651     INIT_LIST_HEAD(&vma->anon_vma_chain);
1652 
1653     if (file) {
1654         if (vm_flags & VM_DENYWRITE) {
1655             error = deny_write_access(file);
1656             if (error)
1657                 goto free_vma;
1658         }
1659         if (vm_flags & VM_SHARED) {
1660             error = mapping_map_writable(file->f_mapping);
1661             if (error)
1662                 goto allow_write_and_free_vma;
1663         }
1664 
1665         /* ->mmap() can change vma->vm_file, but must guarantee that
1666          * vma_link() below can deny write-access if VM_DENYWRITE is set
1667          * and map writably if VM_SHARED is set. This usually means the
1668          * new file must not have been exposed to user-space, yet.
1669          */
1670         vma->vm_file = get_file(file);
1671         error = file->f_op->mmap(file, vma);
1672         if (error)
1673             goto unmap_and_free_vma;
1674 
1675         /* Can addr have changed??
1676          *
1677          * Answer: Yes, several device drivers can do it in their
1678          *         f_op->mmap method. -DaveM
1679          * Bug: If addr is changed, prev, rb_link, rb_parent should
1680          *      be updated for vma_link()
1681          */
1682         WARN_ON_ONCE(addr != vma->vm_start);
1683 
1684         addr = vma->vm_start;
1685         vm_flags = vma->vm_flags;
1686     } else if (vm_flags & VM_SHARED) {
1687         error = shmem_zero_setup(vma);
1688         if (error)
1689             goto free_vma;
1690     }
1691 
1692     vma_link(mm, vma, prev, rb_link, rb_parent);
1693     /* Once vma denies write, undo our temporary denial count */
1694     if (file) {
1695         if (vm_flags & VM_SHARED)
1696             mapping_unmap_writable(file->f_mapping);
1697         if (vm_flags & VM_DENYWRITE)
1698             allow_write_access(file);
1699     }
1700     file = vma->vm_file;
1701 out:
1702     perf_event_mmap(vma);
1703 
1704     vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1705     if (vm_flags & VM_LOCKED) {
1706         if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1707                     vma == get_gate_vma(current->mm)))
1708             mm->locked_vm += (len >> PAGE_SHIFT);
1709         else
1710             vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1711     }
1712 
1713     if (file)
1714         uprobe_mmap(vma);
1715 
1716     /*
1717      * New (or expanded) vma always get soft dirty status.
1718      * Otherwise user-space soft-dirty page tracker won't
1719      * be able to distinguish situation when vma area unmapped,
1720      * then new mapped in-place (which must be aimed as
1721      * a completely new data area).
1722      */
1723     vma->vm_flags |= VM_SOFTDIRTY;
1724 
1725     vma_set_page_prot(vma);
1726 
1727     return addr;
1728 
1729 unmap_and_free_vma:
1730     vma->vm_file = NULL;
1731     fput(file);
1732 
1733     /* Undo any partial mapping done by a device driver. */
1734     unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1735     charged = 0;
1736     if (vm_flags & VM_SHARED)
1737         mapping_unmap_writable(file->f_mapping);
1738 allow_write_and_free_vma:
1739     if (vm_flags & VM_DENYWRITE)
1740         allow_write_access(file);
1741 free_vma:
1742     kmem_cache_free(vm_area_cachep, vma);
1743 unacct_error:
1744     if (charged)
1745         vm_unacct_memory(charged);
1746     return error;
1747 }
1748 
1749 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1750 {
1751     /*
1752      * We implement the search by looking for an rbtree node that
1753      * immediately follows a suitable gap. That is,
1754      * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1755      * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1756      * - gap_end - gap_start >= length
1757      */
1758 
1759     struct mm_struct *mm = current->mm;
1760     struct vm_area_struct *vma;
1761     unsigned long length, low_limit, high_limit, gap_start, gap_end;
1762 
1763     /* Adjust search length to account for worst case alignment overhead */
1764     length = info->length + info->align_mask;
1765     if (length < info->length)
1766         return -ENOMEM;
1767 
1768     /* Adjust search limits by the desired length */
1769     if (info->high_limit < length)
1770         return -ENOMEM;
1771     high_limit = info->high_limit - length;
1772 
1773     if (info->low_limit > high_limit)
1774         return -ENOMEM;
1775     low_limit = info->low_limit + length;
1776 
1777     /* Check if rbtree root looks promising */
1778     if (RB_EMPTY_ROOT(&mm->mm_rb))
1779         goto check_highest;
1780     vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1781     if (vma->rb_subtree_gap < length)
1782         goto check_highest;
1783 
1784     while (true) {
1785         /* Visit left subtree if it looks promising */
1786         gap_end = vma->vm_start;
1787         if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1788             struct vm_area_struct *left =
1789                 rb_entry(vma->vm_rb.rb_left,
1790                      struct vm_area_struct, vm_rb);
1791             if (left->rb_subtree_gap >= length) {
1792                 vma = left;
1793                 continue;
1794             }
1795         }
1796 
1797         gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1798 check_current:
1799         /* Check if current node has a suitable gap */
1800         if (gap_start > high_limit)
1801             return -ENOMEM;
1802         if (gap_end >= low_limit && gap_end - gap_start >= length)
1803             goto found;
1804 
1805         /* Visit right subtree if it looks promising */
1806         if (vma->vm_rb.rb_right) {
1807             struct vm_area_struct *right =
1808                 rb_entry(vma->vm_rb.rb_right,
1809                      struct vm_area_struct, vm_rb);
1810             if (right->rb_subtree_gap >= length) {
1811                 vma = right;
1812                 continue;
1813             }
1814         }
1815 
1816         /* Go back up the rbtree to find next candidate node */
1817         while (true) {
1818             struct rb_node *prev = &vma->vm_rb;
1819             if (!rb_parent(prev))
1820                 goto check_highest;
1821             vma = rb_entry(rb_parent(prev),
1822                        struct vm_area_struct, vm_rb);
1823             if (prev == vma->vm_rb.rb_left) {
1824                 gap_start = vma->vm_prev->vm_end;
1825                 gap_end = vma->vm_start;
1826                 goto check_current;
1827             }
1828         }
1829     }
1830 
1831 check_highest:
1832     /* Check highest gap, which does not precede any rbtree node */
1833     gap_start = mm->highest_vm_end;
1834     gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1835     if (gap_start > high_limit)
1836         return -ENOMEM;
1837 
1838 found:
1839     /* We found a suitable gap. Clip it with the original low_limit. */
1840     if (gap_start < info->low_limit)
1841         gap_start = info->low_limit;
1842 
1843     /* Adjust gap address to the desired alignment */
1844     gap_start += (info->align_offset - gap_start) & info->align_mask;
1845 
1846     VM_BUG_ON(gap_start + info->length > info->high_limit);
1847     VM_BUG_ON(gap_start + info->length > gap_end);
1848     return gap_start;
1849 }
1850 
1851 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1852 {
1853     struct mm_struct *mm = current->mm;
1854     struct vm_area_struct *vma;
1855     unsigned long length, low_limit, high_limit, gap_start, gap_end;
1856 
1857     /* Adjust search length to account for worst case alignment overhead */
1858     length = info->length + info->align_mask;
1859     if (length < info->length)
1860         return -ENOMEM;
1861 
1862     /*
1863      * Adjust search limits by the desired length.
1864      * See implementation comment at top of unmapped_area().
1865      */
1866     gap_end = info->high_limit;
1867     if (gap_end < length)
1868         return -ENOMEM;
1869     high_limit = gap_end - length;
1870 
1871     if (info->low_limit > high_limit)
1872         return -ENOMEM;
1873     low_limit = info->low_limit + length;
1874 
1875     /* Check highest gap, which does not precede any rbtree node */
1876     gap_start = mm->highest_vm_end;
1877     if (gap_start <= high_limit)
1878         goto found_highest;
1879 
1880     /* Check if rbtree root looks promising */
1881     if (RB_EMPTY_ROOT(&mm->mm_rb))
1882         return -ENOMEM;
1883     vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1884     if (vma->rb_subtree_gap < length)
1885         return -ENOMEM;
1886 
1887     while (true) {
1888         /* Visit right subtree if it looks promising */
1889         gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1890         if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1891             struct vm_area_struct *right =
1892                 rb_entry(vma->vm_rb.rb_right,
1893                      struct vm_area_struct, vm_rb);
1894             if (right->rb_subtree_gap >= length) {
1895                 vma = right;
1896                 continue;
1897             }
1898         }
1899 
1900 check_current:
1901         /* Check if current node has a suitable gap */
1902         gap_end = vma->vm_start;
1903         if (gap_end < low_limit)
1904             return -ENOMEM;
1905         if (gap_start <= high_limit && gap_end - gap_start >= length)
1906             goto found;
1907 
1908         /* Visit left subtree if it looks promising */
1909         if (vma->vm_rb.rb_left) {
1910             struct vm_area_struct *left =
1911                 rb_entry(vma->vm_rb.rb_left,
1912                      struct vm_area_struct, vm_rb);
1913             if (left->rb_subtree_gap >= length) {
1914                 vma = left;
1915                 continue;
1916             }
1917         }
1918 
1919         /* Go back up the rbtree to find next candidate node */
1920         while (true) {
1921             struct rb_node *prev = &vma->vm_rb;
1922             if (!rb_parent(prev))
1923                 return -ENOMEM;
1924             vma = rb_entry(rb_parent(prev),
1925                        struct vm_area_struct, vm_rb);
1926             if (prev == vma->vm_rb.rb_right) {
1927                 gap_start = vma->vm_prev ?
1928                     vma->vm_prev->vm_end : 0;
1929                 goto check_current;
1930             }
1931         }
1932     }
1933 
1934 found:
1935     /* We found a suitable gap. Clip it with the original high_limit. */
1936     if (gap_end > info->high_limit)
1937         gap_end = info->high_limit;
1938 
1939 found_highest:
1940     /* Compute highest gap address at the desired alignment */
1941     gap_end -= info->length;
1942     gap_end -= (gap_end - info->align_offset) & info->align_mask;
1943 
1944     VM_BUG_ON(gap_end < info->low_limit);
1945     VM_BUG_ON(gap_end < gap_start);
1946     return gap_end;
1947 }
1948 
1949 /* Get an address range which is currently unmapped.
1950  * For shmat() with addr=0.
1951  *
1952  * Ugly calling convention alert:
1953  * Return value with the low bits set means error value,
1954  * ie
1955  *  if (ret & ~PAGE_MASK)
1956  *      error = ret;
1957  *
1958  * This function "knows" that -ENOMEM has the bits set.
1959  */
1960 #ifndef HAVE_ARCH_UNMAPPED_AREA
1961 unsigned long
1962 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1963         unsigned long len, unsigned long pgoff, unsigned long flags)
1964 {
1965     struct mm_struct *mm = current->mm;
1966     struct vm_area_struct *vma;
1967     struct vm_unmapped_area_info info;
1968 
1969     if (len > TASK_SIZE - mmap_min_addr)
1970         return -ENOMEM;
1971 
1972     if (flags & MAP_FIXED)
1973         return addr;
1974 
1975     if (addr) {
1976         addr = PAGE_ALIGN(addr);
1977         vma = find_vma(mm, addr);
1978         if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
1979             (!vma || addr + len <= vma->vm_start))
1980             return addr;
1981     }
1982 
1983     info.flags = 0;
1984     info.length = len;
1985     info.low_limit = mm->mmap_base;
1986     info.high_limit = TASK_SIZE;
1987     info.align_mask = 0;
1988     return vm_unmapped_area(&info);
1989 }
1990 #endif
1991 
1992 /*
1993  * This mmap-allocator allocates new areas top-down from below the
1994  * stack's low limit (the base):
1995  */
1996 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1997 unsigned long
1998 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1999               const unsigned long len, const unsigned long pgoff,
2000               const unsigned long flags)
2001 {
2002     struct vm_area_struct *vma;
2003     struct mm_struct *mm = current->mm;
2004     unsigned long addr = addr0;
2005     struct vm_unmapped_area_info info;
2006 
2007     /* requested length too big for entire address space */
2008     if (len > TASK_SIZE - mmap_min_addr)
2009         return -ENOMEM;
2010 
2011     if (flags & MAP_FIXED)
2012         return addr;
2013 
2014     /* requesting a specific address */
2015     if (addr) {
2016         addr = PAGE_ALIGN(addr);
2017         vma = find_vma(mm, addr);
2018         if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2019                 (!vma || addr + len <= vma->vm_start))
2020             return addr;
2021     }
2022 
2023     info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2024     info.length = len;
2025     info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2026     info.high_limit = mm->mmap_base;
2027     info.align_mask = 0;
2028     addr = vm_unmapped_area(&info);
2029 
2030     /*
2031      * A failed mmap() very likely causes application failure,
2032      * so fall back to the bottom-up function here. This scenario
2033      * can happen with large stack limits and large mmap()
2034      * allocations.
2035      */
2036     if (offset_in_page(addr)) {
2037         VM_BUG_ON(addr != -ENOMEM);
2038         info.flags = 0;
2039         info.low_limit = TASK_UNMAPPED_BASE;
2040         info.high_limit = TASK_SIZE;
2041         addr = vm_unmapped_area(&info);
2042     }
2043 
2044     return addr;
2045 }
2046 #endif
2047 
2048 unsigned long
2049 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2050         unsigned long pgoff, unsigned long flags)
2051 {
2052     unsigned long (*get_area)(struct file *, unsigned long,
2053                   unsigned long, unsigned long, unsigned long);
2054 
2055     unsigned long error = arch_mmap_check(addr, len, flags);
2056     if (error)
2057         return error;
2058 
2059     /* Careful about overflows.. */
2060     if (len > TASK_SIZE)
2061         return -ENOMEM;
2062 
2063     get_area = current->mm->get_unmapped_area;
2064     if (file) {
2065         if (file->f_op->get_unmapped_area)
2066             get_area = file->f_op->get_unmapped_area;
2067     } else if (flags & MAP_SHARED) {
2068         /*
2069          * mmap_region() will call shmem_zero_setup() to create a file,
2070          * so use shmem's get_unmapped_area in case it can be huge.
2071          * do_mmap_pgoff() will clear pgoff, so match alignment.
2072          */
2073         pgoff = 0;
2074         get_area = shmem_get_unmapped_area;
2075     }
2076 
2077     addr = get_area(file, addr, len, pgoff, flags);
2078     if (IS_ERR_VALUE(addr))
2079         return addr;
2080 
2081     if (addr > TASK_SIZE - len)
2082         return -ENOMEM;
2083     if (offset_in_page(addr))
2084         return -EINVAL;
2085 
2086     error = security_mmap_addr(addr);
2087     return error ? error : addr;
2088 }
2089 
2090 EXPORT_SYMBOL(get_unmapped_area);
2091 
2092 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2093 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2094 {
2095     struct rb_node *rb_node;
2096     struct vm_area_struct *vma;
2097 
2098     /* Check the cache first. */
2099     vma = vmacache_find(mm, addr);
2100     if (likely(vma))
2101         return vma;
2102 
2103     rb_node = mm->mm_rb.rb_node;
2104 
2105     while (rb_node) {
2106         struct vm_area_struct *tmp;
2107 
2108         tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2109 
2110         if (tmp->vm_end > addr) {
2111             vma = tmp;
2112             if (tmp->vm_start <= addr)
2113                 break;
2114             rb_node = rb_node->rb_left;
2115         } else
2116             rb_node = rb_node->rb_right;
2117     }
2118 
2119     if (vma)
2120         vmacache_update(addr, vma);
2121     return vma;
2122 }
2123 
2124 EXPORT_SYMBOL(find_vma);
2125 
2126 /*
2127  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2128  */
2129 struct vm_area_struct *
2130 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2131             struct vm_area_struct **pprev)
2132 {
2133     struct vm_area_struct *vma;
2134 
2135     vma = find_vma(mm, addr);
2136     if (vma) {
2137         *pprev = vma->vm_prev;
2138     } else {
2139         struct rb_node *rb_node = mm->mm_rb.rb_node;
2140         *pprev = NULL;
2141         while (rb_node) {
2142             *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2143             rb_node = rb_node->rb_right;
2144         }
2145     }
2146     return vma;
2147 }
2148 
2149 /*
2150  * Verify that the stack growth is acceptable and
2151  * update accounting. This is shared with both the
2152  * grow-up and grow-down cases.
2153  */
2154 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2155 {
2156     struct mm_struct *mm = vma->vm_mm;
2157     struct rlimit *rlim = current->signal->rlim;
2158     unsigned long new_start, actual_size;
2159 
2160     /* address space limit tests */
2161     if (!may_expand_vm(mm, vma->vm_flags, grow))
2162         return -ENOMEM;
2163 
2164     /* Stack limit test */
2165     actual_size = size;
2166     if (size && (vma->vm_flags & (VM_GROWSUP | VM_GROWSDOWN)))
2167         actual_size -= PAGE_SIZE;
2168     if (actual_size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2169         return -ENOMEM;
2170 
2171     /* mlock limit tests */
2172     if (vma->vm_flags & VM_LOCKED) {
2173         unsigned long locked;
2174         unsigned long limit;
2175         locked = mm->locked_vm + grow;
2176         limit = READ_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2177         limit >>= PAGE_SHIFT;
2178         if (locked > limit && !capable(CAP_IPC_LOCK))
2179             return -ENOMEM;
2180     }
2181 
2182     /* Check to ensure the stack will not grow into a hugetlb-only region */
2183     new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2184             vma->vm_end - size;
2185     if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2186         return -EFAULT;
2187 
2188     /*
2189      * Overcommit..  This must be the final test, as it will
2190      * update security statistics.
2191      */
2192     if (security_vm_enough_memory_mm(mm, grow))
2193         return -ENOMEM;
2194 
2195     return 0;
2196 }
2197 
2198 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2199 /*
2200  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2201  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2202  */
2203 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2204 {
2205     struct mm_struct *mm = vma->vm_mm;
2206     int error = 0;
2207 
2208     if (!(vma->vm_flags & VM_GROWSUP))
2209         return -EFAULT;
2210 
2211     /* Guard against wrapping around to address 0. */
2212     if (address < PAGE_ALIGN(address+4))
2213         address = PAGE_ALIGN(address+4);
2214     else
2215         return -ENOMEM;
2216 
2217     /* We must make sure the anon_vma is allocated. */
2218     if (unlikely(anon_vma_prepare(vma)))
2219         return -ENOMEM;
2220 
2221     /*
2222      * vma->vm_start/vm_end cannot change under us because the caller
2223      * is required to hold the mmap_sem in read mode.  We need the
2224      * anon_vma lock to serialize against concurrent expand_stacks.
2225      */
2226     anon_vma_lock_write(vma->anon_vma);
2227 
2228     /* Somebody else might have raced and expanded it already */
2229     if (address > vma->vm_end) {
2230         unsigned long size, grow;
2231 
2232         size = address - vma->vm_start;
2233         grow = (address - vma->vm_end) >> PAGE_SHIFT;
2234 
2235         error = -ENOMEM;
2236         if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2237             error = acct_stack_growth(vma, size, grow);
2238             if (!error) {
2239                 /*
2240                  * vma_gap_update() doesn't support concurrent
2241                  * updates, but we only hold a shared mmap_sem
2242                  * lock here, so we need to protect against
2243                  * concurrent vma expansions.
2244                  * anon_vma_lock_write() doesn't help here, as
2245                  * we don't guarantee that all growable vmas
2246                  * in a mm share the same root anon vma.
2247                  * So, we reuse mm->page_table_lock to guard
2248                  * against concurrent vma expansions.
2249                  */
2250                 spin_lock(&mm->page_table_lock);
2251                 if (vma->vm_flags & VM_LOCKED)
2252                     mm->locked_vm += grow;
2253                 vm_stat_account(mm, vma->vm_flags, grow);
2254                 anon_vma_interval_tree_pre_update_vma(vma);
2255                 vma->vm_end = address;
2256                 anon_vma_interval_tree_post_update_vma(vma);
2257                 if (vma->vm_next)
2258                     vma_gap_update(vma->vm_next);
2259                 else
2260                     mm->highest_vm_end = address;
2261                 spin_unlock(&mm->page_table_lock);
2262 
2263                 perf_event_mmap(vma);
2264             }
2265         }
2266     }
2267     anon_vma_unlock_write(vma->anon_vma);
2268     khugepaged_enter_vma_merge(vma, vma->vm_flags);
2269     validate_mm(mm);
2270     return error;
2271 }
2272 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2273 
2274 /*
2275  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2276  */
2277 int expand_downwards(struct vm_area_struct *vma,
2278                    unsigned long address)
2279 {
2280     struct mm_struct *mm = vma->vm_mm;
2281     int error;
2282 
2283     address &= PAGE_MASK;
2284     error = security_mmap_addr(address);
2285     if (error)
2286         return error;
2287 
2288     /* We must make sure the anon_vma is allocated. */
2289     if (unlikely(anon_vma_prepare(vma)))
2290         return -ENOMEM;
2291 
2292     /*
2293      * vma->vm_start/vm_end cannot change under us because the caller
2294      * is required to hold the mmap_sem in read mode.  We need the
2295      * anon_vma lock to serialize against concurrent expand_stacks.
2296      */
2297     anon_vma_lock_write(vma->anon_vma);
2298 
2299     /* Somebody else might have raced and expanded it already */
2300     if (address < vma->vm_start) {
2301         unsigned long size, grow;
2302 
2303         size = vma->vm_end - address;
2304         grow = (vma->vm_start - address) >> PAGE_SHIFT;
2305 
2306         error = -ENOMEM;
2307         if (grow <= vma->vm_pgoff) {
2308             error = acct_stack_growth(vma, size, grow);
2309             if (!error) {
2310                 /*
2311                  * vma_gap_update() doesn't support concurrent
2312                  * updates, but we only hold a shared mmap_sem
2313                  * lock here, so we need to protect against
2314                  * concurrent vma expansions.
2315                  * anon_vma_lock_write() doesn't help here, as
2316                  * we don't guarantee that all growable vmas
2317                  * in a mm share the same root anon vma.
2318                  * So, we reuse mm->page_table_lock to guard
2319                  * against concurrent vma expansions.
2320                  */
2321                 spin_lock(&mm->page_table_lock);
2322                 if (vma->vm_flags & VM_LOCKED)
2323                     mm->locked_vm += grow;
2324                 vm_stat_account(mm, vma->vm_flags, grow);
2325                 anon_vma_interval_tree_pre_update_vma(vma);
2326                 vma->vm_start = address;
2327                 vma->vm_pgoff -= grow;
2328                 anon_vma_interval_tree_post_update_vma(vma);
2329                 vma_gap_update(vma);
2330                 spin_unlock(&mm->page_table_lock);
2331 
2332                 perf_event_mmap(vma);
2333             }
2334         }
2335     }
2336     anon_vma_unlock_write(vma->anon_vma);
2337     khugepaged_enter_vma_merge(vma, vma->vm_flags);
2338     validate_mm(mm);
2339     return error;
2340 }
2341 
2342 /*
2343  * Note how expand_stack() refuses to expand the stack all the way to
2344  * abut the next virtual mapping, *unless* that mapping itself is also
2345  * a stack mapping. We want to leave room for a guard page, after all
2346  * (the guard page itself is not added here, that is done by the
2347  * actual page faulting logic)
2348  *
2349  * This matches the behavior of the guard page logic (see mm/memory.c:
2350  * check_stack_guard_page()), which only allows the guard page to be
2351  * removed under these circumstances.
2352  */
2353 #ifdef CONFIG_STACK_GROWSUP
2354 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2355 {
2356     struct vm_area_struct *next;
2357 
2358     address &= PAGE_MASK;
2359     next = vma->vm_next;
2360     if (next && next->vm_start == address + PAGE_SIZE) {
2361         if (!(next->vm_flags & VM_GROWSUP))
2362             return -ENOMEM;
2363     }
2364     return expand_upwards(vma, address);
2365 }
2366 
2367 struct vm_area_struct *
2368 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2369 {
2370     struct vm_area_struct *vma, *prev;
2371 
2372     addr &= PAGE_MASK;
2373     vma = find_vma_prev(mm, addr, &prev);
2374     if (vma && (vma->vm_start <= addr))
2375         return vma;
2376     if (!prev || expand_stack(prev, addr))
2377         return NULL;
2378     if (prev->vm_flags & VM_LOCKED)
2379         populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2380     return prev;
2381 }
2382 #else
2383 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2384 {
2385     struct vm_area_struct *prev;
2386 
2387     address &= PAGE_MASK;
2388     prev = vma->vm_prev;
2389     if (prev && prev->vm_end == address) {
2390         if (!(prev->vm_flags & VM_GROWSDOWN))
2391             return -ENOMEM;
2392     }
2393     return expand_downwards(vma, address);
2394 }
2395 
2396 struct vm_area_struct *
2397 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2398 {
2399     struct vm_area_struct *vma;
2400     unsigned long start;
2401 
2402     addr &= PAGE_MASK;
2403     vma = find_vma(mm, addr);
2404     if (!vma)
2405         return NULL;
2406     if (vma->vm_start <= addr)
2407         return vma;
2408     if (!(vma->vm_flags & VM_GROWSDOWN))
2409         return NULL;
2410     start = vma->vm_start;
2411     if (expand_stack(vma, addr))
2412         return NULL;
2413     if (vma->vm_flags & VM_LOCKED)
2414         populate_vma_page_range(vma, addr, start, NULL);
2415     return vma;
2416 }
2417 #endif
2418 
2419 EXPORT_SYMBOL_GPL(find_extend_vma);
2420 
2421 /*
2422  * Ok - we have the memory areas we should free on the vma list,
2423  * so release them, and do the vma updates.
2424  *
2425  * Called with the mm semaphore held.
2426  */
2427 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2428 {
2429     unsigned long nr_accounted = 0;
2430 
2431     /* Update high watermark before we lower total_vm */
2432     update_hiwater_vm(mm);
2433     do {
2434         long nrpages = vma_pages(vma);
2435 
2436         if (vma->vm_flags & VM_ACCOUNT)
2437             nr_accounted += nrpages;
2438         vm_stat_account(mm, vma->vm_flags, -nrpages);
2439         vma = remove_vma(vma);
2440     } while (vma);
2441     vm_unacct_memory(nr_accounted);
2442     validate_mm(mm);
2443 }
2444 
2445 /*
2446  * Get rid of page table information in the indicated region.
2447  *
2448  * Called with the mm semaphore held.
2449  */
2450 static void unmap_region(struct mm_struct *mm,
2451         struct vm_area_struct *vma, struct vm_area_struct *prev,
2452         unsigned long start, unsigned long end)
2453 {
2454     struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2455     struct mmu_gather tlb;
2456 
2457     lru_add_drain();
2458     tlb_gather_mmu(&tlb, mm, start, end);
2459     update_hiwater_rss(mm);
2460     unmap_vmas(&tlb, vma, start, end);
2461     free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2462                  next ? next->vm_start : USER_PGTABLES_CEILING);
2463     tlb_finish_mmu(&tlb, start, end);
2464 }
2465 
2466 /*
2467  * Create a list of vma's touched by the unmap, removing them from the mm's
2468  * vma list as we go..
2469  */
2470 static void
2471 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2472     struct vm_area_struct *prev, unsigned long end)
2473 {
2474     struct vm_area_struct **insertion_point;
2475     struct vm_area_struct *tail_vma = NULL;
2476 
2477     insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2478     vma->vm_prev = NULL;
2479     do {
2480         vma_rb_erase(vma, &mm->mm_rb);
2481         mm->map_count--;
2482         tail_vma = vma;
2483         vma = vma->vm_next;
2484     } while (vma && vma->vm_start < end);
2485     *insertion_point = vma;
2486     if (vma) {
2487         vma->vm_prev = prev;
2488         vma_gap_update(vma);
2489     } else
2490         mm->highest_vm_end = prev ? prev->vm_end : 0;
2491     tail_vma->vm_next = NULL;
2492 
2493     /* Kill the cache */
2494     vmacache_invalidate(mm);
2495 }
2496 
2497 /*
2498  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2499  * munmap path where it doesn't make sense to fail.
2500  */
2501 static int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2502           unsigned long addr, int new_below)
2503 {
2504     struct vm_area_struct *new;
2505     int err;
2506 
2507     if (is_vm_hugetlb_page(vma) && (addr &
2508                     ~(huge_page_mask(hstate_vma(vma)))))
2509         return -EINVAL;
2510 
2511     new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2512     if (!new)
2513         return -ENOMEM;
2514 
2515     /* most fields are the same, copy all, and then fixup */
2516     *new = *vma;
2517 
2518     INIT_LIST_HEAD(&new->anon_vma_chain);
2519 
2520     if (new_below)
2521         new->vm_end = addr;
2522     else {
2523         new->vm_start = addr;
2524         new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2525     }
2526 
2527     err = vma_dup_policy(vma, new);
2528     if (err)
2529         goto out_free_vma;
2530 
2531     err = anon_vma_clone(new, vma);
2532     if (err)
2533         goto out_free_mpol;
2534 
2535     if (new->vm_file)
2536         get_file(new->vm_file);
2537 
2538     if (new->vm_ops && new->vm_ops->open)
2539         new->vm_ops->open(new);
2540 
2541     if (new_below)
2542         err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2543             ((addr - new->vm_start) >> PAGE_SHIFT), new);
2544     else
2545         err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2546 
2547     /* Success. */
2548     if (!err)
2549         return 0;
2550 
2551     /* Clean everything up if vma_adjust failed. */
2552     if (new->vm_ops && new->vm_ops->close)
2553         new->vm_ops->close(new);
2554     if (new->vm_file)
2555         fput(new->vm_file);
2556     unlink_anon_vmas(new);
2557  out_free_mpol:
2558     mpol_put(vma_policy(new));
2559  out_free_vma:
2560     kmem_cache_free(vm_area_cachep, new);
2561     return err;
2562 }
2563 
2564 /*
2565  * Split a vma into two pieces at address 'addr', a new vma is allocated
2566  * either for the first part or the tail.
2567  */
2568 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2569           unsigned long addr, int new_below)
2570 {
2571     if (mm->map_count >= sysctl_max_map_count)
2572         return -ENOMEM;
2573 
2574     return __split_vma(mm, vma, addr, new_below);
2575 }
2576 
2577 /* Munmap is split into 2 main parts -- this part which finds
2578  * what needs doing, and the areas themselves, which do the
2579  * work.  This now handles partial unmappings.
2580  * Jeremy Fitzhardinge <jeremy@goop.org>
2581  */
2582 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2583 {
2584     unsigned long end;
2585     struct vm_area_struct *vma, *prev, *last;
2586 
2587     if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2588         return -EINVAL;
2589 
2590     len = PAGE_ALIGN(len);
2591     if (len == 0)
2592         return -EINVAL;
2593 
2594     /* Find the first overlapping VMA */
2595     vma = find_vma(mm, start);
2596     if (!vma)
2597         return 0;
2598     prev = vma->vm_prev;
2599     /* we have  start < vma->vm_end  */
2600 
2601     /* if it doesn't overlap, we have nothing.. */
2602     end = start + len;
2603     if (vma->vm_start >= end)
2604         return 0;
2605 
2606     /*
2607      * If we need to split any vma, do it now to save pain later.
2608      *
2609      * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2610      * unmapped vm_area_struct will remain in use: so lower split_vma
2611      * places tmp vma above, and higher split_vma places tmp vma below.
2612      */
2613     if (start > vma->vm_start) {
2614         int error;
2615 
2616         /*
2617          * Make sure that map_count on return from munmap() will
2618          * not exceed its limit; but let map_count go just above
2619          * its limit temporarily, to help free resources as expected.
2620          */
2621         if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2622             return -ENOMEM;
2623 
2624         error = __split_vma(mm, vma, start, 0);
2625         if (error)
2626             return error;
2627         prev = vma;
2628     }
2629 
2630     /* Does it split the last one? */
2631     last = find_vma(mm, end);
2632     if (last && end > last->vm_start) {
2633         int error = __split_vma(mm, last, end, 1);
2634         if (error)
2635             return error;
2636     }
2637     vma = prev ? prev->vm_next : mm->mmap;
2638 
2639     /*
2640      * unlock any mlock()ed ranges before detaching vmas
2641      */
2642     if (mm->locked_vm) {
2643         struct vm_area_struct *tmp = vma;
2644         while (tmp && tmp->vm_start < end) {
2645             if (tmp->vm_flags & VM_LOCKED) {
2646                 mm->locked_vm -= vma_pages(tmp);
2647                 munlock_vma_pages_all(tmp);
2648             }
2649             tmp = tmp->vm_next;
2650         }
2651     }
2652 
2653     /*
2654      * Remove the vma's, and unmap the actual pages
2655      */
2656     detach_vmas_to_be_unmapped(mm, vma, prev, end);
2657     unmap_region(mm, vma, prev, start, end);
2658 
2659     arch_unmap(mm, vma, start, end);
2660 
2661     /* Fix up all other VM information */
2662     remove_vma_list(mm, vma);
2663 
2664     return 0;
2665 }
2666 
2667 int vm_munmap(unsigned long start, size_t len)
2668 {
2669     int ret;
2670     struct mm_struct *mm = current->mm;
2671 
2672     if (down_write_killable(&mm->mmap_sem))
2673         return -EINTR;
2674 
2675     ret = do_munmap(mm, start, len);
2676     up_write(&mm->mmap_sem);
2677     return ret;
2678 }
2679 EXPORT_SYMBOL(vm_munmap);
2680 
2681 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2682 {
2683     int ret;
2684     struct mm_struct *mm = current->mm;
2685 
2686     profile_munmap(addr);
2687     if (down_write_killable(&mm->mmap_sem))
2688         return -EINTR;
2689     ret = do_munmap(mm, addr, len);
2690     up_write(&mm->mmap_sem);
2691     return ret;
2692 }
2693 
2694 
2695 /*
2696  * Emulation of deprecated remap_file_pages() syscall.
2697  */
2698 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2699         unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2700 {
2701 
2702     struct mm_struct *mm = current->mm;
2703     struct vm_area_struct *vma;
2704     unsigned long populate = 0;
2705     unsigned long ret = -EINVAL;
2706     struct file *file;
2707 
2708     pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2709              current->comm, current->pid);
2710 
2711     if (prot)
2712         return ret;
2713     start = start & PAGE_MASK;
2714     size = size & PAGE_MASK;
2715 
2716     if (start + size <= start)
2717         return ret;
2718 
2719     /* Does pgoff wrap? */
2720     if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2721         return ret;
2722 
2723     if (down_write_killable(&mm->mmap_sem))
2724         return -EINTR;
2725 
2726     vma = find_vma(mm, start);
2727 
2728     if (!vma || !(vma->vm_flags & VM_SHARED))
2729         goto out;
2730 
2731     if (start < vma->vm_start)
2732         goto out;
2733 
2734     if (start + size > vma->vm_end) {
2735         struct vm_area_struct *next;
2736 
2737         for (next = vma->vm_next; next; next = next->vm_next) {
2738             /* hole between vmas ? */
2739             if (next->vm_start != next->vm_prev->vm_end)
2740                 goto out;
2741 
2742             if (next->vm_file != vma->vm_file)
2743                 goto out;
2744 
2745             if (next->vm_flags != vma->vm_flags)
2746                 goto out;
2747 
2748             if (start + size <= next->vm_end)
2749                 break;
2750         }
2751 
2752         if (!next)
2753             goto out;
2754     }
2755 
2756     prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2757     prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2758     prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2759 
2760     flags &= MAP_NONBLOCK;
2761     flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2762     if (vma->vm_flags & VM_LOCKED) {
2763         struct vm_area_struct *tmp;
2764         flags |= MAP_LOCKED;
2765 
2766         /* drop PG_Mlocked flag for over-mapped range */
2767         for (tmp = vma; tmp->vm_start >= start + size;
2768                 tmp = tmp->vm_next) {
2769             /*
2770              * Split pmd and munlock page on the border
2771              * of the range.
2772              */
2773             vma_adjust_trans_huge(tmp, start, start + size, 0);
2774 
2775             munlock_vma_pages_range(tmp,
2776                     max(tmp->vm_start, start),
2777                     min(tmp->vm_end, start + size));
2778         }
2779     }
2780 
2781     file = get_file(vma->vm_file);
2782     ret = do_mmap_pgoff(vma->vm_file, start, size,
2783             prot, flags, pgoff, &populate);
2784     fput(file);
2785 out:
2786     up_write(&mm->mmap_sem);
2787     if (populate)
2788         mm_populate(ret, populate);
2789     if (!IS_ERR_VALUE(ret))
2790         ret = 0;
2791     return ret;
2792 }
2793 
2794 static inline void verify_mm_writelocked(struct mm_struct *mm)
2795 {
2796 #ifdef CONFIG_DEBUG_VM
2797     if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2798         WARN_ON(1);
2799         up_read(&mm->mmap_sem);
2800     }
2801 #endif
2802 }
2803 
2804 /*
2805  *  this is really a simplified "do_mmap".  it only handles
2806  *  anonymous maps.  eventually we may be able to do some
2807  *  brk-specific accounting here.
2808  */
2809 static int do_brk(unsigned long addr, unsigned long request)
2810 {
2811     struct mm_struct *mm = current->mm;
2812     struct vm_area_struct *vma, *prev;
2813     unsigned long flags, len;
2814     struct rb_node **rb_link, *rb_parent;
2815     pgoff_t pgoff = addr >> PAGE_SHIFT;
2816     int error;
2817 
2818     len = PAGE_ALIGN(request);
2819     if (len < request)
2820         return -ENOMEM;
2821     if (!len)
2822         return 0;
2823 
2824     flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2825 
2826     error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2827     if (offset_in_page(error))
2828         return error;
2829 
2830     error = mlock_future_check(mm, mm->def_flags, len);
2831     if (error)
2832         return error;
2833 
2834     /*
2835      * mm->mmap_sem is required to protect against another thread
2836      * changing the mappings in case we sleep.
2837      */
2838     verify_mm_writelocked(mm);
2839 
2840     /*
2841      * Clear old maps.  this also does some error checking for us
2842      */
2843     while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2844                   &rb_parent)) {
2845         if (do_munmap(mm, addr, len))
2846             return -ENOMEM;
2847     }
2848 
2849     /* Check against address space limits *after* clearing old maps... */
2850     if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2851         return -ENOMEM;
2852 
2853     if (mm->map_count > sysctl_max_map_count)
2854         return -ENOMEM;
2855 
2856     if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2857         return -ENOMEM;
2858 
2859     /* Can we just expand an old private anonymous mapping? */
2860     vma = vma_merge(mm, prev, addr, addr + len, flags,
2861             NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2862     if (vma)
2863         goto out;
2864 
2865     /*
2866      * create a vma struct for an anonymous mapping
2867      */
2868     vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2869     if (!vma) {
2870         vm_unacct_memory(len >> PAGE_SHIFT);
2871         return -ENOMEM;
2872     }
2873 
2874     INIT_LIST_HEAD(&vma->anon_vma_chain);
2875     vma->vm_mm = mm;
2876     vma->vm_start = addr;
2877     vma->vm_end = addr + len;
2878     vma->vm_pgoff = pgoff;
2879     vma->vm_flags = flags;
2880     vma->vm_page_prot = vm_get_page_prot(flags);
2881     vma_link(mm, vma, prev, rb_link, rb_parent);
2882 out:
2883     perf_event_mmap(vma);
2884     mm->total_vm += len >> PAGE_SHIFT;
2885     mm->data_vm += len >> PAGE_SHIFT;
2886     if (flags & VM_LOCKED)
2887         mm->locked_vm += (len >> PAGE_SHIFT);
2888     vma->vm_flags |= VM_SOFTDIRTY;
2889     return 0;
2890 }
2891 
2892 int vm_brk(unsigned long addr, unsigned long len)
2893 {
2894     struct mm_struct *mm = current->mm;
2895     int ret;
2896     bool populate;
2897 
2898     if (down_write_killable(&mm->mmap_sem))
2899         return -EINTR;
2900 
2901     ret = do_brk(addr, len);
2902     populate = ((mm->def_flags & VM_LOCKED) != 0);
2903     up_write(&mm->mmap_sem);
2904     if (populate && !ret)
2905         mm_populate(addr, len);
2906     return ret;
2907 }
2908 EXPORT_SYMBOL(vm_brk);
2909 
2910 /* Release all mmaps. */
2911 void exit_mmap(struct mm_struct *mm)
2912 {
2913     struct mmu_gather tlb;
2914     struct vm_area_struct *vma;
2915     unsigned long nr_accounted = 0;
2916 
2917     /* mm's last user has gone, and its about to be pulled down */
2918     mmu_notifier_release(mm);
2919 
2920     if (mm->locked_vm) {
2921         vma = mm->mmap;
2922         while (vma) {
2923             if (vma->vm_flags & VM_LOCKED)
2924                 munlock_vma_pages_all(vma);
2925             vma = vma->vm_next;
2926         }
2927     }
2928 
2929     arch_exit_mmap(mm);
2930 
2931     vma = mm->mmap;
2932     if (!vma)   /* Can happen if dup_mmap() received an OOM */
2933         return;
2934 
2935     lru_add_drain();
2936     flush_cache_mm(mm);
2937     tlb_gather_mmu(&tlb, mm, 0, -1);
2938     /* update_hiwater_rss(mm) here? but nobody should be looking */
2939     /* Use -1 here to ensure all VMAs in the mm are unmapped */
2940     unmap_vmas(&tlb, vma, 0, -1);
2941 
2942     free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
2943     tlb_finish_mmu(&tlb, 0, -1);
2944 
2945     /*
2946      * Walk the list again, actually closing and freeing it,
2947      * with preemption enabled, without holding any MM locks.
2948      */
2949     while (vma) {
2950         if (vma->vm_flags & VM_ACCOUNT)
2951             nr_accounted += vma_pages(vma);
2952         vma = remove_vma(vma);
2953     }
2954     vm_unacct_memory(nr_accounted);
2955 }
2956 
2957 /* Insert vm structure into process list sorted by address
2958  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2959  * then i_mmap_rwsem is taken here.
2960  */
2961 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2962 {
2963     struct vm_area_struct *prev;
2964     struct rb_node **rb_link, *rb_parent;
2965 
2966     if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2967                &prev, &rb_link, &rb_parent))
2968         return -ENOMEM;
2969     if ((vma->vm_flags & VM_ACCOUNT) &&
2970          security_vm_enough_memory_mm(mm, vma_pages(vma)))
2971         return -ENOMEM;
2972 
2973     /*
2974      * The vm_pgoff of a purely anonymous vma should be irrelevant
2975      * until its first write fault, when page's anon_vma and index
2976      * are set.  But now set the vm_pgoff it will almost certainly
2977      * end up with (unless mremap moves it elsewhere before that
2978      * first wfault), so /proc/pid/maps tells a consistent story.
2979      *
2980      * By setting it to reflect the virtual start address of the
2981      * vma, merges and splits can happen in a seamless way, just
2982      * using the existing file pgoff checks and manipulations.
2983      * Similarly in do_mmap_pgoff and in do_brk.
2984      */
2985     if (vma_is_anonymous(vma)) {
2986         BUG_ON(vma->anon_vma);
2987         vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2988     }
2989 
2990     vma_link(mm, vma, prev, rb_link, rb_parent);
2991     return 0;
2992 }
2993 
2994 /*
2995  * Copy the vma structure to a new location in the same mm,
2996  * prior to moving page table entries, to effect an mremap move.
2997  */
2998 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2999     unsigned long addr, unsigned long len, pgoff_t pgoff,
3000     bool *need_rmap_locks)
3001 {
3002     struct vm_area_struct *vma = *vmap;
3003     unsigned long vma_start = vma->vm_start;
3004     struct mm_struct *mm = vma->vm_mm;
3005     struct vm_area_struct *new_vma, *prev;
3006     struct rb_node **rb_link, *rb_parent;
3007     bool faulted_in_anon_vma = true;
3008 
3009     /*
3010      * If anonymous vma has not yet been faulted, update new pgoff
3011      * to match new location, to increase its chance of merging.
3012      */
3013     if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3014         pgoff = addr >> PAGE_SHIFT;
3015         faulted_in_anon_vma = false;
3016     }
3017 
3018     if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3019         return NULL;    /* should never get here */
3020     new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3021                 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3022                 vma->vm_userfaultfd_ctx);
3023     if (new_vma) {
3024         /*
3025          * Source vma may have been merged into new_vma
3026          */
3027         if (unlikely(vma_start >= new_vma->vm_start &&
3028                  vma_start < new_vma->vm_end)) {
3029             /*
3030              * The only way we can get a vma_merge with
3031              * self during an mremap is if the vma hasn't
3032              * been faulted in yet and we were allowed to
3033              * reset the dst vma->vm_pgoff to the
3034              * destination address of the mremap to allow
3035              * the merge to happen. mremap must change the
3036              * vm_pgoff linearity between src and dst vmas
3037              * (in turn preventing a vma_merge) to be
3038              * safe. It is only safe to keep the vm_pgoff
3039              * linear if there are no pages mapped yet.
3040              */
3041             VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3042             *vmap = vma = new_vma;
3043         }
3044         *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3045     } else {
3046         new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3047         if (!new_vma)
3048             goto out;
3049         *new_vma = *vma;
3050         new_vma->vm_start = addr;
3051         new_vma->vm_end = addr + len;
3052         new_vma->vm_pgoff = pgoff;
3053         if (vma_dup_policy(vma, new_vma))
3054             goto out_free_vma;
3055         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3056         if (anon_vma_clone(new_vma, vma))
3057             goto out_free_mempol;
3058         if (new_vma->vm_file)
3059             get_file(new_vma->vm_file);
3060         if (new_vma->vm_ops && new_vma->vm_ops->open)
3061             new_vma->vm_ops->open(new_vma);
3062         vma_link(mm, new_vma, prev, rb_link, rb_parent);
3063         *need_rmap_locks = false;
3064     }
3065     return new_vma;
3066 
3067 out_free_mempol:
3068     mpol_put(vma_policy(new_vma));
3069 out_free_vma:
3070     kmem_cache_free(vm_area_cachep, new_vma);
3071 out:
3072     return NULL;
3073 }
3074 
3075 /*
3076  * Return true if the calling process may expand its vm space by the passed
3077  * number of pages
3078  */
3079 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3080 {
3081     if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3082         return false;
3083 
3084     if (is_data_mapping(flags) &&
3085         mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3086         /* Workaround for Valgrind */
3087         if (rlimit(RLIMIT_DATA) == 0 &&
3088             mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3089             return true;
3090         if (!ignore_rlimit_data) {
3091             pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits or use boot option ignore_rlimit_data.\n",
3092                      current->comm, current->pid,
3093                      (mm->data_vm + npages) << PAGE_SHIFT,
3094                      rlimit(RLIMIT_DATA));
3095             return false;
3096         }
3097     }
3098 
3099     return true;
3100 }
3101 
3102 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3103 {
3104     mm->total_vm += npages;
3105 
3106     if (is_exec_mapping(flags))
3107         mm->exec_vm += npages;
3108     else if (is_stack_mapping(flags))
3109         mm->stack_vm += npages;
3110     else if (is_data_mapping(flags))
3111         mm->data_vm += npages;
3112 }
3113 
3114 static int special_mapping_fault(struct vm_area_struct *vma,
3115                  struct vm_fault *vmf);
3116 
3117 /*
3118  * Having a close hook prevents vma merging regardless of flags.
3119  */
3120 static void special_mapping_close(struct vm_area_struct *vma)
3121 {
3122 }
3123 
3124 static const char *special_mapping_name(struct vm_area_struct *vma)
3125 {
3126     return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3127 }
3128 
3129 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3130 {
3131     struct vm_special_mapping *sm = new_vma->vm_private_data;
3132 
3133     if (sm->mremap)
3134         return sm->mremap(sm, new_vma);
3135     return 0;
3136 }
3137 
3138 static const struct vm_operations_struct special_mapping_vmops = {
3139     .close = special_mapping_close,
3140     .fault = special_mapping_fault,
3141     .mremap = special_mapping_mremap,
3142     .name = special_mapping_name,
3143 };
3144 
3145 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3146     .close = special_mapping_close,
3147     .fault = special_mapping_fault,
3148 };
3149 
3150 static int special_mapping_fault(struct vm_area_struct *vma,
3151                 struct vm_fault *vmf)
3152 {
3153     pgoff_t pgoff;
3154     struct page **pages;
3155 
3156     if (vma->vm_ops == &legacy_special_mapping_vmops) {
3157         pages = vma->vm_private_data;
3158     } else {
3159         struct vm_special_mapping *sm = vma->vm_private_data;
3160 
3161         if (sm->fault)
3162             return sm->fault(sm, vma, vmf);
3163 
3164         pages = sm->pages;
3165     }
3166 
3167     for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3168         pgoff--;
3169 
3170     if (*pages) {
3171         struct page *page = *pages;
3172         get_page(page);
3173         vmf->page = page;
3174         return 0;
3175     }
3176 
3177     return VM_FAULT_SIGBUS;
3178 }
3179 
3180 static struct vm_area_struct *__install_special_mapping(
3181     struct mm_struct *mm,
3182     unsigned long addr, unsigned long len,
3183     unsigned long vm_flags, void *priv,
3184     const struct vm_operations_struct *ops)
3185 {
3186     int ret;
3187     struct vm_area_struct *vma;
3188 
3189     vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
3190     if (unlikely(vma == NULL))
3191         return ERR_PTR(-ENOMEM);
3192 
3193     INIT_LIST_HEAD(&vma->anon_vma_chain);
3194     vma->vm_mm = mm;
3195     vma->vm_start = addr;
3196     vma->vm_end = addr + len;
3197 
3198     vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3199     vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3200 
3201     vma->vm_ops = ops;
3202     vma->vm_private_data = priv;
3203 
3204     ret = insert_vm_struct(mm, vma);
3205     if (ret)
3206         goto out;
3207 
3208     vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3209 
3210     perf_event_mmap(vma);
3211 
3212     return vma;
3213 
3214 out:
3215     kmem_cache_free(vm_area_cachep, vma);
3216     return ERR_PTR(ret);
3217 }
3218 
3219 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3220     const struct vm_special_mapping *sm)
3221 {
3222     return vma->vm_private_data == sm &&
3223         (vma->vm_ops == &special_mapping_vmops ||
3224          vma->vm_ops == &legacy_special_mapping_vmops);
3225 }
3226 
3227 /*
3228  * Called with mm->mmap_sem held for writing.
3229  * Insert a new vma covering the given region, with the given flags.
3230  * Its pages are supplied by the given array of struct page *.
3231  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3232  * The region past the last page supplied will always produce SIGBUS.
3233  * The array pointer and the pages it points to are assumed to stay alive
3234  * for as long as this mapping might exist.
3235  */
3236 struct vm_area_struct *_install_special_mapping(
3237     struct mm_struct *mm,
3238     unsigned long addr, unsigned long len,
3239     unsigned long vm_flags, const struct vm_special_mapping *spec)
3240 {
3241     return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3242                     &special_mapping_vmops);
3243 }
3244 
3245 int install_special_mapping(struct mm_struct *mm,
3246                 unsigned long addr, unsigned long len,
3247                 unsigned long vm_flags, struct page **pages)
3248 {
3249     struct vm_area_struct *vma = __install_special_mapping(
3250         mm, addr, len, vm_flags, (void *)pages,
3251         &legacy_special_mapping_vmops);
3252 
3253     return PTR_ERR_OR_ZERO(vma);
3254 }
3255 
3256 static DEFINE_MUTEX(mm_all_locks_mutex);
3257 
3258 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3259 {
3260     if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3261         /*
3262          * The LSB of head.next can't change from under us
3263          * because we hold the mm_all_locks_mutex.
3264          */
3265         down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3266         /*
3267          * We can safely modify head.next after taking the
3268          * anon_vma->root->rwsem. If some other vma in this mm shares
3269          * the same anon_vma we won't take it again.
3270          *
3271          * No need of atomic instructions here, head.next
3272          * can't change from under us thanks to the
3273          * anon_vma->root->rwsem.
3274          */
3275         if (__test_and_set_bit(0, (unsigned long *)
3276                        &anon_vma->root->rb_root.rb_node))
3277             BUG();
3278     }
3279 }
3280 
3281 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3282 {
3283     if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3284         /*
3285          * AS_MM_ALL_LOCKS can't change from under us because
3286          * we hold the mm_all_locks_mutex.
3287          *
3288          * Operations on ->flags have to be atomic because
3289          * even if AS_MM_ALL_LOCKS is stable thanks to the
3290          * mm_all_locks_mutex, there may be other cpus
3291          * changing other bitflags in parallel to us.
3292          */
3293         if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3294             BUG();
3295         down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3296     }
3297 }
3298 
3299 /*
3300  * This operation locks against the VM for all pte/vma/mm related
3301  * operations that could ever happen on a certain mm. This includes
3302  * vmtruncate, try_to_unmap, and all page faults.
3303  *
3304  * The caller must take the mmap_sem in write mode before calling
3305  * mm_take_all_locks(). The caller isn't allowed to release the
3306  * mmap_sem until mm_drop_all_locks() returns.
3307  *
3308  * mmap_sem in write mode is required in order to block all operations
3309  * that could modify pagetables and free pages without need of
3310  * altering the vma layout. It's also needed in write mode to avoid new
3311  * anon_vmas to be associated with existing vmas.
3312  *
3313  * A single task can't take more than one mm_take_all_locks() in a row
3314  * or it would deadlock.
3315  *
3316  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3317  * mapping->flags avoid to take the same lock twice, if more than one
3318  * vma in this mm is backed by the same anon_vma or address_space.
3319  *
3320  * We take locks in following order, accordingly to comment at beginning
3321  * of mm/rmap.c:
3322  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3323  *     hugetlb mapping);
3324  *   - all i_mmap_rwsem locks;
3325  *   - all anon_vma->rwseml
3326  *
3327  * We can take all locks within these types randomly because the VM code
3328  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3329  * mm_all_locks_mutex.
3330  *
3331  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3332  * that may have to take thousand of locks.
3333  *
3334  * mm_take_all_locks() can fail if it's interrupted by signals.
3335  */
3336 int mm_take_all_locks(struct mm_struct *mm)
3337 {
3338     struct vm_area_struct *vma;
3339     struct anon_vma_chain *avc;
3340 
3341     BUG_ON(down_read_trylock(&mm->mmap_sem));
3342 
3343     mutex_lock(&mm_all_locks_mutex);
3344 
3345     for (vma = mm->mmap; vma; vma = vma->vm_next) {
3346         if (signal_pending(current))
3347             goto out_unlock;
3348         if (vma->vm_file && vma->vm_file->f_mapping &&
3349                 is_vm_hugetlb_page(vma))
3350             vm_lock_mapping(mm, vma->vm_file->f_mapping);
3351     }
3352 
3353     for (vma = mm->mmap; vma; vma = vma->vm_next) {
3354         if (signal_pending(current))
3355             goto out_unlock;
3356         if (vma->vm_file && vma->vm_file->f_mapping &&
3357                 !is_vm_hugetlb_page(vma))
3358             vm_lock_mapping(mm, vma->vm_file->f_mapping);
3359     }
3360 
3361     for (vma = mm->mmap; vma; vma = vma->vm_next) {
3362         if (signal_pending(current))
3363             goto out_unlock;
3364         if (vma->anon_vma)
3365             list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3366                 vm_lock_anon_vma(mm, avc->anon_vma);
3367     }
3368 
3369     return 0;
3370 
3371 out_unlock:
3372     mm_drop_all_locks(mm);
3373     return -EINTR;
3374 }
3375 
3376 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3377 {
3378     if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3379         /*
3380          * The LSB of head.next can't change to 0 from under
3381          * us because we hold the mm_all_locks_mutex.
3382          *
3383          * We must however clear the bitflag before unlocking
3384          * the vma so the users using the anon_vma->rb_root will
3385          * never see our bitflag.
3386          *
3387          * No need of atomic instructions here, head.next
3388          * can't change from under us until we release the
3389          * anon_vma->root->rwsem.
3390          */
3391         if (!__test_and_clear_bit(0, (unsigned long *)
3392                       &anon_vma->root->rb_root.rb_node))
3393             BUG();
3394         anon_vma_unlock_write(anon_vma);
3395     }
3396 }
3397 
3398 static void vm_unlock_mapping(struct address_space *mapping)
3399 {
3400     if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3401         /*
3402          * AS_MM_ALL_LOCKS can't change to 0 from under us
3403          * because we hold the mm_all_locks_mutex.
3404          */
3405         i_mmap_unlock_write(mapping);
3406         if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3407                     &mapping->flags))
3408             BUG();
3409     }
3410 }
3411 
3412 /*
3413  * The mmap_sem cannot be released by the caller until
3414  * mm_drop_all_locks() returns.
3415  */
3416 void mm_drop_all_locks(struct mm_struct *mm)
3417 {
3418     struct vm_area_struct *vma;
3419     struct anon_vma_chain *avc;
3420 
3421     BUG_ON(down_read_trylock(&mm->mmap_sem));
3422     BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3423 
3424     for (vma = mm->mmap; vma; vma = vma->vm_next) {
3425         if (vma->anon_vma)
3426             list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3427                 vm_unlock_anon_vma(avc->anon_vma);
3428         if (vma->vm_file && vma->vm_file->f_mapping)
3429             vm_unlock_mapping(vma->vm_file->f_mapping);
3430     }
3431 
3432     mutex_unlock(&mm_all_locks_mutex);
3433 }
3434 
3435 /*
3436  * initialise the VMA slab
3437  */
3438 void __init mmap_init(void)
3439 {
3440     int ret;
3441 
3442     ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3443     VM_BUG_ON(ret);
3444 }
3445 
3446 /*
3447  * Initialise sysctl_user_reserve_kbytes.
3448  *
3449  * This is intended to prevent a user from starting a single memory hogging
3450  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3451  * mode.
3452  *
3453  * The default value is min(3% of free memory, 128MB)
3454  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3455  */
3456 static int init_user_reserve(void)
3457 {
3458     unsigned long free_kbytes;
3459 
3460     free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3461 
3462     sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3463     return 0;
3464 }
3465 subsys_initcall(init_user_reserve);
3466 
3467 /*
3468  * Initialise sysctl_admin_reserve_kbytes.
3469  *
3470  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3471  * to log in and kill a memory hogging process.
3472  *
3473  * Systems with more than 256MB will reserve 8MB, enough to recover
3474  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3475  * only reserve 3% of free pages by default.
3476  */
3477 static int init_admin_reserve(void)
3478 {
3479     unsigned long free_kbytes;
3480 
3481     free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3482 
3483     sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3484     return 0;
3485 }
3486 subsys_initcall(init_admin_reserve);
3487 
3488 /*
3489  * Reinititalise user and admin reserves if memory is added or removed.
3490  *
3491  * The default user reserve max is 128MB, and the default max for the
3492  * admin reserve is 8MB. These are usually, but not always, enough to
3493  * enable recovery from a memory hogging process using login/sshd, a shell,
3494  * and tools like top. It may make sense to increase or even disable the
3495  * reserve depending on the existence of swap or variations in the recovery
3496  * tools. So, the admin may have changed them.
3497  *
3498  * If memory is added and the reserves have been eliminated or increased above
3499  * the default max, then we'll trust the admin.
3500  *
3501  * If memory is removed and there isn't enough free memory, then we
3502  * need to reset the reserves.
3503  *
3504  * Otherwise keep the reserve set by the admin.
3505  */
3506 static int reserve_mem_notifier(struct notifier_block *nb,
3507                  unsigned long action, void *data)
3508 {
3509     unsigned long tmp, free_kbytes;
3510 
3511     switch (action) {
3512     case MEM_ONLINE:
3513         /* Default max is 128MB. Leave alone if modified by operator. */
3514         tmp = sysctl_user_reserve_kbytes;
3515         if (0 < tmp && tmp < (1UL << 17))
3516             init_user_reserve();
3517 
3518         /* Default max is 8MB.  Leave alone if modified by operator. */
3519         tmp = sysctl_admin_reserve_kbytes;
3520         if (0 < tmp && tmp < (1UL << 13))
3521             init_admin_reserve();
3522 
3523         break;
3524     case MEM_OFFLINE:
3525         free_kbytes = global_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3526 
3527         if (sysctl_user_reserve_kbytes > free_kbytes) {
3528             init_user_reserve();
3529             pr_info("vm.user_reserve_kbytes reset to %lu\n",
3530                 sysctl_user_reserve_kbytes);
3531         }
3532 
3533         if (sysctl_admin_reserve_kbytes > free_kbytes) {
3534             init_admin_reserve();
3535             pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3536                 sysctl_admin_reserve_kbytes);
3537         }
3538         break;
3539     default:
3540         break;
3541     }
3542     return NOTIFY_OK;
3543 }
3544 
3545 static struct notifier_block reserve_mem_nb = {
3546     .notifier_call = reserve_mem_notifier,
3547 };
3548 
3549 static int __meminit init_reserve_notifier(void)
3550 {
3551     if (register_hotmemory_notifier(&reserve_mem_nb))
3552         pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3553 
3554     return 0;
3555 }
3556 subsys_initcall(init_reserve_notifier);