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 ucounts *ucounts = NULL;
1603                 struct hstate *hs;
1604
1605                 hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1606                 if (!hs)
1607                         return -EINVAL;
1608
1609                 len = ALIGN(len, huge_page_size(hs));
1610                 /*
1611                  * VM_NORESERVE is used because the reservations will be
1612                  * taken when vm_ops->mmap() is called
1613                  * A dummy user value is used because we are not locking
1614                  * memory so no accounting is necessary
1615                  */
1616                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1617                                 VM_NORESERVE,
1618                                 &ucounts, HUGETLB_ANONHUGE_INODE,
1619                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1620                 if (IS_ERR(file))
1621                         return PTR_ERR(file);
1622         }
1623
1624         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1625 out_fput:
1626         if (file)
1627                 fput(file);
1628         return retval;
1629 }
1630
1631 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1632                 unsigned long, prot, unsigned long, flags,
1633                 unsigned long, fd, unsigned long, pgoff)
1634 {
1635         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1636 }
1637
1638 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1639 struct mmap_arg_struct {
1640         unsigned long addr;
1641         unsigned long len;
1642         unsigned long prot;
1643         unsigned long flags;
1644         unsigned long fd;
1645         unsigned long offset;
1646 };
1647
1648 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1649 {
1650         struct mmap_arg_struct a;
1651
1652         if (copy_from_user(&a, arg, sizeof(a)))
1653                 return -EFAULT;
1654         if (offset_in_page(a.offset))
1655                 return -EINVAL;
1656
1657         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1658                                a.offset >> PAGE_SHIFT);
1659 }
1660 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1661
1662 /*
1663  * Some shared mappings will want the pages marked read-only
1664  * to track write events. If so, we'll downgrade vm_page_prot
1665  * to the private version (using protection_map[] without the
1666  * VM_SHARED bit).
1667  */
1668 int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1669 {
1670         vm_flags_t vm_flags = vma->vm_flags;
1671         const struct vm_operations_struct *vm_ops = vma->vm_ops;
1672
1673         /* If it was private or non-writable, the write bit is already clear */
1674         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1675                 return 0;
1676
1677         /* The backer wishes to know when pages are first written to? */
1678         if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1679                 return 1;
1680
1681         /* The open routine did something to the protections that pgprot_modify
1682          * won't preserve? */
1683         if (pgprot_val(vm_page_prot) !=
1684             pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1685                 return 0;
1686
1687         /* Do we need to track softdirty? */
1688         if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
1689                 return 1;
1690
1691         /* Specialty mapping? */
1692         if (vm_flags & VM_PFNMAP)
1693                 return 0;
1694
1695         /* Can the mapping track the dirty pages? */
1696         return vma->vm_file && vma->vm_file->f_mapping &&
1697                 mapping_can_writeback(vma->vm_file->f_mapping);
1698 }
1699
1700 /*
1701  * We account for memory if it's a private writeable mapping,
1702  * not hugepages and VM_NORESERVE wasn't set.
1703  */
1704 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1705 {
1706         /*
1707          * hugetlb has its own accounting separate from the core VM
1708          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1709          */
1710         if (file && is_file_hugepages(file))
1711                 return 0;
1712
1713         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1714 }
1715
1716 unsigned long mmap_region(struct file *file, unsigned long addr,
1717                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1718                 struct list_head *uf)
1719 {
1720         struct mm_struct *mm = current->mm;
1721         struct vm_area_struct *vma, *prev, *merge;
1722         int error;
1723         struct rb_node **rb_link, *rb_parent;
1724         unsigned long charged = 0;
1725
1726         /* Check against address space limit. */
1727         if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1728                 unsigned long nr_pages;
1729
1730                 /*
1731                  * MAP_FIXED may remove pages of mappings that intersects with
1732                  * requested mapping. Account for the pages it would unmap.
1733                  */
1734                 nr_pages = count_vma_pages_range(mm, addr, addr + len);
1735
1736                 if (!may_expand_vm(mm, vm_flags,
1737                                         (len >> PAGE_SHIFT) - nr_pages))
1738                         return -ENOMEM;
1739         }
1740
1741         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
1742         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
1743                 return -ENOMEM;
1744         /*
1745          * Private writable mapping: check memory availability
1746          */
1747         if (accountable_mapping(file, vm_flags)) {
1748                 charged = len >> PAGE_SHIFT;
1749                 if (security_vm_enough_memory_mm(mm, charged))
1750                         return -ENOMEM;
1751                 vm_flags |= VM_ACCOUNT;
1752         }
1753
1754         /*
1755          * Can we just expand an old mapping?
1756          */
1757         vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1758                         NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1759         if (vma)
1760                 goto out;
1761
1762         /*
1763          * Determine the object being mapped and call the appropriate
1764          * specific mapper. the address has already been validated, but
1765          * not unmapped, but the maps are removed from the list.
1766          */
1767         vma = vm_area_alloc(mm);
1768         if (!vma) {
1769                 error = -ENOMEM;
1770                 goto unacct_error;
1771         }
1772
1773         vma->vm_start = addr;
1774         vma->vm_end = addr + len;
1775         vma->vm_flags = vm_flags;
1776         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1777         vma->vm_pgoff = pgoff;
1778
1779         if (file) {
1780                 if (vm_flags & VM_SHARED) {
1781                         error = mapping_map_writable(file->f_mapping);
1782                         if (error)
1783                                 goto free_vma;
1784                 }
1785
1786                 vma->vm_file = get_file(file);
1787                 error = call_mmap(file, vma);
1788                 if (error)
1789                         goto unmap_and_free_vma;
1790
1791                 /* Can addr have changed??
1792                  *
1793                  * Answer: Yes, several device drivers can do it in their
1794                  *         f_op->mmap method. -DaveM
1795                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1796                  *      be updated for vma_link()
1797                  */
1798                 WARN_ON_ONCE(addr != vma->vm_start);
1799
1800                 addr = vma->vm_start;
1801
1802                 /* If vm_flags changed after call_mmap(), we should try merge vma again
1803                  * as we may succeed this time.
1804                  */
1805                 if (unlikely(vm_flags != vma->vm_flags && prev)) {
1806                         merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
1807                                 NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX);
1808                         if (merge) {
1809                                 /* ->mmap() can change vma->vm_file and fput the original file. So
1810                                  * fput the vma->vm_file here or we would add an extra fput for file
1811                                  * and cause general protection fault ultimately.
1812                                  */
1813                                 fput(vma->vm_file);
1814                                 vm_area_free(vma);
1815                                 vma = merge;
1816                                 /* Update vm_flags to pick up the change. */
1817                                 vm_flags = vma->vm_flags;
1818                                 goto unmap_writable;
1819                         }
1820                 }
1821
1822                 vm_flags = vma->vm_flags;
1823         } else if (vm_flags & VM_SHARED) {
1824                 error = shmem_zero_setup(vma);
1825                 if (error)
1826                         goto free_vma;
1827         } else {
1828                 vma_set_anonymous(vma);
1829         }
1830
1831         /* Allow architectures to sanity-check the vm_flags */
1832         if (!arch_validate_flags(vma->vm_flags)) {
1833                 error = -EINVAL;
1834                 if (file)
1835                         goto unmap_and_free_vma;
1836                 else
1837                         goto free_vma;
1838         }
1839
1840         vma_link(mm, vma, prev, rb_link, rb_parent);
1841         /* Once vma denies write, undo our temporary denial count */
1842 unmap_writable:
1843         if (file && vm_flags & VM_SHARED)
1844                 mapping_unmap_writable(file->f_mapping);
1845         file = vma->vm_file;
1846 out:
1847         perf_event_mmap(vma);
1848
1849         vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1850         if (vm_flags & VM_LOCKED) {
1851                 if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
1852                                         is_vm_hugetlb_page(vma) ||
1853                                         vma == get_gate_vma(current->mm))
1854                         vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1855                 else
1856                         mm->locked_vm += (len >> PAGE_SHIFT);
1857         }
1858
1859         if (file)
1860                 uprobe_mmap(vma);
1861
1862         /*
1863          * New (or expanded) vma always get soft dirty status.
1864          * Otherwise user-space soft-dirty page tracker won't
1865          * be able to distinguish situation when vma area unmapped,
1866          * then new mapped in-place (which must be aimed as
1867          * a completely new data area).
1868          */
1869         vma->vm_flags |= VM_SOFTDIRTY;
1870
1871         vma_set_page_prot(vma);
1872
1873         return addr;
1874
1875 unmap_and_free_vma:
1876         fput(vma->vm_file);
1877         vma->vm_file = NULL;
1878
1879         /* Undo any partial mapping done by a device driver. */
1880         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1881         charged = 0;
1882         if (vm_flags & VM_SHARED)
1883                 mapping_unmap_writable(file->f_mapping);
1884 free_vma:
1885         vm_area_free(vma);
1886 unacct_error:
1887         if (charged)
1888                 vm_unacct_memory(charged);
1889         return error;
1890 }
1891
1892 static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1893 {
1894         /*
1895          * We implement the search by looking for an rbtree node that
1896          * immediately follows a suitable gap. That is,
1897          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1898          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1899          * - gap_end - gap_start >= length
1900          */
1901
1902         struct mm_struct *mm = current->mm;
1903         struct vm_area_struct *vma;
1904         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1905
1906         /* Adjust search length to account for worst case alignment overhead */
1907         length = info->length + info->align_mask;
1908         if (length < info->length)
1909                 return -ENOMEM;
1910
1911         /* Adjust search limits by the desired length */
1912         if (info->high_limit < length)
1913                 return -ENOMEM;
1914         high_limit = info->high_limit - length;
1915
1916         if (info->low_limit > high_limit)
1917                 return -ENOMEM;
1918         low_limit = info->low_limit + length;
1919
1920         /* Check if rbtree root looks promising */
1921         if (RB_EMPTY_ROOT(&mm->mm_rb))
1922                 goto check_highest;
1923         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1924         if (vma->rb_subtree_gap < length)
1925                 goto check_highest;
1926
1927         while (true) {
1928                 /* Visit left subtree if it looks promising */
1929                 gap_end = vm_start_gap(vma);
1930                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1931                         struct vm_area_struct *left =
1932                                 rb_entry(vma->vm_rb.rb_left,
1933                                          struct vm_area_struct, vm_rb);
1934                         if (left->rb_subtree_gap >= length) {
1935                                 vma = left;
1936                                 continue;
1937                         }
1938                 }
1939
1940                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1941 check_current:
1942                 /* Check if current node has a suitable gap */
1943                 if (gap_start > high_limit)
1944                         return -ENOMEM;
1945                 if (gap_end >= low_limit &&
1946                     gap_end > gap_start && gap_end - gap_start >= length)
1947                         goto found;
1948
1949                 /* Visit right subtree if it looks promising */
1950                 if (vma->vm_rb.rb_right) {
1951                         struct vm_area_struct *right =
1952                                 rb_entry(vma->vm_rb.rb_right,
1953                                          struct vm_area_struct, vm_rb);
1954                         if (right->rb_subtree_gap >= length) {
1955                                 vma = right;
1956                                 continue;
1957                         }
1958                 }
1959
1960                 /* Go back up the rbtree to find next candidate node */
1961                 while (true) {
1962                         struct rb_node *prev = &vma->vm_rb;
1963                         if (!rb_parent(prev))
1964                                 goto check_highest;
1965                         vma = rb_entry(rb_parent(prev),
1966                                        struct vm_area_struct, vm_rb);
1967                         if (prev == vma->vm_rb.rb_left) {
1968                                 gap_start = vm_end_gap(vma->vm_prev);
1969                                 gap_end = vm_start_gap(vma);
1970                                 goto check_current;
1971                         }
1972                 }
1973         }
1974
1975 check_highest:
1976         /* Check highest gap, which does not precede any rbtree node */
1977         gap_start = mm->highest_vm_end;
1978         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1979         if (gap_start > high_limit)
1980                 return -ENOMEM;
1981
1982 found:
1983         /* We found a suitable gap. Clip it with the original low_limit. */
1984         if (gap_start < info->low_limit)
1985                 gap_start = info->low_limit;
1986
1987         /* Adjust gap address to the desired alignment */
1988         gap_start += (info->align_offset - gap_start) & info->align_mask;
1989
1990         VM_BUG_ON(gap_start + info->length > info->high_limit);
1991         VM_BUG_ON(gap_start + info->length > gap_end);
1992         return gap_start;
1993 }
1994
1995 static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1996 {
1997         struct mm_struct *mm = current->mm;
1998         struct vm_area_struct *vma;
1999         unsigned long length, low_limit, high_limit, gap_start, gap_end;
2000
2001         /* Adjust search length to account for worst case alignment overhead */
2002         length = info->length + info->align_mask;
2003         if (length < info->length)
2004                 return -ENOMEM;
2005
2006         /*
2007          * Adjust search limits by the desired length.
2008          * See implementation comment at top of unmapped_area().
2009          */
2010         gap_end = info->high_limit;
2011         if (gap_end < length)
2012                 return -ENOMEM;
2013         high_limit = gap_end - length;
2014
2015         if (info->low_limit > high_limit)
2016                 return -ENOMEM;
2017         low_limit = info->low_limit + length;
2018
2019         /* Check highest gap, which does not precede any rbtree node */
2020         gap_start = mm->highest_vm_end;
2021         if (gap_start <= high_limit)
2022                 goto found_highest;
2023
2024         /* Check if rbtree root looks promising */
2025         if (RB_EMPTY_ROOT(&mm->mm_rb))
2026                 return -ENOMEM;
2027         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
2028         if (vma->rb_subtree_gap < length)
2029                 return -ENOMEM;
2030
2031         while (true) {
2032                 /* Visit right subtree if it looks promising */
2033                 gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
2034                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
2035                         struct vm_area_struct *right =
2036                                 rb_entry(vma->vm_rb.rb_right,
2037                                          struct vm_area_struct, vm_rb);
2038                         if (right->rb_subtree_gap >= length) {
2039                                 vma = right;
2040                                 continue;
2041                         }
2042                 }
2043
2044 check_current:
2045                 /* Check if current node has a suitable gap */
2046                 gap_end = vm_start_gap(vma);
2047                 if (gap_end < low_limit)
2048                         return -ENOMEM;
2049                 if (gap_start <= high_limit &&
2050                     gap_end > gap_start && gap_end - gap_start >= length)
2051                         goto found;
2052
2053                 /* Visit left subtree if it looks promising */
2054                 if (vma->vm_rb.rb_left) {
2055                         struct vm_area_struct *left =
2056                                 rb_entry(vma->vm_rb.rb_left,
2057                                          struct vm_area_struct, vm_rb);
2058                         if (left->rb_subtree_gap >= length) {
2059                                 vma = left;
2060                                 continue;
2061                         }
2062                 }
2063
2064                 /* Go back up the rbtree to find next candidate node */
2065                 while (true) {
2066                         struct rb_node *prev = &vma->vm_rb;
2067                         if (!rb_parent(prev))
2068                                 return -ENOMEM;
2069                         vma = rb_entry(rb_parent(prev),
2070                                        struct vm_area_struct, vm_rb);
2071                         if (prev == vma->vm_rb.rb_right) {
2072                                 gap_start = vma->vm_prev ?
2073                                         vm_end_gap(vma->vm_prev) : 0;
2074                                 goto check_current;
2075                         }
2076                 }
2077         }
2078
2079 found:
2080         /* We found a suitable gap. Clip it with the original high_limit. */
2081         if (gap_end > info->high_limit)
2082                 gap_end = info->high_limit;
2083
2084 found_highest:
2085         /* Compute highest gap address at the desired alignment */
2086         gap_end -= info->length;
2087         gap_end -= (gap_end - info->align_offset) & info->align_mask;
2088
2089         VM_BUG_ON(gap_end < info->low_limit);
2090         VM_BUG_ON(gap_end < gap_start);
2091         return gap_end;
2092 }
2093
2094 /*
2095  * Search for an unmapped address range.
2096  *
2097  * We are looking for a range that:
2098  * - does not intersect with any VMA;
2099  * - is contained within the [low_limit, high_limit) interval;
2100  * - is at least the desired size.
2101  * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
2102  */
2103 unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
2104 {
2105         unsigned long addr;
2106
2107         if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
2108                 addr = unmapped_area_topdown(info);
2109         else
2110                 addr = unmapped_area(info);
2111
2112         trace_vm_unmapped_area(addr, info);
2113         return addr;
2114 }
2115
2116 /* Get an address range which is currently unmapped.
2117  * For shmat() with addr=0.
2118  *
2119  * Ugly calling convention alert:
2120  * Return value with the low bits set means error value,
2121  * ie
2122  *      if (ret & ~PAGE_MASK)
2123  *              error = ret;
2124  *
2125  * This function "knows" that -ENOMEM has the bits set.
2126  */
2127 #ifndef HAVE_ARCH_UNMAPPED_AREA
2128 unsigned long
2129 arch_get_unmapped_area(struct file *filp, unsigned long addr,
2130                 unsigned long len, unsigned long pgoff, unsigned long flags)
2131 {
2132         struct mm_struct *mm = current->mm;
2133         struct vm_area_struct *vma, *prev;
2134         struct vm_unmapped_area_info info;
2135         const unsigned long mmap_end = arch_get_mmap_end(addr);
2136
2137         if (len > mmap_end - mmap_min_addr)
2138                 return -ENOMEM;
2139
2140         if (flags & MAP_FIXED)
2141                 return addr;
2142
2143         if (addr) {
2144                 addr = PAGE_ALIGN(addr);
2145                 vma = find_vma_prev(mm, addr, &prev);
2146                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2147                     (!vma || addr + len <= vm_start_gap(vma)) &&
2148                     (!prev || addr >= vm_end_gap(prev)))
2149                         return addr;
2150         }
2151
2152         info.flags = 0;
2153         info.length = len;
2154         info.low_limit = mm->mmap_base;
2155         info.high_limit = mmap_end;
2156         info.align_mask = 0;
2157         info.align_offset = 0;
2158         return vm_unmapped_area(&info);
2159 }
2160 #endif
2161
2162 /*
2163  * This mmap-allocator allocates new areas top-down from below the
2164  * stack's low limit (the base):
2165  */
2166 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2167 unsigned long
2168 arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
2169                           unsigned long len, unsigned long pgoff,
2170                           unsigned long flags)
2171 {
2172         struct vm_area_struct *vma, *prev;
2173         struct mm_struct *mm = current->mm;
2174         struct vm_unmapped_area_info info;
2175         const unsigned long mmap_end = arch_get_mmap_end(addr);
2176
2177         /* requested length too big for entire address space */
2178         if (len > mmap_end - mmap_min_addr)
2179                 return -ENOMEM;
2180
2181         if (flags & MAP_FIXED)
2182                 return addr;
2183
2184         /* requesting a specific address */
2185         if (addr) {
2186                 addr = PAGE_ALIGN(addr);
2187                 vma = find_vma_prev(mm, addr, &prev);
2188                 if (mmap_end - len >= addr && addr >= mmap_min_addr &&
2189                                 (!vma || addr + len <= vm_start_gap(vma)) &&
2190                                 (!prev || addr >= vm_end_gap(prev)))
2191                         return addr;
2192         }
2193
2194         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2195         info.length = len;
2196         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2197         info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
2198         info.align_mask = 0;
2199         info.align_offset = 0;
2200         addr = vm_unmapped_area(&info);
2201
2202         /*
2203          * A failed mmap() very likely causes application failure,
2204          * so fall back to the bottom-up function here. This scenario
2205          * can happen with large stack limits and large mmap()
2206          * allocations.
2207          */
2208         if (offset_in_page(addr)) {
2209                 VM_BUG_ON(addr != -ENOMEM);
2210                 info.flags = 0;
2211                 info.low_limit = TASK_UNMAPPED_BASE;
2212                 info.high_limit = mmap_end;
2213                 addr = vm_unmapped_area(&info);
2214         }
2215
2216         return addr;
2217 }
2218 #endif
2219
2220 unsigned long
2221 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2222                 unsigned long pgoff, unsigned long flags)
2223 {
2224         unsigned long (*get_area)(struct file *, unsigned long,
2225                                   unsigned long, unsigned long, unsigned long);
2226
2227         unsigned long error = arch_mmap_check(addr, len, flags);
2228         if (error)
2229                 return error;
2230
2231         /* Careful about overflows.. */
2232         if (len > TASK_SIZE)
2233                 return -ENOMEM;
2234
2235         get_area = current->mm->get_unmapped_area;
2236         if (file) {
2237                 if (file->f_op->get_unmapped_area)
2238                         get_area = file->f_op->get_unmapped_area;
2239         } else if (flags & MAP_SHARED) {
2240                 /*
2241                  * mmap_region() will call shmem_zero_setup() to create a file,
2242                  * so use shmem's get_unmapped_area in case it can be huge.
2243                  * do_mmap() will clear pgoff, so match alignment.
2244                  */
2245                 pgoff = 0;
2246                 get_area = shmem_get_unmapped_area;
2247         }
2248
2249         addr = get_area(file, addr, len, pgoff, flags);
2250         if (IS_ERR_VALUE(addr))
2251                 return addr;
2252
2253         if (addr > TASK_SIZE - len)
2254                 return -ENOMEM;
2255         if (offset_in_page(addr))
2256                 return -EINVAL;
2257
2258         error = security_mmap_addr(addr);
2259         return error ? error : addr;
2260 }
2261
2262 EXPORT_SYMBOL(get_unmapped_area);
2263
2264 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2265 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
2266 {
2267         struct rb_node *rb_node;
2268         struct vm_area_struct *vma;
2269
2270         mmap_assert_locked(mm);
2271         /* Check the cache first. */
2272         vma = vmacache_find(mm, addr);
2273         if (likely(vma))
2274                 return vma;
2275
2276         rb_node = mm->mm_rb.rb_node;
2277
2278         while (rb_node) {
2279                 struct vm_area_struct *tmp;
2280
2281                 tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2282
2283                 if (tmp->vm_end > addr) {
2284                         vma = tmp;
2285                         if (tmp->vm_start <= addr)
2286                                 break;
2287                         rb_node = rb_node->rb_left;
2288                 } else
2289                         rb_node = rb_node->rb_right;
2290         }
2291
2292         if (vma)
2293                 vmacache_update(addr, vma);
2294         return vma;
2295 }
2296
2297 EXPORT_SYMBOL(find_vma);
2298
2299 /*
2300  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2301  */
2302 struct vm_area_struct *
2303 find_vma_prev(struct mm_struct *mm, unsigned long addr,
2304                         struct vm_area_struct **pprev)
2305 {
2306         struct vm_area_struct *vma;
2307
2308         vma = find_vma(mm, addr);
2309         if (vma) {
2310                 *pprev = vma->vm_prev;
2311         } else {
2312                 struct rb_node *rb_node = rb_last(&mm->mm_rb);
2313
2314                 *pprev = rb_node ? rb_entry(rb_node, struct vm_area_struct, vm_rb) : NULL;
2315         }
2316         return vma;
2317 }
2318
2319 /*
2320  * Verify that the stack growth is acceptable and
2321  * update accounting. This is shared with both the
2322  * grow-up and grow-down cases.
2323  */
2324 static int acct_stack_growth(struct vm_area_struct *vma,
2325                              unsigned long size, unsigned long grow)
2326 {
2327         struct mm_struct *mm = vma->vm_mm;
2328         unsigned long new_start;
2329
2330         /* address space limit tests */
2331         if (!may_expand_vm(mm, vma->vm_flags, grow))
2332                 return -ENOMEM;
2333
2334         /* Stack limit test */
2335         if (size > rlimit(RLIMIT_STACK))
2336                 return -ENOMEM;
2337
2338         /* mlock limit tests */
2339         if (vma->vm_flags & VM_LOCKED) {
2340                 unsigned long locked;
2341                 unsigned long limit;
2342                 locked = mm->locked_vm + grow;
2343                 limit = rlimit(RLIMIT_MEMLOCK);
2344                 limit >>= PAGE_SHIFT;
2345                 if (locked > limit && !capable(CAP_IPC_LOCK))
2346                         return -ENOMEM;
2347         }
2348
2349         /* Check to ensure the stack will not grow into a hugetlb-only region */
2350         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2351                         vma->vm_end - size;
2352         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2353                 return -EFAULT;
2354
2355         /*
2356          * Overcommit..  This must be the final test, as it will
2357          * update security statistics.
2358          */
2359         if (security_vm_enough_memory_mm(mm, grow))
2360                 return -ENOMEM;
2361
2362         return 0;
2363 }
2364
2365 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2366 /*
2367  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2368  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2369  */
2370 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2371 {
2372         struct mm_struct *mm = vma->vm_mm;
2373         struct vm_area_struct *next;
2374         unsigned long gap_addr;
2375         int error = 0;
2376
2377         if (!(vma->vm_flags & VM_GROWSUP))
2378                 return -EFAULT;
2379
2380         /* Guard against exceeding limits of the address space. */
2381         address &= PAGE_MASK;
2382         if (address >= (TASK_SIZE & PAGE_MASK))
2383                 return -ENOMEM;
2384         address += PAGE_SIZE;
2385
2386         /* Enforce stack_guard_gap */
2387         gap_addr = address + stack_guard_gap;
2388
2389         /* Guard against overflow */
2390         if (gap_addr < address || gap_addr > TASK_SIZE)
2391                 gap_addr = TASK_SIZE;
2392
2393         next = vma->vm_next;
2394         if (next && next->vm_start < gap_addr && vma_is_accessible(next)) {
2395                 if (!(next->vm_flags & VM_GROWSUP))
2396                         return -ENOMEM;
2397                 /* Check that both stack segments have the same anon_vma? */
2398         }
2399
2400         /* We must make sure the anon_vma is allocated. */
2401         if (unlikely(anon_vma_prepare(vma)))
2402                 return -ENOMEM;
2403
2404         /*
2405          * vma->vm_start/vm_end cannot change under us because the caller
2406          * is required to hold the mmap_lock in read mode.  We need the
2407          * anon_vma lock to serialize against concurrent expand_stacks.
2408          */
2409         anon_vma_lock_write(vma->anon_vma);
2410
2411         /* Somebody else might have raced and expanded it already */
2412         if (address > vma->vm_end) {
2413                 unsigned long size, grow;
2414
2415                 size = address - vma->vm_start;
2416                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2417
2418                 error = -ENOMEM;
2419                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2420                         error = acct_stack_growth(vma, size, grow);
2421                         if (!error) {
2422                                 /*
2423                                  * vma_gap_update() doesn't support concurrent
2424                                  * updates, but we only hold a shared mmap_lock
2425                                  * lock here, so we need to protect against
2426                                  * concurrent vma expansions.
2427                                  * anon_vma_lock_write() doesn't help here, as
2428                                  * we don't guarantee that all growable vmas
2429                                  * in a mm share the same root anon vma.
2430                                  * So, we reuse mm->page_table_lock to guard
2431                                  * against concurrent vma expansions.
2432                                  */
2433                                 spin_lock(&mm->page_table_lock);
2434                                 if (vma->vm_flags & VM_LOCKED)
2435                                         mm->locked_vm += grow;
2436                                 vm_stat_account(mm, vma->vm_flags, grow);
2437                                 anon_vma_interval_tree_pre_update_vma(vma);
2438                                 vma->vm_end = address;
2439                                 anon_vma_interval_tree_post_update_vma(vma);
2440                                 if (vma->vm_next)
2441                                         vma_gap_update(vma->vm_next);
2442                                 else
2443                                         mm->highest_vm_end = vm_end_gap(vma);
2444                                 spin_unlock(&mm->page_table_lock);
2445
2446                                 perf_event_mmap(vma);
2447                         }
2448                 }
2449         }
2450         anon_vma_unlock_write(vma->anon_vma);
2451         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2452         validate_mm(mm);
2453         return error;
2454 }
2455 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2456
2457 /*
2458  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2459  */
2460 int expand_downwards(struct vm_area_struct *vma,
2461                                    unsigned long address)
2462 {
2463         struct mm_struct *mm = vma->vm_mm;
2464         struct vm_area_struct *prev;
2465         int error = 0;
2466
2467         address &= PAGE_MASK;
2468         if (address < mmap_min_addr)
2469                 return -EPERM;
2470
2471         /* Enforce stack_guard_gap */
2472         prev = vma->vm_prev;
2473         /* Check that both stack segments have the same anon_vma? */
2474         if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2475                         vma_is_accessible(prev)) {
2476                 if (address - prev->vm_end < stack_guard_gap)
2477                         return -ENOMEM;
2478         }
2479
2480         /* We must make sure the anon_vma is allocated. */
2481         if (unlikely(anon_vma_prepare(vma)))
2482                 return -ENOMEM;
2483
2484         /*
2485          * vma->vm_start/vm_end cannot change under us because the caller
2486          * is required to hold the mmap_lock in read mode.  We need the
2487          * anon_vma lock to serialize against concurrent expand_stacks.
2488          */
2489         anon_vma_lock_write(vma->anon_vma);
2490
2491         /* Somebody else might have raced and expanded it already */
2492         if (address < vma->vm_start) {
2493                 unsigned long size, grow;
2494
2495                 size = vma->vm_end - address;
2496                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2497
2498                 error = -ENOMEM;
2499                 if (grow <= vma->vm_pgoff) {
2500                         error = acct_stack_growth(vma, size, grow);
2501                         if (!error) {
2502                                 /*
2503                                  * vma_gap_update() doesn't support concurrent
2504                                  * updates, but we only hold a shared mmap_lock
2505                                  * lock here, so we need to protect against
2506                                  * concurrent vma expansions.
2507                                  * anon_vma_lock_write() doesn't help here, as
2508                                  * we don't guarantee that all growable vmas
2509                                  * in a mm share the same root anon vma.
2510                                  * So, we reuse mm->page_table_lock to guard
2511                                  * against concurrent vma expansions.
2512                                  */
2513                                 spin_lock(&mm->page_table_lock);
2514                                 if (vma->vm_flags & VM_LOCKED)
2515                                         mm->locked_vm += grow;
2516                                 vm_stat_account(mm, vma->vm_flags, grow);
2517                                 anon_vma_interval_tree_pre_update_vma(vma);
2518                                 vma->vm_start = address;
2519                                 vma->vm_pgoff -= grow;
2520                                 anon_vma_interval_tree_post_update_vma(vma);
2521                                 vma_gap_update(vma);
2522                                 spin_unlock(&mm->page_table_lock);
2523
2524                                 perf_event_mmap(vma);
2525                         }
2526                 }
2527         }
2528         anon_vma_unlock_write(vma->anon_vma);
2529         khugepaged_enter_vma_merge(vma, vma->vm_flags);
2530         validate_mm(mm);
2531         return error;
2532 }
2533
2534 /* enforced gap between the expanding stack and other mappings. */
2535 unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2536
2537 static int __init cmdline_parse_stack_guard_gap(char *p)
2538 {
2539         unsigned long val;
2540         char *endptr;
2541
2542         val = simple_strtoul(p, &endptr, 10);
2543         if (!*endptr)
2544                 stack_guard_gap = val << PAGE_SHIFT;
2545
2546         return 1;
2547 }
2548 __setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2549
2550 #ifdef CONFIG_STACK_GROWSUP
2551 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2552 {
2553         return expand_upwards(vma, address);
2554 }
2555
2556 struct vm_area_struct *
2557 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2558 {
2559         struct vm_area_struct *vma, *prev;
2560
2561         addr &= PAGE_MASK;
2562         vma = find_vma_prev(mm, addr, &prev);
2563         if (vma && (vma->vm_start <= addr))
2564                 return vma;
2565         /* don't alter vm_end if the coredump is running */
2566         if (!prev || expand_stack(prev, addr))
2567                 return NULL;
2568         if (prev->vm_flags & VM_LOCKED)
2569                 populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2570         return prev;
2571 }
2572 #else
2573 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2574 {
2575         return expand_downwards(vma, address);
2576 }
2577
2578 struct vm_area_struct *
2579 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2580 {
2581         struct vm_area_struct *vma;
2582         unsigned long start;
2583
2584         addr &= PAGE_MASK;
2585         vma = find_vma(mm, addr);
2586         if (!vma)
2587                 return NULL;
2588         if (vma->vm_start <= addr)
2589                 return vma;
2590         if (!(vma->vm_flags & VM_GROWSDOWN))
2591                 return NULL;
2592         start = vma->vm_start;
2593         if (expand_stack(vma, addr))
2594                 return NULL;
2595         if (vma->vm_flags & VM_LOCKED)
2596                 populate_vma_page_range(vma, addr, start, NULL);
2597         return vma;
2598 }
2599 #endif
2600
2601 EXPORT_SYMBOL_GPL(find_extend_vma);
2602
2603 /*
2604  * Ok - we have the memory areas we should free on the vma list,
2605  * so release them, and do the vma updates.
2606  *
2607  * Called with the mm semaphore held.
2608  */
2609 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2610 {
2611         unsigned long nr_accounted = 0;
2612
2613         /* Update high watermark before we lower total_vm */
2614         update_hiwater_vm(mm);
2615         do {
2616                 long nrpages = vma_pages(vma);
2617
2618                 if (vma->vm_flags & VM_ACCOUNT)
2619                         nr_accounted += nrpages;
2620                 vm_stat_account(mm, vma->vm_flags, -nrpages);
2621                 vma = remove_vma(vma);
2622         } while (vma);
2623         vm_unacct_memory(nr_accounted);
2624         validate_mm(mm);
2625 }
2626
2627 /*
2628  * Get rid of page table information in the indicated region.
2629  *
2630  * Called with the mm semaphore held.
2631  */
2632 static void unmap_region(struct mm_struct *mm,
2633                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2634                 unsigned long start, unsigned long end)
2635 {
2636         struct vm_area_struct *next = vma_next(mm, prev);
2637         struct mmu_gather tlb;
2638
2639         lru_add_drain();
2640         tlb_gather_mmu(&tlb, mm);
2641         update_hiwater_rss(mm);
2642         unmap_vmas(&tlb, vma, start, end);
2643         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2644                                  next ? next->vm_start : USER_PGTABLES_CEILING);
2645         tlb_finish_mmu(&tlb);
2646 }
2647
2648 /*
2649  * Create a list of vma's touched by the unmap, removing them from the mm's
2650  * vma list as we go..
2651  */
2652 static bool
2653 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2654         struct vm_area_struct *prev, unsigned long end)
2655 {
2656         struct vm_area_struct **insertion_point;
2657         struct vm_area_struct *tail_vma = NULL;
2658
2659         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2660         vma->vm_prev = NULL;
2661         do {
2662                 vma_rb_erase(vma, &mm->mm_rb);
2663                 mm->map_count--;
2664                 tail_vma = vma;
2665                 vma = vma->vm_next;
2666         } while (vma && vma->vm_start < end);
2667         *insertion_point = vma;
2668         if (vma) {
2669                 vma->vm_prev = prev;
2670                 vma_gap_update(vma);
2671         } else
2672                 mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2673         tail_vma->vm_next = NULL;
2674
2675         /* Kill the cache */
2676         vmacache_invalidate(mm);
2677
2678         /*
2679          * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2680          * VM_GROWSUP VMA. Such VMAs can change their size under
2681          * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2682          */
2683         if (vma && (vma->vm_flags & VM_GROWSDOWN))
2684                 return false;
2685         if (prev && (prev->vm_flags & VM_GROWSUP))
2686                 return false;
2687         return true;
2688 }
2689
2690 /*
2691  * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2692  * has already been checked or doesn't make sense to fail.
2693  */
2694 int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2695                 unsigned long addr, int new_below)
2696 {
2697         struct vm_area_struct *new;
2698         int err;
2699
2700         if (vma->vm_ops && vma->vm_ops->may_split) {
2701                 err = vma->vm_ops->may_split(vma, addr);
2702                 if (err)
2703                         return err;
2704         }
2705
2706         new = vm_area_dup(vma);
2707         if (!new)
2708                 return -ENOMEM;
2709
2710         if (new_below)
2711                 new->vm_end = addr;
2712         else {
2713                 new->vm_start = addr;
2714                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2715         }
2716
2717         err = vma_dup_policy(vma, new);
2718         if (err)
2719                 goto out_free_vma;
2720
2721         err = anon_vma_clone(new, vma);
2722         if (err)
2723                 goto out_free_mpol;
2724
2725         if (new->vm_file)
2726                 get_file(new->vm_file);
2727
2728         if (new->vm_ops && new->vm_ops->open)
2729                 new->vm_ops->open(new);
2730
2731         if (new_below)
2732                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2733                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2734         else
2735                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2736
2737         /* Success. */
2738         if (!err)
2739                 return 0;
2740
2741         /* Clean everything up if vma_adjust failed. */
2742         if (new->vm_ops && new->vm_ops->close)
2743                 new->vm_ops->close(new);
2744         if (new->vm_file)
2745                 fput(new->vm_file);
2746         unlink_anon_vmas(new);
2747  out_free_mpol:
2748         mpol_put(vma_policy(new));
2749  out_free_vma:
2750         vm_area_free(new);
2751         return err;
2752 }
2753
2754 /*
2755  * Split a vma into two pieces at address 'addr', a new vma is allocated
2756  * either for the first part or the tail.
2757  */
2758 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2759               unsigned long addr, int new_below)
2760 {
2761         if (mm->map_count >= sysctl_max_map_count)
2762                 return -ENOMEM;
2763
2764         return __split_vma(mm, vma, addr, new_below);
2765 }
2766
2767 static inline void
2768 unlock_range(struct vm_area_struct *start, unsigned long limit)
2769 {
2770         struct mm_struct *mm = start->vm_mm;
2771         struct vm_area_struct *tmp = start;
2772
2773         while (tmp && tmp->vm_start < limit) {
2774                 if (tmp->vm_flags & VM_LOCKED) {
2775                         mm->locked_vm -= vma_pages(tmp);
2776                         munlock_vma_pages_all(tmp);
2777                 }
2778
2779                 tmp = tmp->vm_next;
2780         }
2781 }
2782
2783 /* Munmap is split into 2 main parts -- this part which finds
2784  * what needs doing, and the areas themselves, which do the
2785  * work.  This now handles partial unmappings.
2786  * Jeremy Fitzhardinge <jeremy@goop.org>
2787  */
2788 int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2789                 struct list_head *uf, bool downgrade)
2790 {
2791         unsigned long end;
2792         struct vm_area_struct *vma, *prev, *last;
2793
2794         if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2795                 return -EINVAL;
2796
2797         len = PAGE_ALIGN(len);
2798         end = start + len;
2799         if (len == 0)
2800                 return -EINVAL;
2801
2802         /*
2803          * arch_unmap() might do unmaps itself.  It must be called
2804          * and finish any rbtree manipulation before this code
2805          * runs and also starts to manipulate the rbtree.
2806          */
2807         arch_unmap(mm, start, end);
2808
2809         /* Find the first overlapping VMA where start < vma->vm_end */
2810         vma = find_vma_intersection(mm, start, end);
2811         if (!vma)
2812                 return 0;
2813         prev = vma->vm_prev;
2814
2815         /*
2816          * If we need to split any vma, do it now to save pain later.
2817          *
2818          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2819          * unmapped vm_area_struct will remain in use: so lower split_vma
2820          * places tmp vma above, and higher split_vma places tmp vma below.
2821          */
2822         if (start > vma->vm_start) {
2823                 int error;
2824
2825                 /*
2826                  * Make sure that map_count on return from munmap() will
2827                  * not exceed its limit; but let map_count go just above
2828                  * its limit temporarily, to help free resources as expected.
2829                  */
2830                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2831                         return -ENOMEM;
2832
2833                 error = __split_vma(mm, vma, start, 0);
2834                 if (error)
2835                         return error;
2836                 prev = vma;
2837         }
2838
2839         /* Does it split the last one? */
2840         last = find_vma(mm, end);
2841         if (last && end > last->vm_start) {
2842                 int error = __split_vma(mm, last, end, 1);
2843                 if (error)
2844                         return error;
2845         }
2846         vma = vma_next(mm, prev);
2847
2848         if (unlikely(uf)) {
2849                 /*
2850                  * If userfaultfd_unmap_prep returns an error the vmas
2851                  * will remain split, but userland will get a
2852                  * highly unexpected error anyway. This is no
2853                  * different than the case where the first of the two
2854                  * __split_vma fails, but we don't undo the first
2855                  * split, despite we could. This is unlikely enough
2856                  * failure that it's not worth optimizing it for.
2857                  */
2858                 int error = userfaultfd_unmap_prep(vma, start, end, uf);
2859                 if (error)
2860                         return error;
2861         }
2862
2863         /*
2864          * unlock any mlock()ed ranges before detaching vmas
2865          */
2866         if (mm->locked_vm)
2867                 unlock_range(vma, end);
2868
2869         /* Detach vmas from rbtree */
2870         if (!detach_vmas_to_be_unmapped(mm, vma, prev, end))
2871                 downgrade = false;
2872
2873         if (downgrade)
2874                 mmap_write_downgrade(mm);
2875
2876         unmap_region(mm, vma, prev, start, end);
2877
2878         /* Fix up all other VM information */
2879         remove_vma_list(mm, vma);
2880
2881         return downgrade ? 1 : 0;
2882 }
2883
2884 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2885               struct list_head *uf)
2886 {
2887         return __do_munmap(mm, start, len, uf, false);
2888 }
2889
2890 static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2891 {
2892         int ret;
2893         struct mm_struct *mm = current->mm;
2894         LIST_HEAD(uf);
2895
2896         if (mmap_write_lock_killable(mm))
2897                 return -EINTR;
2898
2899         ret = __do_munmap(mm, start, len, &uf, downgrade);
2900         /*
2901          * Returning 1 indicates mmap_lock is downgraded.
2902          * But 1 is not legal return value of vm_munmap() and munmap(), reset
2903          * it to 0 before return.
2904          */
2905         if (ret == 1) {
2906                 mmap_read_unlock(mm);
2907                 ret = 0;
2908         } else
2909                 mmap_write_unlock(mm);
2910
2911         userfaultfd_unmap_complete(mm, &uf);
2912         return ret;
2913 }
2914
2915 int vm_munmap(unsigned long start, size_t len)
2916 {
2917         return __vm_munmap(start, len, false);
2918 }
2919 EXPORT_SYMBOL(vm_munmap);
2920
2921 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2922 {
2923         addr = untagged_addr(addr);
2924         profile_munmap(addr);
2925         return __vm_munmap(addr, len, true);
2926 }
2927
2928
2929 /*
2930  * Emulation of deprecated remap_file_pages() syscall.
2931  */
2932 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2933                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2934 {
2935
2936         struct mm_struct *mm = current->mm;
2937         struct vm_area_struct *vma;
2938         unsigned long populate = 0;
2939         unsigned long ret = -EINVAL;
2940         struct file *file;
2941
2942         pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.rst.\n",
2943                      current->comm, current->pid);
2944
2945         if (prot)
2946                 return ret;
2947         start = start & PAGE_MASK;
2948         size = size & PAGE_MASK;
2949
2950         if (start + size <= start)
2951                 return ret;
2952
2953         /* Does pgoff wrap? */
2954         if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2955                 return ret;
2956
2957         if (mmap_write_lock_killable(mm))
2958                 return -EINTR;
2959
2960         vma = vma_lookup(mm, start);
2961
2962         if (!vma || !(vma->vm_flags & VM_SHARED))
2963                 goto out;
2964
2965         if (start + size > vma->vm_end) {
2966                 struct vm_area_struct *next;
2967
2968                 for (next = vma->vm_next; next; next = next->vm_next) {
2969                         /* hole between vmas ? */
2970                         if (next->vm_start != next->vm_prev->vm_end)
2971                                 goto out;
2972
2973                         if (next->vm_file != vma->vm_file)
2974                                 goto out;
2975
2976                         if (next->vm_flags != vma->vm_flags)
2977                                 goto out;
2978
2979                         if (start + size <= next->vm_end)
2980                                 break;
2981                 }
2982
2983                 if (!next)
2984                         goto out;
2985         }
2986
2987         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2988         prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2989         prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2990
2991         flags &= MAP_NONBLOCK;
2992         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2993         if (vma->vm_flags & VM_LOCKED)
2994                 flags |= MAP_LOCKED;
2995
2996         file = get_file(vma->vm_file);
2997         ret = do_mmap(vma->vm_file, start, size,
2998                         prot, flags, pgoff, &populate, NULL);
2999         fput(file);
3000 out:
3001         mmap_write_unlock(mm);
3002         if (populate)
3003                 mm_populate(ret, populate);
3004         if (!IS_ERR_VALUE(ret))
3005                 ret = 0;
3006         return ret;
3007 }
3008
3009 /*
3010  *  this is really a simplified "do_mmap".  it only handles
3011  *  anonymous maps.  eventually we may be able to do some
3012  *  brk-specific accounting here.
3013  */
3014 static int do_brk_flags(unsigned long addr, unsigned long len, unsigned long flags, struct list_head *uf)
3015 {
3016         struct mm_struct *mm = current->mm;
3017         struct vm_area_struct *vma, *prev;
3018         struct rb_node **rb_link, *rb_parent;
3019         pgoff_t pgoff = addr >> PAGE_SHIFT;
3020         int error;
3021         unsigned long mapped_addr;
3022
3023         /* Until we need other flags, refuse anything except VM_EXEC. */
3024         if ((flags & (~VM_EXEC)) != 0)
3025                 return -EINVAL;
3026         flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3027
3028         mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
3029         if (IS_ERR_VALUE(mapped_addr))
3030                 return mapped_addr;
3031
3032         error = mlock_future_check(mm, mm->def_flags, len);
3033         if (error)
3034                 return error;
3035
3036         /* Clear old maps, set up prev, rb_link, rb_parent, and uf */
3037         if (munmap_vma_range(mm, addr, len, &prev, &rb_link, &rb_parent, uf))
3038                 return -ENOMEM;
3039
3040         /* Check against address space limits *after* clearing old maps... */
3041         if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3042                 return -ENOMEM;
3043
3044         if (mm->map_count > sysctl_max_map_count)
3045                 return -ENOMEM;
3046
3047         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3048                 return -ENOMEM;
3049
3050         /* Can we just expand an old private anonymous mapping? */
3051         vma = vma_merge(mm, prev, addr, addr + len, flags,
3052                         NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
3053         if (vma)
3054                 goto out;
3055
3056         /*
3057          * create a vma struct for an anonymous mapping
3058          */
3059         vma = vm_area_alloc(mm);
3060         if (!vma) {
3061                 vm_unacct_memory(len >> PAGE_SHIFT);
3062                 return -ENOMEM;
3063         }
3064
3065         vma_set_anonymous(vma);
3066         vma->vm_start = addr;
3067         vma->vm_end = addr + len;
3068         vma->vm_pgoff = pgoff;
3069         vma->vm_flags = flags;
3070         vma->vm_page_prot = vm_get_page_prot(flags);
3071         vma_link(mm, vma, prev, rb_link, rb_parent);
3072 out:
3073         perf_event_mmap(vma);
3074         mm->total_vm += len >> PAGE_SHIFT;
3075         mm->data_vm += len >> PAGE_SHIFT;
3076         if (flags & VM_LOCKED)
3077                 mm->locked_vm += (len >> PAGE_SHIFT);
3078         vma->vm_flags |= VM_SOFTDIRTY;
3079         return 0;
3080 }
3081
3082 int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3083 {
3084         struct mm_struct *mm = current->mm;
3085         unsigned long len;
3086         int ret;
3087         bool populate;
3088         LIST_HEAD(uf);
3089
3090         len = PAGE_ALIGN(request);
3091         if (len < request)
3092                 return -ENOMEM;
3093         if (!len)
3094                 return 0;
3095
3096         if (mmap_write_lock_killable(mm))
3097                 return -EINTR;
3098
3099         ret = do_brk_flags(addr, len, flags, &uf);
3100         populate = ((mm->def_flags & VM_LOCKED) != 0);
3101         mmap_write_unlock(mm);
3102         userfaultfd_unmap_complete(mm, &uf);
3103         if (populate && !ret)
3104                 mm_populate(addr, len);
3105         return ret;
3106 }
3107 EXPORT_SYMBOL(vm_brk_flags);
3108
3109 int vm_brk(unsigned long addr, unsigned long len)
3110 {
3111         return vm_brk_flags(addr, len, 0);
3112 }
3113 EXPORT_SYMBOL(vm_brk);
3114
3115 /* Release all mmaps. */
3116 void exit_mmap(struct mm_struct *mm)
3117 {
3118         struct mmu_gather tlb;
3119         struct vm_area_struct *vma;
3120         unsigned long nr_accounted = 0;
3121
3122         /* mm's last user has gone, and its about to be pulled down */
3123         mmu_notifier_release(mm);
3124
3125         if (unlikely(mm_is_oom_victim(mm))) {
3126                 /*
3127                  * Manually reap the mm to free as much memory as possible.
3128                  * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3129                  * this mm from further consideration.  Taking mm->mmap_lock for
3130                  * write after setting MMF_OOM_SKIP will guarantee that the oom
3131                  * reaper will not run on this mm again after mmap_lock is
3132                  * dropped.
3133                  *
3134                  * Nothing can be holding mm->mmap_lock here and the above call
3135                  * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3136                  * __oom_reap_task_mm() will not block.
3137                  *
3138                  * This needs to be done before calling munlock_vma_pages_all(),
3139                  * which clears VM_LOCKED, otherwise the oom reaper cannot
3140                  * reliably test it.
3141                  */
3142                 (void)__oom_reap_task_mm(mm);
3143
3144                 set_bit(MMF_OOM_SKIP, &mm->flags);
3145                 mmap_write_lock(mm);
3146                 mmap_write_unlock(mm);
3147         }
3148
3149         if (mm->locked_vm)
3150                 unlock_range(mm->mmap, ULONG_MAX);
3151
3152         arch_exit_mmap(mm);
3153
3154         vma = mm->mmap;
3155         if (!vma)       /* Can happen if dup_mmap() received an OOM */
3156                 return;
3157
3158         lru_add_drain();
3159         flush_cache_mm(mm);
3160         tlb_gather_mmu_fullmm(&tlb, mm);
3161         /* update_hiwater_rss(mm) here? but nobody should be looking */
3162         /* Use -1 here to ensure all VMAs in the mm are unmapped */
3163         unmap_vmas(&tlb, vma, 0, -1);
3164         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3165         tlb_finish_mmu(&tlb);
3166
3167         /*
3168          * Walk the list again, actually closing and freeing it,
3169          * with preemption enabled, without holding any MM locks.
3170          */
3171         while (vma) {
3172                 if (vma->vm_flags & VM_ACCOUNT)
3173                         nr_accounted += vma_pages(vma);
3174                 vma = remove_vma(vma);
3175                 cond_resched();
3176         }
3177         vm_unacct_memory(nr_accounted);
3178 }
3179
3180 /* Insert vm structure into process list sorted by address
3181  * and into the inode's i_mmap tree.  If vm_file is non-NULL
3182  * then i_mmap_rwsem is taken here.
3183  */
3184 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3185 {
3186         struct vm_area_struct *prev;
3187         struct rb_node **rb_link, *rb_parent;
3188
3189         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3190                            &prev, &rb_link, &rb_parent))
3191                 return -ENOMEM;
3192         if ((vma->vm_flags & VM_ACCOUNT) &&
3193              security_vm_enough_memory_mm(mm, vma_pages(vma)))
3194                 return -ENOMEM;
3195
3196         /*
3197          * The vm_pgoff of a purely anonymous vma should be irrelevant
3198          * until its first write fault, when page's anon_vma and index
3199          * are set.  But now set the vm_pgoff it will almost certainly
3200          * end up with (unless mremap moves it elsewhere before that
3201          * first wfault), so /proc/pid/maps tells a consistent story.
3202          *
3203          * By setting it to reflect the virtual start address of the
3204          * vma, merges and splits can happen in a seamless way, just
3205          * using the existing file pgoff checks and manipulations.
3206          * Similarly in do_mmap and in do_brk_flags.
3207          */
3208         if (vma_is_anonymous(vma)) {
3209                 BUG_ON(vma->anon_vma);
3210                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3211         }
3212
3213         vma_link(mm, vma, prev, rb_link, rb_parent);
3214         return 0;
3215 }
3216
3217 /*
3218  * Copy the vma structure to a new location in the same mm,
3219  * prior to moving page table entries, to effect an mremap move.
3220  */
3221 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3222         unsigned long addr, unsigned long len, pgoff_t pgoff,
3223         bool *need_rmap_locks)
3224 {
3225         struct vm_area_struct *vma = *vmap;
3226         unsigned long vma_start = vma->vm_start;
3227         struct mm_struct *mm = vma->vm_mm;
3228         struct vm_area_struct *new_vma, *prev;
3229         struct rb_node **rb_link, *rb_parent;
3230         bool faulted_in_anon_vma = true;
3231
3232         /*
3233          * If anonymous vma has not yet been faulted, update new pgoff
3234          * to match new location, to increase its chance of merging.
3235          */
3236         if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3237                 pgoff = addr >> PAGE_SHIFT;
3238                 faulted_in_anon_vma = false;
3239         }
3240
3241         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
3242                 return NULL;    /* should never get here */
3243         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3244                             vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3245                             vma->vm_userfaultfd_ctx);
3246         if (new_vma) {
3247                 /*
3248                  * Source vma may have been merged into new_vma
3249                  */
3250                 if (unlikely(vma_start >= new_vma->vm_start &&
3251                              vma_start < new_vma->vm_end)) {
3252                         /*
3253                          * The only way we can get a vma_merge with
3254                          * self during an mremap is if the vma hasn't
3255                          * been faulted in yet and we were allowed to
3256                          * reset the dst vma->vm_pgoff to the
3257                          * destination address of the mremap to allow
3258                          * the merge to happen. mremap must change the
3259                          * vm_pgoff linearity between src and dst vmas
3260                          * (in turn preventing a vma_merge) to be
3261                          * safe. It is only safe to keep the vm_pgoff
3262                          * linear if there are no pages mapped yet.
3263                          */
3264                         VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3265                         *vmap = vma = new_vma;
3266                 }
3267                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3268         } else {
3269                 new_vma = vm_area_dup(vma);
3270                 if (!new_vma)
3271                         goto out;
3272                 new_vma->vm_start = addr;
3273                 new_vma->vm_end = addr + len;
3274                 new_vma->vm_pgoff = pgoff;
3275                 if (vma_dup_policy(vma, new_vma))
3276                         goto out_free_vma;
3277                 if (anon_vma_clone(new_vma, vma))
3278                         goto out_free_mempol;
3279                 if (new_vma->vm_file)
3280                         get_file(new_vma->vm_file);
3281                 if (new_vma->vm_ops && new_vma->vm_ops->open)
3282                         new_vma->vm_ops->open(new_vma);
3283                 vma_link(mm, new_vma, prev, rb_link, rb_parent);
3284                 *need_rmap_locks = false;
3285         }
3286         return new_vma;
3287
3288 out_free_mempol:
3289         mpol_put(vma_policy(new_vma));
3290 out_free_vma:
3291         vm_area_free(new_vma);
3292 out:
3293         return NULL;
3294 }
3295
3296 /*
3297  * Return true if the calling process may expand its vm space by the passed
3298  * number of pages
3299  */
3300 bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3301 {
3302         if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3303                 return false;
3304
3305         if (is_data_mapping(flags) &&
3306             mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3307                 /* Workaround for Valgrind */
3308                 if (rlimit(RLIMIT_DATA) == 0 &&
3309                     mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3310                         return true;
3311
3312                 pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3313                              current->comm, current->pid,
3314                              (mm->data_vm + npages) << PAGE_SHIFT,
3315                              rlimit(RLIMIT_DATA),
3316                              ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3317
3318                 if (!ignore_rlimit_data)
3319                         return false;
3320         }
3321
3322         return true;
3323 }
3324
3325 void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3326 {
3327         mm->total_vm += npages;
3328
3329         if (is_exec_mapping(flags))
3330                 mm->exec_vm += npages;
3331         else if (is_stack_mapping(flags))
3332                 mm->stack_vm += npages;
3333         else if (is_data_mapping(flags))
3334                 mm->data_vm += npages;
3335 }
3336
3337 static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3338
3339 /*
3340  * Having a close hook prevents vma merging regardless of flags.
3341  */
3342 static void special_mapping_close(struct vm_area_struct *vma)
3343 {
3344 }
3345
3346 static const char *special_mapping_name(struct vm_area_struct *vma)
3347 {
3348         return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3349 }
3350
3351 static int special_mapping_mremap(struct vm_area_struct *new_vma)
3352 {
3353         struct vm_special_mapping *sm = new_vma->vm_private_data;
3354
3355         if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3356                 return -EFAULT;
3357
3358         if (sm->mremap)
3359                 return sm->mremap(sm, new_vma);
3360
3361         return 0;
3362 }
3363
3364 static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3365 {
3366         /*
3367          * Forbid splitting special mappings - kernel has expectations over
3368          * the number of pages in mapping. Together with VM_DONTEXPAND
3369          * the size of vma should stay the same over the special mapping's
3370          * lifetime.
3371          */
3372         return -EINVAL;
3373 }
3374
3375 static const struct vm_operations_struct special_mapping_vmops = {
3376         .close = special_mapping_close,
3377         .fault = special_mapping_fault,
3378         .mremap = special_mapping_mremap,
3379         .name = special_mapping_name,
3380         /* vDSO code relies that VVAR can't be accessed remotely */
3381         .access = NULL,
3382         .may_split = special_mapping_split,
3383 };
3384
3385 static const struct vm_operations_struct legacy_special_mapping_vmops = {
3386         .close = special_mapping_close,
3387         .fault = special_mapping_fault,
3388 };
3389
3390 static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3391 {
3392         struct vm_area_struct *vma = vmf->vma;
3393         pgoff_t pgoff;
3394         struct page **pages;
3395
3396         if (vma->vm_ops == &legacy_special_mapping_vmops) {
3397                 pages = vma->vm_private_data;
3398         } else {
3399                 struct vm_special_mapping *sm = vma->vm_private_data;
3400
3401                 if (sm->fault)
3402                         return sm->fault(sm, vmf->vma, vmf);
3403
3404                 pages = sm->pages;
3405         }
3406
3407         for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3408                 pgoff--;
3409
3410         if (*pages) {
3411                 struct page *page = *pages;
3412                 get_page(page);
3413                 vmf->page = page;
3414                 return 0;
3415         }
3416
3417         return VM_FAULT_SIGBUS;
3418 }
3419
3420 static struct vm_area_struct *__install_special_mapping(
3421         struct mm_struct *mm,
3422         unsigned long addr, unsigned long len,
3423         unsigned long vm_flags, void *priv,
3424         const struct vm_operations_struct *ops)
3425 {
3426         int ret;
3427         struct vm_area_struct *vma;
3428
3429         vma = vm_area_alloc(mm);
3430         if (unlikely(vma == NULL))
3431                 return ERR_PTR(-ENOMEM);
3432
3433         vma->vm_start = addr;
3434         vma->vm_end = addr + len;
3435
3436         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3437         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3438
3439         vma->vm_ops = ops;
3440         vma->vm_private_data = priv;
3441
3442         ret = insert_vm_struct(mm, vma);
3443         if (ret)
3444                 goto out;
3445
3446         vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3447
3448         perf_event_mmap(vma);
3449
3450         return vma;
3451
3452 out:
3453         vm_area_free(vma);
3454         return ERR_PTR(ret);
3455 }
3456
3457 bool vma_is_special_mapping(const struct vm_area_struct *vma,
3458         const struct vm_special_mapping *sm)
3459 {
3460         return vma->vm_private_data == sm &&
3461                 (vma->vm_ops == &special_mapping_vmops ||
3462                  vma->vm_ops == &legacy_special_mapping_vmops);
3463 }
3464
3465 /*
3466  * Called with mm->mmap_lock held for writing.
3467  * Insert a new vma covering the given region, with the given flags.
3468  * Its pages are supplied by the given array of struct page *.
3469  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3470  * The region past the last page supplied will always produce SIGBUS.
3471  * The array pointer and the pages it points to are assumed to stay alive
3472  * for as long as this mapping might exist.
3473  */
3474 struct vm_area_struct *_install_special_mapping(
3475         struct mm_struct *mm,
3476         unsigned long addr, unsigned long len,
3477         unsigned long vm_flags, const struct vm_special_mapping *spec)
3478 {
3479         return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3480                                         &special_mapping_vmops);
3481 }
3482
3483 int install_special_mapping(struct mm_struct *mm,
3484                             unsigned long addr, unsigned long len,
3485                             unsigned long vm_flags, struct page **pages)
3486 {
3487         struct vm_area_struct *vma = __install_special_mapping(
3488                 mm, addr, len, vm_flags, (void *)pages,
3489                 &legacy_special_mapping_vmops);
3490
3491         return PTR_ERR_OR_ZERO(vma);
3492 }
3493
3494 static DEFINE_MUTEX(mm_all_locks_mutex);
3495
3496 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3497 {
3498         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3499                 /*
3500                  * The LSB of head.next can't change from under us
3501                  * because we hold the mm_all_locks_mutex.
3502                  */
3503                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3504                 /*
3505                  * We can safely modify head.next after taking the
3506                  * anon_vma->root->rwsem. If some other vma in this mm shares
3507                  * the same anon_vma we won't take it again.
3508                  *
3509                  * No need of atomic instructions here, head.next
3510                  * can't change from under us thanks to the
3511                  * anon_vma->root->rwsem.
3512                  */
3513                 if (__test_and_set_bit(0, (unsigned long *)
3514                                        &anon_vma->root->rb_root.rb_root.rb_node))
3515                         BUG();
3516         }
3517 }
3518
3519 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3520 {
3521         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3522                 /*
3523                  * AS_MM_ALL_LOCKS can't change from under us because
3524                  * we hold the mm_all_locks_mutex.
3525                  *
3526                  * Operations on ->flags have to be atomic because
3527                  * even if AS_MM_ALL_LOCKS is stable thanks to the
3528                  * mm_all_locks_mutex, there may be other cpus
3529                  * changing other bitflags in parallel to us.
3530                  */
3531                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3532                         BUG();
3533                 down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3534         }
3535 }
3536
3537 /*
3538  * This operation locks against the VM for all pte/vma/mm related
3539  * operations that could ever happen on a certain mm. This includes
3540  * vmtruncate, try_to_unmap, and all page faults.
3541  *
3542  * The caller must take the mmap_lock in write mode before calling
3543  * mm_take_all_locks(). The caller isn't allowed to release the
3544  * mmap_lock until mm_drop_all_locks() returns.
3545  *
3546  * mmap_lock in write mode is required in order to block all operations
3547  * that could modify pagetables and free pages without need of
3548  * altering the vma layout. It's also needed in write mode to avoid new
3549  * anon_vmas to be associated with existing vmas.
3550  *
3551  * A single task can't take more than one mm_take_all_locks() in a row
3552  * or it would deadlock.
3553  *
3554  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3555  * mapping->flags avoid to take the same lock twice, if more than one
3556  * vma in this mm is backed by the same anon_vma or address_space.
3557  *
3558  * We take locks in following order, accordingly to comment at beginning
3559  * of mm/rmap.c:
3560  *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3561  *     hugetlb mapping);
3562  *   - all i_mmap_rwsem locks;
3563  *   - all anon_vma->rwseml
3564  *
3565  * We can take all locks within these types randomly because the VM code
3566  * doesn't nest them and we protected from parallel mm_take_all_locks() by
3567  * mm_all_locks_mutex.
3568  *
3569  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3570  * that may have to take thousand of locks.
3571  *
3572  * mm_take_all_locks() can fail if it's interrupted by signals.
3573  */
3574 int mm_take_all_locks(struct mm_struct *mm)
3575 {
3576         struct vm_area_struct *vma;
3577         struct anon_vma_chain *avc;
3578
3579         BUG_ON(mmap_read_trylock(mm));
3580
3581         mutex_lock(&mm_all_locks_mutex);
3582
3583         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3584                 if (signal_pending(current))
3585                         goto out_unlock;
3586                 if (vma->vm_file && vma->vm_file->f_mapping &&
3587                                 is_vm_hugetlb_page(vma))
3588                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3589         }
3590
3591         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3592                 if (signal_pending(current))
3593                         goto out_unlock;
3594                 if (vma->vm_file && vma->vm_file->f_mapping &&
3595                                 !is_vm_hugetlb_page(vma))
3596                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
3597         }
3598
3599         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3600                 if (signal_pending(current))
3601                         goto out_unlock;
3602                 if (vma->anon_vma)
3603                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3604                                 vm_lock_anon_vma(mm, avc->anon_vma);
3605         }
3606
3607         return 0;
3608
3609 out_unlock:
3610         mm_drop_all_locks(mm);
3611         return -EINTR;
3612 }
3613
3614 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3615 {
3616         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3617                 /*
3618                  * The LSB of head.next can't change to 0 from under
3619                  * us because we hold the mm_all_locks_mutex.
3620                  *
3621                  * We must however clear the bitflag before unlocking
3622                  * the vma so the users using the anon_vma->rb_root will
3623                  * never see our bitflag.
3624                  *
3625                  * No need of atomic instructions here, head.next
3626                  * can't change from under us until we release the
3627                  * anon_vma->root->rwsem.
3628                  */
3629                 if (!__test_and_clear_bit(0, (unsigned long *)
3630                                           &anon_vma->root->rb_root.rb_root.rb_node))
3631                         BUG();
3632                 anon_vma_unlock_write(anon_vma);
3633         }
3634 }
3635
3636 static void vm_unlock_mapping(struct address_space *mapping)
3637 {
3638         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3639                 /*
3640                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3641                  * because we hold the mm_all_locks_mutex.
3642                  */
3643                 i_mmap_unlock_write(mapping);
3644                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3645                                         &mapping->flags))
3646                         BUG();
3647         }
3648 }
3649
3650 /*
3651  * The mmap_lock cannot be released by the caller until
3652  * mm_drop_all_locks() returns.
3653  */
3654 void mm_drop_all_locks(struct mm_struct *mm)
3655 {
3656         struct vm_area_struct *vma;
3657         struct anon_vma_chain *avc;
3658
3659         BUG_ON(mmap_read_trylock(mm));
3660         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3661
3662         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3663                 if (vma->anon_vma)
3664                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3665                                 vm_unlock_anon_vma(avc->anon_vma);
3666                 if (vma->vm_file && vma->vm_file->f_mapping)
3667                         vm_unlock_mapping(vma->vm_file->f_mapping);
3668         }
3669
3670         mutex_unlock(&mm_all_locks_mutex);
3671 }
3672
3673 /*
3674  * initialise the percpu counter for VM
3675  */
3676 void __init mmap_init(void)
3677 {
3678         int ret;
3679
3680         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3681         VM_BUG_ON(ret);
3682 }
3683
3684 /*
3685  * Initialise sysctl_user_reserve_kbytes.
3686  *
3687  * This is intended to prevent a user from starting a single memory hogging
3688  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3689  * mode.
3690  *
3691  * The default value is min(3% of free memory, 128MB)
3692  * 128MB is enough to recover with sshd/login, bash, and top/kill.
3693  */
3694 static int init_user_reserve(void)
3695 {
3696         unsigned long free_kbytes;
3697
3698         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3699
3700         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3701         return 0;
3702 }
3703 subsys_initcall(init_user_reserve);
3704
3705 /*
3706  * Initialise sysctl_admin_reserve_kbytes.
3707  *
3708  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3709  * to log in and kill a memory hogging process.
3710  *
3711  * Systems with more than 256MB will reserve 8MB, enough to recover
3712  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3713  * only reserve 3% of free pages by default.
3714  */
3715 static int init_admin_reserve(void)
3716 {
3717         unsigned long free_kbytes;
3718
3719         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3720
3721         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3722         return 0;
3723 }
3724 subsys_initcall(init_admin_reserve);
3725
3726 /*
3727  * Reinititalise user and admin reserves if memory is added or removed.
3728  *
3729  * The default user reserve max is 128MB, and the default max for the
3730  * admin reserve is 8MB. These are usually, but not always, enough to
3731  * enable recovery from a memory hogging process using login/sshd, a shell,
3732  * and tools like top. It may make sense to increase or even disable the
3733  * reserve depending on the existence of swap or variations in the recovery
3734  * tools. So, the admin may have changed them.
3735  *
3736  * If memory is added and the reserves have been eliminated or increased above
3737  * the default max, then we'll trust the admin.
3738  *
3739  * If memory is removed and there isn't enough free memory, then we
3740  * need to reset the reserves.
3741  *
3742  * Otherwise keep the reserve set by the admin.
3743  */
3744 static int reserve_mem_notifier(struct notifier_block *nb,
3745                              unsigned long action, void *data)
3746 {
3747         unsigned long tmp, free_kbytes;
3748
3749         switch (action) {
3750         case MEM_ONLINE:
3751                 /* Default max is 128MB. Leave alone if modified by operator. */
3752                 tmp = sysctl_user_reserve_kbytes;
3753                 if (0 < tmp && tmp < (1UL << 17))
3754                         init_user_reserve();
3755
3756                 /* Default max is 8MB.  Leave alone if modified by operator. */
3757                 tmp = sysctl_admin_reserve_kbytes;
3758                 if (0 < tmp && tmp < (1UL << 13))
3759                         init_admin_reserve();
3760
3761                 break;
3762         case MEM_OFFLINE:
3763                 free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3764
3765                 if (sysctl_user_reserve_kbytes > free_kbytes) {
3766                         init_user_reserve();
3767                         pr_info("vm.user_reserve_kbytes reset to %lu\n",
3768                                 sysctl_user_reserve_kbytes);
3769                 }
3770
3771                 if (sysctl_admin_reserve_kbytes > free_kbytes) {
3772                         init_admin_reserve();
3773                         pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3774                                 sysctl_admin_reserve_kbytes);
3775                 }
3776                 break;
3777         default:
3778                 break;
3779         }
3780         return NOTIFY_OK;
3781 }
3782
3783 static struct notifier_block reserve_mem_nb = {
3784         .notifier_call = reserve_mem_notifier,
3785 };
3786
3787 static int __meminit init_reserve_notifier(void)
3788 {
3789         if (register_hotmemory_notifier(&reserve_mem_nb))
3790                 pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3791
3792         return 0;
3793 }
3794 subsys_initcall(init_reserve_notifier);