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