mm/mmap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * mm/mmap.c
   4  *
   5  * Written by obz.
   6  *
   7  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/kernel.h>
  13 #include <linux/slab.h>
  14 #include <linux/backing-dev.h>
  15 #include <linux/mm.h>
  16 #include <linux/vmacache.h>
  17 #include <linux/shm.h>
  18 #include <linux/mman.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/swap.h>
  21 #include <linux/syscalls.h>
  22 #include <linux/capability.h>
  23 #include <linux/init.h>
  24 #include <linux/file.h>
  25 #include <linux/fs.h>
  26 #include <linux/personality.h>
  27 #include <linux/security.h>
  28 #include <linux/hugetlb.h>
  29 #include <linux/shmem_fs.h>
  30 #include <linux/profile.h>
  31 #include <linux/export.h>
  32 #include <linux/mount.h>
  33 #include <linux/mempolicy.h>
  34 #include <linux/rmap.h>
  35 #include <linux/mmu_notifier.h>
  36 #include <linux/mmdebug.h>
  37 #include <linux/perf_event.h>
  38 #include <linux/audit.h>
  39 #include <linux/khugepaged.h>
  40 #include <linux/uprobes.h>
  41 #include <linux/rbtree_augmented.h>
  42 #include <linux/notifier.h>
  43 #include <linux/memory.h>
  44 #include <linux/printk.h>
  45 #include <linux/userfaultfd_k.h>
  46 #include <linux/moduleparam.h>
  47 #include <linux/pkeys.h>
  48 #include <linux/oom.h>
  49 #include <linux/sched/mm.h>
  50
  51 #include <linux/uaccess.h>
  52 #include <asm/cacheflush.h>
  53 #include <asm/tlb.h>
  54 #include <asm/mmu_context.h>
  55
  56 #define CREATE_TRACE_POINTS
  57 #include <trace/events/mmap.h>
  58
  59 #include "internal.h"
  60
  61 #ifndef arch_mmap_check
  62 #define arch_mmap_check(addr, len, flags)       (0)
  63 #endif
  64
  65 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  66 const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  67 const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  68 int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  69 #endif
  70 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  71 const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  72 const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  73 int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  74 #endif
  75
  76 static bool ignore_rlimit_data;
  77 core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  78
  79 static void unmap_region(struct mm_struct *mm,
  80                 struct vm_area_struct *vma, struct vm_area_struct *prev,
  81                 unsigned long start, unsigned long end);
  82
  83 /* description of effects of mapping type and prot in current implementation.
  84  * this is due to the limited x86 page protection hardware.  The expected
  85  * behavior is in parens:
  86  *
  87  * map_type     prot
  88  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
  89  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
  90  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
  91  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
  92  *
  93  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
  94  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
  95  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
  96  *
  97  * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
  98  * MAP_PRIVATE (with Enhanced PAN supported):
  99  *                                                              r: (no) no
 100  *                                                              w: (no) no
 101  *                                                              x: (yes) yes
 102  */
 103 pgprot_t protection_map[16] __ro_after_init = {
 104         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
 105         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
 106 };
 107
 108 #ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
 109 static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
 110 {
 111         return prot;
 112 }
 113 #endif
 114
 115 pgprot_t vm_get_page_prot(unsigned long vm_flags)
 116 {
 117         pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
 118                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
 119                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
 120
 121         return arch_filter_pgprot(ret);
 122 }
 123 EXPORT_SYMBOL(vm_get_page_prot);
 124
 125 static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
 126 {
 127         return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
 128 }
 129
 130 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
 131 void vma_set_page_prot(struct vm_area_struct *vma)
 132 {
 133         unsigned long vm_flags = vma->vm_flags;
 134         pgprot_t vm_page_prot;
 135
 136         vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
 137         if (vma_wants_writenotify(vma, vm_page_prot)) {
 138                 vm_flags &= ~VM_SHARED;
 139                 vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
 140         }
 141         /* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
 142         WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
 143 }
 144
 145 /*
 146  * Requires inode->i_mapping->i_mmap_rwsem
 147  */
 148 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 149                 struct file *file, struct address_space *mapping)
 150 {
 151         if (vma->vm_flags & VM_SHARED)
 152                 mapping_unmap_writable(mapping);
 153
 154         flush_dcache_mmap_lock(mapping);
 155         vma_interval_tree_remove(vma, &mapping->i_mmap);
 156         flush_dcache_mmap_unlock(mapping);
 157 }
 158
 159 /*
 160  * Unlink a file-based vm structure from its interval tree, to hide
 161  * vma from rmap and vmtruncate before freeing its page tables.
 162  */
 163 void unlink_file_vma(struct vm_area_struct *vma)
 164 {
 165         struct file *file = vma->vm_file;
 166
 167         if (file) {
 168                 struct address_space *mapping = file->f_mapping;
 169                 i_mmap_lock_write(mapping);
 170                 __remove_shared_vm_struct(vma, file, mapping);
 171                 i_mmap_unlock_write(mapping);
 172         }
 173 }
 174
 175 /*
 176  * Close a vm structure and free it, returning the next.
 177  */
 178 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 179 {
 180         struct vm_area_struct *next = vma->vm_next;
 181
 182         might_sleep();
 183         if (vma->vm_ops && vma->vm_ops->close)
 184                 vma->vm_ops->close(vma);
 185         if (vma->vm_file)
 186                 fput(vma->vm_file);
 187         mpol_put(vma_policy(vma));
 188         vm_area_free(vma);
 189         return next;
 190 }
 191
 192 static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags,
 193                 struct list_head *uf);
 194 SYSCALL_DEFINE1(brk, unsigned long, brk)
 195 {
 196         unsigned long newbrk, oldbrk, origbrk;
 197         struct mm_struct *mm = current->mm;
 198         struct vm_area_struct *next;
 199         unsigned long min_brk;
 200         bool populate;
 201         bool downgraded = false;
 202         LIST_HEAD(uf);
 203
 204         if (mmap_write_lock_killable(mm))
 205                 return -EINTR;
 206
 207         origbrk = mm->brk;
 208
 209 #ifdef CONFIG_COMPAT_BRK
 210         /*
 211          * CONFIG_COMPAT_BRK can still be overridden by setting
 212          * randomize_va_space to 2, which will still cause mm->start_brk
 213          * to be arbitrarily shifted
 214          */
 215         if (current->brk_randomized)
 216                 min_brk = mm->start_brk;
 217         else
 218                 min_brk = mm->end_data;
 219 #else
 220         min_brk = mm->start_brk;
 221 #endif
 222         if (brk < min_brk)
 223                 goto out;
 224
 225         /*
 226          * Check against rlimit here. If this check is done later after the test
 227          * of oldbrk with newbrk then it can escape the test and let the data
 228          * segment grow beyond its set limit the in case where the limit is
 229          * not page aligned -Ram Gupta
 230          */
 231         if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
 232                               mm->end_data, mm->start_data))
 233                 goto out;
 234
 235         newbrk = PAGE_ALIGN(brk);
 236         oldbrk = PAGE_ALIGN(mm->brk);
 237         if (oldbrk == newbrk) {
 238                 mm->brk = brk;
 239                 goto success;
 240         }
 241
 242         /*
 243          * Always allow shrinking brk.
 244          * __do_munmap() may downgrade mmap_lock to read.
 245          */
 246         if (brk <= mm->brk) {
 247                 int ret;
 248
 249                 /*
 250                  * mm->brk must to be protected by write mmap_lock so update it
 251                  * before downgrading mmap_lock. When __do_munmap() fails,
 252                  * mm->brk will be restored from origbrk.
 253                  */
 254                 mm->brk = brk;
 255                 ret = __do_munmap(mm, newbrk, oldbrk-newbrk, &uf, true);
 256                 if (ret < 0) {
 257                         mm->brk = origbrk;
 258                         goto out;
 259                 } else if (ret == 1) {
 260                         downgraded = true;
 261                 }
 262                 goto success;
 263         }
 264
 265         /* Check against existing mmap mappings. */
 266         next = find_vma(mm, oldbrk);
 267         if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
 268                 goto out;
 269
 270         /* Ok, looks good - let it rip. */
 271         if (do_brk_flags(oldbrk, newbrk-oldbrk, 0, &uf) < 0)
 272                 goto out;
 273         mm->brk = brk;
 274
 275 success:
 276         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 277         if (downgraded)
 278                 mmap_read_unlock(mm);
 279         else
 280                 mmap_write_unlock(mm);
 281         userfaultfd_unmap_complete(mm, &uf);
 282         if (populate)
 283                 mm_populate(oldbrk, newbrk - oldbrk);
 284         return brk;
 285
 286 out:
 287         mmap_write_unlock(mm);
 288         return origbrk;
 289 }
 290
 291 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
 292 {
 293         unsigned long gap, prev_end;
 294
 295         /*
 296          * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
 297          * allow two stack_guard_gaps between them here, and when choosing
 298          * an unmapped area; whereas when expanding we only require one.
 299          * That's a little inconsistent, but keeps the code here simpler.
 300          */
 301         gap = vm_start_gap(vma);
 302         if (vma->vm_prev) {
 303                 prev_end = vm_end_gap(vma->vm_prev);
 304                 if (gap > prev_end)
 305                         gap -= prev_end;
 306                 else
 307                         gap = 0;
 308         }
 309         return gap;
 310 }
 311
 312 #ifdef CONFIG_DEBUG_VM_RB
 313 static unsigned long vma_compute_subtree_gap(struct vm_area_struct *vma)
 314 {
 315         unsigned long max = vma_compute_gap(vma), subtree_gap;
 316         if (vma->vm_rb.rb_left) {
 317                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
 318                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
 319                 if (subtree_gap > max)
 320                         max = subtree_gap;
 321         }
 322         if (vma->vm_rb.rb_right) {
 323                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
 324                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
 325                 if (subtree_gap > max)
 326                         max = subtree_gap;
 327         }
 328         return max;
 329 }
 330
 331 static int browse_rb(struct mm_struct *mm)
 332 {
 333         struct rb_root *root = &mm->mm_rb;
 334         int i = 0, j, bug = 0;
 335         struct rb_node *nd, *pn = NULL;
 336         unsigned long prev = 0, pend = 0;
 337
 338         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 339                 struct vm_area_struct *vma;
 340                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 341                 if (vma->vm_start < prev) {
 342                         pr_emerg("vm_start %lx < prev %lx\n",
 343                                   vma->vm_start, prev);
 344                         bug = 1;
 345                 }
 346                 if (vma->vm_start < pend) {
 347                         pr_emerg("vm_start %lx < pend %lx\n",
 348                                   vma->vm_start, pend);
 349                         bug = 1;
 350                 }
 351                 if (vma->vm_start > vma->vm_end) {
 352                         pr_emerg("vm_start %lx > vm_end %lx\n",
 353                                   vma->vm_start, vma->vm_end);
 354                         bug = 1;
 355                 }
 356                 spin_lock(&mm->page_table_lock);
 357                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
 358                         pr_emerg("free gap %lx, correct %lx\n",
 359                                vma->rb_subtree_gap,
 360                                vma_compute_subtree_gap(vma));
 361                         bug = 1;
 362                 }
 363                 spin_unlock(&mm->page_table_lock);
 364                 i++;
 365                 pn = nd;
 366                 prev = vma->vm_start;
 367                 pend = vma->vm_end;
 368         }
 369         j = 0;
 370         for (nd = pn; nd; nd = rb_prev(nd))
 371                 j++;
 372         if (i != j) {
 373                 pr_emerg("backwards %d, forwards %d\n", j, i);
 374                 bug = 1;
 375         }
 376         return bug ? -1 : i;
 377 }
 378
 379 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
 380 {
 381         struct rb_node *nd;
 382
 383         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 384                 struct vm_area_struct *vma;
 385                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 386                 VM_BUG_ON_VMA(vma != ignore &&
 387                         vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
 388                         vma);
 389         }
 390 }
 391
 392 static void validate_mm(struct mm_struct *mm)
 393 {
 394         int bug = 0;
 395         int i = 0;
 396         unsigned long highest_address = 0;
 397         struct vm_area_struct *vma = mm->mmap;
 398
 399         while (vma) {
 400                 struct anon_vma *anon_vma = vma->anon_vma;
 401                 struct anon_vma_chain *avc;
 402
 403                 if (anon_vma) {
 404                         anon_vma_lock_read(anon_vma);
 405                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 406                                 anon_vma_interval_tree_verify(avc);
 407                         anon_vma_unlock_read(anon_vma);
 408                 }
 409
 410                 highest_address = vm_end_gap(vma);
 411                 vma = vma->vm_next;
 412                 i++;
 413         }
 414         if (i != mm->map_count) {
 415                 pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
 416                 bug = 1;
 417         }
 418         if (highest_address != mm->highest_vm_end) {
 419                 pr_emerg("mm->highest_vm_end %lx, found %lx\n",
 420                           mm->highest_vm_end, highest_address);
 421                 bug = 1;
 422         }
 423         i = browse_rb(mm);
 424         if (i != mm->map_count) {
 425                 if (i != -1)
 426                         pr_emerg("map_count %d rb %d\n", mm->map_count, i);
 427                 bug = 1;
 428         }
 429         VM_BUG_ON_MM(bug, mm);
 430 }
 431 #else
 432 #define validate_mm_rb(root, ignore) do { } while (0)
 433 #define validate_mm(mm) do { } while (0)
 434 #endif
 435
 436 RB_DECLARE_CALLBACKS_MAX(static, vma_gap_callbacks,
 437                          struct vm_area_struct, vm_rb,
 438                          unsigned long, rb_subtree_gap, vma_compute_gap)
 439
 440 /*
 441  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
 442  * vma->vm_prev->vm_end values changed, without modifying the vma's position
 443  * in the rbtree.
 444  */
 445 static void vma_gap_update(struct vm_area_struct *vma)
 446 {
 447         /*
 448          * As it turns out, RB_DECLARE_CALLBACKS_MAX() already created
 449          * a callback function that does exactly what we want.
 450          */
 451         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
 452 }
 453
 454 static inline void vma_rb_insert(struct vm_area_struct *vma,
 455                                  struct rb_root *root)
 456 {
 457         /* All rb_subtree_gap values must be consistent prior to insertion */
 458         validate_mm_rb(root, NULL);
 459
 460         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 461 }
 462
 463 static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
 464 {
 465         /*
 466          * Note rb_erase_augmented is a fairly large inline function,
 467          * so make sure we instantiate it only once with our desired
 468          * augmented rbtree callbacks.
 469          */
 470         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 471 }
 472
 473 static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
 474                                                 struct rb_root *root,
 475                                                 struct vm_area_struct *ignore)
 476 {
 477         /*
 478          * All rb_subtree_gap values must be consistent prior to erase,
 479          * with the possible exception of
 480          *
 481          * a. the "next" vma being erased if next->vm_start was reduced in
 482          *    __vma_adjust() -> __vma_unlink()
 483          * b. the vma being erased in detach_vmas_to_be_unmapped() ->
 484          *    vma_rb_erase()
 485          */
 486         validate_mm_rb(root, ignore);
 487
 488         __vma_rb_erase(vma, root);
 489 }
 490
 491 static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
 492                                          struct rb_root *root)
 493 {
 494         vma_rb_erase_ignore(vma, root, vma);
 495 }
 496
 497 /*
 498  * vma has some anon_vma assigned, and is already inserted on that
 499  * anon_vma's interval trees.
 500  *
 501  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 502  * vma must be removed from the anon_vma's interval trees using
 503  * anon_vma_interval_tree_pre_update_vma().
 504  *
 505  * After the update, the vma will be reinserted using
 506  * anon_vma_interval_tree_post_update_vma().
 507  *
 508  * The entire update must be protected by exclusive mmap_lock and by
 509  * the root anon_vma's mutex.
 510  */
 511 static inline void
 512 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 513 {
 514         struct anon_vma_chain *avc;
 515
 516         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 517                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 518 }
 519
 520 static inline void
 521 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 522 {
 523         struct anon_vma_chain *avc;
 524
 525         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 526                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 527 }
 528
 529 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
 530                 unsigned long end, struct vm_area_struct **pprev,
 531                 struct rb_node ***rb_link, struct rb_node **rb_parent)
 532 {
 533         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
 534
 535         mmap_assert_locked(mm);
 536         __rb_link = &mm->mm_rb.rb_node;
 537         rb_prev = __rb_parent = NULL;
 538
 539         while (*__rb_link) {
 540                 struct vm_area_struct *vma_tmp;
 541
 542                 __rb_parent = *__rb_link;
 543                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 544
 545                 if (vma_tmp->vm_end > addr) {
 546                         /* Fail if an existing vma overlaps the area */
 547                         if (vma_tmp->vm_start < end)
 548                                 return -ENOMEM;
 549                         __rb_link = &__rb_parent->rb_left;
 550                 } else {
 551                         rb_prev = __rb_parent;
 552                         __rb_link = &__rb_parent->rb_right;
 553                 }
 554         }
 555
 556         *pprev = NULL;
 557         if (rb_prev)
 558                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 559         *rb_link = __rb_link;
 560         *rb_parent = __rb_parent;
 561         return 0;
 562 }
 563
 564 /*
 565  * vma_next() - Get the next VMA.
 566  * @mm: The mm_struct.
 567  * @vma: The current vma.
 568  *
 569  * If @vma is NULL, return the first vma in the mm.
 570  *
 571  * Returns: The next VMA after @vma.
 572  */
 573 static inline struct vm_area_struct *vma_next(struct mm_struct *mm,
 574                                          struct vm_area_struct *vma)
 575 {
 576         if (!vma)
 577                 return mm->mmap;
 578
 579         return vma->vm_next;
 580 }
 581
 582 /*
 583  * munmap_vma_range() - munmap VMAs that overlap a range.
 584  * @mm: The mm struct
 585  * @start: The start of the range.
 586  * @len: The length of the range.
 587  * @pprev: pointer to the pointer that will be set to previous vm_area_struct
 588  * @rb_link: the rb_node
 589  * @rb_parent: the parent rb_node
 590  *
 591  * Find all the vm_area_struct that overlap from @start to
 592  * @end and munmap them.  Set @pprev to the previous vm_area_struct.
 593  *
 594  * Returns: -ENOMEM on munmap failure or 0 on success.
 595  */
 596 static inline int
 597 munmap_vma_range(struct mm_struct *mm, unsigned long start, unsigned long len,
 598                  struct vm_area_struct **pprev, struct rb_node ***link,
 599                  struct rb_node **parent, struct list_head *uf)
 600 {
 601
 602         while (find_vma_links(mm, start, start + len, pprev, link, parent))
 603                 if (do_munmap(mm, start, len, uf))
 604                         return -ENOMEM;
 605
 606         return 0;
 607 }
 608 static unsigned long count_vma_pages_range(struct mm_struct *mm,
 609                 unsigned long addr, unsigned long end)
 610 {
 611         unsigned long nr_pages = 0;
 612         struct vm_area_struct *vma;
 613
 614         /* Find first overlapping mapping */
 615         vma = find_vma_intersection(mm, addr, end);
 616         if (!vma)
 617                 return 0;
 618
 619         nr_pages = (min(end, vma->vm_end) -
 620                 max(addr, vma->vm_start)) >> PAGE_SHIFT;
 621
 622         /* Iterate over the rest of the overlaps */
 623         for (vma = vma->vm_next; vma; vma = vma->vm_next) {
 624                 unsigned long overlap_len;
 625
 626                 if (vma->vm_start > end)
 627                         break;
 628
 629                 overlap_len = min(end, vma->vm_end) - vma->vm_start;
 630                 nr_pages += overlap_len >> PAGE_SHIFT;
 631         }
 632
 633         return nr_pages;
 634 }
 635
 636 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 637                 struct rb_node **rb_link, struct rb_node *rb_parent)
 638 {
 639         /* Update tracking information for the gap following the new vma. */
 640         if (vma->vm_next)
 641                 vma_gap_update(vma->vm_next);
 642         else
 643                 mm->highest_vm_end = vm_end_gap(vma);
 644
 645         /*
 646          * vma->vm_prev wasn't known when we followed the rbtree to find the
 647          * correct insertion point for that vma. As a result, we could not
 648          * update the vma vm_rb parents rb_subtree_gap values on the way down.
 649          * So, we first insert the vma with a zero rb_subtree_gap value
 650          * (to be consistent with what we did on the way down), and then
 651          * immediately update the gap to the correct value. Finally we
 652          * rebalance the rbtree after all augmented values have been set.
 653          */
 654         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 655         vma->rb_subtree_gap = 0;
 656         vma_gap_update(vma);
 657         vma_rb_insert(vma, &mm->mm_rb);
 658 }
 659
 660 static void __vma_link_file(struct vm_area_struct *vma)
 661 {
 662         struct file *file;
 663
 664         file = vma->vm_file;
 665         if (file) {
 666                 struct address_space *mapping = file->f_mapping;
 667
 668                 if (vma->vm_flags & VM_SHARED)
 669                         mapping_allow_writable(mapping);
 670
 671                 flush_dcache_mmap_lock(mapping);
 672                 vma_interval_tree_insert(vma, &mapping->i_mmap);
 673                 flush_dcache_mmap_unlock(mapping);
 674         }
 675 }
 676
 677 static void
 678 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 679         struct vm_area_struct *prev, struct rb_node **rb_link,
 680         struct rb_node *rb_parent)
 681 {
 682         __vma_link_list(mm, vma, prev);
 683         __vma_link_rb(mm, vma, rb_link, rb_parent);
 684 }
 685
 686 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 687                         struct vm_area_struct *prev, struct rb_node **rb_link,
 688                         struct rb_node *rb_parent)
 689 {
 690         struct address_space *mapping = NULL;
 691
 692         if (vma->vm_file) {
 693                 mapping = vma->vm_file->f_mapping;
 694                 i_mmap_lock_write(mapping);
 695         }
 696
 697         __vma_link(mm, vma, prev, rb_link, rb_parent);
 698         __vma_link_file(vma);
 699
 700         if (mapping)
 701                 i_mmap_unlock_write(mapping);
 702
 703         mm->map_count++;
 704         validate_mm(mm);
 705 }
 706
 707 /*
 708  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
 709  * mm's list and rbtree.  It has already been inserted into the interval tree.
 710  */
 711 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 712 {
 713         struct vm_area_struct *prev;
 714         struct rb_node **rb_link, *rb_parent;
 715
 716         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
 717                            &prev, &rb_link, &rb_parent))
 718                 BUG();
 719         __vma_link(mm, vma, prev, rb_link, rb_parent);
 720         mm->map_count++;
 721 }
 722
 723 static __always_inline void __vma_unlink(struct mm_struct *mm,
 724                                                 struct vm_area_struct *vma,
 725                                                 struct vm_area_struct *ignore)
 726 {
 727         vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
 728         __vma_unlink_list(mm, vma);
 729         /* Kill the cache */
 730         vmacache_invalidate(mm);
 731 }
 732
 733 /*
 734  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 735  * is already present in an i_mmap tree without adjusting the tree.
 736  * The following helper function should be used when such adjustments
 737  * are necessary.  The "insert" vma (if any) is to be inserted
 738  * before we drop the necessary locks.
 739  */
 740 int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
 741         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
 742         struct vm_area_struct *expand)
 743 {
 744         struct mm_struct *mm = vma->vm_mm;
 745         struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
 746         struct address_space *mapping = NULL;
 747         struct rb_root_cached *root = NULL;
 748         struct anon_vma *anon_vma = NULL;
 749         struct file *file = vma->vm_file;
 750         bool start_changed = false, end_changed = false;
 751         long adjust_next = 0;
 752         int remove_next = 0;
 753
 754         if (next && !insert) {
 755                 struct vm_area_struct *exporter = NULL, *importer = NULL;
 756
 757                 if (end >= next->vm_end) {
 758                         /*
 759                          * vma expands, overlapping all the next, and
 760                          * perhaps the one after too (mprotect case 6).
 761                          * The only other cases that gets here are
 762                          * case 1, case 7 and case 8.
 763                          */
 764                         if (next == expand) {
 765                                 /*
 766                                  * The only case where we don't expand "vma"
 767                                  * and we expand "next" instead is case 8.
 768                                  */
 769                                 VM_WARN_ON(end != next->vm_end);
 770                                 /*
 771                                  * remove_next == 3 means we're
 772                                  * removing "vma" and that to do so we
 773                                  * swapped "vma" and "next".
 774                                  */
 775                                 remove_next = 3;
 776                                 VM_WARN_ON(file != next->vm_file);
 777                                 swap(vma, next);
 778                         } else {
 779                                 VM_WARN_ON(expand != vma);
 780                                 /*
 781                                  * case 1, 6, 7, remove_next == 2 is case 6,
 782                                  * remove_next == 1 is case 1 or 7.
 783                                  */
 784                                 remove_next = 1 + (end > next->vm_end);
 785                                 VM_WARN_ON(remove_next == 2 &&
 786                                            end != next->vm_next->vm_end);
 787                                 /* trim end to next, for case 6 first pass */
 788                                 end = next->vm_end;
 789                         }
 790
 791                         exporter = next;
 792                         importer = vma;
 793
 794                         /*
 795                          * If next doesn't have anon_vma, import from vma after
 796                          * next, if the vma overlaps with it.
 797                          */
 798                         if (remove_next == 2 && !next->anon_vma)
 799                                 exporter = next->vm_next;
 800
 801                 } else if (end > next->vm_start) {
 802                         /*
 803                          * vma expands, overlapping part of the next:
 804                          * mprotect case 5 shifting the boundary up.
 805                          */
 806                         adjust_next = (end - next->vm_start);
 807                         exporter = next;
 808                         importer = vma;
 809                         VM_WARN_ON(expand != importer);
 810                 } else if (end < vma->vm_end) {
 811                         /*
 812                          * vma shrinks, and !insert tells it's not
 813                          * split_vma inserting another: so it must be
 814                          * mprotect case 4 shifting the boundary down.
 815                          */
 816                         adjust_next = -(vma->vm_end - end);
 817                         exporter = vma;
 818                         importer = next;
 819                         VM_WARN_ON(expand != importer);
 820                 }
 821
 822                 /*
 823                  * Easily overlooked: when mprotect shifts the boundary,
 824                  * make sure the expanding vma has anon_vma set if the
 825                  * shrinking vma had, to cover any anon pages imported.
 826                  */
 827                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
 828                         int error;
 829
 830                         importer->anon_vma = exporter->anon_vma;
 831                         error = anon_vma_clone(importer, exporter);
 832                         if (error)
 833                                 return error;
 834                 }
 835         }
 836 again:
 837         vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
 838
 839         if (file) {
 840                 mapping = file->f_mapping;
 841                 root = &mapping->i_mmap;
 842                 uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 843
 844                 if (adjust_next)
 845                         uprobe_munmap(next, next->vm_start, next->vm_end);
 846
 847                 i_mmap_lock_write(mapping);
 848                 if (insert) {
 849                         /*
 850                          * Put into interval tree now, so instantiated pages
 851                          * are visible to arm/parisc __flush_dcache_page
 852                          * throughout; but we cannot insert into address
 853                          * space until vma start or end is updated.
 854                          */
 855                         __vma_link_file(insert);
 856                 }
 857         }
 858
 859         anon_vma = vma->anon_vma;
 860         if (!anon_vma && adjust_next)
 861                 anon_vma = next->anon_vma;
 862         if (anon_vma) {
 863                 VM_WARN_ON(adjust_next && next->anon_vma &&
 864                            anon_vma != next->anon_vma);
 865                 anon_vma_lock_write(anon_vma);
 866                 anon_vma_interval_tree_pre_update_vma(vma);
 867                 if (adjust_next)
 868                         anon_vma_interval_tree_pre_update_vma(next);
 869         }
 870
 871         if (file) {
 872                 flush_dcache_mmap_lock(mapping);
 873                 vma_interval_tree_remove(vma, root);
 874                 if (adjust_next)
 875                         vma_interval_tree_remove(next, root);
 876         }
 877
 878         if (start != vma->vm_start) {
 879                 vma->vm_start = start;
 880                 start_changed = true;
 881         }
 882         if (end != vma->vm_end) {
 883                 vma->vm_end = end;
 884                 end_changed = true;
 885         }
 886         vma->vm_pgoff = pgoff;
 887         if (adjust_next) {
 888                 next->vm_start += adjust_next;
 889                 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
 890         }
 891
 892         if (file) {
 893                 if (adjust_next)
 894                         vma_interval_tree_insert(next, root);
 895                 vma_interval_tree_insert(vma, root);
 896                 flush_dcache_mmap_unlock(mapping);
 897         }
 898
 899         if (remove_next) {
 900                 /*
 901                  * vma_merge has merged next into vma, and needs
 902                  * us to remove next before dropping the locks.
 903                  */
 904                 if (remove_next != 3)
 905                         __vma_unlink(mm, next, next);
 906                 else
 907                         /*
 908                          * vma is not before next if they've been
 909                          * swapped.
 910                          *
 911                          * pre-swap() next->vm_start was reduced so
 912                          * tell validate_mm_rb to ignore pre-swap()
 913                          * "next" (which is stored in post-swap()
 914                          * "vma").
 915                          */
 916                         __vma_unlink(mm, next, vma);
 917                 if (file)
 918                         __remove_shared_vm_struct(next, file, mapping);
 919         } else if (insert) {
 920                 /*
 921                  * split_vma has split insert from vma, and needs
 922                  * us to insert it before dropping the locks
 923                  * (it may either follow vma or precede it).
 924                  */
 925                 __insert_vm_struct(mm, insert);
 926         } else {
 927                 if (start_changed)
 928                         vma_gap_update(vma);
 929                 if (end_changed) {
 930                         if (!next)
 931                                 mm->highest_vm_end = vm_end_gap(vma);
 932                         else if (!adjust_next)
 933                                 vma_gap_update(next);
 934                 }
 935         }
 936
 937         if (anon_vma) {
 938                 anon_vma_interval_tree_post_update_vma(vma);
 939                 if (adjust_next)
 940                         anon_vma_interval_tree_post_update_vma(next);
 941                 anon_vma_unlock_write(anon_vma);
 942         }
 943
 944         if (file) {
 945                 i_mmap_unlock_write(mapping);
 946                 uprobe_mmap(vma);
 947
 948                 if (adjust_next)
 949                         uprobe_mmap(next);
 950         }
 951
 952         if (remove_next) {
 953                 if (file) {
 954                         uprobe_munmap(next, next->vm_start, next->vm_end);
 955                         fput(file);
 956                 }
 957                 if (next->anon_vma)
 958                         anon_vma_merge(vma, next);
 959                 mm->map_count--;
 960                 mpol_put(vma_policy(next));
 961                 vm_area_free(next);
 962                 /*
 963                  * In mprotect's case 6 (see comments on vma_merge),
 964                  * we must remove another next too. It would clutter
 965                  * up the code too much to do both in one go.
 966                  */
 967                 if (remove_next != 3) {
 968                         /*
 969                          * If "next" was removed and vma->vm_end was
 970                          * expanded (up) over it, in turn
 971                          * "next->vm_prev->vm_end" changed and the
 972                          * "vma->vm_next" gap must be updated.
 973                          */
 974                         next = vma->vm_next;
 975                 } else {
 976                         /*
 977                          * For the scope of the comment "next" and
 978                          * "vma" considered pre-swap(): if "vma" was
 979                          * removed, next->vm_start was expanded (down)
 980                          * over it and the "next" gap must be updated.
 981                          * Because of the swap() the post-swap() "vma"
 982                          * actually points to pre-swap() "next"
 983                          * (post-swap() "next" as opposed is now a
 984                          * dangling pointer).
 985                          */
 986                         next = vma;
 987                 }
 988                 if (remove_next == 2) {
 989                         remove_next = 1;
 990                         end = next->vm_end;
 991                         goto again;
 992                 }
 993                 else if (next)
 994                         vma_gap_update(next);
 995                 else {
 996                         /*
 997                          * If remove_next == 2 we obviously can't
 998                          * reach this path.
 999                          *
1000                          * If remove_next == 3 we can't reach this
1001                          * path because pre-swap() next is always not
1002                          * NULL. pre-swap() "next" is not being
1003                          * removed and its next->vm_end is not altered
1004                          * (and furthermore "end" already matches
1005                          * next->vm_end in remove_next == 3).
1006                          *
1007                          * We reach this only in the remove_next == 1
1008                          * case if the "next" vma that was removed was
1009                          * the highest vma of the mm. However in such
1010                          * case next->vm_end == "end" and the extended
1011                          * "vma" has vma->vm_end == next->vm_end so
1012                          * mm->highest_vm_end doesn't need any update
1013                          * in remove_next == 1 case.
1014                          */
1015                         VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
1016                 }
1017         }
1018         if (insert && file)
1019                 uprobe_mmap(insert);
1020
1021         validate_mm(mm);
1022
1023         return 0;
1024 }
1025
1026 /*
1027  * If the vma has a ->close operation then the driver probably needs to release
1028  * per-vma resources, so we don't attempt to merge those.
1029  */
1030 static inline int is_mergeable_vma(struct vm_area_struct *vma,
1031                                 struct file *file, unsigned long vm_flags,
1032                                 struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1033 {
1034         /*
1035          * VM_SOFTDIRTY should not prevent from VMA merging, if we
1036          * match the flags but dirty bit -- the caller should mark
1037          * merged VMA as dirty. If dirty bit won't be excluded from
1038          * comparison, we increase pressure on the memory system forcing
1039          * the kernel to generate new VMAs when old one could be
1040          * extended instead.
1041          */
1042         if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
1043                 return 0;
1044         if (vma->vm_file != file)
1045                 return 0;
1046         if (vma->vm_ops && vma->vm_ops->close)
1047                 return 0;
1048         if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1049                 return 0;
1050         return 1;
1051 }
1052
1053 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1054                                         struct anon_vma *anon_vma2,
1055                                         struct vm_area_struct *vma)
1056 {
1057         /*
1058          * The list_is_singular() test is to avoid merging VMA cloned from
1059          * parents. This can improve scalability caused by anon_vma lock.
1060          */
1061         if ((!anon_vma1 || !anon_vma2) && (!vma ||
1062                 list_is_singular(&vma->anon_vma_chain)))
1063                 return 1;
1064         return anon_vma1 == anon_vma2;
1065 }
1066
1067 /*
1068  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1069  * in front of (at a lower virtual address and file offset than) the vma.
1070  *
1071  * We cannot merge two vmas if they have differently assigned (non-NULL)
1072  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1073  *
1074  * We don't check here for the merged mmap wrapping around the end of pagecache
1075  * indices (16TB on ia32) because do_mmap() does not permit mmap's which
1076  * wrap, nor mmaps which cover the final page at index -1UL.
1077  */
1078 static int
1079 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1080                      struct anon_vma *anon_vma, struct file *file,
1081                      pgoff_t vm_pgoff,
1082                      struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1083 {
1084         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1085             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1086                 if (vma->vm_pgoff == vm_pgoff)
1087                         return 1;
1088         }
1089         return 0;
1090 }
1091
1092 /*
1093  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1094  * beyond (at a higher virtual address and file offset than) the vma.
1095  *
1096  * We cannot merge two vmas if they have differently assigned (non-NULL)
1097  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1098  */
1099 static int
1100 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1101                     struct anon_vma *anon_vma, struct file *file,
1102                     pgoff_t vm_pgoff,
1103                     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1104 {
1105         if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1106             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1107                 pgoff_t vm_pglen;
1108                 vm_pglen = vma_pages(vma);
1109                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1110                         return 1;
1111         }
1112         return 0;
1113 }
1114
1115 /*
1116  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1117  * whether that can be merged with its predecessor or its successor.
1118  * Or both (it neatly fills a hole).
1119  *
1120  * In most cases - when called for mmap, brk or mremap - [addr,end) is
1121  * certain not to be mapped by the time vma_merge is called; but when
1122  * called for mprotect, it is certain to be already mapped (either at
1123  * an offset within prev, or at the start of next), and the flags of
1124  * this area are about to be changed to vm_flags - and the no-change
1125  * case has already been eliminated.
1126  *
1127  * The following mprotect cases have to be considered, where AAAA is
1128  * the area passed down from mprotect_fixup, never extending beyond one
1129  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1130  *
1131  *     AAAA             AAAA                   AAAA
1132  *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
1133  *    cannot merge    might become       might become
1134  *                    PPNNNNNNNNNN       PPPPPPPPPPNN
1135  *    mmap, brk or    case 4 below       case 5 below
1136  *    mremap move:
1137  *                        AAAA               AAAA
1138  *                    PPPP    NNNN       PPPPNNNNXXXX
1139  *                    might become       might become
1140  *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
1141  *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
1142  *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
1143  *
1144  * It is important for case 8 that the vma NNNN overlapping the
1145  * region AAAA is never going to extended over XXXX. Instead XXXX must
1146  * be extended in region AAAA and NNNN must be removed. This way in
1147  * all cases where vma_merge succeeds, the moment vma_adjust drops the
1148  * rmap_locks, the properties of the merged vma will be already
1149  * correct for the whole merged range. Some of those properties like
1150  * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1151  * be correct for the whole merged range immediately after the
1152  * rmap_locks are released. Otherwise if XXXX would be removed and
1153  * NNNN would be extended over the XXXX range, remove_migration_ptes
1154  * or other rmap walkers (if working on addresses beyond the "end"
1155  * parameter) may establish ptes with the wrong permissions of NNNN
1156  * instead of the right permissions of XXXX.
1157  */
1158 struct vm_area_struct *vma_merge(struct mm_struct *mm,
1159                         struct vm_area_struct *prev, unsigned long addr,
1160                         unsigned long end, unsigned long vm_flags,
1161                         struct anon_vma *anon_vma, struct file *file,
1162                         pgoff_t pgoff, struct mempolicy *policy,
1163                         struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
1164 {
1165         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1166         struct vm_area_struct *area, *next;
1167         int err;
1168
1169         /*
1170          * We later require that vma->vm_flags == vm_flags,
1171          * so this tests vma->vm_flags & VM_SPECIAL, too.
1172          */
1173         if (vm_flags & VM_SPECIAL)
1174                 return NULL;
1175
1176         next = vma_next(mm, prev);
1177         area = next;
1178         if (area && area->vm_end == end)                /* cases 6, 7, 8 */
1179                 next = next->vm_next;
1180
1181         /* verify some invariant that must be enforced by the caller */
1182         VM_WARN_ON(prev && addr <= prev->vm_start);
1183         VM_WARN_ON(area && end > area->vm_end);
1184         VM_WARN_ON(addr >= end);
1185
1186         /*
1187          * Can it merge with the predecessor?
1188          */
1189         if (prev && prev->vm_end == addr &&
1190                         mpol_equal(vma_policy(prev), policy) &&
1191                         can_vma_merge_after(prev, vm_flags,
1192                                             anon_vma, file, pgoff,
1193                                             vm_userfaultfd_ctx)) {
1194                 /*
1195                  * OK, it can.  Can we now merge in the successor as well?
1196                  */
1197                 if (next && end == next->vm_start &&
1198                                 mpol_equal(policy, vma_policy(next)) &&
1199                                 can_vma_merge_before(next, vm_flags,
1200                                                      anon_vma, file,
1201                                                      pgoff+pglen,
1202                                                      vm_userfaultfd_ctx) &&
1203                                 is_mergeable_anon_vma(prev->anon_vma,
1204                                                       next->anon_vma, NULL)) {
1205                                                         /* cases 1, 6 */
1206                         err = __vma_adjust(prev, prev->vm_start,
1207                                          next->vm_end, prev->vm_pgoff, NULL,
1208                                          prev);
1209                 } else                                  /* cases 2, 5, 7 */
1210                         err = __vma_adjust(prev, prev->vm_start,
1211                                          end, prev->vm_pgoff, NULL, prev);
1212                 if (err)
1213                         return NULL;
1214                 khugepaged_enter_vma_merge(prev, vm_flags);
1215                 return prev;
1216         }
1217
1218         /*
1219          * Can this new request be merged in front of next?
1220          */
1221         if (next && end == next->vm_start &&
1222                         mpol_equal(policy, vma_policy(next)) &&
1223                         can_vma_merge_before(next, vm_flags,
1224                                              anon_vma, file, pgoff+pglen,
1225                                              vm_userfaultfd_ctx)) {
1226                 if (prev && addr < prev->vm_end)        /* case 4 */
1227                         err = __vma_adjust(prev, prev->vm_start,
1228                                          addr, prev->vm_pgoff, NULL, next);
1229                 else {                                  /* cases 3, 8 */
1230                         err = __vma_adjust(area, addr, next->vm_end,
1231                                          next->vm_pgoff - pglen, NULL, next);
1232                         /*
1233                          * In case 3 area is already equal to next and
1234                          * this is a noop, but in case 8 "area" has
1235                          * been removed and next was expanded over it.
1236                          */
1237                         area = next;
1238                 }
1239                 if (err)
1240                         return NULL;
1241                 khugepaged_enter_vma_merge(area, vm_flags);
1242                 return area;
1243         }
1244
1245         return NULL;
1246 }
1247
1248 /*
1249  * Rough compatibility check to quickly see if it's even worth looking
1250  * at sharing an anon_vma.
1251  *
1252  * They need to have the same vm_file, and the flags can only differ
1253  * in things that mprotect may change.
1254  *
1255  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1256  * we can merge the two vma's. For example, we refuse to merge a vma if
1257  * there is a vm_ops->close() function, because that indicates that the
1258  * driver is doing some kind of reference counting. But that doesn't
1259  * really matter for the anon_vma sharing case.
1260  */
1261 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1262 {
1263         return a->vm_end == b->vm_start &&
1264                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1265                 a->vm_file == b->vm_file &&
1266                 !((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1267                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1268 }
1269
1270 /*
1271  * Do some basic sanity checking to see if we can re-use the anon_vma
1272  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1273  * the same as 'old', the other will be the new one that is trying
1274  * to share the anon_vma.
1275  *
1276  * NOTE! This runs with mm_sem held for reading, so it is possible that
1277  * the anon_vma of 'old' is concurrently in the process of being set up
1278  * by another page fault trying to merge _that_. But that's ok: if it
1279  * is being set up, that automatically means that it will be a singleton
1280  * acceptable for merging, so we can do all of this optimistically. But
1281  * we do that READ_ONCE() to make sure that we never re-load the pointer.
1282  *
1283  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1284  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1285  * is to return an anon_vma that is "complex" due to having gone through
1286  * a fork).
1287  *
1288  * We also make sure that the two vma's are compatible (adjacent,
1289  * and with the same memory policies). That's all stable, even with just
1290  * a read lock on the mm_sem.
1291  */
1292 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1293 {
1294         if (anon_vma_compatible(a, b)) {
1295                 struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1296
1297                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1298                         return anon_vma;
1299         }
1300         return NULL;
1301 }
1302
1303 /*
1304  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1305  * neighbouring vmas for a suitable anon_vma, before it goes off
1306  * to allocate a new anon_vma.  It checks because a repetitive
1307  * sequence of mprotects and faults may otherwise lead to distinct
1308  * anon_vmas being allocated, preventing vma merge in subsequent
1309  * mprotect.
1310  */
1311 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1312 {
1313         struct anon_vma *anon_vma = NULL;
1314
1315         /* Try next first. */
1316         if (vma->vm_next) {
1317                 anon_vma = reusable_anon_vma(vma->vm_next, vma, vma->vm_next);
1318                 if (anon_vma)
1319                         return anon_vma;
1320         }
1321
1322         /* Try prev next. */
1323         if (vma->vm_prev)
1324                 anon_vma = reusable_anon_vma(vma->vm_prev, vma->vm_prev, vma);
1325
1326         /*
1327          * We might reach here with anon_vma == NULL if we can't find
1328          * any reusable anon_vma.
1329          * There's no absolute need to look only at touching neighbours:
1330          * we could search further afield for "compatible" anon_vmas.
1331          * But it would probably just be a waste of time searching,
1332          * or lead to too many vmas hanging off the same anon_vma.
1333          * We're trying to allow mprotect remerging later on,
1334          * not trying to minimize memory used for anon_vmas.
1335          */
1336         return anon_vma;
1337 }
1338
1339 /*
1340  * If a hint addr is less than mmap_min_addr change hint to be as
1341  * low as possible but still greater than mmap_min_addr
1342  */
1343 static inline unsigned long round_hint_to_min(unsigned long hint)
1344 {
1345         hint &= PAGE_MASK;
1346         if (((void *)hint != NULL) &&
1347             (hint < mmap_min_addr))
1348                 return PAGE_ALIGN(mmap_min_addr);
1349         return hint;
1350 }
1351
1352 int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1353                        unsigned long len)
1354 {
1355         unsigned long locked, lock_limit;
1356
1357         /*  mlock MCL_FUTURE? */
1358         if (flags & VM_LOCKED) {
1359                 locked = len >> PAGE_SHIFT;
1360                 locked += mm->locked_vm;
1361                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1362                 lock_limit >>= PAGE_SHIFT;
1363                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1364                         return -EAGAIN;
1365         }
1366         return 0;
1367 }
1368
1369 static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1370 {
1371         if (S_ISREG(inode->i_mode))
1372                 return MAX_LFS_FILESIZE;
1373
1374         if (S_ISBLK(inode->i_mode))
1375                 return MAX_LFS_FILESIZE;
1376
1377         if (S_ISSOCK(inode->i_mode))
1378                 return MAX_LFS_FILESIZE;
1379
1380         /* Special "we do even unsigned file positions" case */
1381         if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1382                 return 0;
1383
1384         /* Yes, random drivers might want more. But I'm tired of buggy drivers */
1385         return ULONG_MAX;
1386 }
1387
1388 static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1389                                 unsigned long pgoff, unsigned long len)
1390 {
1391         u64 maxsize = file_mmap_size_max(file, inode);
1392
1393         if (maxsize && len > maxsize)
1394                 return false;
1395         maxsize -= len;
1396         if (pgoff > maxsize >> PAGE_SHIFT)
1397                 return false;
1398         return true;
1399 }
1400
1401 /*
1402  * The caller must write-lock current->mm->mmap_lock.
1403  */
1404 unsigned long do_mmap(struct file *file, unsigned long addr,
1405                         unsigned long len, unsigned long prot,
1406                         unsigned long flags, unsigned long pgoff,
1407                         unsigned long *populate, struct list_head *uf)
1408 {
1409         struct mm_struct *mm = current->mm;
1410         vm_flags_t vm_flags;
1411         int pkey = 0;
1412
1413         *populate = 0;
1414
1415         if (!len)
1416                 return -EINVAL;
1417
1418         /*
1419          * Does the application expect PROT_READ to imply PROT_EXEC?
1420          *
1421          * (the exception is when the underlying filesystem is noexec
1422          *  mounted, in which case we dont add PROT_EXEC.)
1423          */
1424         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1425                 if (!(file && path_noexec(&file->f_path)))
1426                         prot |= PROT_EXEC;
1427
1428         /* force arch specific MAP_FIXED handling in get_unmapped_area */
1429         if (flags & MAP_FIXED_NOREPLACE)
1430                 flags |= MAP_FIXED;
1431
1432         if (!(flags & MAP_FIXED))
1433                 addr = round_hint_to_min(addr);
1434
1435         /* Careful about overflows.. */
1436         len = PAGE_ALIGN(len);
1437         if (!len)
1438                 return -ENOMEM;
1439
1440         /* offset overflow? */
1441         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1442                 return -EOVERFLOW;
1443
1444         /* Too many mappings? */
1445         if (mm->map_count > sysctl_max_map_count)
1446                 return -ENOMEM;
1447
1448         /* Obtain the address to map to. we verify (or select) it and ensure
1449          * that it represents a valid section of the address space.
1450          */
1451         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1452         if (IS_ERR_VALUE(addr))
1453                 return addr;
1454
1455         if (flags & MAP_FIXED_NOREPLACE) {
1456                 if (find_vma_intersection(mm, addr, addr + len))
1457                         return -EEXIST;
1458         }
1459
1460         if (prot == PROT_EXEC) {
1461                 pkey = execute_only_pkey(mm);
1462                 if (pkey < 0)
1463                         pkey = 0;
1464         }
1465
1466         /* Do simple checking here so the lower-level routines won't have
1467          * to. we assume access permissions have been handled by the open
1468          * of the memory object, so we don't do any here.
1469          */
1470         vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1471                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1472
1473         if (flags & MAP_LOCKED)
1474                 if (!can_do_mlock())
1475                         return -EPERM;
1476
1477         if (mlock_future_check(mm, vm_flags, len))
1478                 return -EAGAIN;
1479
1480         if (file) {
1481                 struct inode *inode = file_inode(file);
1482                 unsigned long flags_mask;
1483
1484                 if (!file_mmap_ok(file, inode, pgoff, len))
1485                         return -EOVERFLOW;
1486
1487                 flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1488
1489                 switch (flags & MAP_TYPE) {
1490                 case MAP_SHARED:
1491                         /*
1492                          * Force use of MAP_SHARED_VALIDATE with non-legacy
1493                          * flags. E.g. MAP_SYNC is dangerous to use with
1494                          * MAP_SHARED as you don't know which consistency model
1495                          * you will get. We silently ignore unsupported flags
1496                          * with MAP_SHARED to preserve backward compatibility.
1497                          */
1498                         flags &= LEGACY_MAP_MASK;
1499                         fallthrough;
1500                 case MAP_SHARED_VALIDATE:
1501                         if (flags & ~flags_mask)
1502                                 return -EOPNOTSUPP;
1503                         if (prot & PROT_WRITE) {
1504                                 if (!(file->f_mode & FMODE_WRITE))
1505                                         return -EACCES;
1506                                 if (IS_SWAPFILE(file->f_mapping->host))
1507                                         return -ETXTBSY;
1508                         }
1509
1510                         /*
1511                          * Make sure we don't allow writing to an append-only
1512                          * file..
1513                          */
1514                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1515                                 return -EACCES;
1516
1517                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1518                         if (!(file->f_mode & FMODE_WRITE))
1519                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1520                         fallthrough;
1521                 case MAP_PRIVATE:
1522                         if (!(file->f_mode & FMODE_READ))
1523                                 return -EACCES;
1524                         if (path_noexec(&file->f_path)) {
1525                                 if (vm_flags & VM_EXEC)
1526                                         return -EPERM;
1527                                 vm_flags &= ~VM_MAYEXEC;
1528                         }
1529
1530                         if (!file->f_op->mmap)
1531                                 return -ENODEV;
1532                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1533                                 return -EINVAL;
1534                         break;
1535
1536                 default:
1537                         return -EINVAL;
1538                 }
1539         } else {
1540                 switch (flags & MAP_TYPE) {
1541                 case MAP_SHARED:
1542                         if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1543                                 return -EINVAL;
1544                         /*
1545                          * Ignore pgoff.
1546                          */
1547                         pgoff = 0;
1548                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1549                         break;
1550                 case MAP_PRIVATE:
1551                         /*
1552                          * Set pgoff according to addr for anon_vma.
1553                          */
1554                         pgoff = addr >> PAGE_SHIFT;
1555                         break;
1556                 default:
1557                         return -EINVAL;
1558                 }
1559         }
1560
1561         /*
1562          * Set 'VM_NORESERVE' if we should not account for the
1563          * memory use of this mapping.
1564          */
1565         if (flags & MAP_NORESERVE) {
1566                 /* We honor MAP_NORESERVE if allowed to overcommit */
1567                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1568                         vm_flags |= VM_NORESERVE;
1569
1570                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1571                 if (file && is_file_hugepages(file))
1572                         vm_flags |= VM_NORESERVE;
1573         }
1574
1575         addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1576         if (!IS_ERR_VALUE(addr) &&
1577             ((vm_flags & VM_LOCKED) ||
1578              (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1579                 *populate = len;
1580         return addr;
1581 }
1582
1583 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1584                               unsigned long prot, unsigned long flags,
1585                               unsigned long fd, unsigned long pgoff)
1586 {
1587         struct file *file = NULL;
1588         unsigned long retval;
1589
1590         if (!(flags & MAP_ANONYMOUS)) {
1591                 audit_mmap_fd(fd, flags);
1592                 file = fget(fd);
1593                 if (!file)
1594                         return -EBADF;
1595                 if (is_file_hugepages(file)) {
1596                         len = ALIGN(len, huge_page_size(hstate_file(file)));
1597                 } else if (unlikely(flags & MAP_HUGETLB)) {
1598                         retval = -EINVAL;
1599                         goto out_fput;
1600                 }
1601         } else if (flags & MAP_HUGETLB) {
1602                 struct hstate *hs;
1603
1604                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1605                 if (!hs)
1606                         return -EINVAL;
1607
1608                 len = ALIGN(len, huge_page_size(hs));
1609                 /*
1610                  * VM_NORESERVE is used because the reservations will be
1611                  * taken when vm_ops->mmap() is called
1612                  * A dummy user value is used because we are not locking
1613                  * memory so no accounting is necessary
1614                  */
1615                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1616                                 VM_NORESERVE,
1617                                 HUGETLB_ANONHUGE_INODE,
1618                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1619                 if (IS_ERR(file))
1620                         return PTR_ERR(file);
1621         }
1622
1623         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1624 out_fput:
1625         if (file)
1626                 fput(file);
1627         return retval;
1628 }
1629
1630 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1631                 unsigned long, prot, unsigned long, flags,
1632                 unsigned long, fd, unsigned long, pgoff)
1633 {
1634         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1635 }
1636
1637 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1638 struct mmap_arg_struct {
1639         unsigned long addr;
1640         unsigned long len;
1641         unsigned long prot;
1642         unsigned long flags;
1643         unsigned long fd;
1644         unsigned long offset;
1645 };
1646
1647 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1648 {
1649         struct mmap_arg_struct a;
1650
1651         if (copy_from_user(&a, arg, sizeof(a)))
1652                 return -EFAULT;
1653         if (offset_in_page(a.offset))
1654                 return -EINVAL;
1655
1656         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1657                                a.offset >> PAGE_SHIFT);
1658 }
1659 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1660
1661 /*
1662  * Some shared mappings will want the pages marked read-only
1663  * to track write events. If so, we'll downgrade vm_page_prot
1664  * to the private version (using protection_map[] without the
1665  * VM_SHARED bit).
1666  */
1667 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1668 {
1669         vm_flags_t vm_flags = vma->vm_flags;
1670         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1671
1672         /* If it was private or non-writable, the write bit is already clear */
1673         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1674                 return 0;
1675
1676         /* The backer wishes to know when pages are first written to? */
1677         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1678                 return 1;
1679
1680         /* The open routine did something to the protections that pgprot_modify
1681          * won't preserve? */
1682         if (pgprot_val(vm_page_prot) !=
1683             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1684                 return 0;
1685
1686         /* Do we need to track softdirty? */
1687         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1688                 return 1;
1689
1690         /* Specialty mapping? */
1691         if (vm_flags & VM_PFNMAP)
1692                 return 0;
1693
1694         /* Can the mapping track the dirty pages? */
1695         return vma->vm_file && vma->vm_file->f_mapping &&
1696                 mapping_can_writeback(vma->vm_file->f_mapping);
1697 }
1698
1699 /*
1700  * We account for memory if it's a private writeable mapping,
1701  * not hugepages and VM_NORESERVE wasn't set.
1702  */
1703 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1704 {
1705         /*
1706          * hugetlb has its own accounting separate from the core VM
1707          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1708          */
1709         if (file && is_file_hugepages(file))
1710                 return 0;
1711
1712         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1713 }
1714
1715 unsigned long mmap_region(struct file *file, unsigned long addr,
1716                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1717                 struct list_head *uf)
1718 {
1719         struct mm_struct *mm = current->mm;
1720         struct vm_area_struct *vma, *prev, *merge;
1721         int error;
1722         struct rb_node **rb_link, *rb_parent;
1723         unsigned long charged = 0;
1724
1725         /* Check against address space limit. */
1726         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1727                 unsigned long nr_pages;
1728
1729                 /*
1730                  * MAP_FIXED may remove pages of mappings that intersects with
1731                  * requested mapping. Account for the pages it would unmap.
1732                  */
1733                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1734
1735                 if (!may_expand_vm(mm, vm_flags,
1736                                         (len >> PAGE_SHIFT) - nr_pages))
1737                         return -ENOMEM;
1738         }
1739
1740         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1741         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1742                 return -ENOMEM;
1743         /*
1744          * Private writable mapping: check memory availability
1745          */
1746         if (accountable_mapping(file, vm_flags)) {
1747                 charged = len >> PAGE_SHIFT;
1748                 if (security_vm_enough_memory_mm(mm, charged))
1749                         return -ENOMEM;
1750                 vm_flags |= VM_ACCOUNT;
1751         }
1752
1753         /*
1754          * Can we just expand an old mapping?
1755          */
1756         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1757                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1758         if (vma)
1759                 goto out;
1760
1761         /*
1762          * Determine the object being mapped and call the appropriate
1763          * specific mapper. the address has already been validated, but
1764          * not unmapped, but the maps are removed from the list.
1765          */
1766         vma = vm_area_alloc(mm);
1767         if (!vma) {
1768                 error = -ENOMEM;
1769                 goto unacct_error;
1770         }
1771
1772         vma->vm_start = addr;
1773         vma->vm_end = addr + len;
1774         vma->vm_flags = vm_flags;
1775         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1776         vma->vm_pgoff = pgoff;
1777
1778         if (file) {
1779                 if (vm_flags & VM_SHARED) {
1780                         error = mapping_map_writable(file->f_mapping);
1781                         if (error)
1782                                 goto free_vma;
1783                 }
1784
1785                 vma->vm_file = get_file(file);
1786                 error = call_mmap(file, vma);
1787                 if (error)
1788                         goto unmap_and_free_vma;
1789
1790                 /* Can addr have changed??
1791                  *
1792                  * Answer: Yes, several device drivers can do it in their
1793                  *         f_op->mmap method. -DaveM
1794                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1795                  *      be updated for vma_link()
1796                  */
1797                 WARN_ON_ONCE(addr != vma->vm_start);
1798
1799                 addr = vma->vm_start;
1800
1801                 /* If vm_flags changed after call_mmap(), we should try merge vma again
1802                  * as we may succeed this time.
1803                  */
1804                 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1805                         merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1806                                 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1807                         if (merge) {
1808                                 /* ->mmap() can change vma->vm_file and fput the original file. So
1809                                  * fput the vma->vm_file here or we would add an extra fput for file
1810                                  * and cause general protection fault ultimately.
1811                                  */
1812                                 fput(vma->vm_file);
1813                                 vm_area_free(vma);
1814                                 vma = merge;
1815                                 /* Update vm_flags to pick up the change. */
1816                                 vm_flags = vma->vm_flags;
1817                                 goto unmap_writable;
1818                         }
1819                 }
1820
1821                 vm_flags = vma->vm_flags;
1822         } else if (vm_flags & VM_SHARED) {
1823                 error = shmem_zero_setup(vma);
1824                 if (error)
1825                         goto free_vma;
1826         } else {
1827                 vma_set_anonymous(vma);
1828         }
1829
1830         /* Allow architectures to sanity-check the vm_flags */
1831         if (!arch_validate_flags(vma->vm_flags)) {
1832                 error = -EINVAL;
1833                 if (file)
1834                         goto unmap_and_free_vma;
1835                 else
1836                         goto free_vma;
1837         }
1838
1839         vma_link(mm, vma, prev, rb_link, rb_parent);
1840         /* Once vma denies write, undo our temporary denial count */
1841 unmap_writable:
1842         if (file && vm_flags & VM_SHARED)
1843                 mapping_unmap_writable(file->f_mapping);
1844         file = vma->vm_file;
1845 out:
1846         perf_event_mmap(vma);
1847
1848         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1849         if (vm_flags & VM_LOCKED) {
1850                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1851                                         is_vm_hugetlb_page(vma) ||
1852                                         vma == get_gate_vma(current->mm))
1853                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1854                 else
1855                         mm->locked_vm += (len >> PAGE_SHIFT);
1856         }
1857
1858         if (file)
1859                 uprobe_mmap(vma);
1860
1861         /*
1862          * New (or expanded) vma always get soft dirty status.
1863          * Otherwise user-space soft-dirty page tracker won't
1864          * be able to distinguish situation when vma area unmapped,
1865          * then new mapped in-place (which must be aimed as
1866          * a completely new data area).
1867          */
1868         vma->vm_flags |= VM_SOFTDIRTY;
1869
1870         vma_set_page_prot(vma);
1871
1872         return addr;
1873
1874 unmap_and_free_vma:
1875         fput(vma->vm_file);
1876         vma->vm_file = NULL;
1877
1878         /* Undo any partial mapping done by a device driver. */
1879         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1880         charged = 0;
1881         if (vm_flags & VM_SHARED)
1882                 mapping_unmap_writable(file->f_mapping);
1883 free_vma:
1884         vm_area_free(vma);
1885 unacct_error:
1886         if (charged)
1887                 vm_unacct_memory(charged);
1888         return error;
1889 }
1890
1891 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1892 {
1893         /*
1894          * We implement the search by looking for an rbtree node that
1895          * immediately follows a suitable gap. That is,
1896          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1897          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1898          * - gap_end - gap_start >= length
1899          */
1900
1901         struct mm_struct *mm = current->mm;
1902         struct vm_area_struct *vma;
1903         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1904
1905         /* Adjust search length to account for worst case alignment overhead */
1906         length = info->length + info->align_mask;
1907         if (length < info->length)
1908                 return -ENOMEM;
1909
1910         /* Adjust search limits by the desired length */
1911         if (info->high_limit < length)
1912                 return -ENOMEM;
1913         high_limit = info->high_limit - length;
1914
1915         if (info->low_limit > high_limit)
1916                 return -ENOMEM;
1917         low_limit = info->low_limit + length;
1918
1919         /* Check if rbtree root looks promising */
1920         if (RB_EMPTY_ROOT(&mm->mm_rb))
1921                 goto check_highest;
1922         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1923         if (vma->rb_subtree_gap < length)
1924                 goto check_highest;
1925
1926         while (true) {
1927                 /* Visit left subtree if it looks promising */
1928                 gap_end = vm_start_gap(vma);
1929                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1930                         struct vm_area_struct *left =
1931                                 rb_entry(vma->vm_rb.rb_left,
1932                                          struct vm_area_struct, vm_rb);
1933                         if (left->rb_subtree_gap >= length) {
1934                                 vma = left;
1935                                 continue;
1936                         }
1937                 }
1938
1939                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1940 check_current:
1941                 /* Check if current node has a suitable gap */
1942                 if (gap_start > high_limit)
1943                         return -ENOMEM;
1944                 if (gap_end >= low_limit &&
1945                     gap_end > gap_start && gap_end - gap_start >= length)
1946                         goto found;
1947
1948                 /* Visit right subtree if it looks promising */
1949                 if (vma->vm_rb.rb_right) {
1950                         struct vm_area_struct *right =
1951                                 rb_entry(vma->vm_rb.rb_right,
1952                                          struct vm_area_struct, vm_rb);
1953                         if (right->rb_subtree_gap >= length) {
1954                                 vma = right;
1955                                 continue;
1956                         }
1957                 }
1958
1959                 /* Go back up the rbtree to find next candidate node */
1960                 while (true) {
1961                         struct rb_node *prev = &vma->vm_rb;
1962                         if (!rb_parent(prev))
1963                                 goto check_highest;
1964                         vma = rb_entry(rb_parent(prev),
1965                                        struct vm_area_struct, vm_rb);
1966                         if (prev == vma->vm_rb.rb_left) {
1967                                 gap_start = vm_end_gap(vma->vm_prev);
1968                                 gap_end = vm_start_gap(vma);
1969                                 goto check_current;
1970                         }
1971                 }
1972         }
1973
1974 check_highest:
1975         /* Check highest gap, which does not precede any rbtree node */
1976         gap_start = mm->highest_vm_end;
1977         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1978         if (gap_start > high_limit)
1979                 return -ENOMEM;
1980
1981 found:
1982         /* We found a suitable gap. Clip it with the original low_limit. */
1983         if (gap_start < info->low_limit)
1984                 gap_start = info->low_limit;
1985
1986         /* Adjust gap address to the desired alignment */
1987         gap_start += (info->align_offset - gap_start) & info->align_mask;
1988
1989         VM_BUG_ON(gap_start + info->length > info->high_limit);
1990         VM_BUG_ON(gap_start + info->length > gap_end);
1991         return gap_start;
1992 }
1993
1994 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1995 {
1996         struct mm_struct *mm = current->mm;
1997         struct vm_area_struct *vma;
1998         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1999
2000         /* Adjust search length to account for worst case alignment overhead */
2001         length = info->length + info->align_mask;
2002         if (length < info->length)
2003                 return -ENOMEM;
2004
2005         /*
2006          * Adjust search limits by the desired length.
2007          * See implementation comment at top of unmapped_area().
2008          */
2009         gap_end = info->high_limit;
2010         if (gap_end < length)
2011                 return -ENOMEM;
2012         high_limit = gap_end - length;
2013
2014         if (info->low_limit > high_limit)
2015                 return -ENOMEM;
2016         low_limit = info->low_limit + length;
2017
2018         /* Check highest gap, which does not precede any rbtree node */
2019         gap_start = mm->highest_vm_end;
2020         if (gap_start <= high_limit)
2021                 goto found_highest;
2022
2023         /* Check if rbtree root looks promising */
2024         if (RB_EMPTY_ROOT(&mm->mm_rb))
2025                 return -ENOMEM;
2026         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2027         if (vma->rb_subtree_gap < length)
2028                 return -ENOMEM;
2029
2030         while (true) {
2031                 /* Visit right subtree if it looks promising */
2032                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2033                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2034                         struct vm_area_struct *right =
2035                                 rb_entry(vma->vm_rb.rb_right,
2036                                          struct vm_area_struct, vm_rb);
2037                         if (right->rb_subtree_gap >= length) {
2038                                 vma = right;
2039                                 continue;
2040                         }
2041                 }
2042
2043 check_current:
2044                 /* Check if current node has a suitable gap */
2045                 gap_end = vm_start_gap(vma);
2046                 if (gap_end < low_limit)
2047                         return -ENOMEM;
2048                 if (gap_start <= high_limit &&
2049                     gap_end > gap_start && gap_end - gap_start >= length)
2050                         goto found;
2051
2052                 /* Visit left subtree if it looks promising */
2053                 if (vma->vm_rb.rb_left) {
2054                         struct vm_area_struct *left =
2055                                 rb_entry(vma->vm_rb.rb_left,
2056                                          struct vm_area_struct, vm_rb);
2057                         if (left->rb_subtree_gap >= length) {
2058                                 vma = left;
2059                                 continue;
2060                         }
2061                 }
2062
2063                 /* Go back up the rbtree to find next candidate node */
2064                 while (true) {
2065                         struct rb_node *prev = &vma->vm_rb;
2066                         if (!rb_parent(prev))
2067                                 return -ENOMEM;
2068                         vma = rb_entry(rb_parent(prev),
2069                                        struct vm_area_struct, vm_rb);
2070                         if (prev == vma->vm_rb.rb_right) {
2071                                 gap_start = vma->vm_prev ?
2072                                         vm_end_gap(vma->vm_prev) : 0;
2073                                 goto check_current;
2074                         }
2075                 }
2076         }
2077
2078 found:
2079         /* We found a suitable gap. Clip it with the original high_limit. */
2080         if (gap_end > info->high_limit)
2081                 gap_end = info->high_limit;
2082
2083 found_highest:
2084         /* Compute highest gap address at the desired alignment */
2085         gap_end -= info->length;
2086         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2087
2088         VM_BUG_ON(gap_end < info->low_limit);
2089         VM_BUG_ON(gap_end < gap_start);
2090         return gap_end;
2091 }
2092
2093 /*
2094  * Search for an unmapped address range.
2095  *
2096  * We are looking for a range that:
2097  * - does not intersect with any VMA;
2098  * - is contained within the [low_limit, high_limit) interval;
2099  * - is at least the desired size.
2100  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2101  */
2102 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2103 {
2104         unsigned long addr;
2105
2106         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2107                 addr = unmapped_area_topdown(info);
2108         else
2109                 addr = unmapped_area(info);
2110
2111         trace_vm_unmapped_area(addr, info);
2112         return addr;
2113 }
2114
2115 #ifndef arch_get_mmap_end
2116 #define arch_get_mmap_end(addr) (TASK_SIZE)
2117 #endif
2118
2119 #ifndef arch_get_mmap_base
2120 #define arch_get_mmap_base(addr, base) (base)
2121 #endif
2122
2123 /* Get an address range which is currently unmapped.
2124  * For shmat() with addr=0.
2125  *
2126  * Ugly calling convention alert:
2127  * Return value with the low bits set means error value,
2128  * ie
2129  *      if (ret & ~PAGE_MASK)
2130  *              error = ret;
2131  *
2132  * This function "knows" that -ENOMEM has the bits set.
2133  */
2134 #ifndef HAVE_ARCH_UNMAPPED_AREA
2135 unsigned long
2136 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2137                 unsigned long len, unsigned long pgoff, unsigned long flags)
2138 {
2139         struct mm_struct *mm = current->mm;
2140         struct vm_area_struct *vma, *prev;
2141         struct vm_unmapped_area_info info;
2142         const unsigned long mmap_end = arch_get_mmap_end(addr);
2143
2144         if (len > mmap_end - mmap_min_addr)
2145                 return -ENOMEM;
2146
2147         if (flags & MAP_FIXED)
2148                 return addr;
2149
2150         if (addr) {
2151                 addr = PAGE_ALIGN(addr);
2152                 vma = find_vma_prev(mm, addr, &prev);
2153                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2154                     (!vma || addr + len <= vm_start_gap(vma)) &&
2155                     (!prev || addr >= vm_end_gap(prev)))
2156                         return addr;
2157         }
2158
2159         info.flags = 0;
2160         info.length = len;
2161         info.low_limit = mm->mmap_base;
2162         info.high_limit = mmap_end;
2163         info.align_mask = 0;
2164         info.align_offset = 0;
2165         return vm_unmapped_area(&info);
2166 }
2167 #endif
2168
2169 /*
2170  * This mmap-allocator allocates new areas top-down from below the
2171  * stack's low limit (the base):
2172  */
2173 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2174 unsigned long
2175 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2176                           unsigned long len, unsigned long pgoff,
2177                           unsigned long flags)
2178 {
2179         struct vm_area_struct *vma, *prev;
2180         struct mm_struct *mm = current->mm;
2181         struct vm_unmapped_area_info info;
2182         const unsigned long mmap_end = arch_get_mmap_end(addr);
2183
2184         /* requested length too big for entire address space */
2185         if (len > mmap_end - mmap_min_addr)
2186                 return -ENOMEM;
2187
2188         if (flags & MAP_FIXED)
2189                 return addr;
2190
2191         /* requesting a specific address */
2192         if (addr) {
2193                 addr = PAGE_ALIGN(addr);
2194                 vma = find_vma_prev(mm, addr, &prev);
2195                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2196                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2197                                 (!prev || addr >= vm_end_gap(prev)))
2198                         return addr;
2199         }
2200
2201         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2202         info.length = len;
2203         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2204         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2205         info.align_mask = 0;
2206         info.align_offset = 0;
2207         addr = vm_unmapped_area(&info);
2208
2209         /*
2210          * A failed mmap() very likely causes application failure,
2211          * so fall back to the bottom-up function here. This scenario
2212          * can happen with large stack limits and large mmap()
2213          * allocations.
2214          */
2215         if (offset_in_page(addr)) {
2216                 VM_BUG_ON(addr != -ENOMEM);
2217                 info.flags = 0;
2218                 info.low_limit = TASK_UNMAPPED_BASE;
2219                 info.high_limit = mmap_end;
2220                 addr = vm_unmapped_area(&info);
2221         }
2222
2223         return addr;
2224 }
2225 #endif
2226
2227 unsigned long
2228 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2229                 unsigned long pgoff, unsigned long flags)
2230 {
2231         unsigned long (*get_area)(struct file *, unsigned long,
2232                                   unsigned long, unsigned long, unsigned long);
2233
2234         unsigned long error = arch_mmap_check(addr, len, flags);
2235         if (error)
2236                 return error;
2237
2238         /* Careful about overflows.. */
2239         if (len > TASK_SIZE)
2240                 return -ENOMEM;
2241
2242         get_area = current->mm->get_unmapped_area;
2243         if (file) {
2244                 if (file->f_op->get_unmapped_area)
2245                         get_area = file->f_op->get_unmapped_area;
2246         } else if (flags & MAP_SHARED) {
2247                 /*
2248                  * mmap_region() will call shmem_zero_setup() to create a file,
2249                  * so use shmem's get_unmapped_area in case it can be huge.
2250                  * do_mmap() will clear pgoff, so match alignment.
2251                  */
2252                 pgoff = 0;
2253                 get_area = shmem_get_unmapped_area;
2254         }
2255
2256         addr = get_area(file, addr, len, pgoff, flags);
2257         if (IS_ERR_VALUE(addr))
2258                 return addr;
2259
2260         if (addr > TASK_SIZE - len)
2261                 return -ENOMEM;
2262         if (offset_in_page(addr))
2263                 return -EINVAL;
2264
2265         error = security_mmap_addr(addr);
2266         return error ? error : addr;
2267 }
2268
2269 EXPORT_SYMBOL(get_unmapped_area);
2270
2271 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2272 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2273 {
2274         struct rb_node *rb_node;
2275         struct vm_area_struct *vma;
2276
2277         mmap_assert_locked(mm);
2278         /* Check the cache first. */
2279         vma = vmacache_find(mm, addr);
2280         if (likely(vma))
2281                 return vma;
2282
2283         rb_node = mm->mm_rb.rb_node;
2284
2285         while (rb_node) {
2286                 struct vm_area_struct *tmp;
2287
2288                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2289
2290                 if (tmp->vm_end > addr) {
2291                         vma = tmp;
2292                         if (tmp->vm_start <= addr)
2293                                 break;
2294                         rb_node = rb_node->rb_left;
2295                 } else
2296                         rb_node = rb_node->rb_right;
2297         }
2298
2299         if (vma)
2300                 vmacache_update(addr, vma);
2301         return vma;
2302 }
2303
2304 EXPORT_SYMBOL(find_vma);
2305
2306 /*
2307  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2308  */
2309 struct vm_area_struct *
2310 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2311                         struct vm_area_struct **pprev)
2312 {
2313         struct vm_area_struct *vma;
2314
2315         vma = find_vma(mm, addr);
2316         if (vma) {
2317                 *pprev = vma->vm_prev;
2318         } else {
2319                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2320
2321                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2322         }
2323         return vma;
2324 }
2325
2326 /*
2327  * Verify that the stack growth is acceptable and
2328  * update accounting. This is shared with both the
2329  * grow-up and grow-down cases.
2330  */
2331 static int acct_stack_growth(struct vm_area_struct *vma,
2332                              unsigned long size, unsigned long grow)
2333 {
2334         struct mm_struct *mm = vma->vm_mm;
2335         unsigned long new_start;
2336
2337         /* address space limit tests */
2338         if (!may_expand_vm(mm, vma->vm_flags, grow))
2339                 return -ENOMEM;
2340
2341         /* Stack limit test */
2342         if (size > rlimit(RLIMIT_STACK))
2343                 return -ENOMEM;
2344
2345         /* mlock limit tests */
2346         if (vma->vm_flags & VM_LOCKED) {
2347                 unsigned long locked;
2348                 unsigned long limit;
2349                 locked = mm->locked_vm + grow;
2350                 limit = rlimit(RLIMIT_MEMLOCK);
2351                 limit >>= PAGE_SHIFT;
2352                 if (locked > limit && !capable(CAP_IPC_LOCK))
2353                         return -ENOMEM;
2354         }
2355
2356         /* Check to ensure the stack will not grow into a hugetlb-only region */
2357         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2358                         vma->vm_end - size;
2359         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2360                 return -EFAULT;
2361
2362         /*
2363          * Overcommit..  This must be the final test, as it will
2364          * update security statistics.
2365          */
2366         if (security_vm_enough_memory_mm(mm, grow))
2367                 return -ENOMEM;
2368
2369         return 0;
2370 }
2371
2372 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2373 /*
2374  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2375  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2376  */
2377 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2378 {
2379         struct mm_struct *mm = vma->vm_mm;
2380         struct vm_area_struct *next;
2381         unsigned long gap_addr;
2382         int error = 0;
2383
2384         if (!(vma->vm_flags & VM_GROWSUP))
2385                 return -EFAULT;
2386
2387         /* Guard against exceeding limits of the address space. */
2388         address &= PAGE_MASK;
2389         if (address >= (TASK_SIZE & PAGE_MASK))
2390                 return -ENOMEM;
2391         address += PAGE_SIZE;
2392
2393         /* Enforce stack_guard_gap */
2394         gap_addr = address + stack_guard_gap;
2395
2396         /* Guard against overflow */
2397         if (gap_addr < address || gap_addr > TASK_SIZE)
2398                 gap_addr = TASK_SIZE;
2399
2400         next = vma->vm_next;
2401         if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2402                 if (!(next->vm_flags & VM_GROWSUP))
2403                         return -ENOMEM;
2404                 /* Check that both stack segments have the same anon_vma? */
2405         }
2406
2407         /* We must make sure the anon_vma is allocated. */
2408         if (unlikely(anon_vma_prepare(vma)))
2409                 return -ENOMEM;
2410
2411         /*
2412          * vma->vm_start/vm_end cannot change under us because the caller
2413          * is required to hold the mmap_lock in read mode.  We need the
2414          * anon_vma lock to serialize against concurrent expand_stacks.
2415          */
2416         anon_vma_lock_write(vma->anon_vma);
2417
2418         /* Somebody else might have raced and expanded it already */
2419         if (address > vma->vm_end) {
2420                 unsigned long size, grow;
2421
2422                 size = address - vma->vm_start;
2423                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2424
2425                 error = -ENOMEM;
2426                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2427                         error = acct_stack_growth(vma, size, grow);
2428                         if (!error) {
2429                                 /*
2430                                  * vma_gap_update() doesn't support concurrent
2431                                  * updates, but we only hold a shared mmap_lock
2432                                  * lock here, so we need to protect against
2433                                  * concurrent vma expansions.
2434                                  * anon_vma_lock_write() doesn't help here, as
2435                                  * we don't guarantee that all growable vmas
2436                                  * in a mm share the same root anon vma.
2437                                  * So, we reuse mm->page_table_lock to guard
2438                                  * against concurrent vma expansions.
2439                                  */
2440                                 spin_lock(&mm->page_table_lock);
2441                                 if (vma->vm_flags & VM_LOCKED)
2442                                         mm->locked_vm += grow;
2443                                 vm_stat_account(mm, vma->vm_flags, grow);
2444                                 anon_vma_interval_tree_pre_update_vma(vma);
2445                                 vma->vm_end = address;
2446                                 anon_vma_interval_tree_post_update_vma(vma);
2447                                 if (vma->vm_next)
2448                                         vma_gap_update(vma->vm_next);
2449                                 else
2450                                         mm->highest_vm_end = vm_end_gap(vma);
2451                                 spin_unlock(&mm->page_table_lock);
2452
2453                                 perf_event_mmap(vma);
2454                         }
2455                 }
2456         }
2457         anon_vma_unlock_write(vma->anon_vma);
2458         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2459         validate_mm(mm);
2460         return error;
2461 }
2462 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2463
2464 /*
2465  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2466  */
2467 int expand_downwards(struct vm_area_struct *vma,
2468                                    unsigned long address)
2469 {
2470         struct mm_struct *mm = vma->vm_mm;
2471         struct vm_area_struct *prev;
2472         int error = 0;
2473
2474         address &= PAGE_MASK;
2475         if (address < mmap_min_addr)
2476                 return -EPERM;
2477
2478         /* Enforce stack_guard_gap */
2479         prev = vma->vm_prev;
2480         /* Check that both stack segments have the same anon_vma? */
2481         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2482                         vma_is_accessible(prev)) {
2483                 if (address - prev->vm_end < stack_guard_gap)
2484                         return -ENOMEM;
2485         }
2486
2487         /* We must make sure the anon_vma is allocated. */
2488         if (unlikely(anon_vma_prepare(vma)))
2489                 return -ENOMEM;
2490
2491         /*
2492          * vma->vm_start/vm_end cannot change under us because the caller
2493          * is required to hold the mmap_lock in read mode.  We need the
2494          * anon_vma lock to serialize against concurrent expand_stacks.
2495          */
2496         anon_vma_lock_write(vma->anon_vma);
2497
2498         /* Somebody else might have raced and expanded it already */
2499         if (address < vma->vm_start) {
2500                 unsigned long size, grow;
2501
2502                 size = vma->vm_end - address;
2503                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2504
2505                 error = -ENOMEM;
2506                 if (grow <= vma->vm_pgoff) {
2507                         error = acct_stack_growth(vma, size, grow);
2508                         if (!error) {
2509                                 /*
2510                                  * vma_gap_update() doesn't support concurrent
2511                                  * updates, but we only hold a shared mmap_lock
2512                                  * lock here, so we need to protect against
2513                                  * concurrent vma expansions.
2514                                  * anon_vma_lock_write() doesn't help here, as
2515                                  * we don't guarantee that all growable vmas
2516                                  * in a mm share the same root anon vma.
2517                                  * So, we reuse mm->page_table_lock to guard
2518                                  * against concurrent vma expansions.
2519                                  */
2520                                 spin_lock(&mm->page_table_lock);
2521                                 if (vma->vm_flags & VM_LOCKED)
2522                                         mm->locked_vm += grow;
2523                                 vm_stat_account(mm, vma->vm_flags, grow);
2524                                 anon_vma_interval_tree_pre_update_vma(vma);
2525                                 vma->vm_start = address;
2526                                 vma->vm_pgoff -= grow;
2527                                 anon_vma_interval_tree_post_update_vma(vma);
2528                                 vma_gap_update(vma);
2529                                 spin_unlock(&mm->page_table_lock);
2530
2531                                 perf_event_mmap(vma);
2532                         }
2533                 }
2534         }
2535         anon_vma_unlock_write(vma->anon_vma);
2536         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2537         validate_mm(mm);
2538         return error;
2539 }
2540
2541 /* enforced gap between the expanding stack and other mappings. */
2542 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2543
2544 static int __init cmdline_parse_stack_guard_gap(char *p)
2545 {
2546         unsigned long val;
2547         char *endptr;
2548
2549         val = simple_strtoul(p, &endptr, 10);
2550         if (!*endptr)
2551                 stack_guard_gap = val << PAGE_SHIFT;
2552
2553         return 0;
2554 }
2555 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2556
2557 #ifdef CONFIG_STACK_GROWSUP
2558 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2559 {
2560         return expand_upwards(vma, address);
2561 }
2562
2563 struct vm_area_struct *
2564 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2565 {
2566         struct vm_area_struct *vma, *prev;
2567
2568         addr &= PAGE_MASK;
2569         vma = find_vma_prev(mm, addr, &prev);
2570         if (vma && (vma->vm_start <= addr))
2571                 return vma;
2572         /* don't alter vm_end if the coredump is running */
2573         if (!prev || expand_stack(prev, addr))
2574                 return NULL;
2575         if (prev->vm_flags & VM_LOCKED)
2576                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2577         return prev;
2578 }
2579 #else
2580 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2581 {
2582         return expand_downwards(vma, address);
2583 }
2584
2585 struct vm_area_struct *
2586 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2587 {
2588         struct vm_area_struct *vma;
2589         unsigned long start;
2590
2591         addr &= PAGE_MASK;
2592         vma = find_vma(mm, addr);
2593         if (!vma)
2594                 return NULL;
2595         if (vma->vm_start <= addr)
2596                 return vma;
2597         if (!(vma->vm_flags & VM_GROWSDOWN))
2598                 return NULL;
2599         start = vma->vm_start;
2600         if (expand_stack(vma, addr))
2601                 return NULL;
2602         if (vma->vm_flags & VM_LOCKED)
2603                 populate_vma_page_range(vma, addr, start, NULL);
2604         return vma;
2605 }
2606 #endif
2607
2608 EXPORT_SYMBOL_GPL(find_extend_vma);
2609
2610 /*
2611  * Ok - we have the memory areas we should free on the vma list,
2612  * so release them, and do the vma updates.
2613  *
2614  * Called with the mm semaphore held.
2615  */
2616 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2617 {
2618         unsigned long nr_accounted = 0;
2619
2620         /* Update high watermark before we lower total_vm */
2621         update_hiwater_vm(mm);
2622         do {
2623                 long nrpages = vma_pages(vma);
2624
2625                 if (vma->vm_flags & VM_ACCOUNT)
2626                         nr_accounted += nrpages;
2627                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2628                 vma = remove_vma(vma);
2629         } while (vma);
2630         vm_unacct_memory(nr_accounted);
2631         validate_mm(mm);
2632 }
2633
2634 /*
2635  * Get rid of page table information in the indicated region.
2636  *
2637  * Called with the mm semaphore held.
2638  */
2639 static void unmap_region(struct mm_struct *mm,
2640                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2641                 unsigned long start, unsigned long end)
2642 {
2643         struct vm_area_struct *next = vma_next(mm, prev);
2644         struct mmu_gather tlb;
2645
2646         lru_add_drain();
2647         tlb_gather_mmu(&tlb, mm);
2648         update_hiwater_rss(mm);
2649         unmap_vmas(&tlb, vma, start, end);
2650         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2651                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2652         tlb_finish_mmu(&tlb);
2653 }
2654
2655 /*
2656  * Create a list of vma's touched by the unmap, removing them from the mm's
2657  * vma list as we go..
2658  */
2659 static bool
2660 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2661         struct vm_area_struct *prev, unsigned long end)
2662 {
2663         struct vm_area_struct **insertion_point;
2664         struct vm_area_struct *tail_vma = NULL;
2665
2666         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2667         vma->vm_prev = NULL;
2668         do {
2669                 vma_rb_erase(vma, &mm->mm_rb);
2670                 mm->map_count--;
2671                 tail_vma = vma;
2672                 vma = vma->vm_next;
2673         } while (vma && vma->vm_start < end);
2674         *insertion_point = vma;
2675         if (vma) {
2676                 vma->vm_prev = prev;
2677                 vma_gap_update(vma);
2678         } else
2679                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2680         tail_vma->vm_next = NULL;
2681
2682         /* Kill the cache */
2683         vmacache_invalidate(mm);
2684
2685         /*
2686          * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2687          * VM_GROWSUP VMA. Such VMAs can change their size under
2688          * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2689          */
2690         if (vma && (vma->vm_flags & VM_GROWSDOWN))
2691                 return false;
2692         if (prev && (prev->vm_flags & VM_GROWSUP))
2693                 return false;
2694         return true;
2695 }
2696
2697 /*
2698  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2699  * has already been checked or doesn't make sense to fail.
2700  */
2701 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2702                 unsigned long addr, int new_below)
2703 {
2704         struct vm_area_struct *new;
2705         int err;
2706
2707         if (vma->vm_ops && vma->vm_ops->may_split) {
2708                 err = vma->vm_ops->may_split(vma, addr);
2709                 if (err)
2710                         return err;
2711         }
2712
2713         new = vm_area_dup(vma);
2714         if (!new)
2715                 return -ENOMEM;
2716
2717         if (new_below)
2718                 new->vm_end = addr;
2719         else {
2720                 new->vm_start = addr;
2721                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2722         }
2723
2724         err = vma_dup_policy(vma, new);
2725         if (err)
2726                 goto out_free_vma;
2727
2728         err = anon_vma_clone(new, vma);
2729         if (err)
2730                 goto out_free_mpol;
2731
2732         if (new->vm_file)
2733                 get_file(new->vm_file);
2734
2735         if (new->vm_ops && new->vm_ops->open)
2736                 new->vm_ops->open(new);
2737
2738         if (new_below)
2739                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2740                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2741         else
2742                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2743
2744         /* Success. */
2745         if (!err)
2746                 return 0;
2747
2748         /* Clean everything up if vma_adjust failed. */
2749         if (new->vm_ops && new->vm_ops->close)
2750                 new->vm_ops->close(new);
2751         if (new->vm_file)
2752                 fput(new->vm_file);
2753         unlink_anon_vmas(new);
2754  out_free_mpol:
2755         mpol_put(vma_policy(new));
2756  out_free_vma:
2757         vm_area_free(new);
2758         return err;
2759 }
2760
2761 /*
2762  * Split a vma into two pieces at address 'addr', a new vma is allocated
2763  * either for the first part or the tail.
2764  */
2765 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2766               unsigned long addr, int new_below)
2767 {
2768         if (mm->map_count >= sysctl_max_map_count)
2769                 return -ENOMEM;
2770
2771         return __split_vma(mm, vma, addr, new_below);
2772 }
2773
2774 static inline void
2775 unlock_range(struct vm_area_struct *start, unsigned long limit)
2776 {
2777         struct mm_struct *mm = start->vm_mm;
2778         struct vm_area_struct *tmp = start;
2779
2780         while (tmp && tmp->vm_start < limit) {
2781                 if (tmp->vm_flags & VM_LOCKED) {
2782                         mm->locked_vm -= vma_pages(tmp);
2783                         munlock_vma_pages_all(tmp);
2784                 }
2785
2786                 tmp = tmp->vm_next;
2787         }
2788 }
2789
2790 /* Munmap is split into 2 main parts -- this part which finds
2791  * what needs doing, and the areas themselves, which do the
2792  * work.  This now handles partial unmappings.
2793  * Jeremy Fitzhardinge <jeremy@goop.org>
2794  */
2795 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2796                 struct list_head *uf, bool downgrade)
2797 {
2798         unsigned long end;
2799         struct vm_area_struct *vma, *prev, *last;
2800
2801         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2802                 return -EINVAL;
2803
2804         len = PAGE_ALIGN(len);
2805         end = start + len;
2806         if (len == 0)
2807                 return -EINVAL;
2808
2809         /*
2810          * arch_unmap() might do unmaps itself.  It must be called
2811          * and finish any rbtree manipulation before this code
2812          * runs and also starts to manipulate the rbtree.
2813          */
2814         arch_unmap(mm, start, end);
2815
2816         /* Find the first overlapping VMA where start < vma->vm_end */
2817         vma = find_vma_intersection(mm, start, end);
2818         if (!vma)
2819                 return 0;
2820         prev = vma->vm_prev;
2821
2822         /*
2823          * If we need to split any vma, do it now to save pain later.
2824          *
2825          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2826          * unmapped vm_area_struct will remain in use: so lower split_vma
2827          * places tmp vma above, and higher split_vma places tmp vma below.
2828          */
2829         if (start > vma->vm_start) {
2830                 int error;
2831
2832                 /*
2833                  * Make sure that map_count on return from munmap() will
2834                  * not exceed its limit; but let map_count go just above
2835                  * its limit temporarily, to help free resources as expected.
2836                  */
2837                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2838                         return -ENOMEM;
2839
2840                 error = __split_vma(mm, vma, start, 0);
2841                 if (error)
2842                         return error;
2843                 prev = vma;
2844         }
2845
2846         /* Does it split the last one? */
2847         last = find_vma(mm, end);
2848         if (last && end > last->vm_start) {
2849                 int error = __split_vma(mm, last, end, 1);
2850                 if (error)
2851                         return error;
2852         }
2853         vma = vma_next(mm, prev);
2854
2855         if (unlikely(uf)) {
2856                 /*
2857                  * If userfaultfd_unmap_prep returns an error the vmas
2858                  * will remain split, but userland will get a
2859                  * highly unexpected error anyway. This is no
2860                  * different than the case where the first of the two
2861                  * __split_vma fails, but we don't undo the first
2862                  * split, despite we could. This is unlikely enough
2863                  * failure that it's not worth optimizing it for.
2864                  */
2865                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2866                 if (error)
2867                         return error;
2868         }
2869
2870         /*
2871          * unlock any mlock()ed ranges before detaching vmas
2872          */
2873         if (mm->locked_vm)
2874                 unlock_range(vma, end);
2875
2876         /* Detach vmas from rbtree */
2877         if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2878                 downgrade = false;
2879
2880         if (downgrade)
2881                 mmap_write_downgrade(mm);
2882
2883         unmap_region(mm, vma, prev, start, end);
2884
2885         /* Fix up all other VM information */
2886         remove_vma_list(mm, vma);
2887
2888         return downgrade ? 1 : 0;
2889 }
2890
2891 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2892               struct list_head *uf)
2893 {
2894         return __do_munmap(mm, start, len, uf, false);
2895 }
2896
2897 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2898 {
2899         int ret;
2900         struct mm_struct *mm = current->mm;
2901         LIST_HEAD(uf);
2902
2903         if (mmap_write_lock_killable(mm))
2904                 return -EINTR;
2905
2906         ret = __do_munmap(mm, start, len, &uf, downgrade);
2907         /*
2908          * Returning 1 indicates mmap_lock is downgraded.
2909          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2910          * it to 0 before return.
2911          */
2912         if (ret == 1) {
2913                 mmap_read_unlock(mm);
2914                 ret = 0;
2915         } else
2916                 mmap_write_unlock(mm);
2917
2918         userfaultfd_unmap_complete(mm, &uf);
2919         return ret;
2920 }
2921
2922 int vm_munmap(unsigned long start, size_t len)
2923 {
2924         return __vm_munmap(start, len, false);
2925 }
2926 EXPORT_SYMBOL(vm_munmap);
2927
2928 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2929 {
2930         addr = untagged_addr(addr);
2931         profile_munmap(addr);
2932         return __vm_munmap(addr, len, true);
2933 }
2934
2935
2936 /*
2937  * Emulation of deprecated remap_file_pages() syscall.
2938  */
2939 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2940                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2941 {
2942
2943         struct mm_struct *mm = current->mm;
2944         struct vm_area_struct *vma;
2945         unsigned long populate = 0;
2946         unsigned long ret = -EINVAL;
2947         struct file *file;
2948
2949         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2950                      current->comm, current->pid);
2951
2952         if (prot)
2953                 return ret;
2954         start = start & PAGE_MASK;
2955         size = size & PAGE_MASK;
2956
2957         if (start + size <= start)
2958                 return ret;
2959
2960         /* Does pgoff wrap? */
2961         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2962                 return ret;
2963
2964         if (mmap_write_lock_killable(mm))
2965                 return -EINTR;
2966
2967         vma = vma_lookup(mm, start);
2968
2969         if (!vma || !(vma->vm_flags & VM_SHARED))
2970                 goto out;
2971
2972         if (start + size > vma->vm_end) {
2973                 struct vm_area_struct *next;
2974
2975                 for (next = vma->vm_next; next; next = next->vm_next) {
2976                         /* hole between vmas ? */
2977                         if (next->vm_start != next->vm_prev->vm_end)
2978                                 goto out;
2979
2980                         if (next->vm_file != vma->vm_file)
2981                                 goto out;
2982
2983                         if (next->vm_flags != vma->vm_flags)
2984                                 goto out;
2985
2986                         if (start + size <= next->vm_end)
2987                                 break;
2988                 }
2989
2990                 if (!next)
2991                         goto out;
2992         }
2993
2994         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2995         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2996         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2997
2998         flags &= MAP_NONBLOCK;
2999         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3000         if (vma->vm_flags & VM_LOCKED)
3001                 flags |= MAP_LOCKED;
3002
3003         file = get_file(vma->vm_file);
3004         ret = do_mmap(vma->vm_file, start, size,
3005                         prot, flags, pgoff, &populate, NULL);
3006         fput(file);
3007 out:
3008         mmap_write_unlock(mm);
3009         if (populate)
3010                 mm_populate(ret, populate);
3011         if (!IS_ERR_VALUE(ret))
3012                 ret = 0;
3013         return ret;
3014 }
3015
3016 /*
3017  *  this is really a simplified "do_mmap".  it only handles
3018  *  anonymous maps.  eventually we may be able to do some
3019  *  brk-specific accounting here.
3020  */
3021 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3022 {
3023         struct mm_struct *mm = current->mm;
3024         struct vm_area_struct *vma, *prev;
3025         struct rb_node **rb_link, *rb_parent;
3026         pgoff_t pgoff = addr >> PAGE_SHIFT;
3027         int error;
3028         unsigned long mapped_addr;
3029
3030         /* Until we need other flags, refuse anything except VM_EXEC. */
3031         if ((flags & (~VM_EXEC)) != 0)
3032                 return -EINVAL;
3033         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3034
3035         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3036         if (IS_ERR_VALUE(mapped_addr))
3037                 return mapped_addr;
3038
3039         error = mlock_future_check(mm, mm->def_flags, len);
3040         if (error)
3041                 return error;
3042
3043         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3044         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3045                 return -ENOMEM;
3046
3047         /* Check against address space limits *after* clearing old maps... */
3048         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3049                 return -ENOMEM;
3050
3051         if (mm->map_count > sysctl_max_map_count)
3052                 return -ENOMEM;
3053
3054         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3055                 return -ENOMEM;
3056
3057         /* Can we just expand an old private anonymous mapping? */
3058         vma = vma_merge(mm, prev, addr, addr + len, flags,
3059                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3060         if (vma)
3061                 goto out;
3062
3063         /*
3064          * create a vma struct for an anonymous mapping
3065          */
3066         vma = vm_area_alloc(mm);
3067         if (!vma) {
3068                 vm_unacct_memory(len >> PAGE_SHIFT);
3069                 return -ENOMEM;
3070         }
3071
3072         vma_set_anonymous(vma);
3073         vma->vm_start = addr;
3074         vma->vm_end = addr + len;
3075         vma->vm_pgoff = pgoff;
3076         vma->vm_flags = flags;
3077         vma->vm_page_prot = vm_get_page_prot(flags);
3078         vma_link(mm, vma, prev, rb_link, rb_parent);
3079 out:
3080         perf_event_mmap(vma);
3081         mm->total_vm += len >> PAGE_SHIFT;
3082         mm->data_vm += len >> PAGE_SHIFT;
3083         if (flags & VM_LOCKED)
3084                 mm->locked_vm += (len >> PAGE_SHIFT);
3085         vma->vm_flags |= VM_SOFTDIRTY;
3086         return 0;
3087 }
3088
3089 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3090 {
3091         struct mm_struct *mm = current->mm;
3092         unsigned long len;
3093         int ret;
3094         bool populate;
3095         LIST_HEAD(uf);
3096
3097         len = PAGE_ALIGN(request);
3098         if (len < request)
3099                 return -ENOMEM;
3100         if (!len)
3101                 return 0;
3102
3103         if (mmap_write_lock_killable(mm))
3104                 return -EINTR;
3105
3106         ret = do_brk_flags(addr, len, flags, &uf);
3107         populate = ((mm->def_flags & VM_LOCKED) != 0);
3108         mmap_write_unlock(mm);
3109         userfaultfd_unmap_complete(mm, &uf);
3110         if (populate && !ret)
3111                 mm_populate(addr, len);
3112         return ret;
3113 }
3114 EXPORT_SYMBOL(vm_brk_flags);
3115
3116 int vm_brk(unsigned long addr, unsigned long len)
3117 {
3118         return vm_brk_flags(addr, len, 0);
3119 }
3120 EXPORT_SYMBOL(vm_brk);
3121
3122 /* Release all mmaps. */
3123 void exit_mmap(struct mm_struct *mm)
3124 {
3125         struct mmu_gather tlb;
3126         struct vm_area_struct *vma;
3127         unsigned long nr_accounted = 0;
3128
3129         /* mm's last user has gone, and its about to be pulled down */
3130         mmu_notifier_release(mm);
3131
3132         if (unlikely(mm_is_oom_victim(mm))) {
3133                 /*
3134                  * Manually reap the mm to free as much memory as possible.
3135                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3136                  * this mm from further consideration.  Taking mm->mmap_lock for
3137                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3138                  * reaper will not run on this mm again after mmap_lock is
3139                  * dropped.
3140                  *
3141                  * Nothing can be holding mm->mmap_lock here and the above call
3142                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3143                  * __oom_reap_task_mm() will not block.
3144                  *
3145                  * This needs to be done before calling munlock_vma_pages_all(),
3146                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3147                  * reliably test it.
3148                  */
3149                 (void)__oom_reap_task_mm(mm);
3150
3151                 set_bit(MMF_OOM_SKIP, &mm->flags);
3152                 mmap_write_lock(mm);
3153                 mmap_write_unlock(mm);
3154         }
3155
3156         if (mm->locked_vm)
3157                 unlock_range(mm->mmap, ULONG_MAX);
3158
3159         arch_exit_mmap(mm);
3160
3161         vma = mm->mmap;
3162         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3163                 return;
3164
3165         lru_add_drain();
3166         flush_cache_mm(mm);
3167         tlb_gather_mmu_fullmm(&tlb, mm);
3168         /* update_hiwater_rss(mm) here? but nobody should be looking */
3169         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3170         unmap_vmas(&tlb, vma, 0, -1);
3171         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3172         tlb_finish_mmu(&tlb);
3173
3174         /*
3175          * Walk the list again, actually closing and freeing it,
3176          * with preemption enabled, without holding any MM locks.
3177          */
3178         while (vma) {
3179                 if (vma->vm_flags & VM_ACCOUNT)
3180                         nr_accounted += vma_pages(vma);
3181                 vma = remove_vma(vma);
3182                 cond_resched();
3183         }
3184         vm_unacct_memory(nr_accounted);
3185 }
3186
3187 /* Insert vm structure into process list sorted by address
3188  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3189  * then i_mmap_rwsem is taken here.
3190  */
3191 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3192 {
3193         struct vm_area_struct *prev;
3194         struct rb_node **rb_link, *rb_parent;
3195
3196         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3197                            &prev, &rb_link, &rb_parent))
3198                 return -ENOMEM;
3199         if ((vma->vm_flags & VM_ACCOUNT) &&
3200              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3201                 return -ENOMEM;
3202
3203         /*
3204          * The vm_pgoff of a purely anonymous vma should be irrelevant
3205          * until its first write fault, when page's anon_vma and index
3206          * are set.  But now set the vm_pgoff it will almost certainly
3207          * end up with (unless mremap moves it elsewhere before that
3208          * first wfault), so /proc/pid/maps tells a consistent story.
3209          *
3210          * By setting it to reflect the virtual start address of the
3211          * vma, merges and splits can happen in a seamless way, just
3212          * using the existing file pgoff checks and manipulations.
3213          * Similarly in do_mmap and in do_brk_flags.
3214          */
3215         if (vma_is_anonymous(vma)) {
3216                 BUG_ON(vma->anon_vma);
3217                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3218         }
3219
3220         vma_link(mm, vma, prev, rb_link, rb_parent);
3221         return 0;
3222 }
3223
3224 /*
3225  * Copy the vma structure to a new location in the same mm,
3226  * prior to moving page table entries, to effect an mremap move.
3227  */
3228 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3229         unsigned long addr, unsigned long len, pgoff_t pgoff,
3230         bool *need_rmap_locks)
3231 {
3232         struct vm_area_struct *vma = *vmap;
3233         unsigned long vma_start = vma->vm_start;
3234         struct mm_struct *mm = vma->vm_mm;
3235         struct vm_area_struct *new_vma, *prev;
3236         struct rb_node **rb_link, *rb_parent;
3237         bool faulted_in_anon_vma = true;
3238
3239         /*
3240          * If anonymous vma has not yet been faulted, update new pgoff
3241          * to match new location, to increase its chance of merging.
3242          */
3243         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3244                 pgoff = addr >> PAGE_SHIFT;
3245                 faulted_in_anon_vma = false;
3246         }
3247
3248         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3249                 return NULL;    /* should never get here */
3250         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3251                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3252                             vma->vm_userfaultfd_ctx);
3253         if (new_vma) {
3254                 /*
3255                  * Source vma may have been merged into new_vma
3256                  */
3257                 if (unlikely(vma_start >= new_vma->vm_start &&
3258                              vma_start < new_vma->vm_end)) {
3259                         /*
3260                          * The only way we can get a vma_merge with
3261                          * self during an mremap is if the vma hasn't
3262                          * been faulted in yet and we were allowed to
3263                          * reset the dst vma->vm_pgoff to the
3264                          * destination address of the mremap to allow
3265                          * the merge to happen. mremap must change the
3266                          * vm_pgoff linearity between src and dst vmas
3267                          * (in turn preventing a vma_merge) to be
3268                          * safe. It is only safe to keep the vm_pgoff
3269                          * linear if there are no pages mapped yet.
3270                          */
3271                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3272                         *vmap = vma = new_vma;
3273                 }
3274                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3275         } else {
3276                 new_vma = vm_area_dup(vma);
3277                 if (!new_vma)
3278                         goto out;
3279                 new_vma->vm_start = addr;
3280                 new_vma->vm_end = addr + len;
3281                 new_vma->vm_pgoff = pgoff;
3282                 if (vma_dup_policy(vma, new_vma))
3283                         goto out_free_vma;
3284                 if (anon_vma_clone(new_vma, vma))
3285                         goto out_free_mempol;
3286                 if (new_vma->vm_file)
3287                         get_file(new_vma->vm_file);
3288                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3289                         new_vma->vm_ops->open(new_vma);
3290                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3291                 *need_rmap_locks = false;
3292         }
3293         return new_vma;
3294
3295 out_free_mempol:
3296         mpol_put(vma_policy(new_vma));
3297 out_free_vma:
3298         vm_area_free(new_vma);
3299 out:
3300         return NULL;
3301 }
3302
3303 /*
3304  * Return true if the calling process may expand its vm space by the passed
3305  * number of pages
3306  */
3307 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3308 {
3309         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3310                 return false;
3311
3312         if (is_data_mapping(flags) &&
3313             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3314                 /* Workaround for Valgrind */
3315                 if (rlimit(RLIMIT_DATA) == 0 &&
3316                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3317                         return true;
3318
3319                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3320                              current->comm, current->pid,
3321                              (mm->data_vm + npages) << PAGE_SHIFT,
3322                              rlimit(RLIMIT_DATA),
3323                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3324
3325                 if (!ignore_rlimit_data)
3326                         return false;
3327         }
3328
3329         return true;
3330 }
3331
3332 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3333 {
3334         WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3335
3336         if (is_exec_mapping(flags))
3337                 mm->exec_vm += npages;
3338         else if (is_stack_mapping(flags))
3339                 mm->stack_vm += npages;
3340         else if (is_data_mapping(flags))
3341                 mm->data_vm += npages;
3342 }
3343
3344 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3345
3346 /*
3347  * Having a close hook prevents vma merging regardless of flags.
3348  */
3349 static void special_mapping_close(struct vm_area_struct *vma)
3350 {
3351 }
3352
3353 static const char *special_mapping_name(struct vm_area_struct *vma)
3354 {
3355         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3356 }
3357
3358 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3359 {
3360         struct vm_special_mapping *sm = new_vma->vm_private_data;
3361
3362         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3363                 return -EFAULT;
3364
3365         if (sm->mremap)
3366                 return sm->mremap(sm, new_vma);
3367
3368         return 0;
3369 }
3370
3371 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3372 {
3373         /*
3374          * Forbid splitting special mappings - kernel has expectations over
3375          * the number of pages in mapping. Together with VM_DONTEXPAND
3376          * the size of vma should stay the same over the special mapping's
3377          * lifetime.
3378          */
3379         return -EINVAL;
3380 }
3381
3382 static const struct vm_operations_struct special_mapping_vmops = {
3383         .close = special_mapping_close,
3384         .fault = special_mapping_fault,
3385         .mremap = special_mapping_mremap,
3386         .name = special_mapping_name,
3387         /* vDSO code relies that VVAR can't be accessed remotely */
3388         .access = NULL,
3389         .may_split = special_mapping_split,
3390 };
3391
3392 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3393         .close = special_mapping_close,
3394         .fault = special_mapping_fault,
3395 };
3396
3397 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3398 {
3399         struct vm_area_struct *vma = vmf->vma;
3400         pgoff_t pgoff;
3401         struct page **pages;
3402
3403         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3404                 pages = vma->vm_private_data;
3405         } else {
3406                 struct vm_special_mapping *sm = vma->vm_private_data;
3407
3408                 if (sm->fault)
3409                         return sm->fault(sm, vmf->vma, vmf);
3410
3411                 pages = sm->pages;
3412         }
3413
3414         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3415                 pgoff--;
3416
3417         if (*pages) {
3418                 struct page *page = *pages;
3419                 get_page(page);
3420                 vmf->page = page;
3421                 return 0;
3422         }
3423
3424         return VM_FAULT_SIGBUS;
3425 }
3426
3427 static struct vm_area_struct *__install_special_mapping(
3428         struct mm_struct *mm,
3429         unsigned long addr, unsigned long len,
3430         unsigned long vm_flags, void *priv,
3431         const struct vm_operations_struct *ops)
3432 {
3433         int ret;
3434         struct vm_area_struct *vma;
3435
3436         vma = vm_area_alloc(mm);
3437         if (unlikely(vma == NULL))
3438                 return ERR_PTR(-ENOMEM);
3439
3440         vma->vm_start = addr;
3441         vma->vm_end = addr + len;
3442
3443         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3444         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3445
3446         vma->vm_ops = ops;
3447         vma->vm_private_data = priv;
3448
3449         ret = insert_vm_struct(mm, vma);
3450         if (ret)
3451                 goto out;
3452
3453         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3454
3455         perf_event_mmap(vma);
3456
3457         return vma;
3458
3459 out:
3460         vm_area_free(vma);
3461         return ERR_PTR(ret);
3462 }
3463
3464 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3465         const struct vm_special_mapping *sm)
3466 {
3467         return vma->vm_private_data == sm &&
3468                 (vma->vm_ops == &special_mapping_vmops ||
3469                  vma->vm_ops == &legacy_special_mapping_vmops);
3470 }
3471
3472 /*
3473  * Called with mm->mmap_lock held for writing.
3474  * Insert a new vma covering the given region, with the given flags.
3475  * Its pages are supplied by the given array of struct page *.
3476  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3477  * The region past the last page supplied will always produce SIGBUS.
3478  * The array pointer and the pages it points to are assumed to stay alive
3479  * for as long as this mapping might exist.
3480  */
3481 struct vm_area_struct *_install_special_mapping(
3482         struct mm_struct *mm,
3483         unsigned long addr, unsigned long len,
3484         unsigned long vm_flags, const struct vm_special_mapping *spec)
3485 {
3486         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3487                                         &special_mapping_vmops);
3488 }
3489
3490 int install_special_mapping(struct mm_struct *mm,
3491                             unsigned long addr, unsigned long len,
3492                             unsigned long vm_flags, struct page **pages)
3493 {
3494         struct vm_area_struct *vma = __install_special_mapping(
3495                 mm, addr, len, vm_flags, (void *)pages,
3496                 &legacy_special_mapping_vmops);
3497
3498         return PTR_ERR_OR_ZERO(vma);
3499 }
3500
3501 static DEFINE_MUTEX(mm_all_locks_mutex);
3502
3503 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3504 {
3505         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3506                 /*
3507                  * The LSB of head.next can't change from under us
3508                  * because we hold the mm_all_locks_mutex.
3509                  */
3510                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3511                 /*
3512                  * We can safely modify head.next after taking the
3513                  * anon_vma->root->rwsem. If some other vma in this mm shares
3514                  * the same anon_vma we won't take it again.
3515                  *
3516                  * No need of atomic instructions here, head.next
3517                  * can't change from under us thanks to the
3518                  * anon_vma->root->rwsem.
3519                  */
3520                 if (__test_and_set_bit(0, (unsigned long *)
3521                                        &anon_vma->root->rb_root.rb_root.rb_node))
3522                         BUG();
3523         }
3524 }
3525
3526 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3527 {
3528         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3529                 /*
3530                  * AS_MM_ALL_LOCKS can't change from under us because
3531                  * we hold the mm_all_locks_mutex.
3532                  *
3533                  * Operations on ->flags have to be atomic because
3534                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3535                  * mm_all_locks_mutex, there may be other cpus
3536                  * changing other bitflags in parallel to us.
3537                  */
3538                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3539                         BUG();
3540                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3541         }
3542 }
3543
3544 /*
3545  * This operation locks against the VM for all pte/vma/mm related
3546  * operations that could ever happen on a certain mm. This includes
3547  * vmtruncate, try_to_unmap, and all page faults.
3548  *
3549  * The caller must take the mmap_lock in write mode before calling
3550  * mm_take_all_locks(). The caller isn't allowed to release the
3551  * mmap_lock until mm_drop_all_locks() returns.
3552  *
3553  * mmap_lock in write mode is required in order to block all operations
3554  * that could modify pagetables and free pages without need of
3555  * altering the vma layout. It's also needed in write mode to avoid new
3556  * anon_vmas to be associated with existing vmas.
3557  *
3558  * A single task can't take more than one mm_take_all_locks() in a row
3559  * or it would deadlock.
3560  *
3561  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3562  * mapping->flags avoid to take the same lock twice, if more than one
3563  * vma in this mm is backed by the same anon_vma or address_space.
3564  *
3565  * We take locks in following order, accordingly to comment at beginning
3566  * of mm/rmap.c:
3567  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3568  *     hugetlb mapping);
3569  *   - all i_mmap_rwsem locks;
3570  *   - all anon_vma->rwseml
3571  *
3572  * We can take all locks within these types randomly because the VM code
3573  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3574  * mm_all_locks_mutex.
3575  *
3576  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3577  * that may have to take thousand of locks.
3578  *
3579  * mm_take_all_locks() can fail if it's interrupted by signals.
3580  */
3581 int mm_take_all_locks(struct mm_struct *mm)
3582 {
3583         struct vm_area_struct *vma;
3584         struct anon_vma_chain *avc;
3585
3586         BUG_ON(mmap_read_trylock(mm));
3587
3588         mutex_lock(&mm_all_locks_mutex);
3589
3590         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3591                 if (signal_pending(current))
3592                         goto out_unlock;
3593                 if (vma->vm_file && vma->vm_file->f_mapping &&
3594                                 is_vm_hugetlb_page(vma))
3595                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3596         }
3597
3598         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3599                 if (signal_pending(current))
3600                         goto out_unlock;
3601                 if (vma->vm_file && vma->vm_file->f_mapping &&
3602                                 !is_vm_hugetlb_page(vma))
3603                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3604         }
3605
3606         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3607                 if (signal_pending(current))
3608                         goto out_unlock;
3609                 if (vma->anon_vma)
3610                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3611                                 vm_lock_anon_vma(mm, avc->anon_vma);
3612         }
3613
3614         return 0;
3615
3616 out_unlock:
3617         mm_drop_all_locks(mm);
3618         return -EINTR;
3619 }
3620
3621 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3622 {
3623         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3624                 /*
3625                  * The LSB of head.next can't change to 0 from under
3626                  * us because we hold the mm_all_locks_mutex.
3627                  *
3628                  * We must however clear the bitflag before unlocking
3629                  * the vma so the users using the anon_vma->rb_root will
3630                  * never see our bitflag.
3631                  *
3632                  * No need of atomic instructions here, head.next
3633                  * can't change from under us until we release the
3634                  * anon_vma->root->rwsem.
3635                  */
3636                 if (!__test_and_clear_bit(0, (unsigned long *)
3637                                           &anon_vma->root->rb_root.rb_root.rb_node))
3638                         BUG();
3639                 anon_vma_unlock_write(anon_vma);
3640         }
3641 }
3642
3643 static void vm_unlock_mapping(struct address_space *mapping)
3644 {
3645         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3646                 /*
3647                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3648                  * because we hold the mm_all_locks_mutex.
3649                  */
3650                 i_mmap_unlock_write(mapping);
3651                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3652                                         &mapping->flags))
3653                         BUG();
3654         }
3655 }
3656
3657 /*
3658  * The mmap_lock cannot be released by the caller until
3659  * mm_drop_all_locks() returns.
3660  */
3661 void mm_drop_all_locks(struct mm_struct *mm)
3662 {
3663         struct vm_area_struct *vma;
3664         struct anon_vma_chain *avc;
3665
3666         BUG_ON(mmap_read_trylock(mm));
3667         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3668
3669         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3670                 if (vma->anon_vma)
3671                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3672                                 vm_unlock_anon_vma(avc->anon_vma);
3673                 if (vma->vm_file && vma->vm_file->f_mapping)
3674                         vm_unlock_mapping(vma->vm_file->f_mapping);
3675         }
3676
3677         mutex_unlock(&mm_all_locks_mutex);
3678 }
3679
3680 /*
3681  * initialise the percpu counter for VM
3682  */
3683 void __init mmap_init(void)
3684 {
3685         int ret;
3686
3687         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3688         VM_BUG_ON(ret);
3689 }
3690
3691 /*
3692  * Initialise sysctl_user_reserve_kbytes.
3693  *
3694  * This is intended to prevent a user from starting a single memory hogging
3695  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3696  * mode.
3697  *
3698  * The default value is min(3% of free memory, 128MB)
3699  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3700  */
3701 static int init_user_reserve(void)
3702 {
3703         unsigned long free_kbytes;
3704
3705         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3706
3707         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3708         return 0;
3709 }
3710 subsys_initcall(init_user_reserve);
3711
3712 /*
3713  * Initialise sysctl_admin_reserve_kbytes.
3714  *
3715  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3716  * to log in and kill a memory hogging process.
3717  *
3718  * Systems with more than 256MB will reserve 8MB, enough to recover
3719  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3720  * only reserve 3% of free pages by default.
3721  */
3722 static int init_admin_reserve(void)
3723 {
3724         unsigned long free_kbytes;
3725
3726         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3727
3728         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3729         return 0;
3730 }
3731 subsys_initcall(init_admin_reserve);
3732
3733 /*
3734  * Reinititalise user and admin reserves if memory is added or removed.
3735  *
3736  * The default user reserve max is 128MB, and the default max for the
3737  * admin reserve is 8MB. These are usually, but not always, enough to
3738  * enable recovery from a memory hogging process using login/sshd, a shell,
3739  * and tools like top. It may make sense to increase or even disable the
3740  * reserve depending on the existence of swap or variations in the recovery
3741  * tools. So, the admin may have changed them.
3742  *
3743  * If memory is added and the reserves have been eliminated or increased above
3744  * the default max, then we'll trust the admin.
3745  *
3746  * If memory is removed and there isn't enough free memory, then we
3747  * need to reset the reserves.
3748  *
3749  * Otherwise keep the reserve set by the admin.
3750  */
3751 static int reserve_mem_notifier(struct notifier_block *nb,
3752                              unsigned long action, void *data)
3753 {
3754         unsigned long tmp, free_kbytes;
3755
3756         switch (action) {
3757         case MEM_ONLINE:
3758                 /* Default max is 128MB. Leave alone if modified by operator. */
3759                 tmp = sysctl_user_reserve_kbytes;
3760                 if (0 < tmp && tmp < (1UL << 17))
3761                         init_user_reserve();
3762
3763                 /* Default max is 8MB.  Leave alone if modified by operator. */
3764                 tmp = sysctl_admin_reserve_kbytes;
3765                 if (0 < tmp && tmp < (1UL << 13))
3766                         init_admin_reserve();
3767
3768                 break;
3769         case MEM_OFFLINE:
3770                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3771
3772                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3773                         init_user_reserve();
3774                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3775                                 sysctl_user_reserve_kbytes);
3776                 }
3777
3778                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3779                         init_admin_reserve();
3780                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3781                                 sysctl_admin_reserve_kbytes);
3782                 }
3783                 break;
3784         default:
3785                 break;
3786         }
3787         return NOTIFY_OK;
3788 }
3789
3790 static struct notifier_block reserve_mem_nb = {
3791         .notifier_call = reserve_mem_notifier,
3792 };
3793
3794 static int __meminit init_reserve_notifier(void)
3795 {
3796         if (register_hotmemory_notifier(&reserve_mem_nb))
3797                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3798
3799         return 0;
3800 }
3801 subsys_initcall(init_reserve_notifier);